Are GSM phones THE solution for localization

Are GSM phones THE solution for localization?



Alex Varshavsky½ , Mike Y. Chen , Eyal de Lara½ , Jon Froehlich¾ , Dirk Haehnel ,

Jeffrey Hightower , Anthony LaMarca , Fred Potter¾ , Timothy Sohn , Karen Tang¿ , and Ian Smith

½ ¾ ¿

Computer Science Computer Science and Engineering Computer Science

University of Toronto University of Washington Carnegie Mellon University

walex,delara @cs.toronto.edu jfroehli, fpotter @cs.washington.edu kptang@cs.cmu.edu



Computer Science and Engineering Intel Research Seattle

University of California, San Diego mike.y.chen, dirk.haehnel, jeffrey.r.hightower,

tsohn@cs.ucsd.edu anthony.lamarca, ian.e.smith @intel.com





Abstract location-enhanced applications are emerging that use in-

formation about places the user visits repeatedly instead

In this paper, we argue that localization solution based of latitude-longitude coordinates [11]. Examples of these

on cellular phone technology, specifically GSM phones, is place detection applications include automatic configura-

a sufficient and attractive option in terms of coverage and tion of wireless network settings based on place, recom-

accuracy for a wide range of indoor, outdoor, and place- mendation systems that learn user’ preferences by tracking

based location-aware applications. We present preliminary the places (e.g., restaurants, bars) the user visits, location-

results that indicate that GSM-based localization systems enhanced instant messengers, systems that allow setting re-

have the potential to detect the places that people visit in minders based on places that are important to the user [20].

their everyday lives, and can achieve median localization In this paper, we argue that localization based on cellular

accuracies of 5 and 75 meters for indoor and outdoor envi- phone technology, specifically GSM phones, can be a suffi-

ronments, respectively. cient and appropriate solution both in terms of coverage and

localization accuracy for this taxonomic spectrum of appli-

cations.

1 Introduction There is a conception that client-based GSM localiza-

tion is inherently inaccurate. GSM’s large cell sizes (GSM

macro-cells have a range of 35km, which can be extended if

Development of location-aware applications and sys-

necessary) seem to make it harder to achieve good localiza-

tems, especially in the social-mobile [19] and health care [6]

tion accuracies. In this paper, we argue that this conception

domains, has been driving the need for more accurate and

is in fact incorrect. We show that GSM localization achieves

pervasive localization technologies. We see a taxonomy

accuracies that are appropriate for many applications. In

emerging where location-enhanced applications can be cat-

indoor environments it is possible to perform room level

egorized into three types according to their needs: indoor

localization with GSM and achieve median localization ac-

localization, outdoor localization and place detection. Dig-

curacies of 2 to 5 meters. We also report our preliminary

ital homes [13], in-building navigation, and in-building co-

results for wide-area outdoor GSM localization, achieving

ordination between peers [21] all desire indoor localiza-

up to 75m median error. Finally, we show how having only

tion with at least room-level accuracy. City-wide tourist

GSM traces reported by a cell phone allows us to detect

guides [7, 2], the US FCC’s E911 mandate, and location-

places people visit in their everyday lives.

based web search [5] need outdoor wide-area localization

where coverage is often paramount to precision. Some ap-

plications benefit from the combination of both indoor and 2 Background

outdoor localization on a single easy to carry device. Those

include Computer Supported Cooperative Care (CSCC) [6], Many available localization technologies have low cov-

social-mobile computing [19] and gaming [3]. Finally, new erage or only work in a specific environment. The most

commonly available location technology today is the Global WiFi localization: (a) phones operates over a licensed band,

Positioning System (GPS). Although accurate and very ef- meaning no interference from microwave ovens and cord-

fective in open environments, GPS typically does not work less phones; (b) phones use a managed network, meaning no

well where people spend their time. For example, GPS does interference from neighboring access points that happened

not work well indoors, in urban canyons, or in similar ar- to be tuned to the same channel; (c) phone networks require

eas with limited view of sky. A recent study showed that significant installation investments, resulting in stable en-

GPS coverage is available only 4.5% of the time for a de- vironment that changes less frequently [15]; and (d) phone

vice carried in users’ pockets or purse during a typical day network coverage is greater than that of WiFi networks.

