The Horus WLAN Location Determination System

Moustafa Youssef and Ashok Agrawala

Department of Computer Science
University of Maryland
College Park, Maryland 20742
{moustafa, agrawala}@cs.umd.edu

Abstract
We present the design and implementation of the Horus
WLAN location determination system. The design of the
Horus system aims at satisfying two goals: high accuracy
and low computational requirements. The Horus system
identifies different causes for the wireless channel variations and addresses them to achieve its high accuracy. It
uses location-clustering techniques to reduce the computational requirements of the algorithm. The lightweight
Horus algorithm helps in supporting a larger number of
users by running the algorithm at the clients.
We discuss the different components of the Horus system and its implementation under two different operating
systems and evaluate the performance of the Horus system on two testbeds. Our results show that the Horus
system achieves its goal. It has an error of less than 0.6
meter on the average and its computational requirements
are more than an order of magnitude better than other
WLAN location determination systems. Moreover, the
techniques developed in the context of the Horus system are general and can be applied to other WLAN location determination systems to enhance their accuracy.
We also report lessons learned from experimenting with
the Horus system and provide directions for future work.

Introduction

Horus is an RF-based location determination system. It

is currently implemented in the context of 802.11 wireless LANs [25]. The system uses the signal strength observed for frames transmitted by the access points to infer the user location. Since the wireless cards measure
the signal strength information of the received frames as
part of their normal operation, this makes the Horus system a software solution on top of the wireless network
infrastructure. There are two classes of WLAN location
determination systems: client-based and infrastructurebased. Both have their own set of applications. Horus is
Also

affiliated with Alexandria University, Egypt.

currently implemented as a client-based system. A large

class of applications [10], including location-sensitive
content delivery, direction finding, asset tracking, and
emergency notification, can be built on top of the Horus
system.
WLAN location determination is an active research
area [5, 6, 8, 9, 12, 13, 15, 17, 2022, 2931, 33, 34].
WLAN location determination systems usually work in
two phases: an offline training phase and an online location determination phase. During the offline phase, the
system tabulates the signal strength received from the access points at selected locations in the area of interest,
resulting in a so-called radio map. During the location
determination phase, the system use the signal strength
samples received from the access points to search the
radio map to estimate the user location.
Radio-map based techniques can be categorized into
two broad categories: deterministic techniques and probabilistic techniques. Deterministic techniques [5, 6, 22]
represent the signal strength of an access point at a location by a scalar value, for example, the mean value, and
use non-probabilistic approaches to estimate the user location. For example, in the Radar system [5, 6] the authors use nearest neighborhood techniques to infer the
user location. On the other hand, probabilistic techniques [8, 9, 13, 17, 20, 21, 2931, 33, 34] store information about the signal strength distributions from the access points in the radio map and use probabilistic techniques to estimate the user location. For example, the
Nibble system [8, 9] uses a Bayesian Network approach
to estimate the user location.
The Horus system lies in the probabilistic techniques
category. The design of the Horus system aims at satisfying two goals: high accuracy and low computational requirements. The Horus system identifies different causes
for the wireless channel variations and addresses them
to achieve its high accuracy. It uses location-clustering
techniques to reduce the computational requirements of
the algorithm. The lightweight Horus algorithm allows

0.5
0.45
0.4
Probability

0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
-52

-50

-48 -46 -44 -42 -40

Signal Strength (dBm)

-38

-36

Figure 1: An example of the normalized signal strength

histogram from an access point.

it to be implemented in energy-constrained devices. This

non-centralized implementation helps in supporting a
larger number of users. In this paper, we present the different components of the Horus system and show how
they work together to achieve its goals. We discuss our
Horus implementation under two different operating systems and evaluate its performance on two different indoor testbeds.
The rest of the paper is structured as follows: in the
next section, we describe the different causes of variations in the wireless channel. In Section 3 we present
the different components of the Horus system that deal
with the noisy characteristics of the wireless channel. We
present the results of testing the Horus system on two
different testbeds in Section 4. Section 5 presents our
experience while building the Horus system. In Section
6 we discuss related work. Finally, Section 7 concludes
the paper and provides directions for future work.

Wireless Channel Characteristics

In this section, we identify the different causes of variations in the wireless channel quality and how they affect the WLAN location determination systems. We are
mainly concerned with the variations that affect the received signal strength. We start by describing our sampling process. Then, we categorize the variations in the
wireless channel as temporal variations and spatial variations. We performed all the experiments in this section in
a typical office building, measured during the day when
people are around.

2.1

Sampling Process

A key function required by all WLAN location determination systems is signal-strength sampling. We used a

Lucent Orinoco silver network interface card (NIC) supporting up to 11 Mbit/s data rate [3]. The Horus system
is implemented under both the Linux and Windows operating systems.
For the Linux OS, we modified [1] the Lucent Wavelan driver so that it returns the signal strength of probe
response frames received from all access points in the
NIC range using active scanning [25]; our driver was the
first to support this feature.
The scanning process output is a list of the MAC addresses of the access points associated with the signal
strength observed in this scan (through probe response
frames). Each scans result set represents a sample.
We also developed a wireless API [1] that interfaces
with any device driver that supports the wireless extensions [2]. The device driver and the wireless API have
been available for public download and have been used
by others in wireless research.
For the Microsoft Windows operating system, we used
a custom-built NDIS driver to obtain the signal strength
from the wireless card (using active scanning). This
gives us more control over the scanning process as described in Section 5.
We now describe the different causes of variations in
a wireless channel. We divide these causes into two categories: temporal variations and spatial variations.

