You are on page 1of 14

Downloaded from https://iranpaper.

ir
https://www.tarjomano.com https://www.tarjomano.com

IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022 23095

Smartphone-Based Pedestrian NLOS


Positioning Based on Acoustics and
IMU Parameter Estimation
Hucheng Wang , Student Member, IEEE, Can Xue , Zhi Wang, Member, IEEE,
Lei Zhang , Student Member, IEEE, Xiaonan Luo , Senior Member, IEEE,
and Xinheng Wang , Senior Member, IEEE

Abstract —This paper proposes an integrated positioning


algorithm for mobile devices and achieves long-term and
high-accuracy indoor pedestrian tracking under severe non-
Line-of-Sight (NLOS) scenarios. The traditional fusion method
for hand-held devices lacks zero-speed correction and cannot
clear the accumulated error of the Pedestrian Dead Reckon-
ing (PDR). Secondly, the PDR algorithm also requires user
privacy data for high positioning accuracy. Hence, we pro-
pose a customized model with acoustic and PDR through
self-updating parameters with two novel fusing strategies:
Kalman Filter with Least-Square (KFLS) and Kalman Filter with
Bayesian Parameter Estimation (KFBPE), which utilize numer-
ical feedback and Bayesian distribution, respectively. Experi-
ments with Huawei Mate 9 show that both methods above can
effectively eliminate the outlier resulting from severe signal
loss, regardless of hand-holding gestures, with no individual
privacy data required. Extensive experimental results demon-
strate that the proposed methods are more efficient for NLOS
and perform much better than the baselines of traditional
fusion frameworks like the standard Kalman Filter. KFBPE
has a relatively smoother tracking result, which guarantees
an average positioning accuracy of up to 25 cm under the
circumstance of nearly thirty percent acoustic signal loss
(or NLOS) at the same time.

Index Terms — Inertial measurement unit, non-line-of-sight,


signal loss, step length.

Manuscript received 11 May 2022; accepted 17 June 2022. Date of I. I NTRODUCTION


publication 29 June 2022; date of current version 30 November 2022.
This work was supported in part by the Bagui Scholar Program Fund of
Guangxi Zhuang Autonomous Region of China under Grant 2019A40,
in part by the National Key Research and Development Program under
I NDOOR positioning plays a vital role in the fields of
robotics, smart homes, healthcare, and generally emerging
sector of the Internet of Things (IoT) [1], [2]. Over the years,
Grant 2018AA100305 and Grant 2021YFB3900800, and in part by the
National Natural Science Foundation of China under Grant 6202780103,
two broad categories of indoor positioning technologies have
Grant 61936002, and Grant 62033001. The associate editor coordi- been developed. One is infrastructure-based, where anchor
nating the review of this article and approving it for publication was nodes have to be deployed to provide positioning services.
Dr. Valerie Renaudin. (Corresponding author: Zhi Wang.)
Hucheng Wang is with the School of Computer Science and Informa-
The other is infrastructure-less such as positioning based
tion Security, Guilin University of Electronic Technology, Guilin 541004, on geomagnetic field [3] or inertial sensors. Results from a
China, and also with the College of Control Science and Engineering, series of Microsoft Indoor Localization Competition [4]–[6]
Zhejiang University, Hangzhou 310027, China.
Can Xue and Zhi Wang are with the State Key Laboratory of Indus-
events show that infrastructure-based positioning typically has
trial Control Technology, Zhejiang University, Hangzhou 310027, China better localization accuracy than infrastructure-free ones. How-
(e-mail: zjuwangzhi@zju.edu.cn). ever, these two categories have their spotlights and flaws.
Lei Zhang is with the School of Construction Machinery, Chang’an
University, Xi’an 710064, China.
Infrastructure-based positioning such as laser radar (Lidar)
Xiaonan Luo is with the Computer and Information Security School, and ultra-wideband (UWB) are usually accurate, reaching
Guilin University of Electronic Technology, Guilin 541004, China. a centimeter-level accuracy. However, the cost of hardware
Xinheng Wang is with the School of Advanced Technology, Xi’an
Jiaotong-Liverpool University, Suzhou 215000, China.
and deployment of anchor nodes is high. Infrastructure-free
Digital Object Identifier 10.1109/JSEN.2022.3185248 positioning does not need the deployment of anchor nodes,

1558-1748 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23096 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

which has lowered production costs. However, centimeter-level absorbed, and the positioning deviation. Furthermore, to obtain
accuracy can not be guaranteed. For location-critical applica- the three-dimensional indoor positioning accurately, the TDoA
tions, such as positioning for people with a sight impairment algorithm requires at least four beacon-to-phone channels to be
or temporary isolation in COVID-19 [7], infrastructure-less unobstructed. However, the acoustic signal is often blocked in
positioning is not suitable, despite its low cost to deploy and complex indoor environments, making the TDoAs suffer great
easiness of operation. disturbance.
Due to the importance of location-critical applications [8], This paper focuses on mitigating the deviation of acoustic
[9] and the pressure of high cost, one often takes the com- localization by data fusion of acoustic localization and PDR
bination of these two categories of positioning systems into with its short-term high-precision characteristics by applying
one integrated solution for granted. One possible solution is a tight coupling approach at each step of the fusion procedure.
to utilize the Inertial Measurement Unit (IMU) sensors. Com- It is well-known that there are two kinds of PDR mod-
mercial off-the-shelf (COTS) intelligent terminal devices like els [14], the general model and the customized model, respec-
smartphones are becoming popular, so exploring smartphones’ tively. The former model’s performance is relatively poor,
indoor location is becoming potentially the best solution. while the latter has to use personal data, which might cause
Since the breakthrough work of acoustic ranging with smart- privacy issues. In [15], several customized models for step
phones from Microsoft Corporation [10], acoustic localization length (S L) were mentioned as Eq. (1), where l is leg length,
is experiencing a fast development, which can be found in T is step duration, h involves height and step frequency, the
work around the world [4]–[6] and our lab [11]. For research parameter K is derived from a default value to calibrate the
in recent years, the transmitted frequency band has shifted estimator for users. The first equation only takes into account
from audible to inaudible. The application scenarios are moved body proportions. In contrast, the second equation considers
from Line-of-Sight (LOS) to NLOS scenarios, the positioning the vertical accelerations avert during the step, and the range
target is moving dynamically, and the signal is moving dynam- a peak,di f f of the vertical accelerations. The third equation is
ically transmission has been changed from active broadcast another empiric method that is based on the correlation of
to passive reacting. These developments have made acoustic the vertical acceleration at the foot a f oot,vert during one step
positioning with high performance in complex environments. with the length of the step. When it comes to location-based
These advances can also be found in a series of indoor local- apps, collecting private data from users is bad. During the
ization competitions. Our lab’s results in the 2018 Microsoft fusion process, the acoustic localization is applied to reduce
Indoor Localization Competition showed acoustic localization the accumulated error of PDR and estimate the PDR general
could reach 70 cm in practical applications [5], which put us model’s individual feature coefficient.
at the first two places in the sound group. If augmented real- 
ity (AR) is implemented to support the localization, accuracy S L = K 2lh − h 2
 
of acoustic localization can reach 50 cm [12]. The positioning  N
 |a | K
S L = 0.1 
errors do not exceed 30 cm in the LOS scenario by deep vert,i
2.7 i=1

