32nd IEEE Conference on Local Computer Networks

Localization for indoor wireless networks using

minimum intersection areas of iso-RSS lines
Begumhan Turgut and Richard P. Martin
Department of Computer Science,
Rutgers University
110 Frelinghuysen Rd, Piscataway, NJ 08854
Email: {bturgut,rmartin}@cs.rutgers.edu
AbstractWe present a new method for localization in wireless
networks based on the measurement of the received signal
strength (RSS) from multiple access points in an indoor setting.
Our approach starts by learning a smoothed RSS surface for
each of the access points from a set of training data. We then
extract the isometric lines of the RSS surface (iso-RSS lines) for
each access point. To perform the localization, the user measures
the incoming signal strength of each access point and identies
the corresponding iso-RSS line. Ideally, the exact location would
be the common intersection point of these lines. However, noise
and measurement imperfections make the lines not intersect in a
single point. We search for the smallest rectangular area which
is intersected by all the selected iso-RSS lines. This rectangle is
interpreted as the most likely location of the user; the area of
the rectangle is an estimate of the localization error. We describe
an efcient method for nding the minimal intersection area
based on recursive grid partitioning. Through experiments over
multiple indoor data sets we show that our approach provides a
better localization accuracy than existing localization algorithms.
Index Termslocalization, wireless network, WLAN, signal
strength

I. I NTRODUCTION
Location estimation methods based on the measurement of
wireless signals allow us to determine the location of a device
in indoor settings where other positioning and ranging systems,
such as the Global Positioning System (GPS), can not be used.
Early localization systems using WLAN infrastructures
were point-based approaches, where the output of the localization is a single point, representing the likely location of
the device. Newer systems are using an area-based approach,
where the output is an area of the likely location.
The environment we are considering is an indoor area with
multiple wireless access points. We assume that from a given
point, the user can measure the signal of more than one
wireless access points. Localization systems are based on the
measurement of various physical properties of the signal, such
as received signal strength (RSS), the angle of the signals
arrival (AoA), time difference of arrival, and so on.
In this paper we present a new method for localization in
wireless networks based on the measurement of the received
signal strength (RSS) from multiple access points in an indoor
setting. Our approach starts by learning a smoothed RSS
surface for each of the access points from a set of training data.
We then extract the isometric lines of the RSS surface (isoRSS lines) for each access point. To perform the localization,

0742-1303/07 $25.00 2007 IEEE

DOI 10.1109/LCN.2007.147

the user measures the incoming signal strength of each access

point and identies the corresponding iso-RSS line. Ideally,
the exact location would be the common intersection point of
these lines. However, noise and measurement imperfections
make the lines not intersect in a single point. We search for
the smallest rectangular area which is intersected by all the
selected iso-RSS lines. This rectangle is interpreted as the
most likely location of the user; the area of the rectangle is
an estimate of the localization error.
The remainder of the paper is organized as follows. Section
II discusses the related work. Section III described our localization algorithm. In Section IV, through an experimental
study involving two real world data sets we measure the
performance of our localization algorithm and compare it
against current state of the art algorithms. Section V concludes
the paper.
II. R ELATED W ORK
Most of the proposed work in both indoor [1] and outdoor
[2] localization can be categorized based on their applications
such as location dependent services, resource allocation and
control, and the techniques used in positioning. Systems at
varying scales ranging from nding a tiny object inside a
room, to locating a car in a city block, have been built to
meet the ever-increasing demand. These varied requirements
have spawned many unique solutions including methodologies
based on measurements of Time of Arrival (ToA) [2], Received
Signal Strength (RSS) [1], [3], Time Difference of Arrival
(TDoA) [4] [5], Angle of Arrival (AoA) [6][8]. Other techniques use Distance Vectors (DV) [9] and Bayesian networks
[10].
Various aspects of the eld are being explored such as
security issues and attack detection. Defense methods against
attacks are described in [11], [12]. An approach for defense
against wormholes are described in [13][15]. Methods for
achieving better location estimation while maintaining other
system factors such as reliability are proposed in [4], [16]
[18]. As systems such as GPS cannot be used in indoor
settings, several approaches specializing in indoor localization
has been proposed in recent years [19][21].
As with most problems, the surrounding environment dictates which solution can be applied. When it comes to outdoor
localization ToA, RSS, TDoA and DV all work best when the