[15], although these numbers are admittedly worst-case and

they rise if only both mobile and stationary times are con- 3 GSM Primer

sidered. GPS-enabled devices are quite valuable and will

become more and more widespread, but it is clear that many

Global System for Mobile Communication (GSM) is the

systems require another technology to meet the coverage

most widespread cellular telephony standard in the world,

and accuracy demands of applications. Infrared [12], ul-

with deployments in more than 210 countries by over 676

trasound [18] and Bluetooth localization systems [1] work

network operators [8]. In North-America, GSM operates

well indoors, but deploying these technologies to the wide-

on the 850 MHz and 1900 MHz frequency bands. Each

area is either cost prohibitive or not technically possible, for

band is subdivided into 200 KHz wide physical channels

example, due to infrared interference from the sun.

using Frequency Division Multiple Access (FDMA). Each

The wide adoption of WiFi-enabled mobile devices and physical channel is then subdivided into 8 logical channels

rapid deployment of WiFi access points make WiFi lo- based on Time Division Multiple Access (TDMA). There

calization attractive. The RADAR project [4] pioneered are 299 non-interfering physical channels available in the

indoor WiFi location, achieving 2-3 meter median accu- 1900 MHz band, and 124 in the 850 MHz band, totaling

racy and inspiring many follow-up efforts [14, 9]. Place 423 physical channels.

Lab [15] introduced wide-area WiFi location, showing me- A GSM base station is typically equipped with a num-

dian accuracies ranging between 15 and 60 meters and ber of directional antennas that define sectors of coverage

high coverage. Examples of recent commercial systems or cells. Each cell is allocated a number of physical chan-

using Place Lab’s approach include Microsoft’s Virtual nels based on the expected traffic load and the operator’s

Earth (http://virtualearth.msn.com) and SkyHook Wireless requirements. Typically, the channels are allocated in a

(http://www.skyhookwireless.com). Finally, the Beacon way that both increases coverage and reduces interference

Print project [11] showed how the places people go to can between cells. Thus, for example, two neighboring cells

be learned and recognized, without relying on a coordinate- will never be assigned the same channel. Channels are,

based localization system. Unfortunately, because of their however, reused across cells that are far-enough away from

high power consumption, current WiFi-enabled devices are each other so that inter-cell interference is minimized while

not frequently used “on-the-go” and, unless a power line channel reuse is maximized. The channel to cell alloca-

is available nearby, are used intermittently. For exam- tion is a complex and costly process that requires careful

ple, Henderson et al. showed that although people do use planning and typically involves field measurements and ex-

their WiFi-enabled devices in several locations, they tend to tensive computer-based simulations of radio signal propa-

power them off before moving to a new place and do not gation. Therefore, once the mapping between cells and fre-

power them on unless necessary [10]. As a result, WiFi- quencies has been established, it rarely changes.

equipped devices cannot be used effectively as a platform Every GSM cell has a special Broadcast Control Channel

for location-enhanced applications that rely on continuous (BCCH) used to transmit, among other things, the identi-

network connectivity or spontaneous interactions, for exam- ties of neighboring cells to be monitored by mobile stations

ple, social-mobile and monitoring applications. for handover purposes. While GSM employs transmission

Fortunately, there are devices that people do carry with power control both at the base station and the mobile de-

them most of the time that have continuous network con- vice, the data on the BCCH is transmitted at a full and con-

nectivity: mobile phones. Mobile phones have low power stant power. This allows mobile stations to compare signal

consumption, ubiquitous connectivity, established interface strength of neighboring cells in a meaningful manner and

metaphors, wide adoption, and, most important for this pa- choose the best one for further communication.

per, research results suggests that they can offer indoor, We collected GSM traces using a Sony Ericsson GM28

outdoor, and place detection capabilities. We believe all GSM modem and an Audiovox SMT 5600 phone, depicted

these characteristics make mobile phones an excellent plat- in Figures 1 and 2. The modem operates as an ordinary

form for developing and deploying location-enhanced ap- GSM cell phone, but exports a richer programming inter-

plications. Phone localization has specific advantages over face. Both the modem and the phone provide two inter-

