Content uploaded by Han Zou
Author content
All content in this area was uploaded by Han Zou on Oct 20, 2015
Content may be subject to copyright.
Platform and Algorithm Development for a RFID-Based Indoor
Positioning System
Han Zou
*,
‡, Lihua Xie
*,
§
, Qing-Shan Jia
†
,
¶
, Hengtao Wang
†
,
||
*
Nanyang Technological University, Singapore 639798
†
Tsinghua University, Beijing, China 100084
In recent years, developing Indoor Positioning System (IPS) has become an attractive research topic due to the increasing demands on
Location-Based Service (LBS) in indoor environment. Several advantages of Radio Frequency Identification (RFID) Technology, such as
anti-interference, small, light and portable size of RFID tags, and its unique identification of different objects, make it superior to other
wireless communication technologies for indoor positioning. However, certain drawbacks of existing RFID-based IPSs, such as high cost of
RFID readers and active tags, as well as heavy dependence on the density of reference tags to provide the LBS, largely limit the application
of RFID-based IPS. In order to overcome these drawbacks, we develop a cost-efficient RFID-based IPS by using cheaper active RFID tags
and sensors. Furthermore, we also proposed three localization algorithms: Weighted Path Loss (WPL), Extreme Learning Machine (ELM)
and integrated WPL-ELM. WPL is a centralized model-based approach which does not require any reference tags and provides accurate
location estimation of the target effectively. ELM is a machine learning fingerprinting-based localization algorithm which can provide
higher localization accuracy than other existing fingerprinting-based approaches. The integrated WPL-ELM approach combines the fast
estimation of WPL and the high localization accuracy of ELM. Based on the experimental results, this integrated approach provides a
higher localization efficiency and accuracy than existing approaches, e.g., the LANDMARC approach and the support vector machine for
regression (SVR) approach.
Keywords: Indoor positioning system; RFID; Weighted Path Loss (WPL); Extreme Learning Machine (ELM).
1. Introduction
Nowadays, the popularity of social networks and the wide-
spread usage of mobile devices stimulate the huge demands
on Location-Based Service (LBS) in both indoor and outdoor
environment. Global Positioning System (GPS) provides
marvelous LBS in outdoor environment. However, due to
the lack of line-of-sight (LoS) transmission channels be-
tween a satellite and a receiver, and the attenuation and
scattering of microwave signals [1], GPS is not capable of
providing positioning service with sufficient localization
accuracy in indoor environment. Hence, developing an
Indoor Positioning System (IPS) to provide reliable and
precise indoor positioning and navigation becomes a hot
research topic recently. It is worth noticing that a lot of
problems, such as multipath effect of signal reflection from
walls and furniture, physical layout changes of furniture
and signal scattering due to large density of obstacles,
make positioning and navigation in indoor environment
much more complicated and challenging than in outdoor
environment.
Various wireless communication technologies have been
proposed and developed in order to provide indoor posi-
tioning and navigation, including Infrared, Bluetooth, ultra-
sound, Wireless Local Area Network (WLAN), Ultra-Wideband
(UWB) and Radio Frequency Identification (RFID) [2–4]. Ac-
tive Badge is an IPS making use of diffuse infrared technology
to realize indoor localization [5]. The major disadvantages
of Infrared-based IPSs such as Active Badge come from the
requirements of LoS and short-range transmission of in-
frared signal. Due to the compatibility of Bluetooth tags and
Received 21 February 2014; Revised 27 May 2014; Accepted 27 May 2014 ;
Published 9 July 2014. This paper was recommended for publication in its
revised form by editorial board member, Wendong Xiao.
Email Addresses: ‡zouhan@ntu.edu.sg,§elhxie@ntu.edu.sg,¶jiaqs@tsinghua.
edu.cn,kwangt07@mails.tsinghua.edu.cn
Unmanned Systems, Vol. 2, No. 3 (2014) 279–291
#
.
cWorld Scientific Publishing Company
DOI: 10.1142/S2301385014400068
279
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
readers with the majority of mobile devices, Bluetooth-
based IPS has also been proposed recently [6]. However, the
communication range of Bluetooth is usually shorter than
10 m, thus largely limiting its application for indoor posi-
tioning. IPSs by adopting ultrasound technology such as
Cricket [7] and Active Bat [8] have been proposed over the
recent decades. Time-of-flight measurement techniques are
leveraged by these IPSs to provide indoor location infor-
mation. The involved huge cost and large scalability of these
systems'infrastructure requirements make them inacces-
sible for common indoor localization applications. With the
rise of wireless communication, WLAN has also been
exploited to estimate the indoor location of a mobile target
for years. By reusing the existing WLAN infrastructure,
WLAN-based IPSs (RADAR [9], Ekahau [10]) requires only
little deployment cost. However, the current optimal local-
ization accuracy of WLAN-based IPS can only reach 3 m,
which fails to meet the positioning precision requirements
in some circumstances. In contrast, UWB-based IPS can
provide a higher localization accuracy. For instance, the
optimal localization accuracy of Ubisense [12] can reach
30 cm. Nevertheless, the interference of metallic and liquid
materials can largely affect its performance. In addition, the
high manufacturing cost of UWB readers also limits its
application [4].
Compared with other technologies, RFID technology has
several advantages, such as no requirement of LoS, anti-
interference, and the fact that RFID tags are small and light
and most importantly, it can uniquely identify different
objects. It has been widely used in asset tracking, industrial
automation and medical care. The application of RFID
technology in developing IPS has become a hot research
topic in recent years. RFID-based IPSs such as SpotON [13]
and LANDMARC [14] can uniquely identify, localize and
track equipment and persons successfully.
However, several drawbacks hinder the further devel-
opment of existing RFID-based IPSs. One is the high cost of
RFID readers and the active RFID tags. In order to overcome
this, we develop a cost-efficient RFID-based IPS by using
cheaper active RFID tags and sensors. Unlike LANDMARC,
the signal strengths emitted from RFID tags in our system
are picked up by RFID sensors instead of RFID readers.
