A transfer kernel learning based strategy for adaptive localization in dynamic indoor environments: poster

Abstract

Existing WiFi fingerprinting-based Indoor Positioning System (IPS) suffers from the vulnerability of environmental dynamics. To address this issue, we propose TKL-WinSMS as a systematic strategy, which is able to realize robust and adaptive localization in dynamic indoor environments. We developed a WiFi-based Non-intrusive Sensing and Monitoring System (WinSMS) that enables COTS WiFi routers as online reference points by extracting real-time RSS readings among them. With these online data and labeled source data from the offline calibrated radio map, we further combine the RSS readings from target mobile devices as unlabeled target data, to design a robust localization model using an emerging transfer learning algorithm, namely transfer kernel learning (TKL). It can learn a domain-invariant kernel by directly matching the source and target distributions in the reproducing kernel Hilbert space instead of the raw noisy signal space. By leveraging the resultant kernel as input for the SVR training, the trained localization model can inherit the information from online phase to adaptively enhance the offline calibrated radio map. Extensive experimental results verify the superiority of TKL-WinSMS in terms of localization accuracy compared with existing solutions in dynamic indoor environments.

Full text

Poster: A Transfer Kernel Learning based Strategy for
Adaptive Localization in Dynamic Indoor Environments
Han Zou1, Yuxun Zhou2, Hao Jiang3, Baoqi Huang4, Lihua Xie1and Costas Spanos2
1Nanyang Technological University, Singapore 2University of California, Berkeley, USA
3Fuzhou University, Fuzhou, China 4Inner Mongolia University, Hohhot, China
{zouh0005, elhxie}@ntu.edu.sg, {yxzhou, spanos}@berkeley.edu, jiangh@fzu.edu.cn,
cshbq@imu.edu.cn
ABSTRACT
Existing WiFi fingerprinting-based Indoor Positioning Sys-
tem (IPS) suffers from the vulnerability of environmental
dynamics. To address this issue, we propose TKL-WinSMS
as a systematic strategy, which is able to realize robust and
adaptive localization in dynamic indoor environments. We
developed a WiFi-based Non-intrusive Sensing and Monitor-
ing System (WinSMS) that enables COTS WiFi routers as
online reference points by extracting real-time RSS readings
among them. With these online data and labeled source da-
ta from the offline calibrated radio map, we further combine
the RSS readings from target mobile devices as unlabeled
target data, to design a robust localization model using an
emerging transfer learning algorithm, namely transfer kernel
learning (TKL). It can learn a domain-invariant kernel by
directly matching the source and target distributions in the
reproducing kernel Hilbert space instead of the raw noisy
signal space. By leveraging the resultant kernel as input for
the SVR training, the trained localization model can inherit
the information from online phase to adaptively enhance the
offline calibrated radio map. Extensive experimental results
verify the superiority of TKL-WinSMS in terms of localiza-
tion accuracy compared with existing solutions in dynamic
indoor environments.
Categories and Subject Descriptors
C.2.1 [Network Architecture and Design]: Wireless Com-
munication
Keywords
IEEE 802.11 WLAN, Localization, Transfer learning
1. INTRODUCTION
WiFi has been recognized as the most promising tech-
nique for indoor positioning services, due to the widely in-
stalled network infrastructures and pervasive WiFi-enabled
COTS mobile devices (MDs) [1, 4]. Existing WiFi-based
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
MobiCom’16 October 03-07, 2016, New York City, NY, USA
c
2016 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-4226-1/16/10.
DOI: http://dx.doi.org/10.1145/2973750.2985278
IPSs for practical large-scale implementation usually adop-
t fingerprinting-based localization algorithm [3]. It localizes
an MD by comparing the real-time RSS readings with a pre-
established RSS fingerprint database, (a.k.a. radio map).
The major bottleneck of WiFi fingerprinting-based IPS is
the vulnerability to environmental dynamics. The real-time
RSS readings may deviate significantly from the fingerprints
stored in the static radio map due to severe multi-path and
shadow fading effects caused by various interferences, which
leads to huge degradation on localization accuracy. To ac-
commodate environmental dynamics, radio map recalibra-
tion is extremely time-consuming and labor-intensive and
deploying fixed reference anchors to obtain fresh RSS read-
ings introducing extra hardware cost.
To overcome this problem, we propose TKL-WinSMS as a
systematic strategy, which is able to construct a robust mod-
el for adaptive localization in dynamic indoor environments.
We developed WinSMS that enables COTS WiFi routers as
online reference points by extracting real-time RSS readings
among them. With these online data and the offline cali-
brated radio map as labeled source data, we further com-
bine the RSS readings from target MDs as unlabeled target
data, to develop a robust localization model using an emerg-
ing transfer learning algorithm TKL [2]. It is able to learn
a domain-invariant kernel by directly matching the source
and target distributions in the reproducing kernel Hilbert
space. By leveraging the resultant kernel as the input for
SVR training, the trained localization model can inherit the
information from online phase to adaptively enhance the of-
fline calibrated radio map. Extensive experiments have been
conducted and demonstrated that TKL-WinSMS can pro-
vide high localization accuracy under various environmental
dynamics consistently.
2. SYSTEM DESIGN
Suppose pWiFi access points (APs) are installed in an
indoor environment, and the signals received by MD from
these APs at an arbitrary two-dimensional location `can be
represented by a signal vector RSS. The data are collected
at totally lCP offline calibration points (CPs) as DSCP =
{(RSSi,`i)}lCP
i=1 . The proposed WinSMS enables APs as
online reference points, and the data collected at online ref-
erence points lAP are denoted as DSAP ={(RSSi,`i)}lAP
i=1 .
Since the locations of CPs and APs are known, we combine
DSCP and DSAP as labeled source data DS={(RSSi,`i)}lS
i=1
in source domain with lS=lCP +lAP . We use DT=
{RSSi}lT
i=1 to represent the RSS data of target MDs in tar-
get domain. With the labeled source data DSand unla-
462

