IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO.

XX, FIRST QUARTER 2023 1

Localization as a key enabler of 6G wireless

systems: A comprehensive survey and an outlook
Stylianos E. Trevlakis, Member, IEEE, Alexandros-Apostolos A. Boulogeorgos, Senior Member, IEEE, Dimitrios
Pliatsios, Member, IEEE, Konstantinos Ntontin, Member, IEEE, Panagiotis Sarigiannidis, Member, IEEE, Symeon
Chatzinotas, Fellow, IEEE, and Marco Di Renzo, Fellow, IEEE,

Abstract—When fully implemented, sixth generation (6G) of the radio and microwave bands [18]–[21], the research,
wireless systems will constitute intelligent wireless networks that innovation and industrial communities turned its attention
arXiv:2302.02107v1 [eess.SP] 4 Feb 2023

enable not only ubiquitous communication but also high-accuracy to communications in the millimeter wave (mmWave) [22]–
localization services. They will be the driving force behind this
transformation by introducing a new set of characteristics and [26], terahertz (THz) [27]–[30], and optical bands [31]–[35].
service capabilities in which location will coexist with commu- Despite the unprecedented bandwidth that high-frequency sys-
nication while sharing available resources. To that purpose, this tems offer, in order to achieve the promised performance
survey investigates the envisioned applications and use cases of excellence in terms of throughput, latency, and reliability, in
localization in future 6G wireless systems, while analyzing the acceptable transmission distances, both the transmitter (TX)
impact of the major technology enablers. Afterwards, system
models for millimeter wave, terahertz and visible light positioning and the receiver (RX) require knowledge of each others
that take into account both line-of-sight (LOS) and non-LOS relative position and orientation [36]–[40]. As a consequence,
channels are presented, while localization key performance indi- localization is expected to become a vital component of high-
cators are revisited alongside mathematical definitions. Moreover, frequency wireless systems.
a detailed review of the state of the art conventional and learning- Scanning the technical literature, several surveys and tu-
based localization techniques is conducted. Furthermore, the
localization problem is formulated, the wireless system design is torials for localization can be identified [41]–[60]. In more
considered and the optimization of both is investigated. Finally, detail, in [41], the authors presented an overview of the
insights that arise from the presented analysis are summarized simultaneously localization and mapping approaches in self-
and used to highlight the most important future directions for driving vehicles. In [42], the authors reported the requirements,
localization in 6G wireless systems. technology enablers, techniques and approaches, as well as
Index Terms—6G, applications, future directions, key perfor- existing system models and their performance in indoor lo-
mance indicators, localization, machine learning, methodologies, calization for emergency response scenarios. In [43], a state-
optimization, use-cases. of-the-art (SotA) vehicle localization approaches comparison
and assessment was conducted and their role to vehicle-to-
I. I NTRODUCTION everything scenarios was discussed. A list of the enabling loca-
lization technologies that were developed for cellular systems
To satisfy the proliferating demands of next generation and wireless local area networks until 2018 was documented
wireless applications, such as multisensory extended reality in [44]. In [45], the authors presented an overview of the
(XR) [1]–[3], connected robots [4]–[6], wireless brain com- standardized localization methods from the first to the fifth
puter interactions [7], digital twins [8], industrial internet-of- generation cellular systems. In [46], the authors revisited and
things (IoT) [9], tactile IoT [10]–[12], internet of underwater discussed visible light positioning (VLP) approaches.
things [13], [14], self-driving ground and air vehicles [15]– The authors of [47] conducted an overview of the device-
[17], and others, while dealing with the spectrum scarcity free localization approaches and presented the 2019 research
and development trends. In [48], the authors surveyed the
Manuscript received January XX, 2023; revised XX; accepted XX
S. E. Trevlakis is with the Department of Research and Development, 2019 indoor localization systems and methods accounting
InnoCube P.C., 17th Noemvriou 79, 55535 Thessaloniki, Greece (email: for different radio technologies. In [49], the authors revis-
[email protected]) ited and explained multidimensional scaling (MDS)-based
A.-A. A. Boulogeorgos, D. Pliatsios, P. Sarigiannidis are with the Depart-
ment of Electrical and Computer Engineering, University of Western Mace- localization techniques. In [50], a review of the SotA of
donia, ZEP Area, 50100 Kozani, Greece (e-mail: [email protected]; localization techniques for massive multiple-input multiple-
[email protected]; [email protected]) output (MIMO) systems was conducted. Moreover, the authors
K. Ntontin and S. Chatzinotas are with the SnT, University
of Luxembourg, Luxembourg. E-mails: [email protected], of [51] reviewed machine learning (ML)-based localization
[email protected]. methodologies that employ radio signals. In [52], a summation
M. Di Renzo is with Université Paris-Saclay, CNRS, CentraleSupélec, of ML-enabled indoor localization approaches was reported.
Laboratoire des Signaux et Systèmes, 3 Rue Joliot-Curie, 91192 Gif-sur-
Yvette, France. ([email protected]) The authors of [53] presented a survey of radio frequency
This work has received funding from the European Unions Horizon- identifier (RFID) based localization techniques. In [54], the
CL4-2021 research and innovation programme under grant agreement No. authors focused on the identification of key enabling technolo-
101070181 (TALON) as well as the the European Unions HORIZON-JU-
SNS-2022 research and innovation programme under grant agreement No. gies and applications of localization in the sixth generation
101096456 (NANCY). (6G) networks. In [55], the vision of cm-level localization
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 2

was presented and map-based approaches was identified as Introduction

possible enablers. In [56], a survey on localization fundamental
approaches accompanied by SotA results and future directions Applications of localization in 6G wireless systems
was reported. The authors of [57] focused on the recent de-
Sustainable development
velopments and applications of simultaneous localization and Massive twinning
Immersive telepresence
mapping (SLAM) with emphasis in complex and unstructured Robots
Local trust zones
agricultural environments. In [58], the authors overviewed
a number of localization methods for autonomous vehicles. Localization use cases
The authors of [59] conducted a survey on the fundamental Indoor
limits of integrated communication and localization systems. Underwater
Outdoor
Finally, in [60], the authors provided a survey concerning THz Nanoscale
Non-terrestrial
empowered localization techniques.
Tables I and II summarize the surveys existing in literature Technology enablers for 6G wireless systems
that investigate localization applications and use cases as well mmWave/THz localization
Channel charting
as the relevant technologies, system models, key performance Hardware
Radars
indicators (KPIs) and approaches. However, the focus and aim Beamforming
Sensors
of each survey is different both in terms of use cases, technolo- RIS
gies and methods enlisted. For instance, almost every survey VLP
AI
includes indoor and outdoor terrestrial use cases [41]–[60],
System models
while only [49], [56], [59] take into account non-terrestrial
networks (NTN) and only [49] investigates underwater use mmWave/THz VLP
cases. As far as technology enablers are concerned, each con- Localization KPIs
tribution focuses on specific subsets of them. Specifically, the
majority of the surveys before 2020 take into account sensors, Accuracy Coverage
radars, and cellular systems [41]–[43], [45]–[49], while the Precision Complexity
contrary is valid for contributions between 2020 and 2022 that Latency Stability
mainly investigate mmWave/THz technologies [54]–[56], [59], Localization methodologies/algorithms
[60]. In addition, enablers like visible light positioning (VLP),
Conventional methods Federated learning
beamforming, channel charting and reconfigurable intelligent
surfaces (RISs) are taken into consideration in a very small Supervised learning Reinforcement learning

number of surveys. From a system model point of view, only Unsupervised learning Transfer learning

selected contributions provide mathematical modelling for the Optimization frameworks for localization
discussed technologies, such as [44], [46], [50], [55], [59],
Problem formulation
[60]. From these tables, it becomes evident that a complete Optimization
System design
survey of the applications, use cases, technology enablers,
system models, KPIs, and methods with focus on the future 6G Insights & future directions
wireless networks is missing from the bibliography. To cover mmWave and THz localization Heterogeneous localization
this gap and with respect to the aforementioned contributions, Optical wireless localization Integrated localization and
the overall objective of this survey is to deliver a timely communication
RIS-assisted localization
and comprehensive review of the localization applications,
use cases, system models, enabling technologies, methods and Conclusion
KPIs in the 6G era as well as to identify research gaps and
highlight possible research directions. In particular, this survey Fig. 1. The overall structure of this survey.
investigates the following:

• The main localization-specific applications and use cases

methodologies/algorithms are investigated and conceptual
of future 6G wireless networks are presented.
figures are provided to aid the reader in understanding
• The localization-enabling technologies are investigated
their basic principles.
with regard to their contribution towards advancing the
• The formulation, system design and optimization of the
SotA of wireless systems.
localization problem is described.
• System models for line-of-sight (LOS) and non-LOS
• Directions and insights for the design and improvement
(NLOS) channels in mmWave/THz and VLP localization
of future 6G wireless systems are discussed.
wireless systems are analyzed taking into account both
classic reflectors and RISs. The reminder of the survey article is organized as follows:
• The major KPIs used by localization methods to measure Section II describes the applications of 6G wireless systems,
the performance of the system are discussed and accom- while Section III delves into the use cases and their relations
panied by mathematical expressions. with the various 6G applications. Next, the technology en-
• The SotA of conventional and learning-based localization ablers of localization are presented in Section IV. Section V
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 3

TABLE I
S UMMARY OF RECENT LOCALIZATION SURVEYS AND TUTORIALS (2017-2019)

Year Ref. Applications Use-cases Technologies System models KPIs Algorithms

2017 [41] Autonomous vehi- Outdoor Sensors – Accuracy, scalability, SLAM
cle availability, recovery,
updatability, dynam-
icity
2017 [42] Emergency Indoor Radio – Accuracy, informa- Proximity, Triangulation, Latera-
response Outdoor signals, tion accessibility, tion/ trilateration/ multilateration,
Inertial adaptability, Maximum likelihood estimation,
measurement scalability, physical Dead reckoning, Kalman filters,
units (IMUs) robustness, Centroid, Particle filters, Finger-
assembling printing, Visual analysis, Map
complexity, matching
equipment’s size
and weight, energy
efficiency, cost
2018 [43] Autonomous vehi- Outdoor GPS/IMU, – Accuracy, Dead Differential GPS, Assisted GPS,
cles Sensors, reckoning, Position Real time kinematic based tech-
Radio and error, Odometry, niques, SLAM, Random sample
microwave Packet loss, consensus, Frequency modulation
signals, VLP Robustness, Time-to- continuous wave (FMCW)-based
first-fix solutions
2018 [44] Network planing Indoor Cellular, Cellular Accuracy Probabilistic radio signal strength
and optimization, Outdoor WLANs, (RSS) fingerprint matching, Semi-
Emergency Sensors, supervised and unsupervised learn-
response mmWave ing, Cell-identifier and GPS po-
sition sequence matching, Hidden
Markov model, Bayesian learning,
Assisted GPS, Advanced forward
link trilateration, RF pattern match-
ing, Intelligent probability hierar-
chy based solutions, SLAM
2018 [45] – Indoor Radio – Accuracy Cell-ID, RF pattern matching,
Outdoor signals, Time of arrival (ToA) and time
WLAN/ difference of arrival (TDoA) based
Bluetooth, methods, Advanced forward link
Sensors trilateration, Stand-alone, Differen-
tial, and Assisted GNSS
2018 [46] – Indoor VLP VLP Cramér-Rao lower Direct positioning, proximity,
bound (CRLB), Root geometric- and statistical-
mean square error, based approaches, fingerprinting,
Power efficiency mapping
2019 [47] Smart agriculture, Indoor Radio – Static and dynamic Directional and ambient radio
Smart grid, Smart Outdoor signals, backscattering model imaging, channel diversity,
media, Smart cities, Sensors Shadowing, Comprehensive
Social computing, sensing, Radio grid, Diffraction
Social IoT, e- theory, Extreme learning
Business, Affective machine, Markov model,
computing, Probabilistic classification,
Cyber physical Gradient fingerprinting, Support
systems, Wearable vector machine (SVM), Deep
applications learning, Dictionary learning
2019 [50] – Indoor mmWave mmWave Accuracy, Complex- Indirect localization, Direct locali-
Outdoor MIMO ity zation, Fingerprinting
2019 [48] Personalized Indoor WiFi, – Availability, Cost, En- Fingerprinting/scene analysis,
context-aware Bluetooth, ergy efficiency, Re- RSS, Channel state information
networks, e- Zigbee, ception range, Accu- (CSI), Angel of arrival (AoA),
Health, Disaster RFID, VLP, racy, Latency, Scala- Time of flight (ToF), TDoA,
management and Acoustic, bility Return time of flight (RFoF), and
recovery, Security, Ultrasound Phase of arrival-based approaches
Asset management
and tracking
2019 [49] Disaster Indoor Cellular, – Accuracy, Complex- MDS
management, Outdoor RFID, ity
Security, IoT, NTN Sensors,
Underwater Underwater Acoustic,
exploration VLP
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 4

TABLE II
S UMMARY OF RECENT LOCALIZATION SURVEYS AND TUTORIALS (2020-2022)

Year Ref. Applications Use-cases Technologies System models KPIs Algorithms

2020 [51] Autonomous Indoor WiFi, – Accuracy K-nearest neighbors (KNN), Gaus-
vehicles, Mission- Outdoor Cellular, sian processes, Ensemble methods,
critical applications, Bluetooth, Neural networks (NNs), Convolu-
IoT, Beam Sensors tion NNs (CNNs), Recurrent NNs
alignment in (RNNs), Autoencoders, Generative
massive MIMO and statistical models, Deep belief
systems networks, MDS, Tranfer learning
2020 [52] Positioning and Indoor WiFi, – Accuracy, latency, Fingerprint
navigation of Bluetooth, complexity, coverage,
indoor environment, VLP, robustness
Nursing people and Magnetic
tracking, People field
management, Fire
rescue and other
safety needs
2020 [61], [62] Social distancing Indoor WiFi, GNSS, – Accuracy, latency, Trilateration, Kalman filters, RSS-
Outdoor cellular, coverage based, TDoA-based, Assisted-
NTN bluetooth, GNSS, Enhanced Cell-ID
UWB, RFID
2021 [53] Vehicle localization, Indoor RFID, Sen- – Accuracy, cost, en- –
navigation, tracking sors ergy efficiency, com-
plexity
2021 [54] Environment Indoor mmWave, – – Supervised, unsupervised,
mapping, Robot Outdoor THz, Re- reinforcement learning
localization, configurable
Tracking, intelligent
Localization surfaces
and sensing for (RISs)
eHealth, Context
awareness, Radar-
based applications
2021 [55] People, objects, and Indoor mmWave, mmWave Accuracy Fingerprinting, Kalman filter, Ex-
vehicles monitoring Outdoor Cellular tended Kalman filter
2021 [63] Nanonetworks Nanoscale THz – Accuraxy, complexity –
2022 [56] Mutisensory Indoor mmWave, – Accuracy, latency, Tracking, SLAM, NNs Extednde
XR, Tele- Outdoor THz, precision, energy Kalman filter
presence, Smart NTN Sensors efficiency
transportation,
Connected robotics
and autonomous
systems, UAVs
2022 [57] Precision Outdoor Sensors – Accuracy Extended Kalman filter SLAM,
agriculture, Monte Carlo localization, Visual
mapping, navigation SLAM, Sensor fusion SLAM
2022 [58] Autonomous vehi- Outdoor Sensors – Accuracy, F1 Score Extended Kalman filters, NNs
cles
2022 [59] IoT connected Indoor WiFi, Cooperative Accuracy, coverage, –
home, Internet Outdoor mmWave and non- complexity, stability
of vehicles, UAV NTN cooperative
communications localization
and navigation, with MIMO
Target detection and virtual
MIMO
2022 [60] – Indoor mmWave, THz with and Accuracy, Coverage, ToA-based, TDoA-based, AoA-
Outdoor THz without RISs Latency, Update rate, based, AoD-based, Expected max-
Stability, Scalability, imization, NNs, Kalman filter, Par-
Mobility, Complexity ticle filter
This work Autonomous supply Indoor mmWave, mmWave/THz, Accuracy, precision, Triangulation, Kalman filters,
chain, Smart cities, Outdoor THz, VLP latency, coverage, Compressive sensing, MDS, Direct
Manufacturing, XR, NTN hardware, complexity, stability localization, Swarm intelligence,
Earth monitoring, Underwater beamform- Fingerprinting, SLAM, kNN,
Network expansion, Nanoscale ing, RIS, SVM, Decision trees, Gaussian
Mapping, Sensor AI, channel processes, NNs, Autoencoders,
infrastructure web, charting, CNNs, RNNs, Unsupervised
Context-aware radars, learning, Semi-supervised learning,
networks, Precision sensors, clustering, Dimensionality
healthcare, Security, VLP reduction, Federated learning,
Gesture/motion Reinforcement learning (RL),
recognition, Robots, Deep RL, Transfer learning,
Nanoscale, e- Manifold learning, Cooperative
Health, Agriculture approaches
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 5

focuses on the system models for mmWave/THz localization A. Sustainable development

and VLP, while the various KPIs of the localization approaches
are presented in Section VI. Section VII describes the SotA In order to achieve the United Nations’ (UN) sustainable
conventional and learning-based localization methodologies development goals (SDGs) and lessen the environmental im-
and algorithms that are envisioned to be applied in future 6G pact, the next generation of radio access technology (NG-RAT)
wireless networks. Section VIII highlights the major aspects of promises to incorporate measures to protect the environment
the localization problem formulation, systems design as well and empower people to improve their own lives and the world
as its optimization. Finally, in Section IX, the findings of this around them. NG-RAT is planned to serve clusters of use cases
contribution are translated into insights and future directions, such as autonomous supply chain, earth monitoring, e-health,
while concluding remarks are provided in Section X. The and personalized context-aware networks in order to meet the
structure of this survey in a glance is reported in Figure 1. major UN SDG. All these use cases have a common need for
localization, which in turn have a set of criteria on major KPIs
Abbreviations and notations: The abbreviations that can
that must be satisfied.
be found throughout this survey are presented in Table III,
given at the top of the next page. Unless stated otherwise, 1) Autonomous supply chain: With the ability to monitor
bold low and capital letters respectively denote vectors and products from the time they are created all the way through
matrices, (·)T stands for the transpose matrix, | · | denotes their final stages of use and disposal, an autonomous supply
absolute value, ex represents the exponential function, (·)−1 chain may guarantee material and energy savings as well as
is the inverse matrix, while ab is a in the power of b. In increased efficiency. Specifically, 6G-enabled micro-tags may
addition, cos(·), sin(·), arctan 2(·), and arcsin(·) represent the be used to (i) promote a flexible and adaptable supply chain;
cosine, sine, 2-argument arctangent, and the arcsine functions, (ii) track and record the whereabouts and condition of all
respectively. Also, log2 (·) stands for the binary logarithm, products; (iii) allow a decentralised, worldwide asset monitor-
Hilb[·] represents the Hilbert function, and E[·] denotes the ing and management system; as well as (iv) link all products
expected value. Moreover, tr(·) is the trace of a matrix and to the internet [64], [65]. For example, consequences for the
I −1 (·) denotes the inverse Fisher matrix. argmin(·) represents supply chain sector may be seen in the novel approaches
the argument of minimum, diag(·) is the diagonal to final-mile delivery made possible by the widespread use
√ matrix, and of computing and mobile communication technologies. Up
rank(·) denotes the rank of a matrix. Finally, · represents
PN
the root and i=1 xi is the sum of xi with i in the range of to half of total distribution expenses are attributable to the
[1, N ]. last mile of the supply chain’s journey. Packages spend much
of their time in transit dead-zone locations, and very little
real-time information is available, as they make their way
II. A PPLICATIONS OF LOCALIZATION IN 6G WIRELESS from a site of storage, like a warehouse to a point of sale
SYSTEMS or straight to the premises of ultimate costumers. Lastly,
a conventional supply chain is unable to meet the demand
Localization has been completely transformed by the ad- for instantaneous delivery for online purchases. Therefore,
vancement of multiple technologies, such as radars, sensors, the management of supply chains has placed a premium on
AI, mmWave, THz, VLC, and RISs. As illustrated in Fig. 2. It innovative technical solutions. Thus, the ongoing maturation of
is expected to play a dominant role in the next generation of technologies based on 6G wireless systems and fog computing
communication networks through its incorporation to a variety will unquestionably provide a plethora of chances for the next
of applications. Some indicative examples include pinpointing wave of technological innovation [66]. The latter may provide
a device’s outdoor location with cm- to mm-level precision, a wide variety of services due to the use of VNFs, SDN, AI,
accurate indoor localization with reliable interpretation of 3D and resource orchestration that works with both physical and
data in addition to the 2D data that are presently available, or virtual resources. This will pave the way for an ideal answer
the incorporation of roll, pitch, and yaw into the localization in terms of logistics management, protocol compatibility, and
process. To achieve these, 6G promises to use a combination of network architecture for the next-generation delivery systems,
radio and VLC localization to achieve cm- to mm-level accu- which rely on UAVs and robot delivery workers.
racy whether the user, object, or device is inside and/or outside. 2) Earth monitoring: The UN SDGs include global real-
A detailed analysis of specific applications of localization time monitoring. In order to achieve this, a large number of
in 6G wireless systems is presented in the following sections. low-cost biodegradable devices are expected to be deployed to
In more detail, the first two sections distinguish between the gauge climate, collect meteorological data, and track species
various applications of localization based on the nature of the distributions around the globe. Autonomous and robust place-
utilized hardware. In Section II-A, the applications that fall un- ment is required to map the gathered measurements from the
der the sustainable development umbrella are analyzed, while massive deployment of sensors to the site where the data will
in Section II-B, the ones that will lead to achieving immersive be used. The so-called “Destination earth” allows for near real-
telepresence are detailed. Section II-C highlights promising 6G time observation and monitoring of environmentally impor-
applications that involve setting up local trust zones for human tant factors including solute concentrations, biodiversity, and
and machine localization. Finally, Sections II-D and II-E weather [67]. Environmentally friendly and pervasive sensors
describe the applications of localization in next-generation linked together in terrestrial wide area networks and wireless
massive twinning and robotics settings, respectively. mesh networks make it possible to conduct such monitoring
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 6

TABLE III
A BBREVIATIONS

1D One dimensional MR Mixture reality

2D Two dimensional MSE Mean square error
3D Three dimensional NASA National aeronautics and space administration
3G Third generation NG-RAT Next generation radio access technology
4D Four dimensional NLOS Non line of sight
4G Forth generation NN Neural network
5G Fifth generation OFDM Orthogonal frequency division multiplexing
6G Sixth generation PARP Peak to average power ratio
AI Artificial intelligence PCA Principal component analysis
AoA Angle of arrival PD Photo-detector
AoD Angle of departure QoE Quality-of-experience
AR Augmented reality QoPE Quality-of-physical-experience
BSE Beam split effect QoS Quality-of-service
CDF Cumulative distribution function RaF Random forest
CMOS Complementary metal oxide semiconductor RBF Radial basis function
CNN Convolutional neural network RF Radio frequency
COA Curvature of arrival RGB Red green blue
CS Compresive sensing RIS Reflective intelligent surfaces
CSI Channel state information RL Reinforcement learning
D2D Device to device RMS Root mean square
DL Deep learning RRN Recurrent neural network
DoA Direction of arrival RSS Received signal strength
DoD Direction of departure RTT Round trip time
DOF Degree of freedom RX Receiver
DT Digital twin SDG Sustainable development goals
EMND Enhanced MND SDN Software defined network
EKF Extended Kalman filter SINR Signal to interference and noise ratio
FDM Frequency division multiplexing SLAM Simultaneous localization and mapping
FIM Fisher information matrix SNR Signal to noise ratio
FL Federated learning SotA State of the art
FMCW Frequency modulated continuous wave SVM Support vector machine
FOV Field of view TDoA Time-difference of arrival
GALILEO European GNSS TDM Time division multiplexing
GDoP Geometrical dilution of precision THz Terahertz
GLONASS Russian GNSS TL Transfer learning
GNSS Global navigation satellite system ToA Time of arrival
GP Gaussian process ToF Time of flight
GPS Global positioning system TX Transmitter
GRU Gated recurrent units UAV Unmanned aerial vehicle
IoT Internet of things UFK Unscented Kalman filter
IVM Import vector machine UGV Unmanned ground vehicle
KF Kalman filter UN United Nations
kNN k-nearest neighbors URLLC Ultra-reliable low latency communications
KPI Key performance indicator UWB Ultra wide band
LED Light-emitting diode VLC Visible light communications
LiDaR Light detection and ranging VLP Visible light positioning
LOS Line of sight VNF Virtual network function
LSTM Long short-term memory WiFi Wireless fidelity
MDS Multidimensional scaling WSN Wireless sensor network
MEM Micro-electro-mechanical XLM Extended learning machine
MIMO Multiple input multiple output XR Extended reality
ML Machine learning
mmWave Milimeter wave
MND Malicious node detection

tasks. This encourages data collecting even in underdeveloped change. To this end, the climate deal aims for net positive
regions with limited access to modern technology. Cases benefits on the climate budget. Due to the exponential growth
include conserving endangered ecosystems through human of data, future networks will need to significantly lower their
activities like preventing the spread of disease via adjusting cli- energy footprint. To this end, green localization approaches
mate models and/or monitoring environmental conditions [68]. are perfectly suited for energy-saving usage in this context.
As a result, the objectives of future 6G network should 3) e-Health: By using a low-cost 6G networks, it is pos-
include the development of earth monitoring systems. The sible to increase access to healthcare for a huge fraction of
network is strengthened by the incorporation of tiny sensors the world’s population. Remote healthcare is widely used all
that work together to build a “sensor infrastructure web”, over the world to serve people who live in remote areas or
which incorporates a worldwide sensor network for tracking who are unable to travel for medical treatment. The use of
the state of the planet’s ecosystem [69]. The sustainability of telemedicine and other forms of remote health monitoring
human existence on earth is specifically threatened by climate make this service possible [70]. Telemedicine is the practise of
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 7

Infrastructure sensor web

Earth monitoring Trust zones Telepresence Smart cities e-Health Manufacturing Agriculture Network extension

Fig. 2. Localization-assisted applications

treating, diagnosing, and evaluating patients remotely. Using sensor data to portions of the network, where storage and
AR/VR/XR, a patient and doctor may have a conversation that computing are quick and practical. For instance, in healthcare
is almost identical to the one that would take place during contexts, it is crucial to fuse sensor data in order to discover
a face-to-face visit. Telemedicine now makes use of high- patterns and deviations, and this is made possible by intelligent
quality video teleconferencing; but, with the fast development storage and distributed processing capabilities [72], [73]. Addi-
of holographic technology, is anticipated to be employed in tionally, multi-modal localization is made possible by context-
the near future. Wireless body area networks and wearable de- awareness, allowing mobile devices to transition between
vices are only two examples of the Internet-of-Medical-Things channels and communication technologies based on their
technologies that have begun to see widespread deployment present location. With multi-modality, devices may choose
for remote health monitoring [71]. The same way that real- between national and private technologies, reducing power
time monitoring systems benefit society as a whole by tracking consumption and improving service quality. This diversity in
emergency situations, future 6G-enabled e-Health systems will communication styles necessitates the use of context-aware
be able to provide immediate treatment. However, there is a localization approaches [74].
significant barrier to this aim caused by the slow rate of accep- The following developments in awareness of context is
tance of new technologies. Also, those who require constant expected to benefit from 6G networks. Context awareness tech-
monitoring, like diabetics, may rest easy es that their dosages niques will be able to keep track of history and interpret tem-
can be modified with the help of digital medication. Finally, poral data, allowing them to spot patterns and outliers, which,
sample collection for testing may be facilitated by deploying in turn, can improve anomaly detection protocols. For instance,
drones to out-of-the-way locations. Strict requirements on this is essential in hyper-personalization, where instantaneous
localization precision in the range of 0.1−0.3 m in the vertical changes in a user’s status are tracked. In addition, there are
and 0.1 − 0.5 m in the horizontal plane need to be satisfied unique privacy risks with sophisticated temporal context de-
to guarantee a valid sample collection by deploying drones to tection algorithms since they integrate private information, i.e.,
improve remote health care. For tracking purposes, especially human physiological data, with publicly available information,
during landing and sample collection/delivery operations, a such a as environmental variables. Intelligent security solutions
high level of availability of up to 99.99% is required. are needed to store and analyze this data in various portions of
4) Personalized context-aware networks: 6G is envisioned the infrastructure. Finally, a lot of standardization of context
to allow contextual interpretations in both communication parameters is needed for distributed context awareness. It is
and localization protocols. Humans are able to comprehend hard to construct networks that can transform the context into
and understand their activities because of the organizational, assets for applications without defined interpretation rules for
cultural, and physical circumstances in which they participate. context parameters. Standardizing context operating rules and
This is accomplished by a combined comprehension of time parameters to achieve a fair data economy that permits the
and activity components of an action along with the quantita- combinations and flow of multiple data sources, which are
tive characterisation of entities, like things, places, or people. regulated by a diverse ecosystem, poses a significant challenge
The ability to recognize human and technological contexts is to the implementation and adoption of such context-aware
equally crucial for the realization of 6G networks [54]. features.
Contextual understanding will be useful in several 6G
applications. By communicating only when it is most efficient
applications may significantly reduce their overall energy B. Immersive telepresence
footprint. Insightful prediction of data transmission based on Immersive telepresence enables users to interact with a
context paves the way for increased throughput on demand. distant environment, including devices, people, and/or objects,
To achieve hyper-personalization, context-awareness may be utilising all of their senses. The foundations of a fully-
utilized to intelligently relocate personalization algorithms and cyberphysical world would be strengthened by its capacity
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 8

to provide unobstructed near-real-time holographic experience critical for optimising use of available resources. Furthermore,
and mixed reality. This enhanced capacity is expected to be with XR, it is possible to test out ideas before committing to
useful for mixed reality work, as well as fully immersive a final design, and it allows corporations to provide customers
entertainment, such as sports, gaming, live events, etc. All of services that are tailored to their own needs and circumstances.
the use cases under the telepresence umbrella have one thing in To offer an XR solution that is both appropriate and realistic
common: the need for localization in order to enable features for the user’s current location, 6G localization of the device
such as gesture and motion recognition and XR. providing the service is essential. Thus, an XR user would
1) Gesture and motion recognition: 6G’s higher frequency require accurate localization in order for the device to receive
range provides improved precision and resolution, allowing the correct data. A shopper, for instance, may employ XR-
for the capturing of finer gestures and actions. With the enhanced advertising while strolling through a retail centre. As
advent of advanced computer technology and pervasive AI another use case, XR can superimpose a computer-generated
methods [75], a new age of gesture and activity recognition image or simulation onto the real world. However, much
has begun. Localization and sensing in the near future will greater precision is needed in this particular setting.
not be confined to a single space, but rather a massive and
complex setting. This allows for widespread gesture recording
C. Local trust zones
and activity detection without the need for any additional
devices. Each base station is envisioned to serve as a sensor Local trust zones related use cases highlight the need of
that works together with user devices to get a more complete a framework that calls for the NG-RAT to enable adaptable
picture of their environment, leading to a significant increase architectures that let enterprises and services to keep their
in performance. Moreover, the base station enables a notably most sensitive data inside their own networks. Some examples
increased range that was not attainable through only user that fall within this broad category are: (i) infrastructure-less
devices, as well as the dissemination of the gathered data to network extensions; (ii) sensor infrastructure web; (iii) preci-
the cloud or to neighbouring buildings. sion healthcare; and (iv) security and privacy. Localization is
A smart environment could use high classification accuracy required in all the aforementioned use cases clusters.
to include features like behaviour detection, gesture and mo- 1) Infrastructure-less network extensions: This use case
tion recognition, intruder detection, and more [76]. Likewise, requires extending the network coverage temporarily, partic-
patients in a smart hospital may benefit from automated pa- ularly if the number of devices to be served are located
tient monitoring thanks to the facility’s medical rehabilitation near the network’s edge. Use cases are almost ubiquitous.
system. In such environments, improper gestures or actions Normal instances take place in rural locations, where cov-
will trigger notifications, greatly enhancing security as well as erage is restricted and performance levels are inconsistent. In
the provided services (i.e., rehabilitation). More accurate loca- production settings, bigger populations of the same vendor’s
lization capabilities and high classification precision would be modules, computers, and vehicles may be linked together
necessary for certain superior contactless activities like playing via infrastructure-free networks that serve as an underlay for
virtual piano. Assuming that the black keys have an average the infrastructure network and require the adaptive and fast
width of 10 − 11 mm, the expected accuracy for determining solution of geometric consensus problems [79]. For example,
the finger’s vertical distance from the board and its horizontal autonomous vehicle platoons have been used for agricultural
location in relation to the virtual keys is 3 mm. In addition, harvesting. Each vehicle should keep its distance from the
the recognition probability should be higher than 99% for a others to prevent accidents and work together to find their way
pleasant tune. around the operational area. It is also important to guarantee
2) XR: Across the whole AR/VR spectrum, XR will pro- the vehicles’ safe operation even when on the borders of the
vide several applications for 6G. Due to their limitation to coverage area. The application will need an update rate propor-
provide extremely low latencies for data rate-heavy applica- tional to the current pace of the cars’ movement. Cooperating
tions, upcoming 5G systems still fall short of offering a fully localization is a term that is occasionally used to describe these
immersive XR experience that captures all sensory inputs. kinds of applications. Finally, most of these use cases include
Creating an immersive AR/VR/XR experience calls for a either a population of nodes that make up the infrastructure-
collaborative design that takes into account not just technical free network and may move independently to enable direct
but also perceptual considerations. Engineering methods have device-to-device connection, or a collection of sensor nodes
to account for both the maximum and minimum boundaries of travelling along strongly linked trajectories.
human perception. To achieve this goal, traditional QoS, i.e., 2) Sensor infrastructure web: The sensor infrastructure
rate and latency, and QoE, like mean opinion score, inputs web was introduced by NASA [80]. It consists of smart in-
need to be combined with physical elements from the users frastructure that enables coordinated localization and sensing.
themselves, necessitating a new notion of quality-of-physical- It is made up of disparate nodes equipped with sensors that
experience (QoPE) metric. QoPE may be influenced by a num- not only gather information, but also communicate with each
ber of actions, such as non-verbal cues, physical processes, and other to modify their behaviour [81]. Tens of millions of
mental processing. In [77], for instance, we demonstrate that mobile, airborne, in-situ, and space-borne sensors have been
under the URLLC regime, the human brain may not be able to deployed in diverse domains and for varied purposes recently
differentiate between various latency metrics. Meanwhile, as due to the growth of IoT, smart cities, and the Global Earth
shown in [78] the sensory input, particularly touch and sight, is Observation System of Systems (GEOSS). Earth’s urban areas
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 9

have never been detected so dynamically and in-depth in rely only on energy harvesting for their operation [84]. These
the past. To accommodate this scenario, 6G is expected to nanonodes are expected to be passively flowing, i.e. without
provide localization data to devices with almost zero sensing the potential of mechanical guiding toward the specified point,
capabilities, such as guiding a vehicle through an unfamiliar due to their limited energy and small form factors.
area that has inadequate on-board sensors to comprehend Based on the aforementioned, localizing the nanonodes is
its surroundings. Accurate device localization is required to of great importance for regulating their operation. To issue
convey relevant sensing information to devices lacking sensing control commands, for example, two way communication
capabilities, which is necessary to enable such use cases. between the nanonodes and the outside world is essential.
For example, in disaster management application, the sensor THz communications constitutes a viable technology enabler
infrastructure web is able to use a combination of in-situ in such scenarios as the primary prerequisite for in-body
and space sensors in order to collect data for volcano hazard nanonodes, namely small transceiver form-factors, is only
monitoring, or use orbital space sensors, like satellites, to possible with transmission at these frequencies. The THz
locate flooding indicators or wildfire hot-spots [82]. band, however, is entangled with various limitations, such as
3) Precision healthcare: Precision medicine has emerged high scattering and spreading losses that limit the propagation
as a mean to treat and prevent illness in an individualised distance of THz waves. The main challenges of THz-enabled
fashion. Dispensing medication with the help of wireless nano- nanoscale localization include:(i) maintaining low complex-
scale robots that move within the human body’s soft tissues is ity and energy demands of nanonodes; (ii) overcoming the
fundamental to the individualised therapy, which differs from substantial attenuation of in-body THz propagation; and (iii)
the standard one-size-fits-all method. To maintain tabs on a allowing network scalability [63].
patient’s health, 6G connection will be used to gather sensor- Research on nanoscale networks and their localization ap-
based data. Utilizing the massive coverage area provided by proaches is currently limited and constrained in tackling a sub-
6G, patient monitoring and tracking may be utilised to ensure set of these challenges. Specifically, a THz-enabled network
the safety of patients regardless of the nature of their illness architecture was proposed in [85], which enables fine-grained
or their current health status. The anticipated 6G localization localization of the energy-harvesting in-body nanonodes, as
capacity benefits all these distinct applications by providing well as their two-way communication with the outside world.
enhanced precision healthcare, such as medical equipment This approach elisted software-defined metamaterials and
placement on the body, small-scale robots, patient monitoring location-aware and wake-up radio-based wireless nanocom-
and tracking, as well as telesurgery. munication paradigms to enable the novel energy-harvesting
From the aforementioned applications, telesurgery is a capabilities for in-body nanonetworks. The authors suggested
unique use case of cutting-edge precision healthcare that em- that the proposed design can take advantage of the large num-
ploys a surgeon and an expert operator stationed in a different ber of nanonodes, surpassing the limitations of limited range
location to operate one or more robotic arms. Data for receiv- of THz in-body propagation and severely confined nanonodes.
ing feedback and controlling the surgical tools must be reliably Moreover, in [86], [87], the authors explore the potential
sent between distant locations. High-data rate URLLCs that for flowing nanosensors in the blood flow to identify and
are expected to be made possible by 6G technology will allow localize and report anomalies in the human body. This work
robotic arms to mirror the surgeon’s natural hand motions and focuses on the identification of quorum sensing molecules
offer haptic input, boosting the surgeon’s senses and dexterity. and evaluate their performance. The authors use a Markov
More advanced technologies, known as cooperative surgical chain model to simulate the nanosensors’ motion through the
systems, let doctors and machines work together to achieve blood arteries, and apply ML models to predict their trajectory.
better results during procedures. The tissues and surgical The results validate the detection and localization capabilities
equipment both need to be precisely localised for successful of the investigated method across a variety of body areas,
remote surgery. demonstrating their utility in identifying vascular anomalies.
4) Nanoscale localization: Recent breakthroughs in nan- 5) Security and privacy: Addressing the security elements
otechnology are bringing light to nanometer-size devices that of localization is a growing area of study alongside enhanc-
will allow a multitude of innovative applications. Among the ing localization precision, sampling efficiency, and resilience
many fascinating fields where nanotechnology is predicted to against intermittent connection loss. Jamming and spoofing
reap benefits is in-body treatment. Antibiotic resistance might are two versions of intentional interference that degrades or
be detected at the molecular level, drugs could be delivered prevents the calculation of a GNSS signal [88]. In this context,
with pinpoint accuracy, and neurosurgery could be performed protecting personal privacy entails blocking the functioning
on specific areas of the brain. In order to implement such of GNSS receivers that may track the user’s whereabouts
applications, a network of nanodevices will need to be created and communicate such information to other parties. In case
within the body, which will then travel through the circulatory jamming signals are not correctly identified and their conse-
system, respond in response to orders at specific areas, and quences are not mitigated, the employment of jammers leads
report their findings to a more robust body area network in reduced positioning accuracy or complete loss of GNSS
(BAN) [83]. Nanodevices in this context are assumed to have signals, which may result in catastrophic damage. Intentional
dimensions similar to those of red blood cells in order to intervention is also done by political activists, cybercriminals,
prevent clotting as a result of their entry into the circulation. and foreign nations to disrupt the networks of others. Spoofing,
Because of their little size, these nanonodes are likely to on the other hand, refers to the malicious broadcast of false
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 10