Figure 1. GM28 Sony Ericsson Figure 2. Audiovox SMT 5600

Modem Smart Phone









faces for accessing signal strength information: cellsAPI appear slightly lower than those achieved using WiFi lo-

and channelsAPI 1 . The cellsAPI interface reports the cell calization, they are comparable and sufficient for the same

ID, signal strength, and associated channel for the n neigh- types of location-enhanced applications.

boring cells. While the modem’s and the phone’s specifica- Two approaches to GSM and WiFi localization are fin-

tion does not set a hard bound on the value of n, in practice gerprinting and centroid, both of which require a

we saw the maximum value of n varying from 6-7. The training phase where given a set of GPS-stamped WiFi or

channelsAPI interface simultaneously provides the signal GSM traces the algorithm builds a model of an environ-

strength for up to 35 channels on the modem and 18 chan- ment, which it later uses for predicting device’s location.

nels on the phone. In practice, 6 of the channels typically Given the training traces, the centroid algorithm learns

correspond to the 6-strongest cells. Unfortunately, channel- the positions of radio beacons in the environment (i.e., WiFi

sAPI reports signal strength but does not report cell IDs. APs or GSM cell towers) by positioning the radio beacon

Although the modem exposes the cellsAPI and channel- in a location where the signal strength for that beacon was

sAPI explicitly, we are not aware of any GSM phone that observed the strongest. During the testing phase, the cen-

makes this information easily available. For example, to troid algorithm predicts a position of a measurement by

access cell and channel information on the Audiovox SMT averaging the positions of the radio beacons that appear in

5600 phone, we had to write a C tool that reads this data the measurement. Typically, giving a higher weight dur-

directly from the phone’s memory. ing averaging to radio beacons with stronger signal strength

We speculate that the fact that current phones do not ex- yields better localization accuracy. Unfortunately, walls,

pose similar interfaces reflects the unwillingness of network doors and other obstacles attenuate radio signals in an un-

operators to make signal strength information public. In- predictable way, making the centroid algorithm inaccu-

deed, by not exposing these interfaces, network operators rate in indoor environments. Therefore, we present cen-

can maintain a monopoly on the provisioning of location- troid results only for outdoor experiments.

based services. What needs to be done to influence network In contrast, the fingerprinting algorithm uses the

operators to allow exposing such interfaces is beyond the training set to build a mapping from measurements to po-

scope of this paper. Instead, in this paper we explore the sitions where those measurements were observed. Then,

opportunities that are made possible by the availability of during the testing phase,

signal strength information on the phone. fingerprinting matches every measurement in the

testing set to one or more measurements observed during

the training phase and then averages the true positions of

4 Indoor and Outdoor Localization the best matched measurements. Once again, weighting by

the signal strength of the best matched measurements yields

We begin this section by describing how GSM and WiFi better results.

localization works. We then present our experimental GSM We also show results for a random algorithm, which

localization results and show that although GSM accuracies predicts position by randomly picking a measurement from

1 The

the training set and assigning its position as the predicted

terms cellsAPI and channelsAPI are used to simplify presenta-

tion. In practice, the cellsAPI correspond to AT*E2EMM=1 command and

position, thus providing a lower bound on the performance

the channelsAPI correspond to the AT*E2NBTS? command on the GM28 of a localization system. The localization error, or the dis-

GSM modem, respectively tance between the true and predicted position, of random

WiFi fingerprinting GSM fingerprinting Random WiFi fingerprinting GSM fingerprinting Random GSM centroid GSM fingerprinting Random

100 100

40 100 97.01









% of successful floor classifications

93.69 1400

Median Localization Error (meters)









35.61 89.08

35 90









Localization Error (meters)

1200

80

30

70 1000

62.16

25 60

50 50

800

20 50

600

15 13.85 40 33

30 400

10

6.46 20

4.4 4.98 4.41 200

5 2.49 3.11 3.66

10

0

0 0

University Research Lab House University Research Lab House Belltown Redmond





Figure 3. Median indoor lo- Figure 4. Floor classification Figure 5. Outdoor localiza-

calization error. accuracy. tion error (median and 95%).







depends on the size of the area covered by the training set. between floors in both wooden and steel-reinforced con-

crete structures, achieving correct floor classifications be-

º½ ÁÒ ÓÓÖ ÄÓ Ð Þ Ø ÓÒ tween % and % of the time.