Another drawback of existing RFID-based IPS is that its
localization accuracy largely depends on the density of ref-
erence tags. Too many reference tags may result in increased
RF interferences. We proposed the Weighted Path Loss
(WPL) localization algorithm which does not require any
reference tags for real-time indoor localization in [15]in
order to address this problem. The WPL approach can be
classified as a centralized model-based localization algo-
rithm and it works as follows. The distance between the
tracking tag and each sensor is calculated based on a mod-
ified International Telecommunication Union (ITU) indoor
path loss (PL) model in the first place. Then the estimated
location of the tracking tag is obtained as the summation of
each sensor's weighting factor (reciprocal of the distance
between the tracking tag and each sensor) multiplied by its
physical location, provided all the physical locations of the
sensors are known. Based on the experimental results
shown in [15], the WPL approach can provide a higher
localization accuracy than existing RFID-based IPSs. In this
paper, we analyze the performance of WPL more compre-
hensively in terms of both localization accuracy and ro-
bustness under different scenarios.
In order to further enhance the localization accuracy of
our RFID-based IPS, another fingerprinting-based localiza-
tion algorithm: Extreme Learning Machine (ELM) was pro-
posed in [15]. It consists of two phases: offline phase and
online phase. During the offline phase, some RFID tags
are adopted as reference tags. We record the historical
Received Signal Strength (RSS) of these reference tags re-
ceived at each sensor as well as their physical locations. The
RSS vector and the corresponding location vector of these
reference tags are adopted as the inputs and the training
targets of ELM, respectively. Then, after the training process of
ELM, we can obtain an ELM model. During the online phase,
unlike the LANDMARC system, the reference tags are not re-
quired anymore. After feeding the RSS vector of the tracking
tag into the ELM model, the output given by ELM is the esti-
mated location of the tracking tag. Since ELM can be classified
as a machine learning fingerprinting-based localization al-
gorithm, we further compare the performance of ELM with
other machine learning localization algorithms to obtain an
overall performance evaluation of ELM in this paper.
Since the model-based approaches can provide a location
estimation of a target in a short time and fingerprinting-
based approaches can provide a higher localization accuracy
in general, following this idea, we proposed another locali-
zation algorithm: WPL-ELM in [16], which integrates the
fast estimation of WPL and the high localization accuracy of
ELM together. During the offline phase, the indoor envi-
ronment is divided into small zones first and an ELM model
is developed for each zone. During the online phase, the
WPL approach is used to determine the zone of the target
primarily, then the ELM model of that zone is deployed to
provide the final estimated location of the target. We eval-
uate and present a more elaborate performance assessment
of WPL-ELM in terms of average localization accuracy, dis-
tance error distribution and cumulative percentile of error
distance in this paper.
The rest of the paper is organized as follows. The state-
of-the-art of RFID-based IPSs and indoor localization algo-
rithms are investigated in Sec. 2. In Sec. 3, the background
knowledge for this paper is provided. Section 4demon-
strates a system overview of our RFID-based IPS first, fol-
lowed by the algorithm formulation of the three proposed
280 H. Zou et al.
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
localization algorithms. In Sec. 5, we present the experi-
mental results and evaluation of the proposed algorithms.
The conclusion and future work are given in Sec. 6.
2. Related Work
2.1. RFID-based IPS
A typical RFID-based IPS consists of three basic compo-
nents: RFID readers, RFID tags and the interconnecting
communication network. Both RFID readers and tags use a
predefined RF frequency and protocol to transmit and re-
ceive data. The RFID reader is able to read the data emitted
from RFID tags. RFID tags can be classified into two cate-
gories: passive and active tags.
Passive RFID tags operate without a battery and are
mainly used to replace the traditional barcode technology.
A variety of RFID-based IPSs by adopting passive RFID
tags have been proposed in recent two decades [17,18].
Although they are lighter and less expensive than active
tags, the range of the passive RFID tags is limited to ap-
proximately 1 to 2 m which largely restricts the coverage
area of their system [19].
Active RFID tags are small transceivers equipped with
button-cell batteries. They can actively transmit their ID
and additional information to RFID readers. In contrast to
passive RFID tags, a typical active RFID tag enables long
transmission range of 30 m or more with the help of an
onboard radio and a small antenna, rendering it quite
suitable for identifying and tracking high-unit-value pro-
ducts or persons in complex indoor environments. RFID-
based IPSs which use active RFID tags such as SpotON
[13] and LANDMARC [14] have been proposed in recent
decades. SpotON is a fine-grained IPS based on RFID signal
strength. SpotON tags are custom devices that operate
standalone or potentially as a plug in card enabling larger
devices to take advantage of location-sensing technology. It
can provide 3D location information of the tag. LANDMARC
is one of the earliest and most famous IPSs by using active
RFID tags and readers. In order to increase accuracy with-
out placing more readers, extra fixed location reference tags
are introduced in LANDMARC to facilitate location calibra-
tion. By collecting the RSS from each tag to readers, an RSS
radio map is built. The system receives the RSS data from
both reference tags and tracking tags in real time. After
comparing the RSSs of the tracking tags with those of ref-
erence tags, the weighted k-nearest neighbor algorithm
is adopted to estimate the locations of the tracking tags. It
is reported that the localization accuracy of LANDMARC is
around 1.5–2 m with 50% probability. An enhanced LAND-
MARC approach has been proposed in [20] aiming to make
the calculated coordinate of the tracking tags closer to the
real-time measurements without extra readers and refer-
ence tags.
2.2. Indoor localization algorithms
With the booming development of leveraging various
wireless communication technologies to provide indoor
positioning and navigation, indoor localization algorithms
have been extensively studied and numbers of approaches
have been proposed over the past two decades [3,4]. In
general, indoor localization algorithms can be classified into
two categories: model-based approaches and fingerprint-
ing-based approaches.
