SPECIAL SECTION ON ADVANCES ON HIGH PERFORMANCE

WIRELESS NETWORKS FOR AUTOMATION AND IIOT

Received 9 March 2023, accepted 28 March 2023, date of publication 31 March 2023, date of current version 5 April 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3263583

A WiFi Indoor Location Tracking Algorithm

Based on Improved Weighted K Nearest
Neighbors and Kalman Filter
JIUSONG HU AND CONGWEI HU
College of Railway Transportation, Hunan University of Technology, Zhuzhou 412007, China
Corresponding author: Jiusong Hu ([email protected])
This work was supported in part by the Natural Science Youth Foundation of Hunan Province under Grant 2020JJ5144, and in part by the
Scientific Research Project of Hunan Education Department under Grant 19C0563.

ABSTRACT The weighted K -nearest neighbors (WKNN) algorithm is a widely adopted lightweight
methodology for indoor WiFi positioning based on location fingerprinting. Nonetheless, it suffers from the
disadvantage of a fixed K value and susceptibility to incorrect reference point matching. To address this issue,
we present a novel algorithm in this paper, referred to as static continuous statistical characteristics-soft range
limited-self-adaptive WKNN (SCSC-SRL-SAWKNN). Our algorithm not only takes into account location
tracking in the motion state but also exploits the continuous statistical features of extended periods of inac-
tivity to enhance localization. In the motion state, we initially employ the self-adaptive WKNN (SAWKNN)
algorithm to determine the optimal K value, followed by the employment of the soft range limited KNN
(SRL-KNN) algorithm to identify the correct reference point and ultimately estimate the position. When a
prolonged stationary state is detected, we first utilize the moving window method to obtain a more stable
position fingerprint, and then proceed with the positioning process in the same motion state. Ultimately,
we use Kalman filter to generate the location trajectory. Our experimental findings demonstrate that the
proposed SCSC-SRL-SAWKNN algorithm outperforms traditional WKNN, SAWKNN, and SRL-KNN
techniques in terms of localization accuracy and location trajectory. Specifically, the localization accuracy
of our algorithm is 56.7% and 36.6% higher than that of traditional WKNN in the static state and overall
situation, respectively.

INDEX TERMS WiFi, indoor location, WKNN, fingerprint.

I. INTRODUCTION Beidou navigation and positioning systems. Conversely, GPS

With the widespread usage and advancement of mobile has significant limitations in indoor environments, where
devices, the demand for location-based services (LBS) has satellite signals are often obstructed by building walls. Conse-
skyrocketed. Applications related to LBS have now become quently, indoor positioning techniques have been extensively
an essential element of both industrial production and every- researched, including Ultra Wide Band (UWB) [1], Blue-
day life. In outdoor environments, GPS and Beidou naviga- tooth [2], WiFi [3], electromyography (EMG) signals [4],
tion and positioning systems have reached a state of matu- and others [5]. These positioning technologies each possess
rity and perfection, providing highly accurate results at a their own unique characteristics, with WiFi indoor position-
lower cost. Notable examples include commonly used map ing technology being one of the most widely studied in
navigation, real-time public transport location broadcasting, recent years. Its advantages include the extensive deployment
and shared vehicle rental and return points. All of these of wireless networks in indoor environments and the fact
applications are developed and implemented based on GPS or that most mobile devices are equipped with built-in WiFi
modules, thus obviating the need for additional hardware.
The associate editor coordinating the review of this manuscript and WiFi indoor positioning is the most universal and inexpensive
approving it for publication was Stefano Scanzio . option when compared to other positioning technologies.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 32907
 J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

Terminal equipment can be located using indoor wire- A. THE LONG-TERM STATIC CASE IS NOT SPECIFICALLY
less network signals. As an illustration, the location CONSIDERED
fingerprint-based method can be considered. Indoor wireless In practical situations, the existence of multipath effects and
networks typically consist of several access points (APs), channel fading leads to variations in RSSIs collected by
each emitting a wireless signal covering a particular area. The terminal devices, even during successive time intervals while
transmission power of an access point (AP) is theoretically situated in the same location. Regrettably, the long-term static
fixed, so the signal strength in a particular area is theoreti- state that is vital for fingerprint stability is often ignored in
cally consistent. The mobile device can determine its loca- the present-day WiFi-based indoor positioning methodology.
tion by evaluating the signal strength received from the AP. Therefore, improving the stability of the fingerprint in the
This fingerprint-based localization method exhibits great static state is often overlooked.
potential [6]. Euclidean distance [7], [8] and Manhattan
distance [9], [10] are commonly employed in localization B. SETTING OF K VALUE: DIFFERENT K VALUE MEANS
methods based on location fingerprints. Time of arrival- MATCHING DIFFERENT NUMBERS OF PRS
based methods [11] and angle of arrival-based methods [12] This will result in different positioning accuracy. Further-
are also available for WiFi indoor positioning, but these more, the optimal value of K for achieving the highest degree
require additional hardware, which is absent in most mobile of accuracy varies according to differences in the indoor and
terminals, and as a result, have not gained significant network environmental conditions.
attention.
Enhancing the accuracy of WiFi positioning, which C. PHYSICAL DISTANCE DOES NOT MATCH FINGERPRINT
remains relatively low, poses the greatest challenge. The SIMILARITY
location-based fingerprint method in WiFi positioning is Some location fingerprints that are physically distant may
categorized into three groups: deterministic methods [13], bear resemblance to those which are physically closer. The
probabilistic methods [14], and deep learning methods [15]. WKNN algorithm may therefore be misled, leading to a
Despite their unparalleled accuracy, deep learning methods relatively significant error.
are inherently complex, requiring significant computational The challenge of setting the K value has been addressed
resources. Conversely, deterministic methods are comparable by the proposed self-adaptive weight K -nearest neighbors
to probabilistic methods in terms of performance but offer (SAWKNN) algorithm [17]. The issue of physical dis-
the added benefit of reduced complexity, rendering them a tance not matching fingerprint similarity is tackled by
favored choice among researchers. the soft range limited K -nearest neighbors (SRL-KNN)
Deterministic methods achieve localization by calculating algorithm [18]. To overcome these challenges, this paper
the similarity of fingerprints between the online and offline proposes a localization algorithm based on the static
phases. In the offline phase, the task involves constructing a continuous statistical characteristics of fingerprint signals.
fingerprint database comprising the received signal strength The proposed algorithm, named static continuous statisti-
index (RSSI) gathered by the mobile terminal at each ref- cal characteristics-soft range limited-self-adaptive WKNN
erence points (RPs). During the online phase, the collected (SCSC-SRL-SAWKNN), retains the recent location infor-
location fingerprints are matched with those in the offline mation of the user and combines SAWKNN and SRL-KNN
phase. The positioning algorithm is then utilized to determine algorithms to achieve the best positioning accuracy in differ-
the positioning. However, in the actual environment, RSSI ent localization states. This algorithm sets itself apart from
is influenced by various factors, such as object movement, its counterparts through its distinctive approach. It discerns
changes in electromagnetic waves, variations in antenna between the stationary and moving states in the user position-
directivity, and radio frequency interference. Therefore, it is ing process, and handles them separately. In the static state,
highly probable that the offline database may not contain the algorithm employs the concept of a moving window to
the location fingerprint corresponding to the online stage. first process the fingerprint signal. Then, the cumulative mean
In this scenario, the RPs with the highest location fingerprint of the signal is used to enhance the stability of RSSI. Finally,
similarity between the online phase and the offline database in conjunction with SAWKNN, the positioning is completed.
is used as a candidate point. The mobile terminal is subse- On the other hand, in the moving state, the soft range limit
quently positioned at the center of the K candidate points. is used to resolve the issue of physical distance not matching
This approach is known as the K -nearest neighbors (KNN) fingerprint similarity. The SAWKNN algorithm is combined
positioning method. On this basis, weights corresponding with the soft range limit to complete the positioning. The pro-
to the candidate points are added, which further improves posed SCSC-SRL-SAWKNN algorithm in this paper retains
the positioning accuracy [16]. This is a prevalent algorithm the qualities of different localization states, and processes
known as the weighted K -nearest neighbors (WKNN). them separately based on their respective characteristics. The
The WKNN algorithm has been the subject of numerous algorithm is capable of achieving optimal positioning accu-
discussions in the academic literature. However, it continues racy in different positioning states, and it doesn’t require
to confront a series of daunting challenges, including the prior knowledge. As a result, there is no need to deploy any
following: additional hardware. Numerous experimental results attest

