Article

An Indoor Continuous Positioning Algorithm on the

Move by Fusing Sensors and Wi-Fi on Smartphones
Huaiyu Li 1, *, Xiuwan Chen 1 , Guifei Jing 2 , Yuan Wang 1 , Yanfeng Cao 1,3 , Fei Li 1 ,
Xinlong Zhang 1 and Han Xiao 1
Received: 5 October 2015; Accepted: 4 December 2015; Published: 11 December 2015
Academic Editor: Leonhard M. Reindl
1 Institute of Remote Sensing and GIS, Peking University, No. 5 Yiheyuan Road, Haidian District,
Beijing 100871, China; [email protected] (X.C.); [email protected] (Y.W.);
[email protected] (Y.C.); [email protected] (F.L.); [email protected] (X.Z.);
[email protected] (H.X.)
2 National Remote Sensing Center of China, No. 8A, Liulinguan Nanli, Haidian District,
Beijing 100036, China; [email protected]
3 Beijing Aerospace Unmanned Vehicles System Engineering Research Institute, No. 1 Fengyingdong Road,
Haidian District, Beijing 100094, China
* Correspondence: [email protected]; Tel.: +86-10-6274-5183; Fax: +86-10-6276-7119

Abstract: Wi-Fi indoor positioning algorithms experience large positioning error and low stability
when continuously positioning terminals that are on the move. This paper proposes a novel indoor
continuous positioning algorithm that is on the move, fusing sensors and Wi-Fi on smartphones. The
main innovative points include an improved Wi-Fi positioning algorithm and a novel positioning
fusion algorithm named the Trust Chain Positioning Fusion (TCPF) algorithm. The improved Wi-Fi
positioning algorithm was designed based on the properties of Wi-Fi signals on the move, which
are found in a novel “quasi-dynamic” Wi-Fi signal experiment. The TCPF algorithm is proposed
to realize the “process-level” fusion of Wi-Fi and Pedestrians Dead Reckoning (PDR) positioning,
including three parts: trusted point determination, trust state and positioning fusion algorithm.
An experiment is carried out for verification in a typical indoor environment, and the average
positioning error on the move is 1.36 m, a decrease of 28.8% compared to an existing algorithm.
The results show that the proposed algorithm can effectively reduce the influence caused by the
unstable Wi-Fi signals, and improve the accuracy and stability of indoor continuous positioning on
the move.

Keywords: indoor positioning; Wi-Fi; PDR; multi-sensor fusion; TCPF

1. Introduction
Indoor positioning technology based on smartphones has many application scenarios since
people stay inside buildings more than 80% of their daily life [1,2]. Due to the advantages of
low cost, high accuracy, and wide popularization, Wi-Fi indoor positioning has become one of the
mainstream indoor positioning technologies. After the RADAR system [3,4] firstly put forward
the Wi-Fi indoor fingerprint positioning scheme, many indoor positioning algorithms based on its
framework were proposed, such as probability positioning algorithms [5–7] based on the Bayesian
estimation, and machine learning algorithms [8–10] including Support Vector Machine (SVM) and
neural network. As a result, The indoor positioning accuracy has been further improved, achieving
about, on average, a 3 m positioning accuracy, which approximates to the size of a typical office.
Current Wi-Fi positioning algorithms impose the impractical prerequisite that users and terminals
remain at a fixed location during the positioning process [11]. However, the application scenarios of

Sensors 2015, 15, 31244–31267; doi:10.3390/s151229850 www.mdpi.com/journal/sensors

Sensors 2015, 15, 31244–31267

indoor positioning are most focused on smartphones, which leads to many positioning problems on
the move, such as large positioning error, positioning jumps and accuracy reduction [12,13]. There
are few researches focusing on the improvement of Wi-Fi positioning algorithms on the move.
Another direct method to improve positioning accuracy is fusing two or more complementary
technologies. With the improvement of integration and the power consumption reduction of
multi-sensors in recent years, more and more sensors are integrated in smartphones. Algorithms
for fusing sensors and Wi-Fi have become a research hotspot [14]. Based on built-in sensors,
the PDR positioning can calculate the relative displacement to realize the indoor positioning on
smartphones. As a relative positioning method, the PDR positioning has a high accuracy over a
short distance, and it needs a reference point as a starting point. However, the most serious problem
is that the error will be accumulated over time. On the contrary, the Wi-Fi positioning belongs
to the absolute positioning method, which will not accumulate the positioning error. These two
positioning algorithms are complementary to each other so that the fusion can prominently improve
the performance of indoor positioning.
The current fusion algorithms mainly include algorithms based on Particle Filter [15,16],
algorithms based on Kalman Filter [17–24], the Cross-Assistive algorithm [12,13,25], and so on. The
algorithm based on Particle Filter has an intuitive process, but the large amount of computation is not
suitable for a handheld device [15,17,18]. The algorithm based on Kalman Filter has a good real-time
performance, but the fusion is on the “result-level” so that positioning results will be easily skewed
by Wi-Fi signal interferences under volatile conditions. The Cross-Assistive algorithm is currently
proposed to achieve deep fusion in the Wi-Fi/PDR positioning process. However, the algorithm is
not stable, and is prone to fall into error cycles.
Overall, we put forward a novel indoor continuous positioning algorithm fusing built-in sensors
and Wi-Fi on smartphones. Compared to traditional algorithms, it has two innovative points. The first
one is an improved Wi-Fi positioning algorithm, and the other is a new positioning fusion algorithm
named TCPF algorithm. Through these two improvements, the proposed algorithm can optimize
indoor positioning performance of targets on the move.
The following sections are arranged as follows. The previous related researches will be reviewed
in Section 2. A novel “quasi-dynamic” Wi-Fi signal experiment is conducted to analyze the properties
of Wi-Fi signals on the move, and an improved Wi-Fi positioning algorithm will be proposed in
Section 3. The novel positioning fusion algorithm, including trusted point determination, trust state
and positioning fusion algorithm, is introduced after the overall framework of the indoor continuous
positioning algorithm is outlined in Section 4. Another field indoor experiment was conducted to
verify the proposed algorithm, and the results are analyzed in three parts in Section 5. Conclusions
and the future research direction are summarized in Section 6.

2. Related Work
Researches on the Wi-Fi positioning field focus on positioning algorithm improvement [10],
signal analysis [26,27], fingerprint database construction [28], and so on. Their fundamental purpose
is to improve Wi-Fi positioning performance, which is greatly affected by the effective restoration
of actual long-time signals from the short-time signals gathered by terminals. Through analyzing the
statistical properties of a large number of Wi-Fi signals, many researchers tried to analyze and explain
the factors affecting indoor positioning accuracy and stability.
Kamol Kaemarungsi et al. analyzed the signal distribution, mean, standard deviation, deviation
and stability of Wi-Fi Received Signal Strength Indicator (RSSI) from the perspective of indoor
positioning systems [26]. Jiayou Luo et al. analyzed the properties of Wi-Fi RSSI distribution
and differences among smartphones with the purpose to improve the Wi-Fi indoor positioning
accuracy [27]. However, current Wi-Fi positioning algorithms are mostly based on the signal
distributions of terminals at stationary state; there is a lack of research on the properties of Wi-Fi
signals on the move and corresponding positioning algorithm improvements.

31245
Sensors 2015, 15, 31244–31267

Fusing with inertial positioning is another focus. The existing fusion algorithms mainly include
the algorithm based on Particle Filter [15,16], the algorithm based on Kalman Filter [17–24], and the
Cross-Assistive algorithm [12,13,25]. The advantages and disadvantages of the three algorithms are
shown as Table 1.

Table 1. A comparison of existing fusion algorithms.

Advantages Disadvantages

1. Clear and intuitive process

Based on Particle Filter 2. Integrated with 1. Large computations
map matching

1. Fusion based on positioning

1. Good real-time performance result level
Based on Kalman Filter 2. Simple but effective 2. Susceptible to
Wi-Fi interference

1. Wi-Fi positioning filtered 1. There is inevitably some

by PDR error cycles
Cross-Assistive Approach 2. Positioning update by 2. Susceptible to
accurate result Wi-Fi interference

The Wi-Fi/INS fusion indoor positioning algorithm based on Particle Filter was successively
proposed by Frederic Evennou et al. [15] and Hui Wang et al. [16]. A large number of particles are used
to fit the discrete probability density function of the target. The spread of the particles is controlled by
the acceleration at a random process, and map information is integrated to filter out the unreasonable
movement, thus final position estimation is obtained. The algorithm has an intuitive and effective
fusion process, however, it is time consuming since every step needs certain operations for each
particle, and the number of particles can reach thousands, or even more. Overall, the algorithm
based on Particle Filter is not suitable for smartphones whose computing power and resources are
limited [15,17,18].
Kalman Filter algorithm is the core of the GPS/INS integrated navigation algorithm, which can
also be specifically applied to the Wi-Fi/PDR fusion positioning. Zhenghua Chen et al. studied the
Wi-Fi/PDR fusion positioning algorithm based on Kalman Filter [17]. Further, Zhi-An Deng et al. [18]
and Veerachai Malyavej et al. [22] respectively studied the improved forms of Kalman Filter, Extended
Kalman Filter (EKF), and Unscented Kalman Filter (UKF). Simo Ali-Loytty et al. [24] proposed an
Fingerprint Kalman Filter (FKF) on the basis of the EKF and UKF. The algorithm based on Kalman
Filter demonstrated a good real-time performance, but the fusion process is based on the level of
positioning results. When Wi-Fi signal fluctuates under the condition of severe interference, hops will
appear in the Wi-Fi positioning results, and deviations will be fused into the final location estimation
so that the positioning stability needs further improvement.
The Cross-Assistive positioning algorithm is proposed for Wi-Fi/PDR fusion in the latest three
years. K. Miyazaki et al. [12] proposed this algorithm to limit the fingerprint search scope based on
PDR estimation and to use accurate Wi-Fi positioning results to correct the PDR positioning in order
to achieve a finer fusion process. Chang Qiang et al. [13] further put forward an Error Distance to
replace Euclidean Distance. However, there are some defects, such as the error cycle. Positioning
fusion is triggered by the “accurate” Wi-Fi positioning results to fix the PDR positioning cumulated
error, but Wi-Fi positioning results may experience a large deviation over a period of time due to large
disturbances. In this case, fusion positioning results cannot be corrected, error is accumulated in the
PDR positioning, giving no “accurate” positioning results, and finally the positioning deviates from
the track. Another problem is using the sector area established by the PDR positioning and the latest

31246
Sensors 2015, 15, 31244–31267

location to determine whether positioning is accurate, because there is no bias in Wi-Fi positioning
error so step length would be overestimated.
Other fusion algorithms include the sequential Monte Carlo filter, which was developed by
the joint research between the University of California, Santa Barbara (UCSB) and Massachusetts
Institute of Technology (MIT), fusing built-in inertial sensors with Wi-Fi to realize indoor navigation
and positioning [29].

