USE OF AGPS CALL DATA RECORDS FOR NON-GPS TERMINAL

POSITIONING IN CELLULAR NETWORKS

Semun Lee, Jeongkeun Lee, Taekyoung Kwon and Yanghee Choi

School of Computer Science & Engineering, Seoul National University

San 56-1 Shilim-dong, Kwanak-gu, Seoul, South Korea

Tae Joon Ha and Tae-il Kim

Radiant Technologies, Inc.

Il-yang Bldg 206, Nonhyeon 1-dong. Gangnam-gu, Seoul, South Korea

Keywords:

Positioning, Pattern Matching, AGPS CDRs, cellular network.

Abstract:

This paper presents a novel method of user terminal positioning in cellular networks. A pattern matching

technology based on received-signal-strength (RSS) from the pilot channels of cell towers has been the most

popular network-based positioning method. In response to a position request, a terminal measures RSSs of

pilot channels from surrounding cell towers and then the terminal’s RSS pattern is compared with a pattern

database to ﬁnd the most correlated one which indicates the position of the terminal. Although the pattern

matching method can provide accurate positioning, its database construction and maintenance require a high

overhead of periodic labor-intensive pattern collection. In this paper, we propose to exploit the call data records

(CDRs) that are uploaded by Assisted Global Positioning System (AGPS) terminals as inputs to the pattern

database, which removes or reduces the pattern collection overhead. In AGPS systems, terminals measure

satellite signals and cellular network parameters (such as RSS) and relay them to the cellular infrastructure,

which in turn calculates the terminal position using both that satellite and cellular network data. The proposed

AGPS CDR based pattern matching method takes advantage of the increasing number of AGPS terminals in

service: non-AGPS terminals can obtain more precise positioning results in areas where more AGPS calls are

generated (e.g. hotspots). To do so, we analyze the characteristics of RSS patterns and AGPS CDRs. Based

on the analysis, a pattern-distance metric and an AGPS CDR based pattern matching system are proposed and

their performances are evaluated by examining ﬁeld data of several urban downtown areas of Seoul, Korea. We

obtain promising results: the position of the user terminal can be estimated with the accuracy (or, positioning

error) at the level of 96.5m and 149.8m for the 67% and the 95% conﬁdence interval, respectively.

1 INTRODUCTION

The geographic position of a user terminal is valu-

able and critical information to provide ubiqui-

tous/intelligent services such as location-based emer-

gency terminal service. Many researchers and engi-

neers have made great efforts to obtain accurate user

terminal position information with or without Global

Positioning System (GPS).

Although GPS is able to provide excellent position

accuracy, position ﬁxes require lines of sight (LOS)

to multiple satellites, long ﬁrst ﬁx time (at least 30s),

and high processing power. Assisted GPS, or AGPS

(Djuknic, 2001), is a technology that uses an assis-

tance server in a cellular infrastructure to cut down

the time needed to ﬁx the position. In AGPS sys-

tems, the terminal, being limited in processing power,

Korea patent pending.

communicates with the assistance server that has high

processing power. In response to position queries,

AGPS terminals measure satellite signals and cellu-

lar network parameters (such as RSS) and relay them

to the assistance server. The server uses those data

to calculate the position of the terminal and send the

calculated position back to the terminal. In urban ar-

eas, however, AGPS does not work (like GPS) under

heavy tree cover or indoors where the terminal cannot

receive a sufﬁcient number of satellite signals. More-

over, equipping a terminal with AGPS module raises

terminal production cost. That’s why some proposed

positioning technologies exploit the inherent radio pa-

rameters of the cellular network rather than relying on

GPS technologies.

The network parameters open up several possibili-

ties for positioning methodology. Simply, every cel-

lular system provides some information identifying

the serving cell (Cell-ID) enabling a coarse-grained

242

Lee S., Lee J., Kwon T., Choi Y., Joon Ha T. and Kim T. (2006).

USE OF AGPS CALL DATA RECORDS FOR NON-GPS TERMINAL POSITIONING IN CELLULAR NETWORKS.