Model-based approaches. This type of localization
algorithms calculates the location of mobile targets based
on geometrical models. For instance, the log-distance PL
model is used to establish the relationship between the
measured RSS and the Radio Frequency (RF) propagation
distance [21,22]. Several model-based approaches employ-
ing radio propagation models have been investigated in
[23]. The average localization accuracy of these IPSs is
around 5 m.
Besides the RSS related model, other geometric models
have been utilized to characterize the relationship between
signal transmitters and receivers. For example, PinPoint
[24] is based on Time of Arrival (ToA), Cricket [7] on Time
Difference of Arrival (TDoA), and VOR [25] on Angle of
Arrival (AoA). In general, model-based approaches consume
less time than fingerprinting-based approaches to estimate
the location of a mobile target.
The WPL approach we propose in this paper can be
classified as a model-based approach, since its location
estimation process involves the modified ITU indoor PL
model.
Fingerprinting-based approaches. Another type of
localization approaches adopt fingerprint matching as the
basic scheme. They usually involve two phases: an offline
training phase and an online localization phase. During the
offline training phase, a site survey dedicated to measuring
the RSS fingerprints at some known locations is performed
in the indoor environment and consequently, a RSS finger-
print database is built up. During the online localization
phase, when a user sends a location query containing his or
her current RSS fingerprint, the location of the user will be
estimated by matching the measured fingerprint with the
fingerprints stored in the database, and the location asso-
ciated with the matching fingerprint will be returned as his
or her location estimate.
A majority of these approaches leverages RF signals for
RSS fingerprinting. To name a few, RADAR [9] and Horus
[11] are based on WiFi signal, while LANDMARC [14] uti-
lizes RFID signal; FM radio [26], geomagnetism [27] and
GSM signals [28] are also adopted as fingerprints for indoor
Platform and Algorithm Development for a RFID-Based IPS 281
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
localization. Although fingerprinting-based approaches re-
quire a site survey to build up a fingerprint database during
the offline phase, they can provide a higher localization
accuracy than model-based approaches generally.
The ELM approach we propose in this paper is classified
as a machine learning fingerprinting-based approach. The
integrated WPL-ELM approach we propose combines the
advantages of both approaches.
3. Background Knowledge
3.1. Indoor PL model
The most commonly used PL model for indoor environ-
ments is the ITU Indoor Propagation Model [29]. It provides
a relation between the total path loss PL (dBm) and dis-
tance d(m) as:
PL ¼20 logðfÞþ10logðdÞþcðk;fÞþX;ð1Þ
where f(MHz) is the radio frequency, cis an empirical floor
loss penetration factor, kis the number of floors between
transmitter and receiver, is the pass loss exponent, and X
normally represents a Gaussian random noise with stan-
dard deviation . The signal propagation conditions depend
on different indoor environments due to multipath fading
and shadow fading. Therefore, the pass loss exponent
which ranges from 2 to 4 dependent on the layout of indoor
environment should be determined empirically.
The operating frequency of our RFID IPS is 2.4 GHz and
kis 1 in our case since all the RFID sensors and tags are
put on the same floor. After calculating the related terms
20 logðfÞand cðk;fÞin (1), and summing with the constant
term, the indoor PL model can be further expressed as:
PLðdÞ¼PL0þ10logðdÞþX:ð2Þ
With PL0as the reference pass loss coefficient.
3.2. Extreme learning machine (ELM)
ELM is a kind of machine learning algorithm based on a
Single-hidden Layer Feedforward neural Network (SLFN)
architecture. It has been proved to provide good generali-
zation performance at an extremely fast learning speed
[30]. In [31], WLAN IPS by using the ELM approach has
been provided to give a better performance in terms of both
the efficiency and the localization accuracy.
The outputs with L hidden nodes in SLFNs can be
represented as:
yNðxÞ¼X
L
i¼1
igiðxÞ¼X
L
i¼1
iGðai;bi;xÞ;ð3Þ
where ai,biare the weights and bias connecting the input
nodes and the ith hidden node, iare the output weights
connecting the ith hidden node and the output nodes, and
Gðai;bi;xÞis the activation function which gives the output
of the ith hidden node with respect to the input vector x.
In order to widen the application range of ELM, [32]
shows that a SLFN with atmost Nhidden nodes and with
almost any nonlinear activation function can exactly learn N
distinct observations. Given Narbitrary distinct training
samples ðxj;tjÞ;j¼1;2;...;N, by substituting xwith xjin
(3) we obtain
H¼T;ð4Þ
where
H¼
Gða1;b1;x1Þ... GðaL;bL;x1Þ
.
.
.
...
.
.
.
Gða1;b1;xNÞ... GðaL;bL;xNÞ
2
6
6
4
3
7
7
5NL
;ð5Þ
¼
T
1
.
.
.
T
L
2
6
6
4
3
7
7
5Lm
and T¼
tT
1
.
.
.
tT
N
2
6
6
4
3
7
7
5Nm
:ð6Þ
In the above, His the hidden layer output matrix of ELM;
the ith column of His the ith hidden node's output vector
with respect to inputs x1;x2;...;xN, and the jth row of His
the output vector of the hidden layer respect to the input
vector of xj.
Unlike the traditional training algorithms for neural
networks which need to adjust the input weights and hid-
den layer biases, [30] has proved that the parameters
of SLFN can be randomly assigned provided that the
activation function is infinitely differentiable. Therefore, the
hidden layer output matrix H remains unchanged once
these parameters are randomly initialized. To train a SLFN
is simply equivalent to finding an optimal solution LS
of ð4Þas:
jjLS Tjj ¼ min
jjHTjj:ð7Þ
The optimal solution of the above equation can be found
as LS ¼H†T, where H†is the Moor–Penrose generalized
inverse of H.