2.2 Temporal variations

This section describes how the wireless channel changes
over time when the user is standing at a fixed position.
2.2.1

Samples from one access point

We measured the signal strength from a single access

point over a five minute period. We took the samples one
second apart for a total of 300 samples. Figure 1 shows
the normalized histogram of the received signal strength.
In our experience, the histogram range can be as large as
10 dBm or more. This time variation of the channel can
be due to changes in the physical environment such as
people moving about [23].
These variations suggest that the radio map should reflect this range of values to increase the accuracy. Moreover, during the online phase, the system should use
more than one sample in the estimation process to have a
better estimate of the signal strength at a location.
2.2.2

Samples Correlation

Figure 2 shows the autocorrelation function of the samples collected from one access point (one sample per
second) at a fixed position. The figure shows that the
autocorrelation of consecutive samples (lag = 1) is as
high as 0.9. This high autocorrelation is expected as over
a short period of time the signal strength received from
an access point at a particular location is relatively stable

0.4

0 1 2 3 4 5 6 7 8 9 10
Time (Seconds)

0.2
0
-0.2

1000

2000

3000

4000

5000

6000

7000

8000

Time (Seconds)

Figure 2: An example of the autocorrelation between

samples from an access point (one sample per second).
The sub-figure shows the autocorrelation for the first 10
seconds.

200
150
100
50
0
-95

-0.4
0

250

Receiver sensitivity

0.6

Number of Samples Collected

Autocorr. coeff.

Autocorrelation coefficient

0.8

300

0.92
0.9
0.88
0.86
0.84
0.82
0.8
0.78

-90

-85 -80 -75 -70 -65 -60

Average Signal Strength (dBm)

-55

Figure 3: Relation between the average signal strength of

an access point and the percentage of samples received
from it during a 5-minute interval.

scale variations and small-scale variations.

(modulo the changes in the environment discussed in the
previous section).
This high autocorrelation value has to be considered
when using multiple samples from one access point to
enhance accuracy. Assuming independence of samples
from the same access point leads to the undesirable result of degraded system performance as the number of
samples is increased (as explained in Section 4) as in a
typical WLAN environment samples from the same AP
are highly correlated.
2.2.3

Samples from different access points

2.3.1

Large-Scale Variations

Figure 4 shows the average signal strength received from

an access point as the distance from it increases. The signal strength varies over a long distance due to attenuation
of the RF signal.
Large-scale variations are desirable in RF-based systems as they lead to changing the signature stored in the
radio map for different locations and, hence, better differentiation between these locations.
2.3.2

Small-Scale Variations

We performed an experiment to test the behavior of access points with different average signal strength at the
same location. During this experiment, we sampled the
signal strength from each access point at the rate of one
sample per second. Figure 3 shows the relation between
the average signal strength received from an access point
and the percentage of samples we receive from it during
a period of 5 minutes. The figure shows that the number
of samples collected from an access point is a monotonically increasing function of the average signal strength
of this access point. Assuming a constant noise level, the
higher the signal strength, the higher the signal to noise
ratio and the more probable it becomes that the 802.11b
card will identify the existence of a frame. The sharp
drop at about -81 dBm can be explained by noting that
the receiver sensitivity (minimum signal power required
to detect a frame) for the card we used was -82 dBm.

These variations happen when the user moves over a

small distance (order of wavelength). This leads to
changes in the average received signal strength. For the
802.11b networks working at the 2.4 GHz range, the
wavelength is 12.5 cm and we measure a variation in the
average signal strength up to 10 dBm in a distance as
small as 7.6 cm (3 inches) (Figure 5).
Dealing with small-scale variations is challenging. To
limit the radio map size and the time required to build
the radio map, selected radio map locations are typically
placed more than a meter apart. This means that the radio map does not capture small-scale variations leading
to decreased accuracy in the current WLAN location systems.
In the next section, we indicate how the Horus system
handles these temporal and spacial variations.

2.3

Spatial characteristics

These variations occur when the receiver position is

changed. We further divide these variations into large-

The Horus System

In this section, we present the different components of

the Horus system.

Figure 4: Large-scale variations: Average signal strength

over distance.

Overview

70

72

6
4

6 6
6 8

68
70

66

10

20

68

66

6
6

2
6 6

66

64

0
0

4
62
6
8

16

75

70

30

X (centimeters)

40

50

Figure 5: Small-scale variations: Signal strength contours from an AP in a 30.4 cm (12 inches) by 53.3 cm
(21 inches) area.

Figure 6 shows the overall system. The Horus system

works in two phases:
1. Offline phase: to build the radio map, cluster radio map locations, and do other preprocessing of the
signal strength models.
2. Online Phase: to estimate the user location based on
the received signal strength from each access point
and the radio map prepared in the offline phase.
The radio map stores the distribution of signal strength
received from each access point at each sampled location. There are two modes for operation of the Horus
system: one uses non-parametric distributions and the
other uses parametric distributions.
The Clustering module is used to group radio map locations based on the access points covering them. Clustering is used to reduce the computational requirements
of the system and, hence, conserve power (Section 3.7).
The Discrete Space Estimator module returns the radio map location that has the maximum probability given
the received signal strength vector from different access
points (Section 3.3).
The Correlation Modelling and Handling modules use
an autoregressive model to capture the correlation between consecutive samples from the same access point.
This model is used to obtain a better discrete location estimate using the average of n correlated samples (Section
3.4).
The Continuous Space Estimator takes as an input the
discrete estimated user location, one of the radio map locations, and returns a more accurate estimate of the user
location in the continuous space (Section 3.5).

70

14

8
10
12
Distance (Meter)

2
7

0
7

10
66

-60

60

-58

66
62

-56

15

65

-54

6
8

4
6

-52

62

20

664
2

-50

Y (centimeters)

-48

6
6

6 6
8 6

4
6

25

-46

60

74

6
4

-44

Average Signal Strength (dBm)

8
6
66

-62

3.1

30

-42

The Small-Scale Compensator module handles the

small-scale variation characteristics of the wireless channel (Section 3.6).
We start by laying out the mathematical framework
for the approach then give details about different components of the system.

3.2 Mathematical Model

Without loss of generality, let X be a 2 dimensional physical space. At each location x X, we can get the
signal strength from k access points. We denote the kdimensional signal strength space as S. Each element in
this space is a k-dimensional vector whose entries represent the signal strength readings from different access
points. We denote samples from the signal strength space
S as s. We also assume that the samples from different
access points are independent.
The problem becomes, given a signal strength vector
s = (s1 , ..., sk ), we want to find the location x X that
maximizes the probability P (x/s).
In the next section, we assume a discrete X space. We
discuss the continuous space case in Section 3.5.