962

radio signals are not obscured by strong environmental hazards

such as metal frames, machine noise or thick walls. While this
might seem to make outdoor localization an easier problem,
most outdoor localization algorithms function without the
benet xed infrastructure and are generally solutions for ad
hoc wireless networks.
Indoor localization must take into account the structure
and inhabitants of the buildings. However, in indoor settings, we frequently have an existing wireless infrastructure,
which the algorithms can take advantage of. Most indoor
algorithms, therefore, focus on generating a map of signal
strength readings from the existing wireless infrastructure.
For instance, [20] describes an approach based on Bayesian
networks trained for the specic environment for determining
the location of a device based on the current measurements.

Perfect !

Fig. 1. An example of localization where the three iso-lines corresponding

to the three measured values intersect in a single point. This point will be
returned as the location of the device.

III. A N AREA BASED APPROACH FOR ESTIMATING

LOCALIZATION
1intersection
(a)

A. Introduction
We present an area based localization algorithm that makes
use of signal strength values to localize a device. Our algorithm will also return an evaluation of the localization error.
We will characterize the quality of the localization by two
measures: its accuracy and its precision. We dene accuracy
as the distance of the point returned by the localization
algorithm to the real location of the device. The precision
of the measurement describes the level of condence of the
algorithm in the result.
In the training phase of our algorithm we create a database
of iso-RSS lines. Let us assume that we want to perform
localization in a building with n wireless access points. First
we collect a series of measurements at a series of locations
spanning the whole building. For each location we collect a
vector of measurements {m1 , m2 . . . mn }. Using these measurements for each access point, we build a smoothed RSS
surface (a radio map). We extract the iso-RSS lines of these
surfaces. Note that due to the obstacles in the radio path, these
lines are usually not simple circles around the wireless access
points, but have a more complex shape, and frequently, they
are composed of several disjoint lines (multi-lines).
During the localization phase we measure the signal
strengths at an unknown location, and try to infer the location
using the database of iso-RSS lines created during training.
First we nd the n iso-RSS lines corresponding to the measurements. If these lines intersect in a single point, this point
would be the correct location, as shown in Figure 1. However,
in practice the lines might not intersect in the same point,
due to factors such as noise, measurement error, systematic
error, reections, or an imperfectly trained model. It is also
possible that the data set contains one or more signals which
were tempered with. Hence, the goal would be to nd a
measurement which is robust and not too reliant on accidental
factors.
In practice, instead of a single point in which all n iso-RSS
lines intersect, we nd a series of points in which two or more
lines intersect. One intuitive approach would be to believe that

2 intersections
(b)

0!
(c)

many intersections
(d)

Fig. 2. Intersections between iso-RSS lines. Two jagged iso-RSS lines can
intersect several times along the curves (d).

the correct location is somewhere in the area of the greatest

density of such partial intersections. However, we nd that
the number of intersections depend on the smoothness of the
iso-RSS lines. Jagged curves might intersect several times
in close proximity, as shown in Figure 2. In the following we
propose a more robust approach to nding the location of the
device.
B. Minimum intersection area
We denote an area A through which all the iso-RSS lines
pass as an intersection area. Naturally, a rectangle covering the
whole building is such an intersection area. We are interested
in nding the smallest of such areas, the minimum intersection
area (MIA). In this context, an intersection is dened as an
intersection between a rectangular area dened by upper right
point, width and height, between a set of multi-lines (iso-RSS
curves) and as a multi-line which is a series of interconnected
segments. We dene that a rectangle and an iso-RSS line
intersects if any of the segments of any of the multi-lines of
the iso-RSS line is inside of the rectangle.
Having n iso-RSS lines, we will call an intersection area of
degree i an area which is intersected by at least i lines. Naturally, an intersection area of degree i is also an intersection
area of degree i 1, but the reverse is not in general true. This
observation is important in situations when one of the signals
is very noisy, weak or has been tempered with. For instance,
if we have a scenario with 5 signals and we nd that the
minimum intersection area of degree 5 is very large, but the
minimum intersection area of degree 4 is much smaller, we can
assume that the signal which does not participate in that area
is unreliable, and we can exclude it from the measurements.
Another observation is that the minimum intersection area