In Proceedings of the International Conference on Wireless Information Networks and Systems, pages 242-249

 SciTePress

position estimate. For more accurate positioning, a

number of proposed solutions have utilized the prop-

agation delay, time-difference-of-arrival (TDOA), an-

tenna orientation or received-signal-strength (RSS).

The propagation time (or time-of-arrival) to the cell

tower is only available in TDMA systems and UMTS.

And in cases without LOS view between the transmit-

ter and the receiver, the distance from the propagation

time is overestimated. TDOA among several cell tow-

ers could be measured and the terminal position can

be obtained by solving the hyperbolic system. Hy-

perbolic systems require time synchronization among

cell towers, which is feasible not in GSM and UMTS

but in only CDMA networks. Additional hardware

can compensate for the time asynchronity but raises

costs. The hyperbolic systems also suffer from the in-

accurate TDOA measurements mainly caused by the

near-far problem: Idle Periods in Downlink (IPDL)

is proposed to mitigate the problem while it costs

downlink capacity and additional complexity (3GPP,

2002). The information of sector antenna orientation

and angle opening can be used to increase the posi-

tioning accuracy. Many service providers, however,

do not maintain the information of each sector an-

tenna. And radio waves may arrive from outside the

opening angle due to antenna side lobes, reﬂections,

and diffractions especially in dense urban areas.

Compared to the above-mentioned parameters,

RSS is the commonly available basic parameter for

all types of wireless cellular systems. Shadow-

ing and multipath fading, however, make RSS an

unreliable metric to estimate the exact transmitter-

receiver distance. Moreover, the RSS-to-distance

function (propagation model) is highly affected by en-

vironment/system speciﬁc factors such as height of

surrounding buildings, walls and operation RF fre-

quency. Thus, rather than RSS-based triangulation,

RSS database pattern matching algorithm is in use. It

overcomes the above-mentioned problems by using a

database built from measurements or predictions. The

position of the terminal is then determined by com-

paring the terminal’s RSS measurements to the data-

base entries and ﬁnding the best-matching position.

Making appropriate measurements over wide cellu-

lar network areas is very expensive and therefore not

considered applicable. Moreover, the change of cell

towers’ locations, antenna angles, and surrounding

buildings mandates frequent update of the database.

Thus, the pattern matching approach is more practical

by wireless LAN based small-area/indoor positioning

services. Prediction data, on the other hand, can be

obtained from wave propagation simulation tools and

remove the database maintenance overhead. Accu-

rate prediction, however, requires precise 3-D maps

over large areas and accurate/detailed network para-

meters including antenna loss, height, tilt, transmis-

sion power, etc., which are not commonly and easily

obtained. And the maps and parameters also require

frequent update.

In this paper, we propose to exploit the call data

records (CDRs) uploaded by Assisted Global Posi-

tioning System (AGPS) terminals as inputs to the pat-

tern database, which removes or reduces the pattern

collection overhead. The characteristics of RSS pat-

terns and AGPS CDRs are analyzed and reﬂected

in designing a pattern-distance metric and pattern

matching system, respectively. The main contribu-

tions of this paper are:

• Database cost reduction: we show that the AGPS

CDRs can be utilized to build and maintain a pat-

tern database for positioning of non-AGPS termi-

nals.

• Pros and cons of the use of the AGPS CDRs

– As the number of AGPS users increase, non-

AGPS terminals obtain more accurate position

results. It means that non-AGPS terminals can

obtain more accurate positioning results in the

area where more AGPS calls are generated (po-

sition information of non-AGPS terminals will

be requested much more frequently).

– Absence of AGPS CDRs in indoor areas may re-

duce the coverage of pattern matching position-

ing area. However, we show that indoor posi-

tioning by the proposed method is possible even

with outdoor AGPS CDRs.

• The analysis of RSS pattern characteristics: we de-

sign a pattern-distance metric based on the analy-

sis.

The performance of the pattern-distance metric and

the AGPS CDRs based pattern matching system are

evaluated by examining the ﬁeld data of urban down-

town area of Seoul, Korea. Field trials provide

promising results: the position of a user can be es-

timated with the accuracy at the level of 96.5m and

149.8m for the 67% and the 95% conﬁdence interval,

respectively.

