2012 26th IEEE International Conference on Advanced Information Networking and Applications

A Wireless LAN Location Estimation System using Center of Gravity as an

Algorithm Selector for enhancing Location Estimation
Quan Jia Cheng, Joseph Kee-Yin Ng, and Kevin Chin Yiu Shum
Department of Computer Science
Hong Kong Baptist University
Kowloon Tong, Hong Kong
Email: qjcheng@comp.hkbu.edu.hk, jng@comp.hkbu.edu.hk, cyshum@comp.hkbu.edu.hk

signal strength readings collected in real-time and stored in

the data store, together with the physical positions of the
Wireless Access Points(APs), we can apply different estimation algorithms to find out the whereabouts of the mobile
devices. Some of the fundamental position techniques are
Location Fingerprinting [3], Propagation Loss Model [4],
Tri-lateration, Tri-angulations and Radio Maps with pattern
recognition approaches [5][7].
Signal Strength-based approaches are popular because
RSS can be easily obtained from APs, more so than other
features like angle and time of signal arrival. On the other
hand such systems are struggling to obtain their signal
propagation models, for the received signal strengths are
affected by specific environments [3] and are sensitive to
infrastructural changes [8]. In addition, the radio signal can
be blocked by the human body (referred to as the Body
Effect). [9]
In general, the relationship between Signal Strength Indicator (RSSI) values and the inter-distance between a mobile
device and the AP obeys the inverse square law but is also
affected by interference, noise, signal blockage, deflection
and reflection, etc.
The main focus in our system is the Centre of Gravity(CG)
method, for its ease of deployment and the low computation
cost in providing location estimation. Thus, the Centre of
Gravity method is adopted in our system just for demonstrating its effectiveness for network surveillance purposes
[10].

AbstractWith the prevalence of mobile Wi-Fi devices and

infrastructures, there are growing interests in mobile surveillance and device tracking for providing better location-aware
services in metropolitan areas. With a good location estimation
integrated into a wireless infrastructure, system administrators
can closely monitor the network traffic as well as the behavior
of the mobile users.
The Received Signal Strength(RSS), easily available information from Access Point(AP) Sensors, has become the
most popular research approach. However, in reality received
signal strength is affected by factors such as occlusion, signal
deflections and reflections. There had been proposed estimation
systems that use the Fingerprinting approach to provide good
and accurate location recommendation. But such systems have
been drawn back by their time-intensive training and retraining process. The solution to Signal Strength-based estimation,
therefore, is to devise a system that minimizes the training and
readaptation process while attaining good accuracy in location
estimation.
This paper proposes a location estimation system whose
estimation method is based on the Center of Gravity(CG)
method. This method also serves as an algorithm selector such
that the system can switch to another estimation algorithm if
need be. The aim of this system is to reduce the high cost of
training and re-calibration but attain an accuracy comparable
to the Fingerprinting location estimation approach.

I. I NTRODUCTION
Under the Governments Digital 21 Strategy [1] to endorse
the idea of a wireless city, the HKSAR government had put
forward a Wi-Fi Programme (GovWiFi) [2] as an initiative
to create a wireless infrastructure to facilitate mobile Internet
access for Hong Kong citizens and businesses. In view of the
heavy usage and progressively growing coverage of WLAN
within indoor environment, there is a need to investigate
the feasibility of using WLAN to locate a mobile user as
well as to pinpoint mobile devices in an indoor environment.
These estimated locations have a variety of uses, including
point-of-interest for infotainment, context-sensitive services
and detecting, locating and tracking hacker devices within a
WLAN.
Many Wireless LAN location estimation systems follow
the Signal Strength-based approaches, i.e. estimation is
performed based on the Received Signal Strengths(RSS) of
either the base stations or the client devices. With these
1550-445X/12 $26.00 2012 IEEE
DOI 10.1109/AINA.2012.110