3. An Improved Wi-Fi Positioning Algorithm

The Wi-Fi fingerprint positioning algorithm usually works in two phases: an offline training
phase and an online positioning phase. During the offline phase, the system tabulates the signal
strength received from the access points (APs) at selected locations (which are called fingerprint
points) in the area of interest, resulting in a so-called radio map. During the online positioning phase,
the system uses the signal strength samples received from the APs to “search” the radio map and
estimate the user location. Classic Wi-Fi positioning algorithm is the Weighted K Nearest Neighbor
(WKNN [9]) algorithm, which is the improved version of the RADAR [3] system’s Nearest Neighbor
(NN) algorithm. The estimated location is the weighted sum of the locations of the K fingerprint
points, which have the minimum signal space Euclidean distance in the fingerprint database. The
signal space distance can be expressed as Equation (1)
¨ ˛1{ q
n
ÿ ` ˘q
Lqi “ ˝ S j ´ Sij ‚ (1)
j “1

where n is the dimension of the Wi-Fi signal in the fingerprint database. S j is the RSSI referred to
the jth AP sampled during the online phase, and Sij is that during offline phase at the ith fingerprint
points. When q is set 1, Lqi is the Manhattan distance. When q is set 2, Lqi is the Euclidean distance,
which is adopted in this paper. Compared to the NN algorithm, the WKNN algorithm improves
the positioning performance by fusing the K nearest fingerprint points (ordered according to Li ), as
shown in Equations (2) and (3).
1
Li
wi “ (2)
řK 1
i “0
Li
K
ÿ ` ˘
px̂, ŷq “ wi xi , yi (3)
i “1

where Li is the signal space Euclidean distance referred to the ith fingerprint point. K is the number
of selected fingerprint points, and wi is the weight of the ith one.
The stationary positioning accuracy is about 3 m [9], by using the classic WKNN algorithm in
an ideal positioning environment without electromagnetic interference and crowds. However, when
the user is on the move, or influenced by small scale fluctuations [7], there will be problems, such as
a big positioning error and poor stability. The properties of Wi-Fi signals, especially at moving state,
need to be deeply analyzed in order to improve the Wi-Fi positioning performance.

3.1. Properties of Wi-Fi Signals on the Move

Wi-Fi follows the 802.11ˆ series standard. The measured value of the RSSI is the received
instantaneous radio energy, which is the baseband IQ power integration in 104 µs after a reverse
channel baseband filter. The calculation of RSSI can be expressed as Equation (4), where the unit of
RSSI is dBm, and P0 is set as 1 mW.
P
RSSI “ 10logp q (4)
P0

31247
Sensors 2015, 15, 31244–31267

However, it is insufficient to describe the properties of Wi-Fi signals on the move. We propose
two new parameters, “refresh rate” and “loss rate”. Two adjacent Wi-Fi RSSI measurements have a
minimum time interval, which is corresponding to the fastest refresh frequency. When the connection
is good,
Sensors the15,
2015, Wi-Fi signal refresh frequency can maintain the fastest rate. The rate between the current
page–page
Wi-Fi signal refresh frequency and the fastest one is calculated as the “refresh rate”. The “loss rate”
is the
is the mean
mean probability
probability of of the
the terminal
terminal losing
losing the
the Wi-Fi
Wi-Fi signal
signal in in the
the continuous
continuous measurements.
measurements.
Through continuously
Through continuously measuring
measuring thethe Wi-Fi
Wi-FiRSSI,
RSSI,the
the number
number of of null
null values
values divides
divides the
the total
total number
number
of measurements to get the “loss rate”. These two parameters are used to
of measurements to get the “loss rate”. These two parameters are used to quantify the instability of quantify the instability of
Wi-Fi signals. The smaller the “refresh rate” is, or the larger the “loss rate” is,
Wi-Fi signals. The smaller the “refresh rate” is, or the larger the “loss rate” is, the less reliable the the less reliable the
current signal
current signal is.
is. This
This paper
paper focuses
focuses onon the
the “refresh
“refresh rate”,
rate”, and
and thethe “loss
“loss rate”
rate” can
can be
be deduced
deduced in in aa
similar way.
similar way.
According totothethe
According measurement
measurement principle
principle of RSSI,
of RSSI, the RSSI themeasurement
RSSI measurementhas a power has threshold
a power
threshold
P r . When the
𝑃𝑟 . When
received thesignal
received signal
power power isthan
is stronger stronger
the Pthan the 𝑃𝑟 , the
r , the signal can signal can be accurately
be accurately captured.
captured. Otherwise, the signal cannot be easily captured, resulting in
Otherwise, the signal cannot be easily captured, resulting in none refreshing RSSI. It is concludednone refreshing RSSI. It is
concluded
that that the
the “refresh rate” “refresh
shouldrate” should proportional
be inversely be inversely proportional to the signal
to the signal strength, which strength, which is
is exponentially
exponentially related to the RSSI. Overall, the relationship model between
related to the RSSI. Overall, the relationship model between the “refresh rate” (η) and the RSSI the “refresh rate” (𝜂) andis
Equation (5), where the parameters 𝜂 , 𝑎, 𝑏, 𝑐
shown in Equation (5), where the parameters η0 , a, b, c are 0all real numbers greater than zero. than
the RSSI is shown in are all real numbers greater The
zero. The
model modelRefresh
is named is named Refresh
Rate modelRate model
of Wi-Fi of Wi-Fi
signals on signals
the move on(RR).
the move (RR).

a aa
η=η0`
=0 + “ η `0 +
a
bP 0 RSSI  RSSI r (5)
RSSI
( ´ RSSI (5)
1Pr ( r )b
) r
1`p q
1`e
1 e
p´ c
c
q
P P
We set
We set up
up aa “quasi-dynamic”
“quasi-dynamic” Wi-Fi Wi-Fi signal
signal experiment
experiment (called
(called Experiment
Experiment 1) 1) at
at Remote
Remote Sensing
Sensing
Building in Peking University.
Building University. The indoorindoor layout
layout ofofexperiment
experimentarea areaisisshown
shownininFigure
Figure1.1.One
One APAPis
put
is putinin
Room
Room 404 with
404 with a marker
a marker “★”
“‹”andandthethe
“quasi-dynamic”
“quasi-dynamic” Wi-Fi
Wi-Fisignal measurement
signal measurement is is
taken at
taken
thethe
at place with
place withthe
themarker
marker“▲”. “N”.Measurements
Measurementsare aremade
made20 20 times
times at at each
each point with 0.5 s intervals.
intervals.
Adjacent point interval is
Adjacent is 11mmandandthe
thetotal number
total number of of
points is 38.
points All points
is 38. are separated
All points by a door
are separated by a
from from
door the AP,
the which is representative
AP, which of the
is representative of most typical
the most indoor
typical environment.
indoor environment. Different from
Different the
from
terminal
the statically
terminal measuring
statically measuring Wi-Fi
Wi-Fisignals
signalsininthe
thetraditional
traditional experiment,
experiment, the “quasi-dynamic”
“quasi-dynamic”
measurement means
measurement means that
that thethe terminal
terminal keeps
keeps moving
moving to measure
measure Wi-Fi
Wi-Fi signals
signals at
at each
each measurement
measurement
point where
point where the
the user
user is.
is. The terminal was was held
held at
at the
the palm,
palm, swinging
swinging with with the
the hand.
hand. The user
user is
is
standing stationary
standing stationary atat the
the fixed
fixed point
point without
withoutmoving
movingthe thefeet
feetand
andthethebody.
body.

Figure 1.
Figure 1. “Quasi-dynamic” Wi-Fi
Wi-Fi signal
signal experimental
experimental indoor
indoor layout.
layout.