The remainder of this paper is organized as follows.

In the next section, we present related work. Section

3 introduces the proposed pattern matching localiza-

tion, and Section 4 details the proposed pattern dis-

tance metric. In Sections 5 and 6, experimental results

and concluding remarks are given.

2 RELATED WORK

Cell-ID, time-difference-of-arrival (TDOA), GPS are

the traditional localization methods in cellular net-

works (3GPP, 2002), (Djuknic, 2001),(Zhao, 2002).

In Cell-ID systems, the position of user terminal is

determined as the position of serving cell tower. This

method is simple and applicable to every cellular net-

work, but it offers only coarse-grained position infor-

mation because the cell area is typically wide. TDOA

uses the time difference of the radio signal propaga-

tion to estimate the distance between the user terminal

and the adjacent cell towers. By using these distance

data, TDOA triangulates the position of the user ter-

minal. This method can provide a more accurate po-

sition than the Cell-ID method, but its application is

limited to synchronized networks or it introduces an

additional hardware cost to measure the asynchronity.

GPS provides specially coded satellite signals that can

be processed in a GPS receiver, enabling the receiver

to calculate the position, the velocity and the time;

accuracy of GPS is fairly high. By attaching a GPS

receiver on a user terminal, GPS can be employed in

cellular networks. However, GPS is not available in

indoors or deep urban canyons, because it requires

LOS to satellites. In cellular networks, on the other

hand, AGPS is used to reduce the time required to

ﬁnd the position of the user terminal.

Pattern matching localization method is proposed

to overcome the limitations of traditional methods

(Bahl, 2000), (Laitinen, 2001), (Ahonen, 2003),

(Borkowski, 2005). Under the pattern matching

method, a user terminal measures the radio signal pat-

tern, and then, seeks for the most similar pattern in the

pattern database, which consists of the radio signal

patterns gathered at the speciﬁed positions a priori. In

this way, the position of the user terminal is estimated.

(Bahl, 2000) proposes a pattern matching method for

wireless local area networks, while (Laitinen, 2001),

(Ahonen, 2003), (Borkowski, 2005) employ the pat-

tern matching in cellular networks.

Under the pattern matching method, because the

signal pattern database should be updated periodically

in order to adapt to the ever-changing radio environ-

ment, the maintenance cost is signiﬁcant. Accord-

ingly, a number of research efforts have been made

to reduce the maintenance cost. (Zhu, 2005), (Roos,

2002) employ the radio signal propagation model to

predict the radio signal patterns at the speciﬁc posi-

tions. Measured ﬁeld data can complement the radio

signal propagation model: it therefore reserves accu-

racy of the signal pattern database with a low pattern

database maintenance cost. This prediction method is

orthogonal to our proposed AGPS CDR base method,

thus, they can be used together with our method. Ac-

curate propagation modeling, however, requires pre-

cise 3-D maps over large areas and detailed network

parameters including antenna loss, height, tilt, trans-

mission power, etc. (Smailagic, 2002), (Lim, 2006)

propose special algorithms exploiting spatial correla-

tion of patterns in wireless LAN environments. Al-

though they are proved to work well in indoor wire-

less LAN systems, we observed that it is inappropri-

ate to apply them to cellular network systems because

of cellular systems’ larger cell coverage and more

dynamic radio environment than those of small-area

wireless LAN systems.

3 PATTERN MATCHING

LOCALIZATION

3.1 Basic Pattern Matching System

Figure 1 depicts the basic pattern matching system ar-

chitecture. In basic pattern matching systems, opera-

tors use dedicated measurement terminals and collect

signal patterns at positions (whose positions are al-

ready known) in advance. And the patterns are stored

in the signal pattern database. Signal patterns at a po-

sition may vary with the change of radio propagation

environment or cell planning. Therefore, operators

are required to periodically measure signal patterns to

maintain the signal pattern database up-to-date. We

call a collected signal pattern stored in signal pattern

database as a seed. That is, the seed is the entry in the

signal pattern database. On the other hand, we call a

signal pattern in a position request from a terminal as

a sample.