II. R ELATED W ORKS

Many researchers have proposed methods in providing
location services using the mobile phone networks, but
few projects have actually been implemented. The RADAR
system [11] was one of the early WLAN-based location
estimation systems. Based on the FreeBSD distribution and
WaveLAN WLAN network, the RADAR system can locate
a user who carries a notebook with an accuracy of 5 meters.
Y. Wang et al [12] studied on the feasibility of making
use of the WLAN to locate a mobile device. They had an
empirical study inside their department building and labs
and reported their findings and simulation results. However,
261

there are no systematic ways to generalize these approaches

and fine tune the necessary environment parameters. Furthermore, the target environment is critical to the accuracy
of WLAN positioning. Radio characteristics in an open
environment are never static and there is not a single
methodology that can fine tune data from time to time and
adapt to the changing environment. It is only recently that
they start to study on post-deployment adaptation. They try
installing special hardware which monitors radio characteristics at different positions and rebuild the propagation model
periodically [8], [13].
Among the commercial products that are available for
WLAN positioning, Ekahau is the company that is taking
a leading position in in-door positioning. Their positioning
system relies on building an accurate radio map according
to the layout of the access points as well as the model and
specific characteristics of these access points. Nonetheless,
it is a costly system; and it falls victim to environmental
changes and post-deployment adaptation problems.

The Location Fingerprinting Approach is widely used in

the indoor location estimation and has been esteemed for its
great accuracy. However, applying the Location Fingerprinting is not practically scalable. It is because it is very difficult
to collect the RSS measurements of all the locations in a
large area and this large amount of measurements induces
high computation and storage costs.
In addition, the Location Fingerprinting is very sensitive
to the surrounding environments, thus re-calibration or recollection of data is often required. The Least-Squares
method was adopted in our experiment as the performance
measurement of fingerprint matching. Least-Squares tries
to derive a solution that minimizes the summation of squared
differences in signal strength between an AP Sensor and
each of the corresponding AP Sensors, i.e.
n
M X
= min( (sTi sti )2 )
j=1

where M is the number of fingerprints registered in the

database, N is the number of AP Sensor in the WLAN, sTi
represents the ith AP Sensor in the Training Data Set and
s( i)t represents the Signal Strength value of the ith AP
Sensor in the Testing Data Set.

A. Location Fingerprinting
Location Fingerprinting is one of the approaches which
utilize signal strength for location estimation. It is divided
into two phases, the off-line phase and the on-line phase.
Consider an AP sensor network consisting N sensors. In
the off-line phase, snapshots of the Received Signal Strength
(RSS) vectors are collected. Each snapshot of the RSS vector
contains the RSS, Si , received from the ith AP sensor
identified by the AP Identify Code, APi . The RSS vector
and its corresponding location (xi , yi ) are then stored in
the database. And in the on-line phase, the RSS vector of
the Hand-held wireless device or mobile station is measured
and generalized weighted distance Lp between the measured
RSS vector of the mobile station and the database entry is
computed with the following formula,
Lp = 1/N

N
X

p 1/p
([1/wi |si sm
i |] )

(2)

i=1

B. Aggregated Signal Layout

While Fingerprinting is an effective location estimation
method, it is not without its problems.
A Fingerprint is unique to its AP sensor network, i.e. the
number of AP sensors as well as their locations. If there is
a change of AP sensor locations, or an AP sensor removed
from the sensor network, we have to collect the signal
(data) and to train the Fingerprints database from scratch.
This drawback makes the Fingerprint method expensive
to maintain for AP sensor networks that changes sensors
frequently.
A similar location estimation approach called Aggregated
Signal Layout sets out to reduce the complexity of Fingerprinting approach. For each AP sensor in an AP sensor
network, we place AP calibrators at different distances from
the AP sensor and measure the received signal strengths. The
average signal strength(measured in RSSI) from the position
marker and the Position Marker Distance is recorded to
the Signal Layout of the AP sensor.Unlike the Fingerprint
method, Aggregated Signal Layout considers the Signal
Layout collected by AP sensors individually. The Signal
Layout of an AP sensor belongs to that AP alone, and is
independent of the surrounding AP sensors.
Aggregated Signal Layout performs location estimation
by overlapping layers of the Signal Layouts of AP sensors
in the sensor network. Consider a single AP sensor, given the
received signal strength from a mobile device, the distance
between the device and the AP sensor could be obtained
from the sensors Signal Layout; the inter-distance draws

(1)

i=1

where N is the number of the AP sensors received from

the location, p and wi are scalar factor and weighting factor
of the ith signal difference respectively. The weighting factor
wi is used as a bias which indicates the reliability of the
database entry or the RSS measurements. The location of
the mobile station is then estimated in the on-line phase by
either
Choosing the location in the database corresponding to
the fingerprint with the minimum distance to the measured fingerprint of the mobile station. For example,
the Manhattan distance, L1 and the Euclidean distance,
L2 with wi = 1 often used for all entries.
Estimating the location of the mobile station by averaging the M closest locations in the database with
smallest Lp.