The variations of the average RSSI, “refresh rate” and “loss rate” of Wi-Fi signals as the distance
The variations of the average RSSI, “refresh rate” and “loss rate” of Wi-Fi signals as the distance
from the AP increases are shown in Figure 2. Two interesting discoveries from the results are
from the AP increases are shown in Figure 2. Two interesting discoveries from the results are
listed below:
listed below:
1. When the Wi-Fi signal quality is good (RSSI is stronger than −70 dBm), the refresh frequency is
1. When the Wi-Fi signal quality is good (RSSI is stronger than ´70 dBm), the refresh frequency is
high, being maintained within 1 s (minimum is 0.667), and the “loss rate” is 0 without packet
high, being maintained within 1 s (minimum is 0.667), and the “loss rate” is 0 without packet
loss. In other words, at this time, the Wi-Fi signal is reliable.
loss. In other words, at this time, the Wi-Fi signal is reliable.
2. When the Wi-Fi signal quality is bad (RSSI is weaker than −80 dBm), the “refresh rate” quickly
decreases and the “loss rate” quickly increases, nearly to zero, as the RSSI decreases. In other
31248
words, at this time, the Wi-Fi signal is unreliable.
Sensors 2015, 15, 31244–31267

2. When the Wi-Fi signal quality is bad (RSSI is weaker than ´80 dBm), the “refresh rate” quickly
decreases and the “loss rate” quickly increases, nearly to zero, as the RSSI decreases. In other
words, at this time, the Wi-Fi signal is unreliable.
Sensors 2015, 15, page–page
Sensors 2015, 15, page–page

Figure2.2.Wi-Fi
Figure Wi-Fivariation
variation on
onthe
themove.
move.
Figure 2. Wi-Fi variation on the move.
The fitted curve of RR model according to Equation (5) is shown in Figure 3, where 𝜂0 = 2.1630,
The fitted curve
The fitted
a = 82.4917, curve
𝑅𝑆𝑆𝐼
of=RR model
modelaccording
of−75.0075,
RR according to Equation
c = 1.8464,toand
Equation (5)iscorrelation
(5)
the fitting
is shown
shown in Figure
in Figure 3,
3, where
coefficient
where
𝜂0 = η
is 0.9285. 0 “ 2.1630,
2.1630,
There is
𝑟
aan=obvious
82.4917,inflection
a “ 82.4917, RSSI 𝑅𝑆𝑆𝐼
r “ 𝑟 = point
−75.0075,
´75.0075, c “
around c =−75
1.8464,
1.8464, and
dBm in the
and fitting
thethe fitting
fitted correlation
correlation
curve. coefficient
coefficient
The “refresh is
rate” is 0.9285.There is is an
There
0.9285.decreases
rapidly
obvious
an inflection
obvious
as the point around
RSSI inflection
decreases. point
It happens´75todBm
around −75 in the
dBm
conform in fitted
the
with curve.
fitted
the curve.
design ofThe “refresh
TheWi-Fi
the “refresh rate”
rate”
signal; rapidly
therapidly isdecreases
defined as
signal decreases
as
as the
the RSSI RSSI decreases.
decreases.
weak when It RSSI
its Itishappens
happens to−75
conform
to conform
between dBmwithwith
and the the
−85 designofofthe
design
dBm. theWi-Fi
Wi-Fi signal;
signal; the
thesignal
signalisisdefined
defined as
weakaswhen
weakits when
RSSI itsisRSSI is between
between ´75 −75
dBm dBmandand
´85 −85dBm.
dBm.

Figure 3. Curve fitting refresh rate function.

Figure3.3.Curve
Figure Curve fitting refreshrate
fitting refresh ratefunction.
function.
The above statistical distribution partly explains the deterioration of Wi-Fi positioning
The above
performance statistical
at moving state.distribution partly explains
There is no different the deterioration
consideration of Wi-Fi
given for stationary positioning
or moving states
performance at moving state. There is no different consideration given for stationary or moving states
31249
6
6
Sensors 2015, 15, 31244–31267

The above statistical distribution partly explains the deterioration of Wi-Fi positioning
performance at moving state. There is no different consideration given for stationary or moving states
in traditional Wi-Fi positioning algorithms. When the terminal is online positioning at stationary
state, Wi-Fi signals can be sampled many times and be well estimated through a mean filter with a
sliding window so that weak signals can still be accurately captured, and they are well matched with
the offline data. However, the “refresh rate” of the weak signal is low, and the “loss rate” is high when
the terminal is moving with a continuous positioning. At this time, due to the cache mechanism of the
base hardware in sampling signals, the RSSI measurement deviates from the actual value. It leads to
the mismatch with the offline data so that increases the probability of unstable positioning and large
positioning errors.

3.2. Algorithm Improvement

According to the properties of Wi-Fi signals on the move, we put forward two methods including
dynamically adjusting the RSSI threshold and AP matching.

3.2.1. Dynamic Adjustment of RSSI Threshold

According to the fitted distribution of the “refresh rate”, Wi-Fi signals on the move vary
dramatically near ´75 dBm. We put forward the method to dynamically adjust the RSSI threshold
based on the user’s motion state. By analyzing the mean and variance of the horizontal acceleration
(the component perpendicular to the gravity acceleration), and variance of the resultant acceleration,
the static state can be accurately identified when all these three parameters are less than a certain
threshold, respectively [30]. The simplest way is to set the threshold as ´85 dBm at stationary state,
and the corresponding ´75 dBm at moving state, respectively. Of course, the threshold can also be
dynamically adjusted according to other motion states.

3.2.2. AP Match
We can see from Figure 2 that the RSSI decreases quickly near the AP (within 10 m). The signal
then decays slowly until reaching a threshold, where the “refresh rate” decreases, the “loss rate”
increases rapidly, and the signal gradually disappears. Considering that adjacent APs usually have a
certain distance in the real set, the strong RSSI of Wi-Fi signal has a significant clustering feature.
Here, we put forward the AP matching method. Considering the relative stability of the strong
signal in dynamic positioning and the overlapping area between the adjacent APs, offline fingerprints
with two strongest APs containing the strongest AP of the online signal are selected to calculate the
Wi-Fi positioning, as shown in Equation (6).
! )
APmax,online P APtop2,o f f line (6)

The clustering is done in the offline training phase to build up the associated fingerprint
database. The codes of two strongest APs are extracted from fingerprint points as their indices. When
the target is at stationary or moving state, it would be easy to use the AP matching method to find
the candidate fingerprint points instead of searching through the database. Only matched fingerprint
points can be selected to calculate the Wi-Fi positioning so that large positioning errors could be
avoided. The AP matching algorithm can effectively reduce the amount of online data that need to be
considered and calculated. It will play an important role in reducing the calculation time when users
require positioning over a large area, and where the fingerprint database is huge.

31250
Sensors 2015, 15, 31244–31267

4. Novel Fusing Positioning Algorithm

4.1. Algorithm Framework for Indoor Continuous Positioning on the Move

The overall algorithm can be divided into four modules: the motion pattern recognition module,
the Wi-Fi positioning module, the PDR positioning module, and the Trust Chain Positioning Fusion
module. They are shown as dashed rectangles with rounded edges in Figure 4. The inputs of the
Sensors 2015, 15, page–page
algorithm include acceleration, magnetic force, angular acceleration, pressure and Wi-Fi signals,
which
which all can be
all can beobtained
obtainedreal
realtime
time
byby a smartphone.
a smartphone. The The
Wi-FiWi-Fi fingerprint
fingerprint database
database is builtisup
built up
during
during the offline
the offline phase. phase.

Figure4.4.Algorithm
Figure Algorithmframework
frameworkof
ofindoor
indoorcontinuous
continuouspositioning.
positioning.

The motion pattern recognition module fuses the sensor data from the acceleration, gyroscope
The motion pattern recognition module fuses the sensor data from the acceleration, gyroscope
and barometric to identify the motion pattern of users in order to adopt different positioning
and barometric to identify the motion pattern of users in order to adopt different positioning
algorithms. The Wi-Fi positioning module uses the improved Wi-Fi positioning algorithm described
algorithms. The Wi-Fi positioning module uses the improved Wi-Fi positioning algorithm described
in Section 3 to get K most possible positioning estimations. The PDR positioning module fuses the
in Section 3 to get K most possible positioning estimations. The PDR positioning module fuses
sensor data from the acceleration, gyroscopes and magnetometers to realize indoor users’
the sensor data from the acceleration, gyroscopes and magnetometers to realize indoor users’
displacement estimation. The TCPF module includes three parts, trusted point determination, trust
displacement estimation. The TCPF module includes three parts, trusted point determination,
state and positioning fusion algorithm. It determines the trusted points by matching Wi-Fi
trust state and positioning fusion algorithm. It determines the trusted points by matching Wi-Fi
positioning results with the annular sector set up by PDR positioning. A finite state machine is used
positioning results with the annular sector set up by PDR positioning. A finite state machine is used
to establish the dynamic states of the continuous positioning locking mechanism. Finally, Wi-Fi and
to establish the dynamic states of the continuous positioning locking mechanism. Finally, Wi-Fi and
PDR positioning results are fused based on a dynamic positioning fusion algorithm to get the best
PDR positioning results are fused based on a dynamic positioning fusion algorithm to get the best
positioning estimation.
positioning estimation.
Multi-sensor data can be used to recognize motion patterns, such as walking with hand-held
Multi-sensor data can be used to recognize motion patterns, such as walking with hand-held
smartphones, walking with hand swinging, static standing, and so on; in total, six of the most
smartphones, walking with hand swinging, static standing, and so on; in total, six of the most
common motion patterns. The recognition accuracy can reach 95% [30]. The upstairs and downstairs
common motion patterns. The recognition accuracy can reach 95% [30]. The upstairs and downstairs
motion can also be identified with the assistance of a barometer [17]. For simplicity, the details of this
motion can also be identified with the assistance of a barometer [17]. For simplicity, the details of this
part are omitted. The Wi-Fi positioning module is explained as Section 3, and the following focuses
part are omitted. The Wi-Fi positioning module is explained as Section 3, and the following focuses
on the other two modules of the algorithm.
on the other two modules of the algorithm.
4.2. PDR Positioning Algorithm
The PDR positioning module monitors the walking action of the user, then estimates their step
length and orientation to estimate the displacement so that positioning estimation can be realized,
expressed as Equation (7).

t|t 1  cos t 
Lt|t 1  Lt 1|t 1  L31251
PDR =Lt 1|t 1  lt   (7)
 sin t 

 X t 1,Yt 1 
T
where Lt 1|t 1 is the latest positioning result at epoch t , Lt|t 1 is the PDR
Sensors 2015, 15, 31244–31267