32908 VOLUME 11, 2023

J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

that the accuracy of the proposed algorithm surpasses that physical distance and fingerprint similarity is ignored, which
of traditional WKNN and the optimized SAWKNN and can result in large positioning errors if the wrong candidate
SRL-KNN algorithms under both static and dynamic con- RPs is chosen.
ditions. As compared to the traditional WKNN, the overall Additionally, a method exists [30] that records the user’s
positioning accuracy has been improved by 36.6%. Moreover, latest location information to determine their speed and direc-
being a lightweight algorithm, it can enable real-time posi- tion, enabling prediction of their next location. However,
tioning while consuming minimal system resources. if the user is moving slowly or in an irregular pattern, the pre-
The remainder of this paper shall be arranged in the fol- diction error may be significant. Studies have also explored
lowing manner: Section II will introduce related research. the mismatch between distance and fingerprint similarity,
Section III shall expound upon the indoor positioning sys- wherein a location fingerprint at a greater physical distance
tem and discuss pertinent optimization issues. In Section IV, may exhibit significant similarity to the fingerprint of the
the SCSC-SRL-SAWKNN algorithm shall be presented. current location, leading to a large positioning error for that
Section V will furnish a comprehensive assessment via an location. The SRL-KNN method proposed in [18] centers the
ample array of real-world experiments. Finally, in Section VI, circle on the user’s previous position and sets the radius as a
we will conclude the paper and outline potential areas of range factor based on the actual movement situation to limit
future research. the range. This solution addresses the issue of physical dis-
tance not matching fingerprint similarity during positioning,
II. RELATE WORK although it does not consider the static positioning scenario.
Over the past decade, there has been a staggering surge in Additionally, the signal trend index (STI) has been used in
the proliferation of mobile terminals. Consequently, various some studies to compare the vector shape of the RSSI of the
indoor positioning systems have been extensively studied. mobile terminal device during the online phase with that of
These systems can broadly be classified into two categories the offline database. The signal trend exponential weighted
based on hardware deployment. The first category includes K -nearest neighbor (STI-WKNN) algorithm [31] has been
infrared [19], ultrasonic [20], Bluetooth [2], Ultra Wide Band proposed as a solution with better performance targeting
(UWB) [21], Radio Frequency Identification (RFID) [22], heterogeneous devices.
which can provide high positioning accuracy but require In recent years, the field of deep learning has made
additional specialized hardware deployment. The second cat- remarkable progress, leading to the emergence of a plethora
egory includes positioning technologies such as WiFi [23] of indoor localization techniques employing deep learning
and geomagnetic field [24], where users can rely on existing methods [15], [32], [33], [34]. Indoor localization can be
wireless networks to enable location functionality with the accomplished by means of probability estimation, which
added advantage of lower costs. The research presented in sets it apart from deterministic techniques such as WKNN.
this paper pertains to the second category of methods. Among Recent breakthroughs [35], [36], [37] have yielded compara-
these, fingerprint-based localization methods are the most ble localization accuracy as deterministic methods but at the
common. expense of higher computational complexity. Deep learning
In the realm of WiFi positioning, distance-based methods technology, while undoubtedly possessing exceptional accu-
have also been utilized [25], [26]. However, deploying addi- racy, is unfortunately accompanied by the highest degree of
tional hardware and prior knowledge of the location of each complexity, leading to a significant requirement for hardware
AP are required for these methods. As such, their practical resources that pose a challenge to its widespread adoption for
application is limited and they are often not prioritized in most real-time positioning.
situations.
The method most commonly used for location-based fin- III. INDOOR POSITIONING SYSTEM
gerprinting is the KNN algorithm proposed by RANDA [27]. In this section, we shall undertake a systematic analysis of the
An optimization algorithm based on KNN was proposed three primary issues with the WKNN algorithm. These issues
in [28]. In addition, the weighted KNN (WKNN) algorithm include a lack of specific considerations for static situations,
was proposed in [29], which slightly improves the position- problems with setting the K value, and a mismatch between
ing accuracy compared to the KNN algorithm through the physical distance and fingerprint similarity. Firstly, we shall
addition of weights. However, determining the optimal K provide an overview of the system configuration, followed
value has always been a challenging problem, and a fixed by an explanation of how each of these three problems can
experience value is typically used. The optimal K value significantly affect the accuracy of indoor positioning.
changes with the environment, and thus cannot be obtained
using traditional WKNN. To address this issue and ensure the A. SYSTEM CONFIGURATION
best K value is obtained for each positioning, the SAWKNN Fig. 1 depicts the comprehensive block diagram of the WiFi
algorithm was proposed in [17]. It enables adaptive adjust- indoor positioning system. The system is comprised of two
ment of the K value through the fingerprint similarity thresh- stages: the offline stage and the online stage. The prime objec-
old, resulting in improved localization accuracy compared tive of the offline stage is to accomplish the establishment of
to WKNN. However, the problem of mismatch between the offline location fingerprint database. The online stage is

VOLUME 11, 2023 32909

 J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

We performed 100 consecutive signal acquisitions on each

RPs at a rate of 1 per second, and experimental data showed
that each RPs can receive WiFi signals from 27 to 61 APs.

B. OBSERVATION
We carry out three sets of experiments using the collected
fingerprint data with the aim of demonstrating three issues
concerning the conventional WKNN positioning algorithm.