262

the possible locations of the mobile device in a circular

perimeter. By aggregating the Signal Layouts of AP sensors,
essentially locating the common points of the perimeters, the
estimated location of the mobile device can be found.
Since it is usual that the received signal strength does
not conform to an AP sensors Signal Layout, we allow
some error to the distance estimation. For instance, we
allow a search window of +/-2dbm to our Aggregated Signal
Layout method, such that when an AP sensor receives signal
strength of -68dbm, the algorithm searches the sensors Signal Layout for Position Marker Distances that falls between
signal strengths -66 and -70dbm. This search window we
call Signal Spread is used to ensure more overlaps between
Signal Layouts of AP sensors.

III. T HE P ROPOSED M ETHOD

The Center of Gravity (CG) algorithm is generally praised
for its low computation cost (relative to the exhaustive
Fingerprint approach) and its ability to nullify the multipath effect within the WLAN set-up. Moreover, as the CG
formula is theoretically universal to the position of the
mobile device and to every AP Sensor, only one SignalDistance relationship is required to perform CG estimation.
While the accuracy of the CG method is good when the
user stays close to the center of the AP Sensor network, it
is bad when the user is close to the rim of the convex hull
defined by the APs involved. In view of this, we propose an
algorithm selector to choose the CG algorithm to estimate
a mobile users location when the user is not near the rim
of the convex hull, and to use another estimation algorithm
otherwise.
Firstly, in order to construct the said algorithm selector,
the approximate Signal-Distance Relationship for the target
environment is needed. For a better calibration result, an
AP calibrator is placed in an AP sensor network in such a
way that each AP sensor receives signals from the calibrator
from a different distance. The received signal strengths are
captured and stored in data store for further calculations.
At the end of calibration, with the average Received
Signal Strength of each AP sensor known, we have a
(signal, distance) data point for each AP sensor. We then
obtain a linear equation y = a + bx that best fit the data
points, with y-offsets defined as
X
R2 =
[yi f (a, b)]2
(5)

C. Center of Gravity
The Center of Gravity (CG) is also a location estimation
algorithm based on Received Signal Strength (RSS). It is a
location estimation algorithm focusing on areas with dense
AP sensors. The location estimation formula for CG is
defined as follows,

x=

b
b
b
(x1 sb
1 + x2 s2 + x3 s3 + ... + xn sn )
b
b
b
(sb
1 + s2 + s3 + ... + sn )

(3)

y=

b
b
b
(y1 sb
1 + y2 s2 + y3 s3 + ... + yn sn )
b
b
b
b
(s1 + s2 + s3 + ... + sn )

(4)

where (x, y) is the estimated location of the mobile user,

(xn , yn ) is the location of AP sensor n, and sn is the RSSI
from AP sensor n.
The CG algorithm based on the fact that received signal
strength is strong when one is near the AP, and signals get
weaker when one increases the inter-distance between the
mobile device and the AP. Hence, the user location can be
estimated by calculating the centroid position according to
the RSSI from all APs. The CG algorithm has proven to
be very effective and can provide outstanding performance
in metropolitan area during our mobile location estimation
experiments using the mobile phone network.
Although CG has proved its outstanding performance in
metropolitan areas, it can only estimate the mobile device
inside a convex hull, defined by the AP Sensor Network.
The estimated position by CG algorithm is restricted within
the polygonal area formed by joining the position of the
outermost AP sensors even though the actual position of the
mobile device is outside the area. Moreover, it seems that the
performance of the CG algorithm is related to the number
of AP sensors received. Thus, CG may lose its accuracy in
suburban areas where the number of AP sensors received
may not be as many as in metropolitan areas.

where yi is the distance in a (signal, distance) data

point and f (a, b) = a + bx. The a and b with minimal
R2 are obtained by linear Least Squares Fitting, in which
the solution to b and a are
b=

ssxy
ssxx

(6)

a = y + b
x

(7)

where
ssxx =

(xi x
)2 , ssxy =

(xi x
)(yi y)

(8)

The Signal-Distance Relationship is then used to derive a

Primary Estimation Zone(PEZ). Given at least three Access
Points circumscribing an area, the theoretical signal strength
received by an AP in a point inside the area can be
known with the Signal-Distance Relationship; then, using
the theoretical signal strength and the CG algorithm, the
estimated location for a point inside the area can also be
known. Thus, by calculating the estimated location along
paths that are slightly inside the edges of the area, the
Primary Estimation Zone is formed.