4.2. PDR Positioning Algorithm

The PDR positioning module monitors the walking action of the user, then estimates their step
length and orientation to estimate the displacement so that positioning estimation can be realized,
expressed as Equation (7).
˜ ¸
t | t ´1 cosθt
Lt|t´1 “ Lt´1|t´1 ` L4 PDR =Lt´1|t´1 ` lt (7)
sinθt

where Lt´1|t´ 1 is the latest positioning result pXt´1 , Yt´1 qT at epoch t, Lt|t´1 is the PDR positioning
result, lt is the estimated step length, and θt is the estimated orientation.
It is important to note that the main researches are based on the hypothesis that smartphones
keep a specific attitude in the positioning process. Main attitudes under consideration include
hand-held [11,31], kept in a pocket [31], or tied on the waist [32], and so on. The real-time coordinate
transformation of the built-in sensor data is an extremely complex process, even without a solution,
because the position and the attitude of smartphones are random and vary over time. It is assumed
that smartphones maintain a hand-held attitude during the whole process of positioning, while it is
assumed in this paper that users keep watching the phone at the same time. It is the most common
attitude when smartphones are used for navigation, and the orientation of the smartphone (y axis)
stays the same with the user’s.
The PDR positioning algorithm can be divided into three parts: the step monitoring, the step
length estimation, and the orientation estimation. The step monitoring algorithm uses the rising
edge of the average acceleration, assisted by the adjacent step time difference limitation to reduce
misjudgment [11], and to identify every step. The step length estimation algorithm under walking
conditions is shown in Equation (8).

1
` ˘
lt “ Kw av,max ´ av,min 4 (8)

where lt is the step length, av,max ´ av,min is the peak-to-peak value of vertical acceleration av
during each step, Kw is a coefficient calibrated for individuals. The step length is a personalized
parameter [31], and has correlation with user’s height, leg length, weight, and habits. Higher accurate
step length estimation needs further personalized correction fusing user’s historical trajectory [11].
The orientation estimation algorithm fuses sensor data from an electronic compass and a
gyroscope by Kalman filter to get the best estimation [31,32]. The estimation algorithm is based
on the hand-held attitude, a specific optimization algorithm is needed in other attitudes [33]. The
angle, angular velocity and angular acceleration at epoch t is labeled as Qt , Vt , and ut , respectively.
The orientation variable is defined as St “ rQt , Vt sT . According to the Newton’s theorem, the system
transfer equation is shown in Equation (9).

St “ ASt´1 ` But ` w (9)

« ff « ff
1 dt dt2 {2
where A “ and B “ are coefficients, and w denotes the Gaussian noise of
0 1 dt
the system with zero mean and variance φ. The observation of the system comes from the azimuth
reading of the compass, which is defined as Ot . The observation function can be expressed as
Equation (10).
Ot “ CSt ` r (10)

where C “ r1 0s and r denotes the Gaussian noise of the magnetometer output with zero mean and
variance ϕ. The fusion equation based on Kalman filter is expressed as Equations (11) and (12).

31252
Sensors 2015, 15, 31244–31267

Predicting
St|t´1 “ ASt´1|t´1 ` But
(11)
Pt|t´1 “ APt´1|t´1 A T ` φ
Updating
´ ¯ ´1
Kt “ Pt|t´1 C T CPt|t´1 C T ` ϕ
´ ¯
St|t “ St|t´1 ` Kt Ot ´ CSt|t´1 (12)
Pt|t “ pI ´ Kt Cq Pt|t´1
Sensors 2015, 15, page–page
PDR Positioning Error Analysis
The PDR positioning results considering errors can be expressed as Equation (13).
 cos   t   ¸
Lt|t 1  Lt 1|t 1   lt  lt   ˜ t  q (13)
Lt|t´1 “ Lt´1|t´1 ` plt ` 4lt qsin t  t  
cos pθt ` ∆θ t
(13)
sin pθt ` ∆θt q
where the main source of the PDR positioning error is the step length estimation error lt and the
where the main source of the PDR positioning error is the step length estimation error 4lt and the
orientation estimation error Δ𝜃𝑡 . When the user is normally walking indoors, the step length can be
orientation estimation error ∆θt . When the user is normally walking indoors, the step length can
regarded as a constant so that lt can be regarded as a constant much less than 𝑙𝑡 . The PDR
be regarded as a constant so that 4lt can be regarded as a constant much less than lt . The PDR
positioning
positioningerror
errorhas
hasaacontinuity
continuityproperty
propertywithin
withinaashort
shortdistance, becauseΔ𝜃
distance,because ∆θ𝑡 t isissatisfied
satisfiedwith
withaa
Gaussian
Gaussiandistribution
distributionfrom
from the
theperspective
perspective of
of the
thestatistics,
statistics,which
whichwill
willbe
bevalidated
validatedin inSection
Section5.1,
5.1,
and
and the orientation variation is continuous and gradual. The PDR positioning error estimationisis
the orientation variation is continuous and gradual. The PDR positioning error estimation
shown
shownin inFigure
Figure5,5,and
andthe
themathematical expectation EEPDR
mathematicalexpectation PDR isisshown
shownininEquation
Equation(14).
(14).

PDRpositioning
Figure5.5.PDR
Figure positioningerror
errorestimation.
estimation.

  2 

1  t 
 2 2    

E PDR =2 2 ż π 1 e  ´p
EPDR =2 2
0 2 ?2
 2ltt qsin  t ˆ d∆θ˙t
 ∆θ 2

¨ e 2σ2 ¨ 2lt ¨ sin 2  t d∆θt

(14)
(14)
0 2πσ 2 2

4.3. Trust Chain Positioning Fusion Algorithm

4.3.1. Algorithm Flow

The positioning fusion algorithm is the key to take respective advantage of the Wi-Fi positioning
and the PDR positioning. The algorithm flow of the proposed fusion algorithm, which is named Trust
Chain Positioning Fusion (TCPF) algorithm, is shown in Figure 6. The TCPF algorithm includes three
31253
parts, trusted point determination, trust state, and positioning fusion. The first step is to determine
the trusted point by matching the Wi-Fi positioning result with the annular sector set up by the PDR
positioning. The next step is to update the trust state defined by a finite state machine and the TC
 Sensors 2015, 15, 31244–31267

4.3. Trust Chain Positioning Fusion Algorithm

4.3.1. Algorithm Flow

The positioning fusion algorithm is the key to take respective advantage of the Wi-Fi positioning
and the PDR positioning. The algorithm flow of the proposed fusion algorithm, which is named Trust
Chain Positioning Fusion (TCPF) algorithm, is shown in Figure 6. The TCPF algorithm includes three
parts, trusted point determination, trust state, and positioning fusion. The first step is to determine
the trusted point by matching the Wi-Fi positioning result with the annular sector set up by the PDR
positioning. The next step is to update the trust state defined by a finite state machine and the TC
parameter based on the trust determination results. The final step is to fuse two positioning results
based on the trust state and the TC parameter. Compared to the traditional Kalman Filter that directly
fuses two positioning results, the proposed algorithm realizes the “process-level” fusion.
Sensors 2015, 15, page–page

Start

Wi-Fi
Positioning

Initialization
Starting Point
Initialization

Wi-Fi PDR
Positioning Positioning

K Nearest Estimated
Location
Candidates & Trusted Area

Trust Chain
Positioning Fusion
Trusted Point
Trust State
Determination

Positioning TC
Fusion parameter
Update State

Wi-Fi/PDR Fused
Positioning Result

Figure 6. Algorithm flow of proposed fusing positioning.

Figure 6. Algorithm flow of proposed fusing positioning.

4.3.2. Trusted Point Determination

Trusted point determination is the method to 31254
recognize the accurate positioning results, and it
is the key to set up the trust chain. The key to improve the PDR/Wi-Fi fusion positioning accuracy is
to periodically eliminate the cumulative error. However, estimating positioning error at each point is
really difficult. The innovative idea is identifying trusted points instead. The key idea behind the
Sensors 2015, 15, 31244–31267

4.3.2. Trusted Point Determination

Trusted point determination is the method to recognize the accurate positioning results, and it
is the key to set up the trust chain. The key to improve the PDR/Wi-Fi fusion positioning accuracy
is to periodically eliminate the cumulative error. However, estimating positioning error at each point
is really difficult. The innovative idea is identifying trusted points instead. The key idea behind
the TCPF algorithm is using the accurate point to support the whole positioning process in order to
improve the positioning performance.
The trusted point is determined if the Wi-Fi positioning result falls within the annular sector
set up by the PDR positioning. The annular sector is named the trusted area. It is based on the
high accuracy
Sensors of the PDR positioning over a a short distance. If the two positioning results meet the
2015, 15, page–page
conditions in Equation (15), the real location located in the trusted area with high probability and
accuracy. The trusted area is shownin 7 withitsmathematic expression as Equation (15).
 Dist
Figure
Dist p w
 # ˇ (15)
 ˇ p w´  
ˇ
 Dirp Dir
Dist Dist wˇ ď ρ
ˇ
ˇ (15)
Dir p ´ Dirw ď β
ˇ ˇ
where Distw and Dirw is respectively the distance and angle between the Wi-Fi positioning
where Dist and Dirw is respectively the distance and angle between the Wi-Fi positioning results
results and w the latest positioning results. Dist p and Dirp are referred to as the PDR positioning.
and the latest positioning results. Dist p and Dir p are referred to as the PDR positioning. β is the
tolerance
is the tolerance
parameterparameter of the orientation
of the orientation estimationestimation
error, set aserror, set as the experimental
the experimental value
value 45˝ , and ρ is45°,
the
and 𝜌 is the
tolerance toleranceofparameter
parameter of the step
the step length length error,
estimation estimation
set aserror, set as the experimental
the experimental value 2 m. value 2 m.