4. Proposed Approaches
4.1. System overview
The main components of the RFID IPS developed on our
own consists of numbers of RFID sensors, active RFID tags,
282 H. Zou et al.
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
a wireless sensor network that enables the communica-
tion between these devices, a RFID coordinator and a
location server. Unlike the LANDMARC system, the signal
strengths emitted from tags are picked up by RFID sen-
sors instead of RFID readers in our system, due to the
high price of RFID readers. The RFID coordinator in our
IPS is modified from a RFID sensor to collect the Received
Signal Strength Indication (RSSI) data from all RFID
sensors. All the RFID sensors, tags and coordinator use
TICC2530 as the wireless module. The manufacturing cost
of each RFID sensor is only $15, much less than the cost
of a typical commercial RFID reader. By using the 3 V coin
cell, the battery life of each tag is around one month.
Figures 1(a)–1(c) shows a typical RFID sensor, a RFID
tag and a RFID coordinator developed by us respectively.
The system communication protocol is based on ZigBee
2.4 GHz. Before system operation, each active RFID tag is
preprogrammed with a unique 4-character ID for identi-
fication by sensors. In addition, we found that the value of
RSS obtained by the same sensor from different tags at an
identical location may be different, possibly due to the
variation of the chips and circuits. Therefore, we made
some adjustments to ensure that the emitted powers of
all tags in our system are in a similar level. The following
is a brief operation procedure of our system.
First of all, RFID tags broadcast their unique ID signal
every second in the indoor environment. Then, RFID sen-
sors pick up the signal strength of each tag. With external
power supply, these sensors are able to send RSS informa-
tion of all tracking tags to the RFID coordinator continu-
ously through the wireless sensor network. The RSSI data
from all RFID sensors are received at the RFID coordinator
which is connected to the location server. In our experi-
ment, it is enough to use one RFID coordinator to cover a
100 m
2
indoor environment. After that, the location server
calculates the estimated location of each tracking tag by
using the proposed localization algorithms.
4.2. Methodology of WPL
Suppose we have ARFID sensors and Btracking tags. Each
sensor can pick up the signal strengths of all Btracking tags.
In order to calculate the estimated location of each tracking
tag, we define the signal strength of the jth tracking tag
received at the ith sensor as sij , where i2½1;A,j2½1;B.
The real position of the ith sensor is defined as ðxi;yiÞ.
Based on the PL Model defined in Sec. 2, the signal strength
sij can be expressed as:
sij ¼PLðdijÞ¼PL0þ10logðdij ÞþX:ð8Þ
Therefore, based on (8), the distance between the jth
tracking tag and the ith sensor can be calculated by:
dij ¼10
sijPL0X
10:ð9Þ
The distances between these ARFID sensors and the jth
tracking tag can be expressed as a dvector, given by
dj¼ðd1j;d2j;...;dAjÞT. The weighting factor of the ith
sensor with respect to the jth tracking tag is defined as:
wij ¼
1
dij
PA
i¼11
dij
:ð10Þ
The unknown location coordinate ðuj;vjÞof the jth tracking
tag is obtained by:
ðuj;vjÞ¼X
A
i¼1
wijðxi;yiÞ:ð11Þ
4.3. Methodology of ELM
The ELM approach considers the localization problem as a
regression problem. It consists of an offline phase and an
(a) (b) (c)
Fig. 1. RFID IPS mean components. (a) RFID sensor, (b) RFID tag and (c) RFID coordinator.
Platform and Algorithm Development for a RFID-Based IPS 283
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
online phase. During the offline phase, some RFID tags are
adopted as reference tags in order to build up an empirical
database. Preference tags will be used and Qhistorical
RSSI samples will be collected for each tag. Moreover, each
RSSI sample is denoted as ððXpq;Ypq Þ;RSSpqÞ,p2ð1;PÞ,
q2ð1;QÞ.ThevectorsRSS
pq,p¼1;2;...;P,q¼1;2;...;
Qare the inputs of the ELM and the corresponding loca-
tion vectors ðXpq;Ypq Þare the training targets of ELM. The
hard-limit transfer function is chosen as the activation
function. The training process of ELM is introduced in
Sec. 2. It can be conducted in the following three main
steps:
Step 1: Randomly assign values to hidden node para-
meters.
Step 2: Calculate the hidden layer output matrix H.
Step 3: Calculate the output weight by:
¼H†L;ð12Þ
where H†is the Moor–Penrose generalized inverse of H.
During the online phase, the only thing we need to do
is to feed the RSS vector which is contained in the RSSI
sample of the tracking tag into the ELM model. The out-
put given by ELM is the estimated location of the tracking
tag.
4.4. Methodology of integrated WPL-ELM
The model-based approaches can provide an location esti-
mation of a target in a short time since they do not require
any site survey during the offline phase [4]. On the other
hand, fingerprinting-based approaches can provide a higher
localization accuracy than the model-based approaches,
with an extra offline calibration [3,4]. Since WPL and ELM
can be classified as a centralized model-based approach and
afingerprinting-based approach, respectively, it is brilliant
if we can make use of both the fast estimation of WPL and
the high localization accuracy of ELM together for indoor
localization. Following this thought, another localization
algorithm: WPL-ELM integrating the advantages of both
WPL and ELM is proposed. The process of WPL-ELM is
shown in Fig. 2.
During the offline phase, big indoor space is divided into
multiple small zones according to the distribution of
the RFID sensors. Then, an ELM model for each zone is
developed.
During the online phase, the WPL approach is used to
determine the zone of the tracking tag primarily in the first
step. After we know the tracking tag is in which zone, the
ELM model of that zone is deployed in the second step to
provide the final estimated location of the target.
5. Experimental Results and Performance Evaluation
In order to evaluate the performance of the proposed
approaches, extensive experiments have been conducted.