963

might not be unique, due to the complex shape of the isoRSS lines. In general, the lower degree we consider, the larger
number of minimum intersection areas we nd.
Finding the exact minimum intersection area is a complex
problem of geometry. We can obtain a good approximation
by nding the smallest division of the measurement area
which would be the intersection area by repeatedly dividing
the measurement area according to a partitioning scheme.
The problem then becomes nding the optimal partitioning
algorithm that will return this minimum approximate area.
C. Finding the minimum intersection area via partitioning
Our goal is to nd the minimum intersection area (MIA)
which we can dene as the smallest rectangle intersected
by all the considered iso-RSS lines. However, the exhaustive
search for the MIA is very expensive. Our approach will be to
start with the largest possible intersection area, covering the
complete measurement area, and gradually consider smaller
and smaller intersection areas:
A1 , A2 , A3 . . . An
where area(Ai )  area(Ai+1 ), A0 is measurement area, and
every new Ai is generated through a partitioning rule from the
previous areas. The process stops when the partitioning rule
is not able to generate a smaller area which is an intersection
area. The expectation is that the partition rule will generate
an intersection area which is a good approximation of the
minimum intersection area. Not every partitioning scheme
is able to deliver on this. Let us consider a partitioning
scheme which partitions the rectangle in half along its longer
dimension as in the case of Figure 3 (a).
As shown in Figure 3 (a), we have an intersection area A
for the iso-RSS lines L1 , and L2 as the starting space. This
area is partitioned in areas A1 and A2 with neither of them
being intersection areas, as each of them contains only one of
the iso-RSS lines. Therefore, the algorithm will return A as
the smallest intersection area found.
Let us consider the proximity points of the two iso-RSS
lines, P1 , and P2 (the points which minimize the distance
dist(P1 ,P2 ) while P1 element of L1 and P2 element of L2 ).
Note from the gure that P1 and P2 are very close, hence we
can create an area Ax which is much smaller than A, which
would still be an intersection area. In fact, we can easily create
an area where the minimal iso-RSS area is arbitrarily small,
and still, the partition rule will insist that smallest intersection
area it can nd is the full rectangle A. Therefore, dividing the
rectangle along the longer edge is a very bad choice for the
partitioning rule.
We are looking for a partitioning rule that provides formal
guarantees that the smallest intersection area found by the
partitioning rule, will approximate the minimal intersection
area within a constant percentage.

subdividing both the height and the width of the rectangle

into three equal segments. Denote these nine segments as
A1 ,A2 ... A9 . We will consider the primary partition of the
rectangle A as to do the four identical segments A1 ,A2 ,A3
and A4 with width(Ai ) = 2/3 width(A) and height(Ai ) =
2/3height(A). It is worthwhile to point out that these are not
disjoint rectangles. Also, note that the area of the segment is
4/9-th of the area of A. Each of these rectangles is composed
of four rectangles of the secondary partition, for instance
A1 = A1 A2 A4 A5 .
The process of subdividing the space and searching for
the minimal intersection areas is as follows. We start with
a rectangle covering the complete measurement space. The
subdivision rectangles will have their linear dimensions as 2/3
of the original rectangle (and in consequence, their is 4/9 of the
original area) and are created according to the rules described
above. Next, we check the existence of iso-lines within the
segments. This refers to nding out how many of the isolines are passing through the space within the given segment.
Since our experimental data consists of ve RSS readings
from ve different access points (APs), corresponding example
is given as searching for a ve line intersection. Segments
which have at least four intersections are retained, while the
other ones are discarded. If at least one segment has the same
number of intersections as its parent, the parent is discarded.
Otherwise, the parent is put in the list of segments which are
candidates for being the minimum intersection area. Let us
also note that when counting intersections, we mark which
one is not intersecting. Later on, this can help us identify the
signal readings that are not reliable.
The segments are evaluated using a breadth-rst search
algorithm using a queue structure where newly generated
segments go to the end of the queue. The segments to be
partitioned next are picked from the beginning of the queue.
We have proved that this partitioning scheme yields an
approximate minimum intersection area which is, in the worst
case, less than two times the linear size of the minimum
intersection area.
E. Generalization: the (n-1)/n partition rule
The approximation provided by the 2/3 partition rule is still
a relatively loose bound. The question then becomes whether
we can devise a partitioning rule which obtains an arbitrarily
tight bound.
We can generalize the 2/3-s partition rule into the (n-1)/npartition rule in the following way. Let us consider a secondary
partition into n2 disjointed rectangles. The primary partition
n1
will be still in four rectangles, each of size ( n1
n w, n h).
We can prove that by increasing the value of n, we can obtain
an arbitrarily close approximation. Note, however, that the
complexity of the algorithm also increases with n.