analyzing the positioning data. N T · a peak,di f f
In contrast, it rises to meter-level in the NLOS scenarios. 
N
To reduce the NLOS error, data fusion from the IMU and 3
i=1 |a f oot,vert,i |
S L = 0.98 (1)
acoustic localization was employed. The localization error N
could be maintained within 28 cm by applying the Extended
Kalman Filter (EKF) to eliminate the acoustic outliers under Meanwhile, by utilizing the PDR’s property of short-term
LOS and NLOS scenarios in [13]. However, under location- stability, the temporary jitter of acoustic localization is elimi-
critical scenarios, they lack systematic analysis of acoustic nated. Combining acoustic and PDR’s superiority can improve
outliers. An in-depth analysis and better solution are needed to overall positioning performance while all-time high-precision
ensure that the positioning system performs well under various positioning is achieved.
scenarios. Our approach: We have introduced two methods in this
In this paper, we have investigated near-ultrasonic Time paper to increase positioning precision, namely (i) the numer-
Difference of Arrival (TDoA)-based localization and IMU nav- ical coefficient in each step in Least-Square after Kalman Filter
igation for positioning under severe NLOS scenarios because fusion (KFLS), and (ii) probability distribution coefficient in
(i) near-ultrasonic is inaudible, which causes no interference each step for the Bayesian Parameter Estimation (KFBPE).
to human beings, (ii) NLOS scenario is more common than Then we consider outliers with a certain probability. Bayesian
LOS scenario, which almost exists everywhere, (iii) TDoA Parameter Estimation can smooth outliers with prior knowl-
eliminates dependence on precise synchronization that nei- edge, so KFBPE is proposed. Experiments show that under the
ther Android nor iOS phones support, (iv) the accuracy of condition of the acoustic position loss rate being higher than
short-term Pedestrian Dead Reckoning (PDR) is relatively 30%, the accuracy of up to 25 cm in mean square error (MSE)
reliable, and (v) it will be widely used rapidly because both is achieved. In summary, this paper has the following technical
technologies can be implemented in smartphones. contributions to knowledge, including:
In practice, relying on near-ultrasonic signals alone for • a complete real-time fusion positioning system that com-
positioning is unreliable because of the NLOS phenomenon bines the advantages of acoustic positioning and PDR,
and its properties of signals being easily interfered with, which is well-integrated into COTS smart devices,

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
WANG et al.: SMARTPHONE-BASED PEDESTRIAN NLOS POSITIONING BASED ON ACOUSTICS AND IMU PARAMETER ESTIMATION 23097

• KFLS and KFBPE methods are proposed to improve per- to replace the long-existing iBeacon [20], known for its low
sonal coefficients based on online self-learning, without power consumption. Some iOS devices support UWB and are
privacy issues, and said to be compatible with AirTags. The news may have a
• even if an external position loss rate is up to 30%, we can specific impact on the IPS industry.
still derive accurate positioning. Unlike RF-based positioning methods, optical-based IPSs
The remaining parts of this paper are organized as follows: are widely used in the industrial field. Representative tech-
Related work is described in Section II. Then we present our niques have been successfully developed and popularly
approach and provide the details for its implementation in applied, such as LIDAR and Visible Light Communications
Sections III and IV, respectively. Performance evaluation is (VLC). Wolfgang [21] presented the LIDAR in a Simultaneous
conducted in Section V. Finally, Section VI concludes this Localization and Mapping (SLAM) platform, which achieved
paper. real-time mapping and loop closure at a 5 cm resolution.
In practice, LIDAR is more expensive than UWB and is not
II. R ELATED W ORK supported on mobile phones, so it is unlikely to be widely used
Numerous Indoor Positioning Systems (IPS) have been by the public. Visible light is also a medium commonly used
proposed to meet the increasing demand for Location-Based for positioning. An image sensor-based indoor visible light
Services (LBS). Traditional indoor positioning technologies positioning system by Singular Value Decomposition (SVD)
can be generally classified into several categories: Radio and enhanced robustness was proposed in [22]. Tango [23]
Frequency (RF), optical, acoustic, and mechanical. Table. I leveraged the ubiquitous, always-on Augmented Reality (AR)
presents a summary of different IPS, comprehensively depict- glasses and collected camera data from the lenses. The
ing the accuracy, cost, coverage, and compatibility of portable- project was soon abandoned for privacy-preserving and power-
device. Unlike the Global Navigation Satellite System (GNSS) consumption concerns, which this paper will carefully deal
dominant technology used in open space, no such dominant with.
technology is used indoors. Each indoor localization technol- Mainly used in the COTS field, an acoustic-based position-
ogy has its inherent advantages and disadvantages. We will ing system (APS) has already become a research hotspot. APS
briefly describe their advantages and drawbacks in the rest of is generally known for its relatively low cost and broad cover-
this section. age with the most reliable compatibility in intelligent devices.
RF-based indoor positioning methods were once treated as Beepbeep [10] emitted the double-beep sounds during a rang-
the most promising IPSs for their strong compatibility and ing session by Round-Trip Time (RTT). Beepbeep artfully
relatively low cost. Among them, Wi-Fi undoubtedly claims avoided many sources of inaccuracy found in other typical ToA
the most attention, with many problems to overcome, in any schemes. BatTracker [25] was proposed to achieve the high
case. Chen [16] developed a large number of fingerprint precision and infrastructure-free mobile device tracking system
databases with Channel State Information (CSI) detection, in 3D space. BatTracker accommodates uncertainty based on
realizing the commercialization of indoor positioning in shop- a probabilistic algorithm for creating, pruning, and evolv-
ping malls and airports. Nevertheless, the accuracy of Chan’s ing multiple hypotheses. Speaker-to-Microphone (S2M) [30]
method still failed to break through the meter-level. Hence, was an acoustic fingerprint method to authenticate devices.
it can hardly meet the request of a high-accuracy situation. In [31], SteerTrack for steering tracking can track the rota-
Researchers in [17] use WiDraw to implement an in-air tion angle in real-time, leveraging audio devices on smart-
handwriting application that allows the user to draw letters, phones. For acoustic signal selection and processing, the lin-
words, and sentences and achieves a mean word recognition ear chirp signal is employed in a dense multi-path environ-
accuracy of 91%. Researchers are also working to reduce the ment with Fractional Fourier Transform (FrFT) [11]. RAM-
number of base stations. PILSBON used iterative numerical TEL [26] was able to achieve a millimeter-level accuracy
estimation to compensate for the nonlinear errors and deployed for localization and motion tracking in multi-path fading by
an effective positioning system with six transmitters on the leveraging phase measurement. However, the efficiency is
rooftop [18]. To modulate the 2.4 GHz carrier frequency, significantly reduced in the occlusion conditions. A novel
PILSBON requires a 46.43 MHz intermediate frequency gen- localization system that fused the position information esti-
erator, raising the total cost. Other RF-based methods like mated by acoustic localization and IMUs was proposed in [32].
Bluetooth Low Energy (BLE) suffer from poor accuracy, high Researchers in [24] improved the accuracy of ultrasonic direc-
installation cost, or both. tion angles and applied it to SLAM. ALPS [12] was very
In comparison, UWB appears to be the more accurate one. considerate in NLOS recognition, Doppler Effect, and floor
The positioning accuracy of UWB can reach a centimeter-level plan while ignoring the motion tracking problem without
with a frequency of approximately 3.1-10.6 GHz. Although IMU utilization. Considering the most significant difference
it has been demonstrated that UWB can penetrate walls in between ALPS and our paper, the system proposed by us can
buildings and resolve multi-path effects due to its wide band- dynamically position the moving target with a novel fusion
width, it is still a challenge to employ it in indoor environ- strategy.
ments. UWB also seems to be advantageous in scalability. The IMU-based system is a classical method for
Unscented Kalman Filter (UKF) was used in [19] to combine infrastructure-less localization. In general, the pre-condition
UWB landmarks and inertial sensors and enhance the tracking of the high-precision IMU strategy is to make the sensors
accuracy. In the same way, AirTag was developed at Apple Inc. perfectly fit the mover during movement. Without such a

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23098 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