The test-bed is the Internet of Things Laboratory in School
of Electrical and Electronic Engineering, Nanyang Techno-
logical University. The area of the test-bed is around
110 m2ð6:4m17:1mÞ. As shown in Fig. 3, there are 19
RFID sensors installed in the room. The positions of nine
tracking tags and the RFID coordinator are also shown in
Fig. 3.
Before the performance evaluation, we define the rea-
sonable range of received signal strength the RFID sensor
can pick up from RFID tags first. We put one tag directly
besides one sensor in order to estimate the maximum signal
strength a sensor can receive from a tag. After collecting
3600 RSSI samples in one hour, we find that the average
signal strength received by the sensor is around 42 dBm
with the standard deviation of 2.8 dBm. We also put one tag
at the right lower corner and one sensor at the left upper
corner of the room (as shown in Fig. 3) in order to estimate
the minimum signal strength a sensor can pick up from a
tag. The average signal strength received by the sensor is
around 98 dBm with the standard deviation of 5.6 dBm.
Therefore, we define the reasonable range of received signal
strength to be from 40 to 100 dBm for our system.
Fig. 2. Flowchart of integrated WPL-ELM approach.
284 H. Zou et al.
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
The distance error is used to measure the localization
accuracy of the system. We define the location estimation
error eto be the distance between the real location coor-
dinates ðx0;y0Þand the system estimated location coordi-
nates ðx;yÞ, as:
e¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðxx0Þ2þðyy0Þ2
p:ð13Þ
During experiment I, as shown Fig. 3, 19 reference tags
are distributed in the room. The main purpose of experi-
ment I is to build up the historical RSSI sample database for
ELM offline training. We keep collecting data of the signal
strength of the 19 reference tags from the 19 RFID sensors
for 10 days. We obtain 637,000 RSSI samples for each tag in
this experiment. We put these samples with their corre-
sponding real location coordinates into the ELM training
process and built up the ELM model in each zone for real-
time localization during the online phase.
During experiment II, we keep collecting data of the
signal strength of nine tracking tags from the 14 RFID
sensors for five days. The main purpose of experiment II is
to evaluate the localization accuracy of WPL approach, ELM
approach and the integrated WPL-ELM approach. We obtain
325,000 RSSI samples for each tag in this experiment.
The detailed experimental results are presented in 5.2, 5.3,
and 5.4.
5.1. Estimation of the PL exponent ®in WPL
The WPL approach largely depends on the PL exponent .
Therefore, an experiment is conducted to measure the RSSI
values of different distances from a RFID sensor in order to
find out the relationship between RSSI and distance. As
shown in Fig. 3, seven reference tags located on the left side
and the RFID sensor at the left upper corner of the test-bed
are selected in this experiment, since there are relatively
clearer LoS between the sensor and these tags. We measure
the signal strength at 1.50, 3.45, 5.06, 7.64, 10.64, 13.54 and
17.09 m. At each location, 3000 RSSI samples are collected
in 1 day. Figure 4shows the average signal strength of the
collected RSSI data at various locations.
Based on the data, we use a curve fitting method to
construct the relationship between RSSI and distance, as:
PLðdiÞ¼52:40 10 3:58 logðdiÞ;ð14Þ
i.e., the pass loss exponent is taken as 3.58 and the ref-
erence pass loss coefficient PL0as 52.40 dBm. We assume
that and PL0remain unchanged in the entire test period.
5.2. Performance evaluation of WPL
5.2.1. Localization accuracy
We keep collecting data of the signal strength of the nine
tracking tags from the 19 RFID sensors for seven days in
Fig. 3. Placement of RFID reference tags, tracking tags, sensors
and coordinator in experiment I.
Fig. 4. Relationship between RSSI and distance.
Platform and Algorithm Development for a RFID-Based IPS 285
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
experiment I in order to evaluate the localization accuracy
of the WPL approach. Since WPL is classified as model-
based approaches, the performances of LANDMARC [14]
and enhanced LANDMARC [20] are chosen to be compared
with WPL. Furthermore, LANDMARC and enhanced LAND-
MARC use the weighted k-nearest neighbor algorithm to
estimate the location of the tracking tags, we choose kequal
to the maximum number of reference tags in order to op-
timize the localization accuracy of these methods.
Based on the RSSI samples of all the tracking tags, we
collected during experiment I, the localization performance
comparison between LANDMARC, enhanced LANDMARC
and WPL are presented in Table 1and Fig. 5. As shown
in Table 1, the average localization accuracy by using
LANDMARC, enhanced LANDMARC and WPL is respectively
2.642, 1.990 and 1.651 m. WPL enhances the precision of
localization accuracy by 38% over LANDMARC and 17%
overenhanced LANDMARC. To conclude, WPL provides the
highest localization accuracy among the three approaches.
5.2.2. Robustness
In order to evaluate the robustness of WPL when the
number of RFID sensors is reduced, we turned offfive RFID
sensors in the test-bed during experiment II. As shown in
Fig. 6, we keep collecting data of the signal strength of the
nine tracking tags from the 14 RFID sensors for seven days
in this experiment. Then we compare the performance of
LANDMARC, enhanced LANDMARC and WPL using 14
sensors with experiment I database, where 19 sensors are
used.
The mean and variance of the location estimation error
of the three approaches with different number of RFID
sensors are presented in Table 2.Figure7demonstrates
the distance error distribution of the three approaches.
Table 1. Localization accuracy statistics.
Average
Approach localization accuracy (m)
LANDMARC 2.642
Enhanced LANDMARC 1.990
WPL 1.651
Fig. 5. Cumulative percentile of error distance for different
methods.
Fig. 6. Placement of RFID reference tags, tracking tags, sensors
and coordinator in experiments II and III.
Table 2. Comparison between WPL and other methods.