D. The 2/3 partition rule

F. Using minimum intersection areas of variable degrees to

identify unreliable signals

As shown in Figure 3 (b), let us consider an area A, with a

secondary partitioning into 9 identical, disjoint rectangles, by

We have already mentioned that by comparing intersection

areas of various degrees we can identify signals which are

964

iso-RSS line L1

iso-RSS line L2

P1

P2

A1

A2

A'1

A'2

A'3

A'4

A'5

A'6

A'7

A'8

A'9

A1

Ax
(a) Partitioning Rules

A3

A2

A4

(b) The 2/3 partition rule

Fig. 3. (a) An incorrect partitioning rule based on the division of the area in half at every step. Despite of the existence of a very small minimum interception
area at the center of the gure, the smallest interception area found by this partitioning scheme is the whole measurement area.
(b) The 2/3 partition rule.

unreliable (for instance, because they are too weak or they

have been tempered with). Let us consider the four iso-RSS
lines in Figure 4. We note that the MIA of degree four is much
larger than the MIA of degree three. The reason behind this is
that iso-RSS line 4 is much farther than the other lines. We can
speculate that there is something wrong with this line, for
instance, it was tempered with. In this case, by removing this
line and considering only the remaining, trusted signals we can
have a much smaller MIA, corresponding to a higher precision
and, if our hypothesis was correct, also a better accuracy.
We will be using this nding later in our experiment to
show how localization results can be made more precise by
eliminating unreliable signal vector readings. More generally,
if
area(M IAr1 ...rn )  area(M IAr1 ...ri1 ,ri+1 ,...,rn )
then the signal ri should be considered unreliable and should
be removed from the localization algorithm.
IV. S IMULATION S TUDY
We have tested our approach through an implementation
which uses a combination of Matlab and Java code. In the
training step we create a grid surface for the RSS eld of
each AP. The grid surfaces are created as an interpolation of
the measured points which are used as the training data. This
is done using a triangle-based cubic interpolation of the data
points, through the griddata() function in Matlab 7. We
found that the errors in the data set lead to a very jagged
surface, therefore we perform a second set of smoothing using
a bi-cubic interpolation through the interp2() function.
This essentially involves a re-sampling of the surface at a lower
resolution, and then re-scaling the resolution of the surface. We
have found that a re-sampling at a resolution of 20ft yields the
smoothest surfaces, but a less aggressive smoothing, at 5ft,

Iso-RSS line
2

Iso-RSS line
1

Iso-RSS line
4

Iso-RSS line
3

Minimal intersection area for

iso-RSS lines {1,2,3}

Minimal intersection area for

iso-RSS lines {1,2,3,4}

Fig. 4.
An example of nding minimum intersection areas for 4 isoRSS lines. The iso-RSS line 4 is unreliable, thus the 3-rd degree minimum
intersection area is much smaller than the 4-th degree minimum intersection
area.

had in fact yielded better localization. An automatic way to

determine the right amount of smoothing is part of our future
work. The resulting surfaces are saved as the training data of
the localization algorithm.
Whenever the algorithm needs to identify the location of a
point, it will compute the ve iso-RSS lines corresponding to
the measurements from the ve APs. The iso-RSS lines are
the isolines on the previously saved grid surfaces.
A. Experimental data sets and an example localization
We have used two data sets collected from two different
buildings as the respective oor maps shown in Figure 5.
Our rst data set contains ve APs which were positioned in
the four corners and the center of a research laboratory (see
Figure 5-a). There were 254 readings from twenty rooms. The