3.3 Discrete Space Estimator

During the offline phase, the Horus system estimates the
signal strength histogram for each access point at each
location. These histograms represent the Horus systems
radio map. Now consider the online phase. Given a signal strength vector s = (s1 , ..., sk ), we want to find the
location x X that maximizes the probability P (x/s),

Applications
Estimated Location

Horus System Components

Location API

Continuous-Space
Estimator

one access point is as high as 0.9. Assuming independence of samples from the same access point leads to the
undesirable result of degraded system performance as the
number of averaged samples is increased (as we demonstrate below, in Section 4).
3.4.1

Radio
Map
and
clusters

Radio Map
Builder

We use an autoregressive model to capture the correlation between different samples from the same AP.
Let st be the stationary time series representing the
samples from an access point, where t is the discrete time
index. st can be represented as a first order autoregressive model [7] as:

Small-Scale
Compensator

Discrete-Space
Estimator

Correlation
Modeler

st = st1 + (1 )vt

Correlation
Handler

Clustering

Signal Strength Acquisition API

(MAC, Signal Strength)
Device Driver

Figure 6: Horus Components: the arrows show information flow in the system. Shadowed blocks represent
modules used during the offline phase.

i.e., we want
argmaxx [P (x/s)]

(1)

(2)

P (s/x) can be calculated using the radio map as:

P (s/x) =

k
Y

P (si /x)

(3)

i=1

The signal-strength histogram can be approximated by

a parametric distribution such as the Gaussian distribution. We compare the performance of the discrete-space
estimator based on the parametric and non-parametric
distributions in the Section 4.

3.4

Correlation Handling

To account for the temporal signal-strength variations, it

is important to average multiple signal strength samples
from the same access point. As we showed in Figure 2,
the autocorrelation of successive samples collected from

;0 1

(4)

where vt is a noise process, independent of st , and

is a parameter that determines the degree of autocorrelation of the original samples. Moreover, different samples from vt are independent and identically distributed
(i.i.d.).
The model in Equation 4 states that the current signal strength value (st ) is a linear aggregate of the previous signal strength value (st1 ) and an independent noise
value (vt ). The parameter gives flexibility to the model
as it can be used to determine the degree of autocorrelation of the original process. For example, if is zero, the
samples of the process st are i.i.d.s, whereas if is 1 the
original samples are identical (autocorrelation=1).
Assuming that the signal strength distribution of samples from an access point is Gaussian with mean and
variance 2 , we have shown in [29, 31] that the distribution of the average of n correlated samples is a Gaussian
distribution with mean and variance given by:
1+ 2

Using Bayes theorem, this can be shown to be equivalent to [33]:

argmaxx [P (x/s)] = argmaxx [P (s/x)]

Mathematical model

3.4.2

(5)

Correlation modeler

The purpose of the correlation modeler component is to

estimate the value of in the autoregressive model and
to pre-calculate the parameters of the distribution of the
average of n correlated samples during the offline phase.
In a previous work [29,31], we have shown that can be
approximated using the autocorrelation coefficient with
lag 1. The variance of the distribution can be calculated
using Equation 5. These distribution parameters (, ,
and ) are then stored in the radio map.
3.4.3

Correlation handler

During the online phase, the correlation handler module averages the value of n consecutive samples from an
access point and passes this information to the discretespace estimator, which uses the distributions stored in the
radio map (taking correlation into account using the information in Section 3.4.1) to estimate the user location.

3.5

Continuous Space Estimator

The discrete-space estimator returns a single location

from the set of locations in the radio map. To increase the
system accuracy, the Horus system uses two techniques
to obtain a location estimate in the continuous space.
3.5.1

Technique 1: Center of Mass of the Top Candidate Locations

This technique is based on treating each location in the

radio map as an object in the physical space whose
weight is equal to the normalized probability1 assigned
by the discrete-space estimator. We then obtain the center of mass of the N objects with the largest mass, where
N is a parameter to the system, 1 N ||X||.
More formally, let p(x) be the probability of a location
be the list of locax X, i.e., the radio map, and let X
tions in the radio map ordered in a descending order according to the normalized probability (the location with
lower ID comes first for locations with equal probability). The center of mass technique estimates the current
location x as:
N
P

x=

p(i)X(i)

i=1
N
P

(6)

p(i)

i=1

is the ith element of X

where X(i)
Note that the estimated location x need not be one of the
radio map locations.
3.5.2

Technique 2: Time-Averaging in the Physical

Space

The second technique uses a time-average window to

smooth the resulting location estimate. The technique
obtains the location estimate by averaging the last W
locations estimates obtained by either the discrete-space
estimator or the continuous-space estimator discussed in
the previous section.
More formally, given a stream of location estimates
x1 , x2 , ..., xt , the technique estimates the current location xt at time t as:
xt =

1
min(W, t)

t
X

xi

(7)

tmin(W,t)+1

We compare the two techniques in Section 4.

3.6

Small-Scale Compensator

Dealing with small-scale variations (Figure 5) is challenging. Since the selected radio map locations are typically placed more than a meter apart, to limit the radio map size, the radio map does not capture small-scale
variations. This contributes significantly to the estimation errors in the current systems. The Horus system

uses the Perturbation technique to handle the small-scale

variations. The technique is based on two sub-functions:
detecting small-scale variations and compensating for
small-scale variations.
3.6.1

Detecting small-scale variations

In order to detect small-scale variations, the Horus system uses the heuristic that users location cannot change
faster than their movement rate. The system calculates
the estimated location using the standard radio map and
the inference algorithm, then calculates the distance between the estimated location and the previous user location. If this distance is above a threshold, based on the
user movement rate and estimation frequency, the system
decides that there are small-scale variations affecting the
signal strength.
3.6.2

Compensating for small-scale variations

To compensate for these small-scale variations, the system perturbs the received vector entries, re-estimates the
location, and chooses the nearest location to the previous
user location as the final location estimate. For example, if one sample includes a signal-strength observation
from each of k access points (s1 , s2 , . . . , sk ), the system
tries all 3k combinations in which each of the k observations i is replaced by one of three values, si , si (1 + d), or
si (1 d); we explore the parameter d in Section 4.2.4.
An enhancement of this approach is to perturb a subset
of the access points. The effect of the number of access
points to perturb and the value of d on accuracy is described in Section 4.