GNSS-like signals that trick the receiver. The hardest part of counterpart of anything in the actual world [98]. The ability
spoofing is its identification. Fortunately, the attacker must to create a trustworthy, efficient, and effective digital duplicate
be able to produce authentic GNSS signals, including data, of any physical object will open up hitherto inconceivable
modulation, and timing, maintain temporal synchronisation avenues of exploration, such as (i) more effective use of
near to authentic GNSS time, and modify the signal power resources has a positive effect on the long-term viability of
levels to correspond to those of the authentic signals in order agriculture; (ii) to facilitate a fully smart city by developing
to be able to conduct a realistic spoofing assault. 4D spatio-temporal interactive maps; and (iii) tracking the
There are a plethora of localization approaches that take se- operations at every corner of a smart factory at once by
curity into account. SotA security-aware techniques frequently constructing a digital map of the space.
have large computing needs, prioritize structure-based assaults, The development of a digital twin (DT) may fundamentally
assume static networks and even make bold assumptions benefit from data about events occurring in the real world at
that may not hold true under real-world circumstances. A a certain time and place. 6G-connected tags may be affixed
common problem is that when harmful conduct from anchor or to real-world items to generate a digital duplicate, with the
unknown nodes is expected, either no explicit attack model is added benefit of gathering sensory data that can be paired
supplied or detection of such hostile nodes is all that is defined, with time instances to show exactly when the data was
with no further treatment of the discovered nodes being speci- gathered [98]. There are, however, difficulties in correctly
fied. For instance, exploiting the noise characteristics induced localising these devices, particularly when they are installed
by external distance assaults, the authors of [89] suggested in GPS-denied locations. 6G WANs are being constructed to
a lightweight secure ToA-based localization technique that handle the communication demands of DTs; these networks
enables protection against impersonation attacks in WSNs. might be used for localization as well, solving this issue [99].
Clustering and consistency evaluation-based malicious node In addition, localization and detection of items that are not
detection (MND) and its improved and secure variant, en- linked to the digital world via 6G tags or other 6G-enabled
hanced MND (EMND), constitute a solution to identify the devices is required in certain use cases. Radio-signal-based
anomalous node clusters by applying spatial clustering based sensing may be used in this context to gather data about
on density [90]. Afterwards, they utilize a sequential probabil- passive items, such as their position and the time they were
ity ratio test to find the bad actors in the networks. Simulations present. Adding data on inactive things is a useful supplement
reveal that the proposed algorithms are more accurate and to the DT, which typically only includes digital copies of
efficient than current SotA approaches. One downside of the network-accessible physical items.
proposed algorithms is that there is no way for maliciously 1) Agriculture: Numerous advancements have been made
labeled anchor nodes to redeem themselves and reclaim their in the study and use of robotic solutions for the agricultural
trusted status in the network. In addition, in [91], the au- industry, and fresh contributions are anticipated in the near
thors present a extremely computationally-lightweight while future [100]. Farmers are becoming more aware of its influence
yet retaining the benefits of range-free techniques including on agriculture, which has increased the need for autonomous
low cost and complexity of deployment as well as increased machinery in this field. A wide range of duties, including
resilience against assaults of secure localization systems. Due application of fertilisers, watering, harvesting, planting, and
to the distributed nature of this approach, it relies on no more, are being carried out by robots [101]. The design and
external resources or single points of failure, thus being able development of technological concepts that enable robots to
to identify and filter out malicious or malfunctioning anchor travel safely across various settings is crucial in this situation.
nodes, improving localization precision. The fundamental necessity imposed by these advancements
Finally, intruders might exploiting security flaws to harass is to localise the robots in various agricultural settings. Using
the users, break into their homes, steal their identities, and Global Navigation Satellite System (GNSS) is the most typical
more. Privacy concerns about the impending proliferation of technology used so far [102]. The use of GNSS is questionable
THz communications on smartphones and wearables have in many agricultural areas due to signal blocking and multi-
been voiced several time in the past [92], [93]. THz-based reflection of satellite signals [103]. Therefore, it is crucial
remote sensing and see-through imaging may be used by an to investigate and create intelligent systems that calculate
attacker or malicious equipment, which might compromise the robot’s localization using a variety of sensor modalities
users’ privacy. Also, radar-like localization systems are being and data sources. The most advanced method for doing this
developed using 4G and 5G signals and developments in is simultaneous localization and mapping (SLAM) [104].
full-duplex communications [94], [95], with the potential for This method involves mapping the immediate area while
expansion beyond 5G. Radio frequency fingerprinting [96], also predicting the status of a robot. The robot model often
[97] may identify user devices even without providing a includes the robot’s position as well as, in certain situations,
device identifier, highlighting the potential privacy risks of its velocity, calibration factors, and sensor offsets. The map
combining high-resolution photography with machine learning is a 3D depiction of the agents that the robot’s onboard
algorithms. sensors have detected and are used as a guide throughout the
localization process. Creating a map is often crucial to provide
D. Massive twinning details about the surroundings. Additionally, maps may be
Future RAN aims to bridge the gap between the physical updated and reused by the robots each time they go across
and digital worlds, making it possible to create a digital the environment. Finally, visual odometry is one of the most
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 11

popular alternatives to SLAM, which is capable of predicting E. Robot coordination and interaction
the motion of the on-board camera using just image data as
input [105]. Local pose tracking and global localization are two basic
types of the localization challenge for mobile robots. Once
2) Smart cities: To improve the quality of life in a city, a the robot’s starting stance has been determined, the local pose
comprehensive smart city must give a wealth of data, such as tracking problem may focus on maintaining that pose over
a digital reproduction of the city with real-time data on control time. In contrast, the global localization issue requires the
utilities, pollution maps, traffic, etc [106]. An interactive 4D robot to independently localise itself and minimise the uncer-
map is a key component of an immersive smart city, since tainties in its posture predictions. A mobile robot’s first posture
it can be used to better organise and design essential infras- estimate at startup is required for relocating the robot in the
tructure and services, like transportation, waste management, event of a pose tracking failure, autonomously navigating, and
plumbing, electrical wiring, and more [107]. Sensors installed more. Moreover, numerous localization strategies have been
to track, monitor, and update status of these services must investigated over the last several decades, but probabilistic
be localised for this kind of dynamic map to be possible. methods using Bayes filters, Monte Carlo particle filters,
Moreover, units may be stationed by the side of the road to aid Markov particle filter, and extended Kalman filter are now the
in communication, while simultaneously operating as traffic most popular and well-proven options [113]. There are two
monitors in various sections of the city. main uses for localization with respect to robots: (i) interaction
with humans in order to accomplish goals in the everyday life
Radio signals received by a user at a certain position are and industrial settings, and (ii) localization and mapping of
impacted by the features of their immediate physical surround- the environment, where robots may determine where they are
ings in a predictable manner [108]. For instance, because of located in space and how they are oriented in relation to one
the existence of urban infrastructure, such as buildings, a user’s another as well as their surroundings.
device in the street level may pick up a signal that has travelled
1) Interactions: It is anticipated that 6G wireless systems
by way of several reflections. The presence of trees near the
will offer the means through which a technological basis
target causes a reflection of radio signals. By inferring the
for human-robot collaboration may be laid. As a result,
sort of topography at the user’s position, we can better map
collaborative robots (cobots) will be used extensively in both
the terrain around the cell towers. Finally, due to the virtual
the domestic and commercial settings. Moreover, cobots will
depiction of each piece of infrastructure, huge digital twining
become more important in the home and consumer sectors,
may be of great use in the smart buildings of the future.
evolving beyond the robotic household helpers that they are
To do this, you must determine the precise location of every
today. Cobots will also be used in the manufacturing sector
heater, lamp, light switch, and other controlled fixture in the
to facilitate flexible production of a wide variety of items,
facility. Not only will it be put to use in the daily running
including those with a high level of personalization and cus-
of the building, but its intelligence and automation will also
tomization. Accurate 3D localization of cobots that supports
allow for a significant decrease in the cost of the building’s
high update rates is essential for their efficient operation, as
commissioning.
are capabilities to perceive the environment, with the former
3) Manufacturing: As we go towards Industry 5.0, ma- in particular allowing for synchronised, uninterrupted, and
chines and robots will become more self-sufficient, with only smooth collaboration. Cobots are expected to use this data as
little human oversight [109]. As a result, cutting-edge control input to plan and conduct actions, either alone or in swarms in
mechanisms are essential for high-precision production pro- order to achieve their goal. Additionally, the cobots use a sym-
cesses. In particular, connection density of 107 /km2 , latency on biotic autonomy technique, where the robots are self-aware
the scale of hundreds of microseconds, and data speeds on the of their perceptual, motor, and cognitive limits and actively
order of Tbps are required for the fast development industrial seek human assistance for tasks like object manipulation. The
systems as well as the automation of their operations [110]. cobots’ success in the 1,000-km challenge demonstrates their
THz networks are a natural solution for achieving such high localization system’s stability and accuracy during long-term
data speeds. However, there are several obstacles that need deployments [114], despite the existence of environmental
to be addressed when using THz networks for Industry 5.0 fluctuations. To achieve this, a monitoring script running on
applications, such as precision-driven control mechanisms, the robots would follow the status of the tasks’ execution and
dense coverage, and zero-latency. Moreover, the positioning contact researchers in the occasions when the robots required
capabilities of 6G networks have the greatest potential to help. As a result, over ninety percent of deployments of robots
advance industrial use cases [111]. Many different industries, did not need human involvement to reset the robot localization.
from logistics and manufacturing to mining and transportation, 2) Mapping: When the GPS signal is poor or unavail-
will benefit from using position data to improve and automate able, a mobile robot may use SLAM to create a consis-
their operations [112]. Both the (mobile) terminals and robots tent map of its surroundings and simultaneously calculate
as well as the whole network will benefit greatly from location its position [115]. In the realm of autonomous navigation
data in factory automation and industrial control. The former and localization, SLAM is widely recognized as a crucial
can better allocate and regulate resources and boost processing challenge [57]. The autonomous capability of mobile robots in
efficiency, while the latter can better direct the movement of unfamiliar surroundings is greatly enhanced by the solution of
the terminals. the robust long-term and real-time SLAM issue, which focuses
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 12

on two primary components: localization and mapping [116], is given at the top of the next page, illustrates each application
[117]. Through accurate and reliable localization, the robot’s alongside its corresponding KPIs and related use cases.
present location in space is determined. Simultaneous mapping
without prior knowledge of the robot’s position unifies the A. Indoor
fragmented understanding of the surroundings into a coherent
whole [118]. Autonomous mobile robots equipped with power- Services such as indoor localization have emerged as a con-
ful SLAM algorithms have played an important role in a wide sequence of the rapid use of wireless devices in recent years.
variety of application scenes, including exploration in hostile Obtaining the position of a person or device within a building
environments such as aerial space, underground mining, rough is referred to as “indoor localization”. Extensive studies on
terrain, and underwater surveillance [119]. indoor device localization have been conducted over the last
Since its introduction in the 1980s, SLAM has evolved decade, mostly in industrial settings and for wireless sensor
significantly over the course of more than three decades [120]. networks and robotics. Since the widespread adoption of
There has been a shift from filter-based to optimization- smartphones and wearable devices with wireless communica-
based algorithms and from single to multi thread methods. tion capabilities, the localization and tracking of such devices
Sonar has been replaced by 2D/3D light detection and ranging has become synonymous with the localization and tracking of
(Lidar) systems and monocular, stereo, RGB-D, time of flight the corresponding users, enabling a vast ecosystem of related
(ToF), and other cameras. There are now three primary SLAM applications and services. The ability to precisely pinpoint the
methods in use: Visual SLAM, multi-Sensor Fusion SLAM, location of a user or device has several practical applications
and Lidar SLAM. Visual SLAM estimates its posture with the in the fields of autonomous supply chain, e-Health, precision
use of prior knowledge of multi-frame pictures and multi-view healthcare, personalized context-aware networks, XR, gesture
geometry, and then utilizes this estimate together with depth and motion recognition, sensor infrastructure web, security and
information derived from the accumulation of pose changes. privacy, agriculture, smart cities, manufacturing, as well as
Most notably, RGBD cameras can get depth data without robotic interactions and mapping.
any further processing. To complete localization, the SLAM
algorithm in Lidar SLAM matches and compares Point Cloud B. Outdoor
at various periods to determine the distance traveled by the Accurate positioning data is essential for many outdoor
robot as it moves relative to the surroundings and as its attitude applications, such as navigating in uncharted territory and
changes. Due to the limitations of individual 2D/3D Lidar or monitoring the location of company vehicles. However, in
cameras, multi-Sensor Fusion SLAM has gained popularity in urban canyons with high-rise buildings or subterranean parking
recent years. The most common sensors for fusion algorithms lots, where building walls obstruct the GNSS reference signal,
include Lidar, cameras, IMUs, wheel odometers, and Global GNSS coverage is limited. Current mmWave cellular networks
Navigation Satellite Systems (GNSS). and upcoming sub-THz cellular networks offer good prospects
SLAM technology is in high demand for use in a wide for localization where GNSS fails to fulfill the application’s
variety of applications, including but not limited to un- KPIs. At mmWave frequencies, adjustable channel bandwidths
manned aerial vehicle (UAVs) [121], unmanned ground ve- of up to 400 MHz are possible in 5G and phased arrays
hicle (UGVs) [122], indoor autonomous mobile robots [123], with hundreds of antenna components and narrow half power
and VR/AR hardware [124]. The vast majority of AR systems beamwidths are commercially available. Future 6G wireless
need pre-existing knowledge about their surroundings in order systems will migrate to frequencies greater than 100 GHz,
to function. In contrast, the advancement of SLAM algorithms when channel allocations lasting several gigahertz are possible.
and hardware has allowed sensors to rebuild free-form interior Strategies that take the estimated condition of the wireless
scenes without the need for real-time map initialization. Since propagation channel into account have been the focus of
GNSS signals cannot completely penetrate buildings, it is more research in wireless systems’ localization. Channel impulse
challenging to place robots inside, and conventional localiza- response, propagation delay, AoA, and RSS are only some
tion relies on community landmarks and signage. With SLAM of the propagation parameters that are measured or estimated
and inertia measurement units working together, we can solve by networks at one or more locations, usually base stations
this issue [123]. For UGVs and UAVs, SLAM is also a crucial and user equipment. Moreover, user-location fingerprinting
tool. Specifically, Sensor Fusion SLAM is quickly replacing has also been interpreted geometrically, compared to existing
other methods for building the map concurrently, positioning databases, or used to locate the user on maps. Finally, the out-
unmanned vehicles, and implementing autonomous navigation door localization use case includes all applications discussed
in an uncharted area [121], [122]. All things considered, in the present contribution.
SLAM has been heavily used across many different disciplines
and has promising future applications.
C. Non-terrestrial
Non-terrestrial localization has become ubiquitous as a
III. L OCALIZATION USE CASES
result of the fast advancement of satellite localization technol-
In this section, the use cases of future wireless systems ogy, making formerly inconvenient tasks easier for the general
are described based on the localization requirements and public. When it comes to non-terrestrial localization, GNSS is
applications mentioned in Section II. To this end, Fig. 3, which the most applied solution and is made up of constellations of
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 13

Sustainable development
1 m horizontal localization accuracy I Indoor NT Non-terrestrial
Autonomous supply chain N Nanoscale
1 Hz update rate
O Outdoor U Underwater
99% availability I O NT
some decades/m3 scalability

Massive twinning
<1 m horizontal localization accuracy Agriculture
1-30 m horizontal localization accuracy 1 Hz update rate
Earth monitoring
1 Hz update rate 99.99% availability I O NT
99% availability O NT U
<1 m localization accuracy
0.1-0.5 m horizontal localization accuracy Smart cities
e-Health 1o angular resolution
0.1-0.3 m vertical localization accuracy
99.99% availability I O NT
1 Hz update rate I O NT some decades/m3 scalability
99.99% availability
1-3 m horizontal localization accuracy cm localization accuracy
Personalized context- hundreds m maximum range
0.1-3 m vertical localization accuracy Manufacturing
aware networks <1o angular resolution
10o orientation accuracy
30 km/h velocity range
1-10 Hz update rate I O NT U 0.5 m/s velocity resolution I O
99% availability
0.01-10 kHz update rate
99.99% availability
Immersive telepresence

cm localization accuracy
0.01 kHz update rate Gesture & motion

Robots
<1 cm localization accuracy Interactions
99.99% availability recognition <1o 3D orientation accuracy
some decades/m3 scalability
0.5 m maximum range 99.99% availability I O NT U
15o angular resolution >5 devices/m3 scalability
I O
10 m/s velocity range cm localization accuracy
0.3 m/s velocity resolution hundreds m maximum range
<1o angular resolution Mapping
<0.5 m localization accuracy
eXtended reality 30 km/h velocity range
1o angular resolution
0.1 kHz update rate 0.5 m/s velocity resolution I O NT U
I O 0.01-10 kHz update rate
99.9% availability
some decades/m3 scalability 99.99% availability
Local trust zones

<1 mm localization accuracy Precision healthcare

1 cm agriculture localization accuracy Infrastructure-less <1o orientation accuracy
30 cm automotive localization accuracy network expansion 99.99999% availability I O NT N
<1o agriculture orientation accuracy 0.01 kHz update rate
0.1 kHz update rate
O NT U <1 mm localization accuracy
99.99% availability Nanoscale localization
hundreds of devices scalability <1o orientation accuracy
N
hundreds of devices scalability
10 cm pedestrian localization accuracy cm localization accuracy
Sensor infrastructure web
30 cm vehicle localization accuracy cm range resolution
hundreds m maximum range Security & privacy
0.01 kHz pedestrian update rate
0.1 kHz vehicle update rate I O NT U <1o orientation accuracy
I O NT U N
99.9% availability 99.99% availability
0.01 kHz update rate

Fig. 3. Localization use cases

artificial satellites in geosynchronous orbit, transmitting real- personalized context-aware networks, agriculture, smart cities,
time position and timing information. During the time of writ- robotics-enabled mapping, infrastructure-less network expan-
ing, four major not-terrestrial systems are in use: (i) Europe’s sion, sensor infrastructure web, precision healthcare, as well
Galileo, (ii) the United States’ GPS, (iii) China’s Beidou, and as security and privacy.
(iv) Russia’s GLONASS. These systems provide of 1 − 10 m
accuracy and low device autonomy. GPS, which was originally D. Underwater
designed for use by the United States Navy, is a radio navi-
In recent years, underwater wireless systems have applied
gation and positioning system that contains location and time
localization techniques for many different purposes. Each
information, while GLONASS, a second-generation military
implementation is crucial in its own field, while some of
satellite navigation system is technologically superior to GPS
them can advance ocean exploration to accommodate a range
in its capacity to withstand interference, yet it lacks GPS’s
of underwater applications, such as underwater surveillance,
pinpoint single-point precision. Moreover, BDS is a Chinese-
natural disasters alert systems, oceanographic data collection,
created satellite navigation system with worldwide coverage
assisted navigation, etc. For instance, sensors may evaluate
and GALILEO, which uses two ground control centres and 30
specific characteristics, such as base intensity and mooring
satellites, is more widespread in Europe. Based on these, non-
tension, for offshore engineering applications to monitor the
terrestrial localization applications of future wireless systems
structural condition of the mooring environment. Therefore,
include autonomous supply chain, earth monitoring, e-Health,
the basic applications of underwater localization wireless
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 14

systems include earth monitoring, sensor infrastructure web, receiver side [127], [128]. As a result, mmWave and THz links
personalized context-aware networks, mapping, security and are highly directional. In order to ensure high-reliability and
privacy, as well as infrastructure-less network expansion. avoid defects, the base station must know and monitor the
orientation of the user equipment (UE). Reduced beamwidth
E. Nanoscale necessitates a much higher cost for current localization and
tracking methods. Because of this, they are not suitable for use
THz nanonetworks are expected to enable a wide variety in mmWave and THz systems. Because of this, it is necessary
of applications, many of which will need localization or even to work on developing efficient localization methods for THz
tracking of the nanonodes. As a result of the nanoscale nature systems.
of the nodes, it is essential that localization and tracking Due to the complementarity between mmWave and THz
capabilities work under very tight energy budgets while yet systems, the need arises to compare the benefits and drawbacks
maintaining a high degree of precision due to the small of these techniques in various diverse localization scenarios
sizes of the nanonodes. Furthermore, because of the limited especially since THz are likely to be utilized as an extension
range of THz band nano-communication, multi-hopping may of mmWave systems [129]. We anticipate bigger array sizes,
be required for localization, which has the disadvantage of smaller footprints, wider bandwidths, and higher frequencies
increasing localization error with each hop. Few efforts have as we move towards 6G networks. Thus, new possibilities
been made towards localization of THz-operating nanonodes, are enabled both in the sense of larger arrays with the
with the most notable ones using range and hop counting same dimensions as well as antennas with constraint sizes.
to make educated guesses about where all nanonodes are Moreover, increased availability of bandwidth in conjunction
positioned in a given region. In the former, the locations of with higher frequencies result in robust delay estimation, lower
the two examined nanonodes are calculated by counting the path loss, and decreased multipath.
number of hops between them, while the latter assumes that The aforementioned changes will impact a plethora of
all nanonodes are clustered together to cut down on overhead features of future networks. For instance, new synchronization
and energy dissipation. One limitation of these techniques is problems emerge and hardware needs to keep up with novel
that, as the number of hops rises, their accuracy will inevitably mmWave and THz-enabled system designs. Another already
decrease due to the spread of localization errors. The main reported problem is the requirement for increased peak to aver-
KPIs of nanoscale localization are low energy consumption age power ratio (PAPR) compared to 5G systems, which may
and high precision, while the related open challenges include render the classic orthogonal frequency division multiplexing
modeling of such systems in non-free space propagation, (OFDM) inappropriate [130]. Another important aspect that
hardware imperfections, and frequency- and angle-dependent needs to be taken into account in the development of next
response of the nanonodes. generation localization algorithms is the impact of near-field
circumstances along with the beam-split effect (BSE) that can
IV. T ECHNOLOGY ENABLERS FOR 6G WIRELESS SYSTEMS greatly influence the geometry of MIMO channel localization
This section aims to shine light into the key technology models. Finally, the high path loss plagued mmWave and THz
enablers of localization in future 6G networks. Specifically, channels can greatly influence the energy-efficiency of the
Section IV-A presents the current SotA of mmWave and overall system and thus, they must be accounted during the
THz localization techniques. Next, in Section IV-B, the main design of future networks. All in all, the increased performance
hardware advancements related to localization approaches are anticipated by mmWave and THz networks comes hand in
investigated, while Sections IV-C and IV-D illustrates the hand with important difficulties in complexity, overhead, area
role of beamforming and RIS in the localization paradigm of coverage, and hardware architectures that this affects.
future 6G networks. The impact of AI and channel charting THz localization is already the subject of active study, with
approaches are presented in Sections IV-E and IV-F, respec- researchers delving into simulation environments, localization
tively. Section IV-G discusses radar-based technologies that methods, and system architectures. For instance, a high level
enable localization applications, while Section IV-H sensor- of estimated accuracy and reduced deafness in 2D settings are
based ones. Finally, visible light-based localization methods sought for with the proposal of cooperation-aided localization
are analyzed in Section IV-I. procedures [125]. Also, Kalman filtering-based indoor locali-
zation approaches have been proven to offer great advantages
for time-variant channel modeling [131]. Another example is
A. mmWave/THz localization the angle of arrival (AoA) estimation by a forward-backward
With the explosion of cellular services in recent years comes algorithm that offers improved localization and human motion
a surge in demand for higher data transfer speeds [125]. characterization [132], while a near-field model with massive
Using the vast amount of unstandardized bandwidth available antenna arrays has been examined for the purpose of using
in the mmWave and THz range, mmWave and THz wireless the curvature of arrival (COA) as a sixth degree of freedom
systems hold great potential for meeting the data rate need. (DOF) in determining the location of the source [133]. Fur-
However, substantial channel attenuations hinder communica- thermore, by adopting a beam zooming mechanism combined
tions at these frequencies [126]. mmWave and THz systems with delay-phase precoding for mmWave and THz beam
counteract these higher losses by using pencil-beams made tracking, it is possible to monitor numerous users and blockers
from massive antenna arrays at both the transmitter and the using a single RF chain with significantly reduced beam
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 15

training overhead [134]. Finally, in addition to geometry- C. Beamforming

based strategies, DL-based strategies show great potential for Data transmission towards a specific direction in order to
3D mmWave and THz indoor localization by enhancing the increase the SNR and consequently the throughput (beamform-
localization accuracy by approximately 60% [135]. Each of the ing) is a fundamental requirement in high-frequency wireless
aforementioned localization efforts addresses a unique issue systems [54]. For mmWave and microwaves, 3D beamforming
associated with mmWave and THz localization, namely near- is a useful tool for mitigating the detrimental effects of high
field effects, tracking, improper alignment, and BSE. Despite pathloss and for improving SNR. In order to localize and
these advances, important gaps in our understanding still need sense objects or people, the gathered channel response can be
to be filled. It is yet unclear how signals in the mmWave- and analyzed for any geographical information about the relation-
THz-band might boost localization accuracy in both active and ship with them. The channel estimation, which includes the
passive approaches [60]. angular and delay domain characteristics that are necessary
for localization algorithms, is a prerequisite for beamspace
processing [143]. Especially in challenging NLOS and high-
B. Infrastructure mobility conditions, channel estimation is indespensible [144].
Beams are dynamically controlled based on the channel esti-
Most of the work being done to close the gap between
mate of the AoA and angle of departure (AoD) to find and
current and potential applications of THz technology is still at
track mobile users in a dynamic environment [145].
the device level. Several competing technologies have recently
Accurate and rapid AoA or AoD estimate of the primary
illustrated great potential for realizing THz components. The
multipath can suffice forin scenarios of mobile users who
most established technology is electronic silicon-based devices
connect with the BS through a LOS path or strong NLOS
that have already been used in mmWave systems [136]. For
route. Uplink AoA estimates from LOS connections may be
instance, complementary metal oxide semiconductor (CMOS)
used to directly infer the position of users, whereas AoA
and its variants are characterized by manufacturing compat-
estimates from NLOS links can be used for the scatterer’s
ibility and small size [137]. However, such approaches are
location. NLOS links originate from multipaths of the signal
plagued by limited maximum available power gain frequency
emitted by an active user and subsequently arrive at the BS.
and power handling capacity. To overcome the power con-
For the latter, NLOS cases, accurate estimates of both AoD
straints of CMOS and increase coverage area, MIMO-based
and AoA are required to pinpoint the user’s position.
systems with high array gains combined with beamforming
Beamspace processing for locating mobile users in NLOS
techniques provide a promising alternative.
and LOS settings is a significant challenge; “device-free”
In contrast to electronic devices, photonic THz systems, localization 1 presents an even greater one [146]. Here,
such as uni-traveling carrier photodiodes, quantum cascade additional environment indicators are required so that the
lasers, photoconductive antennas, and optical downconversion target(s)’ spatial characterisation may be differentiated from
systems, provide both data speeds and larger carrier frequen- that of the background objects. This is made more challenging
cies (over 300 GHz) [138]. The downside of such approaches as the number of targets increase [144] and the complexity
is the larger form factor that poses power and integration and computational requirements for identifying said targets
constraints. To this end, fully integrated hybrid electronic- increase exponentially [147]. Very fine spatial resolution may
photonic systems have been proposed. Such systems use be achieved by using a pencil-shaped beam operating in the
photonic transmitters and electronic receivers and come with mmWave and THz spectrum. Beamspace channels’ combined
increased requiremetns for synchronization between transmit- angular-delay characteristics are used to detect and locate
ters and receivers [139]. Furthermore, photonics can provide passive objects. In this example, the instantaneous beamspace
extensive network coverage in the base station level, but may be compared with a reference to detect a passive target,
the aforementioned constraints of photonic devices make it and with a prior sample (on-line or real-time) to track a moving
difficult to implement MIMO approaches in the user device target. Target identification relies on learning algorithms that
level. can tell the difference between the varying angular-delay
Graphene and other novel plasmonic materials are being profiles of various targets. High-precision localization and
explored for manufacturing THz devices due to their capacity detection of scattering objects may be enabled by the use
to provide solutions with great reconfigurability [140]. In of specialized beamspace processing signals and the strategic
plasmonics, antenna arrays may be made considerably smaller placement of monostatic/multistatic MAS.
and more versatile since they are characterized by substantially
lower wavelengths [141]. Plasmonic solutions are advanta- D. RIS
geous for adaptable system designs and on-site reconfiguration
due to their small size and frequency-interleaving features. RISs’ role is ever increasing in the communication paradigm
Unlike conventional transceivers, plasmonic ones do not need where their main goal is to increase the data rate via achieving
upconversion or downconversion to function at THz frequen- higher SNR. However, when it comes to localization ap-
cies, where the creation of energy-efficient brief pulses is proaches, RISs are utlized on the one hand to augment the
especially advantageous [142]. The downside of such devices information gathered by the system by introducing curvature
is that due to the power output constraint, they have a short 1 Passive localization methods are capable of localizing targets without
range making them impractical for localization. requiring them to carry a device or tag.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 16

of arrival (COA), while on the other hand they provide geo- of RISs, due to their low-complexity biasing circuits and
metrical diversity by acting as a passive anchor. Specifically, low power requirements [157]. Finally, multiple researcher
the requirement of localization algorithms for numerous base have investigated the plausibility of micro-electro-mechanical
stations (passive anchors) in order to provide an accurate esti- (MEM) devices as a candidate echnology for the realization
mate can be further enhance by the introduction of RISs. This of RISs. However, several issues must be overcome, such as
notion has been investigated in the recent published literature, relatively large footprints, control signals, and switching rates,
which takes advantage of the phenomenon of multipath to to before they can be a viable option.
pinpoint the origin of high-frequency signals [148]. Although
multipath signals can be destructive for communications, re-
E. AI
searchers have been able to take advantage of them and enable
localization and mapping with a single base station (BS) [149]. As we enter the data-rich 6G era, AI is envisioned to play an
It becomes obvious that it is possible to reduce infrastructure increasingly vital role. The design and development intelligent
cost by a great deal by exploiting RISs that act as passive systems and beings that can think, plan, and make optimum
reconfigurable BSs with low energy consumption [150], [151]. decisions based on probabilistic foundations, sometimes in
In mmwave/THz bands high accuracy localization necessi- unpredictable settings, is a vast field [158]. To train models
tates the need to densify the network with a massive num- beyond explicitly coded rules, most modern AI systems rely
ber of anchors in order to counteract blockages. This is of on ML, which enables data-driven interdisciplinary techniques.
course due to the fact that such bands are more susceptible Such data-driven algorithms will be essential to 6G systems
to blockages compared with their sub-6 GHz counterparts. and beyond, opening up new possibilities for not just wireless
However, densifying the network only with active nodes, such communication but also sophisticated localization methods
as small cells and active relays that host power amplifiers, using mmWave, THz, and optical frequencies.
will induce an immense energy consumption due to the power Fingerprinting and the use of classification and regression
amplifier, which is the most power consuming component of techniques were the primary focus of ML approaches to
active nodes. Hence, RISs can also notably reduce the network localization [48], [159]. We anticipate that ML will be used
energy consumption for high-accuracy localization. Morever, more frequently in data-rich and complex localization appli-
active nodes inevitably need to be connected to the power cations; especially, for the GNSS poor indoor and urban out-
grid. There are places though where the power grid cannot door channel conditions, as traditional signal processing and
reach to for reasons of heavy planning/maintenance costs and mathematical techniques are not sufficient to address complex
aesthetics. In such scenarios it’s much more preferable to use issues where we have a large number of noisy and multi-
RISs that may not even need to be connected to the power modal observations, as well as the non-linear characteristics
grid if they are supplied by alternative sources of energy of the signal. Instead, we may mimic the behavior of the
such as solar radiation through solar panels attached to the system—sensor noise and all—using AI techniques. In addi-
RISs in outdoor scenarios [152]. This would of course be tion, in many circumstances, pattern recognition and predictive
feasible due to the much lower consumption of RISs compared models based on ML approaches are needed to achieve high-
with active nodes. Thus, RISs are expected to enable novel level sensing and localization from perhaps high-dimensional
localization approaches that will revolutionize future high- low-level raw observations, such as CSI in massive MIMO
frequency (mmWave, THz, optical wireless) networks. systems. Consequently, mapping systems and localization will
Despite the aforementioned advances, breakthroughs made rely increasingly heavily on the application of statistical AI
in material research focused on mmWave, THz, and optical- approaches to model complicated radio signal properties and
operating materials could provide immense advances in RIS- fuse multiple complementary but noisy sensors. These are
enabled localization [153]. mmWave, THz, and optical- expected to be supplemented by hybrid signal propagation
operating metasurfaces that offer improved reconfigurabil- models that combine conventional physics-based models with
ity and sensing precision are essential for both mmWave, sequential Bayesian models and data-driven learning method-
THz, and optical localization and communications due to the ologies. Furthermore, it will be hard to manually construct
strong blockage and directionality exhibited in the mmWave, the mathematical model, due to the enormous complexity
THz, and optical frequency band. Recent research shows of the predictive function of mapping low-level observations
that mmWave, THz, and optical-operating hypersurfaces that to high-level goal ideas. In order to build approaches with
use a stack of physical and virtual components to provide unprecedented precision and adaptation capabilities for 6G
lens effects and individualized reflections can offer significant localization, ML offers an alternative framework that is based
performance increase [154]. In addition, similar small-scale on cutting-edge DL and probabilistic methods and learns
metasurfaces have been proven capable of efficient mmWave, from data by optimizing or inferring previously unknown
THz, and optical signal steering at a broad variety of angles, parameters [160].
polarization conversion, as well as generation of orbital an- A number of strategies to improve conventional localization
gular momentum [155]. On another note, despite their slow are currently available in the current AI toolkit, based on
clock speeds and capacitance leakage, CMOS-enabled RISs methods, like probabilistic learning and reasoning, as well
offer improved energy efficiency and seamless integration as DNNs. However, a lot of these tried-and-true methods
to existing systems [156]. Similarly, graphene-based meta- have their limits, because of how data-hungry they are and
surfaces offer a promising alternative for the construction how much processing power and labeled training data they
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 17

need. Novel inference approaches and hierarchical models It has been shown that sub 5 GHz functioning ultra-wideband
need to be created and cleverly coupled in order to learn (UWB) radars are useful in such scenarios. Most non-metallic
from restricted, arbitrarily structured, and noisy data. Several materials exhibit minimal attenuation of electromagnetic ra-
current methods might be improved in order to reduce the diation in this frequency region, making it possible to detect
expense of gathering labeled training data. In order to improve devices that otherwise would be obstructed. Moreover, UWB
a supervised solution, semi-supervised learning blends a large radars, by taking advantage of the ultra-wide bandwidth, may
number of unlabeled with a limited number of labeled ones. provide great precision in target localization thanks to their
Additionally, next-generation wireless systems’ localization superior range resolution [164]. These features make UWB
features are expected to be more autonomous and time- radars useful for a wide variety of applications, including
evolving, necessitating the use of adaptive ML algorithms. but not limited to: beamforming in cellular communications;
For example, cooperative localization based on RL [161] and vehicle localization; monitoring critical infrastructure; indoor
crowdsourcing based on FL [160] might assist in overcoming patient monitoring; search and rescue operations; etc.
the difficulties posed by adaptable settings and limited data. Beamforming improvements of the narrow beams are es-
These issues must be resolved in order to properly use AI and sential for terahertz (THz) and millimeter wave (mmWave)
ML approaches as an enabler for the highly dynamic, large- communications systems to generate enough receive signal
scale localization of the future. strength. High beam training overhead, however, makes it
difficult for these systems to serve highly mobile applications
F. Channel charting like extended reality (XR), drone, or vehicle communica-
Applying traditional unsupervised ML-based dimensionality tions [165]. Beam selection is notoriously sensitive to changes
reduction techniques to the field of CSI is what channel chart- in transmitter and receiver locations as well as environmental
ing is all about [54]. Users may be discovered and tracked on a geometry and features. This suggests that learning more about
chart that is generated automatically based on a big collection the environment and the positions of the transmitters and
of CSI samples collected in a certain setting. The lack of receivers can assist solve the beam selection problem. Using
supervision of channel charting makes it useful in cases, when inexpensive radar sensors, such those used for automotive
there is insufficient data to construct an accurate geometric applications, is a quick and easy approach to get this under-
representation of the user’s position and surroundings [162]. standing [166]. This paper’s objective is to solve the beam
Although it is not possible to directly correlate a position on selection issue using radar sensory data, and to do so, it
the chart with a user’s physical location, it provides a pseudo- presents the first AI-based demonstration of radar-based beam
location within the cell that is stable over time and users. prediction under realistic vehicular communication scenarios.
Therefore, the need for specific measurements of the labeled Frequency-modulated continuous wave (FMCW) radars,
CSI datasets that are necessary for fingerprinting methods which can function in a wide range of environmental con-
are eliminated. Several network features, including mmWave ditions, have recently seen widespread use in self-driving
beam tracking and association, grouping in device-to-device automobiles and autonomous robotics. Therefore, it is an
(D2D) situations, resource management, predictive rate and intriguing open topic whether or not these radars can be
cell-to-cell handoffs may benefit from keeping tabs on the employed for reliable simultaneous localization and mapping
pseudo-location. (SLAM) in large-scale areas under harsh weather conditions.
Pseudo-locations have certain advantages over an actual Moreover, due to these advances in radar technologies other
position, but they cannot completely substitute it. Initially, breakthroughs are anticipated in radar odometry [167], [168],
channel charting’s lack of supervision means that there is no estimating mobility [169], landmarks extraction [170], [171],
need for prior information (i.e., geometrical/geographical) in and more [172], [173].
order to achieve self-configuration, which aids in deploying
emergency and temporary networks. Another aspect of pseudo- H. Sensors
locations is privacy, due to the fact that it enables contact A continuously increasing number of IoT devices with a
tracking without the knowledge of the user’s precise location. wide variety of sensors are being deployed to cover a vast array
To begin with, channel mapping was created for scenarios of use cases, ranging between smart cities [174] to industrial
with intense scattering phenomena and where major RF prop- applications [175] and decentralized initiatives to protect the
agation characteristics are anticipated to remain stable (such as planet [176], [177]. Information on the specific location of
massive MIMO). In order to use this method LOS application the measurements taken is crucial for making sense of the
found at high frequencies, it will likely need to be expanded to results. Due to the fact that many applications depend on
jointly handle signals from numerous transmission locations. mobile sensors, the network needs to be able to dynamically
Directly related to the ML foundations of channel charting detect their positions.
are other unresolved concerns, such as the feasibility of When working on a wireless sensor network (WSN) pro-
implementing lifelong learning and the appropriate feature tocol or application, localization is crucial. For instance, in
design for CSI signals. environmental monitoring applications, it is essential to know
the precise location from where readings are taken in order to
G. Radars conduct proper statistical and scientific analysis; without this
Extensive research and development into techniques for location information, the acquired data are essentially worth-
device localization has been performed in recent years [163]. less. Node locations are also crucial to the development of
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 18

such methods. In target identification and geographical routing, Furthermore, ceiling-mounted LEDs have a very high density
node locations are required [178], [179]. Techniques like mo- since their main job is to provide illumination. Therefore, it
bility management [180], mobile sensor deployment [181], and is anticipated that VLP systems would increase accuracy and
topology mapping may also benefit from their utilization [182]. fulfill the tougher standards of indoor settings, while RF-based
There has been a recent uptick in the usage of sensor networks systems are restricted in sensitive regions [189].
for indoor localization [183], [184]. The main VLP techniques utilize multiple image and light
In order to take use of location-based services like navi- sensors, as well as positioning-optimization methods, for
gation, tracking, and monitoring, users may determine their instance filtering, spring and normalization. The proximity
location anywhere in the world using the Global Navigation approach used in [186] predicts the relative position based
Satellite System (GNSS) [49]. The GNSS is a collection of on a known base station or access point and is proven to
navigational satellite systems from various countries, including have a huge variable in proximity, hence it does not match
the Europe’s Global Navigation Satellite System (GALILEO), localization needs in many interior environments. Other ef-
United States’ Global Positioning System (GPS), China’s forts used multiple VLP triangulation methodologies, such
BeiDou, and Russia’s Global Navigation Satellite System as received signal strength (RSS), time difference of arrival
(GLONASS). To locate a user inside or in a hostile envi- (TDoA), time of arrival (ToA), and AoA, to boost position
ronment, GPS and GNSS fall short of expectations. Indoor estimate accuracy without added hardware. Although these
environments are more difficult and intricate than their outside procedures are accurate, they are computationally intensive.
counterparts for a number of reasons. Multi-path propagation Most of the aforementioned contributions agree that nu-
error is caused by interference from several sources, including merous access mechanisms have been suggested to separate
but not limited to ceilings, walls, equipment, and humans. As and analyze transmitted mixed signals, a necessary opera-
a result of obstacles unique to the inside environment, design- tion for a VLP system. Time-division multiplexing (TDM)
ing accurate indoor positioning systems for next-generation and frequency-division multiplexing (FDM) are two common
wireless networks is a formidable task. methods that have been studied in the literature [190]. Demand
Utilizing a satellite-based navigation system is the standard for tighter coordination between individual LEDs, however,
method used for this. However, there are a few problems that increases this complexity cost. When the location data is based
come along with using GPS [91]. For example, they are only on light of varying wavelengths, precision is increased since
effective in areas with good reception for satellite signals, the signals are more easily distinguished. In order to limit
which limits their use in certain outdoor settings and prevents the amount of LEDs utilized for localization, some traditional
them from being used inside. Limiting the number of nodes VLP systems use square waveforms whose spectral content
that have GPS sensors may help with the first two issues. include odd-integer harmonic frequencies. However, a sinu-
These nodes then serve as anchor or seed nodes, guiding soidal waveform has significantly less complicated spectrum
subsequent nodes to their correct locations. Static anchors with information [191]. This allows for a greater number of LEDs
predetermined positions are often used in place of movable to be used, which in turn greatly improves both localization
anchor nodes outfitted with GPS receivers. precision and breadth.

V. S YSTEM MODELS
I. VLP
This section focuses on the system and channel models that
A lot of work has been done to assess the efficacy of various
lay the foundation of the discussed mmWave/THz and optical
tracking approaches suited for mobility control in indoor
localization systems. In more detail, Section V-A discusses the
settings from the perspective of RF technology [185]. None of
proposed mmWave/THz system model, which includes both
the aforementioned improvements may be used in applications,
LOS and RIS-aided NLOS scenarios, while Section V-B delves
where privacy is paramount, such hospitals and petrol stations,
into the VLC-based localization techniques that can be split
since those establishments must instead use isolated VLC
into LOS and optical RIS-aided NLOS system models.
systems. As a further drawback, their precision is limited
by electromagnetic interference and the significant multipath
phenomena brought on by many reflections in barriers, like A. mmWave/THz
walls, furniture, and moving persons [186]. THz and mmWave techniques have been researched ex-
New light-based ways for indoor location without the re- tensively during the past couple of years aiming to reduce
strictions of the RF-based traditional techniques have been the “THz gap”. From a hardware perspective, small-scale
published as a reaction to the recent developments in graphene-based THz transceivers have been demonstrated with
semiconductor-based lighting systems that have made LEDs groundbreaking sensitivity, precision and power. Specifically,
the dominating choice for illumination [187]. Since the optical SotA photonic- and electronic-enabled transceivers have ac-
energy is concentrated on the LOS connection, visible light complished unprecedented signal transmission and modula-
positioning (VLP) devices are more resistant to the multi- tion [139], [192], [193].
path propagation effect. Consequently, the accuracy of the In order to properly configure and analyze the performance
AoA improves. However, in scenarios with LOS blockage, of a system, a channel model with robust accuracy is required.
novel reflecting surfaces, such as optical RISs and mirrors, Channel modelling may make use of either stochastic, deter-
can be used to reinstate high quality communications [188]. ministic, or a combination of such approaches. Initially, the
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 19

RIS and
dBS−U E
τ= + B. (7)
HRIS c
Likewise, B and c are the clock synchronization offset and
the speed of light, respectively, while fk and ck denote the
HBR-RIS HRIS-UE frequency of the k-th subcarrier and the ratio between the
BS central and the subcarrier frequency, respectively. Finally, dBU
represents the distance between the BS and the UE.
HLOS 2) NLOS: Without RIS: When all LOS paths are ob-
UE
structed, the transmitted mmWave/THz signal may arrive
at the user’s device after being reflected on random (i.e.,
HBR-R
obstacles, walls, metallic structures, etc) or carefully placed
HR-UE
(i.e., RISs) objects. Specifically, THz NLOS communication
HNLOS
scenarios are characterized by increased losses and sparsity,
while scattering and diffraction phenomena are introduced
Fig. 4. mmWave/THz localization system architecture.
from rough surfaces [196]. Thus, such channels are commonly
modelled with ray-tracing and stochastic techniques. On the
simplest model, i.e., deterministic, is analyzed for the far-field other hand, mmWave NLOS can be characterized by modeling
MIMO channel in the localization system, which assumes a the stochastic reflection loss and geometrical modeling [197].
planar waves in contrast to the near-field models that assume If we assume M reflectors between the BS and the UE, the
spherical waves. As depicted in Fig. 4, in the uplink, we can channel matrix for each reflector can be expressed as
split the channel matrix, H, of the channel into two distinct (m) (m) (m)

submatrices as HNLOS = ck ρ(m) e−jξ e−j2π∆fk τ aR aTUE , (8)

H = HLOS + HNLOS , (1) where

(m) (m) (m)
with HNLOS and HLOS denoting the NLOS and LOS channels, (m)
λc Kα (R−U E) GBS GUE
ρ = (m)
, (9)
respectively. 4πdR−U E
1) LOS: The LOS channel coefficient is valid in scenarios
where the BS and the user equipment (UE) share an unob- ξ (m) = 2πfc τ (m) (10)
structed communication path. In such cases, the channel matrix and
can be written as in [194] (m) (m)
dBS−R + dR−U E
HLOS = ρe −j2π(f τ −νt)
GBS GUE aBS aTUE , (2) τ (m) = + B. (11)
c
where aBS , aUE , GBS , and GU E represent the steering (m) (m)
In (8), dBS−R and dR−U E denote the distances between the
vectors and the gains of the BS and the UE, respectively. BS and the m-th reflector as well as between the m-th reflector
Moreover, τ , ν, f , and ρ denote the signal delay, Doppler and the UE, respectively. Also, aR represents the steering
shift, frequency, and path gain, respectively. (m)
vector of the reflector, while Kα (R−U E) is the attenuation
In more detail, the path gain can be expressed as in [195] coefficient for the RIS-UE channel.
cKα With RIS: In the case of RIS-aided systems, the previous
ρ= , (3)
4πf d analysis is augmented with RIS-reflected signal coefficients.
with d being the propagation distance and Kα denoting the The reflection of the signal on the RIS modifies its phase and
attenuation coefficient, which is a function of the distance and amplitude with the channel matrix given by [198]
transmission frequency. It is important to highlight that, in HRIS = c2k ρe−jξ e−j2π∆fk τ aBS aTRIS ΩaRIS aTUE , (12)
THz systems Kα can be evaluated based on high-resolution
transmission molecular absorption caused by water vapor and where
other gases in the atmosphere, while in mmWave ones it λ2c Kα(BS−RIS) Kα(RIS−U E)
denotes the atmospheric attenuation. ρ= , (13)
16π 2 dBS−RIS dRIS−U E
Furthermore, by assuming negligible Doppler effect and
ξ = 2πfc τ (14)
identical Kα between the subcarriers, the channel becomes
frequency-flat and can be obtained as and
−jξ −j2π∆fk τ
HLOS = ck ρe e aBS aTUE , (4) τ=
dBS−RIS + dRIS−U E
+ B, (15)
c
where
In (12), Kα(BS−RIS) is the attenuation coefficient for the BS-
λc Kα GBS GUE
ρ= , (5) RIS channel, while Kα(RIS−U E) is the attenuation coefficient
4πdBU for the RIS-UE channel. In addition, dBS−RIS and dRIS−U E
ξ = 2πfc τ, (6) denote the distances between the BS and the RIS as well
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 20

RIS

pRIS
LED RIS
pris
ψk
nk φk
pue UE
BS z' z(z')
z
y (y') y hLOS UE

x' hNLOS
pBS x x
x' pUE y'

Fig. 6. VLC localization system architecture.

Fig. 5. Global vs local coordinate system.

as between the RIS and the UE, respectively. Finally, aRIS an azimuth, φ, and an elevation, θ, angles. The former is
represents the steering vector of the RIS. defined as the angle between the y axis and the projection
Finally, the complete NLOS channel matrix can be obtained of the direction vector on the xy plane, while the latter is the
by summing all the individual reflection components, both angle between the direction vector and the xy plane. Due to
random and RIS ones, as the fact that the AoA/AoD angles can only be measured in
the local coordinate system, the direction vector needs to be
M
X (m) rewritten in the local coordinate system as
HNLOS = HNLOS + HRIS . (16)
m=1
 
cos(φ̃BU )cos(θ̃BU )
3) Localization geometry: From a localization perspective, ṽBU = f (ϕ̃BU ) =  sin(φ̃BU )cos(θ̃BU )  , (22)
the system consists of a base station, a RIS and the end user’s sin(θ̃BU )
device. In order to determine the position of the latter in space,
we need to define a global coordinate system defined as where ϕ and ϕ̃ denote the AoA/AoD angles in the global and
local coordinate system, respectively, while f (ϕ̃BU ) represents
p = [x, y, z]T . (17) the mapping between the direction vector and the AoA/AoD
As discussed in Section IV-D, the channel contains LOS angles. On the contrary, (22) can be expressed in terms of the
as well as NLOS paths generated by RIS, reflector and/or direction vector as
diffractor elements.
   