263

WRT54G WLAN router. An AP Sensor uses TCP socket

to send the data to control server, control server analyzes
the data collected and estimates the location on fly and
the information is stored into the database with a current
timestamp for logging purpose.
The experimental data is collected from the set-up as
follows. Our test bed is situated in lab FSC801D at HKBU
with monitors installed throughout the lab. Four AP Sensors,
APs 1-4, is placed at four corners at the ceiling forming an
AP Sensor Network(Figure 2). An AP Calibrator, placed
at desk level, is inside the network such that APs 1 to
4 continuously receive packets from the AP calibrator.
To capture signal strengths at different distances, in our
experiment the AP calibrator is placed at a position such
that the distances D1 to D4 are 4.3266m, 2.6832m, 7.2996m
and 9.8496m respectively.
While distances D1 to D4 are sufficient for calibration,
we intend to experiment the effect of including more
(signal, distance) data points in our algorithm. We include
3 additional distances from AP1 to AP4(D5 = 9.6m), AP1
to AP2(D6 = 4.8m) and AP2 to AP4(D7 = 12.5328m)
in our experiment. The total of 7 (signal, distance) data
points will form another Signal-Distance Relationship. We
collected Packet information within one-hour sessions.

As illustrated in Figure 1, the Primary Estimation Zone

formed by the 4 APs according to the above method would
divide the effective AP Sensor range into five zone: Zone
0, Zone 1, Zone 2, Zone 3 and Zone 4. With the help of
the APs and AP calibrators, live received signal strength
can be captured and real-time adaptation/calibration of the
various zones defined by the PEZ can be realized by making
a first estimation using the CG algorithm; if the estimated
location lies inside the PEZ, i.e. in Zone 0, the CG algorithm
is a good enough estimation algorithm; otherwise, another
algorithm is used.

Figure 1.

Zones defined by the CG Primary Estimation Zone

In addition, for better visualization and understanding

of signal strength characteristics within our test bed, we
construct a software module for visualizing the said PEZ, the
estimated mobile users location as well as signal-strengthbased Radio Maps.
IV. E XPERIMENT S ETUP
In our experiment, a programmable router - Linksys
WRT54G [14], which is burned with an open-source custommade Firmware, acts as an AP Sensor to obtain information
from data packets transmitted within the WLAN environment. Packet information such as Service Set Identifier
(SSID), Extended Service Set Identifier (ESSID), Received
Signal Strength Indicator (RSSI), Noise Level, Traffic rate
and Traffic Frequency are collected by a custom-made
wireless data acquisition application written for the Linksys

Figure 2. Experimental Setup. D1 = 4.3266m; D2 = 2.6832m; D3 =

7.2996m; D4 = 9.8496m; D5 = 9.6m; D6 = 4.8m

V. E XPERIMENT R ESULTS
In our experiment we used data samples from five onehour sessions; with distances and average RSSI of each AP
shown in Table 1. As seen by comparing the distances of

264

Table I
AP C ALIBRATION R ESULT

D1
D2
D3
D4
D5
D6
D7

Distance(m)
4.3266
2.6832
7.2996
9.8496
9.6
4.8
12.5328

Average RSSI(dbm)
-46.29
-44.22
-58.66
-57.96
-59.28
-59.28
-69.34

the AP calibrator from the AP Sensor and their respective Signal Strengths, while Received Signal Strengths, in
general, display an increasing trend with the decrease in
distance, there are certain points that do not follow this
general trend. Our experiments in the past has shown that
though deviations occur due to constructive and destructive
interference of radio signals, Received Signal Strengths are
inversely proportional to increasing distances (Figure 3).