1) THE IMPORTANCE OF LONG-TERM STATIC STATES IS

OFTEN OVERLOOKED
The RSSI collected by the mobile terminal under static condi-
tions can be easily affected by various environmental factors,
such as the movement of people or objects, changes in elec-
FIGURE 1. Block diagram of WiFi indoor positioning system. tromagnetic waves, and variations in antenna directionality.
These factors can cause fluctuations in the collected RSSI,
responsible for utilizing the proposed positioning algorithm making it challenging to accurately match location finger-
to match the location fingerprints of the online phase with prints between the online and offline stages.
those of the offline phase to achieve accurate positioning. We consecutively acquired signals 100 times at a specific
Lastly, the Kalman filter is employed to accomplish position- test points (TPs). From dozens of available APs, we randomly
ing tracking. selected the RSSI values of 5 APs. The RSSI values of these
We have constructed an indoor positioning system in China 5 selected APs are presented in Fig. 3.
Power Valley located in Zhuzhou City, which is illustrated
in Fig. 2. In the building, there is no centralized deployment
of WiFi devices. Instead, these WiFi networks are distributed
across various locations in each room. During our field tests,
we identified a total of 339 APs in the building, with a grid
spacing of 1 meter.

FIGURE 3. The RSS values over time using same devices from the diverse
AP and at the fixed location.

The Fig. 3 reveals that there are significant fluctuations in

the RSSI collected by the same device even when it is in a
static state. Directly using an RSSI with significant fluctua-
FIGURE 2. Layout of the indoor positioning system. Each blue dot
represents a reference point. tions as the position fingerprint for positioning will result in
poor accuracy. However, if the static conditions during the
We conducted our indoor positioning system experiment positioning process are taken into account, the stability of the
using an ASUS-FX50J mobile terminal device equipped position fingerprint can be considerably enhanced, leading to
with a Realtek 8811CU Wireless LAN 802.11ac USB NIC, improved positioning accuracy.
as illustrated in Fig. 2. The experiment was conducted in a
building containing 10 rooms and a corridor, where WiFi 2) SETTING OF K VALUE
networks were distributed throughout the building with no We carried out a comprehensive set of 860 online localization
centralized deployment of APs. A total of 339 APs were experiments utilizing the WKNN algorithm, wherein each
found in the building with a grid spacing of 1 meter. In order test point (TP) was scanned at least once. The value of K was
to facilitate data statistics, the corridor was divided into two varied from 1 to 20. The resulting positioning errors for the
parts, 101 and 102, starting from 26 meters. A total of 155 RPs WKNN algorithm, at varying K values, have been graphically
fingerprints were collected and marked with blue stars, with represented in Fig. 4.
adjacent RPs separated by 1 meter. Due to the presence of It is evident that the average error is significantly influ-
walls and desks, the distribution of RPs was non-uniform enced by the variation of the K value. Within the range
in the rooms, with each room containing 10 RPs except for of K between 5 and 10, the average error remains rela-
101 and 102, which contained 29 and 36 RPs, respectively. tively stable. However, the average error steadily increases

32910 VOLUME 11, 2023

J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

TP lies between RP1 and RP2, but the position fingerprint of

TP matches the position fingerprint of RP2 and RP3 most.
Consequently, the position of TP is erroneously estimated at
point B. This demonstrates that the accuracy of localization
is heavily influenced by the disparity between physical dis-
tance and fingerprint resemblance. Hence, a certain level of
physical constraints is necessary.
Next we will focus on how to solve these three problems.

IV. SCSC-SRL-SAWKNN ALGORITHM

FIGURE 4. Position result of traditional WKNN algorithm with different
K value. The error is measured in box charts.Each box chart contains six In this section, we present the SCSC-SRL-SAWKNN algo-
values from top to bottom: the upper edge (maximum value), the upper rithm for indoor localization. Initially, we will provide a brief
quartile (75%), the average (small circle), the median (red line), the lower
quartile (25%), and the lower edge (minimum value).
overview of the relevant knowledge of WKNN. Subsequently,
we will elaborate on how to leverage the static properties
in the localization process. Next, the analysis focuses on
as K values ascend from 11 to 20. In tandem, the upper addressing the challenges related to setting the K value and
limit of the error box also increases, indicating a surge in resolving the mismatch between physical distance and finger-
the number of larger errors. The K value of 7 is optimal for print similarity. Additionally, we will review the SAWKNN
minimizing the average error, and its values fluctuate within a and SRL-KNN algorithms separately. Ultimately, we will
narrow error box (representing 75% of the sample). The error amalgamate the aforementioned three characteristics and
box for K values 1-4 is larger compared to K value 7. Here, present the SCSC-SRL-SAWKNN algorithm.
the upper quantile of the box is rising, while the lower quan-
tile is decreasing, suggesting that the overall positioning error A. WKNN BASICS
fluctuates greatly. Although many situations result in large In this section, we first provide an overview of the relevant
positioning errors, WKNN exhibits more precise positioning concepts related to WKNN. Let rssi,j denote the average
accuracy than when K is set to 7 at certain times. number of RSSIs collected from the ith AP at the jth reference
point (RP), where the total number of APs is denoted by N.
3) PHYSICAL DISTANCE DOES NOT CORRESPOND TO In situations where a mobile terminal can detect multiple APs
FINGERPRINT SIMILARITY at one RP, the fingerprint vector collected by the jth RP can
Due to the occurrence of channel fading and the multipath be represented by r⃗ssj :
effect in the signal propagation process, in addition to the
r⃗ssj = rss1,j , rss2,j , . . . , rssi,j , . . . , rssN ,j .
 
(1)
impact of people’s movements, the opening and closing of
doors and windows, and changes in the location of objects During the online testing phase, the RSSI fingerprint
in the actual environment. The phenomenon in question may vector r⃗ssr received at the rth TP can be expressed as:
result in the fingerprint at a distant physical location appear-
r⃗ssr = rss1,r , rss2,r , . . . , rssi,r , . . . , rssN ,r .
 
(2)
ing similar to that of the current location at a particular
instance. This could lead to erroneous outcomes in WKNN, The location fingerprint similarity difference between the
consequently amplifying the positioning error. The situation jth RP and rth TP can be defined as:
is aptly demonstrated in Fig. 5. 2
r⃗ssr · PT − r⃗ssj · PT
difr,j = (3)
M
where || || represents the Manhattan distance. Equation (3)
defines the Manhattan distance Drj between the jth PR and
the rth TP. M in Equation (3) represents the number of APs
that can be collected in the online phase, and P is represented
by:
 
p1
 p2 
 .. 
 
 . 
P=   pi  . (4)

 .. 
 