Figure 1: System Architecture of WinSMS.
beled target data DT, our goal is to estimate the locations
of MDs by constructing a robust localization model using
TKL-WinSMS.
2.1 WinSMS
We develop WinSMS, which enables COTS WiFi router-
s as online reference points by overhearing the data pack-
ets transmitted between each MD and other routers, and
precisely retrieve the RSS values and corresponding MAC
addresses as identifiers without introducing any extra hard-
ware infrastructure. Fig. 1 presents the system architecture
of WinSMS, which includes a main AP, remote APs, a back-
end server, users and their MDs. The main AP provides the
basic WLAN Internet services, receives UDP packets sen-
t by remote APs, and forwards the data to a server. The
server is responsible to store and parse the data. We up-
grade the firmware of remote APs with OpenWrt, and use
Libpcap to capture and analyze RSS packets in the exist-
ing WiFi traffic, extract relevant data and forward them to
the main AP. Since these remote APs can overhear packet-
s of other remote APs as well, all of them become natural
online reference points with their physical coordinates and
real-time RSS readings. In this manner, WinSMS is able to
collect the data among the APs as the online labeled data
DSAP and the RSS data associated with MDs as the unla-
beled data DTwithout introducing extra infrastructure or
any intrusiveness on user side.
2.2 TKL-WinSMS
With DSand DT, TKL aims to reveal the shared knowl-
edge in signal space across different domains, select a new
domain-invariant kernel as input data for SVR training, and
then construct an adaptive localization model to precise-
ly estimate the locations of MDs in dynamic indoor envi-
ronments. The methodology of TKL-WinSMS consists of 6
steps. Step 1 is to compute the source kernel KS, the target
kernel KTand the cross-domain kernel matrix KST using a
predefined input kernel function k. The next step is to eval-
uate the distribution differences between DSand DTin the
kernel Hilbert space. However, KSand KThave distinc-
t dimensions. Thus, we construct an extrapolated source
kernel KSusing the eignesystem of KT. The eignesys-
tem of KTcan be easily obtained by eigendecomposition
KTΦT=ΦTΛTas Step 2. Then, Step 3 is to compute
the eigenvector matrix of the extrapolated source kernel KS
by ΦS'KST ΦTΛ−1
T. We formalize the distribution dis-
Figure 2: Layout of the testbed.
crepancy between KSand the ground truth source kernel
KSin terms of the Nystr¨
om approximation error. As such,
to minimize the distribution divergence is equal to minimize
the approximation error, which can be achieved by using
squared loss as follows:
min
ΛkKS−KSk2=kΦSΛΦ>
S−KSk2,
λi≥ζλi+1 , i = 1, . . . ,lT−1, λi≥0, i = 1,...,lT,
(1)
where Λ= diag{λ1,...,λlT}are the lTnonnegative eigen-
spectrum parameters and ζis the eigenspectrum damping
factor. Step 4 is to estimate Λby translating the opti-
mization problem (1) into a standard convex quadratic pro-
gramming (QP) with linear constraints, which can be easily
solved by well-established convex optimization algorithms,
such as the interior-point algorithm. Step 5 is to build up
the domain-invariant kernel KAon both the source and tar-
get data using KA=ΦAΛΦ−1
A, where ΦA.
= [ΦS;ΦT].
Then, the estimated domain-invariant kernel KAcan be di-
rectly adopted as the input for SVR training to construct the
adaptive localization model via LIBSVM package as Step 6.
The output of the model are the estimated locations of MDs.
3. IMPLEMENTATION AND EVALUATION
Extensive experiments were conducted in a 600m2real
multi-functional lab across 6 months to validate the perfor-
mance of TKL-WinSMS. The experiments include two phas-
es: the initial phase (T1) and a phase conducted 6 months
later (T2). During T1, RSS samples of a Nexus 6 were col-
lected at 47 offline CPs using WinSMS. At each point, we
collected 100 RSS samples and denoted all data by DSCP .
During T2, we collected RSS data (100 RSS samples per lo-
cation) at 15 testing points (TPs) as well as the 14 APs using
WinSMS. Denote the data at 14 APs as DSAP and the RSS
samples at 15 TPs as DT. The locations of the CPs, TPs
and APs are shown in Fig. 2. The localization model built
up by TKL-WinSMS is trained on all labeled data DSCP
and DSAP , as well as unlabeled target data DTto learn the
domain-invariant feature representation. The performance
of TKL-WinSMS is compared with basic SVR (use DSCP
only for SVR training) and SVR-WinSMS (use both DSCP
and DSAP for SVR training). The RBF kernel is adopted
in all the three approaches. The overall performance com-
parison is presented in Fig. 3, with the detailed location
error on each TP using the three approaches demonstrated
463