In order to determine the position of a user termi-

nal, the user terminal ﬁrst measures the signals from

surrounding cell towers, and sends the sample pattern

to the infrastructure to ﬁnd the most correlated pat-

tern which is used to estimate the position where the

terminal’s pattern is measured.

Collecting seed patterns over the wide area of the

cellular network is labor-intensive work. Suppose we

collect seeds at every 50 m grid point in 1km

range,

400 times of measurement are needed, and in case of

5km

urban area range, 10,000 times. Furthermore,

operators should measure seed patterns periodically

to maintain the database up-to-date. Moreover, in or-

der to obtain the more accurate position ﬁx, the more

and the denser seed patterns are needed. Therefore,

we propose a novel pattern matching system to auto-

mate the construction of the signal pattern database.

3.2 Proposed Pattern Matching

System

The proposed AGPS CDR based pattern matching

architecture is illustrated in Figure 2. In the pro-

posed system, we make use of the CDRs uploaded

by AGPS terminals as seed patterns for positioning

non-AGPS terminals. In general, AGPS is accurate

within 50 meters when users are indoors if GPS sig-

nals are received and 15 meters when they are out-

doors (Djuknic, 2001), so that we can leverage the

AGPS result as the actual position of a terminal. An

Seed Pattern

Terminal

User

B1. In response to a

position request, a

sample pattern is

measured and sent to

the cellular

infrastructure.

B2. Find the most

correlated seed in the

pattern DB, and reply

with the position of the

most correlated pattern

B3. User gets the fixed

position.

Measurement

Terminal

Operator

A1. Operators collect

signal patterns and

store them in signal

pattern DB.

Cellular Infrastructure

Figure 1: Basic pattern matching architecture. Process A is the pattern database construction process, and process B is the

user terminal positioning process.

Seed Pattern

Terminal

User

B1. In response to a

position request, a

sample pattern is

measured and sent to

the cellular

infrastructure.

B2. Find the most

correlated seeds in

pattern DB and

calculate the position,

and reply with the result

position

User

A1. In response to a

position request, an

AGPS terminal

measures satellite

signals and cell tower

signal patterns, and

reports them to the

cellular infrastructure.

A2. The position fix is

returned to the AGPS

terminal, and the CDR

is stored in pattern DB.

B3. User gets the fixed

position.

AGPS

Terminal

Cellular Infrastructure

Figure 2: AGPS CDR based pattern matching architecture. Process A is the pattern database construction process using AGPS

user terminals, and process B is the user terminal positioning process.

AGPS CDR includes the call time as well as the sig-

nal pattern and an AGPS positioning result (3GPP2,

2001). From the CDMA (IS-95) system of the com-

pany A in Korea, we have obtained this log data with-

out any modiﬁcation on the system. In this way, we

are able to construct and maintain a pattern database

at low cost.

In some cases, the size of the seed pattern data-

base may be too large to manage, since the number of

AGPS terminals is increasing and many users exploit

AGPS for the location based service. Then, the data-

base lookup time and, eventually, position ﬁx time

will increase. In this case, we can reduce the time

by doing spatial and temporal ﬁltering for the incom-

ing CDRs. For example, we can designate regional

databases for each area unit, and each database main-

tains CDRs generated from its assigned area. When

we choose candidate seed patterns from the database

to compare with the sample pattern, we can ﬁlter out

old seeds. We can also consider the speciﬁc time in-

terval (a day of the week, a time in a day) at which

the sample and seed patterns are measured in choos-

ing most appropriate seed patterns in a manageable

size.

On the other hand, in a particular area, or a partic-

ular cellular network where AGPS terminals are gen-

erating calls not so frequently, there may not be suf-

Table 1: The result error distances (m) on the time varia-

tion. This table shows the mean positioning accuracy and

the accuracies (in meters) for the 67% (1 sigma) and 95%

(2 sigma) conﬁdence interval, respectively.