No. of RFID sensor 19 14
Approach Average Variance Average Variance
LANDMARC 2.642 0.108 2.973 0.191
Enhanced
LANDMARC
1.990 0.192 2.214 0.420
WPL 1.651 0.290 1.782 0.337
286 H. Zou et al.
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
As shown in Table 2, the localization performances of all
the three approaches become worse with lower density of
RFID sensors. However, it can be observed that the locali-
zation accuracy of WPL still remains the best among the
three approaches when five RFID sensors are removed
from the test-bed. Under this circumstance, WPL still
enhances the precision of localization accuracy by 40%
over LANDMARC and 19% over enhanced LANDMARC.
Compared with the results when 19 sensors are used, the
localization accuracy of LANDMARC, enhanced LANDMARC
and WPL decreases to 13%, 11% and 8%, respectively. The
performance decay of WPL is the smallest. As shown in
Fig. 7, the distance error distribution of LANDMARC in
Fig. 7(a) and enhanced LANDMARC in Fig. 7(b) are much
more scattered, while that of WPL is mainly limited within
2.7 m as shown in Fig. 7(c).
Thus we can safely conclude that WPL remains more
reliable and robust when the number of RFID sensors is
reduced.
5.3. Performance evaluation of ELM
Since ELM is a fingerprinting-based localization algorithm,
we build up a historical RSSI fingerprints database for ELM
offline training during experiment II in the first place. As
shown in Fig. 6, we keep collecting data of the signal
strength of the 19 reference tags and the nine testing tags
from the 14 RFID sensors for five days during the offline
phase. 318,500 RSSI samples for each tags are obtained in
this experiment.
For each of 19 reference tags, 5000 RSSI samples are
randomly chosen as training fingerprints from the experi-
ment II database. Here we choose 5000 RSSI samples of
each reference tag for ELM offline training process, con-
sidering the limitation of the number of input variables in
ELM. When the number of input variables is too large,
unnecessary hidden nodes parameters will be introduced
and cause ELM to be unstable and overfitted easily. In our
system, we found that 5000 input variables (RSSI samples
in our case) is appropriate for ELM training.
Besides the number of input variables, another param-
eter that could affect the localization accuracy of ELM is
the number of hidden nodes in the ELM hidden layer. The
localization accuracy of ELM can be improved with the in-
crease of the number of hidden nodes in the ELM hidden
layer. However, both training time and testing time also
increase. For instance, the ELM approach with 2500 hidden
nodes enhances the precision of localization accuracy by
33% over WPL but the testing time is as long as 1.937s,
which is too long for real-time localization. Thus, there is a
tradeoffbetween the localization accuracy and the testing
time when applying the ELM approach. Based on our
evaluation, we choose 2000 hidden nodes in the ELM hid-
den layer for ELM in our system.
Two classical machine learning algorithms, Back-propa-
gation (BP) algorithm and support vector machine for re-
gression (SVR) algorithm, are chosen in comparison with
ELM. After building up the ELM model, we evaluate the
performance of these three approaches based on the RSSI
samples of the nine tracking tags from the experiment II
database. The cumulative percentile of error distance for
the three approaches is shown in Fig. 8. Table 3demon-
strates the performance comparison between the three
approaches in terms of the training time, testing time and
average localization accuracy of the samples.
As observed from Table 3, ELM has an enormous ad-
vantage in training time and learning speed. It learns up to
391.94 times and 5.66 times faster than BP and SVR,
respectively. On the other hand, BP and SVR obtain shorter
testing time than ELM. The testing time of 1.825s is really a
drawback of ELM because it will introduce certain delay
in real-time localization of the tracking tags. It can be seen
in Table 3, the average localization accuracy of all nine
(a) (b) (c)
Fig. 7. Comparison of distance error distribution for different methods. (a) LANDMARC, (b) Enhanced LANDMARC and (c) WPL.
Platform and Algorithm Development for a RFID-Based IPS 287
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
tracking tags by using BP, SVR and ELM is 2.084, 1.769 and
1.198 m. ELM enhances the precision of localization accu-
racy by 43% over BP and 32% over SVR, respectively.
Figure 9demonstrates the distance error distribution of
the three different approaches. The distance error distri-
bution of ELM as shown in Fig. 9(c) ranges mainly within
3 m. In contrast, the distance error distribution of BP in
Fig. 9(a) and SVR in Fig. 9(b) are much more scattered.
To conclude, ELM has a tremendous advantage in offline
training time and online localization accuracy compared
with other approaches.
5.4. Performance evaluation of WPL-ELM
We conduct experiment III to evaluate the performance
of the integrated WPL-ELM approach. During the offline
phase, As shown in Fig. 6, the entire room is divided into
three small zones first. Zone 1 contains seven sensors,
eight reference tags and four tracking tags. Zone 2 contains
eight sensors, nine reference tags and three tracking tags.
There are five sensors, five reference tags and two tracking
tags in Zone 3. We keep collecting data of the signal
strength of the 19 reference tags and the nine testing tags
from the 14 RFID sensors for five days during the offline
phase. 318,500 RSSI samples for each tag are obtained in
this experiment. After that, for each of 19 reference tags,
5000 RSSI samples are randomly chosen as training fin-
gerprints from the experiment III database. These RSSI
samples with their corresponding physical location coor-
dinates are put into the ELM training process and are
adopted to build up the ELM model in each zone for real-
time localization.
Until all the ELM models for each zone are established,
we evaluate the performance of WPL-ELM based on the
RSSI samples of nine tracking tags from the experiment III
database. Since WPL is adopted as the preliminary esti-
mation of the tracking tag, we first analyze the reliability
Fig. 8. Cumulative percentile of error distance for different
methods.
Table 3. Comparison between ELM and other methods.
Approach Training time (s) Testing time (s) Accuracy (m)
BP 97200 0.007 2.084
SVR 1402.887 0.013 1.769
ELM 248.026 1.825 1.198
(a) BP (b) SVR
Fig. 9. Comparison of Distance Error Distribution for different methods.
