Accuracy Characterization for Metropolitan-scale Wi-Fi Localization Ying Wang, Xia Li Ying Wang, Xia Li

Outline Introduction Methodology Results Evaluation Summary

Introduction- What does the paper do  Outdoor Location mechanism based on Wi-Fi  Explore the question of how accurately a user's device can estimate its location using existing hardware and infrastructure and with minimal calibration overhead slide3

Introduction- Why We Need Location Context-aware applications are prevalent – Maps – Location-enhanced content – Social applications – Emergency services (E911) A key enabler: location systems – Must have high coverage Work wherever we take the devices – Low calibration overhead Scale with the coverage – Low cost Commodity devices

Introduction- 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

Introduction- Why Wi-Fi Wi-Fi is everywhere now – No new infrastructure – Low cost – APs broadcast beacons – “War drivers” already build AP maps Calibrated using GPS Constantly updated Position using Wi-Fi – Indoor Wi-Fi positioning gives 2- 3m accuracy – But requires high calibration overhead: 10+ hours per building Manhattan (Courtesy of Wigle.net)

Methodology 1. Training phase (war driving) Position 1 Position 2 Position 3 GPS Wifi card (x 1, y 1 ) (x 3, y 3 ) (x 2, y 2 ) A GPS coordinate List of Access Points

Methodology 2. Positioning phase (x 1, y 1 ) (x 3, y 3 ) (x 2, y 2 ) Position 1 Position 2 Position 3 Use radio map to position the user (x’, y’) A B C

Methodology Problem: How to make position estimation? (x’, y’) (x 3, y 3 ) Answer: By using positioning algorithms

Methodology- Positioning Algorithm 1.Centroid Algorithm Basic Centroid Weighted Centroid 2. Fingerprinting Algorithm Radar Fingerprinting Ranking Fingerprinting 3. Particle Filters

Methodology- Positioning Algorithm 1.Centroid Algorithm  Basic Centroid AP1(x 1,y 1 ) AP3(x 3,y 3 ) AP2(x 2,y 2 ) (x’, y’) Estimated

Methodology- Positioning Algorithm 1.Centroid Algorithm  Weighted Centroid AP1 (x 1,y 1 ) AP3 (x 3,y 3 ) AP2 (x 2,y 2 ) ss 1 ss 2 ss 3 (x’, y’)

Methodology- Positioning Algorithm 2.Fingerprinting Algorithm  What is Fingerprinting? (x 1, y 1 ) ss

Methodology- Positioning Algorithm 2.Fingerprinting Algorithm  Radar Fingerprinting A C B ss A ss B ss C ss’ A ss’ B ss’ C choose “4” nearest GPS coordinates GPS coordinate Access Points New user

Methodology- Positioning Algorithm 2.Fingerprinting Algorithm  Ranking Fingerprinting  All hardware will not give same signal strength  Instead of comparing signal strength directly, this method compares the rank of signal strength  is spearman coefficient. Higher -> more similar rankings SS = (-20, -90, -40)  R = (1,3,2)

Methodology- Positioning Algorithm 3.Particle Filters Key point of Particle Filter: Fusion Sensor Model Motion Model Note: The actual fusion calculation is more complicated, not this linear equation

Results -AP Density Downtown (Seattle) Urban Residential (Ravenna) Suburban (Kirkland)

Results- Table Median error in meters for all of algorithms across the three areas

Results- Histogram Algorithms matter less (except rank) AP density (horizontal/vertical) matters

Evaluation Choice of algorithms – Naïve, Fingerprint, Particle Filter Environmental Factors – AP density: do more APs help? – AP churn: does AP turnover hurt? – GPS noise: what if GPS is inaccurate? – Scanning rate?

Effect of APs per scan More APs/scan  lower median error Rank does not work with 1 AP/scan

Effects of AP Turnovers Minimal effect on accuracy even with 60% AP turnover

Effects of GPS noise Particle filter & Centroid are insensitive to GPS noise

Scanning density 1 scan per 10 meters is good == 25 mph driving speed at 1 scan/sec More war-drives do not help

Summary Wi-Fi-based location with low calibration overhead – 1 city neighborhood in 1 hour Positioning accuracy depends mostly on AP density – Urban 13~20m, Suburban ~40m – Dense AP records get better accuracy – In urban area, simple (Centroid) yields same accuracy as other complex ones AP turnovers & low training data density do not degrade accuracy significantly – Low calibration overhead Noise in GPS only affects fingerprint algorithms

Q & A Any Questions? *The slides were edited based on the original ppt from Yu-Chung Cheng