In our previous work [17], we presented the first accu- 4.1.2 Room level accuracy

rate GSM indoor localization system that achieves median

accuracy ranging from ¾ Ñ to Ñ in large multi-floor

To test room level localization accuracy, we first collected

buildings. We will first briefly summarize this work and a training trace by walking around with an Audiovox 5600

then present our new preliminary results for room level lo- SMT cell phone in 8 rooms on the 6th floor of Intel Re-

calization. search Seattle building. One hour later, we gathered an ad-

ditional similar trace using the same phone. We used fin-

gerprinting to match training and testing measurements.

4.1.1 Coordinate level accuracy First, we have repeatedly broken the training set down

into two random sets with 90% of points in the training set

To test our system, we collected GSM measurements on the

and 10% in the testing set. In all cases, we obtained 100%

5th and 6th floors of the Intel Research Seattle Lab build-

classification accuracy, which suggests that two consecutive

ing, the 5th and 7th floors of Bahen Center (the home to

measurements taken on the phone are typically very similar.

the Department of Computer Science of the University of

We then tried to match the measurements taken within an

Toronto) and at the basement, 1st and 2nd floors of a private

hour and saw 70% accuracy (87 of 126). More detailed

house located in a suburban Seattle area. The measurements

examination revealed that in most incorrect classifications

were collected about 1.5 meters apart. We tested the accu-

the predicted room was the next closest room.

racy with which GSM and WiFi localization systems based

Although we plan to perform more elaborate testing in

on fingerprinting [4] are able to differentiate floors of the

the future, these results suggest that room level localization

buildings and to localize mobile devices within the floor.

using GSM traces is feasible and GSM phones could sup-

Figure 3 and Figure 4 summarize our results. In our

port applications like in-building navigation and in-building

experiments, WiFi achieved within-floor localization accu-

coordination between peers.

racies consistent with previous findings [4, 9]. Also, be-

cause of reinforced concrete floors in the Research Lab

and the University buildings, WiFi was able to differentiate º¾ ÇÙØ ÓÓÖ ÄÓ Ð Þ Ø ÓÒ

floors perfectly. In the house environment, however, WiFi

achieves low classification accuracy as the house’s wood This section presents our preliminary GSM outdoor lo-

structure presents little obstacle to radio propagation, mak- calization results. We collected traces from a vehicle driv-

ing it harder to differentiate floors. ing in two neighborhoods in the Seattle metropolitan area:

GSM localization system performs well, achieving (a) Belltown, a mix of commercial and residential urban

within-floor localization results comparable to 802.11 sys- high-rises and (b) Redmond, a medium density residential

tem. Moreover, our GSM system effectively differentiates neighborhood. We collected GPS-stamped GSM traces us-

Figure 6. Place detection accuracy using GSM mobile phones.





ing a laptop connected to a GPS device and the Sony Erics- places they went to for a month. Figure 6 shows the effec-

son GSM modem. tiveness of our GSM place prediction algorithm in reconi-

The median and the 95% accuracy results for the cen- gizing when the phone is stable or mobile between places.

troid, fingerprinting and random algorithms are Precision is how often the algorithm’s prediction matches

summarized in Figure 5. As expected, fingerprinting the true state while recall is how many of the true states

achieves the best accuracy, with median error below 75m in were correctly identified by the algorithm. With this high

both areas we tested. The centroid algorithm performs accuracy, we argue it should be possible to extend this ca-

worse, achieving a 213m median error. These preliminary pability to a full place learning and recognition system for

results are encouraging, as 75m or even 213m median error GSM phones, analogous to what BeaconPrint [11] did with

is comparable to what is possible with WiFi and likely more WiFi. This capability would allow GSM phones to sup-

than sufficient for many wide-area location-enhanced appli- port applications like visit-driven recommendation systems,

cations such as social coordination and local web search. place-based device configuration, and context-aware notes

and reminders.