TABLE I
C OMPARISON OF I NDOOR P OSITIONING M ETHODS

pre-condition, errors may be reduced to at least 1-2 meters


in the Indoor Positioning and Indoor Navigation (IPIN) com-
petition dataset [33]. Exacerbating the problem, the IMU
in our paper differs from most others in that we mainly
apply embedded IMUs of a smartphone. Considering the
smartphone’s cost, the COTS IMU was lightweight, and its
tracking error accumulated rapidly over time with double
integration. Zero Velocity Update (ZUPT) corrects speed with
standstill determination to avoid the error growth by double
integration. When the ZUPT detects that the user is stationary,
the velocity calculated by the double-integration seems to
be the system error’s observation. [27]. ZUPT requires that
Fig. 1. Diagram of proposed methodology.
specific sensors be mounted on foot, then divides pedestrian
walking into several processes. When the left and right feet are
Machine (ELM) method and detected breathing successfully.
switched, the foot which supports the body is static relative to
However, they consume too much battery energy emitting and
the ground (meaning zero velocity), so the cumulative error
receiving the uninterrupted Frequency Modulated Continuous
can be corrected in each step. In practice, phones are often
Wave (FMCW) and could not work in far-field positioning
placed in hands, pockets, or bags rather than mounted on the
even under LOS conditions.
waist or foot, which is irrational for commercial use. Authors
in [28] realize pedestrian tracking for unconstrained carrying
III. S YSTEM A RCHITECTURE
the smartphone in all gestures by identifying three typical
modes in real-time. Since [28] has accomplished a set of This paper has proposed a novel approach that is fully
complete work in multi-mode PDR, this paper will mainly automated by using existing COTS devices. The schematic
focus on the situation with handhold gestures. The PDR in diagram of the method is shown in Fig. 1. The most significant
IMU has a visible effect on plane positioning, but there is still advantage of this approach is that it can be directly applied to
a significant error in the vertical direction. A barometer was smartphones without extra hardware, except signal emission
used to deal with vertical positioning errors with the Bayesian beacons need to be placed by the location service provider in
Network inference method [29]. advance.
To summarize the methods mentioned above, none of them
can give the ground and stable location alone. Combin- A. TDoA Based Acoustic Localization
ing sounds and IMU has been discussed in a few papers. We have implemented our system in a real-world environ-
Researchers in [32] ingeniously proposed the Kalman filter ment. The acoustic TDoA passive positioning is adopted, and
method. It integrated the advantages of both approaches and the conceptual architecture is shown in Fig. 2. The synchro-
obtained fusion signals with tight coupling, which led to better nization of inter-beacon is modulated with 433 MHz. The
stability and accuracy. However, it can only temporarily deal beacons are fixed, and the phone is held in hand. Different
with a single NLOS scenario but not solve the problem of beacons transmit chirp signals in a Time Division Multiple
regional signal loss (or NLOS) and improve PDR estimation Access (TDMA) mode. The proposed algorithm is imple-
performance. mented on the phone. Then the phone collects acoustic signals
For standard Kalman filters, the consideration for acoustic and IMU data without using other devices.
signal loss is insufficient. We have improved the algorithm The beacons are usually fixed on walls or ceilings to emit
in this paper, taking the fused locations as the input value modulated acoustic signals. The microphone embedded in the
and solving the modified step coefficient according to the smartphone receives and demodulates signals with a General-
least-squares method. CAT in [34] combined the Doppler ized Cross-Correlation PHAse Transformation (GCC-PHAT)
Shift, the IMU, and acoustic characteristics to achieve the algorithm [35], and calculates the Time of Flight (ToF). Based
millimeter-level near-field positioning (coverage within one on the calculated ToF, the distance between the beacons and
meter). Authors in [26] further refined the CAT accuracy the phone can be obtained, and the indoor locations of the
by employing phase ranging using the Extreme Learning phone are determined.

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
WANG et al.: SMARTPHONE-BASED PEDESTRIAN NLOS POSITIONING BASED ON ACOUSTICS AND IMU PARAMETER ESTIMATION 23099

spots represent the estimated positions of our developed


acoustic positioning, the red line indicates the errors. The
black circle labels the severity of outliers. The competition
arena is a double-floor building, including a hall, corridor, and
stairs. The Lidar scheme gives the ground truth. We deployed
ten beacons distributed evenly in the test area. As can be
seen from Fig. 3(a), except for a few outliers, which are
extremely far away, most of the other errors are tolerable.
Due to the inaccurate sound velocity calibration, the outer-ring
positioning results are generally smaller than the ground truth.
Then, the signal loss and errors from the competition results
Fig. 2. Conceptual architecture of acoustic TDoA passive positioning.
are analyzed in Table. II. It can be seen that most of the
positioning results are relatively accurate except for one severe
error drifting from the outlier point. Floor shielding produces
the NLOS area, resulting in the loss of acoustic signals and
outliers.

C. IMU Offset and Orientation


We employ the nine-axis IMUs embedded in Android
smartphones for the inertial sensor, including gyroscopes,
accelerometers, and magnetometers. The magnetometer deter-
mines the initial direction, the accelerometer and gyroscope
Fig. 3. (a) The performance of our acoustic prototype in 2018 Microsoft jointly measure the angular velocity, and the accelerometer is
Indoor Localization Competition. (b) The outlier data are circled in black.
The error analysis of competition results. mainly used to measure acceleration.
We collect the acceleration and angular velocity data from
TABLE II the smartphone with a sampling frequency of 30 Hz. Because
T HE A NALYSIS OF C OMPETITION R ESULTS
B ASED ON LOS AND NLOS C ONDITIONS
these data are calculated with a double integration, even the
minor systematic deviation will be magnified. To calibrate the
minor offset of the phone IMU system, we have analyzed
the sensor data with the help of Allan variance modeling [37].
The Allan variance modeling measures the stability of pre-
cision instruments and identifies various noise sources present
aco is calculated as in fixed sensor measurements. Data of length of four hours is
The acoustics positioning Xm
collected with a sampling frequency of 100 Hz, and the results
ctm,i1 = Xm
aco
− Bi 2 − Xm
aco
− B1 2 +ψi , i ≥ 2, (2) are shown in Fig. 4.
 We analyze the Allan variance curve of the tri-axial average
T Do Am,i1
in segments from the log-log curve shown in Fig. 4. The
where ψi ∼ N (0, σi2 ),  • 2 is Euclidean distance in 2-norm, random angle walk is defined as the white noise spectrum of
and subscript i 1 means from Beacon Bi to Beacon B1 in the IMU output, shown in a black line with a slope of z0.5 in
TDoA. Assume that the speed of sound c is constant at Fig. 4. The random rate walk is defined as the Brownian
340 m/s. tm,i1 is derived from the GCC-PHAT of received noise of the IMU output, shown in a black line with a slope
acoustics in m step and i beacon, and not completely accurate of 0.5. Since the vertical axis has been adjusted, the slope
due to the random loss and NLOS in acoustic signal propaga- does not match the actual log-log curve. Moreover, the line of
tion. Bi means the position of i th beacon and B1 denotes the slope : 0 represents the data of bias instability, which is the
position of 1st beacon. The first beacon is set as the principle source of the Flicker noise of IMU.
beacon, and subtracted by other beacon. Eq. (2) can be solved Then we amend the sensor bias as
by Chan in [36].
â(t) = a(t) + bi as Na (t) + bi as aB (t) + bi as Ka (t)
ω̂(t) = ω(t) + bi as Nω (t) + bi as ωB (t) + bi as Kω (t) (3)
B. Performance Analysis of Near-Ultrasonic Positioning
A typical acoustic positioning system is composed of a where â(t) and ω̂(t) denote the raw accelerometers and gyro-
smartphone and more than four loudspeakers. The loudspeaker scopes reading with bias respectively. The subscripts N, B, K
emits modulated acoustic signals as a beacon, and the smart- represent angle random walk, bias instability, and rate random
phone receives modulated signals as a tag and calculates the walk.
TDoA in real-time. The raw gyroscopes ω after bias removal are also not
Figs. 3(a), 3(b) and Table. II show the results and analysis handled directly and require conversion. Euler Angle is the tra-
of the 2018 Microsoft Indoor Localization Competition [5], ditional way of estimating the rotation angle except for Gimbal
where the blue curve is the trace of ground truth, the green Lock. To avoid Gimbal Lock, the quaternion is employed to

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23100 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

Fig. 5. The comparison with right hand and left foot, while handhold
data are slight and delay.

where ait = [0, ât,x , ât,y , ât,z ]T is denoted tri-axis accelerome-


ter reading by reduced bias. The value of ait is always affected
by gravitation constant, so gg = [0, 0, 0, g] is subtracted
in Eq. (9).