φ̃BU arctan2(ṽBU,x , ṽBU,y )
To simplify the analysis, a local coordinate system needs to f (ϕ̃BU ) = = , (23)
θ̃BU arcsin(ṽBU,z )
be used, which can be translated to the global one by using
an Euler angle vector and a xyz rotation sequence as with arctan2 (·) denoting the four-quadrant inverse tangent.
This procedure can be used to calculate the global AoA/AoD
pi = Rp̃i + p (18)
values by substituting ṽBU with vBU , while the same method-
with pi and p̃i representing the position of the i-th element ology can be applied to any of the aforementioned channels,
with i = BS, UE, RIS in the global and local coordinate either LOS or NLOS.
system, respectively, while R denotes the rotation matrix and
p is the center of the array. The global and local coordinate
systems are presented in Fig. 5. B. VLP
Furthermore, in the case that the user transmits a signal Visible-light based positioning (VLP) systems are distinct
towards the BS, the direction vector in the global coordinate from their radio-based counterparts in a number of ways. As a
system can be written as result, it is important to develop appropriately designed VLP
pUE − pBS system that incorporate the optical transmission particularities
vBS−UE = , (19) and are both precise and reliable. As depicted in Fig. 6, signals
dBS−UE
in VLP systems travel from an light emmiting diode (LED)
where source to a photodiode (PD) receiver via a wireless channel.
dBS−UE = ||pU − pBS || (20) The LED is often categorised as a Lambertian source [199],
meaning it disseminates signals in accordance with Lambert’s
denotes the distance between the UE and the BS. (19) can be emission law (i.e., when observing a perfect diffuse reflecting
translated in the local coordinate system as surface, the radiant intensity is proportional to the cosine of the
ṽBS−UE = R−1 vBS−UE , (21) angle between the direction of incoming light and the surface).
The following analysis, assumes that the channel between the
The AoA/AoD pairs can be derived based on the definitions receiver and the transmitter may be obstructed, in which case
provided for the global and local coordinate systems by using it enlists an optical RIS to extend the VLP system coverage.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 21

1) LOS: The LOS process may be represented by a Lam- k-th RIS element are located in (xr , yr , zr ), (x, y, z), and
bertian model, which means that after being reflected off of xk , yk , zk , correspondingly, φrk and ψk , which are based on
an object like a wall, the light’s intensity drops to such a low the roll and yaw angles of the k-th RIS element, can be
level that it may be safely disregarded. Changing the reflection evaluated as
route of the optical connection and achieving beamforming (xr − xk )
through a multi-mirror reflection link are both possible thanks cos (φrk ) = sin(γk ) cos(ωk )
drk
to RIS based on mirror arrays, as described in [200]. In the
(yr − yk )
LOS channel the system model can be expressed as in [189] + cos(γk ) cos(ωk ) (29)
drk
y = h x + n, (24) (zk − zr )
+ sin(ωk ),
where x and y represent the transmission and received signal, drk
respectively. Also, n denotes the additive white Gaussian noise and
(AWGN), while h is the channel gain. In more detail, the LOS
(x − xk )
channel gain can be written as cos ψki = sin (γk ) cos (ωk )
dk
A(m + 1) (y − yk )
hLOS = cosm (φ) cos (ψ) To G (ψ) , (25) + cos (γk ) cos (ωk ) (30)
2πd2 dk
where (zk − z)
+ sin (ωk ) .
1 dk
m = − log2 (cos(Φ1/2 )), (26)
2 Finally, the k-th element’s norm vector can be written as
and T
nk = (sin(ψk ) cos(φrk ), cos(ψk )cos(φrk ), −sin(φrk )) . (31)
α2
G (ψ) = , for 0 ≤ ψ ≤ Ψ. (27) As a result, the complete channel gain of the VLP system can
sin2 (Ψ)
be expressed as
Moreover, φ and ψ denote the irradiance and incidence angles,
K
respectively. In addition, G (ψ) is the optical concentrator gain, X
while To represents the optical filter gain. Finally, m is the h = hLOS + hNLOS
k (ωk , γk ). (32)
k=1
Lambertian model index, Φ1/2 is the LED’s half-power angle,
Ψ is the field of view (FOV), α is the PD’s refractive index 3) Localization geometry: To dig a little deeper into the
and A is the PD’s area. VLP approaches, the most common geometrical approaches
2) NLOS: In the NLOS scenario, a RIS that is made up can be categorized in to RSS-based, ToA/TDoA-based and
of Nx × Ny identical rectangular mirrors, each of which is d AoA-based [189]. Specifically, the most well known locali-
in size. Each mirror is mounted such that its optical centre zation RSS-based technique for VLP is trilateration, where,
is in a plane perpendicular to the y axis. To determine the after the RX measures the strength of the transmitted signals,
roll ω and the yaw γ angles, which spin clockwise in the the distance between the RX and each TX can be calculated,
positive direction of the x axis and counterclockwise in the allowing for the drawing of concentric circles whose radii
positive direction of the z axis, respectively, we need two correspond to the measured distances. Existing measurement
mirrors placed at right angles to one another. flaws often cause the recorded distance to be just slightly
The RIS-enabled NLOS pathways can be split into two cat- off from the real ones. Thus, almost always some overlap
egories: (i) the RIS component is a RX that handles the signals exists between the areas where the circles are drawn. In this
from LED; and (ii) to re-transmit the optical signal collected case, the trilateration problem transforms into the least-squares
in the prior stage, the RIS component is assumed to be a estimate [202], the solution to which is provided as a solution
point source and its magnitude is multiplied by the reflection to the system
coefficient. As higher-order reflections have an unaccounted- 2 2
 (x − x1 ) + (y − y1 ) = d21

for effect on VLC systems, we restrict the following analysis 2 2
(x − x2 ) + (y − y2 ) = d22 . (33)
to first-order reflections [201]. Moreover, we assume that the 2 2
(x − x3 ) + (y − y3 ) = d23

LED’s incident light falls on each of the RIS elements that
are called meta-atoms (MAs) at its geometric centre as well As far as the ToA/TDoA-based methods are concerned, they
as that there are K MAs of area D. As a result, the channel also leverage trilateration algorithms since, as light travels at
gain of the k-th RIS element can be expressed as in [188] a fixed speed over a given distance, it generates phase delay
(m + 1)AD in the RX. Such systems need at least three TX, i.e., at least
hNLOS
k =ρRIS
k cosm (φk ) three phase differences, with each one having its own unique
2π 2 (dk )2 (drk )2 (28)
frequency id. The power of each TX can be expressed as
× cos (ψkr ) cos (ψk ) cos (φrk ) To G (ψkr ) ,
Pi (t) = Pc + Ps cos(2πfi t + φ0 ), (34)
where dk and drk represent the distances of the k-th RIS
element from the LED and the PD, respectively, while, ρRISk where Ps , Ps , and φ0 denote modulated signal power, the
denotes the reflection coefficient of the k-th element. More- continuous signal power, and the initial phase of the signal,
over, if we assume that the receiver, the transmitter and the respectively. Also, fi is the modulation frequency of each
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 22

transmitter (i = 1, 2, 3). As a result, the power recieved by the expected value. In more detail, kE[ŵ] − wk2 represents
the PD can be written as the bias, while tr(C(ŵ, w)) is the variance of the location
3
X estimate. Moreover, it is possible to get a worst case estimate
E(t) = K R Pi (t)hi (t) , (35) of the accuracy by assuming that no bias exists and using
i=1 the Cramér-Rao bound, which is illustrated in the following
with R, K, and hi (t) represent the PD’s responsivity, the equation, provided the noise and signal models are known.
constant proportionality, and the channel impulse response, ac- C(ŵ, w)  I−1 (w), (40)
cordingly. The received signal is processed by band pass filters
specialized on each modulated frequency and, afterwards, it is where I−1 (w) is the inverse of the Fisher matrix [204]. Fi-
passed by a frequency down converter comprised by a mixer nally, after performing the substitution between (39) and (40),
and a band pass filter to unify the frequency. Thus, the system based on the available/selected measurement such as ToA,
is capable to use Hilbert transform to calculate the two phase AoA, RSS, and more, the localization lower bound can be
differences (from three transmitters). However, a third phase extracted as
difference is required for accurately obtaining the position of eMSE (ŵ) = E kŵ − wk2
 
the user, which can be obtained by altering the frequency of ≥ tr(C(ŵ, w)) (41)
two out of the three transmitters [203]. As a result, the phase
≥ tr I−1 (w) .


differences can be obtained by
 d −3d It should be highlighted that various published works utilize
 ∆ϕ12 = 2πf1 1c12 2 = arctan (I12 /Q12 )
the root MSE (RMSE) instead of the MSE, which results in
∆ϕ13 = 2πf1 d1 −5dc1
3
= arctan (I13 /Q13 ) , (36)
 d2 −3d1 the following lower bound:
∆ϕ21 = 2πf1 c = arctan (I21 /Q21 ) p
eRMSE (ŵ) = E [kŵ − wk2 ]
with f1 denoting the reference frequency, and p
 ≥ tr(C(ŵ, w)) (42)
Iab = Ea (t)Hilb[Eb (t)] − Hilb[Ea (t)]Eb (t)
, (37) p
≥ tr (I−1 (w)).
Qab = Ea (t)Eb (t) + Hilb[Ea (t)]Hilb[Eb (t)]
where E represents the RSS after the band pass filter and the However, the performance of the system may fall short of the
down conversion, while Hilb[·] denotes the Hilbert transform. Cramér-Rao bound since many realistic estimators are skewed
Finally, the user’s position can be calculated based on the due to signal NLoS propagation and other causes. To this end,
following distances tighter constraints have been introduced, such as the Zik-Zakai,
 −1 Weiss-Weinstein, and Bayesian Cramer Rao bounds [56].
d1 = − 18 2πf
c
tan (I12 /Q12 )+3 tan−1 (I21 /Q21 )
 

 h 1 i

d2 = 13 d1 − tan−1 (I12 /Q12 ) 2πf
c
.
 h 1i B. Precision
1 −1 c
 d3 = d1 − tan (I13 /Q13 )

5 2πf1 In terms of localization precision, it reveals the variance of
(38) the estimated location. The KPI of localization precision was
created to statistically characterize the accuracy as it fluctuates
VI. L OCALIZATION KPI S throughout numerous localization attempts. One example of
Naturally, better location and orientation accuracy and pre- precision KPIs is the geometrical dilution of precision (GDoP)
cision would be at the top of the list, when developing a that measures the variation in localization errors and can be
localization system. But other goals, including complexity, expressed as
coverage and mobility, are crucial to the system’s success as eRMSE (ŵ)
a whole. This section provides the fundamental KPIs shared GDoP = , (43)
eRMSE (d̂)
between SotA localization systems.
with the estimated range and location RMSEs serving as
numerator and denominator, correspondingly. Minimizing the
A. Accuracy
GDoP can provide the optimal fixed node selection and
The accuracy with which a system can estimate its location placement as it model the precision of the location estimation
and orientation is one of the most important KPIs of its with regard to the distribution of the fixed nodes in space.
localization capability. In more detail, the accuracy of a system Another KPI of the localization precision is the localization
is often evaluated by calculating the distance between the real error, which can be measured by two metrics, namely its CDF
location and the predicted one. It can be often measured as and outage. The former represents the likelihood that location
the mean squared error (MSE) or the CDF of measurements estimates will be precise to a specified degree and is given by
bound by a predefined threshold, which can be expressed as
Fe (eth ) = 1 − Pout (eth ), (44)
eMSE (ŵ) =E kŵ − wk2
 
(39) while the latter expresses the likelihood of the localization
= tr(C(ŵ, w)) + kE[ŵ] − wk2 ,
error to overcome a predefined value and can be expressed as
with w, C(ŵ, w) and tr(·) denoting the location, the covari-
ance matrix, and the matrix trace, respectively. Also, E[·] is Pout (eth ) = Pr {kŵ − wk ≥ eth } . (45)
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 23

localization request of the wireless network is defined as the number of nodes,

after which the residual capacity decreases for every new node
added in the system. By setting the residual capacity at k equal
time to 0, the scalability can be obtained as in [205]
response
CR (N = k) = 0, (47)
latency
which leads to
Fig. 7. Localization latency.
X
η Ca = (1 + Γi ) Li (1 + Yi ), (48)
i
In actual situations, the outage and/or CDF reflects the like-
lihood of confidence in the estimated location. In the case where Ca is the available capacity, η denotes the efficiency,
when two localization methods have equivalent accuracy, the while Li , Γi , and Yi represent the average load, the contention
approach that produces the smaller outage and/or larger CDF factor, and the transit factor, respectively.
values is more precise and thus preferred.

C. Latency E. Complexity
In a localization system, the term latency refers to the delay Various aspects of the localization approach, such as the
that occurs between a device making a location request and hardware, the signal detection method, and the algorithm’s
receiving the response, as shown in Fig. 7. This is determined computation requirements, contribute to the overall system’s
by the processing speed of the chosen localization method and complexity. Depending on the specifics of the intended use
the length of the position reference signals utilized. However, case, all or a subset of these types of complexity may be
latency only describes part of the time-related performance. relevant. With regard to hardware, the complexity of hardware
The pace at which a localization indicator is updated is known realization and deployment is directly related to the complexity
as its update rate. This KPI is latency-dependent, in the sense of communication and localization algorithms. While, as far
that its maximum value cannot exceed (latency)−1 , and use- as software complexity is concerned, it can be challenging
case-dependent. to analytically determine the complexity formula of various
localization methods. Thus, it is often considered equal to the
D. Coverage computational complexity of the location estimators. More-
over, in the context of localization, complexity is intertwined
As the distance between each fixed and the mobile network with accuracy and precision. In addition, there is a crucial
node grows, localization performance often decreases. The relation between complexity and the time elapsed between
system’s coverage is the largest geographic region across location updates for a specific mobile node, which is measured
which reliable localization services may be provided with by two KPIs, namely update rate and latency.
specified levels of latency, precision, accuracy, and other KPIs.
Depending on the available localization infrastructures, the
coverage may be broadly broken down into three categories: F. Stability
indoor, local, and global. The probability of coverage, which
represents the probability that a specific user achieves an SINR In high-frequency systems with small beamwidths, where
higher than a certain threshold, is the most common KPI of pointing errors might cause an outage, loss of tracking or deaf-
coverage and can be expressed as ness can be a critical issue. Thus, deafness, which is defined
as the power leakage brought on by the estimate error, is often
Pc = P [γ ≥ γth ], (46) used as a metric in the assessment of localization algorithms.
with γ and γth denoting the SINR of the wireless system and Deafness is represented as a percentage and normalised to the
the threshold that must be satisfied in order to consider the half power beamwidth (HPBW) as in [125]
UE inside the coverage area, respectively.
kx − x̂k
However, the coverage KPI disregards the system’s adapt- D= , (49)
HPBW
ability when expanding the scope of localization. There is a
risk of wireless channel congestion as localization coverage and, thus, the stability can be written as
expands, and the localization system will need to execute
increased computations and measurements. To this end, the S = 1 − D. (50)
KPI of scalability has been introduced, which quantifies the
capacity of a system to accommodate a growing number of end If it reaches 100%, the localization process must restart since
devices. If we consider a wireless networks with N nodes, the estimate failed because the UE is outside the beam.
there is a point after which the residual capacity becomes Therefore, a system’s stability may be thought of as the
negative. This point is the maximum number of nodes, k, variation in localization accuracy over time, particularly in a
that can be supported by the system. Thus, the scalibility mobile setting.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 24

TABLE IV
S OTA L OCALIZATION M ETHODS

Algorithm Application scenarios Use cases Technology enablers Metrics

Triangulation Local trust zones [206], [207] Outdoor [210] mmWave/THz [209] AoA, ToA,
Mapping [208]–[210] Indoor [206], [207], [209], [211]– VLP [212] AoD, ToF,
Sustainable develop. [211]–[213] [216] Sensors [210], [214], [216] RSS, RMSE
Security [214]
Massive twinning [215], [216]
Kalman Local trust zones [217]–[221] Outdoor [217], [219]–[223] mmWave/THz [217], [219]–[224] ToA, DoD,
Filters Massive twinning [222], [223] Indoor [218], [224] RIS [221] RSS, RMSE
Robots [224] Sensors [218]
Compressive Local trust zones [225]–[230] Indoor [231] mmWave/THz [226], [228], [229] RSS, TDoA,
sensing Mapping [231] Outdoor [228]–[230] Sensors [225], [227], [229]–[231] DoA, RMSE
Massive twinning [232] Beamforming [226]
Radar [228] , VLP [232]
Multidimensional Robots [233] Indoor [233], [240], [243] Sensors [233]–[243] ToF, TDoA,
Conventional

Scaling Local trust zones [234]–[242] Outdoor [235]–[237], [239], [242] RSS, RMSE
Massive twinning [243] Non-terrestrial [238], [241]
Direct Local trust zones [244]–[252] Outdoor [244]–[254] mmWave/THz [245], [248], [253] ToA, TDoA,
Localization Sustainable develop. [253], [254] Sensors [244], [246], [247], [249]– AoA, RMSE
[252], [254]
Radar [252]
Swarm Robots & Mapping [255] Indoor [255], [258], [259], [259] Sensors [255]–[268] AoA, ToA,
Intelligence Local trust zones [256]–[267] Outdoor [255]–[257], [260]–[262] RMSE, RSS
Sustainable develop. [268]
Fingerprinting Sustainable develop. [269]–[281] Indoor [269]–[282] Sensors [269]–[282] RSS, RMSE
Mapping [282] Outdoor [283] mmWave/THz [283]
Local trust zones [283]
SLAM Mapping [284]–[291] Outdoor [284], [287]–[290] Sensors [284], [286], [288], [289], RMSE, RSS
Non-terrestial [285] [291]
Underwater [286] Radar [285], [290]
Indoor [291]
K-nearest Sustainable develop. [292], [293] Indoor [292]–[295], [298], [299] mmWave/THz [296]–[298] RMSE, RSS
neighbor Massive twinning [294], [295] Outdoor [296], [297] VLP [299]
Local trust zones [296], [297] Sensors [292]–[295]
Mapping [298], [299]
Support vector Sustainable develop. [300]–[302] Indoor [300], [302], [304], [307]– Sensors [300], [304], [307]–[315] ToA, TDoA,
machine Mapping [303] [317] Radars [302], [303], [305], [317] RSS, RSME
Massive twinning [304]–[306] Outdoor [301], [303], [305] mmWave/THz [301]
Local trust zones [51], [307]–[317] Non-terrestrial [303]
Decision Local trust zones [318]–[322] Indoor [318]–[320], [323], [325]– Radars [318], [323], [324] AoA, RSS,
trees Massive twinning [323], [324] [327] Sensors [319], [320], [325]–[327] RMSE
Sustainable develop. [325]–[327] Outdoor [321], [322], [324] mmWave/THz [321], [322]
Gaussian Local trust zones [328]–[334] Indoor [331], [332], [335], [336], Sensors [331], [335], [336], [338] ToA, RSS,
Processes Sustainable develop. [335], [336] [338] Radars [328], [332], [337] RMSE
Mapping [337], [338] Outdoor [328]–[330], [333], [334] mmWave/THz [329], [330], [333],
[334]
Neural Sustainable develop. [339]–[344] Indoor [339], [341]–[344], [348], Sensors [339]–[344], [347]–[349], RSS, ToA,
networks Local trust zones [345]–[353] [349], [354] [354] RMSE
Massive twinning [298], [354] Outdoor [345]–[347], [350]–[353] mmWave/THz [345], [346], [350]–
[353]
Learning-based

Autoencoders Sustainable develop. [355]–[361] Indoor [355]–[362] Sensors [355]–[361] RSS, RMSE
Local trust zones [362] mmWave/THz [362]
Convollutional Local trust zones [363]–[372] Indoor [159], [363], [365]–[370], Sensors [159], [363], [365]–[370], ToA, RSS
NNs Sustainable develop. [159], [373]– [373]–[376] [373]–[377]
[375] Outdoor [364], [371], [372], [377] mmWave/THz [364], [371], [372]
Massive twinning [376], [377]
Recurrent Local trust zones [378]–[382] Indoor [378], [381], [383]–[387] Sensors [378], [379], [383]–[388] RSS, RSME
NNs Sustainable develop. [383]–[388] Outdoor [379], [380], [382], [388] Radar [381], [382]
mmWave/THz [380]
Unsupervised Local trust zones [48], [49], [316], Indoor [48], [49], [316], [389]– Sensors [48], [49], [316], [389], ToA, RSS,
learning [389]–[391] [393] [392]–[394], [396] RMSE
Massive twinning [392]–[395] Outdoor [390] mmWave/THz [390], [391], [395]
Sustainable develop. [396]
Federated Local trust zones [160], [397]– Indoor [160], [397]–[402] Sensors [398]–[402] ToA, TDoA,
learning [402] Outdoor [160], [397], [398], [400] mmWave/THz [397] AoA, RMSE
Reinforcement Local trust zones [161], [403]– Indoor [161], [403]–[405] Sensors [161], [403]–[405] RMSE, RSS
Learning [405] Outdoor [161], [404], [406], [407] mmWave/THz [406], [407]
Mapping [406], [407]
Transfer Local trust zones [408]–[414] Indoor [409], [410], [412], [413], Sensors [410], [412]–[419] AoD, AoA,
learning Sustainable develop. [415]–[417] [415]–[419] Radar [409] RSS
Mapping [418], [419] Outdoor [408], [414] mmWave/THz [408]
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 25

Prior location knowledge Prediction step




Next timestep Measurements

Location estimate Update step



Fig. 9. Kalman filters architecture.

Fig. 8. Triangulation architecture.

Moreover, Landstrom and Beek in [209] proposed an approach
for transmitter localization suitable for multipath mmWave 5G
VII. L OCALIZATION ALGORITHMS scenarios. The proposed approach combines stochastic ray-
shooting with the triangulation concept, to provide accurate
This section delves deeper into the various localization
localization based on the path relations between the receivers
methodologies and algorithms that exist. Differentiating be-
and the TX. Tazawa et al., in [211], investigated the impact
tween the conventional and learning-based methods is selected
of extending the number of antenna elements on localization
for the following analysis. Learning-based methods refers to
accuracy. In this respect, the triangulation technique is applied
methods that use ML frameworks, hence NNs would fall
to the AoD and RSS measurements originating from multiple
under this umbrella. Keep in mind that distinguishing between
receivers. Steendam in [212] presented a 3D triangulation
the two may be challenging; thus, some of the publications
algorithm based on the maximum likelihood principle. Using
described here might be categorized in multiple ways. Not
the proposed iterative algorithm, each AoA estimate is up-
only that, but fingerprint matching makes use of probabilistic
dated based on the previous position estimate. A method that
approaches. As a result, due to the ambiguous position be-
detects unreliable AoA measurements is presented in [214].
tween ML and analytical solutions, they are categorized under
The proposed method uses the characteristics of estimated
the conventional methods umbrella.
locations of various nodes to detect unreliable nodes and
mitigate localization error.
A. Conventional Moreover, the authors of [215] presented a localization so-
This section provides an overview of conventional algo- lution for WiFi wireless networks, which leverages multipath
rithms for deriving the location of a mobile node in the reflection for estimating a node’s location with respect to
considered area. As also summarized in Table IV, these the receiver. The proposed technique offers orientation infor-
methods include triangulation, Kalman filters, compressive mation and decimeter-level localization based on the AoA,
sensing, multidimensional scaling, direct localization, swarm AoD, and ToF measurements. In [213], the authors presented
intelligence algorithms, and fingerprinting. a RSS-based indoor triangulation approach that employs the
1) Triangulation: Triangulation uses the AoA or AoD incoherent reception of the transmission from a node of
measurements of radio signals exchanged between the mobile unknown location by several identical indoor nodes. Also, the
node and the access points to estimate the location of a nodes are able to communicate with each other in order to
node at the intersection of lines through simple geometric construct an indoor propagation model. Adama and Asutkar, in
relationships [44]. A minimum of two access points is required [216], combined the weighted prediction and grey prediction
to estimate a node’s 2D coordinates. An indicative example of algorithms in order to develop a triangulation approach for
triangulation is shown in Fig. 8. The mobile node estimates the reducing the estimation error of unknown nodes and improving
AoA measurements from three different access points, namely the positioning accuracy. The authors of [210] proposed a
φ1 , φ2 , and φ3 . Using these measurements and the known method consisting of a range-free approach for detecting
locations of the access points, the location coordinates of the symmetric triangulations, which, combined with semidefinite
mobile node can be obtained. programming, can enhance localization accuracy.
The authors of [206] presented a UWB triangulation scheme 2) Kalman filters: KF is a fundamental technique for
that leverages the geometrical properties of the propagation the study of noisy or inaccurate measurements, offering a
path based on the AoD, AoA, and ToA measurements. Ottoy clear knowledge of how a signal evolves over time. The
and Strycker, in [207], developed a triangulation algorithm that KF eliminates random noise before estimating the state of
autonomously selects the appropriate AoA measurements out the monitored process over time. KFs are widely used in
of a set of possible values. In [208], the authors presented navigation, radar applications, and movement control. KF
a geometric triangulation approach that leverages the AoA involves two steps: a) the prediction step, which projects the
measurement between the mobile node and the anchor nodes. current state of the model and associated uncertainties into the
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 26

subsequent time step; b) the update step, where the projection Observation space
is adjusted by calculating the weighted average of the projected

(L1 norm minimization)

state and the measurements. KF-based methods estimate and

Compression

Recovery
fuse multiple approximations of an unknown value in order to
generate an accurate approximation. In this respect, KFs are
able to exploit the estimates from multiple anchor nodes in
order to accurately obtain the location of a mobile node.
The authors of [217] presented an extended KF (EKF) x Φ y x^
method that tracks the location of a node based on the fusion of
time of arrival (ToA) and direction of arrival (DoA) estimates. Fig. 10. Compressive sensing architecture.
The simulation results showed that the joint Doa and ToA
estimation outperforms the DoA-only estimation. In [218], the
authors reported the use of an unscented KF (UKF) for the
ToA-based localization of a mobile node in non-line-of-sight
ments it is possible to obtain a compressed form of the signal
scenarios. The evaluation results revealed a high localization
directly. As a result, even with fast recovery techniques, CS
accuracy in presence of low-variance noise, making it suitable
predicts the reconstructed signal from severely under-sampled
for high-resolution UWB localization. Moreover, Koivisto et
non-adaptive data adequately and can be employed to recover
al. in [222] documented two approaches based on EKF and
various estimates, such as TDOA, RSS, AoA, from the signals
UKF for 3D localization scenarios. The proposed approaches
using fewer signal samples.
leverage a number of densely deployed anchor nodes that
enable the fusion of DoA and ToA estimates. Based on An approach for estimating the TDOA using CS theory
the simulations, both methods can achieve sub-meter scale was presented in [225]. Specifically, the approach utilizes
accuracy, with the EKF-based method slightly outperforming the discrete Fourier transform and a maximum likelihood
the UFK-based one. Similarly, a two-stage cascaded EKF estimator for obtaining accurate TDOA estimates. In [231],
method for mmWave localization scenarios is adopted in [219]. the authors presented an indoor localization approach based
In the first stage, the direction of departure (DoD) is estimated on the received signal strength (RSS) obtained CS process-
at each anchor node, whereas in the second stage, all DoD ing. Additionally, the authors of [226] designed a multi-
estimates are fused into the final 3D location estimates. In level beamforming technique for LoS mmWave links, which
[223], the authors presented a cascaded EKF-based solution leverages CS theory to estimate the signal AoD. An approach
to track the DoA and ToA of various devices. In more detail, for deriving the 2D location coordinates in a VLC environment
the first EKF is employed to generate the individual DoA and was presented in [232]. In particular, CS was employed to
ToA estimates from the anchor nodes based on the reference obtain the location estimates from multiple overlapping light
signals, whereas the second EKF is employed to fuse the DoA beams. Furthermore, Salari et al., in [227], documented a CS-
and ToA estimates from multiple anchor nodes. based approach for obtaining TDOA estimates, which requires
Furthermore, Bai et al. in [224] develop an UKF-based a limited number of signal samples, leading to a significantly
algorithm for obtaining the 2D coordinates of a mobile node lower computational complexity. A CS-aided MIMO scheme
that exploits the packet loss errors to estimate and improve for aerial scenarios was reported in [228]. In more detail, the
the localization accuracy. The simulation results indicate an authors combined MIMO techniques for 2D antenna arrays
average localization error of 0.39m when the packet loss rate and CS-based DoA estimation for 3D target tracking. An aerial
is lower than 90%. Moreover, Kanhere and Rappaport in [220] scenario leveraging CS theory was also investigated in [229]
leverage an EKF to combine the DOA and time of flight (ToF) and [230]. The Bayesian CS approach was applied to estimate
from multipath components in LoS and NLoS environments. ground node locations using RSS samples collected by the
The respective experimental results report an average error of UAV sensors. Particularly, in [229], a trajectory optimization
24.8 cm at 140 GHz. Finally, the authors of [221] consider a method was employed to minimize the errors of CS-based
mmWave localization scenario involving a base station and a location estimation, whereas, in [230], the Bayesian CS-based
RIS and design an EKF-based approach to obtain the location approach was enhanced using an a priori knowledge-aided
of a node using the time-difference of arrival (TDOA) and algorithm.
round trip time (RTT) measurements.
3) Compressive Sensing: Compressive sensing (CS) theory, 4) Multidimensional Scaling: Multidimensional scaling
as illustrated in Fig. 10, provides a novel framework for recov- (MDS) is a mathematical approach for classification and visu-
ering sparse signals under a certain basis, with a significantly alization of high-dimensional data. MDS is applied on a matrix
lower number of samples compared to conventional methods formed by the distances between nodes and, for each node,
(e.g., the Nyquist theorem) [420], [421]. Nonetheless, the it generates a vector of coordinates in the Euclidean space
acquisition of the original signal requires a time-consuming [422]. Consequently, MDS is widely utilized in localization
or expensive process, which appears futile as, during com- problems, in which several nodes have to estimate their relative
pression, the vast majority of the recovered information is locations from the measured distances between themselves.
discarded. CS provides and alternative procedure, which by
acquiring only the required linear and non-adaptive measure- Assuming a network consisting of a number of anchor nodes
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 27

and N mobile nodes, let D denote the N × N distance matrix localization algorithm for obtaining the inverse of the distance

0 d21,2 d21,3 ... d21,N
 between any pair of communicating nodes, based on the RSS
 d22,1 0 d22,3 ... d21,N  metric using MDS, and deriving the network topology. Aiming
 2
 d3,1 d23,1 to solve the geometrical uncertainties imposed by the con-
... d23,N 

D= 0 , (51)
 ..  ventional MDS algorithm, an enhanced MDS algorithm was
 .  presented in [234]. In [235], Saeed and Nam combined MDS
d2N,1 d2N,2 d2N,3 ... 0 and Procrustes analysis to develop a two-phase localization
algorithm for cognitive networks. Specifically, an approxi-
where di,j stands for the shortest distance between the i-th and
mated distance-based approach is employed to maximize the
j-th mobile nodes. Classical all-pairs shortest-path algorithms,
accuracy based on the nodes’ proximity information available
such as the Floyd–Warshall or Dijkstra algorithms, can be used
in the network. The authors of [236] designed a MDS-based
for obtaining the distances.
3D localization algorithm that aims to reduce communication
The rank of D can be reduced by double centering as
and computational overheads towards minimizing energy con-
1 sumption. In [243], Cui et al. introduces a 3D localization
C = − J D J, (52)
2 algorithm that is based on the combination of polynomial
where data fitting and MDS. The algorithm is able to achieve high
1
J = I − ee| , (53) localization accuracy without requiring information about the
N channel noise and the node movement model.
I is the N × N identity matrix, e is a column vector of length Moreover, Jiang et al. in [237] presented a solution for
N with all elements equal to 1, and | is the matrix transpose finding a node’s 2D coordinates through the TDOA mea-
operator. Additionally, surements by minimizing the MDS-based cost function. A
C ' X X| , (54) similar approach was presented in [238], in which the authors
introduce an approach that integrates the relative velocity in-
where X is a N × 2 matrix with each nodes’ coordinates formation, available via Doppler shift measurements between
translated so as the centroid of the network is at (0, 0). mobile nodes, into the MDS cost function. Furthermore, a low-
Therefore, the matrix containing the relative node locations complexity majorization algorithm was leveraged to minimize
can be estimated by minimizing the following formula: the MDS cost function. Morral and Bianchi in [239] reported a
X̂MDS = argmin kC − XX| k2 . (55) distributed algorithm for node self-localization in WSNs based
x on sporadic measurements of the RSS metric. The nodes’
Eigenvalue decomposition can be applied for solving (55) positions can be recovered from the principal components of
as the similarity matrix which is constructed from the squared
C = VΛV| , (56) inter-nodes distances. The authors of [240] designed an indoor
UAV localization algorithm that is based on distance measure-
where V stands for the eigenvector matrix and ments between the access points and the UAV. The distance
Λ = diag(λ1 , λ2 , ..., λN ), (57) measurements are derived through the RSS samples, while two
different localization techniques are utilized, namely the MDS
stands for the respective eigenvalue matrix. Since C is and the weighted centroid localization. In [241], Fan et al. a
positive-definite and symmetric, and combination of fast-clustering and MDS is presented for 2D
rank(C) ≤ 2, (58) localization in mobile networks. In more detail, in the inner-
cluster relative localization stage, the advantages of combining
it can be represented by its two largest eigenvalues. Conse- classical MDS and iterative MDS are leveraged, while in the
quently, inter-cluster coordinate registration stage, the least squares
C ' Vs Λs Vs , (59) method is used to reduce the registration error. In [242],
the authors presented a cooperative 3D localization algorithm
and the solution to (55) can be obtained as based on filtering and MDS techniques, which are employed
X̂signal 1/2 to mitigate any abnormal measurements due to errors.
M DS = Vs Λs , (60)
5) Direct Localization: In the aforementioned approaches,
1/2
with Λs is a diagonal matrix with the two largest eigen- the location of a node is estimated by various parameters such
values. Of note, the obtained position estimate is not referred as the AoA, AoD, TDOA, and ToA. An alternative approach to
to any particular reference system, and will be, in general, addressing localization challenges is direct localization [244].
rotated, translated, and reflected from the physical node ar- In direct localization, the source’s location is estimated straight
rangement. Therefore, the actual mobile node locations can from the data, with no need to estimate any intermediate
be derived by the known positions of the anchor nodes. parameters, like the AoAs or others. Specifically, the echo
MDS-based localization offers two significant advantages; signals from all node pairs are accumulated and the target
the approach is robust to individual error ranges, provided that location is extracted directly. To achieve this, signals, or some
there exists data redundancy in the square distance matrix. function thereof, must be sent to a fusion centre that makes
Also, it provides a closed-form localization estimate, regard- location estimations. The fusion operation is carried out on
less of network size. The authors of [233] designed a relative the signal level, resulting in less computational overhead and,
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 28

therefore, improved performance. Such a topology is often based on the dichotomy and interplay of two conventional
simpler to implement in scenarios, where reduced distances procedures, namely exploration and exploitation [255]. The
are involved. However, for cellular networks, cloud radio goal of exploration is to increase the fraction of the search
access networks (C-RAN) is expected to provide the necessary space that has been uncovered and, hence, reduced to work-
backbone as BSs forward incoming signals to a centralised unit able solutions. The term “exploitation” refers to the swarm’s
that performs the baseband processing [253]. effort to enhance the quality of previously found solutions by
In [245], the authors presented a direct source localization increasing the intensity of its search efforts in the area around
approach based on AoA and ToA estimates, which determines those solutions.
the source location by jointly processing data snapshots ac- In [256] and [257], the authors respectively utilized the
quired at each BS. Moreover, the authors of [246] developed dragonfly and fireworks algorithms for estimating the node
a single-step direct localization algorithm for obtaining the locations that were randomly deployed in a specific area. The
location of multiple stationary nodes by combining AoA and authors of [258] presented a multi-stage localization method
Doppler information. The numerical results indicate that the that leverages the artificial bee colony algorithm. Moreover,
proposed direct localization algorithm can attain the corre- they compared the proposed method against the particle swarm
sponding Cramer-Rao constraint. A direct passive localization optimization in terms of localization accuracy, number of
method that jointly exploits the AoA and TDOA information nodes, and computation time. Monica and Ferrari in [259]
was presented in [247]. By exploiting spatio-temporal pro- investigated the problem of UWB indoor localization in WSNs
cessing, the proposed method did not require prior knowledge by employing the particle swarm optimization algorithm. A
about the number of source nodes. Han et al. in [248] reported two-stage WSN localization approach based on the firefly
a joint near-field and far-field direct localization approach algorithm is proposed in [260]. In the first stage, four anchor
for massive MIMO environments. Specifically, the approach nodes, that were placed at the area edges, were considered,
uses the AoA measurements and divides the BSs into far- whereas, in the second stage, the closest anchor nodes are
field and near-field ones to mitigate localization errors. The considered. Furthermore, the authors of [261] utilized the
authors of [249] developed a TDOA-based direct localization monarch butterfly optimization algorithm for addressing the
approach under the assumption that multiple nodes work with localization problem in WSNs. Also, the same authors in [262]
independent clocks. The approach leverages an expectation- presented an adaptation of the hybridized moth search algo-
maximization (EM) algorithm and a Gauss-Newton algorithm rithm for solving the node localization problem and compared
for coarse and refined parameter estimation, respectively. it against alternative state-of-art algorithms. Singh et al. in
Moreover, in [250], the authors presented a framework [268] focused on the 2D localization problem and compared
for direct localization comprised of beamspace design and the performance of four swarm intelligence algorithms, namely
position determination. In addition, they derived the respective the biogeography-based optimization, the firefly algorithm, the
beamspace direct localization performance bound and present particle swarm optimization, and the H-best particle swarm
a localization algorithm with low computational complexity optimization algorithms.
and communication overhead. A direct localization approach Moreover, in [263], the authors introduced a fuzzy logic
that leverages random spatial spectrum was presented in [251]. model for localization in WSNs. Towards increasing the ef-
According to the approach of [251], the co-channel signals ficiency of the model, the concept of resultant force vectors
were transformed into the spatial spectrum, enabling the is applied, while the particle swarm optimization algorithm
suppression of multipath components. Moreover, a spatial is employed to minimize the irregular deployment effects.
sparse clustering algorithm was utilized to distinguish the The authors of [264] utilized the firefly algorithm and the
nodes in the spatial domain. In [252], the authors developed artificial bee colony algorithm for approximating the distance
a factor graph-based approach for direct localization in dis- of a mobile node from the anchor nodes. Also, a comparison
tributed MIMO radar environments. A graph representation between the two algorithms in terms of computation time and
was carefully designed for the direct localization problem, localization accuracy is presented. Similarly, in [265], the au-
which was solved efficiently through message passing based thors investigated the performance of two swarm intelligence
on expectation propagation and belief propagation. Finally, Bai algorithms, namely the elephant herd optimization and tree
et al. in [254] presented a direct localization algorithm for growth algorithms, with respect to solving the localization
multipath orthogonal frequency division multiplexing (OFDM) problem in WSN environments. In [266], Arafat and Moh
environments. The algorithm utilizes the data observed by utilized the particle swarm optimization algorithm for the
multiple nodes and estimatest the source node in one step by localization of UAV nodes. To increase the convergence speed
exploiting the orthogonality between the noise subspace and of the algorithm, the boundary box technique is leveraged,
the array response vector. leading to smaller initial search space. Akram et al. in [267]
6) Swarm Intelligence: Swarm intelligence is a subset of al- presented an adaptation of the multi-objective particle swarm
gorithms that studies group dynamics, where most approaches optimization algorithm for jointly maximizing the number of
take inspiration from the social behaviours of animals, includ- localized nodes and minimizing the energy consumption and
ing bees, birds, ants, and others. Complex patterns often form computation time.Finally, in [255], the authors compared three
through the coordination of the activities of multiple entities swarm intelligence algorithms, namely the firefly algorithm,
that are relatively unsophisticated in their mode of operation. the artificial bee colony algorithm, and the particle swarm
All swarm intelligence-based approaches use a search strategy optimization algorithm, in SLAM scenarios.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 29

7) Fingerprinting: The fundamental idea behind finger- localization robustness. A crowdsourcing method for indoor
printing is to estimate the mobile node location by compar- fingerprinting-based localization is proposed in [276]. The
ing the received signal fingerprint to a database of known method leverages the RSS data collected from multiple smart-
signal locations that have already been recorded [423]. The phones and generates the database of fingerprints. A similar
database of signal fingerprints has to be generated in advance. method is presented in [277]. Specifically, the method aims
Consequently, fingerprinting consists of two phases, namely to improve localization accuracy by generating the fingerprint
the populating or training phase and the matching phase. In database using the channel frequency response, which can be
the first phase, the area of interest is explored to create a obtained from the CSI.
signal map, whereas, in the second phase, the node location In [278], the authors designed a fingerprinting scheme based
is estimated by comparing the current fingerprint signal with on the achieved data rates, which can be directly obtained by
the fingerprints database. the access points. To mitigate data rate fluctuations, the scheme
Various metrics can be used for generating the signal finger- employs a time-window mechanism that takes into account
print database during the training phase. For instance, using multiple access points and various transmission power levels.
the reference RSS, denoted by RSSR , the (x, y) coordinates In [279], the authors developed a fingerprinting method that
of a node can be expressed as integrates information about the signal propagation effects and
architecture of an indoor area to generate fingerprints using a
F(x, y) = [RSSR (1), RSSR (2), ..., RSSR (N)], (61)
low number of RSS measurements. Li et al. [282] proposed a
where N is the number of access points. crowdsourcing fingerprinting approach that integrates wireless,
For the matching phase, an Euclidean distance matching inertial. Also, they defined the fingerprinting accuracy indi-
method can be leveraged. The method calculates the Euclidean cators and assessed their performance in predicting location
distances between the measured signal fingerprint, denoted by estimation errors and outliers. In [280], the authors introduced
RSSM , and each reference location in the database as follows: a phase decomposition method to calculate the multipath
v
uN phase and apply principal component analysis to derive the
uX 2 fingerprint based on the CSI. The authors of [281] developed a
D=t (RSSR (j) − RSSM (j)) (62) fingerprinting prototype that eavesdrops on smartphone signals
j=1
in order to acquire the CSI. Furthermore, it employs a joint
Finally, a nearest neighbor algorithm can be used for finding outlier detection and clustering approach to detect signal
the closest fingerprint based on (62). changes. Finally, Shen et al. in [283] proposed a fingerprint
Zhou et al. in [269] and [270] investigated the correlation training scheme that leverages information entropy theory
between the access point deployment and the localization and maximum likelihood estimation. The scheme uses uplink
precision based on the RSS and designed a simulated an- channels to extract the link states and the AoA measurements
nealing algorithm for optimizing the access point positions. and applies a weighted mean square error algorithm to estimate
The authors of [271] designed an approach that automatically the node location.
updates the fingerprint database by employing a clustering 8) Simultaneous localization and mapping: The challenge
algorithm to filter out the altered signals to achieve high that simultaneous localization and mapping (SLAM) is tasked
localization accuracy. Using the observed signals and the with solving is building a map of a moving agent’s surround-
calculated location of a node, the fingerprint database is ings, alongside estimating its trajectory. Due to its pivotal
constantly updated without manual intervention. In [272], role in the development of autonomous robots, SLAM has
the authors combined RSS-based fingerprinting with context- attracted a lot of attention from the research community during
aware data regarding the node’s environment (i.e., building the last couple of decades. In simultaneous localization and
floor plan) and demonstrated that the proposed method can mapping (SLAM), the map, which can be found in various
reduce the required access point number and mitigate the formats, such as occupancy grid, point cloud, and more, serves
effects of wireless interference. The authors of [273] developed as the model of the physical world, while its estimate is
a collaboration-based fingerprinting approach, in which several entangled with the inference of the robot’s trajectory. Building
assistant nodes around an unknown node are selected, based a consistent representation of the world requires collecting
on the RSS sequences similarity, and distances between them data from a wide range of sources, regardless of the shape
are used as auxiliary information to improve the positioning that representation takes. For the purpose of analysing the
accuracy. connections between the available data, they are split into data
Moreover, Chen et al. in [274] designed a cooperative frames, each of which can be anything from a collection of
fingerprinting approach that takes into consideration the phys- laser range scan to a series of photographs. The collection of
ical constraint of pairwise distances to refine and improve such data frames presents an additional difficulty in that the
the estimated positions of multiple nodes, thereby increasing obtained measurements are always realtive to the “uncertain”
the robustness against outdated database entries and distance location of the mobile entity that captures them.
errors. In [275], the authors proposed a fingerprint localization Schlegel et al. in [284] presented an open source SLAM
method based on the path-loss measurement for the training system that uses localization and image processing techniques
and matching phase. The two-step method consists of a to generate a virtual 3D map. Four processes were em-
path-loss-based fingerprinting scheme, which aims to improve ployed for map generation, namely image-based triangulation,
precision, and a dual-scanned fingerprinting that guarantees motion estimation, pose-graph-based map management, and
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 30