3.7 Clustering Module

This section describes the Incremental Triangulation
(IT) clustering technique used by the Horus system to reduce the computational requirements of the location determination algorithm. We define a cluster as a set of
locations sharing a common set of access points. We call
this common set of access points the cluster key. The
problem can be stated as: Given a location x, we want
to determine the cluster to which x belongs. The noisy
characteristics of the wireless channel described in Section 2 make clustering a challenging problem because the
number of access points covering a location varies with
time.
The IT approach is based on the idea that each access
point defines a subset of the radio map locations that
are covered by this access point. These locations can
be viewed as a cluster of locations whose key is the access point covering the locations in this cluster. If during
the location determination phase we use the access points
incrementally, one after the other, then starting with the
first access point, we restrict our search space to the locations covered by this access point. The second access
point chooses only the locations in the range of the first

Figure 7: Floor plan for the first testbed. Readings were

collected in the corridors and inside the rooms.

Figure 8: Floor plan of the office space where the second

experiment was conducted. Readings were collected in
the corridors and inside the rooms.

access point and covered by the second access point and

so on, leading to a multi-level clustering process.

4.1.1

Notice that no preprocessing is required in the offline

training phase. During the online phase, a location x belongs to a cluster whose key is access point a if there
is information about access point a at location x in the
radio map.
The algorithm works as follows. Given a sequence of
observations from each access point, we start by sorting the access points in descending order according to
the average received signal strength. For the first access
point, the one with the strongest average signal strength,
we calculate the probability of each location in the radio
map set given the observation sequence from this access
point alone. This gives us a set of candidate locations
(locations that have non-zero probability). If the probability of the most probable location is significantly
higher (according to a threshold) than the probability of
the second most probable location, we return that most
probable location as our location estimate, after consulting only one access point. If this is not the case, we go to
the next access point in the sorted access point list. For
this access point, we repeat the same process again, but
only for the set of candidate locations obtained from the
first access point. We study the performance of the IT
approach in Section 4.

Experimental Evaluation

In this section we start by showing the effect of each

module independently on the the accuracy of the basic
algorithm. We then show the effect of using all the components together on the performance of the Horus system.

4.1

Experimental Testbed

We performed our experiment in two different testbeds.

Testbed 1

We performed our first experiment in the south wing of

the fourth floor of the A. V. Williams building in the University of Maryland at College Park. The layout of the
floor is shown in Figure 7. The wing has a dimension of
68.2 meters by 25.9 meters. The technique was tested
in the University of Maryland wireless network using
Cisco access points. 21 access points cover the multistory wing and were involved in testing.
The radio map has 110 locations along the corridors
and 62 locations inside the rooms. On the average, each
location is covered by 6 access points. The Horus system
was running under the Windows XP professional operating system.
4.1.2

Testbed 2

We performed the second experiment in another office

space (Figure 8). The area of the experiment site is approximately 11.8 meters by 35.9 meters covering corridors, cubicles, and rooms. Five LinkSys access points
and one Cisco access point cover the test area.
We have a total of 110 locations in the radio map. On
the average, each location is covered by 4 access points.
The Horus system was running under the Linux (kernel
2.5.7) operating system.
4.1.3

Data collection

The radio map locations were marked on the floor before

the experiment and the user clicked on the map to point
the location of the radio map points. We collected 100
samples, spaced 300 ms apart, at each radio map location. We expect an error of about 15-20 cm due to the
inaccuracies in clicking the map.
The training data was placed 1.52 meters (5 feet) apart
for the first testbed and 2.13 meters apart for the second
testbed (7 feet).
For each testbed, we selected 100 test locations to random cover the entire test area (none of them coincide
with a training point). For both testbeds, the test set was

Table 1: Summary of the percentage enhancement of different components on the basic algorithm
Technique
Testbed 1 Testbed 2
Correlation Handling
19%
11%
Center of Mass
13%
6%
Time Averaging
24%
15%
Small-Scale Compensator 25%
21%

0.9
0.8

Probability

0.7
0.6
0.5
0.4
0.3

0.8

0.1

Non-Parametric
Parametric

0
0

Distance Error (Meter)

Figure 9: Performance of the basic algorithm of the Horus system for the first testbed.

collected by different persons on different days and times

of day than the training set. This difference presents
a realistic testbed and should, if anything, decrease the
measured accuracy of our approach because it lessens
the likelihood that the test data is a close match to the
training data.

4.2

Discrete-space estimator (Basic algorithm)

Figure 9 shows the performance of the basic algorithms

of the Horus system for the first testbed. The system
can achieve an accuracy of 1.4 meters 90% of the time.
The performance of the parametric and non-parametric
methods is comparable with a slight advantage for the
parametric method. Using a parametric distribution to
estimate the signal-strength distribution smooths the distribution shape to account for missing signal strength values in the training phase (due to the finite training time).
This smoothing avoids obtaining a zero probability for
any signal strength value that was not obtained in the
training phase and hence enhances the accuracy.
Table 1 shows the summary of the results for the two
testbeds. Details for the second testbed can be found in
[28].
4.2.2

0.7
0.6
0.5
0.4
0.3
0.2
0.1

Without Corr.
With Corr.

0
1

2
3
4
Number of samples (n)

Figure 10: Average distance error with and without taking correlation into account for the first testbed.

