Review: Accuracy Characterization for Metropolitan-scale Wi-Fi Localization
Authors: Cheng, Chawathe, LaMacra, Krumm 2005Slides Adapted from Cheng, MobiSys 2005
Review by: Jonathan Odom
Location, Location, Location• Require accurate location for many applications but GPS
only works well outdoors and drains battery• Wi-Fi APs are commonly found in populated areas and
hardware is low cost/low power compared to GPS• Use Wi-Fi APs as location beacons• Requires map of APs
– Indoor version RADAR has high overhead– Only need accuracy on the order of 10 m
Manhattan (Wigle.net)
War-driving
• Used to create training data• Drive a laptop with Wi-Fi card and GPS through
the streets of city and collect information• Data – “radio map”
– AP unique ID– GPS location of received signal– Signal strength– Response Rate
Experimental Data Sets
Downtown(Seattle)
Urban Residential(Ravenna)
Suburban(Kirkland)
Algorithm - Centroid
• 1st of 3/4 algorithms used• Use arithmetic mean of positions of all AP’s• Not actually use centroid
AP
AP
APEstimate
Algorithm – Fingerprinting SS
• Use 4 closest APs in the Euclidean distance defined by signal strength (k-nearest neighbor)
• Assuming is the signal strength from the th AP from the map and is from the received data
• Weighting showed only marginal improvement• Allow +/- 2 APs for robustness over time• Based on Bahl 00
Algorithm –Fingerprinting Rank
• All hardware will not give same signal strength• Instead rank signal strength and use correlation with
3 points from radio map
• Where and denotes the mean• Based on Krumm 03
=(-20, -90, -40) -> =(1,3,2)
Algorithm - Particle Filter
• Particle filters, or a Sequential Monte Carlo method, is a recursive Bayesian estimator
• Empirical data model, using training data– Signal strength as function distance to AP– Response rate as function of distance to AP
• Random walk assumed for motion• Often used for noisy non-linear or non-
Gaussian models
Full Results
• Rank algorithm does not work with sparse APs
0
10
20
30
40
50
60
70
Downtown UrbanResidential
Suburban
Med
ian
Erro
r (m
eter
s)
Centroid (Basic)
Fingerprint (Radar)
Fingerprint (Rank)
Particle Filter
More APs Lowers Error
• Rank requires more than 1
AP Reduction
• Localization works well even with 60% APs lost
0
20
40
60
80
100
0% 20% 40% 60% 80% 100%
Med
ian
erro
r (m
eter
s)
AP Turnovers
centroid
particle filter
radar
rank
Adding Noise to GPS Data
• Centroid and particle filter work with noise
Reducing Map Density
• Works well up to 25 mph, 1 scan/sec