965

second data collection was performed on the third oor of the

CoRE building at Rutgers University with four APs that were
somewhat linearly placed over the oor plan. There were 286
readings taken from hallways and rooms.
We start with looking at how the localization performs by
closely examining a single point location estimation example
from our rst data set. Recall that there were ve APs in the
rst data set, so given an RSS-vector which consists of ve signal strength values, each from a different AP, our goal is to nd
out which location these RSS readings belong to. Let us now
see what we can expect on a successful localization. Taking
the measurements at the location A=(153,120) and running the
localization algorithm we obtain an estimate for the location
a rectangle with the center at (154.296875, 120.703125), with
an area of 9.1552734375 square ft for a ve line intersection
requirement. This is a very good approximation of the correct
value of (154, 120). The output is represented with the isoRSS lines and the intersection areas. The high quality of
the approximation is due to the fact that this point is near
the center of the area, where the signal from all the APs is
comparatively strong.
Next, we examine how having factors such as noise, bias
or systematic errors would affect our result. We have seen
that if we can recognize that one or more of the AP readings
is unreliable, we can exclude them from the algorithm, and
thus obtain a better estimation. To mimic a systematic error,
we have articially changed the RSS values from one of
the APs and run the algorithm on these values. As from
the previous example, we have used node A to run this
test, by changing AP3s reading by 3dB. Once, we ran the
partitioning method, the program returned the estimate point as
(148.4375, 117.1875), with an area of 149.484375 square ft for
a ve line intersection requirement. Clearly, this is not a very
accurate result. Having such a large returned area indicates
that there can be an error in the readings, that is throwing us
off. Hence, we should perhaps eliminate this reading. We reran the algorithm with a four line intersection requirement.
The program returned the estimate point as (154.296875,
120.703125), with an area of 9.1552734375 square ft for a four
line intersection requirement with the returned area missing
AP3s iso-line.
B. Errors and error estimation on the research lab dataset
We have run a series of experiments using both of our
data sets. We have measured the absolute and the estimated
error of the data. We dene as the most probable location
based on the ve line intersection area as the center of the
smallest of the candidate intersection areas. The estimated
error of the measurement is the semidiagonal of the candidate
intersection area. If the area has the height h and width w
then the estimated error is equal to the semidiagonal of the
intersection area:

estimated =

h2 + w 2
2

(1)

The absolute error of the measurement is the euclidian

distance between the most probable location and the (known)
real location of the point.
There are two objectives we need to keep in mind when
performing localization: (a) minimize the absolute error and
(b) improve the precision of the estimation, while ensuring that
the estimation is higher than the absolute error. Note that in a
real deployment scenario the absolute error is not measurable.
C. Experiments over large set of points from the two data sets
In this experiment, we have run both the data sets with the
ve and the four line minimal intersection area algorithm for
a large set of points in order to identify outliers, unreliable
signal strength values and attempt to improve the cumulative
distribution function (CDF) of precision by means of returning
a smaller area. Table I shows the statistics for the research
laboratory data set only.
Most of the localization CDFs look similar in shape as our
CDF of precision, where there is a tail of 510% for most of
these algorithms. The goal is to decrease the value of this tail
as much as possible.
In this scenario, we ran the localization for 150 points out
of the 254 total available points in the research laboratory
data set. By looking at the maps of ve line intersection
areas, we were able to dene some points as poor, which
refers to points having poor accuracy or precision or both.
However, some of the points were not as clear cut to identify
as poor, hence we called them maybe poor ones and
incorporated accordingly in our statistics. Out of the total of
150 points localized as ve line intersection areas, we were
able to identify eleven points as poor points, meaning the
returned area was either too big or did not contain the point
of real location. In addition, ve points were labeled as being
maybe poor.
The algorithm was ran with four line intersections. We have
considered two cases. In the rst case, the points initially
marked as uncertain of being poor were in fact not poor, so
they were not included in the statistics. In the second case, the
points marked as uncertain of being poor were actually poor,
in return not included in the statistics.
Through visual observation of the iso-RSS lines we have
concluded that for seven out of the eleven points the returned
area can be made much smaller by removing one of the
lines from the localization algorithm. This would decrease the
percentage of the bad localizations from 7.3% to 2% which
is a substantial improvement. Similarly, if we considered the
points which are either poor or maybe poor the percentage
of bad localizations was improved from 10.6% to 4%.
Figures 6-8 present one of the best, medium, worst cases
from the research lab (left hand column) and the CoRE3
(right hand column) data sets. Please note that because all the
candidate sets returned from both the ve line, continuous line
rectangles, and the four line, dotted line rectangles, intersection
results are shown on the maps, we have multiple rectangles
showing the intersection areas.