Results

We show the effect of each module independently on the

performance of the discrete-space estimator and present
the overall system performance in Section 4.3.
4.2.1

Average distance error (Meter)

0.2

Correlation handler

Figure 10 shows the performance of the Horus system

when taking the correlation into account and without taking the correlation into account for the first testbed. We
estimated the value of to be 0.9. The figures show that
under the independence assumption, as the number of

averaged samples increases, the performance degrades.

The minimum value at n = 2 can be explained by noting
that there are two opposing factors affecting the system
accuracy:
1. as the number of averaged samples n increases, the
accuracy of the system should increase.
2. as n increases, the estimation of the distribution of
the average of the n samples becomes worse due to
the wrong independence assumption.
At low values of n (n = 1, 2) the first factor is the dominating factor and hence the accuracy increases. Starting from n = 3, the effect of the bad estimation of the
distribution becomes the dominating factor and accuracy
degrades.
Using the modified technique, the system average accuracy is enhanced by more than 19% using five signalstrength samples.
4.2.3

Continuous space estimator

Center of Mass Technique: Figure 11 shows the effect of

increasing the parameter N (number of locations to interpolate between) on the performance of the center of
mass technique for the first testbed. Note that the special

0.7
Average Distance Error (Meter)

Average distance error (Meter)

0.7
0.6
0.5
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

0.4
0.3
0.2
0.1

20

40

Average distance error (Meter)

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
2

4
6
8
Averaging window size (W)

10

Figure 12: Average distance error using the timeaveraging technique for the first testbed.

0.3
0.2
0.1

case of N = 1 is equivalent to the discrete-space estimator output. The figures show that the performance of the
Horus system is enhanced by more than 13% for N = 6.
Time-averaging Technique: Figure 12 shows the effect of increasing the parameter W (size of the averaging window) on the performance of the time-averaging
technique. The figures show that the larger the averaging window, the better the accuracy. The performance
of the Horus system is enhanced by more than 24% for
W = 10.
Small-scale compensator

For the purpose of detecting small-scale variations, we

assume a maximum user speed of two meters per second.
Figure 13 shows the effect of changing the perturbation fraction (d, which is the amount by which to perturb

0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Perturbation Fraction

Figure 13: Effect of changing the perturbation fraction

on average distance error.

each access point) on average error. We can see from this

figure that the best value for the perturbation fraction is
0.05 for the first testbed. We use these values for the rest
of this section.
Figure 14 shows the effect of increasing the number
of perturbed access points on the average distance error.
The access points chosen at a location are the strongest
access points in the set of access points that cover that
location. The figure shows that perturbation technique is
not sensitive to the number of access points. This means
that perturbing one access point only is sufficient to enhance the performance.
Figures 15 shows the effect of using the perturbation
technique on the basic Horus system. The perturbation
technique reduces the average distance error by more
than 25% and the worst-case error is reduced by more
than 30%.
4.2.5

4.2.4

0.4

60 80 100 120 140 160 180

Number of locations (N)

Figure 11: Average distance error using the center of

mass technique for the first testbed.

0.5

10

0
0

0.6

Clustering module

Figures 16 and 17 shows the effect of the parameter

T hreshold on the performance. For small values of the
T hreshold parameter, the decision is taken quickly after examining a small number of access points. As the
threshold value increases, more access points are consulted to reach a decision. As the number of access points
consulted increases, the number of operations (multiplications) per location estimate increases and so does the
accuracy.

4.3 Overall System Performance

In the previous sections, we studied the effect of each
component of the Horus system separately on the performance. In this section, we compare the performance of
the full Horus system, to the performance of a deterministic technique (the Radar system [5]) and a probabilistic

0.7
Average distance error (Meter)

Average Distance Error (Meter)

0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
1

2
Number of Access Points Used

0.5
0.4
0.3
0.2
0.1
0
No Clust. 0.1

Figure 14: Effect of changing the number of perturbed

access points on average distance error.

0.6

0.3
Threshold

0.5

0.7

Figure 16: Effect of the parameter T hreshold on the

average distance error for the first testbed.

Table 2: Estimation parameters for the two testbeds

Parameter
Test. 1 Test. 2
Correlation Degree ()
0.9
0.7
Num. of avg. samples (n)
10
10
Num. of loc. used in interp. (N ) 6
6
Averaging window (W )
10
10
Threshold
0.1
0.1

0.9
0.8
Probability

0.7
0.6
0.5
0.4
0.3
0.2
Perturbation
Basic

0.1
0
0

2
4
Error Distance (Meter)

Figure 15: CDF for the distance error for the first testbed.

technique [21]. We use the parametric distribution technique. Table 2 shows the values of different parameters.
Figures 18 shows the comparison for the two testbeds
(the curve for the Radar system is truncated). Tables 3
summarizes the results. The table shows that the average
accuracy of the Horus system is better than the Radar
system by more than 89% for the first testbed and 82%
for the second testbed. The worst case error is decreased
by more than 93% for the first testbed and 70% for the
second testbed.
Comparing the probabilistic system in [21] to the Horus system shows that the average error is decreased by
more than 35% for the first testbed and 27% for the second testbed. The worst case error is decreased by more
than 78% for the first testbed and 70% for the second
testbed. These results show the effectiveness of the pro-

posed techniques.
The performance of the three systems is better in the
first testbed than the second testbed as the first testbed
has a higher density of APs per location and the calibration points were closer for the first testbed.
Moreover, the Horus system leads to more than an
order of magnitude savings in the number of multiplications required per location estimate compared to the
other systems (250 multiplications for Horus compared
to 2708 for the other two systems).
We also applied the enhancement discussed in this paper (without correlation handling) to the Radar system.
We summarize the results is Table 3. These results show
the effectiveness of the techniques proposed in the paper
and that these techniques are general and can be applied
to other WLAN location determination systems to enhance their accuracy.

Discussion

In this section, we highlight some of our experience with

the Horus system.