Figure
Figure7.7.The
Thetrusted
trustedarea
areaset
setup
upby
byPDR
PDRpositioning.
positioning.

4.3.3.
4.3.3.Trust
Trust State
State
Trust
Trust state
state is
is established
established to to mark
mark the
the state
state of
of the
the trust
trust chain,
chain, and
and isis realized
realized byby aa finite
finite state
state
machine
machine shown as Figure 8. Three states are defined: Trusted Locked State, Locked State and Unlocked
shown as Figure 8. Three states are defined: Trusted Locked State, Locked State and Unlocked
State.
State. The
TheTrusted
TrustedLocked
Locked State
State isis set
set when
when current
current point
point isis determined
determined as as the
the trusted
trusted point.
point. AtAtthis
this
time, the positioning result is regarded as highly accurate, and the trust chain
time, the positioning result is regarded as highly accurate, and the trust chain is locked. Until the is locked. Until the
positioning
positioning point
point isis no
no longer
longer consider
consider thethe trusted
trusted point,
point, the
the state
state changes
changes to to the
the Locked
Locked State.
State. At
At
this
this time,
time, the
thetrust
trustchain
chainlocks
locksthethecurrent
currentpoint
pointwith
withthe
thelatest
latesttrusted
trustedpoint
pointso sothat
that the
thepositioning
positioning
still
stillmainly
mainlyrelies
reliesononthe
thePDRPDRpositioning
positioning result
resultdue
dueto to
its its
high accuracy
high over
accuracy a short
over range.
a short The Wi-
range. The
Fi positioning candidate points are filtered by the trust area. If the point remains untrusted to a certain
extent, the state would change to the Unlocked State. At this time, the Wi-Fi positioning is mainly relied
on, and its candidate points are not filtered. As long 31255as the current positioning point is determined as
the trusted point, the state is change back to the Trusted Locked State, forming a full loop.
Sensors 2015, 15, 31244–31267

Wi-Fi positioning candidate points are filtered by the trust area. If the point remains untrusted to a
certain extent, the state would change to the Unlocked State. At this time, the Wi-Fi positioning is
mainly relied on, and its candidate points are not filtered. As long as the current positioning point
is determined as the trusted point, the state is change back to the Trusted Locked State, forming a
Sensors 2015, 15, page–page
full loop.

Figure 8.
Figure Finite state
8. Finite state machine
machine of
of trust
trust state.
state.

Where λt tisisthe

theTCTCparameter,
parameter, which
which isisintroduced
introducedandandanalyzed
analyzedininthe
thenext
nextsection, andε is the
section,and is
determination factor to change the state from the Locked State to the Unlocked
the determination factor to change the state from the Locked State to the Unlocked State. The State. The parameter
ε characterizes
parameter  characterizes
the locking extent from the
the locking latest
extent trusted
from pointtrusted
the latest to the point
current positioning
to the point. The
current positioning
point. The larger  is, the more trust is given to the PDR positioning based on the trusted point,
larger ε is, the more trust is given to the PDR positioning based on the trusted point, and vice versa.andIn
the paper, the empirical value 1 is selected so that two positioning algorithms is
vice versa. In the paper, the empirical value 1 is selected so that two positioning algorithms isbalanced equivalently.
balanced equivalently.
4.3.4. Positioning Fusion
4.3.4.The positioning
Positioning fusion estimation is the dynamic weighted sum of the PDR positioning
Fusion
estimation and the improved Wi-Fi positioning estimation, as shown in Equation (16). The PDR
The positioning fusion estimation is the dynamic weighted sum of the PDR positioning
positioning is based on the built-in sensors data and the Wi-Fi positioning module is based on the
estimation and the improved Wi-Fi positioning estimation, as shown in Equation (16). The PDR
RSSI of the Wi-Fi. They are independent of each other so that they can be regarded as the unrelated
positioning is based on the built-in sensors data and the Wi-Fi positioning module is based on the
positioning estimations for the linear weighted fusion.
RSSI of the Wi-Fi. They are independent of each other so that they can be regarded as the unrelated
positioning estimations for the linear weighted
Lt|t´fusion. t
1 ` λt|t´1 LWi´ Fi
Lt|t “ (16)
Lt|t 11`λt |tt |t1´L1tWi  Fi
Lt|t  (16)
t 1   
where Lt|t-1 and LWi´Fi is respectively the PDR and Wi-Fi positioning result at epoch t, and Lt|t is the
t |t 1

positioning fusion result

t pXt , Yt qT at epoch t. λt|t´1 is the key weighting parameter, which is named
where Lt|t-1 and LWi  Fi is respectively the PDR and Wi-Fi positioning result at epoch t, and Lt|t is
the TC parameter.
 Xalgorithm
t , Yt  t . parameter,
T
The first core
the positioning issue
fusion of the
result is the TC
at epoch which is used to characterize the
t |t 1 is the key weighting parameter, which is
inaccuracy (or mistrust) of the positioning point. On one hand, it is defined as increasing linearly
named the TC parameter.
as the positioning target moves according to the continuity property of the PDR positioning error,
The first core issue of the algorithm is the TC parameter, which is used to characterize the
as Equation (17).
inaccuracy (or mistrust) of the positioning point. On one hand, it is defined as increasing linearly as
λt|t´1 “ λt´1|t´1 ` d PDR ¨ α (17)
the positioning target moves according to the continuity property of the PDR positioning error, as
Equation
where (17).
d PDR is the distance target moves during epoch t ´ 1 an t, and α is a constant coefficient, which
is the ratio of the accumulated error of the PDR positioning moving one meter and the Wi-Fi average
t|t 1  t 1|t 1  d PDR   (17)
positioning error. The Wi-Fi positioning error is regarded as a Gaussian random error so that the
average
where disPDRset is
asthe
thedistance
metric of the inaccuracy.
target moves during Onepochthe other t  1hand,
and the
t , and  is a constant
TC parameter is updated after
coefficient,
fusing the Wi-Fi positioning result. When the positioning point is no longer considered as the trusted
which is the ratio of the accumulated error of the PDR positioning moving one meter and the Wi-Fi
average positioning error. The Wi-Fi positioning error is regarded as a Gaussian random error so that
the average is set as the metric of the inaccuracy. On the other hand, the TC parameter is updated
after fusing the Wi-Fi positioning result. When the positioning point is no longer considered as the
trusted point, the TC parameter is iterated after31256 fusion, as Equation (18). The inaccuracy of the fused
result is calculated as the (1  t|t 1 ) times the original accumulated inaccuracy t|t 1 .

t|t  t|t 1 (1  t|t 1 ) (18)

Sensors 2015, 15, 31244–31267

point, the TC parameter is iterated after fusion, as Equation (18). The inaccuracy of the fused result is
calculated as the p1 ` λt|t´1 q times the original accumulated inaccuracy λt|t´1 .

λt|t “ λt|t´1 p1 ` λt|t´1 q (18)

Otherwise, the trust state is changed back to Trusted Locked State, and the TC parameter is
changed to zero.
The second core issue of the algorithm is the relationship between the trust state and the Wi-Fi
t
positioning. The Wi-Fi positioning result LWi ´ Fi is defined as follows.

1. At the locked state, including Trusted Locked State and Locked State, the Wi-Fi positioning
candidate points are limited within the trust area, and LWit
´ Fi is the weighted sum of these
selected points’ locations.
2. At the Unlocked State, the Wi-Fi positioning is not filtered, and the candidate points are all
t
weighted added as the positioning fusion estimation LWi ´ Fi .

Taking overall Equation (7) and Equations (16)´(18) into account, the proposed TCPF algorithm is
expressed as Equation (19).

Lt|t´1 “ Lt´1|t´1 ` Lt|t-1

∆PDR
λt|t´1 “ λt´1|t´1 ` d PDR ¨ α
¯ Lt|t-1 ` λ t
λ t | t ´1 ´ t t|t´1 LWi-Fi
Lt|t =Lt|t-1 ` LWi-Fi ´ Lt|t´1 “ (19)
1 ` λt|t´1 1 ` λ t | t ´1
λt|t “ 0, i f Trusted Locked State
λt|t “ λt|t´1 p1 ` λt|t´1 q, otherwise

Compared to the traditional fusion algorithm, the proposed TCPF algorithm has the following
two significant features.

1. The trusted point determination method and the trust state machine are established to
multi-dimensionally adjust the weight of the dynamic positioning fusion so that the fused result
is the optimal estimation.
2. The algorithm has a strong anti-interference performance without error cycles. At the locked
state, the TC parameter is set as a small value so that the fused positioning result is focused on
the PDR positioning, which is highly accurate over a short distance. At the unlocked state, the
TC parameter increases rapidly iteratively so that the fused positioning result is focused on the
Wi-Fi positioning. The weight parameter is dynamically adjusted so that there is no error cycle.

5. Experimental Evaluation

5.1. Experimental Setup

The continuous positioning on the move experiment (called Experiment 2) is set up to evaluate
the performance of the proposed fusing positioning algorithm, including the improved Wi-Fi
positioning, the PDR positioning, and the novel TCPF algorithm.
In order to sufficiently test positioning performance, including turning on the move, the
experimental field is changed to the 4th floor in the LiJiao Building, Peking University, which is a
typical modern teaching building with a larger area than where Experiment 1 was performed. The
layout of the experiment area is shown in Figure 9. Altogether, 30 APs, marked as hollow dots, are
evenly distributed in the experimental area, which is about 430 m2 . Altogether, 184 fingerprint points,
marked as black and red solid dots, are evenly distributed and the adjacent interval, aligned at the
floor tile, is 1.6 m. The moving path for testing, marked as a red line and red points, includes 20 test
points with a length of 60.6 m.