FIGURE 5. The problem of mismatch between physical distance and  . 
fingerprint similarity.
pM
In Fig. 5, point A represents the location of TP, while The vector pi is a row vector of dimension [1 ∗ N ] influ-
RP1, RP2, and RP3 are three reference points respectively. enced by N APs. In our approach, pi is defined as a sparse

VOLUME 11, 2023 32911

 J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

vector where only one element is 1 and all other elements As illustrated in Fig. 6, the signal processing method based
are set to 0. This indicates that only one corresponding AP on moving window in static state is divided into two steps.
is chosen in pi . Due to the variability in the number of APs In the first step, during the k + L − 1th signal acquisition
that can be detected at different RPs, it is advisable to set the in the online stage, the position fingerprint of the ith AP
K value to be lower than the total number of available APs. collected between the kth time and the k + L − 1th time is
After computing the difr,j for the rth TP and all RPs using used as the old window. The earliest received test signal in
formula (3), WKNN selects the K RPs with the smallest this window r⃗ssi,k is discarded. In the second step, during the
difference in fingerprint similarity for position estimation: k + L th signal acquisition in the online phase, we process the
PK newly obtained r⃗ssi,k+L signal through equation (7) to obtain
ωl LOC(xl , yl ) ⃗ i,k+L . Then add it to the excessive state and get the new
Rss
LOC ′ (xr , yr ) = l=1 PK (5)
l=1 ωl window. The Rss⃗ i,k+L is the average of the sum of r⃗ss of the
ith AP from r⃗ssi,k to r⃗ssi,k+L :
In the formula, LOC ′ (xr , yr ) represents the estimated posi- Pk+L
tion after performing the positioning algorithm. The position rss
⃗ i,k+L = P n=k+1 i,n
Rss (7)
of the lth RP among the K candidate points is denoted by k+L
n=k+1 Num(n)i
LOC(xl , yl ). The weight of the lth position is represented by
ωth , which is determined by 1/difr,j . A weighted strategy Next, we proceed with a further update of Rss ⃗ i,k+L using
was suggested in [29], which yielded a slight improvement formula (8) to obtain a more refined and smoother signal. The
in accuracy compared to the KNN algorithm. formula is defined as follows:
( k+L ⃗
P
⃗ i,k+L = n=k+1 rssi,n ) + Rssi,k+L
B. STATIC SITUATION OF ONLINE POSITIONING RSS (8)
L
In this section, we present a method to enhance the stabil-
ity of WiFi location fingerprints in stationary states during ⃗ i,k+L is the RSSI information of ith access point
The RSS
the online localization phase. First, we distinguish between after the k + L th signal acquisition is refined using the mov-
stationary and moving states in the positioning process and ing window. This refined information is then used to replace
process them separately. Then, we apply specific techniques the original signal r⃗ssi,k+L , signifying the completion of the
to handle the RSSI values in the stationary state. signal processing task. To further solidify the stability of
In the second section, we analyzed the fluctuations in RSSI, the cumulative average method is employed based on
signal strength when the device is in a stationary state. the moving window signal processing approach. The steps
To improve the stability of the RSSI in static conditions, involved in this method can be summarized as follows:
( ns=1 r⃗ssi,s )
P
we first determine the state of static based on the condition 1
⃗ i,k+L
RSS = (9)
that the number of consecutive single-point acquisitions is cnt
greater than L. Then we analyze the statistical properties of The r⃗ssi,s represents the location fingerprint of the ith AP’s
the RSSI in the static state to improve positioning accuracy RSSI that has undergone static processing and was collected
by addressing the issue of signal fluctuation and obtaining a at the sth online stage. The variable cnt represents the number
stable fingerprint signal. To achieve this, we preprocess the of signal acquisitions in the current static state. In this par-
RSSI using a moving window approach when the condition ticular scenario, we focus on the AP corresponding to Fig. 3.
in equation (6) is met. To obtain the RSSI data, 100 static signal acquisitions were
Pk+L−1
Num(n)i performed, and the parameter L was set to 10. The resulting
n=k
>σ (6) RSSI data after undergoing SCSC processing is depicted
L
in Fig. 7.
In this case, Num(n)i denotes the validity of the ith AP
signal when the nth signal is collected. Any signal that is
recorded as -100 dB during the signal collection process is
considered invalid. We define σ be the threshold for signal
validity, where signals exceeding the threshold are considered
valid.

FIGURE 7. RSSI processed by SCSC in static state.

Upon comparing Fig. 3 and Fig. 7, it is evident that the

FIGURE 6. Signal preprocessing based on moving window in static state. stability of RSSI has been substantially enhanced through

32912 VOLUME 11, 2023

J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

the implementation of SCSC processing. As the number of The similarity between two pairs of dif is represented by γ .
acquisitions increases, the stability of the location fingerprint The threshold of fingerprint similarity, γt h, can be determined
will also improve. Thus, by strategically utilizing the static through experimentation. Using this threshold, we can filter
conditions in the positioning process, we can obtain more sta- out the number of RPs that meet the required conditions,
bility location fingerprints, leading to enhanced positioning thus establishing the value of K . The positioning can then be
accuracy. achieved using formula (5). In Algorithm 1, the maximum K
value that can be obtained through experimentation is denoted
C. K VALUE SETTING AND PHYSICAL DISTANCE DO NOT by Kmax .
CORRESPOND TO FINGERPRINT SIMILARITY
In this subsection, we will briefly discuss the solutions to Algorithm 1 The Pseudo-Code for Obtaining the K
two problems: setting the K value and addressing the mis- Value
match between physical distance and location fingerprint 1 for j = 2 : Kmax + 1 do
similarity. The setting of the K value is described in detail 2 if difr,other difr,j1 − 1 > γth then

in SAWKNN [17]. Interested readers can refer to the original 3 break;
source for more information. Here is a brief overview of how 4 end
the K value is dynamically obtained. Firstly, the classical path 5 end
loss model for radio signals is used: 6 K = j − 1;
d 7 return K ;
P(d)[dBm] = P(d0 )[dBm] − 10n log[ ]+ω (10)
d0
The path loss coefficient is denoted by n, and the RSSI Clearly, the K value is subject to dynamic changes based
measured at distances d and d0 from the AP are represented on the fingerprint similarity between the online and offline
by P(d) and P(d0 ), respectively. Here, d0 serves as the refer- stages. Therefore, the K value can be adjusted adaptively and
ence distance and is based on 1 m. In addition, ω represents utilized flexibly within the WKNN algorithm. This leads to
the ambient noise. Typically, the value of P(d) in the same the automatic determination of the optimal K value.
area is largely dependent on the distance d. The location Aiming to address the issue of discrepancy between phys-
fingerprint similarity difference dif(r,j) between the RP and ical distance and fingerprint similarity in the positioning
the TP in formula (3) is likely to be consistent with dr,j to a process, we have referenced the soft range limiting concept
large extent. of SRL-KNN [18]. Here, we provide a brief overview. The
1) IF difA,O < difD,O THEN dAO < dDO , fundamental concept is that there is a limit to the velocity at
2) IF difA,O ≈ difD,O THEN dAO ≈ dDO , which a person or object can move during localization. It is
3) IF difA1 ,O ≈ difA2 ,O ≈ difA3 ,O · · · < difD,O THEN implausible for objects to suddenly emerge beyond a particu-
dA1 O ≈ dA2 O ≈ dA3 O · · · < dDO . lar range in the ongoing online positioning process. A circle
According to the three properties mentioned above, the is drawn with the location of the previous moment as the
RPs with the smallest difr,j between the location fingerprint center and d⃗ as the radius. The positioning space is restricted
of the offline stage and the location fingerprint of the online within this circle. d⃗ is determined by the individual’s or
stage is more likely to be close to the TP. Therefore, the object’s movement velocity and positioning frequency during
positioning problem can be transformed into finding the RPs localization. With soft range restriction, the position near the
fingerprints that are most similar to the TP fingerprints. previous moment is more likely to be chosen as the candidate
WKNN selects K RPs with the highest similarity to the TP position for localization results. However, unlike [18], where
fingerprints as candidate points. However, the number of the Euclidean distance expression is transformed, we trans-
RPs with high similarity location fingerprints varies under form the expression for the Manhattan distance.
different TPs, especially in cases where K is 1 or 2, which j
Wr + Dr
j
are more easily ignored. Hence, the approach of dynami- ⃗ jr = P
D (12)
s j
cally acquiring the optimal K value can be implemented to i=1 Wr
enhance the accuracy of the positioning system. The issue of (xr − xpre )2 + (yr − ypre )2
Wrj = exp (13)
setting the K value can be transformed into determining the 4d⃗ 2
number of RPS fingerprints that exhibit high similarity to TP j
fingerprints. Suppose that difr,j1 represents the smallest dif , The function Wr represents the penalty imposed on posi-
then the correlation between difr,j1 and the dissimilarity in tion r. The variable s denotes the number of RPs in the offline
fingerprint similarity values of other locations is established database, and (xpre , ypre ) represents the user’s previous posi-
as follows: tion. The maximum distance the user can move between two
consecutive positioning events is defined as d. ⃗ Typically, the
 γ = difr,other difr,j1 − 1
 