Measurement date Mean 1 Sigma 2 Sigma

of sample

12/15/2005 87.7 110 169

2/8/2006 93.2 121 196

ﬁcient number of seeds in the pattern database. In

this case, the seeds should be accumulated in the pat-

tern database for a long time to offer required accu-

racy in that area. This incurs a question on the seed

valid time. We investigate this question: how long

the seed is valid, by using two sample sets of 55-day

difference, which is somewhat long time if we con-

sider the rapid change of outdoor radio environment

especially in urban area. We collected the seed set

on Dec. 15, 2005, and used two sample set: one was

measured on Dec. 15, 2005, and the other on Feb.

8, 2006. The seeds and samples were collected from

about 100 spots of downtown area of Seoul (near the

Gangnam station, one of the most crowded areas in

Korea, with many high buildings) and the area size

is approximately 1km × 1k m. The result error dis-

Table 2: An example of RSS pattern (unit: dBm).

Location RSS RSS RSS RSS

from from from from

cell 1 cell 2 cell 3 cell 4

(37.5085, -4.43 -7.23 -10.22 -20.46

127,0335)

tance of our proposed method is shown in Table 1.

The mean error of former sample set (with ’fresh’

seeds) is 87.7 m, while that of the latter sample set

(with ’55-day-old’ seeds) is 93.2 m. Despite of the

55 day gap between the two sets, the positioning ac-

curacy results do not make a considerable difference.

This demonstrates that the duration of the seed valid-

ity could be very long than our presumption. There-

fore, when AGPS terminals are generating calls not so

frequently, we can store seeds in the pattern database

for a long time to maintain the density of seeds in the

pattern database.

The basic pattern matching picks up only one seed

(and its measurement position) in determining the po-

sition of a sample pattern. Consequently, the result

of the basic pattern matching will become unstable

as the variations of the radio signals increase (due to

slow fading and fast fading). In order to mitigate the

effect of the variations, we select multiple seeds and

use them together in determining the position of user

terminals. We introduce a metric of the pattern dis-

tance, which will be detailed in Section 4, and select

multiple seed patterns in terms of the pattern distance.

We then estimate the centroid of the selected seeds as

the position of a user terminal. The ﬁeld trial test in

Section 5 shows that this centroid method exhibits

with less deviating positioning results.

4 PATTERN DISTANCE METRIC

We deﬁne a radio signal pattern as a set of the

received-signal-strengths (RSSs) of adjacent cell tow-

ers (Table 2). User terminals commonly measure the

RSS of the pilot channel to determine when to hand-

off to other cells, or to control transmit power in any

cellular network, i.e., GSM, CDMA, WCDMA net-

works.

Before designing a pattern distance metric, we an-

alyze the characteristics of RSS patterns. First we

look at the distribution of RSS values measured from

one cell tower as illustrated in Figure 3. X, Y, and

Z axes represent latitude, longitude, and RSS value

in dBm, respectively. Circular points indicate mea-

surement locations. In Figure 3, as the user terminal

goes far from the cell tower, the measured RSS value

Cell Tower

Figure 3: Signal strength at various positions received

from a cell sector antenna. Cell tower is at (37.50504,

127.02527).

decreases while the RSS values measured near the

cell tower are not decreasing fast. Although there are

some ﬂuctuation of RSS values due to slow and fast

fading, in general, the difference between two RSS

values measured at two positions increases in propor-

tion to the distance between the positions. With these

characteristics of the RSS, we reach the following ob-

servation.

• Observation 1. In general, as the distance between

two measurement positions increases, the differ-

ence between the two measured RSS values in-

creases. The pattern distance metric should reﬂect

the difference between RSSs of two patterns (i.e.,

a seed and a sample), then eventually indicate the

distance between the two measurement positions.

Therefore, in order to take RSS difference into an

account to pattern distance metric, we use the Euclid-

ean distance as follows.

k=1

(RSS

− RSS

)

where n is the number of RSSs in two patterns and

RSS

is the k

RSS value of the pattern A and

RSS

is the k

RSS value of the pattern B.

Because RSSs of the points at the same distance

from a cell tower are similar (as shown in Figure 3),

if a pattern (sample or seed or both of them) contains

only one RSS measurement value from one cell, the

pattern with a single RSS may appear at a number of

positions. In that case, the picked-up position of the

most correlated seed based on the above metric may

be far from the position of the sample pattern. How-

ever, as the number of RSS values from different cells

increases, an RSS pattern will have a fewer number of

candidate positions. In addition to Observation 1, we

come to another observation.

