1. Practical Metropolitan-Scale Positioning for GSM Phones
Presented by Khushnood Irani

2. Authors Mike Y. Chen Intel Research Seattle, USA
Timothy Sohn University of California at San Diego, USA
Dmitri Chmelev University of Washington, USA
Dirk Haehnel Intel Research Seattle, USA
Jeffrey Hightower Intel Research Seattle, USA
Jeff Hughes University of Washington, USA
Anthony LaMarca Intel Research Seattle, USA
Fred Potter University of Washington, USA
Ian Smith Intel Research Seattle, USA
Alex Varshavsky University of Toronto, Canada

3. Need for Localization?
The E911/E112 initiatives in the US & Europe specify requirements on Localization Accuracy for mobile phones placing emergency calls.
These initiatives have catalyzed a market for network operator-provided location capabilities and services like:
AT&T Wireless’ friend-finder.
Sprint-NexTel’s fleet management tools.

4. Introduction
To examine the location accuracy of a positioning system in a metropolitan environment we need to answer one question:
Which platform serves best for location aware applications?

5. Introduction
Global Positioning System (GPS) though the most common location technology, serves well for outdoor environments but provides inadequate information indoors.
GSM & WiFi based location techniques are better in overcoming the shortcoming in GPS.
Operators provide services to calculate mobile phone positions using:
Hybrid Network Client Techniques
Assisted GPS (AGPS)
Network Only Techniques
Enhanced Observed Time Difference(EOTD)
Angle Of Arrival (AOA)
Time Difference of arrival (TDOA)

6. Introduction
WiFi beacon-based positioning system results in good indoor as well as outdoor location system with high coverage & good accuracy.
Wide area beacon-based approach complement many indoor positioning systems that provide high precision in indoor environment but require specialized hardware or have high installation costs.

7. Mobile Phones: A Better Platform
Introduction
GSM is the most widespread cellular telephony standard in the world & has a subscriber base which is far greater than internet users.
GSM mobile phones have a longer battery life, constant connectivity & are usually handheld.
Range of GSM is 35km which is 70 times larger than the 500m range that WiFi offers.
GSM network is well planned & stable as compared to the ad-hoc deployment of WiFi access points.
GSM uses licensed frequency bands which is less prone to interference.

8. Methodology The Methodology is divided into three parts:
Data Collection
Description of the underlying hardware & software of the system as well as the type of data collected.
2. Trace Characteristics
Description of the trace drive for the training & testing periods.
Positioning Algorithms
Examination & Comparison of the performance of the three position algorithms.

9. Data Collection Methodology Hardware
1. IBM Thinkpad T30 laptop (with a WiFi card)
2. Two GPS units
3. Three Sony Ericsson GM28 GSM modems
4. Three Audiovox SMT5600 phones
(The phones & GSM modems contain SIM cards)

10. Data Collection Methodology Software
The data collection Software is implemented in C#.
This software records the Cell ID & the signal strength (in terms of dBm) for seven cells once every three seconds.
Readings from the GPS are recorded every second.

11. Trace Characteristics
Methodology
Setup was implemented in a car with roof mounted antennas for better reception.
Car driven through public accessible street in Seattle metropolitan area.
For better accuracy the training data required to be large enough.
Duration of the drive was 208 hours & covered 4350 km.
24 GB of traces collected which contain 6756 unique cells.
The calibration trace was used to train three position algorithms.

12. Trace Characteristics
Methodology
To measure the accuracy of the algorithm, three test neighborhoods were selected for the testing phase of the analyses.
(Downtown with high cell tower density & Residential with low cell tower density)

13. Trace characteristics
Methodology
Training Trace
Testing Trace
Name
Downtown
Residential
Duration
208hr
70 min
169 min
Distance
4350 km
24 km
89 km
Dimension
25.0 km x 18.6 km
2.7 km x 2.3 km
2.6 km x 4.1 km +
4.6 km x 5.5 km
Area
Greater Seattle
Downtown Seattle
Ravenna
East Bellevue
Avg. Cell Density
28 cells/km2
66 cells/km2
26 cells/km2
Properties of the training & testing traces.

14. Positioning Algorithms
Methodology
The performance of three positioning algorithms are measured.
Centroid Algorithm
Radio Fingerprinting Algorithm
Gaussian Process-based Monte Carlo Localization
(These algorithms were chosen because they represent the spectrum of positioning algorithms & vary in complexity & accuracy)

15. Positioning Algorithms
Methodology
The algorithms were implemented in a C# location toolkit that runs on Microsoft Windows platforms such as:
Smartphones
PDAs
PCs
The toolkit can poll GSM readings & calculate the location four times per second using the centroid algorithm & the cell tower maps for the various cells observed.