Initial data Calculate distance Voting for labels classification issue can be provided by
X
ŷo = max 1 (yi = c) , (63)
c
i∈Ko

where 1 (yi = c) is equal to unity in the case of the i-th neigh-

bor’s label is equal to c, while Ko denotes the neighbors of the
Fig. 11. kNN architecture. observed data point. Using the available data and background
information, we can determine an appropriate value for the
design parameter k and specify the set of individuals who
relocalization. A radar-based SLAM method was presented qualify as “neighbors”. As a result the kNNs can be selected
in [285]. The method constructs a map from radar measure- based on a variety of criteria, including the Euclidean distance
ments combining the pose graph and the iterative closes point between locations, which can be expressed as
algorithm for matching the scans. Also, GNSS information v
uN
is leveraged to improve localization accuracy. The authors of uX 2
d (xi , x0 ) = t |xi,k − x0,k | , (64)
[286] reported a SLAM method for autonomous underwater
k=1
vehicles. The method utilizes simulated annealing and applies
an EKF on the depth, compass, and acoustic sensor measure- with xi,k denoting the measurement between the i-th mobile
ments. Rodriguez et al. in [287] emphasized the criticality of and the k-th fixed network node, while N represents the total
the accurate representation and quantification of uncertainties number of mobile nodes. It is important to remember that
to correctly report the associated confidence of the robot’s kNN is a non-parametric model. The new labels are derived
location estimate in every step of the SLAM process. In by comparing the fresh observations to the training examples,
[288], the authors developed an intelligent filter-based SLAM hence it is in the same category as instance-based learning.
approach for enhancing the localization performance of mo- Several cutting-edge ML systems use weighted kNN as a
bile robots. The proposed approach can be also applied to last-ditch effort to forecast the location, with varying proposals
underwater and aerial scenarios using the appropriate radar- on how to define the “distance” between fingerprints and
based measurements. In [289], the authors investigated the weights. Weighted kNN uses kNN as a fusing method, but
relations among the graphical structure of pose-graph SLAM, assigns different weights to the positions of the k-nearest
Fisher information matrix (FIM), Cramer Rao lower bound, neighbors before averaging them. Moreover, kernel functions
and the optimal design metrics. A SLAM framework that and euclidean distance are examples of methods that may be
employs light detection and ranging (LIDAR) and image used to compute the distances and weights, respectively. In
processing techniques for localization was presented in [290]. addition, kNN has been taken into consideration in various
The framework considers both light intensity information and localization methods, notably fingerprint based localization,
geometry information to improve localization accuracy. Chen because of its low computational cost and straightforward
et al. in [291] presented an active SLAM algorithm for mobile structure. Methods for building fingerprint databases, for adap-
robots able to derive a collision-free trajectory. The algorithm tively choosing k, and for generating meaningful similarity
is based on a predictive control framework that leverages metrics have all been the subject of studies pertaining to
graph topology that approximates the uncertainty minimization kNN. Finally, various features have also been employed in
problem as a constrained non-linear least squares problem. kNN. The authors of [292] developed a kNN-based cascading
After applying convex relaxation to the original problem, a localization system, whereby a kNN is first used to identify
convex optimization algorithm and a rounding process based the environment, and then a kNN is used to localization
on singular value decomposition are employed for solving the using a variety of characteristics. They demonstrate that hybrid
problem. features outperform RSS alone in a variety of settings and
highlight the relevance of doing so.
The gathered RSS data are then used to build the radio
B. Supervised learning map, an early example of this method. When the localization
step takes place online, the online RSS measurement from the
This section examines contributions that use supervised three indoor fixed nodes is compared to the previously stored
learning based ML methodologies for tackling the localization RSS dataset. Typically, the Euclidean distance is used for
problem. In more detail, these techniques discussed entail the comparison. However, considerable localization estimate
both traditional ML approaches, such as kNN, SVM, decision errors may occur when employing a fixed k value, as is the
trees, and Gaussian processes, and DL based ones, like NNs, case with the typical kNN, if the network nodes are moved
autoencoders, CNNs, and RNNs. or the RSS value changes. Several papers have investigated
1) K-nearest neighbors: With k-nearest neighbors (kNN), tweaks to kNN as a potential solution to this problem, with
the average position of the k-nearest fixed network nodes a number of techniques being available, including adjusting
from the fingerprint database is used to estimate the target’s the similarity metric [293], [294], and sub-selecting a subset
coordinate. In the localization process as presented in Fig. 11, of the kNN [294], [295]. For instance, the method in [294]
the selected region is where the majority of the k-neighbors clusters the k-nearest fixed network nodes and then uses their
are placed. The label for an observed point in an M -class mean in the ”delegate” cluster to determine the final location.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 31

Original dataset Added separator Transformed data measurement timing and random synchronization mistakes,
the method suggested in [303] employs ToA as fingerprints
and is based on the Gaussian Kernel. Moreover, the energy
decay time is only one example of characteristics that may
be extracted from the channel impulse response and used
as part of a UWB localization system [304]. The goal is
Fig. 12. SVM architecture. to reduce the size of the impulsive response while still
capturing relevant environmental information. For hierarchical
area classification, the features are sent to a support vector
Both the Euclidean distance and the cosine similarity are machine algorithm. Using CSI from multiple sub-carriers, the
proposed as similarity metrics and used to build the weights authors of [307] built a visibility graph to capture frequency
for a weighted kNN in the work cited in [293]. Lastly, the correlations between neighboring sub-carriers for SVM-based
crowd-sourced indoor localization technique in [299] uses localization. next, in [308], the authors documented an SVM
a smartphone’s orientation sensor and an optical camera to and kernel regression based method for localizing and rec-
pinpoint a user’s precise location within a building. In this ognizing activities based on CSI. In this approach, SVM
approach, a crude estimate of the position is derived via performs classification of the target into an activity class while
kNN over RSS from WiFi, speeding up the search space of localization is accomplished by a regression model. Based on
image-based localization, which may be further limited by the CSI of a MIMO system, [309] was able to gather CSI and
information from orientation sensors. transmit it to a centralized server in a device-free localization
Another feature that has been employed for the kNN al- system. Finally, in [310], localization using an accelerometer
gorithm is the AoA. For example, in [296], AoA is used and RSS was employed in WSN target tracking; the RSS based
to pick a subset of k fixed network nodes in a massive- kernel approach produces a coarse position estimate, which is
MIMO OFDM system. However, the channel observations then passed along with the accelerometer measurements to a
are numerous and computationally intensive to store and sort Kalman filter that performs instantaneous localization.
through. To this end, a sparse channel representation based on Localization has seen widespread use of SVM with a great
the angle-delay domain that allows for effective compression variety of kernel approaches. Applying SVM in WSN locali-
has been proposed for database development. It employs two zation for a small number of sensors with known positions
tiers of fingerprint clustering and categorization in order to was one of the first research efforts that use kernel-based
search across multiple fingerprints. A combined angle and localization [51]. The RSS between different sensors is used
delay similarity measure, which is dependent on the amount in SVM trained with a Gaussian kernel to classify the area.
of overlap between the scatterers, is suggested to quantify the At last, more precise coordinates are calculated by averaging
channel observation-fingerprints similarity. Finally, a weighted the sensors’ respective area centers. For indoor mobile phone
kNN is used to accomplish localization based on ToA, AoA, positioning using RSS, [311] employed Gaussian and linear
and the corresponding weights. In this respect, localization kernels. Another approach combined SVM with a radial basis
is achieved by minimizing the Euclidean distance between the function (RBF) kernel and PCA to perform localization [312],
features, while the representation of the angle-delay domain is while the authors of [313] offered a strategy to handle the
created by applying a fast method to retrieve the fixed network devices’ variety by first ranking the RSSs only from trustwor-
nodes and compressing the database [297]. thy fixed nodes and then using SVM to categorize the desired
2) Support vector machine: The support vector machine location. To reduce the number of correlated fixed network
(SVM) is a very potent ML kernel based algorithm. Similar- nodes, [314] presented using real-time node selection. Then,
ity between vectors may be captured by a kernel function, it uses a kernel to assign weights with regard to the similarity
which can be included into instant based learning approaches. between fixed nodes and the corresponding RSS values; based
Intuitively, the distance between two vectors may be used to on the assigned weights it estimates the location.
represent their similarity to one another. One such kernel is The localization problem is very sensitive to the choice of
the radial basis function (RBF), which changes as one moves kernel. To better capture the connection between the coordi-
further away and is used in the Gaussian Kernel. Other kernel nates before localization and the RSS space, kernel canonical
are sigmoid, polymonial, and linear ones. The kernel func- correlation analysis is presented, where the Matern kernel is
tion seeks to identify the optimal hyperplane for classifying utilized for the physical space and the Gaussian kernel is used
entities. Specifically, the hyperplane is selected to increase for the signal space [315]. To account for the training set’s
the gap between the groups. The kernels provide an efficient spatial organization, a kernel that takes use of any potential
means of locating the product of observed features and training connection between the coordinates has been proposed [316].
sessions in higher dimensional spaces, which can improve According to [317], a hybrid kernel consists of two separate
the features’ distinguishability. A high-level illustration of the kernels: a global and a local kernel that together account for
aforementioned SVM method is presented in Fig. 12. the influence of both nearby and far-off fixed network nodes.
Other characteristics have been applied to SVM with various In [305], Import vector machine (IVM) was employed for
kernel approaches. Using Gaussian kernel ridge regression, the NLOS classification for ToA-based ranging in UWB systems;
authors of [300] presented a localization approach based on however, the authors noted that an ISM has less complexity
RSS and ToA. Of note, in order to avoid being affected by and offers a greater classification probability.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 32

Online measurements benefited from the use of decision trees [318], [323]. Many of
1 2 N the articles, however, resort to extensive use of decision trees
to solve the difficulties at hand or apply them to relatively
straightforward categorization issues.
The employment of numerous learning solutions to arrive at
the final solution has been demonstrated to deliver outstand-
ing performance even when constructed as an ensemble of
fundamental ML solutions, such as decision trees. Adaptive
and gradient boosting are two ensemble learning approaches
that have been extensively applied [325]–[327]. When the RSS
Final predicted location distance is not indicative of the actual physical separation
between two points, [326] provided a fingerprinting approach
Fig. 13. Decision trees architecture. that uses gradient boosting to convert raw RSS into features
using a learnt non-linear mapping function. The main idea
is to first build negative and positive pairs of fixed nodes
Other facets of localization had also made use of kernel and afterwards develop a loss function to guarantee that the
approaches, such as a SVM for estimating range errors with similarity is maintained between them. Finally, after training
regard to CSI in an UWB system and an SVM for recognizing the mapping function, weighted kNN was used with the
poses, which is subsequently utilized to find a good match revised mapping function to carry out the localization. On the
with fixed network node [306]. Prior to utilizing particle other hand, adaptive boosting was used for passive localization
filters for location determination, [301] classified spaces as in [325], where phase information from CSI was employed
indoor or outdoor. Using SVR to recreate the RSS values of to build a fingerprint map. This method iteratively refines
the unselected fixed nodes, [302] improves the localization the sample weights of the training sets in order to facilitate
process’s resilience against noise by proposing a method for classification. The best four anticipated nodes’ positions are
classification and node selection. weighted equally to provide a final location estimate.
3) Decision trees: Rule sets for categorizing data are com- Another of the most prominent approaches of ensemble
monly derived via decision trees, which do so by segmenting learning is random forest (RaF), which uses a forest of
the space of possible labels and possible observations. Specif- decision trees to make a prediction. For instance, [324] em-
ically, Fig. 13 outlines a our two-stage positioning approach ployed classification-based localization with multi-path infor-
that makes use of various weighted decision trees. During mation and a RaF to obtain the TDoA through volume cross-
the first stage (training), data is gathered to create a database correlation between CSI values. As a means of improving
known as the radio map. For R reference and A access nodes, localization accuracy, the suggested technique combines ray-
N location and θ direction measurements per node, the radio tracing with empirical data. In order to determine both room
map can be expressed as in [424] predictions and spatial coordinates, the RSS of WiFi was in-
 θ
ϕ1,1 [τ ] ϕθ1,2 [τ ] · · · ϕθ1,R [τ ]
 vestigated in [319]. The offline phase comprises the training of
 ϕθ2,1 [τ ] ϕθ2,2 [τ ] · · · ϕθ2,R [τ ]  ensemble classifiers and a RaF regressor, as well as fingerprint
θ data preprocessing for location prediction. Room prediction,
Ψ = . (65)
 
.. .. .. ..
 . . . .  establishing membership in a soft cluster, and preprocessing
ϕθA,1 [τ ] ϕθA,2 [τ ] · · · ϕθA,R [τ ] are all performed in the online phase. Another example uses a
This data is comprised of location and direction measurements multi-antenna system to construct fingerprints from a variety
from fixed nodes. The ensemble model is constructed using of parameters, including power spectral density (PSD), RSS,
this dataset. In the second stage, which is called testing, the and other statistical data [320]. A RaF is then trained to
mobile nodes make localization requests utilizing online mea- serve as a classifier for each individual feature. The approach
surements. In more detail, inputs such as online measurements employs a number of samples and classifiers to boost the
from the fixed nodes are used to estimate the mobile node’s reliability of position estimations, and an entropy measure is
location, which can be represented as utilized to choose both a stable time instant and a reliable
 θ  classifier. The location is therefore the median of the location
ϕ1,r [q] predictions, which is limited to fall inside the union of the
 ϕθ2,r [q]  projected locations generated by the chosen classifier and the
Ψθr =  , (66)
 
.. chosen time instant.
 . 
4) Gaussian processes: Gaussian process (GP) regression
ϕθA,r [q]
models have been successfully used to forecast the spatial
with r denoting the mobile node’s location. distribution of RSS measurements using minimal known la-
Both the localization system’s accuracy and the time re- beled fixed nodes, which has been proven to increase the
quired to construct and assess the ensemble model are affected precision of localization schemes. Fig. 14 depicts the overall
by the dataset’s size. Moreover, a variety of localization structure of an indicative localization system. In particular,
issues, such as localization based on fixed node selection and the radio map used for localization in this example requires
clustering, coordinate prediction, and LOS identification, have very few training measurements. Therefore, it significantly
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 33

Σ σ
Observation space Σ σ

Optimal prediction model selection

Measurement map construction

Gaussian process regretion

Localization algorithm
Σ σ

Location estimate
Data acquisition
Σ σ

Output
Input
Σ σ
Σ σ

Σ σ

Offline Online
Hidden layers
Fig. 14. Gaussian processes architecture.
Fig. 15. NN architecture.

reduces the required time and energy for deployment and radio with a set of parallel neurons. After calculating a weighted
map construction procedures. Moreover, during the operation combination of the input based on the summation function,
phase, RSS measurements are collected and used to pinpoint Σ, the neuron feeds it into the activation function, σ. Thus,
an exact location of the mobile nodes. we may think of a neuron as a perceptron or an extension of
Multiple alternative ML strategies rely on GPs, which can logistic regression. Each node in one layer is linked to all other
substitute Bayesian methods in the case of posterior distribu- nodes of the next layer in the simplest structures, often called
tion estimation of the labeling function [328]–[330], [335], fully-connected NNs. With non-linear activation functions and
[337]. During the training process, it is important to find at least one hidden layer, NNs may estimate any function with
optimal values for the GP’s covariance and mean, while kernel arbitrary precision, according to the universal approximation
functions, like Matern and RBF, are often used to capture theorem. Much like the logistic regression example, we require
it. For example, in [338], GP regression was used for the an optimization technique even to train a single neuron.
construction of an RSS-based continuous distribution. The Efficiency in training is especially crucial when dealing with
authors used a MLE technique to determine the location of the deep NNs. Backpropagation is a typical technique for training
target based on the RSS data available to them. Researchers NNs, where the loss function’s gradient is back-propagated
examined the quadratic, Matern, and Gaussian kernels in through the layers using the chain rule and then techniques
order to determine which one best captures the relationship like gradient descent may be used to adjust the parameters.
between geographical points. Moreover, in [331], a GP was Various NN-based localization approaches have been of-
used to represent the probability distribution of the RSS values, fered in past literature. Among these reported techniques, NNs
and then for a particular RSS observation, the location was were employed for passive localization, where the calibrated
estimated using a weighted combination of the fixed network phase and amplitude of CSI were used as a hybrid complex
node’s positions in accordance with Bayes rule. An approach input feature to the NN to localize the user [339]. Also,
that applies GP-based dynamic calibration and estimation of in [345], localization was achieved in a mmWave communica-
the radio map with regard to RSS measurements is described tion system by integrating NNs into a collaborative weighted
in [336]. The standard deviation of the trained GP model was least square estimator, while, in [340], the authors employed
used to quantify the accuracy of the predicted position, and RSS data as input to a NN, which subsequently calculates
the final location is calculated using weighted kNN. an estimate for the ranging error; the modified range values
In order to achieve a RMSE near the Cramer-Rao constraint may then be used by a localization technique like least
in massive-MIMO systems, the authors of [330] presented squares. Another instance, [346] utilizes the user’s position
numerical approximation GP techniques, which were validated in an cellular network along with the RSRP of the three
using simulations. Also, several UWB channel characteristics, strongest fixed nodes for the user’s localization. Moreover,
including RMS delay spread, RSS and ToA, were retrieved a three hidden layer NN that leverages RSS from several
from the power delay profile and put to use in a ranging fixed nodes has been recently proposed with its robustness
scheme in [332]. The initial stage is to employ a kernel PCA being increased by data augmentation methods. In addition,
to project the chosen channel parameters onto a nonlinear two separate networks are developed based on the RSS and
orthogonal high-dimensional space; from there, a subset of ToA measurements of a WiFi system, and the final user’s
these projections is used as an input for GP regression to get location was calculated as the weighted average of the two
a ranging estimate. NNs’ outputs [347]. A hierarchical localization technique that
5) Neural networks: Today, neural networks (NNs) are utilized a NN trained for cellular and WiFi indoor location
widely used in ML. This is because they form the basis of has been implemented in [348]. This technique is based on
many of the most effective learning designs. NNs are robust real-time RSS data to recognize the environment during the
computational models that aim to mimic human cognitive online phase before being fed to the matching trained NN.
abilities. Figure 15 presents a visual representation of the As far as MIMO-OFDM systems are concerned, [341]
structure of a NN, which comprises of many layers, each one presented a fingerprint-based localization approach that relies
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 34

on the CSI magnitude from three antennas. This technique Fixed node #1

Localization data preprocessing

Localization predictions
represents fingerprints using the weights of a four hidden
layer NN. For the NNs’ training, a greedy learning approach Fixed node #N

Output
Input
was used that stacks radial basis functions. Following initial
training and supervised tweaking, a NN’s output is a faithful
recreation of the input data. During the online phase, the
observed data likelihood probability is represented by a set Encoder Decoder
of RSS values when the i-th location is true, and then the
Bayes rule is used to get the posterior probability of location i. Fig. 16. Autoencoder architecture.
The destination is determined by averaging all fixed network
nodes equally. A linear modification of the phase value as
a calibration step increases the stability of the phase values, level representation of an device-free localization autoencoder,
as proposed in [342]. Fingerprints may be created from the where the area is partitioned into N grids, each of which
weights of a NN with three hidden layer. has a fixed node. The goal is to predict the position of the
With the use of directional antennas in a WiFi network, the target inside one of these N grids. Pretrained features are
authors of [354] constructed localization fingerprints based on extracted from data using an autoencoder as the foundation
RSS measurements and employed a classifier in the form of a of the proposed network. The localization readings from
two hidden layer NN to determine which fingerprint is most the stationary nodes are used to build the input topology.
similar to the observed values. It is noted that, an improve- Interestingly, the autoencoder’s output is a “copy” of its input,
ment in localization precision may be achieved by using an which delegates autoencoders into the unsupervised learning
approximation of the AoA. For example, the CSI amplitude architectures. The autoencoder may be broken down into three
of a MIMO system can be used as input to several NNs with distinct parts. Firstly, the encoder maps the input to the second
various hyperparameters, with the final location estimation part, namely the hidden state, h, and afterwards, the decoder
being a fusion of the output from all NNs [349]. From the maps h to the output, which can be expressed as
numerous combining strategies that were analyzed, the best y = f (g(x)), (67)
result was provided by taking both a weighted average and a
median of the location estimations. Furthermore, in [343], a where x and y denote the input and output, while f (·) and
framework was presented that integrates characteristics from g(·) represent the encoder and decoder functions.
many communication protocols, including XBee, Bluetooth, The output of the autoencoder is often sent into a loca-
and WiFi, to aid in the tracking of targets. Yaw readings lization module, making this method a common choice for
and RSS values from various nodes are used as inputs to the extracting strong feature representations. The autoencoder is
proposed NN. After the fingerprint probabilities have been cal- utilized to extract high-level features, and the encoder’s output
culated, the output of the Gaussian outliers filtering technique substitutes the random projection of an XLM [355]. Different
can be aggregated to generate the position estimate, which can regularization algorithms that rely on the encoder output may
subsequently be fed into a particle filter for target tracking. be easily included in the training procedure, which involves
Extreme learning machine (XLM) is a NN architecture with a hierarchical tuning procedure where the classifier parameter
a single hidden layer that is presented as an alternative to back- is adjusted before moving on to the encoder parameters. On
propagation, which is used by the aforementioned designs a different approach, a multi-output network that makes use
and requires significantly more time. First-layer weights are of the encoded information to predict the coordinates, floor,
often assigned at random, whereas hidden-layer weights are and building was presented in [356]. The building structure
typically determined using a least-squares fit. Several studies, along with the target’s coordinates were used to generate the
including [350] and [344], aim to exploit the low-complexity location using stacked autoencoders [357]. Another application
of XLM design in localization. In [344], the data was first of stacked autoencoders combined with two fully-connected
grouped offline using k-means clustering, then an XLM was layers of a deep network applies multi-label classification
employed to classify the data to one of the clusters, and finally prediction in an indoor localization scenario [358], while,
a dedicated XLM was trained for each cluster. During the in [359], an autoencoder was used to reduce the dimensionality
active online phase, the XLM of the corresponding cluster of the magnetic field signals and RSS.
was employed once the RSS values have been first categorised. Autoencoders have also been used to evaluate the degree
As part of the offline process in [350], an XLM was trained of similarity between the gathered fingerprints and observed
using data collected from a database of WiFi RSS fingerprints. characteristics in fingerprint-based localization. Bi-modal DL
Fingerprints were still gathered at certain previously identified localization (BiLoc) is a fingerprint localization solution that
hotspots throughout the online phase, and the solution was provides the average amplitude across two antennas and
revised to account for the additional data. The newest XLM AoA estimations as input to autoencoders and utilizes the
was then utilized for coordinate prediction. output weights as fingerprints [362]. In the online phase,
6) Autoencoders: The localization literature makes use the degree of similarity determines the likelihood that the
of autoencoders with a number of pre-existing supervised target is located at a given time and place based on the
solutions, despite the fact that autoencoders are typically observed characteristics recnostructed by the autoencoders.
applied in unsupervised learning. Figure 16 presents a high- The final position of the target is determined by averaging
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 35

despite changes to the input data. Due to the achievable

data and energy efficiency, more complex structures may be
developed, allowing for the benefits of deeper networks to be

Convolution layer #N
Convolution layer #1
Measurement data

Location estimate
Linear layer
realized. Overfitting is also mitigated by limiting the amount

Output
of trainable parameters, allowing the network to be trained on
fewer samples of data.
Since CNNs have been proven effective in computer vision,
various recent publications have developed CNN-based locali-
zation strategies. Several studies restructure the recorded RSS
array into 2D or 3D pictures and pass it to a CNN because
Fig. 17. CNN architecture. of the high number of reported RSS values in metropolitan
regions. With an applied CNN on the RSS picture, the
authors of [363] were able to assign the coordinates to one
the locations of the fingerprints according to their respective of numerous floors and buildings. Improvements to the RSS
weights. In [360], the closest static point was used to determine pictures were achieved by adding correlation values to the
the location by reconstructing the observed signal based on the characteristics in [376]. The augmented image was fed into
autoencoder’s latent variable and position. Moreover, stacked a hierarchical localization structure, where CNNs first predict
autoencoders matching to each fingerprint are constructed the floor number, based on which the corridor number was
to provide a probabilistic position estimate in real-time. To estimated, and finally the coordinates. In order to quantify the
achieve this, this method first attempts to rebuild the observed relative importance of the RSS value to the fingerprint, the
characteristics and then compares the similarity with the radial authors of [373] suggested a set of hybrid RSS characteristics.
basis function [361]. In a smart parking system, time-series bluetooth RSS measure-
Autoencoders have also been used in device-free localiza- ments were utilized for pedestrian and car localization [364].
tion [425]–[427]. The encoder’s output is sent into a classifier Another hierarchical localization approach predicts the coor-
in convolutional autoencoder, which then makes a prediction dinates, floor, and building by combining RSS values from all
about the predicted location of the target [426]. In this exam- fixed nodes over many time occurrences to create an RSS-
ple, the convolutional autoencoder takes as input a picture built time 2D representation [365]. From a temporal perspective,
from the disparity between the target and real-time RSS values. in [366], continuous wavelet transform was employed to
Based on the detected signals from the RSS, the authors generate a 2D time-frequency picture, which was fed in a CNN
of [425] estimated the gesture, activity, and position of the user that predicts the closest reference points to the target. Then,
by denoising the signal through wavelet decomposition, using kNN was used to infer the target’s coordinates.
a sparse autoencoder for dimentionality reduction, and finally High dimensional features may be used thanks to CNNs’
incorporating the learnt features into a regression model and increased efficiency in NNs. A multi-layer CNN that forms
softmax classification. Data transformation is another possible an image using the channel frequency response values at
use of autoencoders. For instance, the device heterogeneity the available sub-carriers at various time instances has been
issue may be addressed by an autoencoder that transforms the presented in [367]. Under this framework, realizations at
characteristics seen by a test device into features that match various antennas can be used at different CNN channels,
to the device used for obtaining the database [428]. whose output was the likelihood of the target localization
7) Convolution neural networks: CNNs are efficient NNs problem. According to [368], by averaging the fixed nodes’
designs that have demonstrated remarkable performance in expected probability and actual locations, it is possible to
computer vision applications. Through the use of parameter estimate the precise position, which is based on the pair-
sharing, CNNs are able to construct more complex networks wise phase difference between antennas. The position was
with a reduced set of parameters, compared to a fully- predicted by combining the locations of the most likely fixed
connected NN, like the one presented in Fig. 17. The con- nodes during the online phase. Next, in [369], the posterior
volution technique as well as the features maps and kernels distribution of the position of the end device was approximated
play a very important role in the performance of CNNs. by a Markov model using the spatio-temporal information
Within the convolution operation, the model’s parameters are learnt by 3D CNNs. It produces two 3D representations of the
multiplied by the input data. Therefore, its data requirements phase and calibrated amplitude of the CSI, where the depth,
and the computational complexity are reduced. At this point, breadth, and height correspond to the phase/amplitude of
it is important to point out two noteworthy characteristics of various TX-RX pairs, sub-carriers, and packets, respectively.
CNNs. On the one hand, connection sparsity is common in Finally, a computationally efficient deep CNN architecture has
CNNs, in which low dimension filters can be used to generate been used for predicting the most likely fixed nodes in a three-
each output value, while, on the other hand, CNN’s sharing antenna system that generates three-channel images. From
of parameters where each pixel uses the same weights. The these channels, two were utilized for the antenna pairwise
translation efficiency and invariance of CNNs are appealing phase differences, while the other for the amplitudes of all
qualities made possible by these attributes. The term “trans- sub-carrier packet samples [370].
lation invariance” is used to describe the CNNs’ resistance to CNNs have been also employed for device-free localization.
translation, which allows them to provide consistent results For instance, in [159], an 1D CNN model that takes advantage
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 36

Regression remember the results of previous calculations. The output of

the RNN can be written as
Fully-connected layer y(t) = σy (Wh h(t) + bh ) , (68)
with Wh denoting the weights of the RNN, while its hidden
state can be expressed as
RNN

RNN
RNN
RNN

RNN

RNN
RNN
RNN
h(t) = σh (Vx x(t) + Vh h(t − 1) + bh ) . (69)
In the previous equations, bh , Vx , and Vh are the RNN’s
parameters, while the activation function is denoted by σ.
RNN

RNN
RNN
RNN

RNN
RNN
RNN

RNN
RNN
RNN
In RNNs, the model’s parameters are preserved across time.
In other words, they are resources that are shared by all links
inside a certain layer and are fine-tuned through time via
back propagation. However, when a gradient is propagated in
RNN

RNN

RNN
RNN
RNN

RNN
RNN
RNN
time, gradient explosion or vanishing can occur and cause said
gradient to increase or decrease, respectively. Another issue
1-to-1 1-to-many many-to-1 many-to-many
is that the model’s effect gradually wanes with time; thus,
it can not keep in mind too much history. Therefore, plain
Extracted features RNNs are not often employed in the real world. In reality,
“gates” in RNN cells are used to account for influence loss or
deterioration over time, with the most prominent architectures
Measurement data being long short-term memory (LSTM) and gated recurrent
units (GRUs). As illustreted in Fig. 18, RNNs come in a wide
Fig. 18. RNN architecture. variety of sizes and shapes, with some even being able to learn
in both directions. According to architecture presented in this
figure, for each fixed node during a specifide time window, a
of either RSS values of packets from multiple antennas has number of features extracted from the raw measurements are
been used or CSI amplitudes from all antennas and sub- computed. Thus a plethora of data is generated from the initial
carries. Another approach, in [374], generated pictures from a measurements. These features are then fed into the LSTM
time series of RSS values and associated continuous wavelet RNNs in order to extract higher level features that capture
transform in order to identify people in a building. From these the essential information of the localization data. Finally,
contributions, it becomes evident that solutions based on CSI the high level features are used as input for the regression,
outperform others when it comes to localization accuracy. which is realized by using fully-connected layers in order to
Moreover, in [375], a CNN was employed to identify NLOS counterbalance overfitting of the algorithm. All in all, using
interference and estimate the range error in ultra-wideband RNNs to monitor a target’s movement over time is a crucial
systems, while, in [377], the authors recommended the use part of localization.
of time-calibrated complicated channel impulse responses to Numerous recent efforts have relied on recurrent neural
build pictures in an industrial context, where the the increased networks that receive a series of RSS values for localization
multi-path propagation limits the communication and locali- purposes. For instance, the floor and building can be estimated
zation range. using a cascaded RNN that receives a succession of RSS
Lastly, CNNs have been successfully applied in massive- values as input [378]. For real-time systems, RSS measure-
MIMO systems. According to [371], a uniform planar array ments accompanied by numerous Bluetooth anchors can be
can benefit from 3D pictures with power values in the delay, used for localization with high efficiency [383]. Another
as well as vertical and horizontal domains. A CNN network is effort monitored the users’ whereabouts from the linked cell
then fed these pictures and trained using the inception module towers’ RSS history and utilized kNN and Markov models
in combination to various kernel sizes, which are imposed to produce synthetic measurements based on the previously
by the distinct sparsity of each domain. In popular Deep NNs observed ones [384], [385]. Moreover, the authors of [386]
designs, such as AlexNet, GoogLeNet, and more, the inception used a sliding window on a series of RSS data, computing five
module is utilized to combine the results of many kernels for features for each access point within that window, which were
more accurate feature extraction [372]. then used as input for an LSTM device as a vector sequence.
8) Recurrent neural networks: Modeling sequential data is In WiFi and cellular networks, RNNs have been used for
a common use of a specific form of NN called a recurrent collecting the RSS fingerprint from several fixed network
neural network (RNN). With natural language processing and nodes over different timesteps and feeding it into a single-
data processing of time series exhibiting a lot of potential. layer RNN, which returns the coordinates [380]. By comparing
RNNs have the ability to carry out tasks, in which the outcome the CSI amplitudes at various sub-carriers and antennas, it is
is reliant on both the past outcomes and the present input. The possible to achieve localization in a MIMO system, where the
results of previous calculations may be viewed as being stored SNR, CSI, and correlation matrix are employed as candidate
in the ”memory” of RNNs. The RNN’s hidden state, h(t), may features for various RNN approaches [381]. After the CSI
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 37

has been pre-processed and shifted, polynomial regression location [48], [390]. To solve this model, a Gaussian mixture
is applied and the output data sequentially into the LSTM model was employed to calculate the received RSS likelihood
system, which is proven to performs better than competing under the assumption of independent normally distributed
systems. Furthermore, predicting the position of UAV BS variables. Next, the models parameters are estimated through
using WiFi and RSS measurements has been reported in [388]. the expectation maximization algorithm that is initialized with
Before being integrated with the other data, the RSS features’ a basic pathloss model and the positions of the fixed network
dimensions are reduced using principal component analysis. nodes in order to address the identifiability issue.
In order to estimate the location, the new feature vector is fed 2) Clustering: On a different note, unsupervised techniques
into a RNN. In a similar arrangement, [382] utilized linear have been successfully applied for constructing the map of a
discriminant analysis to choose the subset of fixed nodes to radio-based system. An instance of such a system develops a
decrease calculation time, using RSS data from the accessible logical floor plan using accelerometer and RSS measurements
WiFi fixed nodes to construct a radio-map. It then uses an from users within the service area [394]. This techniques is
LSTM system to provide a position prediction based on the based on clustering techniques like k-means on RSS stacking
RSS sequences. difference, while each virtual room is assigned a representa-
RNN and CNN hybrids have been mulled upon in the tive localization fingerprint. Another proposed framework for
recent literature [379], [387]. A smooth trajectory of projected unsupervised localization utilizes WiFi, gyroscope, compass,
locations can be generated from sequential measurements in gyroscope, and accelerometer values of naturally moving users
a cellular system [379]. At first, the area is split into cells, and tries to identify some fixed structures in the building (i.e.,
while afterwards pictures are generated with dimensions that elevators, columns, stairs, and more) that greatly influence
correspond to the grid structure and populated based on the their movement. Afterwards, it employs k-means clustering
measurements. The photos are then propagated to a CNN, to extract unique sensor signatures that can increase the
which uses the retrieved spatial characteristics to predict a localization accuracy for users starting from a known spot in
score for each possible location. Next, the outcomes are then the building [316]. Localization can be also achieved by the
used as input to a multi-layer LSTM, which is responsible combination of global-local optimization and a Markov model
for creating the trajectory. With the use of consecutive RSS that fits the RSS traces into the structure of the environment
values from neighboring fixed nodes, localization is achieved. based on unlabeled data and considers the solutions that do
On another note, a 1D-CNN is used to extract features from not violate the signal propagation [395].
consecutive RSS values and is followed by a RNN to record 3) Dimensionality reduction: Since many approaches rely
the temporal correlation [387]. Afterwards, a mixture density on fusion to fulfil the localization procedure, an unsupervised
network is trained on the RNN’s output to discover the learning approach has been proposed that concatenates the
conditional probability distributions of the locations. outputs of the best classifiers to develop a joint location-
weight estimate and an extended candidate location set within
C. Unsupervised learning an unsupervised optimization framework [295]. Forecasts that
are considered more accurate are given more weight, thus
In contrast to supervised learning approaches, unsuper- the user’s actual location should be near the most accu-
vised learning methodologies take advantage of the features’ rate predictions. Lastly, multidimensional scaling has been
underlying structure and distributions enclosed in unlabeled widely employed in dimensionality reduction schemes for
data. Multiple techniques for pretraining NNs reside in the localization [49], [391], [396]. Such schemes create a low
general category of unsupervised learning and are discussed dimensionality projection of the known distances between
in this section. Two examples of architectures that have been wireless nodes that yields the spatial map of the network con-
employed, are deep belief networks and restricted Boltzman figuration. Cooperative localization utilizing wireless signals,
machines. Specifically, the former has been used to learn such as ToA, AoA, RSS and more, has lately been used for
deep features from RSS data, and then feed those features cognitive radio, IoT, and RFID to estimate distances between
into a different ML solution for location estimation [389]. devices [391], [396].
Another application of unsupervised learning techniques is to
be applied in various localization tasks like device mapping,
data filtering, or access point selection in order to provide
D. Federated learning
increased performance.
1) Semi-supervised: On the way towards applying unsu- Federated learning (FL) decouples the capacity to perform
pervised techniques of localization, several researcher have ML from the need of storing the data in the cloud, allowing
developed methods of both semi-supervised and unsupervised edge nodes to develop a shared prediction model cooperatively
nature. One application can estimate the distance of the user while maintaining all the training data locally. This extends
from the access points using RSS values based on a variety the usage of edge device-based prediction models by bringing
of cost functions, while, at the same time, evaluating the model training to the edge node itself. Specifically, FL is
localization precision in terms of the difference between actual implemented by deploying a plethora of mobile nodes, each
and estimated locations [392], [393]. Another solution is of which is in charge of a certain region that may or may not
provided by developing a graph-based model, where latent overlap with its neighbors. In order to perform local training
variables are considered with regard to power levels and of the global parameters, each mobile node gathers a dataset
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 38

Environment
Core network
Reward
Agent

Action
Observable state

State reformulation

Device Device Device

Fig. 19. Federated learning architecture.

based on which it performs the minimization of the NN’s

Fig. 20. Reinforcement learning architecture.
parameters, θk , by solving
X 2
θk = arg min kyi − f (xi ; θ)k2 , (70)
θ
∀{xi ,yi }∈Dk
of mobile devices, the suggested architecture combines cen-
tralized indoor localization with FL to lighten the burden of
2
with kak2 denoting the sum of the square of a’s components fingerprint collecting and save network computational costs
and while maintaining users’ privacy. In order to construct a global
statistical localization model, a DL model is run on each node
yi = f (xi ; θ) + ni , (71)
based on unlabeled crowdsourcing and labeled fingerprint data,
where f denoting the regression function. After this is carried and then shares the calculated updates with a central server.
out by every mobile node working together, the parameters can When compared to other methods of disseminating fingerprint
be defined as the global objective can be expressed in the form data, FL definitely outperforms them in terms of stability,
of a sum. The mobile terminals encrypt the communications privacy protection, and transmission cost [400].
before sending them to the core network, which then uses Similarly, a federated localization strategy for WiFi net-
homomorphic algorithms to decode the messages. Figure 19 works is constructed in [401] using FL. WiFi signals rep-
depicts the FL concept for localization. resenting known landmarks may be used by mobile devices
Advantages of this collaborative framework include less to create local fingerprints, which can then be fed into a
time spent on training and, more crucially, the protection DNN model and subjected to a deep autoencoder to filter
of individual privacy by removing the need to provide sen- out background noise. A centralized server then compiles all
sitive information to the network [397]. Typically, in FL, of the regional weights into one global one. Local updates
it is assumed that a centralized entity would combine the are encrypted using a homomorphic encryption method to
models trained at the edge devices and broadcast back an safeguard the communication channel during the offloading
updated version of the model. Due to the unpredictable nature phase. High precision in localization estimates and safety
of mobile surroundings and the privacy leakage limitations are shown in a laboratory corridor experiment. As another
introduced by the centralized nature of the data processing example, in [402] a federated localization architecture was
architecture employed in object localization, traditional ML suggested for precise IoT localization. FL reduces the privacy
localization approaches may be surprisingly fragile [398]. FL risks associated with location estimation by having numerous
has the potential to provide intriguing solutions for allowing users work together to develop a model using local fingerprint
accurate and private location services. data. Emerging localization services, such as wireless traffic
There is an abundance of literature exploring the usage of prediction using BSs, and mobile indoor GPS localization,
FL in localization contexts; especially, since location data is mobile target navigation and tracking using inertial sensors,
one of the most fundamentally private bits of information. may all benefit from FL’s utility.
For example, a FL technique that uses NNs to anticipate user
coordinates in order to increase the reliability and robustness
of RSS fingerprint-based localization without compromising E. Reinforcement learning
participant privacy has been recently presented [160]. When Reinforcement learning (RL) is among ML’s burgeoning
the users have finished training the model, the central node research fields and it involves teaching an algorithm how to
adjusts the NN weights based on the amount of samples make judgments by letting it make mistakes, while interacting
they were exposed to and sends the resulting model back with the world around it. To optimize their rewards over time,
to the users. In addition, [399] confirmed FL’s promise in agents conduct a series of behaviors in a predefined setting.
indoor localization services by addressing the problems of A conceptual outline of RL-based localization is provided
privacy leakage threats and task learning associated with in Fig. 20. The agent’s abstracted state of the world, or
using a centralized AI server. Using the computing power environment, from which it draws its action choices at time t
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 39

is comprised of both the agent’s location and the RSS reading Distribution without Distribution with knowledge
Physical deployment
knowledge transfer transfer
make up the state. Moreover, the agent evaluates the state
and uses that information to choose what courses of action
to conduct. The “good” activities explored in localization
schemes include traversing the grid while remaining on it.
In accordance with the reward function, the agent will get a
favorable response after taking the desired course of action. Knowledge transfer from the source domain
Theoretically, incentives may be determined by the distance
between the target position and the agent. However, it is im- Fig. 21. Transfer learning architecture.
possible to utilize this approach to handle unsupervised data. A
technique, which build on the premise of isolating landmarks
with accurate position labels and powerful RSS characteristics up the TL-based localization system, namely offline training
for establishing rewards, is proposed as a solution to this and online localization. To be more precise, a number of fixed
problem [403]. In this respect, the agent receives a reward and mobile nodes are installed in advance, taking into account
if and only if the measured RSS matches the known RSS the physical characteristics of the area and the needs of the
characteristic at the landmark location. individual application. RSS readings from the mobile device
1) Markov decision process: For dynamic systems, like may be sent to the surrounding fixed nodes by positioning
resource allocation, wireless communication and localization, the mobile ones. Each fixed node then uploads its data to the
it has attracted a lot of interest. For instance, a bluetooth server, which uses the signal characteristics acquired in order
based system that uses past locations information and RSS to create a fingerprint database.
measurements to determine whether or not the reward is
Another interesting case of TL is dynamic systems, such
achieved, when the solution approaches known RSS values or
as localization applications, that require configuration and
an RP [404]. In order to determine a progressive localization
retraining on each distinct application. In this scenario, the
algorithm, the WiFi localization problem in a WiFi system
overhead of retraining the model when the data is accessible
is viewed as a Markov decision process with a model-free
in both domains is challenging and can be avoided with TL
approach [405]. In this example, states may contain the RSS,
approaches. Based on these, the challenges of TL enabled
action history, and center coordinate values. There are five
localization span across many dimensions, including time,
possible motion directions in the action space, each of which
space, and hardware, where the data distribution may vary
may be used to shift the predicted location window by a dif-
depending on the dimension. Specifically, TL is presented as
ferent set of radius values. Another localization problem that
a pair of optimization problems: the first involves extracting
can be tackled with RL methods is scheduling the exchange of
the underlying semantic manifold of the signal to serve as con-
signals during cooperative localization [161]. As a result, the
straints for the second, which involves labeling the unlabeled
solution treats the measurement choices as operational tasks
data in the target domain. The challenge of TL across devices
and the linkages as autonomous agents. The observations may
might be stated as one of multitask learning. For instance, a
include the nodes that failed to meet a localization quality
manifold regularization-based solution for a temporal TL can
criteria, covariance values, and distance.
be applied to localization approaches where data distributions
2) Deep RL: Another application of RL is to enable self-
vary between time instances, but are often consistent with one
learning localization scenarios in which new information is
another in low dimensions.
learned from the outcomes of previous choices. Compared to
unsupervised multilateration localization, the accuracy of Q- 1) NNs: One of the more recent applications of TL involves
network in a deep RL structure increases by 37% when used the utlization of SotA NN techniques. For instance, deep
to predict the position of a pedestrian device [404]. To better CNNs employ TL between many antenna configurations based
understand human traffic, predicting pedestrian paths can play on the time-domain CSI. After the network is first trained, it
a very important role. This issue has been tackled by designing is possible to retrain the network’s lower layers with less data
a deep RL method that aims at maximizing entropy [407] points to accommodate the altered antenna configuration [408].
and has achieved performance similar to LSTM localization. Another example, the transfer of knowledge across settings is
Another approach, used semi-supervised deep RL for estimat- explored in order to address the issue of localization based
ing the pedestrian distance in localization applications and on RSS [415]. The NN model is pre-trained in one domain
exhibited a 23% increase in precision with regard to supervised before being transferred to another domain, with the two do-
learning approaches. mains being characterized by distinct propagation phenomena.
Specifically, the domains are two floors in the same building,
both of which have the same access point placement and
F. Transfer learning architectural design. It is proven that with only 30% of the
Transfer learning (TL) is utilized to take advantage of the data in the second floor the TL enabled system can achieve the
knowledge learned in one domain to a new domain. This is same level of performance. Finally, deep CNNs for pedestrian
especially useful when it is challenging to acquire data in localization can be combined with TL approaches with great
the target domain, but an abundance of data exists in the success (i.e., 45% increase of the training data and cut down
source domain. Figure 21 depicts the two stages that make on training time by half) [409].
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 40