Table 3: An example of hole RSS. (unit : dBm).

RSS RSS RSS RSS

from from from from

cell 1 cell 2 cell 3 cell 4

Pattern A -3.42 -5.23 hole -16.78

Pattern B -6.23 -13.25 -8.43 hole

• Observation 2. Let S

and S

denote the set of

cells whose pilot signals are received by user A

and user B, respectively. As the number of cells

in their intersection increases, we can say that the

similarity (the distance between A and B) between

two patterns becomes more substantial. Therefore,

comparing two patterns, we need to consider the

number of cells common in the two patterns.

Patterns may have RSSs from a different set of

cells. Let us take an example of Table 3, in which

pattern A has RSSs from cell 1, cell 2, and cell 4, and

pattern B has RSSs from cell 1, cell 2, and cell 3. In

this example, we cannot calculate Euclidean distance

between patterns A and B. We call the RSS from the

cell whose RSS is measured by only one pattern (not

in the other pattern) as hole. In the example of Ta-

ble 3, the pattern A contains a RSS from cell 4 but

the pattern B does not: the RSS entry of cell 4 is a

hole in pattern B. Likewise, the RSS entry of cell 3 is

a hole in pattern A. As we have observed, if the num-

ber of holes of a pattern pair is large, the two patterns’

measurement positions are far apart.

In order to calculate the Euclidean distance be-

tween them, we assign a certain constant value to

holes: hole RSS. Through an experiment given in

Section 5, we ﬁnd that a constant somewhat lower

than the smallest RSS in the pattern database (also in

the sample patterns) is appropriate for the hole RSS.

According to Observation 2, we should give some

penalty to holes, and the proposed hole RSS method

gives penalty by assigning a small value to the hole:

as the hole RSS becomes smaller, a distance metric

value becomes larger.

We have tested our proposed metric by observing

the correlation coefﬁcient between the pattern dis-

tance metric and the actual geographic distance. High

coefﬁcient value (close to one) indicates the proposed

pattern distance metric reﬂects the actual geographic

distance well. From the data set (both seed and sam-

ple) of section 5, all possible pattern pairs are exam-

ined: we calculate the correlation coefﬁcient between

the vector of pattern distance metric and the vector

of geographic distance (between measurement posi-

tions). The results show that a promising value of a

coefﬁcient of 0.7.

The another application of the proposed distance

100

150

200

250

300

-50 -47.5 -45 -42.5 -40 -38 -35 -33 -30 -28 -25

Hole RSS (dB)

Error Distance (m)

mean

1sigma

2sigma

Figure 4: Experiment result with various hole RSS. 1 sigma

means the accuracy for the 67% conﬁdence interval and 2

sigma means the accuracy for the 95% conﬁdence interval.

metric is to infer conﬁdence level of a position ﬁx.

If the sample and the picked-up most correlated seed

has a small pattern distance metric, the measurement

position of the seed patten is close to the sample mea-

surement position, i.e., the position ﬁx is accurate

with a high probability.

5 EXPERIMENTAL RESULTS

We have performed extensive experiments with data

gathered from the commercial CDMA network of the

company A in the urban area of Seoul, Korea. The

gathered data sets are as follows.

• Seed pattern database: The pairs of the posi-

tion and the signal pattern were gathered by the

CDRs uploaded by AGPS terminals in the urban

area (near the Gangnam station) of Seoul. The area

size is 1km×1.8km, and we obtained seed patterns

of 283 positions in that area.

• Sample data set: We measured signal patterns at

30 points in the area covered by the signal pattern

database. We measured 5 samples at one point, in-

doors/outdoors separately at the same point

; 300

samples in total were measured. We retrieved the

real latitude and longitude of each point using the

digital map, which enabled us to calculate the error

distances of localization results.

Choice of Hole RSS: First, we have performed an

experiment to determine the appropriate value of the

hole RSS. We ﬁgured out under which value it shows

the best accuracy, varying the hole RSS from -50 dB

In our experiments, the CDRs contain the SINR values

(Ec/Io) from the adjacent cell towers, and we have used the

SINR values as signal pattern instead of the RSS.