D. Step Judgment
After calibrating the acceleration and rotation, we have
two options: acceleration to displacement, separately double
integration method with Inertial Navigation System (INS),
Fig. 4. Acceleration and angular velocity in Allan deviation of the log-log
curve. Tri-axial and average data are shown.
and PDR method. Since the acceleration data are inaccurate,
double integration is equivalent to double-expansion offset.
update orientation. A convert quaternion can be regarded as a Then we select the PDR method to obtain displacement to
hyper-vector as q = [q0 , q1 , q2 , q3 ]T , usually the initial value avoid double integration. The test phone is hand-held instead
of q0 is zero. The first-order differential equation gives the of tied to our system’s human body or feet. Whether walking
generation of q as or calling, the sensor data always contains buffer and delay
from the wrist and shoulder, so the sensors reading can hardly
dq
= i q, (4) reflect the actual posture. Fig. 5 compares hand-held phone
dt ⎛ ⎞ data in red and foot-tie-phone data in purple. The handhold
0 −ω̂x −ω̂ y −ω̂z data has a significant delay compared to the foot data. In line
1⎜
⎜ω̂x 0 ω̂z −ω̂ y ⎟

 = ⎝
i
(5) with the kinematics model, it satisfies the following conditions
2 ω̂ y −ω̂z 0 ω̂x ⎠ after removing delay.
ω̂z ω̂ y −ω̂x 0,
|a f oot | ≥ |ahand | (10)
where ω̂ = [0, ω̂x , ω̂ y , ω̂z ]T is gyroscope amendment reading.
Then q is updated by first order Runge-Kutta Principle as where a f oot is the data of IMU on foot, whereas ahand is the
data of IMU on hand.
qt +1 = exp(it dt)qt . (6)
At first, we should distinguish each step boundary. In Fig. 6,
Ri and R g are denoted IMU system and geodetic frame the orange line shows the raw accelerometer data, and the
system transformed by thick red line shows data after the average smooth filter
  while the window size is set to 5. Then the peak threshold
Rg = q Ri qT = Q(Ri ). (7) is set in advance in the gray dotted line. Once the red line
We express transform as Q(•). The Eq. (8) is none-matrix peak exceeds, the threshold is considered a practical step.
multiplication, but Hamilton products. Furthermore, the peak-peak interval cannot be too dense [38].
 Otherwise, it will still be justified as a jitter.
[w1 , x 1 , y1 , z 1 ] [w2 , x 2 , y2 , z 2 ] As far as we know that step length is individual from
= [w1 w2 − x 1 x 2 − y1 y2 − z 1 z 2 , person to person, and the experiments should be conducted
w1 x 2 + x 1 w2 + y1 z 2 − z 1 yz 2 , in different populations. However, due to the arbitrariness and
complexity of human movement, the threshold and segment
w1 y2 − x 1 z 2 + y1 w2 + z 1 x 2 , may still not accurately judge the pace. For example, when
w1 z 2 + x 1 y2 − y1 x 2 + z 1 w2 ] (8) the arm lifts the smartphone straight, it is almost uniformly
moved, resulting in nearly zero acceleration. The loss is
The accelerator reading from i to g is converted as
inevitable unless the phone is tied directly to the shoe (like
g
at = Q t (ait ) − gg , (9) ZUPT). Merely changing the threshold for acceleration may

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
WANG et al.: SMARTPHONE-BASED PEDESTRIAN NLOS POSITIONING BASED ON ACOUSTICS AND IMU PARAMETER ESTIMATION 23101

Fig. 7. The event points of acoustics and PDR. The frequency of PDR
may be non-constant. The point of acoustics will lost occasionally.

Fig. 6. Example of step judgment. The peak marked with a red A. Strategy of Fusion
pentagram represents a step of the pedestrian. The black dotted line We assume the ground truth state at time k is evolved from
denotes the threshold, and the green dotted circle below the threshold
the state k − 1 by PDR model from Eq. (13)
is a missing step. The blue line shows the v H of the missing step.

Xk+1 = Xk + Sk T · uk + wk , (13)
be counterproductive. For example, we set the cut-off peak
at 0.5 m/s 2 , while the lost peak in Fig. 6 appears in around
0.3 m/s 2 , which even lower than some sub-peaks. A false where Xk = [x k yk ]T is denoted as plane coordinate at k in
peak method was proposed in [39], except for dealing with indoor frame system, Sk is denoted as step length in k, and
the lost peak. We found that the velocity within a single step uk = [sin θ cos θ ]T is controller vector where θ is azimuth
is calculated as Eq. (11). angle from Eq. (7). The PDR noise which follows a normal
distribution with wk ∼ N (0, Qk ).
vtH = vtH−1 + atH t, (11) The indoor frame system usually obeys the right-hand rule,
and the building determines the orientation. The orientation
where H is denoted as a horizontal component, for instance leads to a transformation between the indoor frame system
y
atH = [(atx )2 , (at )2 ]. t is the IMU sampling interval, shown and the geodetic frame system, denoted as T in
in Fig. 6. The threshold in Eq. (11) is assigned by practical
experience.  
cos φ − sin φ
T= , (14)
sin φ cos φ
E. Step Length Measurement
The traditional PDR method implements personal data, where φ is the rotation angle.
including but not limited to height, age, and leg length, to gen- The observation matrix in our method is different from
erate crucial parameters, raising tremendous privacy concerns. the usual Kalman filter. We define the moment of received
Authors in [40] raise constant paces for different motion acoustics as tm and of walk step tk . Unfortunately, the fre-
modes, including walking, jogging, and running. We propose a quency of pace is not equal to acoustic frequency, as shown
feedback step length estimation. Initially, an empirical formula in Fig. 7. We set a time gap interval (tgap ), when |tm − tk | is
as Eq. (12) designed by Weinberg describes the corresponding less than tgap , tm is approximately the same as tk . Otherwise,
relationship between step length and peak-valley of accel- when |tm − tk+1 | is less than tgap , this moment is considered
eration [41]. Weinberg only needs a fourth-root calculation to be none of observation, so skip to k + 1. If else, tm is
compared with other models, and it can be accelerated by just between tk and tk+1 , and we estimate the data of Xkaco by
Newton’s Method with rapid iteration. first-order Newton interpolation. The synchronization criterion
of acoustic signal and PDR is shown as Eqs. (15) and (16).
g g 1
Sk = γ (|ak |max − |ak |min ) 4 (12) ⎧
g

⎨Xm , |tm − tk | ≤ tgap ,
aco
|ak |is denoted as the estimation scalar of acceleration in k
Xkaco = skip, |tm − tk+1 | ≤ tgap , (15)
step, γ represents the correction parameters of an individual. ⎪

In fact, the preset value γ is not suitable for everyone in αk Xm
aco + (1 − α )Xaco , other wi se,
k m−1
the motion model. γ cannot be collected, which involves the tk − tm−1
αk = . (16)
privacy of personal body type [42]. It is also impractical to tm − tm−1
modify it by handling IMU value again. In order to protect
user privacy, we did not collect individual data to obtain the Then, we define the observation matrix as
step coefficient. Considering many loss data in observation,
we propose employing the Kalman Filter to compose IMU zk+1 = Xk+1
aco
= Xk+1|k + vk , (17)
and acoustic data.