Figure 4.
used(blue)

S-D Relationship 4-point calibration(red) and data points

Figure 5. S-D Relationship with 7-point calibration(red) and data points

used(blue)
Figure 3.

A Signal-Distance Relationship

try to discover the optimal constant value for our PEZ. We

have simulated with four estimation methods, namely,
1) Estimate with the original S-D Relationships
2) Estimate with the constant set to a small number, i.e.
-0.01
3) Estimate with the constant set to a value that best fits
the calibrators physical location
4) Estimate with the constant set to a value that gives the
least average error
For methods 3 and 4, we probe along the values between
constants between the original constant and -0.01, for apparently the Primary Estimation Zone in question cannot be
larger than that from method 2 and smaller than that from
method 1. Also, when we inspect the resulting error along
these values, we discover that there is a minimum value each
for methods 3 and 4(as indicated in Figures 8 and 9), which
proves the plausibility of both methods.
Simulation results are given in Table II, the error being the
average Euclidean distance between the estimated location

According to the calibration data, using linear Least

Squares Fitting, the approximate Signal(S)-Distance(D) Relationships can be written as s = 1.79d 35.80 for the
4 distances, D1 to D4(henceforth named S-D Relationship
with 4-point calibration) and s = 1.77d 38.12 for the
7 distances, D1 to D7(henceforth named S-D Relationship
with 7-point calibration). The graphical representation of SD relationships A and B are shown in Figures 4 and 5.
With the Signal-Distance Relationships established in real
time, we try to construct our algorithm selector i.e. the
Primary Estimation Zone derived from the S-D Relationship.
Figure 6 shows the PEZ conscribed by the S-D Relationship,
which shows a very restricted Zone 0. Since we want the
CG algorithm to be our main estimation algorithm, this is
not a desirable PEZ defined by the CG method under the
S-D Relationship.
As we observe that altering the constant term in our S-D
Relationship provides a more effective PEZ(Figure 7), we

265

Figure 8. Probing result for Method 3(4-point calibration); note that the
error is the distance between the estimated location of the calibrator and
the calibrators physical location

Figure 6.

Derived PEZ using the original S-D Relationship

Figure 9. Probing result for Method 4(4-point calibration); note that the
error is the average distance between the estimated location and the actual
location

Figure 7.

and the actual location.

From the table we can infer that, while method 2 gives
a Primary Estimation Zone that resembles our desired PEZ,
it may not provide the least error. We also observe that the
S-D Relationship with more data points only yields larger
error in method 1 but does not make much of a difference
in the others, as they have already adjusted the PEZ to give
a minimal error.
We also notice that, for methods 2 to 4, the most accurate
locations come from the corners and the center of the testing
area, while the others contribute the majority of the error.
Therefore, we intuitively categorize the points into two types
- points that are accurate(have error less than half of average

Derived PEZ using the modified S-D Relationship

266

Table II
P RIMARY E STIMATION Z ONE S IMULATION R ESULTS
Method
1
1
2
2
3
3
4
4

S-D Relationship
Constant
A
-35.80
B
-38.12
A
-0.01
B
-0.01
A
-5.36
B
-5.28
A
-3.24
B
-3.20
A: 4-point calibration
B: 7-point calibration

Error(m)
1.95324
2.00892
0.85266
0.85266
0.76422
0.76416
0.70338
0.70338

error), and points that are inaccurate(have error more than

or equal to half of average error), as indicated in Figure 10.

Figure 11.

Points estimated with Method4A and point adjustment

The Center of Gravity algorithm is accurate and efficient

but only so if the mobile device is well inside of the
convex hull circumscribed by AP sensors. With our proposed
system, if the first estimated location (by CG method) is
outside of the Primary Estimation Zone, we will switch to
another location estimation algorithm for those positions. We
will propose more appropriate algorithms as we progress
and aim at a hybrid system that can use the best of each
algorithm has to offer.
For better visualization and better understanding of signal
strength characteristics, we have constructed a software tool
for displaying information about mobile users as well as
Signal Layouts for investigation in order to discover better
algorithms for locating and tracking mobile users. Also, to
test the systems adaptability to different environments (e.g.
libraries), we will test our system with more variations in
terms of environment in the future.