288 H. Zou et al.
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
of WPL in classifying the tracking tags into the correct
zone. By evaluating all the 318,500 RSSI samples for each
tracking tags in experiment II, WPL can determine the
zone of the tracking tag with a 97.8% accuracy. With
fast estimation and 1.782 m localization accuracy of the
tracking tag, WPL is fully capable of providing the correct
zone of the tracking tag, or equally an estimate of its
location.
After we get the preliminary location estimation of the
tracking tag (tracking tag is in which zone), ELM is adopted
to provide final estimated location of the target by using
the ELM model of that zone which is developed during
the offline phase. The performance comparison between
ELM and WPL-ELM is shown in Fig. 10. The distance error
distribution of WPL-ELM as shown in Fig. 9(d) ranges
mainly within 2.3 m which is the best among the four
approaches.
Table 4demonstrates the performance comparison
between ELM and WPL-ELM in terms of the training
time, average testing time and average localization accu-
racy of the samples in each zone followed by the overall
performance. As observed in Table 4, the overall average
localization accuracy of WPL-ELM is 0.799 m, which
enhances the precision of localization accuracy by 62%
over BP, 55% over SVR and 33% over ELM, respectively.
In addition, the more noteworthy point is that WPL-ELM
largely reduces both training time during the offline phase
and testing time during the online phase as compared
with ELM. The overall training time of WPL-ELM is
147.089 s, saving 41% less time than ELM. The overall
testing time of WPL-ELM is 0.432 s, 4.22 times faster than
ELM. Therefore, WPL-ELM can overcome the drawback of
ELM, namely the tedious testing time during the online
phase.
Table 4. Comparison between WPL-ELM and ELM.
Approach
Training
time (s)
Testing
time (s) Accuracy (m)
ELM 248.026 1.825 1.198
WPL-ELM
Zone 1 59.338 0.524 0.763
Zone 2 62.057 0.428 0.901
Zone 3 25.694 0.252 0.719
Overall 147.089 0.432 0.799
Improvement 41% 76% 33%
(c) ELM (d) WPL-ELM
Fig. 9. (Continued )
Fig. 10. Cumulative percentile of error distance for different
methods.
Platform and Algorithm Development for a RFID-Based IPS 289
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
In summary, WPL-ELM can provide not only a higher
localization accuracy than other approaches, but also a
more efficient location estimation of the target than ELM.
6. Conclusion and Future Work
In this paper, we proposed a cost-efficient RFID IPS by using
cheaper active RFID tags and sensors. In addition to the
experimental results in [15,16] a more elaborate and com-
prehensive performance evaluation of the proposed three
localization algorithms: WPL, ELM, WPL-ELM was demon-
strated in this paper. Our experimental results show that
the WPL approach enhances the precision of localization
accuracy by 38% over LANDMARC and 17% over enhanced
LANDMARC. In addition, WPL is more robust when the
number of RFID sensors is reduced than existing approa-
ches. The ELM approach has tremendous advantages in off-
line training time and online localization accuracy compared
with other approaches. It improves the precision of indoor
localization by 43% over BP and 32% over SVR, respectively.
Considering the fast estimation by WPL and the high
localization accuracy by ELM, another localization algo-
rithm: WPL-ELM which integrates the advantages of both
approaches was also proposed in this paper. Based on our
experimental results, the training time and testing time of
WPL-ELM are 1.69 times and 4.22 times faster than ELM.
Furthermore, it improves the precision of indoor localiza-
tion by 62% over the BP approach, 55% over the SVR ap-
proach and 33% over the ELM approach, respectively. In
conclusion, WPL-ELM can provide a higher localization ac-
curacy of the target in a more efficient way than existing
approaches. Moreover, WPL-ELM can greatly reduce the
deployment cost of the entire system because it requires
less RFID sensors than WPL to maintain the same locali-
zation accuracy.
Future work can be focused on the exploration and
analysis of applying the WPL approach, the ELM approach
and the WPL-ELM approach on other IPSs, such as Infrared-
based IPS and WiFi-based IPS.
Acknowledgments
This research is funded by the Republic of Singapore Na-
tional Research Foundation through a grant to the Berkeley
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 educa-
tion in Singapore. The work of Q.-S. Jia is partially supported
by the National Science Foundation of China under grants
(Nos. 61174072, 61222302, 91224008, and U1301254),
the Tsinghua National Laboratory for Information Science
and Technology (TNLIST) Cross-discipline Foundation, and
the 111 International Collaboration Program of China (No.
B06002).
References
[1] B. Buchli, F. Sutton and J. Beutel, GPS-equipped wireless sensor
network node for high-accuracy positioning applications, in Wireless
Sensor Networks (Springer, 2012), pp. 179–195.
[2] H. Liu, H. Darabi, P. Banerjee and J. Liu, Survey of wireless indoor
positioning techniques and systems, IEEE Trans. Syst. Man, Cybern. C:
Appl. Rev. 37 (2007) 1067–1080.
[3] Y. Gu, A. Lo and I. Niemegeers, A survey of indoor positioning
systems for wireless personal networks, IEEE Commun. Surveys Tut.
11 (2009) 13–32.
[4] G. Deak, K. Curran and J. Condell, A survey of active and passive
indoor localisation systems, Comput. Commun. (2012).
[5] R. Want, A. Hopper, V. Falcao and J. Gibbons, The active badge
location system, ACM Trans. Information Systems 10 (1992)
91–102.
[6] F. Forno, G. Malnati and G. Portelli, Design and implementation of a
Bluetooth ad hoc network for indoor positioning, IEE Proc. Software
(2005), pp. 223–228.
[7] N. B. Priyantha, A. Chakraborty and H. Balakrishnan, The cricket lo-
cation-support system, in Proc. Sixth Ann. Intl. Conf. Mobile Comput-
ing and Networking (2000), pp. 32–43.