where vk ∼ N (0, Rk ). Due to the uncertain error distribution


IV. S TEP L ENGTH C ORRECTION B ASED F USION of the acoustics, the outliers caused by signal loss or NLOS
In Weinberg’s model, the step coefficient γ is determined by are considered unsuitable for fusion. Therefore, we need to
prior knowledge such as the user’s height, leg length, gender, predict the acoustics positioning certainty L as to the weight.
and weight; all the data are users’ privacy. The advantage of We calculate the L and  L from Xm aco of Eq. (2) to Eqs. (18)

our approach is self-iteration without privacy problems. and (19) into the updating process of weight by a Kalman

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23102 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

filter, where the σi is from ψi ∼ N (0, σi2 ) in Eq. (2). C. Updating γ From KFBPE
KFLS method is calculated and overwritten when a new
aco
L(Xm |Bi , · · ·, Bn )
fusion position is obtained. Once the user pace is unstable, the
n
1 (ctm,i1 − T Do Am,i1 )2 step coefficient will endure a substantial deviation from the last
= √ ex p{− }, i ≥ 2,
i=2
2πσi 2σi2 datum, KFLS also changes dramatically. Because the method
(18) only considers single loss signals, they will be eliminated by
 the next fusion even if upheaval appears. In KFBPE, we are

L
, if L < L max , caring about large-scale loss signal problems. A Bayesian
L = L+1 (19)
1, if L ≥ L max . estimation-based step coefficient correction scheme is adopted,
which significantly enhances the robustness of positioning.
In general, L max is the maximum likelihood, which is set
aco is considered loss and We consider that positioning is a Hidden Markov
to 100. When L exceeds L max , Xm
Model (HMM) model, that the transition of each location
will not be updated z.
depends on the last n locations. So γ probability model is
Then Kalman gain Kk+1 , Xk+1 at state k + 1, and covari-
simplified as P(γ |Xk , Xk−1 , . . . , Xk−n+1 ) = P(γ |X K ), (k ≥
ance Pk+1 from (20–23) are updated. The measurement H is
n), according to Bayesian estimation as
temporarily considered as a 2-order identity matrix.
P(X K |γ )P(γ )
Kk+1 = Pk+1|k H T (HPk+1|k H T + Rk+1 )−1 , (20) P(γ |X K ) = ∝ P(X K |γ )P(γ ) (27)
P(X K )
Xk+1 = Xk+1|k + Kk+1 [(1 −  L k+1 )yk+1 + L k+1 Xk+1|k ],
(21) where from Xk to Xk−4 are positions we obtained in Kalman
Filter, prior probability P(γ ) is known by variable PDR
yk+1 = zk+1|k − HXk+1|k , (22) database. Then we calculate the likelihood P(X K |γ ) as
Pk+1 = (I − Kk+1 H)Pk+1|k . (23)
P( K |γ )P(X K | K,γ)
So far, we have estimated the actual movement path of P(X K |γ ) = (28)
P( K |X K , γ )
the target, but it is still based on a free acoustic propagation
channel. It is practically impossible to maintain such an ideal where K is a set of observation data Xaco K , likelihood
situation all the time. In other words, the acoustic signal is of acoustics L K , and IMU data, which is independent of
often lost, or the observation data sometimes are incredibly γ . Then we simplify Eq. (28) and eliminate the evidence
inaccurate. We can only trust IMU at that moment. However, P( K |X K , γ ),
neither the issue of step coefficient nor accumulated error
in IMU is resolved. The next subsection task guarantees P(X K |γ ) ∝ P( K )P(X K |γ , K ). (29)
the personal step coefficient through existing fusion data,
Now we have a prior probability of P( K ) from IMU and
videlicet, guarantees long-term accuracy without relying on
acoustics, and just calculate P(X K |γ , K ) from Kalman Filter
stable acoustic data, and eliminates the accumulated error.
in Eqs. (21) and (22) as Eq. (30), as shown at the bottom of
the next page.
B. Updating γ From KFLS It means the state value Xk+1 update is regarded as
As mentioned in Eq. (12), Weinberg’s pace estimation the weight of observation and prediction, and correspond-
B P E and X B P E . Calculate the
involves a specific parameter of γ . In order to update γ , ing weights are respectively Zk+1 k+1
we push back step length with Least Square after Kalman spectral radius and normalize them of ρ(Zk+1 B P E ), ρ(X B P E ) to
k+1
Filter according to the fusion coordinates. Assuming a con- approximate the probability P(z K |γ , K ) and P(X K |γ , K ).
secutive two-step model based on Eq. (13), the cost function Finally, we integrate Eqs. (27) and (29),
of Sk is denoted as
P(γ |X K ) ∝ P( K )P(X K |γ , K )P(γ )
Js (Sk ) = Xk+1 − Xk − Sk T · uk 22 . (24) = P( −1
K )P(X K |γ , K )P(γ ) (31)
To minimize the cost function Ŝk = arg min Js (Sk ), the step
Sk ∈R where  is a normalization coefficient, which makes (31)
length estimation Ŝk must satisfy as a distribution. Through the (31), we combine the prior
probability and Kalman filter to convert the single-valued γ
Ŝk = (T · uk )† (Xk+1 − Xk ), (25) into a conditional distribution, which enhances the robustness
where † denotes the pseudo-inverse matrix. Then γ is calcu- of the system and stabilizes the motion model.
lated as (26) on the basis of (12) We summarize the above method together. Table. III shows
the pseudo-code of our final algorithm.
1
γ̂ = Ŝk /(|âk |max − |âk |min ) 4 (26)
KFLS accomplishes the Least Squares (LS) specific parame- V. E VALUATION
ters update γ through the above process. Generally speaking, γ This section describes the experiment scenario, device
is relatively stable for the same person with the same posture, preparation, experiment processing, and result analysis. All
with maybe no change in the short term. the tests are actual measurements.

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
WANG et al.: SMARTPHONE-BASED PEDESTRIAN NLOS POSITIONING BASED ON ACOUSTICS AND IMU PARAMETER ESTIMATION 23103

TABLE III TABLE IV


P SEUDO -C ODE OF O UR M ETHODS N UMBER OF B EACONS AND P OSITION L OSS
R ATE (PLR) IN LOS AND NLOS

Fig. 10. Hold the phone and walk along the pre-prepared ground truth
line. The smartphone calculates real-time positioning coordinates.

The above limitation is that some commercial smartphones’


maximum sampling frequency is only 44.1 kHz or 48 kHz.
Although some Android smartphones are preset with 96 kHz
or higher, the smartphone’s hardware can only support up to
48 kHz. The sound pressure level (SPL) is less than 60 dB,
guaranteeing human safety. Under 60 dB SPL, the acoustic
signal can effectively propagate over 30 m, so we adjust the
Fig. 8. Different chirp triggering strategies by division time.
time-division gap to 150 ms, as shown in Fig. 8.
Note that we can receive and process only one signal
for each interval when the beacons emit chirps succes-
sively. It avoids the classification of multi-signals, such as
multi-signal cocktail issues. We hold a smartphone in hand and
move according to a pre-defined route. Meanwhile, we also
evaluated the influence of the number of beacons on the
position loss rate under different indoor scenarios (LOS and
NLOS). The result is shown in Table. IV. When the number
of beacons is only four, nearly 30% of the signals may be lost
and cause tremendous errors. Beacons are usually fixed on
Fig. 9. Power supply of beacons. The integrated lithium battery inside, the walls and are powered either by batteries or 5V adapters,
so the width exceeds two coins.
as shown in Fig. 9.