31257
Sensors 2015, 15, page–page

Sensors 2015, 15, 31244–31267

Sensors 2015, 15, page–page

Figure 9. The layout of the experimental area.

Figure 9.
Figure The layout
9. The layout of
of the
the experimental
experimental area.
area.

The Experiment 2 is divided into two parts: the static positioning part and the continuous
The
The Experiment
positioning on the 2
Experiment is divided
move into
part. In the two part,
former parts:thethe
the static
static
static positioning
positioning
test data is sampled part
part and
and the
10 times the continuous
continuous
at each test
positioning
positioning on the
onthe
point, and move
movepart.
thesampling part. InIn
intervalthethe
is former
0.5 former part,
s. In the thethe
part,
latter static
part, test
static
the data
test
tester is sampled
data
walks is
at sampled10 times
a constant 10 atalong
times
speed each test
at each
point,the moving
and the
test point, path,
andsampling and starts
interval
the sampling Wi-Fi positioning
is 0.5iss. 0.5
interval In the once
latter
s. In passing
part, the
the latter a test
part, point.
tester
the walks The built-in sensor
at a constant
tester walks data is
speed along
at a constant speed
the
along automatically
moving path, and
the moving recorded
starts
path, andinWi-Fi
the background
starts positioning during
Wi-Fi positioning the
once passing
onceentirea moving
test point.
passing process,
a test The
point. andThe
thebuilt-in
built-in sampling
sensor data
sensoris
interval is recorded
automatically set at 50 ms. inThe
themoving test is repeated
background during 10
thetimes, andmoving
entire the orientation
process, of the
andsmartphone
the sampling
data is automatically recorded in the background during the entire moving process, and the sampling
conforms with the user’s all the time. The attitude of the smartphone and the tester’s gesture are
interval
interval is
is set
set at
at 5050 ms. The
The moving
moving test test isis repeated
repeated 10 10 times,
times, and and the
the orientation
orientation of of the
the smartphone
smartphone
shown in Figure 10. The train data is the same for the two parts, and is sampled 10 times at each
conforms
conforms with the user’s
user’s all
all the time. The The attitude
attitude of
of the
the smartphone
smartphone and
and the
the tester’s
tester’s gesture
gesture areare
fingerprint point with 0.5 s intervals.
shown
shown in in Figure 10. The train data is the same for the two parts, and is
The train data is the same for the two parts, and is sampled 10 times at eachsampled 10 times at each
fingerprint
fingerprint point with 0.5 s intervals.

Figure 10. Attitude of the smartphone and the user’s gesture in the experiment.

Figure 10. Attitude of

10. Attitude of the
the smartphone 15
smartphone and
and the
the user’s
user’s gesture
gesture in
in the
the experiment.
experiment.
Figure

15
31258
Sensors 2015, 15, 31244–31267
Sensors 2015, 15, page–page
Sensors 2015, 15, page–page
TheThefollowing
following analysis
analysisisisdivided
dividedinto
intothree
three parts,
parts, including
including thethe improved
improvedWi-Fi
Wi-Fipositioning
positioning
The following
evaluation,
evaluation, thethe
PDRPDR analysis
positioning is divided
positioning intoand
evaluation
evaluation three
and parts,
the
the fusing
fusing including the evaluation.
positioning
positioning improved Wi-Fi positioning
evaluation.
evaluation, the PDR positioning evaluation and the fusing positioning evaluation.
5.2.5.2.
Improved
ImprovedWi-Fi
Wi-FiPositioning
Positioning Evaluation
Evaluation
5.2. Improved Wi-Fi Positioning Evaluation
Firstly,
Firstly, a positioning
a positioning performancecomparison
performance comparisonat at stationary
stationary andand moving
movingstates
stateswas
wascarried
carriedout.
out.
Firstly, a positioning performance comparison
wereat stationary and moving states was carried out.
TheTheNNNN andand WKNN
WKNN positioning
positioning algorithms
algorithms were adopted.
adopted. Compared
Comparedtotothe thepositioning
positioningerror
erroratat
The NN and
stationary WKNN
state, error positioning algorithms were adopted. Compared to theusingpositioning error at
stationary state, error ononthethemove
moveincreased
increasedbyby 141%
141% and 133%,
and 133%, respectively,
respectively, usingNN NNand
andWKNN
WKNN
stationary
positioning state, error on The
algorithms. the move increased
average by 141%
positioning and
error at 133%,
movingrespectively, using NN and WKNN
positioning algorithms. The average positioning error at moving state
staterespectively
respectivelydeteriorated
deterioratedtoto
positioning algorithms. Thecomparison
average positioning error at moving state respectively deteriorated to
5.355.35 m and
m and 4.274.27
m,m,and andthethe
comparison isisshownshownin in Figure
Figure 11.
11.
5.35 m and 4.27 m, and the comparison is shown in Figure 11.

Figure Wi-Fi
11.11.
Figure positioning
Wi-Fi positioningperformance
performanceat
atmoving/stationary states.
moving/stationary states.
Figure 11. Wi-Fi positioning performance at moving/stationary states.
Secondly, we adjusted the RSSI threshold step by step to find the variation tendency of the
Secondly, wewe
Secondly, adjusted
adjustedthethe
RSSI threshold
RSSI step by by
threshold step to find the the
variation tendency of the Wi-Fi
Wi-Fi positioning error at stationary and movingstep step
states, as to find
shown variation
in Figure 12. tendency of the
positioning error at error
Wi-Fi positioning stationary and moving
at stationary states, states,
and moving as shown in Figure
as shown 12. 12.
in Figure

Figure 12. Wi-Fi positioning performance with RSSI threshold.

Figure12.
Figure 12.Wi-Fi
Wi-Fipositioning
positioning performance
performance with
withRSSI
RSSIthreshold.
threshold.
16
16
31259
Sensors 2015, 15, 31244–31267
Sensors 2015, 15, page–page

Two
Twoconclusions
conclusionscan
canbe
besummarized
summarizedfrom
fromthe
thetendency
tendencyofofthe
thepositioning
positioningerrors.
errors.
1.1. For
ForWi-Fi
Wi-Fistationary
stationarypositioning,
positioning,weak
weaksignals
signalsless
lessthan
than´75−75 dBm
dBmalso
alsocontributed
contributedto topositioning
positioning
so
so that
that the
the excessive
excessive threshold
threshold reduced
reduced signal
signal information
information andand increased
increased the
the positioning
positioning error.
error.
Overall,
Overall, Wi-Fi
Wi-Fistatic
staticpositioning
positioningachieved
achievedaastrongest
strongestperformance
performanceatataround
around´85−85 dBm.
dBm.
2.2. For Wi-Fi
For Wi-Fimoving
movingpositioning,
positioning,the
thepositioning
positioningerrorerrorfirstly
firstlydecreased
decreasedas asthe
thethreshold
thresholdincreased,
increased,
which was
which was between
between ´90−90 and
and ´75
−75 dBm. Then, Then, asas the
the threshold increased continuously,
continuously, thethe
positioning error increases instead.
positioning instead.ItItcan canbebe
explained
explained that weak
that signals
weak lessless
signals thanthan
−75 dBm
´75 cause
dBm
a decline
cause in performance
a decline due due
in performance to the great
to the instability
great on on
instability thethe
move.
move.The
Thebest performance
best performanceis
obtained
is obtainedatataround
around−75 ´75dBm.
dBm.This
Thisisisininaccordance
accordancewithwiththethe properties
properties of Wi-Fi signals
signals on
onthe
the
move analyzed in Section
move analyzed in Section 3.1. 3.1.
Next, the
Next, the clustering
clustering property
property of
of the
the strongest
strongest AP
AP was
was validated
validated through
through thethe strongest
strongest AP
AP
distribution in
distribution in the
the experimental
experimental area,
area, as
as shown
shown inin Figure
Figure13.
13.We
Wefound
foundthat thethe
that distribution hashas
distribution an
extremely strong clustering property.
an extremely strong clustering property.

Figure 13.Strongest
Figure13. StrongestAP
APdistribution
distributionin
inexperimental
experimentalarea.
area.

Finally,the
Finally, theimprovement
improvementtotothetheclassic
classicpositioning
positioningalgorithm
algorithmby bythe
thetwo
twoproposed
proposedmethods
methodsisis
analyzed step by step, and the positioning accuracies and distributions that were compared
analyzed step by step, and the positioning accuracies and distributions that were compared are shown are
shown
in Figurein14.
Figure 14. The “WKNN
The “WKNN + THR” algorithm
+ THR” algorithm represents
represents the classicthe classicalgorithm
WKNN WKNN algorithm fusing
fusing with the
with the dynamically adjusting RSSI threshold method, and it fuses the AP matching
dynamically adjusting RSSI threshold method, and it fuses the AP matching method representing the method
representing the “WKNN + THR + AP” algorithm. The experimental results show that the proposed

31260
17
Sensors 2015, 15, 31244–31267
Sensors 2015, 15, page–page

“WKNN + THR + AP” algorithm. The experimental results show that the proposed two methods
two methods have a significant effect in improving the Wi-Fi positioning performance on the move.
have a significant effect in improving the Wi-Fi positioning performance on the move. The average
The average error of the improved Wi-Fi positioning was reduced to 2.68 m.
error of the improved Wi-Fi positioning was reduced to 2.68 m.

(a)

(b)
Figure 14.
Figure 14. Performance
Performance of
of improved
improved Wi-Fi
Wi-Fi positioning.
positioning. (a)
(a) Accuracy;
Accuracy; (b)
(b) distribution.
distribution.