16. Positioning Algorithms
Methodology
Analyzing the positioning accuracy of the algorithms was done in two phases:
Training trace
Time-stamped GSM & GPS measurements are used to build models specific to an algorithm.
Testing trace
GSM measurements are used to estimate the position & provide the estimated latitude & longitude values.
(NOTE: The positioning error was computed by calculating the distance between the estimated position (GSM) & the true ground position (GPS))

17. Centroid Algorithm Methodology | Positioning Algorithms
The centroid algorithm is very fast to compute & does not employ a radio propagation model.
Using a lookup table (Cell ID, Latitude & Longitude), the algorithm estimates the phones location to be the geometric center of all the cells seen in the measurement.
The algorithm also approximates the tower positions by averaging the places where the highest signal strengths for each was observed.
(NOTE: In the USA, true cell tower locations are not publicly available.)
To evaluate the placement accuracy, six cell towers were selected at random whose locations were physically verified.
This method gave an average error of 56m & a maximum error of 76m.
Weighting by the observed signal strength could improve the accuracy.

18. Estimated tower positions for the two test areas.
Centroid Algorithm
Methodology | Positioning Algorithms
Estimated tower positions for the two test areas.

19. Radio Fingerprinting Algorithm
Methodology | Positioning Algorithms
The fingerprinting algorithm does not create a map of estimated tower positions, nor does it model radio propagation. In fact there is a radio profile that is feature rich in space and reasonably consistent with time.
This method assumes that the radio beacons & the associated signal strengths observed at a location are stable over time.
During the training phase the algorithm constructs a search index that maps the radio fingerprints to locations in terms of latitude & longitude coordinates.
During the testing phase the algorithm uses the constructed search index to deduce the phone’s location by calculating the Euclidian distance in signal strength space between the current fingerprint & all available fingerprints in the index.
The algorithm then selects K fingerprints with the smallest Euclidian distance & estimate the location of the device by averaging the latitude & longitude coordinates of the K matches.

20. Radio Fingerprinting Algorithm
Methodology | Positioning Algorithms
How does it work?
The algorithm uses the K-nearest neighbors technique for matching fingerprints which estimates the location of the testing point in two stages.
First, the algorithm scans through all training points and calculates the Euclidean distance in signal space between the testing point and each of the training points.
The algorithm produces an estimate of the testing point’s location by averaging the locations of the K training points with the smallest Euclidean distance.
To compute the Euclidean distance, the algorithm uses readings for all available radio sources in the fingerprint whose accuracy depends on the density of the collected fingerprints
Example: if a training fingerprint contains signal-strength readings for 3 sources {Rtr1,Rtr2,Rtr3} and a testing fingerprint has signal-strength readings for the same 3 sources {Rtst1,Rtst2,Rtst3} then the Euclidean distance between the two fingerprints will be calculated as:
(Rtr1 – Rtst1)2 + (Rtr2 – Rtst2)2 + (Rtr3 – Rtst3)2

21. Monte Carlo Localization
Methodology | Positioning Algorithms
This algorithm uses an abstract parametric radio propagation model with Markov localization to predict the location (estimated using a Bayesian particle filter).
Gaussian processes are used to model the signal propagation that estimate the Gaussian distribution over functions based on training data.
In order to achieve fast execution, the signal propagation function is pre-processed to a grid with 15m grid cells.
The computation of the signal propagation can be implemented by a lookup function in the maps of the cell towers.
The likelihood of an observation can be computed from the predicted signal strength provided the phone is at a particular location.

22. Monte Carlo Localization
Methodology | Positioning Algorithms
Standard Monte Carlo localization (called particle filtering) is applied to represent the posterior probability distribution about the positioning of the phone.
In Monte Carlo Localization the phone’s position is represented by a set of random samples.
Each sample consists of a state vector (which is the position of the device) & a weighting factor.
The weight is the likelihood of the measurement at the particle’s location represented by the distribution of the samples & their importance factors.

23. Results
The analyses presented explores the effects of the five factors on positioning accuracy.
Choice of Algorithm
Scan set size
Simultaneous use of cells from different service providers.
Training & testing on different devices.
Calibrating drive density.
NOTE: To characterize the GSM accuracy, the three service providers have been made anonymous. Also the results presented are based on HTC Typhoon phone used for testing & training purposes.

24. Effect of Algorithm Selection
Results
The positioning error was evaluated for each of the three algorithms using the test traces from the test areas.
Downtown (Higher density)
Residential (Lower density)
50%
90%
Centroid
232m
574m
760m
2479m
Fingerprinting
94m
291m
277m
984m
Gaussian Processes
126m
358m
196m
552m
Median & 90th – percentile positioning errors

25. Effect of Scan Set Size Results
At any given time a GSM phone is in the range of a number of cells.
The GSM phones & modems used provide not only information about the current cell with which they are associated but also six nearby cells thus making a total of seven cells.
The number of cells are varied between one & seven by sorting them according to their signal strengths & using only the n-strongest cells.