A. Implementation B. Benchmark and Ground Truth


The proposed methods are evaluated by a Huawei Mate As shown in Fig. 10 (a), we stuck white tape on the floor
9 with Android 9.0, Hisilicon Kirin 960 CPU, and 4 GB in advance as the ground truth, and the actor walked along the
memory as a middle-level configuration COTS phone. The white route. There are a few centimeters in error between the
beacon consists of an amplified speaker, a power supply trace of a phone and the ground truth line, which is difficult
module, a sync module, and an STM 32 chip (about five to avoid.
dollars each). We modulate the chirp signal with a duration of
40 ms [11]. To ensure that most people are insensible to the C. Positioning Experiment of LOS
modulated chirp signals, we use a 4 kHz bandwidth, sweeping The experiment scenario set no artificial obstacles, and
from 18 kHz to 22 kHz, transmitting near-ultrasonic signals. all areas could theoretically be considered unobstructed. The

Xk+1 = Xk+1|k + Kk+1 [(1 − L k+1 )(zk+1|k − HXk+1|k ) + L k+1 Xk+1|k ]


= (1 − 
L k+1 )Kk+1 zk+1|k + [I − (1 − L k+1 )Kk+1 H + L k+1 Kk+1 ] Xk+1|k (30)
 
BPE
Zk+1 BPE
Xk+1

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23104 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

Fig. 11. The test scenario and ground truth route. The illustrated route
can traverse every part of the room. All of the tables and chairs show no
more than 1 meter high.

Fig. 14. Comparison between KFLS and KFBPE for different numbers
of beacons.

Fig. 12. Comparison of the LOS results. The axis is room size.

Fig. 15. Probability density of error, mean μ, and standard deviation δ


comparison between KFLS, KFBPE and normal kalman.

Fig. 13. Error distribution in LOS. The maximum error in pure PDR is
too large and out of the bound.

beacons are uniformly distributed according to the optimal


Geometric Dilution Precision (GDOP). In this experiment, all
six Beacons were turned on. We defined it as an entire LOS
scenario, as shown in Fig. 11. Nevertheless, the actor will
inevitably block some signals behind himself during the walk.
It can be seen from Figs. 12 and 13 that acoustics, the KFLS,
and the KFBPE are similar in LOS performance. Fig. 16. The γ varies with the step number comparing with KFLS, KFLS
We got high accuracy positioning in this experiment, except after Moving Average (MA), and KFBPE. The red line of KFBPE shows
the expectation.
for drift points and outliers in several specific areas. Two
cabinet air-conditioners are fixed in the working model at the
left and right corners, affecting the corner temperature. Tem- in the 46 s walking, conforming to the theoretical value of
perature changes will inevitably lead to sound velocity changes cumulative error. Here, the step size of the PDR algorithm is
and affect localization. Especially in the right bottom corner, estimated as the Weinberg algorithm, the quaternion updates
the yellow line showed the outliers with severe multi-path the orientation, and the trajectory follows the Attitude and
fused temperature change. Due to body buffering, the accuracy Heading Reference System (AHRS) principle. The advantage
of the PDR algorithm has been decreased by more than 2 m of the proposed method did not fully reflect in the LOS

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
WANG et al.: SMARTPHONE-BASED PEDESTRIAN NLOS POSITIONING BASED ON ACOUSTICS AND IMU PARAMETER ESTIMATION 23105

Fig. 17. Screen board position and ground truth route under two scenarios of NLOS. The blue circle highlights the screening board.

Fig. 18. Coordinate of S1 and S2, different amount of beacons, mixed gender.

scenario. Six beacons are fixed to achieve a 150% redundancy D. Positioning Experiment of NLOS
rate, whereas minimum positioning requires four beacons. To further verify the robustness of the KFBPE method,
Fig. 14 shows a different number of beacons required we set up a screen board occlusion experiment. The occlu-
for positioning accuracy. Both KFLS and KFBPE obtain a sion environment has always been a tricky problem in
superior accuracy, 4, 5, 6 represents the number of beacons. infrastructure-based localization. Most papers exert map con-
We switched off the B5 and B6 beacons in Fig. 12 sequen- straints or floor matching methods to solve them. KFBPE
tially. KFBPE will improve accuracy more than the KFLS method uses IMU and fusion data effectively and is theoreti-
method, whereas different numbers of beacons have a slight cally robust to occlusion.
impact. Fig. 17 shows the two occlusion scenarios set in our experi-
The probability density of error for our proposed method ments. On the left is Scenario 1 (S1), in which the screen board
and standard Kalman filter is shown in Fig. 15. KFBPE has a blocks the halfway aisle. Scenario 2 (S2) is on the right. The
concentrated density distribution with a small mean value and screening board almost wholly cut off the lab, leaving only
strong robustness. All of them are distributed below 0.5 m, a gap to walk. Due to the screen board barrier, we adjusted
of which 0.25 m error is the most concentrated. KFLS also has the route to a loop for the S2. We notice that almost half
a high probability distribution within 0.5 m, but its density is of the S2 suffers from regional signal loss and NLOS. The
relatively sparse. The error probability density of the standard floor is 2.7 m in height, and the screen panel is 2.65 m in
Kalman filter in the other paper approximately presents a height. Acoustic signals can hardly pass through from above.
normal distribution with a mean value of 0.4 m, and its In contrast to the screen board, the office desk and seats
performance improvement is slightly worse than that of the in the middle have a maximum height of 1 m, making up
above two methods. Fig. 16 shows the variation trend of γ in a weak occlusion environment. The positioning accuracy in
KFLS and KFBPE. After moving averaging, the line of KFLS such a scenario is more convincing and closer to the actual
serves as a contrast to show that even smoothing the γ in application.
KFLS is far less effective than KFBPE. The results in Fig. 18 with both male and female users
Measured by a large amount of previous data, the γ of show the positioning performance for two scenarios. The
this user is 0.59, which is very close to the value evaluated gender difference has no discernible effect on positioning,
by KFBPE. Although KFLS fluctuates wildly, the average but the beacon number plays a significant role. It represents
is slightly close to 0.59. The experiment proves that both a different amount and distribution of beacons, which will
methods effectively correct the personal coefficients γ based result in different outliers and performance. Compared with
on online self-learning, while KFBPE ensures low variance Fig. 19, although the positioning performance under the NLOS
and stability. condition is not as ideal as LOS, we still control the maximum

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23106 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

Fig. 19. Cumulative error distribution comparison between different settings. The error in S1 is obviously lower than S2.

TABLE V
C OMPARISON OF P OSITIONING AND T IME -C OSTUMING R ESULTS IN S1, S2, AND LOS T ESTS (6 B EACONS )

Box plot, as Fig. 20 of positioning error more vividly


expresses the error distribution. The first six columns represent
six sets of experiments for S1, and the last six represent six
sets of experiments for S2. The median overall error is between
15 cm and 40 cm, whereas the lower quartile is mostly
within 20 cm. The maximum outlier is less than 1 m, and
the distribution is relatively flat. The sixth set, corresponding
to S1-6-actress, is one of the best data besides the maximum
outliers. Overall, the beacon quantity is positively correlated
with the positioning performance.
Table. V shows a result comparison of the methods pro-
posed in this paper. Due to the NLOS, the acoustics accuracy
Fig. 20. Box plot of the error distribution. The first six columns correspond of S1 and S2 is the worst, especially in the S2 scene. The
to Figs. 18(a) and 19(a) in half-occluded, and the last six columns
correspond to Figs. 18(b) and 19(b) in full-occluded. trace of PDR is not limited by occlusion but only by moving
distance, leading the similar results in the three scenarios. The
standard Kalman filter is better than acoustics and PDR but
not as good as KFLS and KFBPE. Among them, KFBPE
error below 1 m, and the average error distribution is less achieves the best positioning results without impacting the
than 30 cm. We find approximate error performance compared NLOS of S1 and S2. The Longest Positioning Time (LPT) in
with S1 and S2, reflecting the strong robustness of the KFBPE the last column of each scenario is in Table. V, both KFLS and
method from other viewings. The KFLS method is not suitable KFBPE algorithms can complete positioning calculation and
for this scenario. synchronous display within 1 s. Due to the loss of acoustics,
Due to the screen barrier, the testers cannot walk a circle there is a situation where there is no positioning point for a
along the same way in S1 and S2. More than three directional long time. Especially in the S2 scenario, the LPT is up to 12 s.
acoustic signals still exist in the left half of the laboratory and
the right half, maybe only remaining less than two beacon VI. C ONCLUSION
signals. Without the KFBPE method, two signals can only We propose a novel positioning system with strong robust-
determine the fuzzy positioning, which means location failure. ness in a complex environment. The core is Kalman Filter
At that time, the KFBPE method can obtain valid information and Bayesian parameter estimation so that (i) it fuses acoustic
from the exact value on the left half, adjust the step coefficient, and IMU data, (ii) users can localize their position only on a
and improve the right half’s positioning accuracy. Even if the smartphone, (iii) it can handle a variety of LOS and NLOS sce-
number of beacons decreased to 4, KFBPE still maintains a narios, (iv) it avoids getting user privacy. A Bayesian method
stable track path without outliers. is applied to estimate the step coefficient, which solves the