For example, samples were measured inside and out-

side of a building entrance with a distance of several meters,

or measured under the roof and on the roof of a building.

100

150

200

250

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Seed Selection Range

Error Distance (m)

mean

1sigma

2sigma

Figure 5: Experiment result with various seed selection

range.

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 30 60 90 120 150 180 210 240 270

Error Distance (m)

Cumulative Probability

Indoor samples

Outdoor samples

Figure 6: Cumulative distribution function (CDF) accuracy

of proposed pattern matching that used with indoor and out-

door samples separately. Hole RSS is -45 dB and seed se-

lection range is 0.37.

100

150

200

250

300

350

400

450

40 70 100 130 160 190 220 250 280 310

Grid Size (m)

Error Distance (m)

mean

1sigma

2sigma

Figure 7: Experiment result with various grid size (seed

density). Hole RSS is -45 dB and seed selection range is

0.37.

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 50 100 150 200 250 300 350 400 450 500 550 600 650

Error Distance (m)

Cumulative Probability

Cell ID

Proposed PM

Figure 8: Cumulative distribution function (CDF) accuracy

of proposed pattern matching and Cell ID method. Hole

RSS is -45 dB, and seed selection range is 0.37, and all

seeds are used.

to -25 dB. Figure 4 shows that, with -45 dB of the hole

RSS, we can achieve the best accuracy; in this case,

the mean error distance is 87.4 m and 2 sigma (the ac-

curacy for the 95% conﬁdence interval) is 175.2m. In

our experiment environment, the minimum RSS that

a user terminal can detect is around -32 dB: -45 dB is

somewhat lower value than the minimum RSS. From

this result, we can infer that the somewhat lower value

than the minimum RSS which a user terminal can de-

tect is eligible when applying our proposed method to

other wireless networks.

Effect of multiple top seeds: Next, we have ﬁg-

ured out the effect of seed selection when determin-

ing the position of a user terminal. The basic pattern

matching uses only the most similar seed (top seed),

and determines the position of that seed as the user

terminal’s position. However, because of the momen-

tary ﬂuctuation of the radio signal caused by shad-

owing, the result of the basic pattern matching is not

stable. This is the reason why we propose to exploit

several similar seeds when estimating the position of

a user terminal in the previous section. We perform an

experiment, varying the seed selection range based on

the top seed metric. The seed selection range will be

denoted by S. Let m be the top seed metric, then we

will consider the positions of the seeds whose metric

is less than m × (1+S). Then we determine the posi-

tion of a user terminal as the centroid of those seeds.

We use -45 dB as the hole RSS in all following exper-

iments.

Figure 5 shows the error distances with varying the

value of the seed selection range. When the seeds are

selected within the appropriate seed selection range,

the accuracy of the result is better than that of the case

that only top seed is selected, i.e., when the seed se-

lection range is zero. From the graph, we ﬁnd that

the accuracy of the result is the best when the seed

selection range is 0.37, in which case, the mean error

distance is 76.8 m and the 2 sigma result was 149.8m.

In particular, the 2 sigma result shows more improve-

ment (180m to 150m) while the mean value exhibits

relatively small improvement (90m to 77m): this cen-

troid method decreases the deviation, and thus, has a

stabilization effect.

Indoor vs. outdoor: Throughout the previous ex-

periments, we have used the mixture of both the in-

door and the outdoor samples. Now, we test whether

our proposed pattern matching is suitable for the in-

door samples. Figure 6 shows the cumulative distri-

bution function (CDF) accuracy of our proposed pat-

tern matching with indoor and outdoor samples sepa-

rately. Although the indoor samples are usually col-

lected from several meters inside from building en-

trances and do not include deep basement samples, it

shows almost the same accuracy values in both indoor

and outdoor samples, from which we could conclude

our proposed pattern matching is suitable for both the

indoor and the outdoor samples.