Figure 10. Points estimated with Method 2A(left) and 4A(right); cyan
points have error less than half of average error

Since the Convex Hull of the AP Sensor network forms a

long rectangular shape, the long edges of the PEZ curve
towards the centre of the testing area, which induces a
larger error in the x-component of the estimated locations.
Averaging the x-coordinate of each inaccurate point, while
nowhere near the ground truth, would result in reduced
average error as shown in Figure 11. When applied to
Method 4(4-point calibration), the new average error would
be 0.53346m, a 24% improvement compared to the initial
error(0.70338m); and when applied to Method 2(4-point
calibration), the new average error would be 0.83526m, a
2% improvement compared to the initial error(0.85266m).

ACKNOWLEDGMENT
The work reported here was supported in part by the
Faculty Research Grant of HKBU. FRG2/09-10/056.

VI. C ONCLUSION

R EFERENCES

Our experiment aims to demonstrate the feasibility of using WLAN to locate and track mobile activities in an indoor
environment, by developing a location system that does not
possess adaptation issues and expensive maintenance. In this
paper, we proposed a location estimation system using two
or more signal strength-based location estimation methods
that, as a result, allows a more stable and more accurate
location estimation.

[1] Hong kong governments digital 21 strategy, http://www.

info.gov.hk/digital21/eng/index.htm.
[2] Government wi-fi programme (govwifi), http://www.gov.hk/
en/theme/wifi/program/index.htm.
[3] W. M. Yeung and J. K. Ng, An enhanced wireless lan
positioning algorithm based on the fingerprint approach,
in Proceedings of IEEE TENCON 2006, IEEE CS Press,
November 2006, pp. 1417.

267

[4] K. M. Chu, K. R. Leung, J. K. Ng, and C. H. Li, A

directional propagation model for locating mobile stations
within a mobile phone network, International Journal of
Wireless and Mobile Computing (IJWMC): Special Issue on
Applications, Services and Infrastructure for Wireless and
Mobile Computing, August 2008.
[5] P. Bahl and V. Padmanabhan, Radar: An in-building rfbased user location and tracking system, in Proceedings of
INFOCOM 2000, vol. 2, March 2000, pp. 775784.
[6] Z. li Wu, C. hung Li, J. K. Ng, and K. R. Leung, Location estimation via support vector kernel regression, IEEE
Transactions on Mobile Computing, March 2007.
[7] W. M. Yeung, J. K. Ng, and K. R. Leung, Large-scale
location estimation over gsm networks: the gear approach,
in Proceedings of the 24th IEEE International Conference
on Distributed Computing Systems Workshops (ICDCS-MDC
2004), March 2004, pp. 574579.
[8] P. Krishnan, A. Krishnakumar, W.-H. Ju, C. Mallows, and
S. Ganu, A system for lease: Location estimation assisted
by stationary emitters for indoor rf wireless networks, in
Proceedings of INFOCOM 2004, vol. 2, March 2004, pp.
10011011.
[9] T. King, S. Kopf, T. Haenselmann, C. Lubberger, and W. Effelsberg, Compass: A probabilistic indoor positioning system
based on 802.11 and digital compasses, in The 1st Int. Workshop on Wireless Network Testbeds, Experimental Evaluation
& Characterization (WiNTECH 06), September 2006, pp. 34
40.
[10] J. Zhou, Enhanced signal propagation models and algorithm
selector for providing location estimation services within
cellular radio networks, Ph.D. dissertation.
[11] P. Bahl and V. Padmanabhan, Radar: An in-building rfbased user location and tracking system, in Proceedings of
INFOCOM 2000, vol. 2, March 2000, pp. 775784.
[12] Y. Wang, H. L. X. Jia, and G. Li, An indoor wireless
positioning system based on wireless local area network
infrastructure, in Proceedings of the 6th International Symposium on Satellite Navigation Technology Including Mobile
Positioning & Location Services (SatNav 2003), July 2003.
[13] S. Ganu, A. Krishnakumar, and P. Krishnan, Infrastructurebased location estimation in wlan networks, in Proceedings
of the IEEE Wireless Communications and Networking Conference (WCNC 2004), vol. 1, March 2004, pp. 465470.
[14] Linksys wrt54g series, http://en.wikipedia.org/wiki/Linksys.

268