31261
18
Sensors 2015,
Sensors 15,15,
2015, 31244–31267
page–page
Sensors 2015, 15, page–page
5.3. PDR Positioning Evaluation
5.3. PDR Positioning
5.3. PDR Evaluation
Positioning Evaluation
As the key part of the PDR positioning, the orientation estimation was firstly analyzed, and the
As the
comparisonAskey
the part
betweenofthe
key partthe PDR
the PDRpositioning,
of orientation
positioning,the
the
estimation orientation
orientation
and estimation
estimation
the true value at thewaswas firstly
firstly
hand-held analyzed,
analyzed,
attitude and and
the the
is shown
comparison
comparison between
between the the orientation
orientation estimationand
estimation andthe
the true
true value
value atatthe
thehand-held
hand-held attitude is shown
attitude is shown
in Figure 15.
in Figure
in Figure 15. 15.

Figure
Figure
Figure15.15.
15. Orientationestimation
Orientation
Orientation estimation and
estimation and
andtrue
truevalue.
true value.
value.

Through the statistical analysis of the orientation estimation error, the distribution approaches
Through
Through thethe statisticalanalysis
statistical analysisofofthe
theorientation
orientationestimation
estimation error,
error, the
the distribution
distribution approaches
approaches a
a Gaussian distribution, and the fitted result is shown in Figure 16.
a Gaussian distribution, and the fitted result is shown in Figure
Gaussian distribution, and the fitted result is shown in Figure 16. 16.

Figure 16. Gaussian fitting orientation measurement error.

Figure16.
16.Gaussian
Gaussian fitting
fitting orientation measurement error.
Figure
The Gaussian fitting correlation coefficientorientation measurement
is 0.9520, and error.
the fitted estimation error probability
distribution function e PDR is shown in Equation (20).
The Gaussian fitting correlation coefficient is 0.9520, and the fitted estimation error probability
The Gaussian fitting correlation coefficient is 0.9520, and the fitted estimation error probability
distribution function e PDR is shown in Equation (20).  ( 0 )2
distribution function ePDR is shown in Equation (20).
1
e PDR  2 2
e (20)
21 2 
( 0 )2
2
 pθ ´2θ0 q 2
PDR e 1 ´e (20)
where σ = 8.5916, 𝑥0 = −1.5086 .e The
PDR “ ? 2e
statistical 2probability
2σ2 is 94.00% when the orientation (20)
2πσ2
estimation error is within ±20°, 87.76% corresponding to within ±15°, and 71.13% corresponding to
where σ ±10°.
within = 8.5916, 𝑥0 = −1.5086 . The statistical probability is 94.00% when the orientation
where σ “ 8.5916,
estimation x0 within
error is “ ´1.5086. The statistical
±20°, 87.76% probability
corresponding is 94.00%
to within ±15°,when the orientation
and 71.13% estimation
corresponding to
error is within ˘20 ˝ , 87.76% corresponding to within ˘15˝ , and 71.13% corresponding to within ˘10˝ .
within ±10°.
Based on the fitted Gaussian distribution and19 Equation (14), the mathematical expectation of the
PDR positioning error is 8.9675 cm when the step length estimation d is estimated as 75 cm. It is the
important coefficient used to calculate the parameter 19 α in Equation (17).

31262
 Sensors 2015, 15, page–page

Based on the fitted Gaussian distribution and Equation (14), the mathematical expectation of the
Sensors 2015, 15, 31244–31267
PDR positioning error is 8.9675 cm when the step length estimation d is estimated as 75 cm. It is the
important coefficient used to calculate the parameter  in Equation (17).
In order
In order to to evaluatethe
evaluate theperformance
performance of the theonly
onlyPDR
PDRpositioning,
positioning, thethe
initial location
initial is setisasset
location theas
the starting
starting point.
point. The
Thepositioning
positioningresults
resultsand andaccumulated
accumulatederrors
errorsare
areshown
showninin Figure
Figure 17.
17. TheThe
accumulated error reaches 6.66 m. The experimental results show that the PDR positioningisis
accumulated error reaches 6.66 m. The experimental results show that the PDR positioning
extremely
extremely accurate
accurate over
over a short
a short distanceininthe
distance thelinear
linearmovement.
movement. Positioning
Positioning errors
errorsmainly
mainlyappear
appearinin
turnings,
turnings, because
because of the
of the orientation
orientation estimationerror.
estimation error.Overall,
Overall, the
the PDR
PDR positioning
positioningerror
errorvaries
variesalmost
almost
linearly over a short distance.
linearly over a short distance.

(a)

(b)
Figure 17. PDR positioning. (a) Distribution; (b) error distribution.
Figure 17. PDR positioning. (a) Distribution; (b) error distribution.

5.4. Fusing Positioning Evaluation

The PDR/Wi-Fi fusing positioning based on the
20 proposed algorithm is compared with the only
Wi-Fi positioning and the only PDR positioning, as shown in Figure 18. The result shows that the
fusion positioning performance is much better than the only Wi-Fi or PDR positioning. The average

31263
 Sensors 2015, 15, page–page

5.4. Fusing Positioning Evaluation

The PDR/Wi-Fi fusing positioning based on the proposed algorithm is compared with the only
Sensors 2015, 15, 31244–31267
Wi-Fi positioning and the only PDR positioning, as shown in Figure 18. The result shows that the
fusion positioning performance is much better than the only Wi-Fi or PDR positioning. The average
positioning
positioning errorerror of the
of the proposed
proposed algorithmisis1.36
algorithm 1.36 m,
m, aa decrease
decreaseofof49.6% compared
49.6% to that
compared of the
to that of the
only improved Wi-Fi positioning.
only improved Wi-Fi positioning.

Figure 18. PDR/Wi-Fi fusion positioning result distribution.

Figure 18. PDR/Wi-Fi fusion positioning result distribution.

The proposed
The proposed algorithm
algorithm demonstrates
demonstrates strong
strong anti-interference performance
anti-interference performance verified
verified by the
by the
behavior of correcting
behavior the the
of correcting PDR PDRaccumulated
accumulatederror error in
in the second
secondhalf
halfofofthe
thepath.
path.TheThe proposed
proposed
algorithm
algorithm can dynamically
can dynamically adjust
adjust thetheweight,
weight,andand take
take advantage
advantageofoftwo
two positioning
positioningalgorithms to to
algorithms
obtain a better positioning performance. The point with a red painted marker “★” is
obtain a better positioning performance. The point with a red painted marker “‹” is the trusted point, the trusted point,
and
and its its average
average positioning
positioning error
error is is 1.10m.
1.10 m.ItItplays
plays aa role
role as
asthe
the“anchor”
“anchor” throughout
throughout the the
positioning
positioning
process due to its higher positioning accuracy than other positioning points.
process due to its higher positioning accuracy than other positioning points.
The proposed algorithm was compared to the similar fusing positioning algorithm, which is the
The proposed algorithm was compared to the similar fusing positioning algorithm, which is
Cross-Assistive algorithm [12]. The positioning comparison results are shown in Figure 19. The
the Cross-Assistive
average positioning algorithm [12].mThe
error is 1.91 usingpositioning comparison
the Cross-Assistive results
algorithm withare
theshown in Figure Wi-
same improved 19. The
average positioning error is 1.91 m using the Cross-Assistive algorithm
Fi positioning algorithm, with 28.8% less accuracy than the proposed algorithm. with the same improved Wi-Fi
Sensors 2015, 15, page–page
positioning algorithm, with 28.8% less accuracy than the proposed algorithm.

21

Figure 19.Lateral
Figure19. Lateral comparison of proposed
comparison of proposedalgorithm.
algorithm.

6. Conclusions and Future Work

31264
This paper has proposed a novel indoor continuous positioning algorithm fusing built-in sensors
and Wi-Fi on smartphones to deal with positioning problems on the move, such as large errors and
poor stability, which cannot be well dealt with in traditional algorithms. Compared to the traditional
Sensors 2015, 15, 31244–31267

6. Conclusions and Future Work

This paper has proposed a novel indoor continuous positioning algorithm fusing built-in sensors
and Wi-Fi on smartphones to deal with positioning problems on the move, such as large errors
and poor stability, which cannot be well dealt with in traditional algorithms. Compared to the
traditional algorithms, the main innovative points include an improved Wi-Fi positioning algorithm
and a novel positioning fusion algorithm called the TCPF algorithm. The improved Wi-Fi positioning
algorithm enhances the Wi-Fi positioning performance on the move by dynamically adjusting the
RSSI threshold and AP matching. It is based on the properties of Wi-Fi signals at moving state,
which are analyzed through a novel “quasi-dynamic” Wi-Fi signal experiment. The TCPF algorithm
is proposed to realize the “process-level” fusion of Wi-Fi and PDR positioning. The algorithm is
described in detail in three parts: Trusted point determination, trust state, and positioning fusion.
The mathematical calculation model is also summarized and verified by a field experiment, which
was carried out in a typical teaching building. The result shows that the average positioning error
on the move using the improved Wi-Fi positioning algorithm decreases by 49.8% compared to the
traditional algorithm. The average positioning error of the novel fusing positioning algorithm reaches
1.36 m, a decrease of 28.8% compared to a current fusion algorithm. The proposed algorithm can
effectively reduce the influence caused by the unstable Wi-Fi signals, eliminating the positioning error
accumulated over time, and improving the accuracy and stability of indoor continuous positioning
on the move.
There is still room for improvement in the algorithm. The main problem is that the algorithm
greatly relies on the smartphone’s attitude and the user’s gesture. Although the attitude under
consideration is among the most common one, the smartphone and user are flexible and their
behabiours vary in daily life. How to accurately recognize the various attitudes and adopt
corresponding compensation algorithms is the key to further increasing the universality of the
algorithm’s application. Other research directions include hardware differentiation, map matching,
and power control. The adaptation problem for different smartphones, especially for the varying
accuracies of built-in sensors, need to be further studied. The map matching technology application
has been mainly used in outdoor scenarios, but lacks application in indoor environmenst. Further
research could focus on motion recognition with indoor spatial analysis.