Avg. num. of oper. per location estimate

Table 3: Comparison of the Horus system and other systems (error in centimeters)
Testbed 1
Testbed 2
Median Avg Stdev 90% Max Median Avg Stdev 90%
Horus
39
42
28
86
121
51
64
53
132
System [21]
48
65
63
143
578
72
86
77
181
Radar
296 400
326
853 1757
341 361
184
611
Radar with Horus tech.
161 193
107
302
423
142 195
106
332

is dependent on the application in use.

Latency can be reduced by presenting the location estimate incrementally using one sample at a time. The
system need not to wait until it acquires the n samples
all at once. Instead, it can give a more accurate estimate
of the location as more samples become available by reporting the estimated location given the partial samples
it has. Other factors that affect the choice of the value of
n are the user mobility rate and the sampling rate. The
higher the user mobility rate or the sampling rate, the
lower the value of n.

4000
3500
3000
2500
2000
1500

80
70
60
50
40
30
20
10
0
0.1

0.3

0.5

0.7

1000
500
0
No Clust. 0.1

0.3
Threshold

0.5

0.7

Figure 17: Effect of the parameter T hreshold on the

average number of operations per location estimate for
the first testbed. The sub-figure shows the same curve
for T hreshold = [0.1, 0.7].

5.1

Parametric vs Non-Parametric Distributions

The Horus system can model the signal strength distributions received from the access points using parametric or
non-parametric distributions. The main advantage of the
non-parametric technique is the efficiency of calculating
the location estimate, while the parametric technique reduces the radio map size and smooths the distribution
shape which leads to a slight computational advantage of
the parametric technique over the non-parametric technique.

5.2

Max
289
991
967
483

Location Estimation Latency

The correlation handling and the continuous space estimator modules each use more than one sample to increase the accuracy of the system. However, a side effect
of this increased accuracy is that the latency of calculating the location estimate increases. In general, we have a
tradeoff between the accuracy required and the latency of
the location estimate. The higher the required accuracy,
the higher the number of samples required and the higher
the latency to obtain the location estimate. This decision

5.3 User Profile

A common assumption of WLAN location determination
systems is that the user position follows a uniform distribution over the set of possible locations. Our analysis
and experimentation [32] show that knowing the probability distribution of the user position can reduce the
number of access points required to obtain a given accuracy. However, with a high density of access points, the
performance of the Horus system is consistent under different probability distributions for the user position, i.e.,
the effect of the user profile is not significant with a high
density of access points. Systems can use this fact to reduce the energy consumed in the location determination
algorithm by not using the user profile in the estimation
process.

5.4 Effect of Different Hardware

One of the hardware related questions is whether different hardware from different manufacturers are compatible. That is, how does using different APs, mobile devices, or wireless cards affect the accuracy?
Our experience with the Horus system shows that the
Laptop or PDA used for the calibration has no effect on
the accuracy if a different device is used in tracking. APs
from different manufacturers can be used without affecting the accuracy since the radio map captures the signature of the AP at each location (note that the second
testbed uses mixed types of APs). The 802.11h specifications, however, require APs to have transmission power
control (TPC) and dynamic frequency selection (DFS).
This presents an open research direction for the current
WLAN location determination systems as they assume
that the AP transmission power does not change over

drivers that support that interface. Under the Windows

OS, NDIS allowed us to perform the same functions.
Our experience with both systems shows that drivers
under Linux conform to the Wireless Extensions APIs
better than Windows Drivers do with the NDIS. For example, under the Windows, some cards, like the Cisco
card, respond to scans with low frequency (every 2-3
seconds) and return only one AP. We hope that future
versions of the driver will have better support for the
NDIS interface. Moreover for both systems, better active scanning techniques needs to be developed to reduce
the scanning overhead.

Probability

0.8
0.6
0.4
0.2

Horus
System [21]
Radar

0
0

2
3
Error Distance (Meter)

(a) First testbed

Probability

0.8
0.6
0.4
0.2

Horus
System [21]
Radar

0
0

2
3
Error Distance (Meter)

(b) Second testbed

Figure 18: CDF of the performance of the Horus system

and the Radar and the probabilistic system.

time.
The main factor that may affect the accuracy when
changing hardware is the wireless card. Our experience
shows that cards from the same manufacture are interchangeable. The good news is that a linear mapping exists between different NICs [13]. Unfortunately, some of
the cards in the market are so noisy [27] that with this
linear mapping the obtained radio map is not representative of the environment. We found that Orinoco cards
and Cisco cards are stable, in terms of signal-strength
measurements.

5.5

Operating System Interface

We implemented the Horus system under both Linux

and Windows. The main functionality we require from
the OS is support for issuing scan requests and returning the results. Under the Linux OS, the wireless extensions [1,2] give us a common interface to query different

Related Work

Many systems over the years have tackled the problem

of determining and tracking the user position. Examples
include GPS [11], wide-area cellular-based systems [24],
infrared-based systems [4, 26], ultrasonic-based systems [19], various computer vision systems [16], and
physical contact systems [18].
Compared with these systems, WLAN location determination systems are software based (do not require
specialized hardware) and may provide more ubiquitous
coverage. This feature adds to the value of the wireless
data network.
The Daedalus project [14] developed a system for
coarse-grained user location. A mobile host estimates
its location to be the same as the base station to which
it is attached. Therefore, the accuracy of the system is
limited by the access point density.
The RADAR system [5] uses the RF signal strength
as an indication of the distance between the transmitter
and receiver. During an offline phase, the system builds
a radio map for the RF signal strength from a fixed number of receivers. During normal operation, the RF signal
strength of the mobile client is measured by a set of fixed
receivers and is sent to a central controller. The central
controller uses a K-nearest approach to determine the location from the radio map that best fits the collected signal strength information.
The Aura system proposed in [22] uses two techniques: pattern matching (PM) and triangulation, mapping and interpolation (TMI). The PM approach is very
similar to the RADAR approach. In the TMI technique,
the physical position of all the access points in the area
needs to be known and a function is required to map signal strength onto distances. They generate a set of training points at each trained position. The interpolation of
the training data allows the algorithm to use less training
data than the PM approach. During the online phase, they
use the approximate function they got from the training
data to generate contours and they calculate the intersection between different contours yielding the signal space
position of the user. The nearest set of mappings from