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
WANG et al.: SMARTPHONE-BASED PEDESTRIAN NLOS POSITIONING BASED ON ACOUSTICS AND IMU PARAMETER ESTIMATION 23107

inaccurate positioning of hand-held smartphones, eliminates [19] F. Hoflinger, R. Zhang, P. Fehrenbach, J. Bordoy, L. Reindl, and
IMU’s accumulated error, and figures out the customized para- C. Schindelhauer, “Localization system based on handheld inertial sen-
sors and UWB,” in Proc. IEEE Int. Symp. Inertial Sensors Syst. (INER-
meter by online self-learning. We compare KFLS and KFBPE TIAL), Mar. 2017, pp. 1–2.
methods to other methods, showing a significant positioning [20] Wikipedia. (2020). 2020airtag. [Online]. Available: https://en.wikipedia.
accuracy, especially in some NLOS conditions. Future work org/wiki/AirTags
[21] W. Hess, D. Kohler, H. Rapp, and D. Andor, “Real-time loop closure
will leverage Long Short-Term Memory (LSTM) to analyze in 2D LIDAR SLAM,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA),
acoustic signal characteristics under NLOS and combine the May 2016, pp. 1271–1278.
clustering to deal with NLOS data further. [22] R. Zhang, W.-D. Zhong, K. Qian, and D. Wu, “Image sensor based
visible light positioning system with improved positioning algorithm,”
IEEE Access, vol. 5, pp. 6087–6094, 2017.
[23] J. Lee, “4–1: Invited paper: Mobile AR in your pocket with Google
R EFERENCES Tango,” in SID Symp. Dig. Tech. Papers, 2017, vol. 48, no. 1, pp. 17–18.
[24] S. Yuan et al., “DSmT-based ultrasonic detection model for estimating
[1] R. C. Shit, S. Sharma, D. Puthal, and A. Y. Zomaya, “Location of things indoor environment contour,” IEEE Trans. Instrum. Meas., vol. 69, no. 7,
(LoT): A review and taxonomy of sensors localization in IoT infrastruc- pp. 4002–4014, Jul. 2020.
ture,” IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 2028–2061, [25] B. Zhou, M. Elbadry, R. Gao, and F. Ye, “BatTracker: High preci-
3rd Quart., 2018. sion infrastructure-free mobile device tracking in indoor environments,”
[2] M. Z. Win, W. Dai, Y. Shen, G. Chrisikos, and H. V. Poor, “Network in Proc. 15th ACM Conf. Embedded Netw. Sensor Syst., Nov. 2017,
operation strategies for efficient localization and navigation,” Proc. p. 13.
IEEE, vol. 106, no. 7, pp. 1224–1254, Jul. 2018. [26] Y. Liu, W. Zhang, Y. Yang, W. Fang, F. Qin, and X. Dai, “RAMTEL:
[3] X. Wang, C. Zhang, F. Liu, and X. Xu, “Exponentially weighted particle Robust acoustic motion tracking using extreme learning machine for
filter for simultaneous localization and mapping based on magnetic smart cities,” IEEE Internet Things J., vol. 6, no. 5, pp. 7555–7569,
field measurements,” IEEE Trans. Instrum. Meas., vol. 66, no. 7, Oct. 2019.
pp. 1658–1667, Jul. 2017. [27] Z. Wang, H. Zhao, S. Qiu, and Q. Gao, “Stance-phase detec-
[4] D. Lymberopoulos and J. Liu, “The Microsoft indoor localization com- tion for ZUPT-aided foot-mounted pedestrian navigation system,”
petition: Experiences and lessons learned,” IEEE Signal Process. Mag., IEEE/ASME Trans. Mechatronics, vol. 20, no. 6, pp. 3170–3181,
vol. 34, no. 5, pp. 125–140, Sep. 2017. Dec. 2015.
[5] Microsoft. (2018). 2018competition. [Online]. Available: [28] Q. Tian, Z. Salcic, K. Wang, and Y. Pan, “A multi-mode dead reckoning
https://www.microsoft.com/enus/research/event/microsoft-indoor- system for pedestrian tracking using smartphones,” IEEE Sensors J.,
localization-competition-ipsn-2018/ vol. 16, no. 7, pp. 2079–2093, Apr. 2016.
[6] F. Zafari, A. Gkelias, and K. K. Leung, “A survey of indoor localization [29] Z. Xu, J. Wei, J. Zhu, and W. Yang, “A robust floor localization method
systems and technologies,” IEEE Commun. Surveys Tuts., vol. 21, no. 3, using inertial and barometer measurements,” in Proc. Int. Conf. Indoor
pp. 2568–2599, 3rd Quart., 2017. Positioning Indoor Navigat. (IPIN), Sep. 2017, pp. 1–8.
[7] V. Chamola, V. Hassija, V. Gupta, and M. Guizani, “A comprehensive [30] D. Chen et al., “S2M: A lightweight acoustic fingerprints-based wireless
review of the COVID-19 pandemic and the role of IoT, drones, AI, device authentication protocol,” IEEE Internet Things J., vol. 4, no. 1,
blockchain, and 5G in managing its impact,” IEEE Access, vol. 8, pp. 88–100, Feb. 2017.
pp. 90225–90265, 2020. [31] X. Xu, J. Yu, Y. Chen, Y. Zhu, and M. Li, “Leveraging acoustic signals
[8] E. Naserian, X. Wang, X. Xu, and Y. Dong, “A framework of loose for vehicle steering tracking with smartphones,” IEEE Trans. Mobile
travelling companion discovery from human trajectories,” IEEE Trans. Comput., vol. 19, no. 4, pp. 865–879, Apr. 2020.
Mobile Comput., vol. 17, no. 11, pp. 2497–2511, Nov. 2018. [32] H. Yang et al., “Smartphone-based indoor localization system using iner-
[9] E. Naserian, X. Wang, K. Dahal, Z. Wang, and Z. Wang, “Person- tial sensor and acoustic transmitter/receiver,” IEEE Sensors J., vol. 16,
alized location prediction for group travellers from spatial–temporal no. 22, pp. 8051–8061, Nov. 2016.
trajectories,” Future Gener. Comput. Syst., vol. 83, pp. 278–292, [33] I. E. Radoi, D. Cirimpei, and V. Radu, “Localization systems reposi-
Jun. 2018. tory: A platform for open-source localization systems and datasets,” in
[10] C. Peng, G. Shen, and Y. Zhang, “BeepBeep: A high-accuracy acoustic- Proc. Int. Conf. Indoor Positioning Indoor Navigat. (IPIN), Sep. 2019,
based system for ranging and localization using COTS devices,” ACM pp. 1–8.
Trans. Embedded Comput. Syst., vol. 11, no. 1, pp. 1–29, 2012. [34] W. Mao, J. He, and L. Qiu, “CAT: High-precision acoustic motion track-
ing,” in Proc. 22nd Annu. Int. Conf. Mobile Comput. Netw., Oct. 2016,
[11] L. Zhang, M. Chen, X. Wang, and Z. Wang, “TOA estimation of
pp. 69–81.
chirp signal in dense multipath environment for low-cost acoustic
[35] S. U. N. Wood, J. Rouat, S. Dupont, and G. Pironkov, “Blind
ranging,” IEEE Trans. Instrum. Meas., vol. 68, no. 2, pp. 355–367,
speech separation and enhancement with GCC-NMF,” IEEE/ACM
Feb. 2019.
Trans. Audio Speech Lang. Process., vol. 25, no. 4, pp. 745–755,
[12] N. Rajagopal, P. Lazik, N. Pereira, S. Chayapathy, B. Sinopoli, and Apr. 2017.
A. Rowe, “Enhancing indoor smartphone location acquisition using floor [36] K. C. Ho, “Bias reduction for an explicit solution of source local-
plans,” in Proc. 17th ACM/IEEE Int. Conf. Inf. Process. Sensor Netw. ization using TDOA,” IEEE Trans. Signal Process., vol. 60, no. 5,
(IPSN), Apr. 2018, pp. 278–289. pp. 2101–2114, May 2012.
[13] H. Wang, L. Zhang, Z. Wang, and X. Luo, “Pals: High-accuracy pedes- [37] J. Nikolic, P. Furgale, A. Melzer, and R. Siegwart, “Maximum likelihood
trian localization with fusion of smartphone acoustics and PDR,” in Proc. identification of inertial sensor noise model parameters,” IEEE Sensors
IPIN, 2019, pp. 291–298. J., vol. 16, no. 1, pp. 163–176, Jan. 2016.
[14] L. E. Díez, A. Bahillo, J. Otegui, and T. Otim, “Step length estimation [38] C. Tudor-Locke et al., “How fast is fast enough? Walking cadence
methods based on inertial sensors: A review,” IEEE Sensors J., vol. 18, (steps/min) as a practical estimate of intensity in adults: A narrative
no. 17, pp. 6908–6926, Sep. 2018. review,” Brit. J. Sports Med., vol. 52, no. 12, pp. 776–788, Jun. 2018.
[15] J. Jahn, U. Batzer, J. Seitz, L. Patino-Studencka, and J. G. Boronat, [39] F. Gu, K. Khoshelham, C. Yu, and J. Shang, “Accurate step length esti-
“Comparison and evaluation of acceleration based step length estimators mation for pedestrian dead reckoning localization using stacked autoen-
for handheld devices,” in Proc. Int. Conf. Indoor Positioning Indoor coders,” IEEE Trans. Instrum. Meas., vol. 68, no. 8, pp. 2705–2713,
Navigat., Sep. 2010, pp. 1–6. Oct. 2018.
[16] K.-H. Chow, H. He, J. Tan, and S.-H. G. Chan, “Efficient locality [40] H. Abdelnasser et al., “SemanticSLAM: Using environment landmarks
classification for indoor fingerprint-based systems,” IEEE Trans. Mobile for unsupervised indoor localization,” IEEE Trans. Mobile Comput.,
Comput., vol. 18, no. 2, pp. 290–304, Feb. 2019. vol. 15, no. 7, pp. 1770–1782, Jul. 2016.
[17] L. Sun, S. Sen, D. Koutsonikolas, and K.-H. Kim, “WiDraw: Enabling [41] H. Weinberg, “Using the ADXL202 in pedometer and personal
hands-free drawing in the air on commodity WiFi devices,” in Proc. 21st navigation applications,” Analog Devices, Norwood, MA, USA,
Annu. Int. Conf. Mobile Comput. Netw., Sep. 2015, pp. 77–89. Tech. Rep., AN-602, 2002, vol. 2, no. 2, pp. 1–6.
[18] J. Su, Z. Yao, and M. Lu, “An improved position determination algorithm [42] V. Renaudin, M. Susi, and G. Lachapelle, “Step length estimation
based on nonlinear compensation for ground-based positioning systems,” using handheld inertial sensors,” Sensors, vol. 12, no. 7, pp. 8507–8525,
IEEE Access, vol. 7, pp. 23675–23689, 2019. Jun. 2012.