966

(a) Research Lab - (225 x 144 feet)

(b) CoRE 3rd Floor - (200 x 80 feet)

Fig. 5. Floor maps of (a) a research lab and (b) the third oor of the CoRE building, showing the position of access points (APs) and points where signal
strengths were collected.
TABLE I
I MPROVING THE QUALITY OF THE LOCALIZATION BY IDENTIFYING BAD RSS
poor points
Number and percentage of bad localizations after the
initial phase (5-th degree minimum intersection area)
Estimate of how many bad readings are due to
one unreliable iso-RSS line (based on human visual
observation)
Number and percentage of points with previously
bad localization which were improved by removing
the unreliable iso-RSS line (thus using 4-th degree
minimum intersection area)
Number and percentage of bad localizations, after
removing the unreliable iso-RSS lines

In the best cases, since the returned area containing the real
location is very small, it is hard to see either the ve line or
the four line intersection areas. In the median cases, we can
visually identify both types of box and even though the real
location is within the returned area, in some cases, the size
of the box is too large, which in turn lowers our precision.
The worse cases clearly stood out as the size of the returned
ve line intersection area is close to the whole search space.
However, in most cases, seven to eight out of eleven points in
the research laboratory data set were improved by the four line
intersection as the dotted line rectangle shown is considerably
smaller in size and either contains the real location point or is
very close to it.
V. C ONCLUSIONS AND FUTURE WORK
In this paper, we presented a new area-based localization
method for wireless networks in an indoor setting. We use
the existing wireless infrastructure to collect RSS readings to
be later matched to an unknown location. We have quantied
our methods performance through a trace-driven experimental
study with accuracy and precision metrics. We have improved
the precision metric from which was previously known to be

READINGS .

11 (7.3%)

poor and maybe poor

points
16 (10.6%)

8 (5.3%)

can not estimate

8 (5.3%)

10 (6.6%)

3 (2%)

6 (4%)

5-10% to 24%. We have achieved an accuracy of 3 to 5 feet

(even though there are still outliers, due to bad readings in the
data set). We were able to recognize bad RSS readings based
on the returned area. Removing bad RSS readings improved
the accuracy of the algorithm.
Our future work includes the development of various techniques to improve the accuracy of our approach. One direction
of research would be, for instance, the automatic recognition
of the unreliable measurements, a work which currently needs
to be done through visual observation. Another promising
direction of research is to develop better smoothing techniques
for the RSS surface, using techniques such as robust estimation
or particle lters. One particularly promising direction is to use
existing knowledge about the structure of the building (such
as obstacles, walls, doors and so on) to improve the accuracy
of the localization.

967

R EFERENCES
[1] P. Bahl and V. Padmanabhan, RADAR: An in-building RF-based user
location and tracking system, in Proceedings of IEEE InfoCom, March
2000, pp. 775784.
[2] I. Getting, The global positioning system, IEEE Spectrum, vol. 30,
no. 12, pp. 3647, December 1993.

(a) Research lab data set. The minimum intersection area is basically a single point.

(b) CoRE3 data set. Although there are multiple minimum intersection areas, they are
very small and close together.

Fig. 6.

Example localization. Iso-RSS lines, 5-th and 4-th degree intersection areas for best localized points

(a) Research lab data set. Note how the 4-th degree intersection areas (dotted line

(b) CoRE3 data set. There are multiple MIAs, but they all contain the real location of

rectangles) are smaller, but there are more of them than 5-th degree MIAs. The real

the point.

location of the point is inside the rectangles.

Fig. 7.

Example localization. Iso-RSS lines, 5-th and 4-th degree intersection areas for medium localized points