5 Place Detection

6 Conclusions and Future Work

In this section, we describe how using a stream of GSM

readings, we were able to effectively detect places people In this paper, we argued that for emerging location-

visit in their everyday lives. We developed an algorithm that enhanced applications, client-based GSM localization can

given a stream of time-stamped GSM readings, outputs the provide an adequate solution both in terms of coverage and

times when a person was at a “place”. Here, we consider a accuracy in a device people already carry. To dispel the

place to be a time interval during which our algorithm pre- notion that location systems using GSM phones are inher-

dicted that the user was stationary for more than 3 minutes. ently less accurate than systems built for WiFi devices, we

To gather data, we developed an application that runs on presented preliminary results showing that using GSM it is

an Audiovox SMT 5600 phone, continuously scans nearby feasible to achieve 2-5 meters median error and room-level

GSM cell towers once per second and allows users to use localization indoors, 70-200 meters median error outdoors,

text entry to name and select the places they went. Having and to detect places people go in their everyday lives. These

the GSM traces labeled with the ground truth data enabled results are comparable to what has been demonstrated pre-

us to test the accuracy of our place detection algorithm. viously for WiFi.

Our place detection algorithm applies a simple principle: In this paper, we discussed localization solutions based

when someone is at a place the stream of GSM readings on GSM cellular network. However, we believe that on-

their phone captures is “stable.” Currently, we measure sta- phone localization based on cellular networks is not spe-

bility by tracking the Euclidean distance in signal strength cific to GSM. Indeed, any cellular technology that transmits

space [4] between consecutive GSM readings. The smaller stable beacons from the cellular towers (e.g., for the need

the signal distance between two readings, the more similar of hand-off purposes) will make the on-phone localization

these readings are. When the phone is stationary, the dis- possible.

tance between consecutive GSM measurements tends to be The main drawback of many existing localization sys-

small, whereas when the cell phone is being carried around, tems, whether WiFi or GSM, is the non-trivial training

the Euclidean distance between consecutive GSM measure- required for the system to become usable. For instance,

ments will oscillate widely. We deployed the system to 5 fingerprinting-based solutions require a tedious training

users in our lab who carried the phones and labeled every data collection, and centroid-based solutions require as-

sembling maps of locations of cellular towers or access [13] J. Krumm, S. Harris, B. Meyers, B. Brumitt, M. Hale, and

points in the target environment. We already made ini- S. Shafer. Multi-camera multi-person tracking for easy liv-

tial steps toward reducing the training needed for wide- ing. In VS ’00: Proceedings of the Third IEEE International

area centroid-based location using a technique called bea- Workshop on Visual Surveillance (VS’2000), Washington,

con self-mapping [16]. Our future plans include investigat- DC, USA, 2000. IEEE Computer Society.

[14] A. Ladd, K. Bekris, G. Marceau, A. Rudys, L. Kavraki, and

ing novel ways of reducing training requirements for both

D. Wallach. Robotics-based location sensing using wireless

indoor and outdoor localization. ethernet. In Proceedings of the Tenth ACM International

Conference on Mobile Computing and Networking (MOBI-

References COM), 2002.

[15] A. LaMarca, Y. Chawathe, S. Consolvo, J. Hightower,

I. Smith, J. Scott, T. Sohn, J. Howard, J. Hughes, F. Potter,

[1] L. Aalto, N. Gothlin, J. Korhonen, and T. Ojala. Bluetooth J. Tabert, P. Powledge, G. Borriello, and B. Schilit. Place

and WAP push based location-aware mobile advertising sys- lab: Device positioning using radio beacons in the wild.

tem. In MobiSYS ’04: Proceedings of the 2nd international In Proceedings of the Third International Conference on

conference on Mobile systems, applications, and services, Pervasive Computing, Lecture Notes in Computer Science.

pages 49–58. ACM Press, 2004. Springer-Verlag, May 2005.