26. Sensitivity analyses of positioning error versus the number of cells.
Effect of Scan Set Size
Results
Sensitivity analyses of positioning error versus the number of cells.

27. Effects of Using Towers from Multiple Service Providers
Results
GSM devices only monitor the cells from towers of their respective service providers even though cell towers from other service providers may be closer & having stronger signals.
Positioning accuracy ideally can be improved by increasing the number of observable cell towers thereby increasing the number of cells to scan.
To evaluate this effect, a cross-provider device was simulated by combining the measurements from the three service providers.

28. Effects of Using Towers from Multiple Service Providers
Results
Downtown
Residential
Single Provider (7)
Cross- Provider (7)
Cross- Provider (All)
Single Provider (7)
Centroid
187m
166m
170m
647m
456m
574m
Fingerprinting
94m
153m
245m
277m
313m
297m
Gaussian Process
126m
87m
65m
196m
147m
134m
Cross-provider median positioning error.

29. Effects of Using Towers from Multiple Service Providers
Results
Cumulative Distribution Function of positioning error for the Gaussian Process Algorithm in Downtown

30. Effects of Training on One Device & Testing on Another
Results
The algorithms require that another device should observe similar towers & its signal strength values must correlate with the HTC Typhoon phone.
A transformation function with a strong correlation can convert the device’s signal strength values into those reported by the HTC Typhoon phone.
Signal strengths of commonly observed cells were compared among the three GSM devices:
A duplicate HTC Typhoon phone.
An HTC Tornado phone.
A Sony Ericsson GM28 modem.
(These devices have different radio & antenna designs.)

31. Effects of Training on One Device & Testing on Another
Results
Devices
Radio
Antenna
Average # of Common Cells
Signal Strength Correlation
Correlation Significance
HTC Typhoon (reference)
Same
7.000
1.000
.000
HTC Typhoon (duplicate)
6.484
0.828
< .001
HTC Tornado
Different
5.018
0.789
Sony Ericsson
GM28 Modem
4.283
0.874
Similarity between different GSM devices and the reference HTC Typhoon phone, showing the number of cells that are the common when two different devices scan at the same time. Pearson correlation coefficient and significance are shown for the signal strengths of these common cells between each device and the reference phone.

32. Effects of Training on One Device & Testing on Another
Results
Downtown
Residential
50%
% change
Centroid
245m
5.6%
818m
7.6%
Fingerprinting
366m
289%
803m
190%
Gaussian Processes
206m
63%
307m
57%
Cross-device median positioning error and % change when training with the trace collected on one device (the HTC Typhoon phones) and testing on another device (the Sony Ericsson GM28 modems)

33. Effects of Reducing Calibration Drive Density
Results
A tradeoff between the calibration drive density & positioning error is characterized by simulating sparser driving patterns from the comprehensive data set.
Understanding this effect is useful to estimate the resource & cost necessary to calibrate a GSM-based positioning system.
Sophisticated algorithms rely on calibration data to improve positioning accuracy because the radio models degrade in quality with less calibration.
Vital street grid patterns are superimposed on the dense calibration trace & measurements that do not fall on the virtual street grid are filtered.
To reduce the systematic error, five random offsets for each grid width are used which are then averaged to estimate the positioning error.

34. Effects of Reducing Calibration Drive Density
Results
Example of a generated, virtual street grid that simulates a drive density equivalent to 10% of the full training trace.

35. Effects of Reducing Calibration Drive Density
Results
Median positioning error as a function of the calibration drive density using simulated street grids for downtown.

36. Conclusion Summary of events:
Examined the practical challenges in deploying a GSM beacon-based location system in a metropolitan environment.
Characterized wide-area positioning accuracy for three classes of algorithms & investigated the effects of several practical issues such as cross-device positioning & calibration drive density.
Presented a novel cross-provider positioning technique that improves positioning accuracy significantly.

37. Conclusion
Storage Required
CPU Usage
Accuracy
(Dense Towers)
Accuracy (Sparse Towers)
Required Density of Training Data
Requires Same Device Training Set
Benefits from Cross- Provider Scanning
Tolerant of Phones Exposing Single Cell
Centroid
Low (445KB)
Low
232m
760m No
Yes
Fingerprinting
High (188MB)
Medium
94m
277m
High
Gaussian Processes
Medium (80MB)
126m
196m
Summary of the characteristics of the three positioning algorithms.

38. Conclusion What the results show:
Analysis of the calibration drive density show that 30% of the data set was sufficient to provide comparable positioning accuracy & suggested that 60hrs of driving can cover a metropolitan area.
Cross-device positioning is possible with a little degradation in accuracy (for centroid algorithm) on devices with different radios & antennas.
Scanning cells across all available service providers can improve the accuracy.

39. Thank You!