[3] S. Saha, K. Chaudhuri, D. Sanghi, and P. Bhagwat, Location determination of a mobile device using IEEE 802.11 access point signals,
in Proceedings of IEEE Wireless Communications and Networking
Conference (WCNC), vol. 3, March 2003, pp. 19871992.
[4] N. Priyantha, A. Chakraborty, and H. Balakrishnan, The CRICKET
location-support system, in Proceedings of the 6th Annual International
Conference on Mobile Computing and Networking, August 2000, pp.
3243.
[5] A. Harter, A. Hopper, P. Steggles, A. Ward, and P. Webster, The
anatomy of a context-aware application, in Proceedings of 5th Annual
International Conference on Mobile Computing and Networking, August
1999, pp. 5968.
[6] D. Niculescu and B. Nath, Ad hoc positioning system (APS), in
Proceedings of Globecom, November 2001, pp. 29262931.
[7] D. Nicelescu and B. Nath, VOR base stations for indoor 802.11
positioning, in Proceedings of the 10th Annual ACM International Conference on Mobile Computing and Networking (MobiCom), September
2004, pp. 29262931.
[8] E. Elnahrawy, J.-A. Francisco, and R. P. Martin, Bayesian localization
in wireless networks using angle of arrival, in Proceedings of ACM
SenSys, November 2005, pp. 272273.
[9] D. Nicelescu and B. Nath, DV based positioning in ad hoc networks,
in Proceedings of Telecommunication Systems, 2003, pp. 267280.

[10] D. Madigan, E. Elnahrawy, R. Martin, W.-H. Ju, P. Krishnan, and A. Krishnakumar, Bayesian indoor positioning systems, in Proceedings of
Infocom, vol. 2, March 2005, pp. 12171227.
[11] S. Capkun and J. P. Hubaux, Secure positioning in
sensor networks, Ecole Polytechnique Federale de Lausanne
(EPFL), Tech. Rep. EPFL/IC/200444, 2004. [Online]. Available:
http://www2.imm.dtu.dk/ sca/secpos JSAC.pdf
[12] Y. C. Hu, A. Perrig, and D. Johnson, Packet leashes: a defense
against wormhole attacks in wireless networks, in Proceedings of IEEE
Infocom, March 2003, pp. 19761986.
[13] Z. Li, W. Trappe, Y. Zhang, and B. Nath, Robust statistical methods
for securing wireless localization in sensor networks, in Proceedings of
the 4th International Symposium on Information Processing in Sensor
Networks (IPSN), April 2005, pp. 9198.
[14] W. Xu, T. Wood, W. Trappe, and Y. Zhang, Channel surng and spatial
retreats: defenses against wireless denial of service, in Proceedings of
the 2004 ACM workshop on Wireless security (WiSe), October 2004, pp.
8089.
[15] N. Sastry, U. Shankar, and D. Wagner, Secure verication of location
claims, in Proceedings of the 2003 ACM Workshop on Wireless Security
(WiSe), September 2003, pp. 110.
[16] P. Krishnan, A. Krishnakumar, W.-H. Ju, C. Mallows, and S. Gamt, A
system for LEASE: Location estimation assisted by stationary emitters

968

(a) Research lab data set. The 5-th degree MIA covers 2/3 of the complete measurement

(b) CoRE3 data set. The 4-th degree MIA contains virtually the whole area. There are

area and it does not contain the point! The 4-th degree MIA, in the lower left corner is

multiple 3-th degree MIAs which provide a reasonably good approximation of the point,

much smaller, and although it still does not contain the point, it shows its approximate

although not all of them contain it.

location. The reason for the poor localization is the peripheral location of the point.

Fig. 8.

Example localization. Iso-RSS lines, 5-th and 4-th degree intersection areas for worst localized points

for indoor, in Proceedings of Infocom, vol. 2, March 2004, pp. 1001

1011.
[17] M. Youssef, A. Agrawal, and A. U. Shankar, WLAN location determination via clustering and probability distributions, in Proceedings
of IEEE Pervasive Computing and Communications (PerCom), March
2003, pp. 143150.
[18] S. Savvides, C.-C. Han, and M. Srivastava, Dynamic ne-grained
localization in ad hoc networks of sensors, in Proceedings of the
7th Annual ACM International Conference on Mobile Computing and
Networking (MobiCom), July 2001, pp. 166179.

[19] K. Kaemarungsi and P. Krishnamurthy, Modeling of indoor positioning

systems based on location ngerprinting, in Proceedings of Infocom,
vol. 2, March 2004, pp. 10121022.
[20] E. Elnahrawy, X. Li, and R. Martin, The limits of localization using
signal strength: A comparative study, in Proceedings of Secon, October
2004, pp. 406417.
[21] M. Youssef and A. Agrawala, The horus WLAN location determination
system, in Proceedings of the 3rd International Conference on Mobile
Systems, Applications, and Services, June 2005, pp. 205218.

969