High dimensionality

Low dimentionality
timestamped traces are used to build the graphs [419]. In
xn
localization data

localization data
f yn addition, to keep the distance estimates based on the wireless
dn,m propagation model, the goal functions can be modified and
|| yn – ym || Gaussian kernels can be used to determine the graph’s weights
f ym between the RSS measurements [412]. Another approach of
xm manifold learning optimizes a time-series graph Laplacian
SVM to generate pseudo-labels, which are then used as part
Fig. 22. Manifold learning architecture. of a learning framework to assist with semi-supervised RSS
localization. In order to strike a middle ground between
the labeled and pseudo-labeled contributions, it incorporates
2) Conventional TL: Conventional TL methods augment manifold regularization into a transductive SVM [413]. In
the real-time measurements gathered from the target domain WiFi, manifold regularization using Laplacian graphs is taken
with labeled fingerprints from the source domain. For instance, into consideration in the solution for extreme learning ma-
a domain-invariant kernel, suitable for use with SVM in chine parameters. Since Bluetooth and WiFi signals display
TL enabled localization, is learned using data from both distinct propagation circumstances, two graph Laplacians are
domains [410]. Metric transfer and metric learning are the developed to reflect the smoothness in Bluetooth and WiFi,
two components of the TL-based methodology that operate respectively [429]. Finally, a Siamese network design, which
together to lessen the burden of offline training in the new consists of two identical NNs used to compare two inputs
setting [417]. The metric learning component maximizes the was used as a semi-supervised or supervised CSI localization
statistical dependency between signal label and feature statis- solution [414]. Because of the impact of geography on the
tics to learn the distance metrics from source domains. Using values of global parameters, a pair of feed-forward NNs is
a method that minimizes the difference in data between the used to transform the input features into more manageable
source and target domains, the metric transfer component location representations.
determines which metric is optimal for the target domain. In
certain cases, the TL result may not take into consideration all VIII. O PTIMIZATION FRAMEWORKS FOR LOCALIZATION
aspects of the surrounding environment in order to recreate the
radio map. To solve this issue, fuzzy C-means clustering has To assess the fundamental bounds of achievable localization
been used to reduce the impact of external factors [416]. performance, problem formulation, design and optimization
are crucial. This section starts by giving the formulation of
3) Manifold learning: The premise that the observed data
the optimization problem based on the intended system goals,
sit on a low-dimensional manifold inside a higher-dimensional
which is presented in Section VIII-A. The system’s high-
space is what distinguishes the discipline of manifold learning
level design considerations are then covered in Section VIII-B.
from others. In the ML world, manifold alignment is widely
Finally, Section VIII-C delves into the optimization of system
used for sharing models and data across disparate collections,
design issues.
assuming a common manifold. For instance, the Siamese
network design presented in Fig. 22 employs two identical
NNs to translate the high-dimensionality input characteristics, A. Problem formulation
xn and xm , to lower dimensional representations, yn and ym , Precision, throughput, signal quality, and energy efficiency
taking into account the fact that location affects the values are among the most important KPIs that must be optimized
of large scale parameters. The network’s objective is to keep in communication systems. When designing a system with
the distance between x and y to a minimum. Finally, in a precision as a goal, the orientation and position error bounds
semi-supervised learning situation, the parameter α is a scaling are put to use for localization. However, these criteria are
value for distance scale matching. still a useful and manageable tool for assessing performance
Data visualization and dimentionality reduction are only two in the asymptotic area, even if it is only applicable when
of the techniques that are applied in manifold learning. For the estimator is efficient [60]. Aside from the goals listed in
example, graphs is one way to approximate the manifold’s Section II, there are several situations in which they become
point relations, while variable manifold alignments can often crucial. However, there is no one, universal definition of an
result in neighborhood graphs with unique distances between aim; rather, objectives must be defined in light of the context
graph nodes [418]. Specifically, in the area of localization, in which they are used in. Compromises are necessary due to
manifold alignment has found multiple applications for con- the various formulations of objective functions, particularly in
structing a radio map based on the observed distances be- the optimization of simultaneous communication and location
tween the nodes. Another example is the Laplacian Eigenmap systems.
manifold alignment method that develops a weighted graph The BSs transmit positioning reference signals in low-
connecting the data points and maintaining the local geome- frequency localization systems, and the accompanying system
try [411]. This technique utilizes the Laplacian eigenvectors architecture are mostly offline. For instance, this is the case
of the constructed graphs that contain the physical relations for antenna array and BS layout design. The improvement in
of the geometrical observations (such as RSS). Moreover, localization performance is a side effect of the beamforming
manifold learning has been utilized under a semi-supervised gain in MIMO systems. However, knowing the location of
approach where both labeled RSS data as well as unlabeled the receivers is essential for beamforming. Thus, combining
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 41

matrix design, online precoding, and resource allocation are however they might also be software processes. Centralized
of paramount relevance. THz systems that rely on array-of- and distributed architectures are often utilized in macro set-
subarrays antenna structure perform precoder/combiner opti- tings with a wide communication distance. The centralized
mization at the subarray level rather than the antenna level. structures may offer greater overall performance with appro-
Therefore, the beamforming angles of the subarray must be priate scheduling, while the scattered ones safeguard user
well-designed together with the data symbols from the RFCs. privacy. Finally, a cluster network comprises of two or more
Also, powerful algorithms are needed for optimizing resource computing devices sharing a single computational task. Such
allocation and RIS coefficients within a dense network. It networks make use of the hardware’s ability to process data
is clear that both online and offline improvements will play in parallel in order to not only boost processing power,
a significant role in the success of future communication but also provide scalability, high availability, and eliminate
systems. The optimization issue is then formulated, and the network dependency on a single device. Nanonetwork settings,
impact of the various factors on the goals of the system is which value low power consumption and short communication
discussed. distances, are well suited to the clustered design [430].
Specifically, the performance requirements for localization There are four possible architectures to increase the system’s
vary depending on the application scenario, as well as the performance in large-scale scenarios: (i) cell-free, (ii) RIS-
related objectives and KPIs. These goals often overlap, and enabled, (iii) 3D and (iv) heterogeneous networks [431]. The
it is necessary to make concessions. Increases in coverage main advantages of cell-free networks include ultramassive
and update rates may have unintended consequences, such as MIMO beamforming and power savings. However, the THz
diminishing accuracy. In order to maximize its effectiveness, a channel’s sparser multipath propagation and low rank restricts
system may have several goals that it must balance. Such a lo- performance [109]. In this scenario, due to the lack of cell
calization system can be modeled under a generic optimization borders, end devices may have a high coverage probability,
problem such as while the localization performance can be augmented by the
X = arg min f (X ) geometrical diversity of the base stations [432]. Next, The
x , (72) network’s coverage area may be expanded and the channel
s. t. g (X ) ≤ 0 reshaped with the use of RISs. THz-frequency RISs are
predicted to have modest footprints due to the utilization of
where f and g denote the objective and constraint functions,
short wavelengths, which can allow a higher versatility in
respectively. Moreover, X represents the variable set that is
deployment. Finally, in a heterogeneous network, many wired
used to achieve the optimal solution for the optimization prob-
and wireless protocols coexist. This is a possible scenario for
lem constructed under each localization scenario. Depending
future networks. In addition to solving the hearing problem,
on the context, a parameter may represent either an aim or a
this kind of multi-band network may also significantly shorten
limitation. For instance, localization accuracy may be used as
the time it takes to go online.
either a goal to achieve or a constraint to be satisfied, such as
The offline design of a system includes hardware selection
spacing, number of devices, beamforming angles, and so on,
to strike a balance between budget and overall system per-
in addition to other KPIs, like throughput, precision, etc.
formance. Antenna polarization, hardware imperfections, and
phase-shifter quantization are all taken into account throughout
B. System design the hardware selection process. The main factors that influence
The localization system’s design is entangled to the use signal quality in THz systems are absorption, blockage, de-
cases in order to fulfil their requirements. These use cases vice density, and antenna design. Understanding the optimal
inform choices on network architectures, cooperative ap- antenna polarization and device density for a THz network
proaches, as well as algorithm development and optimization. has been investigated in [433]. As a rule, directional antennas
1) Network architectures: Under the umbrella of networks, are utilized for localization and transmissions, whereas unidi-
as discussed in the previous sections, three types of chan- rectional antennas are employed during the service discovery
nels have been covered LOS, NLOS, and RIS-enabled ones. phase [434]. Additionally, a quantized model should be con-
However, densification is a key characteristic of future com- sidered in system design due to the discrete nature of phase
munication networks, thus numerous fixed and mobile nodes control and amplitude signals in RISs [435].
should be engaged. There are three distinct kinds of topologies 2) Cooperative approaches: Although energy consumption,
for THz communication systems: centralized, distributed, and as well as computational and time requirements are increased
clustered. In a centralised system, one or more client nodes are due to frequent contacts between the end devices, the loca-
linked directly to a centralised server, which uses client/server lization coverage and precision are enhanced by cooperative
architecture. This is the most often utilised form of wireless localization [56]. For a fair compromise, it is necessary to
system, in which the clients submit requests to the server and specify the related performance KPIs. Collaboration between
receive the answer. On the contrary, distributed networks aim fixed and mobile nodes, data fusion from several kinds of
to eliminate single points of failure by distributing processing sensors [436], [437], UAV-assisted localization [438] are all
across the network and coordinating their efforts. Distributed crucial. Furthermore, in many real-world use cases, the ex-
systems rely on the communication and synchronization be- istence of NLOS propagation makes it difficult for certain
tween the nodes that are dispersed throughout a network. agent nodes to directly connect with adequate fixed nodes for
These nodes often consist of discrete pieces of hardware, localization purposes, which may reduce localization preci-
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 42

sion. This phenomenon can be mitigated by using cooperative established. The design may be optimized for the greatest
localization strategies [56]. On the contrary, non-cooperative localization performance by taking into account environmental
localization requires all mobile nodes to be in constant com- data. It is also important to ensure optimal orientation of
munication with fixed ones, necessitating either a densely base station antennas. Furthermore, higher beamforming and
packed network or a wide coverage area. By facilitating com- angular resolution performance is achieved by increasing the
munication between mobile nodes, cooperative localization array dimensions of the antenna. However, as the number of
enhances accuracy and increases localization coverage beyond antennas increases a more complicated and expensive setup
that of non-cooperative localization. The main problem of is created. Designing an appropriate size subarray is crucial
cooperative localization, which is a parameters estimation when selecting the antenna architecture. Beamforming gain
issue, may be addressed with probabilistic or deterministic and accuracy may be improved by increasing the number
methods. In the first case, the most prominent solutions of active elements per subarray, although coverage can be
involve multidimensional scaling, multilateration, traditional reduced and hearing loss can occur with narrow beamwidths.
linear scaling, and other techniques that often fail in practice To get from an unconnected state to an active one, a
because they presume a Gaussian model for all measurement new device must go through the initial access operation of
errors [439]. As far as probabilistic approaches are concerned, establishing a physical connection with a fixed node [446].
they not only provide location estimates, but also quantify the The initial access process may be seen as localization without
degree of uncertainty associated with those predictions. For any input from the end device. Blockage and deafness are
instance, estimation theory and factor graphs are among the two problems that arise in THz systems because to the small
most notable techniques of probabilistic approaches, with re- beams, making initial access difficult [447]. This necessitates
cent efforts being focused on belief propagation and equivalent well thought-out codebook construction and efficient first-
Fisher information [440]. time-in processes. Therefore, the codebook layout is deter-
3) Algorithm development: It is still unclear whether THz mined by search techniques, which can split into science-
systems benefit from the use of multi- or single-carrier aware, iterative and exhaustive search ones. In the first cat-
modulation schemes. The former appear to be advantageous egory, the beams may be learnt for each partitioned region
in applications, where frequency-flat channels exist, while, to lessen initial access procedure’s delay, given the fact that
in spite of the poor power efficiency and high complexity, the previous user device position or environmental information
multi-carrier systems are still favoured in frequency-dependent is provided [448]. Next, for iterative search, in order to
molecule absorption loss and multipath scenarios. The uti- determine the optimal angular space, hierarchical codebooks
lization of discrete-Fourier transform spread OFDM may be may be created for transmission across progressively narrower
employed as a ready-made solution in order to mitigate the beams [449]. In the last category, the fixed and mobile
PAPR impact [130]. Spatial [441], index [442], hierarchi- nodes utilize beamforming to send and receive messages in
cal bandwidth [443], and orthogonal time-frequency space various directions [450]. In terms of hardware and coverage
modulations [444] are also taken into account for specific practicality, an exhaustive search is the best option for these
scenarios as additional multi-carrier modulations. Additionally, algorithms [451]. However, the discovery latency increases
research on non-orthogonal multiple access at THz is being linearly with beamforming strength. While iterative search
conducted [445]. shortens the time it takes to make a discovery, its narrow focus
When designing wireless systems, the spectrum efficiency comes at a cost.
or data rate may be impacted by tweaking signal characteristics On the contrary to aforementioned optimization approaches,
including packet length, bandwidth, and carrier frequency. when end device location information are available, the three
These values are also critical for localization to accomplish major concepts are active beamforming, RIS coefficient, and
targeted goals. For better route separation in the delay domain, resource allocation. Regarding the first case, if you know
more bandwidth is preferable, but the higher data size and the exact position of the receiver ahead of time, you may
sampling rate must be kept within the hardware’s capabilities. boost the received signal strength by adjusting the beam-
When possible, the messages should be as lengthy as possible forming angles such that they point directly at the receiver.
to maximize energy acquisition while being as short as possi- However, this boost in SNR does not automatically result
ble to minimize delay, particularly in mobile settings. Finally, in better localization results. For more realistic approaches
the localization system’s effectiveness is heavily influenced by the Cramér-Rao bound is applied. For instance, beamforming
the design choices made. optimization and location estimation can be iteratively applied
by using numerous observations [452]. Moreover, the optimum
precoders for tracking the angle of departure and arrival
C. Optimization can be determined by solving a stated convex optimization
When a network architecture with no knowledge regarding problem, if the uncertainty range of the target directions is
the location of the end devices is assumed, it is possible to available [453]. In order to minimize beam assignments, space
optimize the layout, antenna, and codebook based on data resource allocation often makes use of directional antennas and
about the immediate fixed surroundings. In this scenario, the array-of-subarray structures. Accuracy and range are directly
optimal placement of fixed and mobile network nodes may impacted by how well the analog beamforming angles are
be calculated from the Cramér-Rao bound using a predeter- optimized. In multi-end-device environments, the beams must
mined codebook after the quantity of these elements has been be strategically distributed across the various end devices so
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 43

that communication and localization KPIs are met. To improve

the overall system performance in a network with several fixed
nodes, cooperative beamforming optimization is required.
Next, improved signal gain in RIS-assisted systems relies
heavily on both adjustment of the RIS coefficients and active 6G
beamforming. Multiple broadcasts using randomized beam- mmWave/THz
forming angles of messages may be employed for localization 1-10 Gbps localization

when the end device’s location is uncertain. Beamforming 0.1-1 Gbps

5G Optical wireless
localization
angles at mobile and fixed nodes may be simultaneously 2 Mbps 4G
improved using knowledge of the end device’s location and 64 Kbps 3G
RIS-assisted
localization
orientation in advance. For improving the localization and 2G
communication quality, it is possible to tweak the coefficients Heterogeneous
localization
of the RIS [454]. Also, a larger data rate is achievable by systems

optimizing the RIS components to increase the SNR at the World wide Internet of Massive Integrated
Voice calling broadband localization &
receiver, while the higher SNR is not always associated with & messaging web Applications
& IoT communication

a low Cramér-Rao bound. One limitation that has yet to be 1990 2000 2010 2020

solved is 3D MIMO system optimization methods, which are

not currently available. Fig. 23. Future directions for 6G localization.
Finally, allocating resources in a communication network
that is servicing numerous end devices or performing different
In this section, as depicted in Fig. 23, we will address the
activities is a crucial step, with bandwidth, energy, and time
insights and future directions with regard to mmWave/THz
being the three key resources. For bandwidth allocation to be
localization, wireless optical localization, RIS-assisted locali-
feasible, the THz spectrum is split into a series of spectral
zation, heterogeneous localization systems, integrated locali-
windows that change in distance due to the fact that the
zation and communication, as well as low earth orbit (LEO)
vapor absorption coefficient varies with frequency. When the
positioning, navigation and timing (PNT).
connection distance increases, the effective bandwidth window
decreases in size [455]. The hierarchical bandwidth modula-
tion, which optimizes device density to maximize capacity, A. mmWave and THz localization
is thought to be possible in THz communications due to mmWave/THz localization research is still in its infancy,
the effective bandwidth [443]. For localization purposes, it with multiple approaches still to be investigated. In this
is necessary to identify and allocate appropriate sub-bands section, future directions based on fundamental research and
and subcarriers to the end devices at various ranges. As intelligent control design are presented.
far as time slot allocation is concerned, a compromise be- 1) Fundamental research: So far, a deterministic channel
tween overhead and transmission speed must be found when model was investigated. However, KPIs of dispersed signals,
communicating with a single recipient. It stands to reason such as AoA, RSS, and more, are random in practical settings.
that data acquired from multiple transmissions will improve Modeling the influence of randomly generated signals is neces-
localization accuracy, but at the expense of an increase in sary, and the impact of reflectors and scatterers on localization
overhead and latency. Additionally, channel coherence and performance must be measured. Furthermore, since geometric
end device capabilities should be taken into account by the localization relies on the channel model as its basis, current
allocation process. Allocating communication time slots to SotA approaches use an extension of mmWave models by
different users helps ensure that all users within the service including high-frequency system characteristics as the AOSA
region receive the best possible positioning QoS. Finally, with structure, the wideband effect, and atmospheric attenuation.
regard to energy consumption, localization precision is often a However, the channel model may not account for the impact
limiting factor instead of a variable to be adjusted. Specifically, of hardware impairments and other THz-specific factors. Since
in an indoor positioning scenario, the end user requires to the localization performance suffers as a result of these model
know its position with accuracy in the order of centime- incompatibilities, it is clear that a more precise channel model
tres. Therefore, increasing the transmission power to achieve is necessary. Finally, in most localization tasks, the position
higher precision is unnecessary. All in all, intelligent energy and orientation of the BS and RIS are assumed as known
allocation can meet the performance needs of user devices anchors and research is focused on the UEs’ location and
with minimum wasted resources. All of the aforementioned orientation. However, this results in neglecting calibration
resource allocation strategies can be written in the form of problems between the reference anchor and additional ones. To
optimization restriction in a plethora of active and passive overcome this issue, joint UE localization and BS/RIS calibra-
beamforming problems. tion is of major importance in applications with many anchors.
2) Intelligent protocol design: While model-based ap-
proaches are simple to deconstruct, AI-based approaches are
IX. I NSIGHTS & FUTURE DIRECTIONS
often favoured for learning from or mitigating the impact of
So far, important concerns of future wireless systems’ loca- unanticipated model mismatches. In this case, it is important
lization have been explored and intriguing findings obtained. to have access to shared databases so that methods will be
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 44

compared and contrasted. The preservation of user security that documents the performance envelop of RIS-empowered
and privacy, the transfer of a taught model into another or assisted localization systems and/or present suitable pro-
domain to save training time, as well as the gathering, sharing, tocols in this direction. In this section, we briefly present
and storage of massive amounts of data, are all pressing the research gap and provide future directions in RIS-assisted
concerns that need to be addressed in order to successfully localization systems.
deploy future intelligent 6G wireless networks. For example, 1) Fundamental research: The initial theoretical studies
optimising the placement of stationary BSs for better coverage on RIS-assisted localization system assume: i) deterministic
or data accuracy is easy when compared to dynamically channel model, ii) continuous RIS phase shift capabilities,
deployment of the BS by attaching it to UAVs, in temporary iii) ideal RIS with no imperfect meta-atom (MA), iv) analog
heavy traffic settings like stadiums or conference halls. To RIS, v) single-RIS, vi) interference-free systems, and viii)
accommodate the UEs’ communication and localization needs, ideal transceivers front-end. Although these assumptions are
the UAVs’ position and path must be optimised. Connectivity necessary in order to determine the performance envelop of
in dynamic mmWave/THz UAV networks, however, presents RIS-assisted localization systems, in several real-world im-
its own unique set of difficulties that must be thoughtfully plementations, they are not realistic. This drives a theoretical
handled. Moreover, the SLAM method might provide access to campaign in this area that would allow the relaxation of the
a map of the immediate area. The PDF of previous access sites aforementioned assumptions. In more detail, the performance
can be employed for scene-aware localization in addition to the of RIS-assisted localization systems that experience different
map. In order to get around obstructions and make the most of type of fading, such as Rayleigh, Rice, Nakagami-m, Weibul,
the strong reflectors for localization, the beamforming vectors generalized Gamma, Malaga, Gamma-Gamma, etc. need to be
at the BS/UE and RIS element coefficients must be tuned. extracted. Notice that several recent works, including [456]–
The intelligent optimization of the procedure to construct and [460], have verified these type of fading even in THz systems.
maintain a map over time is an open question in this area. In a similar direction, the assumption of continuous phase shift
is not valid in practice, since RISs have a discrete number of
B. Optical wireless localization phases that can use [461]–[464]. As a result, research questions
like how many RIS phases we need for localization arises.
The vast majority of VLP techniques use at least three
Likewise, according to [465], the assumption that all the MAs
LEDs, with the voltage driving them being sinusoidal with
of the RIS are in perfect condition may not be always valid.
very slight frequency changes. These strategies make use of
This motivates an analysis towards the localization approach
the generally valid linear response between the LEDs and
tolerance to MA failures. Different types of RIS exist, such
the PD, as detailed in most manufacturers’ data-sheets. By
as analog, hybrid, and digital. Current contributions only deal
using the RSS parameters and supposing a LOS attenuation
with analog RISs. The benefits from using hybrid and digital
model, these methods can determine the PD’s static location.
RISs need to be analytically quantified. Another research
In addition, similar techniques have been adapted to carry out
question that waits to be answer is concerning the feasibility
position estimate even while the UE is in motion at speeds
and benefits of cascaded and parallel multi-RIS localization
that are more representative of real-world scenarios. However,
systems. Moreover, the impact of interference and transceivers
by averaging the computed location using a small number
hardware imperfections on the performance of RIS-assisted
of substantially overlapping received signals, the possible
localization systems needs to be analytically assessed. Finally,
accuracy may be greatly improved. The reconstructed route
most published contributions present closed-form or analytical
can be smoothed down with the use of AI techniques like
expressions for the Cramer-Rao lower bound or a Fisher
Kalman filter-based tracking. With this in mind, a robust
information analysis. However, a number of other insightful
modelling framework is needed to examine performance over
KPIs exist, such as localization energy efficiency, latency,
a wide range of LED frequency, noise level, UE speed, and
coverage, etc. that are necessary to be assessed.
other design parameters. According to the findings reported
2) Intelligent protocol design: To make the most out of
in the literature, VLP approaches are effective at accurately
RIS-assisted localization system, while providing adaptability
localising the UE, which bodes well for their use both inside
to ever changing conditions, an intelligent protocol that al-
and outdoors. Finally, in order to go beyond the SotA of
lows the exploitation of the systems resources, namely space,
VLP wireless systems, it is necessary to pay close attention to
frequency, and time need to be developed. In more detail,
positional errors, computing efficiency, as well as localization
the protocol should allow co-design of beam code-words for
precision and latency.
both the source and the RIS, selects the optimum bandwidth
and timeslot for localization. Moreover, it should coordinate
C. RIS-assisted localization the overall localization process allowing changing between
RISs are expected to play a key role in the 6G era by ma- beam codewords when needed. Such a protocol should aim at
nipulating the propagation condition, creating and/or canceling maximizing the localization system energy efficiency without
propagation paths, and practically enabling the “propagation- compromising its accuracy and stability. Likewise, the pro-
as-a-service” vision. Although a great amount of effort was tocol should take into account the nature of the system, i.e.
put on analyzing the benefits and performance of RIS as operation frequency band, as well as the type of the RIS, i.e.,
well as designing RIS-empowered wireless communication analog, hybrid, digital. Finally, the protocol should provide
protocols, there are only a limited number of contributions guidelines for the optimal RIS placement. To design such
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 45

protocol, researchers should formulate and solve a number been the different design objectives that these functionali-
of optimization problems and combine their solution to a ties have. In more detail, the localization system aims at
reinforcement learning strategy that will ensure the system’s maximizing its accuracy, while the communication system
adaptability. overall objective is to jointly maximize its reliability, data rate,
and transmission range through the wireless channel, which
experience fading and noise. Therefore, localization systems
D. Heterogeneous localization system design
used to operate in high frequency bands, where the short
To counterbalance the physical limitations of high frequency wavelength and antenna beamwidth allowed them to pin-point
wireless localization systems, simultaneous exploitation of the target with acceptable accuracy. On the other hand, com-
microwave and mmWave, THz, and optical bands is required. munication have only recently started to exploiting mmWave
However, localization systems in different bands have different and THz transmissions. Thus, the time of a localization and
localization range and accuracy. Moreover, their range and communication synergy has come [93], [466]–[468]. In order
accuracy depends on different phenomena, like atmospheric to devise efficient and low-cost integrated localization and
conditions, size of the scatterers and obstracles, and design communication systems a number of strategically designed
parameters, such as antenna gains, beamforming type, etc. steps, starting from suitable KPIs definition and finishing
Additionally, localization mechanisms have to facilitate the to the design of suitable protocols capable of operating in
co-existence of several technologies with different coverage heterogeneous network environments, need to be taken. Next,
that follow different standards. As a result, localization system we identify and analyze these steps.
heterogeneity can be: i) spectrum, and ii) technology hetero- 1) KPIs definition: As discussed in [469]–[472], there are
geneity. several trade-offs between localization and communication
1) Spectrum heterogeneity: refers to scenarios in which system goals. In more detail, in [469], the authors discussed
the localization agent uses both high (e.g., mmWave, THz, the localization accuracy and communication data rate trade-
optical) and low frequencies (e.g., microwaves). The higher the off. In [470], the authors revealed the localization accuracy,
localization frequency, the higher the localization accuracy and power efficiency, data rate trade-off. Similarly, in [471], the
the lower the localization range. This observation drives the authors highlighted the localization accuracy data rate trade-
idea of designing multi-band localizers capable of determining off in single-user mmWave systems, while, in [472], the
the position of the object of interest in a hierarchical manner. prior contribution was extended in multi-user scenarios. The
Specifically, such a localizer would use lower frequency to find aforementioned contributions revealed that there are a number
the location of an object that is outside the high-frequency of trade-offs that need to be understand and balanced. To
system range with low accuracy, and high frequencies to achieve this, the current localization and communication KPIs
increase the localization accuracy of object that are inside may be not enough. This motivates rethinking and redefining
their coverage area. A hierarchical localization approach is or creating new KPIs capable of providing insights that in
required for such a localizer that will allow efficient and fast turns optimize the localization and communication synergy.
localization of the object of interest and boost its accuracy. Moreover, new KPIs may need to be defined that quantify
Finally, an experimentally verified theoretical framework that the performance of specific applications. For example, QoS
quantify the hierarchical localizer performance and provides may be a better KPI instead of localization accuracy. Finally,
design guidelines is needed. for a number of applications in which the reliability of
2) Technology heterogeneity: introduces two scenarios, position information is a key design parameter, such as RIS or
namely: i) stand-alone and ii) integrated localization systems. UAV-empowered wireless systems, the position integrity and
In stand-alone scenarios only one type of localization tech- availability may be the most important KPIs.
nology, such as sensors, radards, mmaWave, THz, optical, 2) Fundamental research: As we move towards mmWave,
is used. On the contrary in integrated localization systems, THz, and optical bands, the performance of integrated locali-
more than one localization technologies are used. In the later, zation and sensing systems are naturally expected to improve.
the identification of the technologies that could combined However, new issues and challenges need to be addressed,
without significantly negatively impacting the localizers en- giving space to fundamental research. Specifically, the re-
ergy efficiency needs to be performed. Another parameter search concerning THz channel modeling as well as integrated
that need to be accounted for is the implementation cost. localization and communication is still at the early stage.
Finally, algorithms for combining localization information Next, we focus in these two topics and provide some possible
from different sources are necessary to be developed. research directions.
Most published contribution in high-frequency integrated
localization and communication neglect the impact of multiple
E. Integrated localization and communication scatters and obstacles and employ deterministic channel mod-
A radio signal can carry both the TX information and els. However, in practice, the AOA as well as the amplitude
localization-related data. As a consequence, a unified frame- of the scattered signals are random processes. It is of high
work for integrated communication and localization seems to importance to model the effect of non-LOS signals and of
be a natural next step. Current cellular networks implement lo- scatters in order to allow the accurate integrated localization
calization functionalities separately from the communication- and communication system performance evaluation. In this
related functions. The main reason behind this approach has direction, new channel modeling methodologies based on
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 46

stochastic geometry, random shape theory and electromagnetic mean square bandwidth that can be obtained as
tools, like ray tracing or the finite element method, need to ´B 2 2
be developed. Other phenomena that should be taken into f |S(f )| df
b = −B
2
´B 2
, (74)
consideration in signal modeling are the wideband effect |S(f )| df
−B
as well as the impact of atmospheric conditions and/or the
transceivers hardware imperfections. Finally, the channel and with S(f ) being the Fourier transform of the transmission sig-
signal model should also account for the existance of smart nal.
materials, such as RISs. From (73) and (74), for the localization perspective, a signal
Building upon the new channel and signal models, a new that has its power concentrated to the edges of the spectrum
theoretical framework for the analysis of the performance of achieves better performance in terms of localization accuracy,
integrated localization and communication systems will be de- i.e., lower Cramer-Rao lower bound, in comparison to signals
veloped. This framework should return not only conventional that their power is either uniformly distributed or concentrated
KPIs that quantify the performance either the localization or at the center of their spectrum. From communications point of
the communication part of the system, but also the novel view, in order to reduce the inter-symbol interference, signals
KPIs that would be defined and allow the assessment of the are designed in a way that their power is concentrated at the
integrated localization and communication system. Moreover, center of their spectrum. This observation reveals that there is
different system models that employ different types of local- a trade-off in terms of the PSD requirements for localization
izers with and without RISs, need to be investigated. The and communication. To find the equilibrium, optimization
derived expressions are expected to bring insights and allow problems for a number of different system models is necessary
the optimization of the integrated localization and communi- to be formulated and suitable policies need to be extracted. Fi-
cation system. nally, adaptability and flexibility in terms of bandwidth, signal
3) Advanced signal processing: Mobile terminals at var- power, PSD shape, etc. should be allowed in order to make
ious tiers of a multi-tiered network may communicate with the most in real-time ever changing wireless environments.
one another using distinct frequency bands and independent 5) Pro-active radio environment mapping and radio re-
communication paths. In such scenarios, efficient and dy- source management: In order to reap the communication
namic maintenance of wireless networks is a key concern for benefits of the proactive radio resource management in areas
integrated localization and communication. Maintenance and like beam alignment, channel prediction, cell selection, and
reconfiguration of wireless connections may be possible via more, high-precision radio environment mappings will be
dynamic resource management and the utilization of multiple necessary. The majority of the radio environment mappings
distinct frequency bands. In addition, cross-layer data ex- models that exist so far has focused on 2D scenarios. With
change depends on efficient signal processing for its operation. this in mind, more accurate 3D localization in nLOS, and
In this respect, the use of common hardware for localization multi-path applications is required for future 6G networks.
and communication services can provide high efficiency for In spite of the fact that a number of effective multi-path
signal processing technologies in integrated communication and NLOS mitigation algorithms have been proposed in the
and localization services, as well as the required reusability of literature, these algorithms are typically quite complex and
channel estimation units for position extraction. As a result, are only practical for remote localization systems as opposed
the most important challenge with respect to advanced signal to self-localization systems where the UE must perform the
processing is achieving efficient co-design of signal processing computation and estimation. However, for the sake of ana-
techniques and hardware architecture. lytical simplicity, these studies often only include first-order
4) Waveform design: Despite the fact that localization and reflection, even though higher-order reflections are common in
communication waveform designs share a number of common dense multi-path settings and might be crucial to localization
key requirements, including low-latency, low transceiver hard- in NLOS situations. As a result, further research is required
ware complexity, and high reliability, in the 4G and the 5G in order to develop more precise and inexpensive localization
eras, signals and systems for localization and communications methods for multi-path scenarios while maintaining the UE’s
were designed separately [473], [474]. As we move towards communication performance and allocating radio resources
higher frequency wireless systems, the idea of designing a based on its current position. Finally,despite the fact that
common signal for localization and communication arises. To precise location data improves communication, a better under-
achieve this we need to determine optimal signal PSDs that standing of the underlying connection between varying levels
satisfies both the requirements of the localization accuracy and of location precision and transmission rates is required.
the communication robustness to interference. In more detail, 6) Design of integrated localization and communication
according to [473] and [475], the Cramer-Rao lower bound of protocols for heterogeneous networks: Future networks will
time-based ranging localization systems can be expressed as become more diverse by using a wide variety of standards
and frequency ranges. For mobile terminals that are constantly
c2 moving under unstable radio conditions, the challenge of
CR = , (73)
8π 2 b2 NPo B how to swiftly transition between protocols is of paramount
importance in order to maintain reliable localization and
where P and B are respectively the transmission power and communication. Since each network layer can use a unique set
bandwidth, while No is the noise PSD. Moreover, b2 is the of protocols, the creation of logical links between them can be
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 47

achieved by considering the protocols as gateways. However, [3] F. Hu, Y. Deng, H. Zhou, T. H. Jung, C.-B. Chae, and A. H. Aghvami,
the need for cross-layer, -module, and -node information trans- “A Vision of an XR-Aided Teleoperation System toward 5G/B5G,”
IEEE Comm. Mag., vol. 59, no. 1, pp. 34–40, Jan. 2021.
fer constitutes the integrated protocol design crucial. Finally, [4] A. Weiss, A.-K. Wortmeier, and B. Kubicek, “Cobots in Industry 4.0: A
a reevaluation of the physical and MAC layers with respect Roadmap for Future Practice Studies on Human–Robot Collaboration,”
to the protocol design that takes into account localization in IEEE Trans. Human-Mach. Syst., vol. 51, no. 4, pp. 335–345, Jul. 2021.
[5] S. Savazzi, M. Nicoli, M. Bennis, S. Kianoush, and L. Barbieri,
addition to the communication metrics is required. “Opportunities of Federated Learning in Connected, Cooperative, and
Automated Industrial Systems,” IEEE Comm. Mag., vol. 59, no. 2, pp.
16–21, Feb. 2021.
F. Low earth orbit (LEO) localization systems [6] K. Sankhe, D. Jaisinghani, and K. Chowdhury, “ReLy: Machine Learn-
ing for Ultra-Reliable, Low-Latency Messaging in Industrial Robots,”
In rural/remote areas that lack terrestrial infrastructure or IEEE Comm. Mag., vol. 59, no. 4, pp. 75–81, Apr. 2021.
when the signals from the terrestrial infrastructure are blocked [7] S. Perdikis and J. d. R. Millan, “Brain-Machine Interfaces: A Tale
of Two Learners,” IEEE Syst., Man, Cybern. Mag., vol. 6, no. 3, pp.
due to obstacles, such as in mmWave/THz transmissions, 12–19, Jul. 2020.
terrestrial- and LEO-based localization can be used in inte- [8] L. U. Khan, Z. Han, W. Saad, E. Hossain, M. Guizani, and C. S. Hong,
grated terrestrial non-terrestrial networks to achieve the best “Digital Twin of Wireless Systems: Overview, Taxonomy, Challenges,
and Opportunities,” IEEE Commun. Surv., vol. 24, no. 4, pp. 2230–
from the two options. In future communication, sensing, and 2254, Aug. 2022.
localization applications, and, especially, positioning applica- [9] H. Zhou, C. She, Y. Deng, M. Dohler, and A. Nallanathan, “Machine
tions with LEO signals, there is currently a dichotomy of Learning for Massive Industrial Internet of Things,” IEEE Wirel.
Commun., vol. 28, no. 4, pp. 81–87, Aug. 2021.
approaches. On the one hand, existing systems can be utilized [10] A. Aijaz, M. Dohler, A. H. Aghvami, V. Friderikos, and M. Frodigh,
to offer novel services to the end users, while, on the other “Realizing the Tactile Internet: Haptic Communications over Next
hand, completely new systems, such as the ongoing efforts Generation 5G Cellular Networks,” IEEE Wirel. Commun., vol. 24,
no. 2, pp. 82–89, Oct. 2017.
towards a LEO-PNT concept, can be designed. In the former, [11] N. Saeed, T. Y. Al-Naffouri, and M.-S. Alouini, “Around the World of
the receiver is responsible for most of the optimization work IoT/Climate Monitoring Using Internet of X-Things,” IEEE Internet of
based on the assumption that the necessary infrastructure is Things Magazine, vol. 3, no. 2, pp. 82–83, Jun. 2020.
[12] H.-C. Yang and M.-S. Alouini, “Data-Oriented Transmission in Future
already in place. In the latter, a three-segment optimization Wireless Systems: Toward Trustworthy Support of Advanced Internet
is required, which can be achieved either simultaneously or of Things,” IEEE Vehicular Techn. Mag., vol. 14, no. 3, pp. 78–83,
in stages. In order to obtain higher precision for localization Sep. 2019.
[13] A. Celik, N. Saeed, B. Shihada, T. Y. Al-Naffouri, and M.-S. Alouini,
with worldwide coverage, future efforts on constructing unique “A Software-Defined Opto-Acoustic Network Architecture for Internet
LEO-PNT constellations through optimization are focused of Underwater Things,” IEEE Comm. Mag., vol. 58, no. 4, pp. 88–94,
on (i) integration of 6G and LEO networks, (ii) intelligent Apr. 2020.
[14] J. I. d. O. Filho, A. Trichili, B. S. Ooi, M.-S. Alouini, and K. N. Salama,
beamforming-based localization, (iii) LEO-based edge com- “Toward Self-Powered Internet of Underwater Things Devices,” IEEE
puting, and (iv) equip LEO satellites with GNSS receivers. Communications Magazine, vol. 58, no. 1, pp. 68–73, Jan. 2020.
[15] A. Pretto, S. Aravecchia, W. Burgard, N. Chebrolu, C. Dornhege,
T. Falck, F. Fleckenstein, A. Fontenla, M. Imperoli, R. Khanna,
X. C ONCLUSION F. Liebisch, P. Lottes, A. Milioto, D. Nardi, S. Nardi, J. Pfeifer,
M. Popović, C. Potena, C. Pradalier, E. Rothacker-Feder, I. Sa,
The potential of the future 6G wireless system for lo- A. Schaefer, R. Siegwart, C. Stachniss, A. Walter, W. Winterhalter,
calization was investigated in this article. In more detail, X. Wu, and J. Nieto, “Building an Aerial–Ground Robotics System for
the key localization-based applications and use cases for Precision Farming: An Adaptable Solution,” IEEE Rob. Autom Mag.,
vol. 28, no. 3, pp. 29–49, Aug. 2021.
the next generation of wireless networks were discussed, [16] X. Liu, S. W. Chen, G. V. Nardari, C. Qu, F. C. Ojeda, C. J. Taylor,
and the technologies that enable localization services were and V. Kumar, “Challenges and Opportunities for Autonomous Micro-
investigated. Moreover, models of mmWave/THz as well as UAVs in Precision Agriculture,” IEEE Micro, vol. 42, no. 1, pp. 61–68,
Feb. 2022.
VLP localization wireless systems were presented , while [17] Y. Zhong, T. Bi, J. Wang, J. Zeng, Y. Huang, T. Jiang, Q. Wu, and
also taking into account both LOS and NLOS channels that S. Wu, “Empowering the V2X Network by Integrated Sensing and
were generated by traditional reflectors as well as RISs. Also, Communications: Background, Design, Advances, and Opportunities,”
IEEE Network, vol. 36, no. 4, pp. 54–60, Oct. 2022.
the major localization KPIs were described and mathemati- [18] A.-A. A. Boulogeorgos and G. K. Karagiannidis, “Low-cost Cognitive
cal expressions were provided when possible. Furthermore, Radios against Spectrum Scarcity,” IEEE Technical Committee on
SotA traditional and learning-based localization techniques Cognitive Networks Newsletter, vol. 3, no. 2, pp. 30–34, Nov. 2017.
[19] ——, “Energy Detection in Full-Duplex Systems With Residual RF
were presented alongside their conceptual figures. Finally, the Impairments Over Fading Channels,” IEEE Wireless Commun. Lett.,
impact of all design aspects of future 6G wireless systems, vol. 7, no. 2, pp. 246–249, Apr. 2018.
namely characteristics, assumptions, localization problem for- [20] A.-A. A. Boulogeorgos, “Interference mitigation techniques in modern
wireless communication systems,” Ph.D. dissertation, Aristotle Univer-
mulation, and optimization, were discussed, while lessons sity of Thessaloniki, Thessaloniki, Greece, Sep. 2016.
learned and future directions were extracted. [21] M. Matthaiou, O. Yurduseven, H. Q. Ngo, D. Morales-Jimenez, S. L.
Cotton, and V. F. Fusco, “The Road to 6G: Ten Physical Layer Chal-
lenges for Communications Engineers,” IEEE Comm. Mag., vol. 59,
R EFERENCES no. 1, pp. 64–69, Jan. 2021.
[22] X. Yang, M. Matthaiou, J. Yang, C.-K. Wen, F. Gao, and S. Jin,
[1] W. Saad, M. Bennis, and M. Chen, “A Vision of 6G Wireless Systems: “Hardware-Constrained Millimeter-Wave Systems for 5G: Challenges,
Applications, Trends, Technologies, and Open Research Problems,” Opportunities, and Solutions,” IEEE Commun. Mag., vol. 57, no. 1, pp.
IEEE Network, vol. 34, no. 3, pp. 134–142, Oct. 2020. 44–50, Jan. 2019.
[2] M. Maier, A. Ebrahimzadeh, S. Rostami, and A. Beniiche, “The [23] J. Yang, S. Jin, C.-K. Wen, X. Yang, and M. Matthaiou, “Fast Beam
Internet of No Things: Making the Internet Disappear and “See the Training Architecture for Hybrid mmWave Transceivers,” IEEE Trans.
Invisible”,” IEEE Comm. Mag., vol. 58, no. 11, pp. 76–82, Oct. 2020. Veh. Technol., vol. 69, no. 3, pp. 2700–2715, Jan. 2020.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 48