0246810
Location Error [m]
0
0.2
0.4
0.6
0.8
1
CDF
SVR
SVR-WinSMS
TKL-WinSMS
Figure 3: CDF of location error.
in Fig. 4. The mean localization accuracy of TKL-WinSMS
is 1.860 m, which obtains a tremendous performance im-
provement of 38.85% and 26.66% over SVR (3.041 m) and
SVR-WinSMS (2.536 m), respectively. TKL-WinSMS also
outperforms others at nearly every TP as presented in Fig. 4.
These consistent performance gains verify that constructing
the robust domain-invariant kernel KAby TKL is much
more advantageous than SVR and SVR-WinSMS that on-
ly leverage raw RSS data. Furthermore, by fully exploring
the eigenspaces of both the source and target domain da-
ta to learn a domain-invariant kernel, TKL-WinSMS is able
to correctly revealing the related knowledge, and kernelizing
the original RSS data across different domains (time period-
s) for adaptive indoor localization. Thus, by leveraging the
online labeled data DSAP obtained by WinSMS as well as
the domain-invariant kernel KAlearned by TKL, an adap-
tive localization model can be constructed to consistently
provide reliable positioning service with various indoor en-
vironment dynamics.
4. CONCLUSION
In this paper, we proposed TKL-WinSMS, which is able
to construct a robust localization model for adaptive lo-
calization in dynamic indoor environments. We develope-
d WinSMS to enable COTS WiFi routers as online refer-
ence points by extracting the real-time RSS readings among
them. With these online data and the offline calibrated
radio map as labeled source data, we further combine the
RSS readings from the target MDs as unlabeled target da-
ta, to design a robust localization model using TKL. It can
construct a domain-invariant kernel that minimizes the d-
ifference between the source and target distributions in the
reproducing kernel Hilbert space instead of the raw noisy
signal space. By leveraging the resultant kernel as input for
the SVR training, the trained localization model can inher-
it the information from online phase to adaptively enhance
the offline calibrated radio map. Extensive experimental re-
sults verify the superiority of TKL-WinSMS. We envision
TKL-WinSMS as a fundamental and indispensable part for
WiFi-based IPS to cope with various environmental dynam-
ics and achieve a robust localization service consistently.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0
2
4
6
8
m
SVR
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0
2
4
m
SVR-WinSMS
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Testing Point Index
0
2
4
m
TKL-WinSMS
Figure 4: Localization accuracy at each testing
point.
Acknowledgement
This research is funded by the Republic of Singapore Nation-
al Research Foundation (NRF) through a grant to the Berke-
ley Education Alliance for Research in Singapore (BEARS)
for the Singapore-Berkeley Building Efficiency and Sustain-
ability in the Tropics (SinBerBEST) Program. BEARS has
been established by the University of California, Berkeley as
a center for intellectual excellence in research and education
in Singapore.
5. REFERENCES
[1] S. Kumar, S. Gil, D. Katabi, and D. Rus. Accurate
indoor localization with zero start-up cost. In
Proceedings of the 20th annual international conference
on Mobile computing and networking, pages 483–494.
ACM, 2014.
[2] M. Long, J. Wang, J. Sun, and P. S. Yu. Domain
invariant transfer kernel learning. Knowledge and Data
Engineering, IEEE Transactions on, 27(6):1519–1532,
2015.
[3] D. Lymberopoulos, J. Liu, X. Yang, R. R. Choudhury,
V. Handziski, and S. Sen. A realistic evaluation and
comparison of indoor location technologies: experiences
and lessons learned. In Proceedings of the 14th
International Conference on Information Processing in
Sensor Networks, pages 178–189. ACM, 2015.
[4] H. Zou, B. Huang, X. Lu, H. Jiang, and L. Xie. A
robust indoor positioning system based on the
procrustes analysis and weighted extreme learning
machine. Wireless Communications, IEEE
Transactions on, 15(2):1252–1266, 2016.
464