the signal-space to the physical space is found by applying a weighted average, based on proximity, to the signal
space position.
The Nibble location system, from UCLA, uses a
Bayesian network to infer a user location [8]. Their
Bayesian network model include nodes for location,
noise, and access points (sensors). The signal to noise
ratio observed from an access point at a given location
is taken as an indication of that location. The system
also quantizes the SNR into four levels: high, medium,
low, and none. The system stores the joint distribution
between all the random variables of the system.
Another system, [21], uses Bayesian inversion to return the location that maximizes the probability of the received signal strength vector. The system stores the signal strength histograms in the radio map and uses them
in the online phase to estimate the user location. Yet,
another system, [17], applies the same technique to the
robotics domain.
The Horus system is unique in defining the possible
causes of variations in the received signal strength vector and devising techniques to overcome them, namely
providing the correlation modeler, correlation handler,
continuous space estimator, and small-space compensator modules. Moreover, it reduces the computational
requirements of the location determination algorithm by
applying location-clustering techniques. This allows the
Horus system to achieve its goals of high accuracy and
low energy consumption.

Conclusions

In this paper, we presented the design of the Horus

system: a WLAN-based location determination system.
We approached the problem by identifying the various
causes of variations in a wireless channel and developed
techniques to overcome them. We also showed the various components of the system and how they interact.
The Horus system models the signal strength distributions received from access points using parametric and
non-parametric distributions. By exploiting the distributions, the Horus system reduces the effect of temporal
variations.
We showed that the correlation of the samples from
the same access point can be as high as 0.9. Experiments showed that under the independence assumption,
as the number of averaged samples increases, the performance degrades. Therefore, we introduced the correlation modeler and handling modules that use an autoregressive model for handling the correlation between
samples from the same access point. Using the modified technique, the system average accuracy is enhanced
by more than 19% for the first testbed and 11% for the
second testbed.
The Horus system uses the Perturbation technique for

handling small-scale variations. The perturbation technique enhances the average distance error by more than
25% for the first testbed and more than 21% for the second testbed. Moreover, the worst-case error is reduced
by more than 30% for the two testbeds.
The basic Horus technique chooses the estimated location from the discrete set of radio map locations. We
described two techniques for allowing continuous-space
estimation: the Center of Mass technique and the TimeAveraging technique. Using the Center of Mass technique, the accuracy of the Horus system was increased
by more than 13% for the first testbed and by more than
6% for the second testbed compared to the basic technique. The Time-Averaging technique enhances the performance of the Horus system by more than 24% for the
first testbed and more than 15% for the second testbed.
The two techniques are independent and can be applied
together.
We also compared the performance of the Horus system to the performance of the Radar system. We showed
that the average accuracy of the Horus system is better
than the Radar system by more than 89% for the first
testbed and 82% for the second testbed. The worst case
error is decreased by more than 93% for the first testbed
and 70% for the second testbed. Comparing the probabilistic system in [21] to the Horus system shows that
the average error is decreased by more than 35% for the
first testbed and 27% for the second testbed. The worst
case error is decreased by more than 78% for the first
testbed and 70% for the second testbed. These results
show the effectiveness of the proposed techniques. In
terms of computational requirements, the Horus system
is more efficient by more than an order of magnitude.
The proposed modules are all applicable to any of
the current WLAN location determination systems. We
show the result of applying the techniques of the Horus
system to the Radar system. The results show that the
average distance error is reduced by more than 58% for
the first testbed and by more than 54% for the second
testbed. The worst case error is decreased by more than
76% for the first testbed and by more than 48% for the
second testbed.
As part of our ongoing work we are experimenting
with different clustering techniques. Automating the
radio-map generation process is a possible research area.
The Horus system provides an API for location-aware
applications and services. We are looking at designing
and developing applications and services over the Horus
system. A possible future extension is to dynamically
change the system parameters based on the environment,
such as changing the averaging window size as the user
speed changes or using a time-dependent radio map. We
are also working on the theoretical analysis of different
components of the system.

Our experience with the Horus system showed that it

has achieved its goals of:
High accuracy: through a probabilistic location determination technique, using a continuous-space estimator, handling the high correlation between samples from the same access point, and the perturbation technique to handle small-scale variations.
Low computational requirements: through the use
of clustering techniques.
The design of Horus also allows it to achieve scalability to large coverage areas, through the use of clustering
techniques, and to large number of users, through the distributed implementation on the mobile devices and due to
the low energy requirements of the algorithms.
Moreover, the techniques presented in this paper may
be applicable to other RF-technologies such as 802.11a,
802.11g, HiperLAN, and BlueTooth.

Acknowledgments
This work was supported in part by the Maryland Information and Network Dynamics (MIND) Laboratory,
its founding partner Fujitsu Laboratories of America,
and by the Department of Defense through a University
of Maryland Institute for Advanced Computer Studies
(UMIACS) contract.

Availability
The MAPI API and the Linux device drivers are available
for download at [1].

References
[1]
[2]
[3]
[4]
[5]
[6]