walking speed of a person is within 2 m/s, and the maximum
difr,j1 < difr,other , γ > γth (11) movement speed can be set according to the actual situation.
difr,j1 ≈ difr,other , γ ⩽ γth

It is unnecessary to know the specific movement speed and

VOLUME 11, 2023 32913

 J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

direction of the user. With this approach, the problem of mis- algorithms in different states. Finally, we will delve into the
match between physical distance and fingerprint similarity application development of location tracking. For the tradi-
can be effectively addressed. tional WKNN and SRL-KNN algorithms, we have set K to
an empirical value of 3. The computer utilized to provide
D. PROPOSED SCSC-SRL-SAWKNN METHOD location services was equipped with an Intel(R) Core(TM)
In this subsection, we shall consolidate the information from i5-4200H CPU and 12 GB of RAM.
sections IV-A, IV-B, and IV-C to provide a comprehensive
overview of the SCSC-SRL-SAWKNN algorithm. A. THE EFFECT OF MOVING WINDOW SIGNAL
PROCESSING
Firstly, we will analyze the impact of key parameters under
static conditions on the performance of the moving window
processing method. As discussed in section IV, the crucial
parameters of this method are denoted by L and σ . Our aim
is to identify the optimal parameter values that would result
in the highest positioning accuracy. Subsequently, we will
evaluate the impact of the moving window signal processing
method on the localization accuracy in the real environment
using these values.
Fig. 2 depicts the real experimental environment that has
been constructed, wherein the blue pentagrams denote RPs
with a total of 155 RPs. The trajectory and direction of
motion in the online phase are represented by the green lines,
FIGURE 8. Outline of the SCSC-SRL-SAWKNN method. with a movement speed of approximately 0.5 m/s. The signal
acquisition frequency has been set to 1 Hz. At each static
Fig. 8 illustrates the flow of the SCSC-SRL-SAWKNN continuous collection point represented by the red pentagram,
algorithm. The localization problem is to find offline fin- we collected 100 fingerprint signals continuously, with an
gerprints that closely match the fingerprints collected during acquisition frequency of 1 Hz. The entire experiment was
the online phase. When a user requests their location, the conducted during a bustling time period with many teachers
mobile terminal collects location fingerprints and processes and students present in the experimental environment. Addi-
different motion states separately. In the case of a static state, tionally, the state of indoor doors and windows opening and
a moving window signal processing method is applied to closing was continuously changing. In order to maintain the
improve the stability of the location fingerprint in the online authenticity of the experiment, all algorithms have utilized
phase, followed by the method of mean accumulation to the real-time collected RSSI as the original fingerprint.
obtain robust online fingerprints. The calculation of the loca-
tion fingerprint similarity difference between the online and 1) SETTING OF L
offline stages is then completed, and an adaptive algorithm Fig. 9 depicts the error pattern of static continuous statistical
is used to calculate the K value and estimate the position. characteristics weight K -nearest neighbor (SCSC-WKNN)
If the user is in a moving state, the signal processing steps are algorithm in a static state when employing different values of
omitted, and the fingerprint similarity difference is directly parameter L. It is discernible that the error curve of the posi-
computed to obtain the K value through Algorithm 1. The tioning accuracy fluctuates considerably when L is within
Manhattan distance in the fingerprint similarity expression the range of [2, 30]. Nonetheless, it is evident that the best
is then transformed using the penalty function of the range positioning accuracy is attained when L is equal to 4. As such,
factor, and the final modified K RPs is used for position the value of L is set to 4.
estimation.

V. EXPERIMENTS
In this section, we will assess the performance of the
proposed SCSC-SRL-SAWKNN algorithm through actual
experiments. Firstly, we will analyze the influence of several
crucial parameters, including L, σ , γrh and Kmax , under vari-
ous scenarios. Secondly, we will compare the performance of
the WKNN algorithm before and after applying the moving
window signal processing method to stabilize the position
fingerprints during the online positioning process in static
states. Then, we will compare the localization accuracies of
WKNN, SAWKNN, SRL-KNN, and SCSC-SRL-SAWKNN FIGURE 9. Position results of SCSC-WKNN under different L.

32914 VOLUME 11, 2023

J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

2) SETTING OF σ σ yielding the utmost localization accuracy were determined

Fig. 10 illustrates the average localization error of the in section V-A. Our objective is to determine the most favor-
SCSC-WKNN algorithm at varying values of σ . It can be able values of γrh and Kmax .
observed that the positioning errors are relatively small and
closely clustered when σ falls within the range of [0.1, 0.7]. 1) SETTING OF γrh
To maximize the preservation of fingerprint data, the value The localization errors of SCSC-SRL-SAWKNN for various
of σ is set to 0.2. γrh values are presented in Fig. 12. It is evident that the most
accurate positioning results can be achieved when the γrh
value falls between the range of 0.4 and 1. This is because
the formula (11) is satisfied by a greater number of location
fingerprints as the value of γrh increases, leading to increased
complexity of the algorithm. Therefore, a value of 0.4 is
ideal for achieving the highest computational efficiency while
maintaining the best positioning accuracy.

FIGURE 10. Position results of SCSC-WKNN under different σ .