Effect of seed density: The accuracy of our pro-

posed pattern matching shows some dependency on

the density of the seeds (the number of seeds in the

seed database per unit area). Through experiments,

we examine the relation between the density of seeds

and the accuracy of our method. In the experiment,

we divide a range into grids, and leave only one seed

in a grid. Figure 7 shows the experiment result. As

the grid size increases, the accuracy of the proposed

pattern matching method becomes lower. Particularly,

above 70 m of the grid size, the accuracy of our pro-

posed pattern matching decreases rapidly. Hence, in

our proposed pattern matching, the appropriate den-

sity of seeds has to be maintained to achieve the high

accuracy.

Comparison with Cell-ID: Finally, we have com-

pared the accuracy of our proposed method with the

Cell-ID method. Figure 8 shows the cumulative dis-

tribution function (CDF) accuracy of our proposed

pattern matching and Cell-ID method. As Figure 8

shows, the accuracy of our proposed pattern match-

ing is much higher than Cell-ID method.

6 CONCLUSION

In this paper, we have proposed a novel pattern match-

ing localization method for the cellular network, by

exploiting the CDRs uploaded by AGPS terminals as

inputs to the pattern database. We have analyzed the

characteristics of RSS patterns from AGPS CDRs,

and designed the pattern-distance metric and the new

pattern matching method using that metric. The pro-

posed positioning method reduces the pattern collec-

tion overhead by automating the construction of the

RSS pattern database. The experiment results demon-

strate that 1) the accuracy of the proposed position-

ing method is much higher than that of the Cell-ID

method 2) the proposed AGPS CDR based method

works well for indoor users as well as for outdoor

users 3) the density of seed patterns is closely related

to the positioning accuracy. For the future work, we

will investigate how to leverage other GPS-free po-

sitioning methods together with the proposed AGPS

CDR based pattern matching method.

REFERENCES

Djuknic, G.; Richton, R. (2001). Geolocation and Assisted

GPS. In IEEE Computer Magazine, Feb. 2001. IEEE

press.

Zhao, Y. (2002). Standardization of Mobile Phone Position-

ing for 3G Systems. In IEEE Communications Maga-

zine, Jul. 2002. IEEE press.

Laitinen, H.; L

ahttenm

aki, J.; Nordstr

om T. (2001). Data-

base Correlation Method for GSM Location. In VTC.

IEEE press.

Ahonen, S.; Laitinen, H. (2003). Database Correlation

Method for UMTS Location. In VTC. IEEE press.

Borkowski, J.; Lempi

ainen, J. (2005). Pilot correlation po-

sitioning method for urban UMTS networks. In Euro-

pean Wireless Conference 2005.

Zhu, J.; Durgin, G.D. (2005). Indoor/outdoor location of

cellular handsets based on received signal strength.

In Electronics Letters, Vol. 41, No. 1, Jan. 2005. IEE

press.

Daniel Wong, K. (2001). Get-Location in Urban Areas Us-

ing Signal Strength Repeatability. In IEEE Communi-

cations Letters, Vol. 5, No. 10, Oct. 2001. IEEE press.

Roos, T.; Myllym

aki, P.; Tirri, H. (2002). A Statistical

Modeling Approach to Location Estimation. In IEEE

Transactions on Mobile Computing, Vol. 1, No. 1,

Jan.-Mar. 2002. IEEE press.

Bahl, P.; Padmanabhan, V. N. (2000). RADAR: An In-

Building RF-based User Location and Tracking Sys-

tem. In INFOCOM 2000, IEEE press.

Smailagic, A.; Kogan, D. (2002). Location Sensing and

Privacy in a Context-Aware Computing Environment.

In IEEE Wireless Communications, Oct. 2002. IEEE

press.

Lim, H.; Kung, L.; Hou, J. C.; Luo, H. (2006). Zero-

Conﬁguration, Robust Indoor Localization: Theory

and Experimentation. In INFOCOM 2006, IEEE

press.

3GPP. (2002). Stage 2 functional speciﬁcation of User

Equipment (UE) positioning in UTRAN (Release 4).

3GPP.

3GPP2. (2001). Position Determination Service Standard

for Dual Mode Spread Spectrum Systems. 3GPP2.