[2] G. D. Abowd, C. G. Atkeson, J. Hong, S. Long, R. Kooper, [16] A. LaMarca, J. Hightower, I. Smith, and S. Consolvo. Self-

and M. Pinkerton. Cyberguide: a mobile context-aware tour mapping in 802.11 location systems. In Proceedings of the

guide. Wirel. Netw., 3(5):421–433, 1997. Seventh International Conference on Ubiquitous Computing

[3] R. Anastasi, N. Tandavanitj, M. Flintham, A. Crabtree, (Ubicomp 2005), Lecture Notes in Computer Science, pages

M. Adams, J. Row-Farr, J. Iddon, S. Benford, T. Hemmings, 87–104. Springer-Verlag, September 2005.

S. Izadi, and I. Taylor. Can you see me now? a citywide [17] V. Otsason, A. Varshavsky, A. LaMarca, and E. de Lara.

mixed-reality gaming experience. In Proceedings of the Ubi- Accurate gsm indoor localization. In the Seventh Inter-

Comp, 2002. national Conference on Ubiquitous Computing (UbiComp

[4] P. Bahl and V. N. Padmanabhan. RADAR: An in-building 2005), September 2005.

RF-based user location and tracking system. In INFOCOM, [18] N. B. Priyantha, A. Chakraborty, and H. Balakrishnan. The

pages 775–784, 2000. cricket location-support system. In Mobile Computing and

[5] O. Buyukkokten, J. Cho, H. Garcia-Molina, L. Gravano, and Networking, pages 32–43, 2000.

N. Shivakumar. Exploiting geographical location informa- [19] I. E. Smith. Social mobile applications. IEEE Computer,

tion of web pages. In WebDB (Informal Proceedings), pages 38(4):84–85, April 2005.

91–96, 1999. [20] T. Sohn, K. Li, G. Lee, I. Smith, J. Scott, and W. G.

[6] S. Consolvo, P. Roessler, and B. Shelton. The carenet dis- Griswold. Place-its: Location-based reminders on mobile

play: Lessons learned from an in home evaluation of an am- phones. In Proceedings of Ubicomp, Tokyo, Japan, Septem-

bient display. In Proceedings of Ubicomp, 2004. ber 2005.

[21] J. C. Tang, N. Yankelovich, J. Begole, M. V. Kleek, F. Li,

[7] N. Davies, K. Cheverst, K. Mitchell, and A. Efrat. Using

and J. Bhalodia. Connexus to awarenex: extending aware-

and determining location in a context-sensitive tour guide.

ness to mobile users. In CHI ’01: Proceedings of the

IEEE Computer, 33(8), August 2001.

SIGCHI conference on Human factors in computing sys-

[8] J. Eberspacher, H.-J. Vogel, and C. Bettstetter. GSM switch-

tems, pages 221–228, New York, NY, USA, 2001. ACM

ing, services and protocols, 2001.

Press.

[9] A. Haeberlen, E. Flannery, A. M. Ladd, A. Rudys, D. S.

Wallach, and L. E. Kavraki. Practical robust localization

over large-scale 802.11 wireless networks. In Proceed-

ings of the Tenth ACM International Conference on Mobile

Computing and Networking (MOBICOM), Philadelphia, PA,

Sept. 2004.

[10] T. Henderson, D. Kotz, and I. Abyzov. The changing us-

age of a mature campus-wide wireless network. In Proc. of

MobiCom, 2004.

[11] J. Hightower, S. Consolvo, A. LaMarca, I. Smith, and

J. Hughes. Learning and recognizing the places we go.

In Proceedings of the Seventh International Conference

on Ubiquitous Computing (Ubicomp 2005), Lecture Notes

in Computer Science, pages 159–176. Springer-Verlag,

September 2005.

[12] A. Hopper, A. Harter, and T. Blackie. The active badge sys-

tem. In Proc. of INTERCHI-93, pages 533–534, Amsterdam,

The Netherlands, 1993.