3) STATIC LOCATION TEST

The positioning accuracy of the WKNN algorithm before
and after the moving window processing in the static state
is depicted in Fig. 11. The figure illustrates the green dots
which represent the average error and the red line indicat-
ing the median error. Under static conditions, the traditional FIGURE 12. Position results of SCSC-SRL-SAWKNN under different γrh .
WKNN and SCSC-WKNN algorithms have average posi-
tioning errors of 2.14 and 0.95 meters, respectively. This 2) SETTING OF Kmax
suggests that the localization accuracy of WKNN is signif- The mean localization error of SCSC-SRL-SAWKNN at dif-
icantly improved after the moving window signal processing, ferent Kmax values is illustrated in Fig. 13. It is observed that
demonstrating an excellent positioning effect. Furthermore, the positioning accuracy diminishes as Kmax increases. The
the localization accuracy of SCSC-WKNN is enhanced by optimal positioning accuracy is achieved when Kmax is set
55.6% compared to traditional WKNN. Simultaneously, the to 3. Additionally, the smaller the value of Kmax , the fewer
overall boxplot of SCSC-WKNN is considerably lower than cycles of the algorithm will be executed, thereby enhancing
that of traditional WKNN, indicating the superiority of the its computational efficiency.
WKNN algorithm after moving window signal processing
in achieving higher positioning accuracy. Additionally, the
SCSC-WKNN box has a lower upper quantile, implying that
there are fewer large error cases compared to WKNN.

FIGURE 13. Position results of SCSC-SRL-SAWKNN under different Kmax .

3) STATIC LOCATION TEST

FIGURE 11. Positioning accuracy of WKNN before and after moving
The positioning error distributions of WKNN, SAWKNN,
window processing. SRL-KNN algorithms and our proposed SCSC-SRL-
SAWKNN algorithm in the static state are illustrated in
B. LOCALIZATION EFFECT OF SCSC-SRL-SAWKNN Fig. 14. The traditional WKNN and the improved SAWKNN
In this section, we shall examine the impact of critical and SRL-KNN algorithms exhibit average localization errors
parameters of SCSC-SRL-SAWKNN on localization accu- of 2.14, 2.06, and 1.40 meters, respectively, under station-
racy. As highlighted in Section IV, the algorithm is char- ary conditions. However, the mean positioning error of the
acterized by four essential parameters. The values of L and SCSC-SRL-SAWKNN algorithm under static conditions is

VOLUME 11, 2023 32915

 J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

state as a whole. Undoubtedly, the diminished precision in

the localization of the mobile state compared to that of the
stationary state can be attributed to the fact that a sole signal
collection is conducted for every location in the mobile state.
This implies that the number of signal acquisition frames
is extremely limited, resulting in the fingerprint signal’s
instability.
5) OVERALL LOCATION TEST
FIGURE 14. Comparing with traditional WKNN, SAWKNN, SRL-KNN. The The static and kinematic scenarios will be integrated next,
figure shows the distribution of positioning errors of WKNN, SAWKNN,
SRL-KNN and SCSC-SRL-SAWKNN in static state. and the overall performance of the SCSC-SRL-SAWKNN
algorithm will be assessed comprehensively.
a mere 0.94 meters, indicative of its significantly improved
localization accuracy in stationary conditions. In fact, its
positioning accuracy is outstanding. Compared with the
traditional WKNN algorithm, SCSC-SRL-SAWKNN has
improved the localization accuracy by 56.7%. Additionally,
SCSC-SRL-SAWKNN showed the lowest box plot compared
to the other three algorithms, indicating an overall better
positioning accuracy. Moreover, the upper quantile of this box
plot is lower than that of the other three algorithms, indicating
that SCSC-SRL-SAWKNN has the least instances of large
FIGURE 16. Comparing with traditional WKNN, SAWKNN, SRL-KNN. The
positioning errors. figure shows the distribution of positioning errors of WKNN, SAWKNN,
SRL-KNN and SCSC-SRL-SAWKNN in Overall state.
4) MOVING LOCATION TEST
The positioning error results of the comprehensive posi-
Subsequently, we will compare the positioning perfor- tioning experiment of WKNN, SAWKNN, SRL-KNN, and
mance of WKNN, SAWKNN, SRL-KNN, and SCSC-SRL- SCSC-SRL-SAWKNN are depicted in Fig. 16. The mean
SAWKNN under mobile conditions. positioning errors of these algorithms are 2.90, 2.73, 2.28, and
The illustration presented in Fig. 15 depicts the per- 1.84 meters, respectively. Notably, SCSC-SRL-SAWKNN
formance of four distinct algorithms, namely WKNN, outperforms the other algorithms in terms of positioning
SAWKNN, SRL-KNN, and SCSC-SRL-SAWKNN, in terms accuracy. Compared with the traditional WKNN algorithm,
of positioning accuracy under mobile conditions. The aver- our proposed algorithm demonstrates a 36.6% improvement
age positioning errors of these algorithms are 4.65, 4.29, in accuracy. The lower section of the box plot of SCSC-
4.32 and 3.94 meters, respectively. It is observed that SCSC- SRL-SAWKNN is significantly lower than that of traditional
SRL-SAWKNN outperforms other algorithms in terms of WKNN, indicating that it performs much better in terms
positioning accuracy in the mobile state. The boxplots of of high-precision positioning. Additionally, the quantile on
SCSC-SRL-SAWKNN exhibit smaller upper and lower quan- the box plot is much lower than that of other algorithms,
tiles in comparison to the other algorithms. This signifies that indicating that the occurrence of large errors in SCSC-
the proposed algorithm outperforms the others in terms of SRL-SAWKNN is much less frequent than in other algo-
high-precision localization during motion, while also exhibit- rithms. Overall, the algorithm exhibits excellent positioning
ing the least number of large-error localizations. However, accuracy.
a comparison of Fig. 15 reveals that the positioning accuracy Fig. 17 presents a comparison of the cumulative distribu-
in motion state is significantly lower than that in stationary tion function (CDF) of the overall localization error between

FIGURE 15. Comparing with traditional WKNN, SAWKNN, SRL-KNN. The

figure shows the distribution of positioning errors of WKNN, SAWKNN, FIGURE 17. CDF of overall positioning error for WKNN, SAWKNN,
SRL-KNN and SCSC-SRL-SAWKNN in Moving state. SRL-KNN, and SCSC-SRL-SAWKNN.

32916 VOLUME 11, 2023

J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

SCSC-SRL-SAWKNN and the other three KNN methods.