Acknowledgments: This work is supported by the Navigation and Location-based service (NAL) Lab,
Peking University.
Author Contributions: Huaiyu Li developed the main algorithm of the improved Wi-Fi positioning and the
sensor fusion. Xiuwan Chen and Guifei Jing supervised the work and complemented the algorithm. Yuan Wang
and Fei Li drew the indoor map and set up the experiment. Yanfeng Cao and Xinlong Zhang conducted the
experiment. Han Xiao revised the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. MIIT National Remote Sensing Center of China. Precision Indoor and Outdoor Positioning and Navigation.
Available online: http://www.nrscc.gov.cn/nrscc/upload/2015/826ab84c22044d999c522447d528b7bf138
0438703328.pdf (accessed on 9 November 2015).
2. Waqar, W.; Chen, Y.; Vardy, A. Incorporating user motion information for indoor smartphone positioning
in sparse Wi-Fi environments. In Proceedings of the 17th ACM International Conference on Modeling,
Analysis and Simulation of Wireless and Mobile Systems, Montreal, QC, Canada, 21–26 September 2014.
3. Bahl, P.; Padmanabhan, V.N. RADAR: An in-building RF-based user location and tracking system. In
Proceedings of the 19th Annual Joint Conference of the IEEE Computer and Communications Societies,
Tel Aviv, Israel, 26–30 March 2000.
4. Bahl, P.; Padmanabhan, V.N. Enhancements to the RADAR User Location and Tracking System. Available
online: http://research.microsoft.com/apps/pubs/default.aspx?id=69861 (accessed on 9 November 2015).

31265
Sensors 2015, 15, 31244–31267

5. Roos, T.; Myllymäki, P.; Tirri, H.; Misikangas, P.; Sievänen, J. A probabilistic approach to Wlan user location
estimation. Int. J. Wirel. Inf. Netw. 2002, 9, 155–164. [CrossRef]
6. Xiang, Z.; Song, S.; Chen, J.; Wang, H.; Huang, J.; Gao, X. A wireless LAN-based indoor positioning
technology. IBM J. Res. Dev. 2004, 48, 617–626. [CrossRef]
7. Youssef, M.; Agrawala, A. The Horus WLAN location determination system. In Proceedings of the 3rd
International Conference on Mobile Systems, Applications, and Services, Seattle, WA, USA, 6–8 June 2005.
8. Wu, C.-L.; Fu, L.-C.; Lian, F.-L. WLAN location determination in e-home via support vector classification.
In Proceedings of the 2004 IEEE International Conference on Networking, Sensing and Control, Taipei,
Taiwan, 21–23 March 2004.
9. Brunato, M.; Battiti, R. Statistical learning theory for location fingerprinting in wireless LANs. Comput.
Netw. 2005, 47, 825–845. [CrossRef]
10. Figuera, C.; Rojo-Alvarez, J.L.; Wilby, M.; Mora-Jiménez, I.; Caamaño, A.J. Advanced support vector
machines for 802.11 indoor location. Signal Process. 2012, 92, 2126–2136. [CrossRef]
11. Kakiuchi, N.; Kamijo, S. Pedestrian dead reckoning for mobile phones through walking and running
mode recognition. In Proceedings of the 16th International IEEE Conference on Intelligent Transportation
Systems: Intelligent Transportation Systems for All Modes, Hague, The Netherlands, 6–9 October 2013.
12. Miyazaki, K.; Mochizuki, M.; Murao, K.; Nishio, N. Cross-assistive approach for PDR and Wi-Fi
positioning. In Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous
Computing, Seattle, WA, USA, 13–17 September 2014.
13. Chang, Q.; Van de Velde, S.; Wang, W.; Li, Q.; Hou, H.; Heidi, S. Wi-Fi Fingerprint Positioning Updated
by Pedestrian Dead Reckoning for Mobile Phone Indoor Localization. In Proceedings of the China Satellite
Navigation Conference, Xi’an, China, 13–15 May 2015.
14. Chu, H.-J.; Tsai, G.-J.; Chiang, K.-W.; Duong, T.-T. GPS/MEMS INS data fusion and map matching in urban
areas. Sensors 2013, 13, 11280–11288. [CrossRef] [PubMed]
15. Evennou, F.; Marx, F. Advanced integration of WiFi and inertial navigation systems for indoor mobile
positioning. Eurasip J. Appl. Signal Process. 2006, 2006, 164. [CrossRef]
16. Wang, H.; Lenz, H.; Szabo, A.; Bamberger, J.; Hanebeck, U.D. WLAN-Based Pedestrian Tracking Using
Particle Filters and Low-Cost MEMS Sensors. In Proceedings of the Positioning, Navigation and
Communication, Hannover, Germany, 22–22 March 2007.
17. Chen, Z.; Zou, H.; Jiang, H.; Zhu, Q.; Soh, Y.C.; Xie, L. Fusion of WiFi, smartphone sensors and landmarks
using the Kalman filter for indoor localization. Sensors 2015, 15, 715–732. [CrossRef] [PubMed]
18. Deng, Z.-A.; Hu, Y.; Yu, J.; Na, Z. Extended Kalman Filter for Real Time Indoor Localization by Fusing WiFi
and Smartphone Inertial Sensors. Micromachines 2015, 6, 523–543. [CrossRef]
19. Yim, J.; Park, C.; Joo, J.; Jeong, S. Extended Kalman Filter for wireless LAN based indoor positioning. Decis.
Support Syst. 2008, 45, 960–971. [CrossRef]
20. Xiao, W.; Ni, W.; Yue, K.T. Integrated Wi-Fi fingerprinting and inertial sensing for indoor positioning. In
Proceedings of the 2011 International Conference on Indoor Positioning and Indoor Navigation (IPIN),
Guimaraes, Portugal, 21–23 September 2011.
21. Panyov, A.A.; Golovan, A.A.; Smirnov, A.S. Indoor Positioning Using Wi-Fi Fingerprinting, Pedestrian
Dead Reckoning and Aided INS. In Proceedigns of the 2014 International Symposium on Inertial Sensors
and Systems, Laguna Beach, CA, USA, 25–26 February 2014.
22. Malyavej, V.; Udomthanatheera, P. RSSI/IMU sensor fusion-based localization using unscented Kalman
filter. In Proceedings of the 20th Asia-Pacific Conference on Communication, Pattaya City, Thailand, 1–3
October 2014.
23. Yim, J.; Jeong, S.; Gwon, K.; Joo, J. Improvement of Kalman filters for WLAN based indoor tracking.
Expert Syst. Appl. 2010, 37, 426–433. [CrossRef]
24. Ali-Loytty, S.; Tommi, P.; Honkavirta, V.; Piche, R. Fingerprint Kalman Filter in indoor positioning
applications. In Proceedings of the 2009 IEEE International Conference on Control Applications, Saint
Petersburg, Russia, 8–10 July 2009.
25. Jin, M.; Koo, B.; Lee, S.; Park, C.; Lee, M.J.; Kim, S. IMU-Assisted Nearest Neighbor Selection for Real-Time
WiFi Fingerprinting Positioning. In Proceedings of the 2014 International Conference on Indoor Positioning
and Indoor Navigation, Busan, Korea, 27–30 October 2014.

31266
Sensors 2015, 15, 31244–31267

26. Kaemarungsi, K.; Krishnamurthy, P. Analysis of WLAN’s received signal strength indication for indoor
location fingerprinting. Pervasive Mob. Comput. 2012, 8, 292–316. [CrossRef]
27. Luo, J.; Zhan, X. Characterization of Smart Phone Received Signal Strength Indication for WLAN Indoor
Positioning Accuracy Improvement. J. Netw. 2014, 9, 739–746. [CrossRef]
28. Huang, J.; Millman, D.; Quigley, M.; Stavens, D.; Thrun, S.; Aggarwal, A. Efficient, generalized indoor
WiFi GraphSLAM. In Proceedings of the 2011 IEEE International Conference on Robotics and Automation
(ICRA), Shanghai, China, 9–13 May 2011.
29. Li, W.L.; Iltis, R.A.; Win, M.Z. A smartphone localization algorithm using RSSI and inertial sensor
measurement fusion. In Proceedings of the 2013 IEEE Global Communications Conference (GLOBECOM),
Atlanta, GA, USA, 9–13 December 2013.
30. Pei, L.; Chen, R.; Liu, J.; Chen, W.; Kuusniemi, H.; Tenhunen, T.; Kröger, T.; Chen, T.; Leppäkoski, H.;
Takala, J. Motion recognition assisted indoor wireless navigation on a mobile phone. In Proceedings of the
23rd International Technical Meeting of the Satellite Division of the Institute of Navigation 2010, Portland,
OR, USA, 21–24 September 2010.
31. Li, F.; Zhao, C.; Ding, G.; Gong, J.; Zhao, F. A Reliable and accurate indoor localization method using phone
inertial sensors. In Proceedings of the 14th International Conference on Ubiquitous Computing, Pittsburgh,
PA, USA, 5–8 September 2012.
32. Sharp, I.; Yu, K. Sensor-based dead-reckoning for indoor positioning. Phys. Commun. 2014, 13, 4–16.
[CrossRef]
33. Hoseinitabatabaei, S.A.; Gluhak, A.; Tafazolli, R. Towards a position and orientation independent approach
for pervasive observation of user direction with mobile phones. Pervasive Mob. Comput. 2015, 17, 23–42.
[CrossRef]

© 2015 by the authors; licensee MDPI, Basel, Switzerland. This article is an open
access article distributed under the terms and conditions of the Creative Commons by
Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).

31267