Transcript Accuracy Characterization for Metropolitan-scale Wi

Accuracy Characterization for
Metropolitan-scale Wi-Fi Localization
Yu-Chung Cheng (UCSD, Intel
Research)
Yatin Chawathe (Intel Research)
Anthony LaMarca (Intel Research)
John Krumm (Microsoft Research)
Introduction

Context-aware applications are prevalent
–
–
–
–

Maps
Location-enhanced content
Social applications
Emergency services (E911)
A key enabler: location systems
–
Must have high coverage

–
Low calibration overhead

–
Work wherever we take the devices
Scale with the coverage
Low cost

Commodity devices
slide2
Riding the Wi-Fi wave

Wi-Fi is everywhere now
–
–
–
–
No new infrastructure
Low cost
APs broadcast beacons
“War drivers” already build AP
maps



Position using Wi-Fi
–
–

Calibrated using GPS
Constantly updated
Indoor Wi-Fi positioning gives 23m accuracy
But requires high calibration
overhead: 10+ hours per building
What if we use war-driving
maps for positioning?
Manhattan (Courtesy of Wigle.net)
slide3
Why not just use GPS?

High coverage and accuracy (<10m)

But, does not work indoors or in urban
canyons

GPS devices are not nearly as prevalent
as Wi-Fi
slide4
Methodology

Training phase
–
–
Collect AP beacons by “war driving”
with Wi-Fi card + GPS
Each scan records


–
–

A GPS coordinate
List of Access Points
Covers one neighborhood in 1 hr
(~1 km2)
Build radio map from AP traces
Positioning phase
–
–
Use radio map to position the user
Compare the estimated position w/
GPS
slide5
Downtown vs. Urban Residential vs.
Suburban
Downtown
(Seattle)
Urban Residential
(Ravenna)
Suburban
(Kirkland)
slide6
Evaluation

Choice of algorithms
–


Naïve, Fingerprint, Particle Filter
Environmental Factors
–
AP density: do more APs help?
–
#APs/scan?
–
AP churn: does AP turnover hurt?
–
GPS noise: what if GPS is inaccurate?
Datasets
–
Scanning rate?
slide7
Compare Accuracy of Different Algorithms

Centroid
–


Estimate position as arithmetic mean of positions of all heard APs
Fingerprinting
–
User hears APs with some signal strength signature
–
Match closest 3 signatures in the radio map
–
RADAR: compare using absolute signal strengths [Bahl00]
–
RANK: compare using relative ranking of signal strengths
[Krumm03]
Particle Filters
–
Probabilistic approximation algorithm for Bayes filter
slide8
Baseline Results
70
Median Error (meters)
60
Centroid (Basic)
50
Fingerprint (Radar)
40
Fingerprint (Rank)
30
Particle Filter
20
10
0
Downtown
Urban
Residential
Suburban
• Algorithms matter less (except rank)
• AP density (horizontal/vertical) matters
slide9
Effect of APs per scan
• More APs/scan  lower median error
• Rank does not work with 1 AP/scan
slide10
Effects of AP Turnovers
Median error (meters)
100
80
centroid
60
particle filter
40
radar
rank
20
0
0%
20%
40%
60%
AP Turnovers
80%
100%
• Minimal effect on accuracy even with 60% AP
turnover
slide11
Effects of GPS noise
• Particle filter & Centroid are insensitive to GPS noise
slide12
Scanning density
• 1 scan per 10 meters is good == 25 mph driving speed at 1 scan/sec
• More war-drives do not help
slide13
Summary

Wi-Fi-based location with low calibration overhead
–

Positioning accuracy depends mostly on AP density
–
–
–

Urban 13~20m, Suburban ~40m
Dense ap records get better acuracy
In urban area, simple (Centroid) algo. yields same accuracy as
other complex ones
AP turnovers & low training data density do not degrade
accuracy significantly
–

1 city neighborhood in 1 hour
Low calibration overhead
Noise in GPS only affects fingerprint algorithms
slide14
Q&A
http://placelab.org
slide15