It is evident that SCSC-SRL-SAWKNN (depicted by the red
line) outperforms the other methods overall in terms of local-
ization accuracy. Furthermore, in terms of high-precision
performance within 1 meter, SRL-KNN and SCSC-SRL-
SAWKNN outperform the other two algorithms. Remarkably,
the maximum positioning error of SCSC-SRL-SAWKNN is
only 4.54 meters, whereas the maximum positioning errors
of WKNN, SAWKNN, and SRL-KNN are 6.95 meters,
5.39 meters and 7.38 meters, respectively. Thus, it can be FIGURE 19. Comparing with traditional WKNN, SAWKNN, SRL-KNN. This
figure shows the error distribution of WKNN, SAWKNN, SRL-KNN and
deduced that SCSC-SRL-SAWKNN excels in both high pre- SCSC-SRL-SAWKNN after Kalman filtering.
cision and stability.
It is evident that SCSC-SRL-SAWKNN outperforms the
C. LOCATION TRACKING
other three algorithms in terms of positioning accuracy,
Upon observing Section V-B, it is evident that the positioning
whether in static or dynamic states, overall performance,
error in the mobile state is significantly higher than that in the
or combined with Kalman filtering. Particularly, in static and
stationary state. In order to enhance the positioning accuracy
overall conditions, its localization accuracy is significantly
in the mobile state, a high-precision tracking function is
better than that of WKNN, SAWKNN, and SRL-KNN. The
implemented by combining the Kalman filter [38] with the
average time of SCSC-SRL-SAWKNN localization is only
localization algorithm. The positioning trajectories of each
6.8ms. This algorithm not only enhances positioning accu-
localization algorithm after Kalman filtering are displayed
racy, but also operates with minimal resource consumption,
in Fig. 18.
thereby offering an advantageous high real-time performance
in practical applications.
VI. CONCLUSION
The current WiFi-based indoor positioning methods typically
focus only on a single situation, either in stationary or moving
states, neglecting the fact that actual positioning involves a
combination of both static and dynamic situations. This paper
proposes the SCSC-SRL-SAWKNN algorithm, which intel-
ligently leverages the characteristics of both states. By pro-
cessing signals based on the features of different positioning
states, our proposed algorithm updates the fingerprint via the
mean accumulation method after processing the signal with a
moving window under static conditions, which significantly
improves the stability of RSSI. Then combined with the
SAWKNN algorithm, the positioning accuracy in static state
FIGURE 18. Ground truth and estimated trajectories. Green line
can be greatly improved. To address the mismatch between
represents the trajectory ground truth. Red lines are estimated physical distance and fingerprint similarity in the moving
trajectories. state, the algorithm transforms the Manhattan distance by
incorporating the soft range limitation concept, and then com-
The green line in Fig. 18 represents the actual trajectory, bines it with SAWKNN to improve positioning accuracy. Par-
while the red line shows the positioning tracking effect. ticularly for overall positioning cases that involve both static
SCSC-SRL-SAWKNN algorithm exhibits the most precise and dynamic states, the state-by-state processing approach
location tracking effect, with its positioning trajectory fitting of SCSC-SRL-SAWKNN algorithm effectively utilizes the
the real path better than other algorithms. The positioning characteristics of each state, leading to a remarkable improve-
errors of the four algorithms combined with Kalman filtering ment in positioning accuracy. Our future research aims to
are ultimately presented in the ensuing figure. develop a more robust fingerprint combination and a signal
After applying Kalman filtering, the mean errors of preprocessing method for variable-step moving windows to
WKNN, SAWKNN, SRL-KNN, and SCSC-SRL-SAWKNN further enhance the positioning accuracy and resolve the issue
decrease to 2.44, 2.40, 1.82, and 1.27 meters, respectively. of equipment heterogeneity.
Notably, the SCSC-SRL-SAWKNN algorithm demonstrates
significantly smaller errors than the other methods. Addi- REFERENCES
tionally, the lower and upper bounds of the box plot for this
[1] A. Poulose and D. S. Han, ‘‘UWB indoor localization using deep
algorithm were considerably smaller than those of the other learning LSTM networks,’’ Appl. Sci., vol. 10, no. 18, p. 6290,
algorithms, indicating its superiority in all directions. Sep. 2020.