[8] M. Addlesee, R. Curwen, S. Hodges, J. Newman, P. Steggles, A. Ward
et al., Implementing a sentient computing system, Computer 34
(2001) 50–56.
[9] P. Bahl and V. N. Padmanabhan, RADAR: An in-building RF based
user location and tracking system, in Proc. IEEE INFOCOM, Vol. 2,
(2000), pp. 775–784.
[10] Ekahau (2013), Available at http://www.ekahau.com/.
[11] M. Youssef and A. Agrawala, The Horus WLAN location determina-
tion system, in Proc. Third Int. Conf. Mobile Systems, Applications, and
Services (2005), pp. 205–218.
[12] Ubisense (2013), Available at http://www.ubisense.net.
[13] J. Hightower, R. Want and G. Borriello, SpotON: An indoor 3D location
sensing technology based on RF signal strength, UW CSE 00-02-02,
Vol. 1, University of Washington, Department of Computer Science
and Engineering, Seattle, WA (2000).
[14] L. M. Ni, Y. Liu, Y. C. Lau and A. P. Patil, LANDMARC: Indoor location
sensing using active RFID, Wireless Netw. 10 (2004) 701–710.
[15] H. Zou, H. Wang, L. Xie and Q. Jia, An RFID indoor positioning system
by using weighted path loss and extreme learning machine, in Proc.
IEEE Int. Conf. Cyber-Physical Systems, Networks and Applications
(CPSNA 2013), August 2013, pp. 66–71.
[16] H. Zou, L. Xie, Q. Jia and H. Wang, An integrative weighted path loss
and extreme learning machine approach to RFID based indoor po-
sitioning, in Proc. IEEE Int. Conf. Indoor Positioning and Indoor
Navigation (IPIN 2013), October 2013, pp. 181–185.
[17] D. Lieckfeldt, J. You and D. Timmermann, Exploiting RF-scatter:
Human localization with bistatic passive UHF RFID-systems, 2009
IEEE Int. Conf. Wireless and Mobile Computing, Networking and
Communications (2009).
[18] B. Wagner, N. Patwari and D. Timmermann, Passive RFID tomo-
graphic imaging for device-free user localization, 9th Workshop on
Positioning, Navigation and Communication (WPNC) (2012).
290 H. Zou et al.
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.
[19] K. Kolodziej and J. Hjelm, Local Positioning Systems: LBS Applications
and Services (CRC Press, 2006).
[20] X. Jiang, Y. Liu and X. Wang, An enhanced approach of indoor location
sensing using active RFID, WASE Int. Conf. Information Engineering,
2009. ICIE'09 (2009), pp. 169–172.
[21] K. Chintalapudi, A. Padmanabha Iyer and V. N. Padmanabhan, Indoor
localization without the pain, in Proc. Sixteenth Annual Int. Conf.
Mobile Computing and Networking (2010), pp. 173–184.
[22] H. Lim, L.-C. Kung, J. C. Hou and H. Luo, Zero-configuration indoor
localization over IEEE 802.11 wireless infrastructure, Wireless Netw.
16 (2010) 405–420.
[23] D. Turner, S. Savage and A. C. Snoeren, On the empirical performance
of self-calibrating WiFi location systems, 2011 IEEE 36th Conf. Local
Computer Networks (LCN) (2011), pp. 76–84.
[24] M. Youssef, A. Youssef, C. Rieger, U. Shankar and A. Agrawala, Pin-
point: An asynchronous time-based location determination system,
in Proc. 4th Int. Conf. Mobile Systems, Applications and Services
(2006), pp. 165–176.
[25] D. Niculescu and B. Nath, Ad hoc positioning system (APS) using AOA,
in INFOCOM 2003. Twenty-Second Annual Joint Conf. IEEE Computer
and Communications (IEEE Societies, 2003), pp. 1734–1743.
[26] A. Matic, A. Papliatseyeu, V. Osmani and O. Mayora-Ibarra, Tuning to
your position: FM radio based indoor localization with spontaneous
recalibration, in Proc. IEEE Int. Conf. Pervasive Computing and Comm.
(PerCom), April 2010, pp. 153–161.
[27] J. Chung, M. Donahoe, C. Schmandt, I. Kim, P. Razavai and M. Wise-
man, Indoor location sensing using geo-magnetism, in Proc. Ninth Int.
Conf. Mobile Systems, Applications, and Services (MobiSys 11) (2011),
pp. 141–154.
[28] A. Varshavsky, E. de Lara, J. Hightower, A. LaMarca and V. Otsason,
GSM indoor localization, in Proc. IEEE Fifth Ann. Conf. Pervasive
Computing and Comm., Vol. 3, No. 6 (2007), pp. 698–720.
[29] T. Chrysikos, G. Georgopoulos and S. Kotsopoulos, Site-specific vali-
dation of ITU indoor path loss model at 2.4 GHz, in IEEE Int. Symp.
World of Wireless, Mobile and Multimedia Networks and Workshops,
2009. WoWMoM 2009 (2009), pp. 1–6.
[30] G.-B. Huang, Q.-Y. Zhu and C.-K. Siew, Extreme learning machine:
Theory and applications, Neurocomputing 70 (2006) 489–501.
[31] W. Xiao, P. Liu, W.-S. Soh and G.-B. Huang, Large scale wireless
indoor localization by clustering and Extreme Learning Machine,
2012 15th Int. Conf. Information Fusion (FUSION) (2012),
pp. 1609–1614.
[32] G.-B. Huang and H. A. Babri, Upper bounds on the number of hid-
den neurons in feedforward networks with arbitrary bounded
nonlinear activation functions, IEEE Trans. Neural Netw. 9(1998)
224–229.
Platform and Algorithm Development for a RFID-Based IPS 291
Un. Sys. 2014.02:279-291. Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/11/14. For personal use only.