[24] K. Ntontin, A.-A. A. Boulogeorgos, D. G. Selimis, F. I. Lazarakis, [44] C. Laoudias, A. Moreira, S. Kim, S. Lee, L. Wirola, and C. Fischione,
A. Alexiou, and S. Chatzinotas, “Reconfigurable Intelligent Surface “A Survey of Enabling Technologies for Network Localization, Track-
Optimal Placement in Millimeter-Wave Networks,” IEEE Open Journal ing, and Navigation,” IEEE Commun. Surv., vol. 20, no. 4, pp. 3607–
of the Communications Society, vol. 2, pp. 704–718, Mar. 2021. 3644, Jul. 2018.
[25] K. Ntontin, D. Selimis, A.-A. A. Boulogeorgos, A. Alexandridis, [45] J. A. del Peral-Rosado, R. Raulefs, J. A. López-Salcedo, and G. Seco-
A. Tsolis, V. Vlachodimitropoulos, and F. Lazarakis, “Optimal Re- Granados, “Survey of Cellular Mobile Radio Localization Methods:
configurable Intelligent Surface Placement in Millimeter-Wave Com- From 1G to 5G,” IEEE Commun. Surv., vol. 20, no. 2, pp. 1124–1148,
munications,” in 2021 15th European Conference on Antennas and Dec. 2018.
Propagation (EuCAP), Apr. 2021, pp. 1–5. [46] M. F. Keskin, A. D. Sezer, and S. Gezici, “Localization via Visible
[26] L. Yang, X. Li, S. Jin, M. Matthaiou, and F.-C. Zheng, “Fine-Grained Light Systems,” Proc. IEEE, vol. 106, no. 6, pp. 1063–1088, May
Analysis of Reconfigurable Intelligent Surface-Assisted mmWave Net- 2018.
works,” IEEE Trans. Commun., vol. 70, no. 9, pp. 6277–6294, Jan. [47] R. C. Shit, S. Sharma, D. Puthal, P. James, B. Pradhan, A. v. Moorsel,
2022. A. Y. Zomaya, and R. Ranjan, “Ubiquitous Localization (UbiLoc): A
[27] A.-A. A. Boulogeorgos, J. M. Jornet, and A. Alexiou, “Directional Survey and Taxonomy on Device Free Localization for Smart World,”
Terahertz Communication Systems for 6G: Fact Check: A Quantitative IEEE Commun. Surv., vol. 21, no. 4, pp. 3532–3564, Oct. 2019.
Look,” IEEE Veh. Technol. Mag., pp. 2–10, Dec. 2021. [48] F. Zafari, A. Gkelias, and K. K. Leung, “A Survey of Indoor Localiza-
[28] A.-A. A. Boulogeorgos and A. Alexiou, “Error Analysis of Mixed THz- tion Systems and Technologies,” IEEE Commun. Surv., vol. 21, no. 3,
RF Wireless Systems,” IEEE Commun. Lett., vol. 24, no. 2, pp. 277– pp. 2568–2599, Apr. 2019.
281, Feb. 2020. [49] N. Saeed, H. Nam, T. Y. Al-Naffouri, and M.-S. Alouini, “A State-
[29] A.-A. A. Boulogeorgos, S. E. Trevlakis, and N. D. Chatzidiaman- of-the-Art Survey on Multidimensional Scaling-Based Localization
tis, “Optical Wireless Communications for In-Body and Transdermal Techniques,” IEEE Commun. Surv., vol. 21, no. 4, pp. 3565–3583, Jun.
Biomedical Applications,” IEEE Commun. Mag., vol. 59, no. 1, pp. 2019.
119–125, Jan. 2021. [50] F. Wen, H. Wymeersch, B. Peng, W. P. Tay, H. C. So, and D. Yang,
[30] T. A. Tsiftsis, C. Valagiannopoulos, H. Liu, A.-A. A. Boulogeorgos, “A Survey on 5G Massive MIMO Localization,” Digit. Signal Process,
and N. I. Miridakis, “Metasurface-Coated Devices: A New Paradigm vol. 94, no. C, p. 21–28, Nov. 2019.
for Energy-Efficient and Secure 6G Communications,” IEEE Veh. [51] D. Burghal, A. T. Ravi, V. Rao, A. A. Alghafis, and A. F. Molisch,
Technol. Mag., vol. 17, no. 1, pp. 27–36, Nov. 2022. “A comprehensive survey of machine learning based localization with
[31] S. E. Trevlakis, A.-A. A. Boulogeorgos, N. D. Chatzidiamantis, and wireless signals,” arXiv preprint arXiv:2012.11171, Dec. 2020.
G. K. Karagiannidis, “All-Optical Cochlear Implants,” IEEE Trans.
[52] X. Zhu, W. Qu, T. Qiu, L. Zhao, M. Atiquzzaman, and D. O. Wu, “In-
Mol. Biol. Multi-Scale Commun., vol. 6, no. 1, pp. 13–24, Jul. 2020.
door Intelligent Fingerprint-Based Localization: Principles, Approaches
[32] A.-A. A. Boulogeorgos and A. Alexiou, “Coverage Analysis of Recon- and Challenges,” IEEE Commun. Surv., vol. 22, no. 4, pp. 2634–2657,
figurable Intelligent Surface Assisted THz Wireless Systems,” IEEE Aug. 2020.
Open Journal of Vehicular Technology, vol. 2, pp. 94–110, Jan. 2021.
[53] A. Motroni, A. Buffi, and P. Nepa, “A Survey on Indoor Vehicle
[33] A. Celik, I. Romdhane, G. Kaddoum, and A. M. Eltawil, “A Top-
Localization Through RFID Technology,” IEEE Access, vol. 9, pp.
Down Survey on Optical Wireless Communications for the Internet of
17 921–17 942, Jan. 2021.
Things,” IEEE Commun. Surv., pp. 1–1, Jun. 2022.
[54] C. De Lima, D. Belot, R. Berkvens, A. Bourdoux, D. Dardari,
[34] A. Chaaban, Z. Rezki, and M.-S. Alouini, “On the Capacity
M. Guillaud, M. Isomursu, E.-S. Lohan, Y. Miao, A. N. Barreto et al.,
of Intensity-Modulation Direct-Detection Gaussian Optical Wireless
“Convergent communication, sensing and localization in 6G systems:
Communication Channels: A Tutorial,” IEEE Commun. Surv., vol. 24,
An overview of technologies, opportunities and challenges,” IEEE
no. 1, pp. 455–491, Oct. 2022.
Access, vol. 9, pp. 26 902–26 925, Jan. 2021.
[35] S. Naser, L. Bariah, S. Muhaidat, P. C. Sofotasios, M. Al-Qutayri,
E. Damiani, and M. Debbah, “Toward Federated-Learning-Enabled [55] O. Kanhere and T. S. Rappaport, “Position Location for Futuristic
Visible Light Communication in 6G Systems,” IEEE Wirel. Commun., Cellular Communications: 5G and Beyond,” IEEE Commun. Mag.,
vol. 29, no. 1, pp. 48–56, Feb. 2022. vol. 59, no. 1, pp. 70–75, Jan. 2021.
[36] M. M. Azari, S. Solanki, S. Chatzinotas, and M. Bennis, “THz- [56] Z. Xiao and Y. Zeng, “An overview on integrated localization and
Empowered UAVs in 6G: Opportunities, Challenges, and Trade-offs,” communication towards 6G,” Science China Information Sciences,
IEEE Commun. Mag., vol. 60, no. 5, pp. 24–30, May 2022. vol. 65, no. 3, pp. 1–46, Dec. 2021.
[37] D. Pliatsios, S. K. Goudos, T. Lagkas, V. Argyriou, A.-A. A. Bouloge- [57] H. Ding, B. Zhang, J. Zhou, Y. Yan, G. Tian, and B. Gu, “Recent de-
orgos, and P. Sarigiannidis, “Drone-Base-Station for Next-Generation velopments and applications of simultaneous localization and mapping
Internet-of-Things: A Comparison of Swarm Intelligence Approaches,” in agriculture,” J Field Robot, vol. 39, no. 6, pp. 956–983, May 2022.
IEEE Open Journal of Antennas and Propagation, vol. 3, pp. 32–47, [58] J. Laconte, A. Kasmi, R. Aufrère, M. Vaidis, and R. Chapuis, “A
Dec. 2022. Survey of Localization Methods for Autonomous Vehicles in Highway
[38] A.-A. A. Boulogeorgos, N. D. Chatzidiamantis, H. G. Sandalidis, Scenarios,” Sensors, vol. 22, no. 1, Dec. 2022.
A. Alexiou, and M. D. Renzo, “Cascaded Composite Turbulence and [59] A. Liu, Z. Huang, M. Li, Y. Wan, W. Li, T. X. Han, C. Liu, R. Du,
Misalignment: Statistical Characterization and Applications to Re- D. K. P. Tan, J. Lu, Y. Shen, F. Colone, and K. Chetty, “A Survey on
configurable Intelligent Surface-Empowered Wireless Systems,” IEEE Fundamental Limits of Integrated Sensing and Communication,” IEEE
Trans. Veh. Technol., vol. 71, no. 4, pp. 3821–3836, Jan. 2022. Commun. Surv., vol. 24, no. 2, pp. 994–1034, Feb. 2022.
[39] A.-A. A. Boulogeorgos, A. Alexiou, and M. D. Renzo, “Outage [60] H. Chen, H. Sarieddeen, T. Ballal, H. Wymeersch, M.-S. Alouini, and
Performance Analysis of RIS-Assisted UAV Wireless Systems Under T. Y. Al-Naffouri, “A tutorial on terahertz-band localization for 6G
Disorientation and Misalignment,” IEEE Trans. Veh. Technol., vol. 71, communication systems,” IEEE Commun. Surv., May 2022.
no. 10, pp. 10 712–10 728, Jun. 2022. [61] C. T. Nguyen, Y. M. Saputra, N. V. Huynh, N.-T. Nguyen, T. V. Khoa,
[40] A.-A. A. Boulogeorgos, J. M. Riera, and A. Alexiou, “On the Joint B. M. Tuan, D. N. Nguyen, D. T. Hoang, T. X. Vu, E. Dutkiewicz,
Effect of Rain and Beam Misalignment in Terahertz Wireless Systems,” S. Chatzinotas, and B. Ottersten, “A Comprehensive Survey of En-
IEEE Access, vol. 10, pp. 58 997–59 012, Jun. 2022. abling and Emerging Technologies for Social Distancing—Part I:
[41] G. Bresson, Z. Alsayed, L. Yu, and S. Glaser, “Simultaneous Loca- Fundamentals and Enabling Technologies,” IEEE Access, vol. 8, pp.
lization and Mapping: A Survey of Current Trends in Autonomous 153 479–153 507, Aug. 2020.
Driving,” IEEE Trans. Intell. Veh., vol. 2, no. 3, pp. 194–220, Sep. [62] C. T. Nguyen, Y. M. Saputra, N. Van Huynh, N.-T. Nguyen, T. V. Khoa,
2017. B. M. Tuan, D. N. Nguyen, D. T. Hoang, T. X. Vu, E. Dutkiewicz,
[42] A. F. G. Goncalves Ferreira, D. M. A. Fernandes, A. P. Catarino, and S. Chatzinotas, and B. Ottersten, “A Comprehensive Survey of En-
J. L. Monteiro, “Localization and Positioning Systems for Emergency abling and Emerging Technologies for Social Distancing—Part II:
Responders: A Survey,” IEEE Commun. Surv., vol. 19, no. 4, pp. 2836– Emerging Technologies and Open Issues,” IEEE Access, vol. 8, pp.
2870, May 2017. 154 209–154 236, Aug. 2020.
[43] S. Kuutti, S. Fallah, K. Katsaros, M. Dianati, F. Mccullough, and [63] F. Lemic, S. Abadal, W. Tavernier, P. Stroobant, D. Colle, E. Alarcón,
A. Mouzakitis, “A Survey of the State-of-the-Art Localization Tech- J. Marquez-Barja, and J. Famaey, “Survey on terahertz nanocommuni-
niques and Their Potentials for Autonomous Vehicle Applications,” cation and networking: A top-down perspective,” IEEE J. Sel. Areas
IEEE Internet Things J., vol. 5, no. 2, pp. 829–846, Mar. 2018. Commun., vol. 39, no. 6, pp. 1506–1543, Apr. 2021.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 49

[64] M. Nemati and J. Choi, “Intelligent-Meta-Surfaces-Aided Wireless harvesting,” in Proceedings of the 9th ACM International Conference
Communications in 6G,” in 2022 IEEE 95th Vehicular Technology on Nanoscale Computing and Communication, Oct. 2022, pp. 1–6.
Conference: (VTC2022-Spring), Aug. 2022, pp. 1–5. [86] J. T. Gómez, A. Kuestner, J. Simonjan, B. D. Unluturk, and F. Dressler,
[65] H. Viswanathan and P. E. Mogensen, “Communications in the 6G Era,” “Nanosensor Location Estimation in the Human Circulatory System
IEEE Access, vol. 8, pp. 57 063–57 074, Mar. 2020. Using Machine Learning,” IEEE Trans. Nanotechnol., vol. 21, pp. 663–
[66] E. Baccarelli, M. Scarpiniti, and A. Momenzadeh, “Ecomobifog–design 673, Oct. 2022.
and dynamic optimization of a 5G mobile-fog-cloud multi-tier ecosys- [87] J. T. Gómez, J. L. G. Rios, and F. Dressler, “Nanosensor location in
tem for the real-time distributed execution of stream applications,” the human circulatory system based on electric circuit representation
IEEE Access, vol. 7, pp. 55 565–55 608, Apr. 2019. of vessels,” in Proceedings of the 9th ACM International Conference
[67] European Commission, “Destination Earth – new digital twin of the on Nanoscale Computing and Communication, Oct. 2022, pp. 1–7.
Earth will help tackle climate change and protect nature,” Mar. 30, [88] A. Elango, A. Al-Tahmeesschi, M. Saukkoriipi, T. Malmivirta, and
2022 [Online]. [Online]. Available: https://ec.europa.eu/commission/ L. Ruotsalainen, “WHITE PAPER: Protecting GNSS Against Inten-
presscorner/detail/en/IP 22 1977 tional Interference,” arXiv preprint arXiv:2208.11555, Aug. 2022.
[68] P. Hofmann, R. Bassoli, F. H. Fitzek, and M. Reisslein, “MC NFV: [89] N. Xie, Y. Chen, Z. Li, and D. O. Wu, “Lightweight Secure Localiza-
Molecular Communication NFV in 6G Networks,” in 2022 IEEE 21st tion Approach in Wireless Sensor Networks,” IEEE Trans. Commun.,
Mediterranean Electrotechnical Conference (MELECON). IEEE, Aug. vol. 69, no. 10, pp. 6879–6893, Jul. 2021.
2022, pp. 1205–1210. [90] X. Liu, S. Su, F. Han, Y. Liu, and Z. Pan, “A Range-Based Secure
[69] E. Torres-Martinez, G. Paules, M. Schoeberg, and M. W. Kalb, “A web Localization Algorithm for Wireless Sensor Networks,” IEEE Sensors
of sensors: Enabling the earth science vision,” Acta Astronaut., vol. 53, J., vol. 19, no. 2, pp. 785–796, Oct. 2019.
no. 4-10, pp. 423–428, Nov. 2003. [91] A. Bochem and H. Zhang, “Robustness Enhanced Sensor Assisted
[70] P. K. D. Pramanik, A. Nayyar, and G. Pareek, “WBAN: Driving e- Monte Carlo Localization for Wireless Sensor Networks and the
healthcare beyond telemedicine to remote health monitoring: Archi- Internet of Things,” IEEE Access, vol. 10, pp. 33 408–33 420, Mar.
tecture and protocols,” in Telemedicine technologies. Elsevier, May 2022.
2019, pp. 89–119. [92] E.-S. Lohan, P. Richter, V. L. Sabola, J. A. Lopez-Salcedo, G. Seco-
[71] K. Wei, L. Zhang, Y. Guo, and X. Jiang, “Health monitoring based Granados, H. Leppäkoski, and E. S. Santiago, “Location privacy
on internet of medical things: architecture, enabling technologies, and challenges and solutions: Part 2: Hybrid-and non-GNSS localization,”
applications,” IEEE Access, vol. 8, pp. 27 468–27 478, Feb. 2020. Inside GNSS, vol. 12, no. 6, pp. 56–64, Nov. 2017.
[72] A. K. Dey, G. D. Abowd, and D. Salber, “A conceptual framework [93] H. Sarieddeen, N. Saeed, T. Y. Al-Naffouri, and M.-S. Alouini, “Next
and a toolkit for supporting the rapid prototyping of context-aware Generation Terahertz Communications: A Rendezvous of Sensing,
applications,” Human–Computer Interaction, vol. 16, no. 2-4, pp. 97– Imaging, and Localization,” IEEE Commun. Mag., vol. 58, no. 5, pp.
166, Dec. 2001. 69–75, May 2020.
[73] P. Dourish, “Seeking a foundation for context-aware computing,” [94] C. B. Barneto, L. Anttila, M. Fleischer, and M. Valkama, “OFDM radar
Human–Computer Interaction, vol. 16, no. 2-4, pp. 229–241, Nov. with LTE waveform: Processing and performance,” in 2019 IEEE Radio
2001. and Wireless Symposium (RWS). IEEE, May 2019, pp. 1–4.
[95] C. B. Barneto, T. Riihonen, M. Turunen, L. Anttila, M. Fleischer,
[74] D. Nüst, F. Bache, A. Bröring, C. Stasch, and S. Jirka, “Visualizing the
K. Stadius, J. Ryynänen, and M. Valkama, “Full-duplex OFDM radar
availability of temporally structured sensor data,” AGILE International
with LTE and 5G NR waveforms: Challenges, solutions, and measure-
Conference on Geographic Information Science 2010, Sep. 2019.
ments,” IEEE Trans. Microwave Theory Tech., vol. 67, no. 10, pp.
[75] A. Behravan, V. Yajnanarayana, M. F. Keskin, H. Chen, D. Shrestha,
4042–4054, Aug. 2019.
T. E. Abrudan, T. Svensson, K. Schindhelm, A. Wolfgang, S. Lindberg,
[96] S. Balakrishnan, S. Gupta, A. Bhuyan, P. Wang, D. Koutsonikolas,
and H. Wymeersch, “Positioning and Sensing in 6G: Gaps, Challenges,
and Z. Sun, “Physical layer identification based on spatial–temporal
and Opportunities,” IEEE Veh. Technol. Mag., Dec. 2022.
beam features for millimeter-wave wireless networks,” IEEE Trans.
[76] D. K. P. Tan, J. He, Y. Li, A. Bayesteh, Y. Chen, P. Zhu, and W. Tong, Inf. Forensics Secur., vol. 15, pp. 1831–1845, Oct. 2019.
“Integrated sensing and communication in 6G: Motivations, use cases, [97] N. Soltanieh, Y. Norouzi, Y. Yang, and N. C. Karmakar, “A review
requirements, challenges and future directions,” in 2021 1st IEEE of radio frequency fingerprinting techniques,” IEEE Journal of Radio
International Online Symposium on Joint Communications & Sensing Frequency Identification, vol. 4, no. 3, pp. 222–233, Jan. 2020.
(JC& S). IEEE, Mar. 2021, pp. 1–6. [98] S. Yuan, B. Han, D. Krummacker, and H. D. Schotten, “Massive
[77] A. T. Z. Kasgari, W. Saad, and M. Debbah, “Human-in-the-loop Twinning to Enhance Emergent Intelligence,” 2022 IEEE Symposium
wireless communications: Machine learning and brain-aware resource on Computers and Communications (ISCC), Oct. 2022.
management,” IEEE Trans. Commun., vol. 67, no. 11, pp. 7727–7743, [99] M. A. Uusitalo, P. Rugeland, M. R. Boldi, E. C. Strinati, P. Demes-
Jul. 2019. tichas, M. Ericson, G. P. Fettweis, M. C. Filippou, A. Gati, M.-H.
[78] J. Park, S. Samarakoon, M. Bennis, and M. Debbah, “Wireless network Hamon et al., “6G vision, value, use cases and technologies from
intelligence at the edge,” Proc. IEEE, vol. 107, no. 11, pp. 2204–2239, european 6G flagship project Hexa-X,” IEEE Access, vol. 9, pp.
Oct. 2019. 160 004–160 020, Nov. 2021.
[79] A. Mostaani, T. X. Vu, S. K. Sharma, Q. Liao, and S. Chatzino- [100] E. Skvortsov, E. Skvortsova, I. Sandu, and G. Iovlev, “Transi-
tas, “Task-oriented communication system design in cyber-physical tion of agriculture to digital, intellectual and robotics technologies,”
systems: A survey on theory and applications,” arXiv preprint Ekonomika Regiona, no. 3, p. 1014, Sep. 2018.
arXiv:2102.07166, Nov. 2021. [101] J. J. Roldán, J. del Cerro, D. Garzón-Ramos, P. Garcia-Aunon,
[80] K. A. Delin and S. P. Jackson, “Sensor web: a new instrument concept,” M. Garzón, J. De León, and A. Barrientos, “Robots in agriculture:
in Functional Integration of Opto-Electro-Mechanical Devices and State of art and practical experiences,” Service Robots, pp. 67–90, Oct.
Systems, vol. 4284. SPIE, Jan. 2001, pp. 1–9. 2018.
[81] C. Xiao and N. Chen, “Sensor Web: An Infrastructure for Real-Time [102] J. Guo, X. Li, Z. Li, L. Hu, G. Yang, C. Zhao, D. Fairbairn, D. Watson,
Smart Applications,” Int J Sens Netw Data Commun, vol. 6, p. e114, and M. Ge, “Multi-GNSS precise point positioning for precision
Sep. 2017. agriculture,” Precision Agriculture, vol. 19, no. 5, pp. 895–911, Oct.
[82] X. Zhang, N. Chen, Z. Chen, L. Wu, X. Li, L. Zhang, L. Di, J. Gong, 2018.
and D. Li, “Geospatial sensor web: A cyber-physical infrastructure for [103] L. C. Santos, A. S. Aguiar, F. N. Santos, A. Valente, J. B. Ventura,
geoscience research and application,” Earth-science reviews, vol. 185, and A. J. Sousa, “Navigation stack for robots working in steep
pp. 684–703, Oct. 2018. slope vineyard,” in Proceedings of SAI Intelligent Systems Conference.
[83] F. Dressler and S. Fischer, “Connecting in-body nano communication Springer, Aug. 2020, pp. 264–285.
with body area networks: Challenges and opportunities of the Internet [104] C. Cadena, L. Carlone, H. Carrillo, Y. Latif, D. Scaramuzza, J. Neira,
of Nano Things,” Nano Communication Networks, vol. 6, no. 2, pp. I. Reid, and J. J. Leonard, “Past, present, and future of simultaneous
29–38, Jun. 2015. localization and mapping: Toward the robust-perception age,” IEEE
[84] G. Piro, G. Boggia, and L. A. Grieco, “On the design of an energy- Trans. Rob., vol. 32, no. 6, pp. 1309–1332, Dec. 2016.
harvesting protocol stack for Body Area Nano-NETworks,” Nano [105] H. Zhan, C. S. Weerasekera, J.-W. Bian, and I. Reid, “Visual odom-
Commun. Networks, vol. 6, no. 2, pp. 74–84, Jun. 2015. etry revisited: What should be learnt?” in 2020 IEEE International
[85] F. Lemic, S. Abadal, A. Stevanovic, E. Alarcón, and J. Famaey, Conference on Robotics and Automation (ICRA). IEEE, Sep. 2020,
“Toward location-aware in-body terahertz nanonetworks with energy pp. 4203–4210.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 50

[106] U. D. Ulusar, G. Celik, and F. Al-Turjman, “Cognitive RF-based loca- [127] A.-A. A. Boulogeorgos, A. Alexiou, D. Kritharidis, A. Katsiotis,
lization for mission-critical applications in smart cities: An overview,” G. Ntouni, J. Kokkoniemi, J. Lethtomaki, M. Juntti, D. Yankova,
Comput Electr Eng, vol. 87, p. 106780, Jul. 2020. A. Mokhtar et al., “Wireless terahertz system architectures for networks
[107] A. A. Brincat, F. Pacifici, S. Martinaglia, and F. Mazzola, “The internet beyond 5G,” arXiv preprint arXiv:1810.12260, Oct. 2018.
of things for intelligent transportation systems in real smart cities [128] A.-A. A. Boulogeorgos, A. Alexiou, T. Merkle, C. Schubert,
scenarios,” in 2019 IEEE 5th World Forum on Internet of Things (WF- R. Elschner, A. Katsiotis, P. Stavrianos, D. Kritharidis, P.-K. Chartsias,
IoT). IEEE, Jul. 2019, pp. 128–132. J. Kokkoniemi, M. Juntti, J. Lehtomaki, A. Teixeira, and F. Rodrigues,
[108] P. Gope, R. Amin, S. H. Islam, N. Kumar, and V. K. Bhalla, “Terahertz Technologies to Deliver Optical Network Quality of Experi-
“Lightweight and privacy-preserving RFID authentication scheme for ence in Wireless Systems Beyond 5G,” IEEE Commun. Mag., vol. 56,
distributed IoT infrastructure with secure localization services for smart no. 6, pp. 144–151, Jun. 2018.
city environment,” Future Gener Comp Sy, vol. 83, pp. 629–637, Mar. [129] N. Kouzayha, M. A. Kishk, H. Sarieddeen, T. Y. Al-Naffouri, and M.-
2018. S. Alouini, “Coverage and rate analysis in coexisting terahertz and RF
[109] C. Chaccour, M. N. Soorki, W. Saad, M. Bennis, P. Popovski, and finite wireless networks,” arXiv preprint arXiv:2109.00646, Sep. 2021.
M. Debbah, “Seven defining features of terahertz (THz) wireless [130] S. Tarboush, H. Sarieddeen, M.-S. Alouini, and T. Y. Al-Naffouri,
systems: A fellowship of communication and sensing,” IEEE Commun. “Single-versus multi-carrier terahertz-band communications: A com-
Surv., vol. 24, no. 2, pp. 967–993, Jan. 2022. parative study,” IEEE Open Journal of the Communications Society,
[110] M. Giordani, M. Polese, M. Mezzavilla, S. Rangan, and M. Zorzi, Aug. 2022.
“Toward 6G networks: Use cases and technologies,” IEEE Commun. [131] S. Nie and I. F. Akyildiz, “Three-dimensional dynamic channel mod-
Mag., vol. 58, no. 3, pp. 55–61, Mar. 2020. eling and tracking for terahertz band indoor communications,” in 2017
[111] F. Mogyorósi, P. Revisnyei, A. Pašić, Z. Papp, I. Törös, P. Varga, and IEEE 28th Annual International Symposium on Personal, Indoor, and
A. Pašić, “Positioning in 5G and 6G networks—A Survey,” Sensors, Mobile Radio Communications (PIMRC). IEEE, Feb. 2017, pp. 1–5.
vol. 22, no. 13, p. 4757, Jun. 2022. [132] B. Peng and T. Kürner, “Three-dimensional angle of arrival estimation
[112] Y. Lu, P. Richter, and E. S. Lohan, “Opportunities and Challenges in in dynamic indoor terahertz channels using a forward–backward algo-
the Industrial Internet of Things based on 5G Positioning,” in 2018 rithm,” IEEE Trans. Veh. Technol., vol. 66, no. 5, pp. 3798–3811, Aug.
8th International Conference on Localization and GNSS (ICL-GNSS). 2016.
IEEE, Aug. 2018, pp. 1–6. [133] A. Guerra, F. Guidi, D. Dardari, and P. M. Djurić, “Near-field tracking
[113] Q.-b. Zhang, P. Wang, and Z.-h. Chen, “An improved particle filter with large antenna arrays: Fundamental limits and practical algo-
for mobile robot localization based on particle swarm optimization,” rithms,” IEEE Trans. Signal Process., vol. 69, pp. 5723–5738, Aug.
Expert Syst. Appl., vol. 135, pp. 181–193, Nov. 2019. 2021.
[114] J. Biswas and M. Veloso, “The 1,000-km challenge: Insights and [134] J. Tan and L. Dai, “Wideband beam tracking in THz massive MIMO
quantitative and qualitative results,” IEEE Intell. Syst., vol. 31, no. 3, systems,” IEEE J. Sel. Areas Commun., vol. 39, no. 6, pp. 1693–1710,
pp. 86–96, May 2016. Apr. 2021.
[115] H. Durrant-Whyte and T. Bailey, “Simultaneous localization and map- [135] S. Fan, Y. Wu, C. Han, and X. Wang, “A structured bidirectional LSTM
ping: part I,” IEEE Rob. Autom Mag., vol. 13, no. 2, pp. 99–110, Jun. deep learning method for 3D terahertz indoor localization,” in IEEE
2006. INFOCOM 2020-IEEE Conference on Computer Communications.
[116] T. Chong, X. Tang, C. Leng, M. Yogeswaran, O. Ng, and Y. Chong, IEEE, Aug. 2020, pp. 2381–2390.
“Sensor Technologies and Simultaneous Localization and Mapping [136] P. Hillger, M. van Delden, U. S. M. Thanthrige, A. M. Ahmed, J. Wit-
(SLAM),” Procedia Computer Science, vol. 76, pp. 174–179, Dec. temeier, K. Arzi, M. Andree, B. Sievert, W. Prost, A. Rennings et al.,
2015, 2015 IEEE International Symposium on Robotics and Intelligent “Toward mobile integrated electronic systems at THz frequencies,” J.
Sensors (IEEE IRIS2015). Infrared Millim. Terahertz Waves, vol. 41, no. 7, pp. 846–869, Jun.
[117] J. Fuentes-Pacheco, J. Ruiz-Ascencio, and J. M. Rendón-Mancha, 2020.
“Visual simultaneous localization and mapping: a survey,” Artificial [137] A. Nikpaik, A. H. M. Shirazi, A. Nabavi, S. Mirabbasi, and S. Shekhar,
intelligence review, vol. 43, no. 1, pp. 55–81, Jan. 2015. “A 219-to-231 GHz frequency-multiplier-based VCO with˜ 3% peak
[118] A. R. Khairuddin, M. S. Talib, and H. Haron, “Review on simulta- DC-to-RF efficiency in 65-nm CMOS,” IEEE J. Solid-State Circuits,
neous localization and mapping (SLAM),” in 2015 IEEE International vol. 53, no. 2, pp. 389–403, Oct. 2017.
Conference on Control System, Computing and Engineering (ICCSCE), [138] T. Nagatsuma, G. Ducournau, and C. C. Renaud, “Advances in terahertz
Jun. 2015, pp. 85–90. communications accelerated by photonics,” Nat. Photonics, vol. 10,
[119] S. Erfani, A. Jafari, and A. Hajiahmad, “Comparison of two data fusion no. 6, pp. 371–379, Mar. 2016.
methods for localization of wheeled mobile robot in farm conditions,” [139] K. Sengupta, T. Nagatsuma, and D. M. Mittleman, “Terahertz integrated
Artificial Intelligence in Agriculture, vol. 1, pp. 48–55, May 2019. electronic and hybrid electronic–photonic systems,” Nature Electronics,
[120] R. Smith, M. Self, and P. Cheeseman, “Estimating Uncertain Spatial vol. 1, no. 12, pp. 622–635, Dec. 2018.
Relationships in Robotics,” in Uncertainty in Artificial Intelligence, ser. [140] H. A. Hafez, S. Kovalev, J.-C. Deinert, Z. Mics, B. Green, N. Awari,
Machine Intelligence and Pattern Recognition, J. F. Lemmer and L. N. M. Chen, S. Germanskiy, U. Lehnert, J. Teichert et al., “Extremely
Kanal, Eds. North-Holland, Jul. 1988, vol. 5, pp. 435–461. efficient terahertz high-harmonic generation in graphene by hot Dirac
[121] F. Demim, A. Nemra, and K. Louadj, “Robust SVSF-SLAM for fermions,” Nature, vol. 561, no. 7724, pp. 507–511, Sep. 2018.
unmanned vehicle in unknown environment,” IFAC-PapersOnLine, [141] A. Singh, M. Andrello, N. Thawdar, and J. M. Jornet, “Design
vol. 49, no. 21, pp. 386–394, Nov. 2016. and operation of a graphene-based plasmonic nano-antenna array for
[122] S. Gupte, P. I. T. Mohandas, and J. M. Conrad, “A survey of communication in the terahertz band,” IEEE J. Sel. Areas Commun.,
quadrotor Unmanned Aerial Vehicles,” in 2012 Proceedings of IEEE vol. 38, no. 9, pp. 2104–2117, Jun. 2020.
Southeastcon, May 2012, pp. 1–6. [142] L. M. Zakrajsek, D. A. Pados, and J. M. Jornet, “Design and perfor-
[123] R. M. Faragher, C. Sarno, and M. Newman, “Opportunistic radio mance analysis of ultra-massive multi-carrier multiple input multiple
SLAM for indoor navigation using smartphone sensors,” in Pro- output communications in the terahertz band,” in Image Sensing
ceedings of the 2012 IEEE/ION Position, Location and Navigation Technologies: Materials, Devices, Systems, and Applications IV, vol.
Symposium, Jul. 2012, pp. 120–128. 10209. SPIE, Apr. 2017, pp. 26–36.
[124] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, [143] I. Ahmed, H. Khammari, A. Shahid, A. Musa, K. S. Kim, E. De Poorter,
A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and A. Fitzgibbon, and I. Moerman, “A survey on hybrid beamforming techniques in 5G:
“KinectFusion: Real-time dense surface mapping and tracking,” in 2011 Architecture and system model perspectives,” IEEE Commun. Surv.,
10th IEEE International Symposium on Mixed and Augmented Reality, vol. 20, no. 4, pp. 3060–3097, Jun. 2018.
Mar. 2011, pp. 127–136. [144] Y. Miao, E. Tanghe, J.-I. Takada, T. Pedersen, P. Laly, D. P. Gaillot,
[125] G. Stratidakis, A.-A. A. Boulogeorgos, and A. Alexiou, “A cooperative M. Liénard, L. Martens, and W. Joseph, “Measurement-based feasi-
localization-aided tracking algorithm for THz wireless systems,” in bility exploration on detecting and localizing multiple humans using
2019 IEEE Wireless Communications and Networking Conference MIMO radio channel properties,” IEEE Access, vol. 8, pp. 3738–3750,
(WCNC). IEEE, Oct. 2019, pp. 1–7. Dec. 2019.
[126] J. Kokkoniemi, J. Lehtomäki, and M. Juntti, “Simplified molecular [145] W. Huang, S. Zhu, S. Wang, J. Xie, and Y. Zhang, “Sparse repre-
absorption loss model for 275–400 gigahertz frequency band,” in 12th sentation for device-free human detection and localization with COTS
European Conference on Antennas and Propagation (EuCAP 2018), RFID,” in International Conference on Algorithms and Architectures
Dec. 2018, pp. 1–5. for Parallel Processing. Springer, Jan. 2019, pp. 639–654.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 51

[146] S. Denis, R. Berkvens, and M. Weyn, “A survey on detection, tracking [166] B. P. Ginsburg, K. Subburaj, S. Samala, K. Ramasubramanian, J. Singh,
and identification in radio frequency-based device-free localization,” S. Bhatara, S. Murali, D. Breen, M. Moallem, K. Dandu, S. Jalan,
Sensors, vol. 19, no. 23, p. 5329, Dec. 2019. N. Nayak, R. Sachdev, I. Prathapan, K. Bhatia, T. Davis, E. Seok,
[147] F. Hong, X. Wang, Y. Yang, Y. Zong, Y. Zhang, and Z. Guo, H. Parthasarathy, R. Chatterjee, V. Srinivasan, V. Giannini, A. Kumar,
“WFID: Passive device-free human identification using WiFi signal,” R. Kulak, S. Ram, P. Gupta, Z. Parkar, S. Bhardwaj, Y. C. Rakesh,
in Proceedings of the 13th International Conference on Mobile and K. A. Rajagopal, A. Shrimali, and V. Rentala, “A multimode 76-to-
Ubiquitous Systems: Computing, Networking and Services, Nov. 2016, 81GHz automotive radar transceiver with autonomous monitoring,” in
pp. 47–56. 2018 IEEE International Solid - State Circuits Conference - (ISSCC),
[148] F. Wen, J. Kulmer, K. Witrisal, and H. Wymeersch, “5G Positioning Feb. 2018, pp. 158–160.
and Mapping With Diffuse Multipath,” IEEE Trans. Wireless Commun., [167] R. Aldera, D. De Martini, M. Gadd, and P. Newman, “What could
vol. 20, no. 2, pp. 1164–1174, Oct. 2020. go wrong? introspective radar odometry in challenging environments,”
[149] Y. Ge, F. Wen, H. Kim, M. Zhu, F. Jiang, S. Kim, L. Svensson, in 2019 IEEE Intelligent Transportation Systems Conference (ITSC).
and H. Wymeersch, “5G SLAM using the clustering and assignment IEEE, Nov. 2019, pp. 2835–2842.
approach with diffuse multipath,” Sensors, vol. 20, no. 16, p. 4656, [168] Y. S. Park, Y.-S. Shin, and A. Kim, “Pharao: Direct radar odometry
Aug. 2020. using phase correlation,” in 2020 IEEE International Conference on
[150] H. Wymeersch, J. He, B. Denis, A. Clemente, and M. Juntti, “Ra- Robotics and Automation (ICRA). IEEE, Sep. 2020, pp. 2617–2623.
dio localization and mapping with reconfigurable intelligent surfaces: [169] A. N. Ramesh, C. M. León, J. C. Zafra, S. Brüggenwirth, and M. A.
Challenges, opportunities, and research directions,” IEEE Veh. Technol. González-Huici, “Landmark-based RADAR SLAM for Autonomous
Mag., vol. 15, no. 4, pp. 52–61, Oct. 2020. Driving,” in 2021 21st International Radar Symposium (IRS). IEEE,
[151] E. C. Strinati, G. C. Alexandropoulos, H. Wymeersch, B. De- Jul. 2021, pp. 1–10.
nis, V. Sciancalepore, R. D’Errico, A. Clemente, D.-T. Phan-Huy, [170] S. H. Cen and P. Newman, “Precise ego-motion estimation with
E. De Carvalho, and P. Popovski, “Reconfigurable, intelligent, and millimeter-wave radar under diverse and challenging conditions,” in
sustainable wireless environments for 6G smart connectivity,” IEEE 2018 IEEE International Conference on Robotics and Automation
Commun. Mag., vol. 59, no. 10, pp. 99–105, Oct. 2021. (ICRA). IEEE, Nov. 2018, pp. 6045–6052.
[152] A. Tishchenko, A. Ali, A. Araghi, P. Botham, B. Burton, and [171] ——, “Radar-only ego-motion estimation in difficult settings via graph
M. Khalily, “Autonomous Reconfigurable Intelligent Surface based matching,” in 2019 International Conference on Robotics and Automa-
on Highly-Efficient Solar Cells,” in 17th European Conference on tion (ICRA). IEEE, Aug. 2019, pp. 298–304.
Antennas and Propagation (EuCAP 2023), Mar. 2023. [172] Ş. Săftescu, M. Gadd, D. De Martini, D. Barnes, and P. Newman,
[153] N. Khalid and O. B. Akan, “Experimental throughput analysis of low- “Kidnapped radar: Topological radar localisation using rotationally-
THz MIMO communication channel in 5G wireless networks,” IEEE invariant metric learning,” in 2020 IEEE International Conference on
Wireless Commun. Lett., vol. 5, no. 6, pp. 616–619, Sep. 2016. Robotics and Automation (ICRA). IEEE, Sep. 2020, pp. 4358–4364.
[154] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis, [173] T. Y. Tang, D. De Martini, D. Barnes, and P. Newman, “Rsl-net:
and I. Akyildiz, “A new wireless communication paradigm through Localising in satellite images from a radar on the ground,” IEEE Robot.
software-controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9, Autom. Lett., vol. 5, no. 2, pp. 1087–1094, Jan. 2020.
pp. 162–169, Sep. 2018. [174] A. Kanev, A. Nasteka, C. Bessonova, D. Nevmerzhitsky, A. Silaev,
[155] X. Fu, F. Yang, C. Liu, X. Wu, and T. J. Cui, “Terahertz Beam A. Efremov, and K. Nikiforova, “Anomaly detection in wireless sensor
Steering Technologies: From Phased Arrays to Field-Programmable network of the “smart home” system,” in 2017 20th Conference of
Metasurfaces,” Advanced optical materials, vol. 8, no. 3, p. 1900628, Open Innovations Association (FRUCT), Oct. 2017, pp. 118–124.
Aug. 2020. [175] M. Liu, K. Yang, N. Zhao, Y. Chen, H. Song, and F. Gong, “Intel-
[156] S. Venkatesh, X. Lu, H. Saeidi, and K. Sengupta, “A high-speed ligent Signal Classification in Industrial Distributed Wireless Sensor
programmable and scalable terahertz holographic metasurface based Networks Based Industrial Internet of Things,” IEEE Trans. Ind. Inf.,
on tiled CMOS chips,” Nature electronics, vol. 3, no. 12, pp. 785–793, vol. 17, no. 7, pp. 4946–4956, Aug. 2021.
Dec. 2020. [176] P. Sommer, J. Liu, K. Zhao, B. Kusy, R. Jurdak, A. McKeown, and
[157] S. Nie, J. M. Jornet, and I. F. Akyildiz, “Intelligent environments D. Westcott, “Information Bang for the Energy Buck: Towards Energy-
based on ultra-massive MIMO platforms for wireless communication and Mobility-Aware Tracking.” in EWSN, Feb. 2016, pp. 193–204.
in millimeter wave and terahertz bands,” in ICASSP 2019-2019 IEEE [177] L. Tõnisson, J. Voigtländer, M. Weger, D. Assmann, R. Käthner,
International Conference on Acoustics, Speech and Signal Processing B. Heinold, and A. Macke, “Knowledge Transfer with Citizen Science:
(ICASSP). IEEE, Apr. 2019, pp. 7849–7853. Luft-Leipzig Case Study,” Sustainability, vol. 13, no. 14, Jul. 2021.
[158] S. Ali, W. Saad, N. Rajatheva, K. Chang, D. Steinbach, B. Sliwa, [178] S. Tomic, M. Beko, and R. Dinis, “3-D Target Localization in Wireless
C. Wietfeld, K. Mei, H. Shiri, H.-J. Zepernick et al., “6G white Sensor Networks Using RSS and AoA Measurements,” IEEE Trans.
paper on machine learning in wireless communication networks,” arXiv Veh. Technol., vol. 66, no. 4, pp. 3197–3210, Jul. 2017.
preprint arXiv:2004.13875, Apr. 2020. [179] X. Xing, G. Wang, and J. Li, “Collaborative Target Tracking in Wireless
[159] C.-H. Hsieh, J.-Y. Chen, and B.-H. Nien, “Deep learning-based indoor Sensor Networks.” Adhoc & Sensor Wireless Networks, vol. 23, Aug.
localization using received signal strength and channel state informa- 2014.
tion,” IEEE Access, vol. 7, pp. 33 256–33 267, Mar. 2019. [180] Y. Xiao, S. Zhang, J. Wang, and C. Zhu, “A Novel Indoor Localization
[160] B. S. Ciftler, A. Albaseer, N. Lasla, and M. Abdallah, “Federated Algorithm for Efficient Mobility Management in Wireless Networks,”
learning for RSS fingerprint-based localization: A privacy-preserving Wireless Communications and Mobile Computing, vol. 2018, Jun. 2018.
crowdsourcing method,” in 2020 International Wireless Communica- [181] Z. Liao, J. Wang, S. Zhang, J. Cao, and G. Min, “Minimizing
tions and Mobile Computing (IWCMC). IEEE, Jul. 2020, pp. 2112– Movement for Target Coverage and Network Connectivity in Mobile
2117. Sensor Networks,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 7,
[161] B. Peng, G. Seco-Granados, E. Steinmetz, M. Fröhle, and H. Wymeer- pp. 1971–1983, Jun. 2015.
sch, “Decentralized scheduling for cooperative localization with deep [182] A. Gunathillake, A. V. Savkin, and A. P. Jayasumana, “Topology
reinforcement learning,” IEEE Trans. Veh. Technol., vol. 68, no. 5, pp. mapping algorithm for 2D and 3D wireless sensor networks based on
4295–4305, Apr. 2019. maximum likelihood estimation,” Comput. Networks, vol. 130, pp. 1–
[162] C. Studer, S. Medjkouh, E. Gonultaş, T. Goldstein, and O. Tirkkonen, 15, Nov. 2018.
“Channel charting: Locating users within the radio environment using [183] F. Ebner, T. Fetzer, F. Deinzer, and M. Grzegorzek, “On prior
channel state information,” IEEE Access, vol. 6, pp. 47 682–47 698, navigation knowledge in multi sensor indoor localisation,” in 2016
Aug. 2018. 19th International Conference on Information Fusion (FUSION), Aug.
[163] D. Kocur, M. Svecova, and D. Novak, “UWB Radar Based Localization 2016, pp. 557–564.
of a Person Changing the Nature of Their Motion State,” in 2018 [184] ——, “On Wi-Fi model optimizations for smartphone-based indoor
International Conference on Radar (RADAR). IEEE, Aug. 2018. localization,” ISPRS International Journal of Geo-Information, vol. 6,
[164] J. Sachs, Handbook of ultra-wideband short-range sensing: theory, no. 8, p. 233, Aug. 2017.
sensors, applications. John Wiley & Sons, Dec. 2013. [185] A.-C. Anastou, K. K. Delibasis, A.-A. A. Boulogeorgos, H. G. San-
[165] A. Alkhateeb, S. Alex, P. Varkey, Y. Li, Q. Qu, and D. Tujkovic, “Deep dalidis, A. Vavoulas, and S. K. Tasoulis, “A Low Complexity Indoor
Learning Coordinated Beamforming for Highly-Mobile Millimeter Visible Light Positioning Method,” IEEE Access, vol. 9, pp. 57 658–
Wave Systems,” IEEE Access, vol. 6, pp. 37 328–37 348, Jun. 2018. 57 673, Apr. 2021.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 52