VOLUME 11, 2023 32917

 J. Hu, C. Hu: WiFi Indoor Location Tracking Algorithm Based on Improved WKNN and Kalman Filter

[2] L. Bai, F. Ciravegna, R. Bond, and M. Mulvenna, ‘‘A low cost indoor [24] M. Sun, Y. Wang, S. Xu, H. Yang, and K. Zhang, ‘‘Indoor geomagnetic
positioning system using Bluetooth low energy,’’ IEEE Access, vol. 8, positioning using the enhanced genetic algorithm-based extreme learning
pp. 136858–136871, 2020. machine,’’ IEEE Trans. Instrum. Meas., vol. 70, pp. 1–11, 2021.
[3] A. Poulose, J. Kim, and D. S. Han, ‘‘A sensor fusion framework for indoor [25] J. Koo and H. Cha, ‘‘Localizing WiFi access points using signal strength,’’
localization using smartphone sensors and Wi-Fi RSSI measurements,’’ IEEE Commun. Lett., vol. 15, no. 2, pp. 187–189, Feb. 2011.
Appl. Sci., vol. 9, no. 20, p. 4379, Oct. 2019. [26] C. Feng, S. Arshad, S. Zhou, D. Cao, and Y. Liu, ‘‘Wi-multi: A three-phase
[4] X. Hou and J. Bergmann, ‘‘Pedestrian dead reckoning with wearable sen- system for multiple human activity recognition with commercial WiFi
sors: A systematic review,’’ IEEE Sensors J., vol. 21, no. 1, pp. 143–152, devices,’’ IEEE Internet Things J., vol. 6, no. 4, pp. 7293–7304, Aug. 2019.
Jan. 2021. [27] P. Bahl and V. N. Padmanabhan, ‘‘Radar: An in-building RF-based user
[5] Y. Yu, R. Chen, L. Chen, S. Xu, W. Li, Y. Wu, and H. Zhou, ‘‘Precise 3- location and tracking system,’’ in Proc. IEEE INFOCOM Conf. Comput.
D indoor localization based on Wi-Fi FTM and built-in sensors,’’ IEEE Commun., 19th Annu. Joint Conf. IEEE Comput. Commun. Societies,
Internet Things J., vol. 7, no. 12, pp. 11753–11765, Dec. 2020. vol. 2, Mar. 2000, pp. 775–784.
[6] J. Xue, J. Liu, M. Sheng, Y. Shi, and J. Li, ‘‘A WiFi fingerprint based high- [28] S. Zhang, X. Li, M. Zong, X. Zhu, and R. Wang, ‘‘Efficient kNN classifi-
adaptability indoor localization via machine learning,’’ China Commun., cation with different numbers of nearest neighbors,’’ IEEE Trans. Neural
vol. 17, no. 7, pp. 247–259, Jul. 2020. Netw. Learn. Syst., vol. 29, no. 5, pp. 1774–1785, May 2018.
[7] B. Huang, R. Yang, B. Jia, W. Li, and G. Mao, ‘‘A theoretical analysis on [29] M. Brunato and R. Battiti, ‘‘Statistical learning theory for location finger-
sampling size in WiFi fingerprint-based localization,’’ IEEE Trans. Veh. printing in wireless LANs,’’ Comput. Netw., vol. 47, no. 6, pp. 825–845,
Technol., vol. 70, no. 4, pp. 3599–3608, Apr. 2021. Apr. 2005.
[30] M. Sun, Y. Wang, S. Xu, H. Qi, and X. Hu, ‘‘Indoor positioning tightly
[8] W. Xue, K. Yu, X. Hua, Q. Li, W. Qiu, and B. Zhou, ‘‘APs’ virtual positions-
coupled Wi-Fi FTM ranging and PDR based on the extended Kalman filter
based reference point clustering and physical distance-based weighting
for smartphones,’’ IEEE Access, vol. 8, pp. 49671–49684, 2020.
for indoor Wi-Fi positioning,’’ IEEE Internet Things J., vol. 5, no. 4,
[31] H. Zou, M. Jin, H. Jiang, L. Xie, and C. J. Spanos, ‘‘WinIPS: WiFi-based
pp. 3031–3042, Aug. 2018.
non-intrusive indoor positioning system with online radio map construc-
[9] B. Wang, X. Gan, X. Liu, B. Yu, R. Jia, L. Huang, and H. Jia, ‘‘A novel tion and adaptation,’’ IEEE Trans. Wireless Commun., vol. 16, no. 12,
weighted KNN algorithm based on RSS similarity and position distance pp. 8118–8130, Dec. 2017.
for Wi-Fi fingerprint positioning,’’ IEEE Access, vol. 8, pp. 30591–30602,
[32] T. Koike-Akino, P. Wang, M. Pajovic, H. Sun, and P. V. Orlik,
2020.
‘‘Fingerprinting-based indoor localization with commercial mmWave
[10] J. Niu, B. Wang, L. Shu, T. Q. Duong, and Y. Chen, ‘‘ZIL: An energy- WiFi: A deep learning approach,’’ IEEE Access vol. 8, pp. 84879–84892,
efficient indoor localization system using ZigBee radio to detect WiFi 2020.
fingerprints,’’ IEEE J. Sel. Areas Commun., vol. 33, no. 7, pp. 1431–1442, [33] X. Song, X. Fan, C. Xiang, Q. Ye, L. Liu, Z. Wang, X. He, N. Yang,
Jul. 2015. and G. Fang, ‘‘A novel convolutional neural network based indoor
[11] S. Li, M. Hedley, K. Bengston, D. Humphrey, M. Johnson, and W. Ni, localization framework with WiFi fingerprinting,’’ IEEE Access, vol. 7,
‘‘Passive localization of standard WiFi devices,’’ IEEE Syst. J., vol. 13, pp. 110698–110709, 2019.
no. 4, pp. 3929–3932, Dec. 2019. [34] C. Liu, C. Wang, and J. Luo, ‘‘Large-scale deep learning framework on
[12] Y. Zheng, M. Sheng, J. Liu, and J. Li, ‘‘Exploiting AoA estimation accu- FPGA for fingerprint-based indoor localization,’’ IEEE Access, vol. 8,
racy for indoor localization: A weighted AoA-based approach,’’ IEEE pp. 65609–65617, 2020.
Wireless Commun. Lett., vol. 8, no. 1, pp. 65–68, Feb. 2019. [35] X. Guo, L. Li, F. Xu, and N. Ansari, ‘‘Expectation maximization indoor
[13] M. R. Nowicki and P. Skrzypczynski, ‘‘Leveraging visual place recognition localization utilizing supporting set for Internet of Things,’’ IEEE Internet
to improve indoor positioning with limited availability of WiFi scans,’’ Things J., vol. 6, no. 2, pp. 2573–2582, Apr. 2019.
Sensors, vol. 19, no. 17, p. 3657, Aug. 2019. [36] Y. Li, S. Williams, B. Moran, and A. Kealy, ‘‘A probabilistic indoor
[14] M. T. Hoang, B. Yuen, X. Dong, T. Lu, R. Westendorp, and K. R. Tarimala, localization system for heterogeneous devices,’’ IEEE Sensors J., vol. 19,
‘‘Semi-sequential probabilistic model for indoor localization enhance- no. 16, pp. 6822–6832, Aug. 2019.
ment,’’ IEEE Sensors J., vol. 20, no. 11, pp. 6160–6169, Jun. 2020. [37] M. Zhou, X. Li, Y. Wang, S. Li, Y. Ding, and W. Nie, ‘‘6G multisource-
[15] Z. Chen, H. Zou, J. Yang, H. Jiang, and L. Xie, ‘‘WiFi fingerprinting indoor information-fusion-based indoor positioning via Gaussian kernel density
localization using local feature-based deep LSTM,’’ IEEE Syst. J., vol. 14, estimation,’’ IEEE Internet Things J., vol. 8, no. 20, pp. 15117–15125,
no. 2, pp. 3001–3010, Jun. 2020. Oct. 2021.
[16] A. Poulose and D. S. Han, ‘‘Performance analysis of fingerprint matching [38] Z. Chen, H. Zou, H. Jiang, Q. Zhu, Y. C. Soh, and L. Xie, ‘‘Fusion of
algorithms for indoor localization,’’ in Proc. Int. Conf. Artif. Intell. Inf. WiFi, smartphone sensors and landmarks using the Kalman filter for indoor
Commun. (ICAIIC), Feb. 2020, pp. 661–665. localization,’’ Sensors, vol. 15, no. 1, pp. 715–732, Jan. 2015.
[17] J. Hu, D. Liu, Z. Yan, and H. Liu, ‘‘Experimental analysis on weight
K-nearest neighbor indoor fingerprint positioning,’’ IEEE Internet Things
J., vol. 6, no. 1, pp. 891–897, Feb. 2019. JIUSONG HU received the M.Sc. and Ph.D.
[18] M. T. Hoang, Y. Zhu, B. Yuen, T. Reese, X. Dong, T. Lu, R. Westendorp, degrees from Hunan University, Changsha, China,
and M. Xie, ‘‘A soft range limited K-nearest neighbors algorithm for in 2014 and 2019, respectively. He is currently
indoor localization enhancement,’’ IEEE Sensors J., vol. 18, no. 24, a Lecturer with the Department of Information
pp. 10208–10216, Dec. 2018. Engineering, Hunan University of Technology,
[19] T. Raharijaona, R. Mawonou, T. Nguyen, F. Colonnier, M. Boyron, Zhuzhou, China. His research interests include
J. Diperi, and S. Viollet, ‘‘Local positioning system using flickering intelligent perception, information processing, and
infrared LEDs,’’ Sensors, vol. 17, no. 11, p. 2518, Nov. 2017. indoor positioning.
[20] M. O. Khyam, M. N.-A. Rahim, X. Li, A. Jayasuriya, M. A. Mahmud,
A. M. T. Oo, and S. S. Ge, ‘‘Simultaneous excitation systems for ultrasonic
indoor positioning,’’ IEEE Sensors J., vol. 20, no. 22, pp. 13716–13725,
Nov. 2020.
[21] X. Zhu, J. Yi, J. Cheng, and L. He, ‘‘Adapted error map based mobile robot CONGWEI HU received the B.S. degree in elec-
UWB indoor positioning,’’ IEEE Trans. Instrum. Meas., vol. 69, no. 9, tronic information engineering from the Hunan
pp. 6336–6350, Sep. 2020. University of Technology, in 2020, where he is
[22] A. A. N. Shirehjini and S. Shirmohammadi, ‘‘Improving accuracy and currently pursuing the master’s degree in elec-
robustness in HF-RFID-based indoor positioning with Kalman filtering tronic information. His current research interests
and Tukey smoothing,’’ IEEE Trans. Instrum. Meas., vol. 69, no. 11, include WSN, intelligent information processing,
pp. 9190–9202, Nov. 2020. fault detection, and indoor location.
[23] W. Zhang, K. Yu, W. Wang, and X. Li, ‘‘A self-adaptive AP selection algo-
rithm based on multiobjective optimization for indoor WiFi positioning,’’
IEEE Internet Things J., vol. 8, no. 3, pp. 1406–1416, Feb. 2021.

32918 VOLUME 11, 2023