[7]
[8]
[9]
[10]
[11]
[12]

http://www.cs.umd.edu/users/moustafa/Downloads.htm.
http://www.hpl.hp.com/personal/Jean Tourrilhes/.
http://www.orinocowireless.com.
A ZUMA , R. Tracking requirements for augmented reality. Communications of the ACM 36, 7 (July 1997).
BAHL , P., AND PADMANABHAN , V. N. RADAR: An InBuilding RF-based User Location and Tracking System. In IEEE
Infocom 2000 (March 2000), vol. 2, pp. 775784.
BAHL , P., PADMANABHAN , V. N., AND BALACHANDRAN , A.
Enhancements to the RADAR User Location and Tracking System. Tech. Rep. MSR-TR-00-12, Microsoft Research, February
2000.
B OX , G. E. P., J ENKINS , G. M., AND R EINSEL , G. C. Time
Series Analysis: Forcasting and Control, third ed. Prentice Hall,
1994.
C ASTRO , P., C HIU , P., K REMENEK , T., AND M UNTZ , R. A
Probabilistic Location Service for Wireless Network Environments. Ubiquitous Computing 2001 (September 2001).
C ASTRO , P., AND M UNTZ , R. Managing Context for Smart
Spaces. IEEE Personal Communications (OCTOBER 2000).
C HEN , G., AND KOTZ , D. A Survey of Context-Aware Mobile
Computing Research. Tech. Rep. Dartmouth Computer Science
Technical Report TR2000-381, 2000.
E NGE , P., AND M ISRA , P. Special issue on GPS: The Global
Positioning System. Proceedings of the IEEE (January 1999),
3172.
G WON , Y., JAIN , R., AND K AWAHARA , T. Robust Indoor Location Estimation of Stationary and Mobile Users. In IEEE Infocom
(March 2004).

[13] H AEBERLEN , A., F LANNERY, E., L ADD , A., RUDYS , A.,

WALLACH , D., AND K AVRAKI , L. Practical Robust Localization over Large-Scale 802.11 Wireless Networks. In 10th ACM
MOBICOM (Philadelphia, PA, September 2004).
[14] H ODES , T. D., K ATZ , R. H., S CHREIBER , E. S., AND ROWE ,
L. Composable ad hoc mobile services for universal interaction.
In 3rd ACM MOBICOM (September 1997), pp. 112.
[15] K RISHNAN , P., K RISHNAKUMAR , A., J U , W. H., M ALLOWS ,
C., AND G ANU , S. A System for LEASE: Location Estimation Assisted by Stationary Emitters for Indoor RF Wireless Networks. In IEEE Infocom (March 2004).
[16] K RUMM , J., ET AL . Multi-camera multi-person tracking for Easy
Living. In 3rd IEEE Intl Workshop on Visual Surveillance (Piscataway, NJ, 2000), pp. 310.
[17] L ADD , A. M., B EKRIS , K., RUDYS , A., M ARCEAU , G.,
K AVRAKI , L. E., AND WALLACH , D. S. Robotics-Based Location Sensing using Wireless Ethernet. In 8th ACM MOBICOM
(Atlanta, GA, September 2002).
[18] O RR , R. J., AND A BOWD , G. D. The Smart Floor: A Mechanism for Natural User Identification and Tracking. In Conference on Human Factors in Computing Systems (CHI 2000) (The
Hague, Netherlands, April 2000), pp. 16.
[19] P RIYANTHA , N. B., C HAKRABORTY, A., AND BALAKRISH NAN , H. The Cricket Location-Support system. In 6th ACM
MOBICOM (Boston, MA, August 2000).
[20] ROOS , T., M YLLYMAKI , P., AND T IRRI , H. A Statistical Modeling Approach to Location Estimation. IEEE Transactions on
Mobile Computing 1, 1 (January-March 2002), 5969.
[21] ROOS , T., M YLLYMAKI , P., T IRRI , H., M ISIKANGAS , P., AND
S IEVANEN , J. A Probabilistic Approach to WLAN User Location Estimation. International Journal of Wireless Information
Networks 9, 3 (July 2002).
[22] S MAILAGIC , A., S IEWIOREK , D. P., A NHALT, J., KOGAN , D.,
AND WANG , Y. Location Sensing and Privacy in a Context Aware
Computing Environment. Pervasive Computing (2001).
[23] S TALLINGS , W. Wireless Communications and Networks,
first ed. Prentice Hall, 2002.
[24] T EKINAY, S. Special issue on Wireless Geolocation Systems and
Services. IEEE Communications Magazine (April 1998).
[25] T HE I NSTITUTE OF E LECTRICAL AND E LECTRONICS E NGI NEERS , I NC . IEEE Standard 802.11 - Wireless LAN Medium
Access Control (MAC) and Physical Layer (PHY) specifications.
[26] WANT, R., H OPPER , A., FALCO , V., AND G IBBONS , J. The Active Badge Location System. ACM Transactions on Information
Systems 10, 1 (January 1992), 91102.
[27] Y EO , J., BANERJEE , S., AND AGRAWALA , A. Measuring traffic
on the wireless medium: experience and pitfalls. In Technical Report, CS-TR 4421, Department of Computer Science, University
of Maryland, College Park (Dec. 2002).
[28] YOUSSEF, M. Horus: A WLAN-Based Indoor Location Determination System. PhD thesis, University of Maryland at College
Park, May 2004. Submitted for SigMobile Dissertation Page.
[29] YOUSSEF, M., A BDALLAH , M., AND AGRAWALA , A. Multivariate Analysis for Probabilistic WLAN Location Determination Systems. In The Second Annual International Conference on
Mobile and Ubiquitous Systems: Networking and Services (July
2005).
[30] YOUSSEF, M., AND AGRAWALA , A. Small-Scale Compensation
for WLAN Location Determination Systems. In IEEE WCNC
2003 (March 2003).
[31] YOUSSEF, M., AND AGRAWALA , A. Handling Samples Correlation in the Horus System. In IEEE Infocom (March 2004).
[32] YOUSSEF, M., AND AGRAWALA , A. On the Optimality of
WLAN Location Determination Systems. In Communication
Networks and Distributed Systems Modeling and Simulation
Conference (January 2004).
[33] YOUSSEF, M., AGRAWALA , A., AND S HANKAR , A. U. WLAN
Location Determination via Clustering and Probability Distributions. In IEEE PerCom 2003 (March 2003).
[34] YOUSSEF, M., AGRAWALA , A., S HANKAR , A. U., AND N OH ,
S. H. A Probabilistic Clustering-Based Indoor Location Determination System. Tech. Rep. UMIACS-TR 2002-30 and CSTR 4350, University of Maryland, College Park, March 2002.
http://www.cs.umd.edu/Library/TRs/.

Notes
1 The normalization is used to ensure that the sum of the probabilities of all locations equals one.