[186] J. Luo, L. Fan, and H. Li, “Indoor Positioning Systems Based on Visible [207] G. Ottoy and L. De Strycker, “An Improved 2D Triangulation Algo-
Light Communication: State of the Art,” IEEE Commun. Surv., vol. 19, rithm for Use With Linear Arrays,” IEEE Sensors J., vol. 16, no. 23,
no. 4, pp. 2871–2893, Aug. 2017. pp. 8238–8243, Jun. 2016.
[187] P. Chen, M. Pang, D. Che, Y. Yin, D. Hu, and S. Gao, “A survey [208] P. Kristalina, A. Pratiarso, T. Badriyah, and E. D. Putro, “A wireless
on visible light positioning from software algorithms to hardware,” sensor networks localization using geometric triangulation scheme
Wireless Communications and Mobile Computing, vol. 2021, Feb. for object tracking in urban search and rescue application,” in 2016
2021. 2nd International Conference on Science in Information Technology
[188] A. M. Abdelhady, A. K. S. Salem, O. Amin, B. Shihada, and M.- (ICSITech), Feb. 2016, pp. 254–259.
S. Alouini, “Visible light communications via intelligent reflecting [209] A. Landström and J. van de Beek, “Transmitter localization for 5G
surfaces: Metasurfaces vs mirror arrays,” IEEE Open J. Commun. Soc., mmWave REMs by stochastic generalized triangulation,” in 2016 23rd
vol. 2, pp. 1–20, Dec. 2020. International Conference on Telecommunications (ICT), Jun. 2016, pp.
[189] Y. Zhuang, L. Hua, L. Qi, J. Yang, P. Cao, Y. Cao, Y. Wu, J. Thompson, 1–5.
and H. Haas, “A survey of positioning systems using visible LED [210] Prateek and R. Arya, “T-LOC: RSSI-based, range-free, triangulation
lights,” IEEE Commun. Surv., vol. 20, no. 3, pp. 1963–1988, Feb. 2018. assisted localization for convex relaxation with limited node range
[190] M. Maheepala, A. Z. Kouzani, and M. A. Joordens, “Light-based indoor under uncertainty skew constraint,” Journal of Ambient Intelligence
positioning systems: A review,” IEEE Sensors J., vol. 20, no. 8, pp. and Humanized Computing, Oct. 2021.
3971–3995, Jan. 2020. [211] R. Tazawa, N. Honma, M. Atsusi, and M. Hiroto, “Improving accuracy
[191] A. Costanzo and V. Loscrı̀, “Error compensation in indoor position- of RSSI-based indoor localization using three-element array,” in 2016
ing systems based on software defined visible light communication,” International Symposium on Antennas and Propagation (ISAP), Oct.
Physical Communication, vol. 34, pp. 235–245, Apr. 2019. 2016, pp. 336–337.
[192] K. Sengupta, “Universal terahertz integrated systems: Bridging the [212] H. Steendam, “A 3-D Positioning Algorithm for AOA-Based VLP With
‘THz’and ‘application’gap in the next decade,” in 2019 IEEE MTT- an Aperture-Based Receiver,” IEEE J. Sel. Areas Commun., vol. 36,
S International Microwave and RF Conference (IMARC). IEEE, Dec. no. 1, pp. 23–33, Nov. 2018.
2019, pp. 1–3. [213] T.-H. Lee, S. Weng, and J. Sanford, “Indoor radio triangulation using
[193] K. Kenneth, W. Choi, Q. Zhong, N. Sharma, Y. Zhang, R. Han, only RSSI data,” in 2020 IEEE International Symposium on Antennas
Z. Ahmad, D.-Y. Kim, S. Kshattry, I. R. Medvedev et al., “Opening and Propagation and North American Radio Science Meeting, Feb.
terahertz for everyday applications,” IEEE Commun. Mag., vol. 57, 2020, pp. 1097–1098.
no. 8, pp. 70–76, Aug. 2019. [214] Q. Yan, J. Chen, G. Ottoy, B. Cox, and L. De Strycker, “An accurate
[194] R. W. Heath, N. Gonzalez-Prelcic, S. Rangan, W. Roh, and A. M. AOA localization method based on unreliable sensor detection,” in
Sayeed, “An overview of signal processing techniques for millimeter 2018 IEEE Sensors Applications Symposium (SAS), Mar. 2018, pp.
wave MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 10, 1–6.
no. 3, pp. 436–453, Feb. 2016.
[215] E. Soltanaghaei, A. Kalyanaraman, and K. Whitehouse, “Multipath
[195] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados, Triangulation: Decimeter-level WiFi Localization and Orientation with
and H. Wymeersch, “Position and orientation estimation through a Single Unaided Receiver,” in Proceedings of the 16th Annual Inter-
millimeter-wave MIMO in 5G systems,” IEEE Trans. Wireless Com- national Conference on Mobile Systems, Applications, and Services.
mun., vol. 17, no. 3, pp. 1822–1835, Dec. 2017. Munich Germany: ACM, Jun. 2018, pp. 376–388.
[196] J. Ma, R. Shrestha, W. Zhang, L. Moeller, and D. M. Mittleman,
[216] V. R. Adama and G. M. Asutkar, “Improved localization with RSSI and
“Terahertz wireless links using diffuse scattering from rough surfaces,”
DTN algorithms in wireless sensor networks,” International Journal of
IEEE Trans. THz Sci. Technol., vol. 9, no. 5, pp. 463–470, Aug. 2019.
Information Technology, vol. 13, no. 2, pp. 603–607, Jan. 2021.
[197] A. Fascista, A. Coluccia, H. Wymeersch, and G. Seco-Granados,
“Millimeter-wave downlink positioning with a single-antenna receiver,” [217] J. Werner, M. Costa, A. Hakkarainen, K. Leppanen, and M. Valkama,
IEEE Trans. Wireless Commun., vol. 18, no. 9, pp. 4479–4490, Jul. “Joint User Node Positioning and Clock Offset Estimation in 5G Ultra-
2019. Dense Networks,” in 2015 IEEE Global Communications Conference
(GLOBECOM), Dec. 2015, pp. 1–7.
[198] W. Tang, M. Z. Chen, X. Chen, J. Y. Dai, Y. Han, M. Di Renzo,
Y. Zeng, S. Jin, Q. Cheng, and T. J. Cui, “Wireless communications [218] S. Yousefi, X.-W. Chang, and B. Champagne, “Mobile Localization in
with reconfigurable intelligent surface: Path loss modeling and experi- Non-Line-of-Sight Using Constrained Square-Root Unscented Kalman
mental measurement,” IEEE Trans. Wireless Commun., vol. 20, no. 1, Filter,” IEEE Trans. Veh. Technol., vol. 64, no. 5, pp. 2071–2083, May
pp. 421–439, Sep. 2020. 2015.
[199] J. M. Kahn and J. R. Barry, “Wireless infrared communications,” Proc. [219] E. Rastorgueva–Foi, M. Koivisto, M. Valkama, M. Costa, and
IEEE, vol. 85, no. 2, pp. 265–298, Feb. 1997. K. Leppänen, “Localization and Tracking in mmWave Radio Networks
[200] L. Qian, X. Chi, L. Zhao, and A. Chaaban, “Secure visible light using Beam-Based DoD Measurements,” in 2018 8th International
communications via intelligent reflecting surfaces,” in ICC 2021-IEEE Conference on Localization and GNSS (ICL-GNSS), Jun. 2018, pp.
International Conference on Communications. IEEE, Aug. 2021, pp. 1–6.
1–6. [220] O. Kanhere and T. S. Rappaport, “Outdoor sub-THz Position Location
[201] T. Tang, T. Shang, and Q. Li, “Impact of multiple shadows on visible and Tracking using Field Measurements at 142 GHz,” in ICC 2021
light communication channel,” IEEE Commun. Lett., vol. 25, no. 2, pp. - IEEE International Conference on Communications, Aug. 2021, pp.
513–517, Oct. 2020. 1–6.
[202] W. Zhang, M. S. Chowdhury, and M. Kavehrad, “Asynchronous indoor [221] M. Ammous and S. Valaee, “Positioning and Tracking Using Recon-
positioning system based on visible light communications,” Opt. Eng., figurable Intelligent Surfaces and Extended Kalman Filter,” in 2022
vol. 53, no. 4, p. 045105, Apr. 2014. IEEE 95th Vehicular Technology Conference: (VTC2022-Spring), Aug.
[203] S.-Y. Jung, S. Hann, and C.-S. Park, “TDOA-based optical wireless 2022, pp. 1–6.
indoor localization using LED ceiling lamps,” IEEE Trans. Consum. [222] M. Koivisto, M. Costa, A. Hakkarainen, K. Leppanen, and M. Valkama,
Electron., vol. 57, no. 4, pp. 1592–1597, Nov. 2011. “Joint 3D Positioning and Network Synchronization in 5G Ultra-Dense
[204] I. Guvenc and C.-C. Chong, “A survey on TOA based wireless Networks Using UKF and EKF,” in 2016 IEEE Globecom Workshops
localization and NLOS mitigation techniques,” IEEE Commun. Surv., (GC Wkshps), Feb. 2016, pp. 1–7.
vol. 11, no. 3, pp. 107–124, Aug. 2009. [223] M. Koivisto, M. Costa, J. Werner, K. Heiska, J. Talvitie, K. Leppänen,
[205] R. Ramanathan, A. Samanta, and T. L. Porta, “Symptotics: A frame- V. Koivunen, and M. Valkama, “Joint Device Positioning and Clock
work for analyzing the scalability of real-world wireless networks,” in Synchronization in 5G Ultra-Dense Networks,” IEEE Trans. Wireless
Proceedings of the 9th ACM symposium on Performance evaluation Commun., vol. 16, no. 5, pp. 2866–2881, Mar. 2017.
of wireless ad hoc, sensor, and ubiquitous networks, Oct. 2012, pp. [224] X. Bai, Z. Zhang, L. Liu, X. Zhai, J. Panneerselvam, and L. Ge,
31–38. “Enhancing Localization of Mobile Robots in Distributed Sensor Envi-
[206] B. Hanssens, D. Plets, E. Tanghe, C. Oestges, D. P. Gaillot, M. Liénard, ronments for Reliable Proximity Service Applications,” IEEE Access,
L. Martens, and W. Joseph, “An indoor localization technique based on vol. 7, pp. 28 826–28 834, Feb. 2019.
ultra-wideband AoD/AoA/ToA estimation,” in 2016 IEEE International [225] H. Cao, Y. T. Chan, and H. C. So, “Maximum Likelihood TDOA Esti-
Symposium on Antennas and Propagation (APSURSI), Oct. 2016, pp. mation From Compressed Sensing Samples Without Reconstruction,”
1445–1446. IEEE Signal Process Lett., vol. 24, no. 5, pp. 564–568, Mar. 2017.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 53

[226] A. Abdelreheem, E. M. Mohamed, and H. Esmaiel, “Location-Based [249] F. Ma, Z.-M. Liu, and F. Guo, “Direct Position Determination in
Millimeter Wave Multi-Level Beamforming Using Compressive Sens- Asynchronous Sensor Networks,” IEEE Trans. Veh. Technol., vol. 68,
ing,” IEEE Commun. Lett., vol. 22, no. 1, pp. 185–188, Oct. 2018. no. 9, pp. 8790–8803, Jul. 2019.
[227] S. Salari, F. Chan, Y.-T. Chan, and W. Read, “TDOA Estimation With [250] H. Zhao, N. Zhang, and Y. Shen, “Beamspace Direct Localization for
Compressive Sensing Measurements and Hadamard Matrix,” IEEE Large-Scale Antenna Array Systems,” IEEE Trans. Signal Process.,
Trans. Aerosp. Electron. Syst., vol. 54, no. 6, pp. 3137–3142, Apr. vol. 68, pp. 3529–3544, May 2020.
2018. [251] N. Xia, B. Li, and J. Wang, “A Direct Localization Method in the
[228] A. Harlakin, J. Mietzner, P. A. Hoeher, and A. Meusling, “Compressive- Presence of Cochannel Interference and Multipath Fading,” IEEE Syst.
Sensing-Aided MIMO Radar Enabling Multi-Functional and Compact J., vol. 15, no. 2, pp. 2855–2858, Jul. 2021.
Sensors in Air Scenarios Using Optimized Antenna Arrays,” IEEE [252] Z. Yu, J. Li, Q. Guo, and J. Ding, “Efficient Direct Target Localization
Access, vol. 9, pp. 41 417–41 429, Mar. 2021. for Distributed MIMO Radar With Expectation Propagation and Belief
[229] X. Jiang, N. Li, Y. Guo, D. Yu, and S. Yang, “Localization of Multiple Propagation,” IEEE Trans. Signal Process., vol. 69, pp. 4055–4068,
RF Sources Based on Bayesian Compressive Sensing Using a Limited Jun. 2021.
Number of UAVs With Airborne RSS Sensor,” IEEE Sensors J., vol. 21, [253] M. F. Hossain, A. U. Mahin, T. Debnath, F. B. Mosharrof, and K. Z.
no. 5, pp. 7067–7079, Nov. 2021. Islam, “Recent research in cloud radio access network (C-RAN) for 5G
[230] X. Jiang, N. Li, Y. Guo, W. Xie, K. Xu, and M. Guo, “Localization cellular systems-A survey,” J Netw Comput Appl, vol. 139, pp. 31–48,
of Multiple RF Sources via Priori Knowledge-Aided Bayesian Com- May 2019.
pressive Sensing in UAV-Based WSN,” IEEE Commun. Lett., vol. 25, [254] J. Bai, G. Wei, S. Wang, X. Wang, and Z. Fei, “Efficient Direct
no. 12, pp. 3848–3852, Oct. 2021. Localization of OFDM Emitters in Multipath Environment With Mobile
[231] R. Ramadan, A. Jwaifel, H. Al-Tous, and I. Barhumi, “Compressive Receivers,” IEEE Trans. Veh. Technol., vol. 71, no. 1, pp. 545–556,
sensing with weighted coefficient approach for indoor source localiza- Nov. 2022.
tion,” in 2017 40th International Conference on Telecommunications [255] N. Nedjah, L. Macedo Mourelle, and P. J. Albuquerque de Oliveira,
and Signal Processing (TSP), Oct. 2017, pp. 243–246. “Simultaneous Localization and Mapping Using Swarm Intelligence
[232] K. Gligoric, M. Ajmani, D. Vukobratović, and S. Sinanović, “Visi- Based Methods,” Expert Syst. Appl., vol. 159, p. 113547, Nov. 2020.
ble Light Communications-Based Indoor Positioning via Compressed [256] P. T. Daely and S. Y. Shin, “Range based wireless node localization
Sensing,” IEEE Commun. Lett., vol. 22, no. 7, pp. 1410–1413, May using Dragonfly Algorithm,” in 2016 Eighth International Conference
2018. on Ubiquitous and Future Networks (ICUFN), Aug. 2016, pp. 1012–
[233] L. Oliveira, H. Li, L. Almeida, and T. E. Abrudan, “RSSI-based 1015.
Relative Localisation for Mobile Robots,” Ad Hoc Networks, vol. 13, [257] A. Arsic, M. Tuba, and M. Jordanski, “Fireworks algorithm applied to
pp. 321–335, Feb. 2014. wireless sensor networks localization problem,” in 2016 IEEE Congress
[234] C. Di Franco, A. Melani, and M. Marinoni, “Solving ambiguities on Evolutionary Computation (CEC), Nov. 2016, pp. 4038–4044.
in MDS relative localization,” in 2015 International Conference on [258] V. R. Kulkarni, V. Desai, and R. V. Kulkarni, “Multistage localization
Advanced Robotics (ICAR), Sep. 2015, pp. 230–236. in wireless sensor networks using artificial bee colony algorithm,” in
[235] N. Saeed and H. Nam, “Robust Multidimensional Scaling for Cognitive 2016 IEEE Symposium Series on Computational Intelligence (SSCI),
Radio Network Localization,” IEEE Trans. Veh. Technol., vol. 64, no. 9, Feb. 2016, pp. 1–8.
pp. 4056–4062, Oct. 2015. [259] S. Monica and G. Ferrari, “A Swarm-Based Approach to Real-Time
3D Indoor Localization: Experimental Performance Analysis,” Appl.
[236] V. K. Chaurasiya, N. Jain, and G. Nandi, “A Novel Distance Estimation
Soft Comput., vol. 43, pp. 489–497, Jun. 2016.
Approach for 3D Localization in Wireless Sensor Network Using Multi
[260] E. Tuba, M. Tuba, and M. Beko, “Two Stage Wireless Sensor Node
Dimensional Scaling,” Information Fusion, vol. 15, pp. 5–18, Jan. 2014.
Localization Using Firefly Algorithm,” in Smart Trends in Systems,
[237] W. Jiang, C. Xu, L. Pei, and W. Yu, “Multidimensional Scaling-Based
Security and Sustainability, X.-S. Yang, A. K. Nagar, and A. Joshi,
TDOA Localization Scheme Using an Auxiliary Line,” IEEE Signal
Eds. Singapore: Springer Singapore, Dec. 2018, vol. 18, pp. 113–
Process Lett., vol. 23, no. 4, pp. 546–550, Mar. 2016.
120.
[238] S. Kumar, R. Kumar, and K. Rajawat, “Cooperative Localization of
[261] I. Strumberger, E. Tuba, N. Bacanin, M. Beko, and M. Tuba, “Monarch
Mobile Networks Via Velocity-Assisted Multidimensional Scaling,”
butterfly optimization algorithm for localization in wireless sensor
IEEE Trans. Signal Process., vol. 64, no. 7, pp. 1744–1758, Dec. 2016.
networks,” in 2018 28th International Conference Radioelektronika
[239] G. Morral and P. Bianchi, “Distributed On-Line Multidimensional (RADIOELEKTRONIKA), Jun. 2018, pp. 1–6.
Scaling for Self-Localization in Wireless Sensor Networks,” Signal [262] ——, “Wireless Sensor Network Localization Problem by Hybridized
Process., vol. 120, pp. 88–98, Mar. 2016. Moth Search Algorithm,” in 2018 14th International Wireless Commu-
[240] B. R. Stojkoska, J. Palikrushev, K. Trivodaliev, and S. Kalajdziski, nications and Mobile Computing Conference (IWCMC), Aug. 2018,
“Indoor localization of unmanned aerial vehicles based on RSSI,” pp. 316–321.
in IEEE EUROCON 2017 -17th International Conference on Smart [263] S. Phoemphon, C. So-In, and D. T. Niyato, “A Hybrid Model Using
Technologies, Aug. 2017, pp. 120–125. Fuzzy Logic and an Extreme Learning Machine with Vector Particle
[241] Y. Fan, X. Qi, B. Li, and L. Liu, “Fast clustering-based multidimen- Swarm Optimization for Wireless Sensor Network Localization,” Appl.
sional scaling for mobile networks localisation,” IET Commun., vol. 14, Soft Comput., vol. 65, pp. 101–120, Apr. 2018.
no. 1, pp. 135–143, Jan. 2020. [264] V. R. Kulkarni and V. Desai, “Computational Intelligence for Lo-
[242] Y. Fan, X. Qi, and L. Liu, “Fault-Tolerant Cooperative Localization of calization of Mobile Wireless Sensor Networks,” in Computational
3D Mobile Networks via Two-Layer Filter Multidimensional Scaling,” Intelligence: Theories, Applications and Future Directions - Volume II,
IEEE Sensors J., vol. 21, no. 6, pp. 8354–8366, Dec. 2021. N. K. Verma and A. K. Ghosh, Eds. Singapore: Springer Singapore,
[243] W. Cui, C. Wu, W. Meng, B. Li, Y. Zhang, and L. Xie, “Dynamic Sep. 2019, vol. 799, pp. 449–463.
Multidimensional Scaling Algorithm for 3-D Mobile Localization,” [265] I. Strumberger, M. Minovic, M. Tuba, and N. Bacanin, “Performance of
IEEE Trans. Instrum. Meas., vol. 65, no. 12, pp. 2853–2865, Oct. 2016. Elephant Herding Optimization and Tree Growth Algorithm Adapted
[244] A. Weiss, “Direct position determination of narrowband radio fre- for Node Localization in Wireless Sensor Networks,” Sensors, vol. 19,
quency transmitters,” IEEE Signal Process Lett., vol. 11, no. 5, pp. no. 11, p. 2515, Jun. 2019.
513–516, Apr. 2004. [266] M. Y. Arafat and S. Moh, “Localization and Clustering Based on
[245] N. Garcia, H. Wymeersch, E. G. Larsson, A. M. Haimovich, and Swarm Intelligence in UAV Networks for Emergency Communica-
M. Coulon, “Direct Localization for Massive MIMO,” IEEE Trans. tions,” IEEE Internet Things J., vol. 6, no. 5, pp. 8958–8976, Jun.
Signal Process., vol. 65, no. 10, pp. 2475–2487, Feb. 2017. 2019.
[246] J. Yin, D. Wang, Y. Wu, and R. Liu, “Direct Localization of Multiple [267] J. Akram, A. Javed, S. Khan, A. Akram, H. S. Munawar, and
Stationary Narrowband Sources Based on Angle and Doppler,” IEEE W. Ahmad, “Swarm Intelligence Based Localization in Wireless Sensor
Commun. Lett., vol. 21, no. 12, pp. 2630–2633, Sep. 2017. Networks,” in Proceedings of the 36th Annual ACM Symposium on
[247] M. Chen, X. Mao, X. Long, and L. Xin, “Underdetermined Direct Applied Computing. Virtual Event Republic of Korea: ACM, Mar.
Localization of Emitters Based on Spatio-Temporal Processing,” IEEE 2021, pp. 1906–1914.
Signal Process Lett., vol. 25, no. 3, pp. 452–456, Jan. 2018. [268] P. Singh, A. Khosla, A. Kumar, and M. Khosla, “Computational
[248] K. Han, Y. Liu, Z. Deng, L. Yin, and L. Shi, “Direct Positioning Intelligence Based Localization of Moving Target Nodes Using Single
Method of Mixed Far-Field and Near-Field Based on 5G Massive Anchor Node in Wireless Sensor Networks,” Telecommun Syst, vol. 69,
MIMO System,” IEEE Access, vol. 7, pp. 72 170–72 181, Mar. 2019. no. 3, pp. 397–411, Nov. 2018.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 54

[269] M. Zhou, F. Qiu, Z. Tian, K. Xu, and Q. Jiang, “Positioning Error approach,” IEEE Antennas Wirel. Propag. Lett., vol. 18, no. 11, pp.
vs. Signal Distribution: An Analysis Towards Lower Error Bound in 2306–2310, May 2019.
WLAN Fingerprint Based Indoor Localization,” in 2015 IEEE Global [293] Y. Peng, W. Fan, X. Dong, and X. Zhang, “An iterative weighted
Communications Conference (GLOBECOM), Dec. 2015, pp. 1–6. KNN (IW-KNN) based indoor localization method in bluetooth low
[270] M. Zhou, F. Qiu, Z. Tian, H. Wu, Q. Zhang, and W. He, “An energy (BLE) environment,” in 2016 Intl IEEE Conferences on
Information-Based Approach to Precision Analysis of Indoor WLAN Ubiquitous Intelligence & Computing, Advanced and Trusted Com-
Localization Using Location Fingerprint,” Entropy, vol. 17, no. 12, pp. puting, Scalable Computing and Communications, Cloud and Big
8031–8055, Dec. 2015. Data Computing, Internet of People, and Smart World Congress
[271] S. He, B. Ji, and S.-H. G. Chan, “Chameleon: Survey-Free Updating (UIC/ATC/ScalCom/CBDCom/IoP/SmartWorld). IEEE, Jul. 2016, pp.
of a Fingerprint Database for Indoor Localization,” IEEE Pervasive 794–800.
Comput., vol. 15, no. 4, pp. 66–75, Oct. 2016. [294] B. Wang, X. Liu, B. Yu, R. Jia, and X. Gan, “An improved WiFi
[272] M. D’Souza, B. Schoots, and M. Ros, “Indoor position tracking using positioning method based on fingerprint clustering and signal weighted
received signal strength-based fingerprint context aware partitioning,” Euclidean distance,” Sensors, vol. 19, no. 10, p. 2300, May 2019.
IET Radar, Sonar & Navigation, vol. 10, no. 8, pp. 1347–1355, Oct. [295] X. Guo, S. Zhu, L. Li, F. Hu, and N. Ansari, “Accurate WiFi
2016. localization by unsupervised fusion of extended candidate location set,”
[273] Q. Li, W. Li, W. Sun, J. Li, and Z. Liu, “Fingerprint and Assistant IEEE Internet Things J., vol. 6, no. 2, pp. 2476–2485, Sep. 2018.
Nodes Based Wi-Fi Localization in Complex Indoor Environment,” [296] X. Sun, X. Gao, G. Y. Li, and W. Han, “Single-site localization based
IEEE Access, vol. 4, pp. 2993–3004, Jun. 2016. on a new type of fingerprint for massive MIMO-OFDM systems,” IEEE
[274] L. Chen, K. Yang, and X. Wang, “Robust Cooperative Wi-Fi Trans. Veh. Technol., vol. 67, no. 7, pp. 6134–6145, Mar. 2018.
Fingerprint-Based Indoor Localization,” IEEE Internet Things J., vol. 3, [297] X. Wang, L. Liu, Y. Lin, and X. Chen, “A fast single-site fingerprint
no. 6, pp. 1406–1417, Sep. 2016. localization method in massive MIMO system,” in 2019 11th Interna-
[275] J. Zhang, G. Han, N. Sun, and L. Shu, “Path-Loss-Based Fingerprint tional Conference on Wireless Communications and Signal Processing
Localization Approach for Location-Based Services in Indoor Environ- (WCSP). IEEE, Dec. 2019, pp. 1–6.
ments,” IEEE Access, vol. 5, pp. 13 756–13 769, Jul. 2017. [298] B. El Boudani, L. Kanaris, A. Kokkinis, M. Kyriacou, C. Chrysoulas,
[276] B. Zhou, Q. Li, Q. Mao, and W. Tu, “A Robust Crowdsourcing-Based S. Stavrou, and T. Dagiuklas, “Implementing Deep Learning Tech-
Indoor Localization System,” Sensors, vol. 17, no. 4, Apr. 2017. niques in 5G IoT Networks for 3D Indoor Positioning: DELTA (DeEp
[277] G. Pecoraro, S. D. Domenico, E. Cianca, and M. D. Sanctis, “CSI-based Learning-Based Co-operaTive Architecture),” Sensors, vol. 20, no. 19,
fingerprinting for indoor localization using LTE Signals,” EURASIP Sep. 2020.
Journal on Advances in Signal Processing, vol. 2018, no. 1, Jul. 2018. [299] W. Chen, W. Wang, Q. Li, Q. Chang, and H. Hou, “A crowd-sourcing
[278] Y. Duan, K.-Y. Lam, V. C. S. Lee, W. Nie, K. Liu, H. Li, and C. J. indoor localization algorithm via optical camera on a smartphone
Xue, “Data Rate Fingerprinting: A WLAN-Based Indoor Positioning assisted by wi-fi fingerprint RSSI,” Sensors, vol. 16, no. 3, p. 410,
Technique for Passive Localization,” IEEE Sensors J., vol. 19, no. 15, Mar. 2016.
pp. 6517–6529, Apr. 2019. [300] L. Zhang, N. Xiao, J. Li, and W. Yang, “Heterogeneous feature machine
[279] V. Moghtadaiee, S. A. Ghorashi, and M. Ghavami, “New Reconstructed learning for performance-enhancing indoor localization,” in 2018 IEEE
Database for Cost Reduction in Indoor Fingerprinting Localization,” 87th Vehicular Technology Conference (VTC Spring). IEEE, Jun. 2018,
IEEE Access, vol. 7, pp. 104 462–104 477, Jul. 2019. pp. 1–5.
[280] L. Zhang, E. Ding, Y. Hu, and Y. Liu, “A novel CSI-based fingerprint- [301] A. Ray, S. Deb, and P. Monogioudis, “Localization of LTE mea-
ing for localization with a single AP,” EURASIP Journal on Wireless surement records with missing information,” in IEEE INFOCOM
Communications and Networking, vol. 2019, no. 1, Feb. 2019. 2016-The 35th Annual IEEE International Conference on Computer
[281] Z. Gao, Y. Gao, S. Wang, D. Li, and Y. Xu, “CRISLoc: Reconstructable Communications. IEEE, Jul. 2016, pp. 1–9.
CSI Fingerprinting for Indoor Smartphone Localization,” IEEE Internet [302] Y.-K. Cheng, H.-J. Chou, and R. Y. Chang, “Machine-learning indoor
Things J., vol. 8, no. 5, pp. 3422–3437, Sep. 2021. localization with access point selection and signal strength recon-
[282] Y. Li, Z. He, Z. Gao, Y. Zhuang, C. Shi, and N. El-Sheimy, “Toward struction,” in 2016 IEEE 83rd Vehicular Technology Conference (VTC
Robust Crowdsourcing-Based Localization: A Fingerprinting Accuracy Spring). IEEE, Jul. 2016, pp. 1–5.
Indicator Enhanced Wireless/Magnetic/Inertial Integration Approach,” [303] J. Li, I.-T. Lu, J. S. Lu, and L. Zhang, “Robust kernel-based machine
IEEE Internet Things J., vol. 6, no. 2, pp. 3585–3600, Dec. 2019. learning localization using NLOS TOAs or TDOAs,” in 2017 IEEE
[283] Z. Shen, K. Xu, and X. Xia, “2D Fingerprinting-Based Localization for Long Island Systems, Applications and Technology Conference (LISAT).
mmWave Cell-Free Massive MIMO Systems,” IEEE Commun. Lett., IEEE, May 2017, pp. 1–6.
vol. 25, no. 11, pp. 3556–3560, Sep. 2021. [304] S. Kram, M. Stahlke, T. Feigl, J. Seitz, and J. Thielecke, “UWB channel
[284] D. Schlegel, M. Colosi, and G. Grisetti, “ProSLAM: Graph SLAM impulse responses for positioning in complex environments: A detailed
from a Programmer’s Perspective,” in 2018 IEEE International Confer- feature analysis,” Sensors, vol. 19, no. 24, p. 5547, Dec. 2019.
ence on Robotics and Automation (ICRA), Sep. 2018, pp. 3833–3840. [305] X. Yang, F. Zhao, and T. Chen, “NLOS identification for UWB loca-
[285] M. Holder, S. Hellwig, and H. Winner, “Real-Time Pose Graph SLAM lization based on import vector machine,” AEU-International Journal
based on Radar,” in 2019 IEEE Intelligent Vehicles Symposium (IV), of Electronics and Communications, vol. 87, pp. 128–133, Apr. 2018.
Aug. 2019, pp. 1145–1151. [306] S. Zhang, J. Guo, N. Luo, L. Wang, W. Wang, and K. Wen, “Improving
[286] J. Cui, D. Feng, Y. Li, and Q. Tian, “Research on Simultaneous Wi-Fi fingerprint positioning with a pose recognition-assisted SVM
Localization and Mapping for AUV by an Improved Method: Variance algorithm,” Remote Sensing, vol. 11, no. 6, p. 652, Mar. 2019.
Reduction FastSLAM with Simulated Annealing,” Defence Technology, [307] Z. Wu, L. Jiang, Z. Jiang, B. Chen, K. Liu, Q. Xuan, and Y. Xiang,
vol. 16, no. 3, pp. 651–661, Jun. 2020. “Accurate indoor localization based on CSI and visibility graph,”
[287] M. L. Rodrı́guez-Arévalo, J. Neira, and J. A. Castellanos, “On the Sensors, vol. 18, no. 8, p. 2549, Aug. 2018.
Importance of Uncertainty Representation in Active SLAM,” IEEE [308] K. Wu, M. Yang, C. Ma, and J. Yan, “CSI-based wireless localization
Trans. Rob., vol. 34, no. 3, pp. 829–834, Apr. 2018. and activity recognition using support vector machine,” in 2019 IEEE
[288] M. Lin, C. Yang, D. Li, and G. Zhou, “Intelligent Filter-Based SLAM International Conference on Signal Processing, Communications and
for Mobile Robots With Improved Localization Performance,” IEEE Computing (ICSPCC). IEEE, Sep. 2019, pp. 1–5.
Access, vol. 7, pp. 113 284–113 297, Aug. 2019. [309] T. F. Sanam and H. Godrich, “An improved CSI based device free
[289] Y. Chen, S. Huang, and R. Fitch, “Active SLAM for Mobile Robots indoor localization using machine learning based classification ap-
With Area Coverage and Obstacle Avoidance,” IEEE/ASME Trans. proach,” in 2018 26th European Signal Processing Conference (EU-
Mechatron., vol. 25, no. 3, pp. 1182–1192, Jan. 2020. SIPCO). IEEE, Dec. 2018, pp. 2390–2394.
[290] H. Wang, C. Wang, and L. Xie, “Intensity-SLAM: Intensity Assisted [310] S. Mahfouz, F. Mourad-Chehade, P. Honeine, J. Farah, and H. Snoussi,
Localization and Mapping for Large Scale Environment,” IEEE Robot. “Non-parametric and semi-parametric RSSI/distance modeling for tar-
Autom. Lett., vol. 6, no. 2, pp. 1715–1721, Feb. 2021. get tracking in wireless sensor networks,” IEEE Sensors J., vol. 16,
[291] Y. Chen, S. Huang, L. Zhao, and G. Dissanayake, “Cramér–Rao no. 7, pp. 2115–2126, Dec. 2015.
Bounds and Optimal Design Metrics for Pose-Graph SLAM,” IEEE [311] M. Petric, A. Neskovic, N. Neskovic, and M. Borenovic, “Indoor
Trans. Rob., vol. 37, no. 2, pp. 627–641, Jan. 2021. localization using multi-operator public land mobile networks and
[292] M. I. AlHajri, N. T. Ali, and R. M. Shubair, “Indoor localization for support vector machine learning algorithms,” Wireless Personal Com-
IoT using adaptive feature selection: A cascaded machine learning munications, vol. 104, no. 4, pp. 1573–1597, Feb. 2019.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 55

[312] L. Zhang, Y. Li, Y. Gu, and W. Yang, “An efficient machine learning [333] X. Wang, M. Patil, C. Yang, S. Mao, and P. A. Patel, “Deep Convo-
approach for indoor localization,” China Commun., vol. 14, no. 11, pp. lutional Gaussian Processes for Mmwave Outdoor Localization,” in
141–150, Nov. 2017. ICASSP 2021 - 2021 IEEE International Conference on Acoustics,
[313] Y. Rezgui, L. Pei, X. Chen, F. Wen, and C. Han, “An efficient Speech and Signal Processing (ICASSP), Jun. 2021, pp. 8323–8327.
normalized rank based SVM for room level indoor WiFi localization [334] K. N. R. S. V. Prasad, E. Hossain, V. K. Bhargava, and S. Mallick,
with diverse devices,” Mob Inf Syst, vol. 2017, Jul. 2017. “Analytical Approximation-Based Machine Learning Methods for User
[314] A. Khalajmehrabadi, N. Gatsis, and D. Akopian, “Modern WLAN Positioning in Distributed Massive MIMO,” IEEE Access, vol. 6, pp.
fingerprinting indoor positioning methods and deployment challenges,” 18 431–18 452, Feb. 2018.
IEEE Commun. Surv., vol. 19, no. 3, pp. 1974–2002, Mar. 2017. [335] E. Homayounvala, M. Nabati, R. Shahbazian, S. A. Ghorashi, and
[315] H. Wang, Z. Zhao, J. Hu, Z. Qu, and H. Feng, “Study on im- V. Moghtadaiee, “A novel smartphone application for indoor position-
provement of fingerprint matching algorithm in wireless LAN based ing of users based on machine learning,” in Adjunct proceedings of the
indoor positioning system,” in 2016 17th IEEE/ACIS international 2019 ACM international joint conference on pervasive and ubiquitous
conference on software engineering, artificial intelligence, networking computing and proceedings of the 2019 ACM international symposium
and parallel/distributed computing (SNPD). IEEE, Jul. 2016, pp. 275– on wearable computers, Sep. 2019, pp. 430–437.
280. [336] W. Sun, M. Xue, H. Yu, H. Tang, and A. Lin, “Augmentation of
[316] X. Guo, N. Ansari, F. Hu, Y. Shao, N. R. Elikplim, and L. Li, “A survey fingerprints for indoor WiFi localization based on Gaussian process
on fusion-based indoor positioning,” IEEE Commun. Surv., vol. 22, regression,” IEEE Trans. Veh. Technol., vol. 67, no. 11, pp. 10 896–
no. 1, pp. 566–594, Nov. 2019. 10 905, Sep. 2018.
[317] J. Yan, L. Zhao, J. Tang, Y. Chen, R. Chen, and L. Chen, “Hybrid kernel [337] J. Zhao, X. Gao, X. Wang, C. Li, M. Song, and Q. Sun, “An efficient
based machine learning using received signal strength measurements radio map updating algorithm based on K-Means and gaussian process
for indoor localization,” IEEE Trans. Veh. Technol., vol. 67, no. 3, pp. regression,” The Journal of Navigation, vol. 71, no. 5, pp. 1055–1068,
2824–2829, Nov. 2017. Apr. 2018.
[318] H. Ahmadi and R. Bouallegue, “RSSI-based localization in wireless [338] S. Yiu, M. Dashti, H. Claussen, and F. Perez-Cruz, “Wireless RSSI
sensor networks using Regression Tree,” in 2015 International Wireless fingerprinting localization,” Signal Process., vol. 131, pp. 235–244,
Communications and Mobile Computing Conference (IWCMC). IEEE, Feb. 2017.
Oct. 2015, pp. 1548–1553. [339] S.-H. Fang, C.-C. Li, W.-C. Lu, Z. Xu, and Y.-R. Chien, “Enhanced
[319] B. A. Akram, A. H. Akbar, and O. Shafiq, “HybLoc: Hybrid indoor device-free human detection: Efficient learning from phase and ampli-
Wi-Fi localization using soft clustering-based random decision forest tude of channel state information,” IEEE Trans. Veh. Technol., vol. 68,
ensembles,” IEEE Access, vol. 6, pp. 38 251–38 272, Jul. 2018. no. 3, pp. 3048–3051, Jan. 2019.
[320] X. Guo, N. Ansari, L. Li, and H. Li, “Indoor localization by fusing a [340] S. Wu, S. Zhang, K. Xu, and D. Huang, “Neural network localization
group of fingerprints based on random forests,” IEEE Internet Things with TOA measurements based on error learning and matching,” IEEE
J., vol. 5, no. 6, pp. 4686–4698, Feb. 2018. Access, vol. 7, pp. 19 089–19 099, Feb. 2019.
[321] M. Malmstr´’om, I. Skog, S. M. Razavi, Y. Zhao, and F. Gunnarsson, [341] X. Wang, L. Gao, S. Mao, and S. Pandey, “CSI-based fingerprinting
“5G Positioning - A Machine Learning Approach,” in 2019 16th for indoor localization: A deep learning approach,” IEEE Trans. Veh.
Workshop on Positioning, Navigation and Communications (WPNC), Technol., vol. 66, no. 1, pp. 763–776, Mar. 2016.
Mar. 2019, pp. 1–6. [342] X. Wang, L. Gao, and S. Mao, “PhaseFi: Phase fingerprinting for indoor
[322] W. Y. Al-Rashdan and A. Tahat, “A Comparative Performance Eval- localization with a deep learning approach,” in 2015 IEEE Global
uation of Machine Learning Algorithms for Fingerprinting Based Communications Conference (GLOBECOM). IEEE, Dec. 2015, pp.
Localization in DM-MIMO Wireless Systems Relying on Big Data 1–6.
Techniques,” IEEE Access, vol. 8, pp. 109 522–109 534, Jun. 2020. [343] A. Belmonte-Hernández, G. Hernández-Peñaloza, D. M. Gutiérrez, and
[323] A. Musa, G. D. Nugraha, H. Han, D. Choi, S. Seo, and J. Kim, “A F. Alvarez, “SWiBluX: Multi-Sensor Deep Learning Fingerprint for
decision tree-based NLOS detection method for the UWB indoor lo- precise real-time indoor tracking,” IEEE Sensors J., vol. 19, no. 9, pp.
cation tracking accuracy improvement,” Int. J. Commun. Syst., vol. 32, 3473–3486, Jan. 2019.
no. 13, p. e3997, Jun. 2019. [344] Z. Feng, Y. Cao, and J. Yan, “A Received Signal Strength Based Indoor
[324] M. N. de Sousa, “Enhanced localization systems with multipath finger- Localization Algorithm Using ELM Technique and Ridge Regression,”
prints and machine learning,” in 2019 IEEE 30th Annual International in 2019 IEEE 2nd International Conference on Electronic Information
Symposium on Personal, Indoor and Mobile Radio Communications and Communication Technology (ICEICT). IEEE, Jan. 2019, pp. 599–
(PIMRC). IEEE, Nov. 2019, pp. 1–6. 603.
[325] Y. Zhang, D. Li, and Y. Wang, “An indoor passive positioning method [345] J. Yang, S. Jin, C.-K. Wen, J. Guo, and M. Matthaiou, “3-D positioning
using CSI fingerprint based on Adaboost,” IEEE Sensors J., vol. 19, and environment mapping for mmWave communication systems,”
no. 14, pp. 5792–5800, Mar. 2019. arXiv preprint arXiv:1908.04142, Aug. 2019.
[326] P. Chen, J. Shang, and F. Gu, “Learning RSSI feature via ranking [346] J.-Y. Lee, C. Eom, Y. Kwak, H.-G. Kang, and C. Lee, “DNN-
model for Wi-Fi fingerprinting localization,” IEEE Trans. Veh. Technol., based wireless positioning in an outdoor environment,” in 2018 IEEE
vol. 69, no. 2, pp. 1695–1705, Dec. 2019. International Conference on Acoustics, Speech and Signal Processing
[327] G. H. Apostolo, I. G. B. Sampaio, and J. Viterbo, “Feature selection (ICASSP). IEEE, Sep. 2018, pp. 3799–3803.
on database optimization for Wi-Fi fingerprint indoor positioning,” [347] H. Ge, F. Jiang, and Z. Zhang, “A hybrid localization algorithm of rss
Procedia Comput. Sci., vol. 159, pp. 251–260, Oct. 2019. and toa based on an ensembled neural network,” in 2019 IEEE 8th
[328] V. Savic, E. G. Larsson, J. Ferrer-Coll, and P. Stenumgaard, “Kernel Joint International Information Technology and Artificial Intelligence
methods for accurate UWB-based ranging with reduced complexity,” Conference (ITAIC). IEEE, Aug. 2019, pp. 1280–1284.
IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 1783–1793, Oct. [348] R. H. Jaafar and S. S. Saab, “A neural network approach for indoor
2015. fingerprinting-based localization,” in 2018 9th IEEE Annual Ubiquitous
[329] K. S. V. Prasad, E. Hossain, and V. K. Bhargava, “A numerical Computing, Electronics & Mobile Communication Conference (UEM-
approximation method for RSS-based user positioning in distributed CON). IEEE, Aug. 2018, pp. 537–542.
massive MIMO,” in 2017 IEEE International Conference on Advanced [349] A. Sobehy, E. Renault, and P. Mühlethaler, “CSI based indoor localiza-
Networks and Telecommunications Systems (ANTS). IEEE, Jun. 2017, tion using ensemble neural networks,” in International Conference on
pp. 1–6. Machine Learning for Networking. Springer, Apr. 2019, pp. 367–378.
[330] K. N. R. S. V. Prasad, E. Hossain, and V. K. Bhargava, “Machine [350] A. Decurninge, L. G. Ordóñez, P. Ferrand, H. Gaoning, L. Bojie,
Learning Methods for RSS-Based User Positioning in Distributed Z. Wei, and M. Guillaud, “CSI-based outdoor localization for massive
Massive MIMO,” IEEE Trans. Wireless Commun., vol. 17, no. 12, pp. MIMO: Experiments with a learning approach,” in 2018 15th Inter-
8402–8417, Oct. 2018. national Symposium on Wireless Communication Systems (ISWCS).
[331] S. Kumar, R. M. Hegde, and N. Trigoni, “Gaussian process regression IEEE, Oct. 2018, pp. 1–6.
for fingerprinting based localization,” Ad Hoc Networks, vol. 51, pp. [351] M. Orabi, A. A. Abdallah, J. Khalife, and Z. M. Kassas, “A Machine
1–10, Sep. 2016. Learning Multipath Mitigation Approach for Opportunistic Navigation
[332] K. Yu, K. Wen, Y. Li, S. Zhang, and K. Zhang, “A novel NLOS miti- with 5G Signals,” in Proceedings of the 34th International Technical
gation algorithm for UWB localization in harsh indoor environments,” Meeting of the Satellite Division of The Institute of Navigation (ION
IEEE Trans. Veh. Technol., vol. 68, no. 1, pp. 686–699, Nov. 2018. GNSS 2021). Institute of Navigation, Oct. 2021.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 56