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com
23108 IEEE SENSORS JOURNAL, VOL. 22, NO. 23, 1 DECEMBER 2022

Hucheng Wang (Student Member, IEEE) Lei Zhang (Student Member, IEEE) received
received the B.Eng. degree in information the B.Eng. and M.Sc. degrees in mechatron-
and communication technology in 2015. He is ics engineering from Chang’an University, Xi’an,
currently pursuing the Ph.D. degree with the China, in 2007 and 2013, respectively, and the
School of Computer Science and Information Ph.D. degree in control science and engineer-
Security, Guilin University of Electronic ing from Zhejiang University, Hangzhou, China,
Technology, Guilin, China. He is a Visiting in 2019. He is currently a Lecturer with the School
Scholar with the College of Control Science and of Construction Machinery, Chang’an University.
Engineering, Zhejiang University, Hangzhou, His academic interests include machine learn-
China. His research interests include indoor ing, acoustic signal processing, time-frequency
localization, acoustic signal processing, and analysis, indoor positioning, and tracking.
data fusion.

Xiaonan Luo (Senior Member, IEEE) was the


Director of the National Engineering Research
Center of Digital Life, Sun Yat-sen University,
Guangzhou, China. He is currently a Professor
Can Xue received the B.Sc. degree in automa- with the School of Computer and Information
tion from the Chongqing University of Posts Security, Guilin University of Electronic Technol-
and Telecommunications, Chongqing, China, ogy, Guilin, China. His current research interests
in 2019. He is currently pursuing the Ph.D. include computer graphics, machine learning,
degree with the Department of Control Science and pattern recognition. He received the National
and Engineering, Zhejiang University, Hangzhou, Science Fund for Distinguished Young Schol-
China. His current research interests include ars granted by the National Natural Science
indoor positioning, mobile systems, and acoustic Foundation of China.
signal processing.

Xinheng Wang (Senior Member, IEEE) received


the B.Eng. and M.Sc. degrees in electrical engi-
neering from Xi’an Jiaotong University, Xi’an,
China, in 1991 and 1994, respectively, and the
Ph.D. degree in electronics and computer engi-
Zhi Wang (Member, IEEE) received the B.Eng. neering from Brunel University, Uxbridge, U.K.,
degree from Shenyang Jianzhu University in 2001. He is currently a Professor with the
in 1991, the M.Sc. degree from Southeast School of Advanced Technology and the Head
University in 1997, and the Ph.D. degree of the Department of Mechatronics and Robot-
from the Shenyang Institute of Automation, ics, Xi’an Jiaotong-Liverpool University (XJTLU),
Chinese Academy of Sciences, in 2000. Suzhou, China. Prior to joining XJTLU, he was a
From 2000 to 2002, he was a Postdoctoral Professor with different universities in the U.K. He has been an Inves-
Fellow with the Institute National Polytechnique tigator or a Co-Investigator of nearly 30 research projects sponsored
de Lorraine, France; and Zhejiang University, from EU, U.K. EPSRC, Innovate U.K., China NSFC, and industry. He has
China. He is currently a Ph.D. Advisor with the authored or coauthored over 170 referred papers. He holds 15 granted
College of Control Science and Engineering, patents, including one U.S., one Japan, four South Korea, and nine
Zhejiang University. His research areas include acoustic signal and array China patents. His current research interests include tactile Internet,
signal processing, acoustic SLAM and robot positioning and navigation, indoor positioning, the Internet of Things (IoT), acoustic localization,
indoor localization and temporal-spatial data mining, device-free human communications and sensing, and big data analytics for intelligent ser-
activity recognition, compressed sensing, deep learning, data fusion vices, where he has developed the world’s first smart trolley with Chigoo
of multiple sensing systems, Industrial Internet of Things (IIoT), and Interactive Technology Company Ltd. His research has led to a few
privacy protection in crowdsourcing systems. He was also the winner of commercial products in condition monitoring, wireless mesh networks,
the 2018 Microsoft Indoor Localization Competition in the Sound Group. and user-centric routing and navigation for group users.

Authorized licensed use limited to: Universitaet Bielefeld. Downloaded on November 25,2023 at 11:32:49 UTC from IEEE Xplore. Restrictions apply.

You might also like