[352] J. Gante, G. Falcão, and L. Sousa, “Deep Learning Architectures for Personal, Indoor, and Mobile Radio Communications (PIMRC). IEEE,
Accurate Millimeter Wave Positioning in 5G,” Neural Process Lett, Feb. 2017, pp. 1–6.
vol. 51, no. 1, pp. 487–514, Aug. 2019. [373] Z. Liu, B. Dai, X. Wan, and X. Li, “Hybrid wireless fingerprint indoor
[353] M. Z. Comiter, M. B. Crouse, and H. T. Kung, “A Data-Driven localization method based on a convolutional neural network,” Sensors,
Approach to Localization for High Frequency Wireless Mobile Net- vol. 19, no. 20, p. 4597, Oct. 2019.
works,” in GLOBECOM 2017 - 2017 IEEE Global Communications [374] H. Huang and S. Lin, “WiDet: Wi-Fi based device-free passive person
Conference, Jan. 2017, pp. 1–7. detection with deep convolutional neural networks,” in Proceedings
[354] Y. Zhang and K. Psounis, “Efficient Indoor Localization via Switched- of the 21st ACM International Conference on Modeling, Analysis and
Beam Antennas,” IEEE Trans. Mob. Comput., vol. 19, no. 9, pp. 2101– Simulation of Wireless and Mobile Systems, Dec. 2018, pp. 53–60.
2115, Jun. 2019. [375] K. Bregar and M. Mohorčič, “Improving indoor localization using
[355] Z. E. Khatab, A. Hajihoseini, and S. A. Ghorashi, “A fingerprint convolutional neural networks on computationally restricted devices,”
method for indoor localization using autoencoder based deep extreme IEEE Access, vol. 6, pp. 17 429–17 441, Mar. 2018.
learning machine,” IEEE Sensors Letters, vol. 2, no. 1, pp. 1–4, Dec. [376] A. Mittal, S. Tiku, and S. Pasricha, “Adapting convolutional neural
2017. networks for indoor localization with smart mobile devices,” in Pro-
[356] K. S. Kim, “Hybrid building/floor classification and location coordi- ceedings of the 2018 on Great Lakes Symposium on VLSI, May 2018,
nates regression using a single-input and multi-output deep neural net- pp. 117–122.
work for large-scale indoor localization based on Wi-Fi fingerprinting,” [377] A. Niitsoo, T. Edelhäußer, E. Eberlein, N. Hadaschik, and C. Mutschler,
in 2018 Sixth International Symposium on Computing and Networking “A deep learning approach to position estimation from channel impulse
Workshops (CANDARW). IEEE, Dec. 2018, pp. 196–201. responses,” Sensors, vol. 19, no. 5, p. 1064, Mar. 2019.
[357] X. Song, X. Fan, C. Xiang, Q. Ye, L. Liu, Z. Wang, X. He, N. Yang, [378] H. Turabieh and A. Sheta, “Cascaded layered recurrent neural network
and G. Fang, “A novel convolutional neural network based indoor for indoor localization in wireless sensor networks,” in 2019 2nd
localization framework with WiFi fingerprinting,” IEEE Access, vol. 7, International Conference on New Trends in COMPUTING Sciences
pp. 110 698–110 709, Aug. 2019. (ICTCS). IEEE, Dec. 2019, pp. 1–6.
[358] S. BelMannoubi, H. Touati, and H. Snoussi, “Stacked auto-encoder [379] Y. Zhang, Y. Xiao, K. Zhao, and W. Rao, “DeepLoc: deep neural
for scalable indoor localization in wireless sensor networks,” in 2019 network-based telco localization,” in Proceedings of the 16th EAI Inter-
15th International Wireless Communications and Mobile Computing national Conference on Mobile and Ubiquitous Systems: Computing,
Conference (IWCMC). IEEE, Jun. 2019, pp. 1245–1250. Networking and Services, Nov. 2019, pp. 258–267.
[359] W. Zhang, R. Sengupta, J. Fodero, and X. Li, “DeepPositioning: [380] L. Wu, C.-H. Chen, and Q. Zhang, “A mobile positioning method
Intelligent fusion of pervasive magnetic field and WiFi fingerprinting based on deep learning techniques,” Electronics, vol. 8, no. 1, p. 59,
for smartphone indoor localization via deep learning,” in 2017 16th Jan. 2019.
IEEE International Conference on Machine Learning and Applications [381] J. Yu, H. M. Saad, and R. M. Buehrer, “Centimeter-level indoor locali-
(ICMLA). IEEE, Dec. 2017, pp. 7–13. zation using channel state information with recurrent neural networks,”
[360] P. Yazdanian and V. Pourahmadi, “DeepPos: Deep supervised au- in 2020 IEEE/ION Position, Location and Navigation Symposium
toencoder network for CSI based indoor localization,” arXiv preprint (PLANS). IEEE, Jun. 2020, pp. 1317–1323.
arXiv:1811.12182, Nov. 2018. [382] G. B. Tarekegn, H.-P. Lin, A. B. Adege, Y. Y. Munaye, and S.-S.
Jeng, “Applying long short-term memory (LSTM) mechanisms for
[361] M. Abbas, M. Elhamshary, H. Rizk, M. Torki, and M. Youssef,
fingerprinting outdoor positioning in hybrid networks,” in 2019 IEEE
“WiDeep: WiFi-based accurate and robust indoor localization system
90th Vehicular Technology Conference (VTC2019-Fall). IEEE, Nov.
using deep learning,” in 2019 IEEE International Conference on
2019, pp. 1–5.
Pervasive Computing and Communications (PerCom). IEEE, Jul.
[383] V. Cantón Paterna, A. Calveras Auge, J. Paradells Aspas, and M. A.
2019, pp. 1–10.
Perez Bullones, “A bluetooth low energy indoor positioning system
[362] X. Wang, L. Gao, and S. Mao, “BiLoc: Bi-modal deep learning for
with channel diversity, weighted trilateration and kalman filtering,”
indoor localization with commodity 5GHz WiFi,” IEEE Access, vol. 5,
Sens., vol. 17, no. 12, p. 2927, Dec. 2017.
pp. 4209–4220, Mar. 2017.
[384] H. Rizk, “Solocell: Efficient indoor localization based on limited cell
[363] J.-W. Jang and S.-N. Hong, “Indoor localization with wifi fingerprinting
network information and minimal fingerprinting,” in Proceedings of
using convolutional neural network,” in 2018 Tenth International
the 27th ACM SIGSPATIAL International Conference on Advances in
Conference on Ubiquitous and Future Networks (ICUFN). IEEE,
Geographic Information Systems, Nov. 2019, pp. 604–605.
Aug. 2018, pp. 753–758.
[385] H. Rizk and M. Youssef, “Monodcell: A ubiquitous and low-overhead
[364] T. Ebuchi and H. Yamamoto, “Vehicle/pedestrian localization system deep learning-based indoor localization with limited cellular infor-
using multiple radio beacons and machine learning for smart parking,” mation,” in Proceedings of the 27th ACM SIGSPATIAL International
in 2019 International Conference on Artificial Intelligence in Informa- Conference on Advances in Geographic Information Systems, Jul. 2019,
tion and Communication (ICAIIC). IEEE, Mar. 2019, pp. 086–091. pp. 109–118.
[365] M. Ibrahim, M. Torki, and M. ElNainay, “CNN based indoor localiza- [386] Z. Chen, H. Zou, J. Yang, H. Jiang, and L. Xie, “WiFi fingerprinting
tion using RSS time-series,” in 2018 IEEE symposium on computers indoor localization using local feature-based deep LSTM,” IEEE Syst.
and communications (ISCC). IEEE, Nov. 2018, pp. 01 044–01 049. J., vol. 14, no. 2, pp. 3001–3010, Jun. 2019.
[366] B. Soro and C. Lee, “Joint time-frequency RSSI features for convolu- [387] W. Qian, F. Lauri, and F. Gechter, “Convolutional mixture density
tional neural network-based indoor fingerprinting localization,” IEEE recurrent neural network for predicting user location with WiFi fin-
Access, vol. 7, pp. 104 892–104 899, Aug. 2019. gerprints,” arXiv preprint arXiv:1911.09344, Nov. 2019.
[367] H. Chen, Y. Zhang, W. Li, X. Tao, and P. Zhang, “ConFi: Convolutional [388] A. B. Adege, H.-P. Lin, and L.-C. Wang, “Mobility predictions for iot
neural networks based indoor Wi-Fi localization using channel state devices using gated recurrent unit network,” IEEE Internet Things J.,
information,” IEEE Access, vol. 5, pp. 18 066–18 074, Sep. 2017. vol. 7, no. 1, pp. 505–517, Oct. 2019.
[368] X. Wang, X. Wang, and S. Mao, “Deep convolutional neural networks [389] D. V. Le, N. Meratnia, and P. J. Havinga, “Unsupervised deep feature
for indoor localization with CSI images,” IEEE Trans. Netw. Sci. Eng., learning to reduce the collection of fingerprints for indoor localization
vol. 7, no. 1, pp. 316–327, Sep. 2018. using deep belief networks,” in 2018 International conference on
[369] Y. Jing, J. Hao, and P. Li, “Learning spatiotemporal features of indoor positioning and indoor navigation (IPIN). IEEE, Sep. 2018,
CSI for indoor localization with dual-stream 3D convolutional neural pp. 1–7.
networks,” IEEE Access, vol. 7, pp. 147 571–147 585, Oct. 2019. [390] D. Sikeridis, B. P. Rimal, I. Papapanagiotou, and M. Devetsikiotis, “Un-
[370] H. Li, X. Zeng, Y. Li, S. Zhou, and J. Wang, “Convolutional neural supervised crowd-assisted learning enabling location-aware facilities,”
networks based indoor Wi-Fi localization with a novel kind of CSI IEEE Internet Things J., vol. 5, no. 6, pp. 4699–4713, Mar. 2018.
images,” China Commun., vol. 16, no. 9, pp. 250–260, Sep. 2019. [391] N. Saeed and H. Nam, “Cluster based multidimensional scaling for
[371] C. Wu, X. Yi, W. Wang, L. You, Q. Huang, X. Gao, and Q. Liu, irregular cognitive radio networks localization,” IEEE Trans. Signal
“Learning to localize: A 3D CNN approach to user positioning in Process., vol. 64, no. 10, pp. 2649–2659, Feb. 2016.
massive MIMO-OFDM systems,” IEEE Trans. Wireless Commun., [392] J. Choi, Y.-S. Choi, and S. Talwar, “Unsupervised learning techniques
vol. 20, no. 7, pp. 4556–4570, Feb. 2021. for trilateration: From theory to android app implementation,” IEEE
[372] J. Vieira, E. Leitinger, M. Sarajlic, X. Li, and F. Tufvesson, “Deep Access, vol. 7, pp. 134 525–134 538, Sep. 2019.
convolutional neural networks for massive MIMO fingerprint-based [393] ——, “Unsupervised learning technique to obtain the coordinates of
positioning,” in 2017 IEEE 28th Annual International Symposium on Wi-Fi access points,” in 2019 International Conference on Indoor
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 57

Positioning and Indoor Navigation (IPIN). IEEE, Nov. 2019, pp. [415] L. Xiao, A. Behboodi, and R. Mathar, “Learning the localization
1–6. function: Machine learning approach to fingerprinting localization,”
[394] B. Jang and H. Kim, “Indoor positioning technologies without offline arXiv preprint arXiv:1803.08153, Mar. 2018.
fingerprinting map: A survey,” IEEE Commun. Surv., vol. 21, no. 1, [416] K. Liu, H. Zhang, J. K.-Y. Ng, Y. Xia, L. Feng, V. C. Lee, and S. H.
pp. 508–525, Aug. 2018. Son, “Toward low-overhead fingerprint-based indoor localization via
[395] Y. Ye and B. Wang, “RMapCS: Radio map construction from crowd- transfer learning: Design, implementation, and evaluation,” IEEE Trans.
sourced samples for indoor localization,” IEEE Access, vol. 6, pp. Ind. Inf., vol. 14, no. 3, pp. 898–908, Sep. 2017.
24 224–24 238, Apr. 2018. [417] X. Zhang, Y. Mei, H. Jin, and D. Liang, “TL-FCMA: Indoor lo-
[396] Z. Gao, Y. Ma, K. Liu, X. Miao, and Y. Zhao, “An indoor multi-tag calization by integrating fuzzy clustering with transfer learning,” in
cooperative localization algorithm based on NMDS for RFID,” IEEE 2018 International Conference on Network Infrastructure and Digital
Sensors J., vol. 17, no. 7, pp. 2120–2128, Feb. 2017. Content (IC-NIDC). IEEE, Aug. 2018, pp. 372–377.
[397] S. Niknam, H. S. Dhillon, and J. H. Reed, “Federated learning for [418] H. Xu, “Semi-supervised manifold learning based on polynomial
wireless communications: Motivation, opportunities, and challenges,” mapping for localization in wireless sensor networks,” Signal Process.,
IEEE Commun. Mag., vol. 58, no. 6, pp. 46–51, Jun. 2020. vol. 172, p. 107570, May 2020.
[398] L. Chen, S. Thombre, K. Järvinen, E. S. Lohan, A. Alén-Savikko, [419] M. Zhou, Y. Tang, Z. Tian, L. Xie, and W. Nie, “Robust neighborhood
H. Leppäkoski, M. Z. H. Bhuiyan, S. Bu-Pasha, G. N. Ferrara, graphing for semi-supervised indoor localization with light-loaded
S. Honkala et al., “Robustness, security and privacy in location-based location fingerprinting,” IEEE Internet Things J., vol. 5, no. 5, pp.
services for future IoT: A survey,” IEEE Access, vol. 5, pp. 8956–8977, 3378–3387, Oct. 2017.
Apr. 2017.
[420] D. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52,
[399] W. Li, C. Zhang, and Y. Tanaka, “Pseudo label-driven federated
no. 4, pp. 1289–1306, Apr. 2006.
learning-based decentralized indoor localization via mobile crowd-
sourcing,” IEEE Sensors J., vol. 20, no. 19, pp. 11 556–11 565, May [421] E. J. Candes and M. B. Wakin, “An Introduction To Compressive
2020. Sampling,” IEEE Signal Process Mag., vol. 25, no. 2, pp. 21–30, Mar.
[400] D. C. Nguyen, M. Ding, P. N. Pathirana, A. Seneviratne, J. Li, and 2008.
H. V. Poor, “Federated learning for internet of things: A comprehensive [422] F. Seco, A. R. Jiménez, and P. Peltola, “A Review of Multidimen-
survey,” IEEE Commun. Surv., vol. 23, no. 3, pp. 1622–1658, Apr. sional Scaling Techniques for RSS-Based WSN Localization,” in 2018
2021. International Conference on Indoor Positioning and Indoor Navigation
[401] Y. Liu, H. Li, J. Xiao, and H. Jin, “FLoc: Fingerprint-based indoor (IPIN), Nov. 2018, pp. 1–8.
localization system under a federated learning updating framework,” [423] Q. D. Vo and P. De, “A Survey of Fingerprint-Based Outdoor Locali-
in 2019 15th International Conference on Mobile Ad-Hoc and Sensor zation,” IEEE Commun. Surv., vol. 18, no. 1, pp. 491–506, Jun. 2016.
Networks (MSN). IEEE, Apr. 2019, pp. 113–118. [424] D. Sánchez-Rodrı́guez, P. Hernández-Morera, J. M. Quinteiro, and
[402] F. Yin, Z. Lin, Q. Kong, Y. Xu, D. Li, S. Theodoridis, and S. R. Cui, I. Alonso-González, “A low complexity system based on multiple
“FedLoc: Federated learning framework for data-driven cooperative weighted decision trees for indoor localization,” Sensors, vol. 15, no. 6,
localization and location data processing,” IEEE Open Journal of pp. 14 809–14 829, Jun. 2015.
Signal Processing, vol. 1, pp. 187–215, Nov. 2020. [425] J. Wang, X. Zhang, Q. Gao, H. Yue, and H. Wang, “Device-free wire-
[403] W. Shao, H. Luo, F. Zhao, Y. Ma, Z. Zhao, and A. Crivello, “Indoor po- less localization and activity recognition: A deep learning approach,”
sitioning based on fingerprint-image and deep learning,” IEEE Access, IEEE Trans. Veh. Technol., vol. 66, no. 7, pp. 6258–6267, Dec. 2016.
vol. 6, pp. 74 699–74 712, Nov. 2018. [426] L. Zhao, H. Huang, X. Li, S. Ding, H. Zhao, and Z. Han, “An accurate
[404] Y. Li, X. Hu, Y. Zhuang, Z. Gao, P. Zhang, and N. El-Sheimy, “Deep and robust approach of device-free localization with convolutional
reinforcement learning (DRL): Another perspective for unsupervised autoencoder,” IEEE Internet Things J., vol. 6, no. 3, pp. 5825–5840,
wireless localization,” IEEE Internet Things J., vol. 7, no. 7, pp. 6279– Mar. 2019.
6287, Dec. 2019. [427] L. Zhao, H. Huang, S. Ding, and X. Li, “An accurate and efficient
[405] F. Dou, J. Lu, Z. Wang, X. Xiao, J. Bi, and C.-H. Huang, “Top-down device-free localization approach based on Gaussian Bernoulli re-
indoor localization with Wi-fi fingerprints using deep Q-network,” in stricted Boltzmann machine,” in 2018 IEEE International Conference
2018 IEEE 15th International Conference on Mobile Ad Hoc and on Systems, Man, and Cybernetics (SMC). IEEE, Oct. 2018, pp.
Sensor Systems (MASS). IEEE, Dec. 2018, pp. 166–174. 2323–2328.
[406] V. F. Mirama, L. E. Diez, A. Bahillo, and V. Quintero, “A Survey of [428] H. Rizk, “Device-invariant cellular-based indoor localization system
Machine Learning in Pedestrian Localization Systems: Applications, using deep learning,” in The ACM MobiSys 2019 on Rising Stars
Open Issues and Challenges,” IEEE Access, vol. 9, pp. 120 138– Forum, Jun. 2019, pp. 19–23.
120 157, Aug. 2021. [429] X. Jiang, Y. Chen, J. Liu, Y. Gu, and L. Hu, “FSELM: fusion semi-
[407] M. Fahad, Z. Chen, and Y. Guo, “Learning how pedestrians navigate: supervised extreme learning machine for indoor localization with Wi-Fi
A deep inverse reinforcement learning approach,” in 2018 IEEE/RSJ and Bluetooth fingerprints,” Soft Comput, vol. 22, no. 11, pp. 3621–
International Conference on Intelligent Robots and Systems (IROS). 3635, Apr. 2018.
IEEE, Jan. 2018, pp. 819–826.
[430] S. Ghafoor, N. Boujnah, M. H. Rehmani, and A. Davy, “MAC
[408] S. De Bast, A. P. Guevara, and S. Pollin, “CSI-based positioning in
protocols for terahertz communication: A comprehensive survey,” IEEE
massive MIMO systems using convolutional neural networks,” in 2020
Commun. Surv., vol. 22, no. 4, pp. 2236–2282, Aug. 2020.
IEEE 91st Vehicular Technology Conference (VTC2020-Spring). IEEE,
Jun. 2020, pp. 1–5. [431] A. Yastrebova, R. Kirichek, Y. Koucheryavy, A. Borodin, and
[409] P. Li, H. Cui, A. Khan, U. Raza, R. Piechocki, A. Doufexi, and A. Koucheryavy, “Future networks 2030: Architecture & requirements,”
T. Farnham, “Deep transfer learning for WiFi localization,” in 2021 in 2018 10th International Congress on Ultra Modern Telecommuni-
IEEE Radar Conference (RadarConf21). IEEE, Jun. 2021, pp. 1–5. cations and Control Systems and Workshops (ICUMT). IEEE, Nov.
[410] H. Zou, Y. Zhou, H. Jiang, B. Huang, L. Xie, and C. Spanos, “Adap- 2018, pp. 1–8.
tive localization in dynamic indoor environments by transfer kernel [432] A. Faisal, H. Sarieddeen, H. Dahrouj, T. Y. Al-Naffouri, and M.-S.
learning,” in 2017 IEEE wireless communications and networking Alouini, “Ultramassive MIMO systems at terahertz bands: Prospects
conference (WCNC). IEEE, Mar. 2017, pp. 1–6. and challenges,” IEEE Veh. Technol. Mag., vol. 15, no. 4, pp. 33–42,
[411] M. Zhou, Y. Tang, W. Nie, L. Xie, and X. Yang, “GrassMA: graph- Oct. 2020.
based semi-supervised manifold alignment for indoor WLAN localiza- [433] V. Petrov, M. Komarov, D. Moltchanov, J. M. Jornet, and Y. Kouch-
tion,” IEEE Sensors J., vol. 17, no. 21, pp. 7086–7095, Sep. 2017. eryavy, “Interference and SINR in millimeter wave and terahertz
[412] A. P. Behera, A. Singh, S. Verma, and M. Kumar, “Manifold learning communication systems with blocking and directional antennas,” IEEE
with localized procrustes analysis based WSN localization,” IEEE Sens. Trans. Wireless Commun., vol. 16, no. 3, pp. 1791–1808, Jan. 2017.
Lett., vol. 4, no. 10, pp. 1–4, Sep. 2020. [434] C. Han, W. Tong, and X.-W. Yao, “MA-ADM: A memory-assisted
[413] J. Yoo, “Time-series Laplacian semi-supervised learning for indoor angular-division-multiplexing MAC protocol in Terahertz communica-
localization,” Sensors, vol. 19, no. 18, p. 3867, Sep. 2019. tion networks,” Nano communication networks, vol. 13, pp. 51–59,
[414] E. Lei, O. Castañeda, O. Tirkkonen, T. Goldstein, and C. Studer, Aug. 2017.
“Siamese neural networks for wireless positioning and channel chart- [435] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment:
ing,” in 2019 57th Annual Allerton Conference on Communication, Intelligent reflecting surface aided wireless network,” IEEE Commun.
Control, and Computing (Allerton). IEEE, Sep. 2019, pp. 200–207. Mag., vol. 58, no. 1, pp. 106–112, Nov. 2019.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 58

[436] R. Zhang and S. Cao, “Extending reliability of mmwave radar tracking Reconfigurable Intelligent Surfaces for Exploitation of the Randomness
and detection via fusion with camera,” IEEE Access, vol. 7, pp. of Wireless Environments, p. 22.
137 065–137 079, Sep. 2019. [459] A.-A. A. Boulogeorgos and A. Alexiou, “Performance evaluation of the
[437] C. X. Lu, M. R. U. Saputra, P. Zhao, Y. Almalioglu, P. P. De Gusmao, initial access procedure in wireless THz systems,” in 16th International
C. Chen, K. Sun, N. Trigoni, and A. Markham, “milliEgo: single-chip Symposium on Wireless Communication Systems (ISWCS), Oct. 2019,
mmWave radar aided egomotion estimation via deep sensor fusion,” in pp. 422–426.
Proceedings of the 18th Conference on Embedded Networked Sensor [460] E. N. Papasotiriou, A.-A. A. Boulogeorgos, and A. Alexiou, “Outdoor
Systems, Nov. 2020, pp. 109–122. THz fading modeling by means of Gaussian Mixture Distribution,”
[438] Z. Wang, R. Liu, Q. Liu, J. S. Thompson, and M. Kadoch, “Energy- Research Square, Jul. 2022.
efficient data collection and device positioning in UAV-assisted IoT,” [461] P. Xu, G. Chen, Z. Yang, and M. D. Renzo, “Reconfigurable Intelligent
IEEE Internet Things J., vol. 7, no. 2, pp. 1122–1139, Nov. 2019. Surfaces-Assisted Communications With Discrete Phase Shifts: How
[439] Y. Wang, K. Gu, Y. Wu, W. Dai, and Y. Shen, “NLOS effect mitigation Many Quantization Levels Are Required to Achieve Full Diversity?”
via spatial geometry exploitation in cooperative localization,” IEEE IEEE Wireless Commun. Lett., vol. 10, no. 2, pp. 358–362, Oct. 2021.
Trans. Wireless Commun., vol. 19, no. 9, pp. 6037–6049, Jun. 2020. [462] J. An, C. Xu, L. Gan, and L. Hanzo, “Low-Complexity Channel
[440] C. Gao, G. Zhao, and H. Fourati, Cooperative localization and navi- Estimation and Passive Beamforming for RIS-Assisted MIMO Systems
gation: theory, research, and Practice. CRC Press, Aug. 2019. Relying on Discrete Phase Shifts,” IEEE Trans. Commun., vol. 70,
[441] H. Sarieddeen, M.-S. Alouini, and T. Y. Al-Naffouri, “Terahertz-band no. 2, pp. 1245–1260, Nov. 2022.
ultra-massive spatial modulation MIMO,” IEEE J. Sel. Areas Commun., [463] P. Xu, G. Chen, G. Pan, and M. D. Renzo, “Ergodic Secrecy Rate
vol. 37, no. 9, pp. 2040–2052, Jul. 2019. of RIS-Assisted Communication Systems in the Presence of Discrete
[442] M. H. Loukil, H. Sarieddeen, M.-S. Alouini, and T. Y. Al-Naffouri, Phase Shifts and Multiple Eavesdroppers,” IEEE Wireless Commun.
“Terahertz-band MIMO systems: Adaptive transmission and blind Lett., vol. 10, no. 3, pp. 629–633, Dec. 2021.
parameter estimation,” IEEE Commun. Lett., vol. 25, no. 2, pp. 641– [464] W. Shi, J. Xu, W. Xu, M. D. Renzo, and C. Zhao, “Secure Outage
645, Oct. 2020. Analysis of RIS-Assisted Communications With Discrete Phase Con-
[443] Z. Hossain and J. M. Jornet, “Hierarchical bandwidth modulation for trol,” IEEE Trans. Veh. Technol., pp. 1–6, Nov. 2022.
ultra-broadband terahertz communications,” in ICC 2019-2019 IEEE [465] H. Taghvaee, A. Cabellos-Aparicio, J. Georgiou, and S. Abadal, “Error
international conference on communications (ICC). IEEE, Jul. 2019, Analysis of Programmable Metasurfaces for Beam Steering,” IEEE
pp. 1–7. Journal on Emerging and Selected Topics in Circuits and Systems,
[444] Z. Wei, W. Yuan, S. Li, J. Yuan, G. Bharatula, R. Hadani, and vol. 10, no. 1, pp. 62–74, Jan. 2020.
L. Hanzo, “Orthogonal time-frequency space modulation: A promising [466] F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and
next-generation waveform,” IEEE Wirel. Commun., vol. 28, no. 4, pp. S. Buzzi, “Integrated Sensing and Communications: Toward Dual-
136–144, Aug. 2021. Functional Wireless Networks for 6G and Beyond,” IEEE J. Sel. Areas
[445] A. Magbool, H. Sarieddeen, N. Kouzayha, M.-S. Alouini, and T. Y. Commun., vol. 40, no. 6, pp. 1728–1767, Mar. 2022.
Al-Naffouri, “Terahertz-band non-orthogonal multiple access: System- [467] J. He, F. Jiang, K. Keykhosravi, J. Kokkoniemi, H. Wymeersch, and
and link-level considerations,” IEEE Wirel. Commun., pp. 1–14, May M. Juntti, “Beyond 5G RIS mmWave Systems: Where Communication
2022. and Localization Meet,” IEEE Access, vol. 10, pp. 68 075–68 084, Jun.
[446] A. Hu and J. He, “Position-Aided Beam Learning for Initial Access in 2022.
mmWave MIMO Cellular Networks,” IEEE Syst. J., Oct. 2020. [468] Y. Liu, X. Liu, X. Gao, X. Mu, X. Zhou, O. A. Dobre, and H. V.
[447] C. N. Barati, S. Dutta, S. Rangan, and A. Sabharwal, “Energy and Poor, “Robotic Communications for 5G and Beyond: Challenges and
latency of beamforming architectures for initial access in mmWave Research Opportunities,” IEEE Commun. Mag., vol. 59, no. 10, pp.
wireless networks,” J. Indian Inst. Sci., vol. 100, no. 2, pp. 281–302, 92–98, Oct. 2021.
May 2020. [469] G. Ghatak, R. Koirala, A. De Domenico, B. Denis, D. Dardari, and
[448] K. Chen, C. Qi, and G. Y. Li, “Two-step codeword design for millimeter B. Uguen, “Positioning Data-Rate Trade-Off in mm-Wave Small Cells
wave massive MIMO systems with quantized phase shifters,” IEEE and Service Differentiation for 5G Networks,” in IEEE 87th Vehicular
Trans. Signal Process., vol. 68, pp. 170–180, Jan. 2019. Technology Conference (VTC Spring), Jun. 2018, pp. 1–5.
[449] X. Sun, C. Qi, and G. Y. Li, “Beam training and allocation for [470] S. Jeong, O. Simeone, A. Haimovich, and J. Kang, “Beamforming
multiuser millimeter wave massive MIMO systems,” IEEE Trans. Design for Joint Localization and Data Transmission in Distributed
Wireless Commun., vol. 18, no. 2, pp. 1041–1053, Jan. 2019. Antenna System,” IEEE Trans. Veh. Technol., vol. 64, no. 1, pp. 62–
[450] A. Alkhateeb, Y.-H. Nam, M. S. Rahman, J. Zhang, and R. W. Heath, 76, Feb. 2015.
“Initial beam association in millimeter wave cellular systems: Analysis [471] G. Destino and H. Wymeersch, “On the trade-off between positioning
and design insights,” IEEE Trans. Wireless Commun., vol. 16, no. 5, and data rate for mm-wave communication,” in 2017 IEEE Interna-
pp. 2807–2821, Mar. 2017. tional Conference on Communications Workshops (ICC Workshops),
[451] C. N. Barati, S. A. Hosseini, M. Mezzavilla, T. Korakis, S. S. Panwar, Jul. 2017, pp. 797–802.
S. Rangan, and M. Zorzi, “Initial access in millimeter wave cellular [472] G. Destino, J. Saloranta, H. Wymeersch, and G.-S. Granados, “Impact
systems,” IEEE Trans. Wireless Commun., vol. 15, no. 12, pp. 7926– of imperfect beam alignment on the rate-positioning trade-off,” in 2018
7940, Feb. 2016. IEEE Wireless Communications and Networking Conference (WCNC),
[452] B. Zhou, A. Liu, and V. Lau, “Successive localization and beamforming Apr. 2018, pp. 1–5.
in 5G mmWave MIMO communication systems,” IEEE Trans. Signal [473] A. Dammann, T. Jost, R. Raulefs, M. Walter, and S. Zhang, “Opti-
Process., vol. 67, no. 6, pp. 1620–1635, Jan. 2019. mizing waveforms for positioning in 5G,” in IEEE 17th International
[453] N. Garcia, H. Wymeersch, and D. T. Slock, “Optimal precoders for Workshop on Signal Processing Advances in Wireless Communications
tracking the AoD and AoA of a mmWave path,” IEEE Trans. Signal (SPAWC), Jul. 2016, pp. 1–5.
Process., vol. 66, no. 21, pp. 5718–5729, Sep. 2018. [474] R. Raulefs, A. Dammann, T. Jost, M. Walter, and S. Zhang, “The
[454] E. Björnson, H. Wymeersch, B. Matthiesen, P. Popovski, L. Sanguinetti, 5G localization Waveforms,” in ETSI Workshop on Future Radio
and E. de Carvalho, “Reconfigurable intelligent surfaces: A signal pro- Technologies Focusing on Air Interfaces. ETSI, Jan. 2016, pp. 1–
cessing perspective with wireless applications,” IEEE Signal Process 6.
Mag., vol. 39, no. 2, pp. 135–158, Feb. 2022. [475] S. Gezici, Z. Tian, G. Giannakis, H. Kobayashi, A. Molisch, H. Poor,
[455] M. Gerasimenko, D. Moltchanov, M. Gapeyenko, S. Andreev, and and Z. Sahinoglu, “Localization via ultra-wideband radios: a look at
Y. Koucheryavy, “Capacity of multiconnectivity mmWave systems positioning aspects for future sensor networks,” IEEE Signal Process.
with dynamic blockage and directional antennas,” IEEE Trans. Veh. Mag, vol. 22, no. 4, pp. 70–84, Jul. 2005.
Technol., vol. 68, no. 4, pp. 3534–3549, Jan. 2019.
[456] E. N. Papasotiriou, A.-A. A. Boulogeorgos, K. Haneda, M. F. de Guz-
man, and A. Alexiou, “An experimentally validated fading model for
THz wireless systems,” Sci. Rep., vol. 11, no. 1, Sep. 2021.
[457] A.-A. A. Boulogeorgos, J. M. Riera, and A. Alexiou, “On the Joint
Effect of Rain and Beam Misalignment in Terahertz Wireless Systems,”
IEEE Access, vol. 10, pp. 58 997–59 012, Jun. 2022.
[458] A.-A. A. Boulgoeorgos and A. Alexiou, Statistical Modeling of Relia-
bility Structures and Industrial Processes. CRC Press, Sep. 2022, ch.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 59

Stylianos E. Trevlakis was born in Thessaloniki, Dimitrios Pliatsios received his diploma degree
Greece in 1991. He received the Electrical and Com- from the Department of Electrical and Computer
puter Engineering (ECE) diploma (5 year) from the Engineering, Aristotle University of Thessaloniki,
Aristotle University of Thessaloniki (AUTh) in 2016. Greece in 2016 and his PhD from the Department
Afterwards, Dr. Trevlakis served in the Hellenic of Electrical and Computer Engineering, University
Army in for nine months in the Research Office as of Western Macedonia, Kozani, Greece in 2022.
well as at the Office of Research and Informatics Currently, he works as a postdoctoral researcher
of the School of Management and Officers. During at the ITHACA Lab, Department of Electrical and
2017, he joined the Information Technologies Insti- Computer Engineering, University of Western Mace-
tute, while from October 2017 until April 2022, he donia, in EU-funded research projects and partic-
was part of WCIP as a PhD candidate in AUTh. ipates in drafting research funding proposals. His
During the same period, he was a teaching assistant at the department of research interests include resource allocation in wireless communications and
ECE of AUTh. From April 2022 until now, Dr. Trevlakis is working at edge computing environments, optimization theory, B5G/6G mobile networks,
InnoCube as a postdoctoral researcher with focus on state-of-the-art research and computer and network security. He is a member of the IEEE and the
in conventional & AI-enabled Wireless Communication Systems. Technical Chamber of Greece and he has served as a reviewer in several
Dr. Trevlakis’s research interests lie in the area of Wireless Communica- scientific journals (IEEE Internet of Things Journal, IEEE Communication
tions, with emphasis on conventional & AI-enabled Wireless Communication Letters, Elsevier Computer Networks, IEEE Access, MDPI Sensors) and
Systems, as well as Communications & Signal Processing for Biomedical conferences (IEEE GLOBECOM, IEEE ICC, IEEE NetSoft, IEEE CAMAD,
Engineering. IEEE INFOCOM, IEEE PIMRC). His PhD research was funded by the Greek
State Scholarship Foundation and he has received the 1st Research Excellence
Award by the Research Committee of the University of Western Macedonia.

Alexandros-Apostolos A. Boulogeorgos (Senior

Member, IEEE) was born in Trikala, Greece, in
1988. He received the Diploma degree in Electrical
and Computer Engineering (ECE) and the Ph.D.
degree in wireless communications from the Aris-
totle University of Thessaloniki (AUTh) in 2012
and 2016, respectively. From 2022, he is an As-
sistant Professor at the Department Electrical and
Computer Engineering of the University of Western
Macedonia, Greece. From 2017 to 2022, he served
as a Senior Researcher at the Department of Digital
Systems, University of Piraeus, where he conducted research in the area
of wireless communications. From October 2012 to September 2016, he
was a Teaching Assistant with the Department of ECE, AUTh, and from
February 2017 to September 2022, he served as an Adjunct Professor
with the Department of ECE, University of Western Macedonia, and as an
Visiting Lecturer with the Department of Computer Science and Biomedical
Informatics, University of Thessaly. He has worked in a number of EU and
national projects.
Dr Boulogeorgos has (co-)authored more than 120 technical papers, which
were published in scientific journals and presented at prestigious international
conferences. Furthermore, he is the holder of two (one national and one
European) patents. His current research interests spans in the area of wireless
communications and networks with emphasis in high frequency communi-
cations, intelligent communication systems, optical wireless communications,
and signal processing and communications for biomedical applications.
Dr. Boulogeorgos was awarded the Distinction Scholarship Award from
the Research Committee of AUTh in 2014, and was recognized as an
Exemplary Reviewer for IEEE COMMUNICATION LETTERS in 2016 (top
3% of reviewers). Moreover, he was named a Top Peer Reviewer (top
1% of reviewers) in Cross-Field and Computer Science in the Global Peer
Konstantinos Ntontin (Member, IEEE) is currently
Review Awards 2019, which was presented by the Web of Science and
a research scientist of the SIGCOM Research Group
Publons. Finally, in 2021, he received the best oral presentation award in
at SnT, University of Luxembourg. In the past, he
the International Conference on Modern Circuits and Systems Technologies
held research associate positions at the Electronic
(MOCAST) 2021. He has been involved as a member of organizational and
Engineering and Telecommunications department of
technical program committees in several IEEE and non-IEEE conferences
the University of Barcelona, the Informatics and
and served as a reviewer in various IEEE journals and conferences. He is an
Telecommunications department of the University
IEEE Senior Member and a Member of the Technical Chamber of Greece.
of Athens, and the National Centre of Scientific
He is currently an Editor for IEEE COMMUNICATIONS LETTERS, an
Research-“Demokritos”. In addition, he held an in-
Associate Editor for the Frontier in Communications and Networks, and for
ternship position at Ericsson Eurolab Gmbh, Ger-
the MDPI Telecom. Finally, Dr Boulogeorgos has participated as a guest
many. He received the Diploma in Electrical and
editor in the organization of a number of special issues in IEEE and non-
Computer Engineering in 2006, the M. Sc. Degree in Wireless Systems in
IEEE journals.
2009, and the Ph. D. degree in 2015 from the University of Patras, Greece, the
Royal Institute of Technology (KTH), Sweden, and the Technical University
of Catalonia (UPC), Spain, respectively. His research interests are related to
the physical layer of wireless telecommunications with focus on performance
analysis in fading channels, MIMO systems, array beamforming, transceiver
design, and stochastic modeling of wireless channels.
IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. XX, NO. XX, FIRST QUARTER 2023 60

Panagiotis Sarigiannidis is the Director of Marco Di Renzo (Fellow, IEEE) received the
the ITHACA lab (https://ithaca.ece.uowm.gr/), co- Laurea (cum laude) and Ph.D. degrees in electrical
founder of the 1st spin-off of the University of engineering from the University of L’Aquila, Italy,
Western Macedonia: MetaMind Innovations P.C. in 2003 and 2007, respectively, and the Habilitation
(https://metamind.gr), and Associate Professor in the à Diriger des Recherches (Doctor of Science) de-
Department of Electrical and Computer Engineering gree from University Paris-Sud (now Paris-Saclay
in the University of Western Macedonia, Kozani, University), France, in 2013. Since 2010, he has
Greece. He received the B.Sc. and Ph.D. degrees been with the French National Center for Scientific
in computer science from the Aristotle University Research (CNRS), where he is a CNRS Research
of Thessaloniki, Thessaloniki, Greece, in 2001 and Director (Professor) with the Laboratory of Signals
2007, respectively. He has published over 260 papers and Systems (L2S) of Paris-Saclay University –
in international journals, conferences and book chapters, including IEEE CNRS and CentraleSupelec, Paris, France. In Paris-Saclay University, he
Communications Surveys and Tutorials, IEEE Transactions on Communica- serves as the Coordinator of the Communications and Networks Research
tions, IEEE Internet of Things, IEEE Transactions on Broadcasting, IEEE Area of the Laboratory of Excellence DigiCosme, and as a Member of the
Systems Journal, IEEE Wireless Communications Magazine, IEEE Open Admission and Evaluation Committee of the Ph.D. School on Information and
Journal of the Communications Society, IEEE/OSA Journal of Lightwave Communication Technologies. He is the Editor-in-Chief of IEEE Communica-
Technology, IEEE Transactions on Industrial Informatics, IEEE Access, and tions Letters and a Distinguished Speaker of the IEEE Vehicular Technology
Computer Networks. He received 5 best paper awards. He has been involved Society. In 2017-2020, he was a Distinguished Lecturer of the IEEE Vehicular
in several national, European and international projects. He is currently the Technology Society and IEEE Communications Society. He has received
project coordinator of three H2020 projects, namely a) H2020-DS-SC7- several research distinctions, which include the SEE-IEEE Alain Glavieux
2017 (DS-07-2017), SPEAR: Secure and PrivatE smArt gRid, b) H2020- Award, the IEEE Jack Neubauer Memorial Best Systems Paper Award, the
LC-SC3-EE-2020-1 (LC-SC3-EC-4-2020), EVIDENT: bEhaVioral Insgihts Royal Academy of Engineering Distinguished Visiting Fellowship, the Nokia
anD Effective eNergy policy acTions, and c) H2020-ICT-2020-1 (ICT-56- Foundation Visiting Professorship, the Fulbright Fellowship, and the 2021
2020), TERMINET: nexT gEneRation sMart INterconnectEd ioT, while he EURASIP Journal on Wireless Communications and Networking Best Paper
coordinates the Operational Program MARS: sMart fArming with dRoneS Award. He is a Fellow of the UK Institution of Engineering and Technology
(Competitiveness, Entrepreneurship, and Innovation) and the Erasmus+ KA2 (IET), a Fellow of the Asia-Pacific Artificial Intelligence Association (AAIA),
ARRANGE-ICT: SmartROOT: Smart faRming innOvatiOn Training. He also an Ordinary Member of the European Academy of Sciences and Arts (EASA),
serves as a principal investigator in the H2020-SU-DS-2018 (SU-DS04-2018), and an Ordinary Member of the Academia Europaea (AE). Also, he is a Highly
SDN-microSENSE: SDN-microgrid reSilient Electrical eNergy SystEm and Cited Researcher.
in three Erasmus+ KA2: a) ARRANGE-ICT: pArtneRship foR AddressiNG
mEgatrends in ICT, b) JAUNTY: Joint undergAduate coUrses for smart
eNergy managemenT sYstems, and c) STRONG: advanced firST RespONders
traininG (Cooperation for Innovation and the Exchange of Good Practices).
His research interests include telecommunication networks, internet of things
and network security. He is an IEEE member and participates in the Editorial
Boards of various journals.

Symeon Chatzinotas (Fellow, IEEE) is currently

Full Professor / Chief Scientist I and Co-Head of
the SIGCOM Research Group at SnT, University of
Luxembourg. In the past, he has been a Visiting
Professor at the University of Parma, Italy and he
was involved in numerous Research and Develop-
ment projects for the National Center for Scientific
Research Demokritos, the Center of Research and
Technology Hellas and the Center of Communica-
tion Systems Research, University of Surrey. He
received the M.Eng. degree in telecommunications
from the Aristotle University of Thessaloniki, Thessaloniki, Greece, in 2003,
and the M.Sc. and Ph.D. degrees in electronic engineering from the University
of Surrey, Surrey, U.K., in 2006 and 2009, respectively. He was a co-recipient
of the 2014 IEEE Distinguished Contributions to Satellite Communications
Award, the CROWNCOM 2015 Best Paper Award and the 2018 EURASIP
JWCN Best Paper Award. He has (co-)authored more than 450 technical
papers in refereed international journals, conferences and scientific books.
He is currently in the editorial board of the IEEE Open Journal of Vehicular
Technology and the International Journal of Satellite Communications and
Networking.