490 likes | 634 Views
What to Do With Thousands of GPS Tracks. John Krumm Microsoft Research Redmond, WA. Location from FM Radio. Commercial FM. SPOT Watch. weather. dining. traffic. movies.
E N D
What to Do With Thousands of GPS Tracks John Krumm Microsoft Research Redmond, WA
Location from FM Radio Commercial FM SPOT Watch weather dining traffic movies Adel Youssef, John Krumm, Ed Miller, Gerry Cermak, and Eric Horvitz, "Computing Location from Ambient FM Radio Signals", IEEE Wireless Communications and Networking Conference (WCNC 2005), March 2005.
GPS Data Microsoft Multiperson Location Survey (MSMLS) 55 GPS receivers 227 subjects 1.77 million points 95,000 miles 153,000 kilometers 12,507 trips Home addresses & demographic data • Garmin Geko 201 • $115 • 10,000 point memory • median recording interval • 6 seconds • 63 meters Seattle Downtown Close-up Greater Seattle
Projects With GPS Tracks • Destination Modeling • Predestination – Destination prediction • Snap-to-Road – Map matching with temporal constraints • Personalized Routes • Location Privacy
Destination Models Destinations of drivers in our location survey John Krumm and Eric Horvitz, "Driver Destination Models", Eleventh International Conference on User Modeling (UM 2007), June 25-27, 2007, Corfu, Greece.
Ground Cover U.S. Geological Survey – Seattle Area What are the most attractive kinds of ground cover?
Trips vs. Time Time of Day Day of Week
New Destinations Drivers reach steady state after about two weeks • Rate of Decline vs. Demographics • single vs. partner – no significant difference • children vs. no children – no significant difference • extended family nearby vs. not – no significant difference • gender – women decline faster than men
Projects With GPS Tracks • Destination Modeling • Predestination – Destination prediction • Snap-to-Road – Map matching with temporal constraints • Personalized Routes • Location Privacy
Predestination Where do you want to go today? We already know, more or less. John Krumm and Eric Horvitz, "Predestination: Inferring Destinations from Partial Trajectories", Eighth International Conference on Ubiquitous Computing (UbiComp 2006), September 2006.
Going to the airport? Park with us for $8/day! Traffic Warning Destination Safeco Field(54% chance): 15-minute delay at I-405 & I-90. Suggest I-5 instead. Destination Seattle Center (31% chance): Broad St. closed. Suggest Denny Way instead. Why? • Anticipatory information • Location-based advertising • Hybrid vehicle efficiency
Basic Approach Greater Seattle, ~ 40km X 40 km 1 km grid
Ground Cover Prior ground cover prior
Personal Destinations Prior All Possible Destinations Destinations of One Subject
New Destinations Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9 Day 10 Day 11 Day 12 Day 13 Day 14 Personal destinations = visited cells + clustering + sparkling
Efficient Driving start Δt current location R r candidate destination
Trip Times From 2001 U.S. National Household Transportation Survey
Fuse Probabilities Final probability: Efficient driving likelihood: Trip time likelihood: Open-world prior: Closed-world prior: Wedding cakes: Ground cover:
Results • Half of trips (3667) for training efficiency distributions • Remaining half for testing • Leave-one-out for personal destinations prior
MSMLS Projects • Destination Modeling • Predestination – Destination prediction • Snap-to-Road – Map matching with temporal constraints • Personalized Routes • Location Privacy
Intelligent Snap-to-Road Goal: Infer actual route from noisy location data Snap-to-nearest road → disaster Spatial and temporal errors John Krumm, Julie Letchner, and Eric Horvitz, "Map Matching with Travel Time Constraints", Society of Automotive Engineers (SAE) 2007 World Congress, April 2007, Paper 2007-01-1102.
Error Distributions Assume that most nearest road snaps are correct Spatial Error (GPS error) Driving Time Error (MapPoint error) μ = 0.0 meters (assumed) σ = 1.4826*MAD(d) = 1.4826*median|di – median(d)| μ = median(Δt) = -0.5690 seconds σ = 1.4826*MAD(Δt) = 2.7725 seconds MAD(d) = median absolute distance
Hidden Markov Model Optimize tradeoff between spatial and temporal errors Spatial errors → observation probabilities Temporal errors → transition probabilities
MSMLS Projects • Destination Modeling • Predestination – Destination prediction • Snap-to-Road – Map matching with temporal constraints • Personalized Routes • Location Prvacy
Personalized Routing Percentage of trips in our data for which the driver’s actual route matched the… Shortest route: 27% Fastest route: 31% MapPoint route: 39% Neither shortest nor fastest: 60% Four routes from A to B, all different: Empirically fastest MapPoint plan Shortest distance Driver’s route Julia Letchner, John Krumm, and Eric Horvitz, "Trip Router with Individualized Preferences (TRIP): Incorporating Personalization into Route Planning", Eighteenth Conference on Innovative Applications of Artificial Intelligence (IAAI-06), July 2006.
Customize Routes Fastest at Midnight Fastest at Rush Hour Make route sensitive to our measured road speeds Reduce cost of roads that a user has driven
Personalized Routing Results Tested on ~2500 trips from MSMLS • Experiment Results: • 46.6% of personalized routes matched actual routes • Only 34.5% of drivers took the fastest route • Only 10.8% of routes show up more than once in the MSMLS data
MSMLS Projects • Destination Modeling • Predestination – Destination prediction • Snap-to-Road – Map matching with temporal constraints • Personalized Routes • Location Privacy
Motivation – Why Send Your Location? Congestion Pricing Pay As You Drive (PAYD) Insurance Location Based Services Collaborative Traffic Probes (DASH) Research (London OpenStreetMap) John Krumm, "Inference Attacks on Location Tracks", Fifth International Conference on Pervasive Computing (Pervasive 2007), May 13-16, 2007, Toronto, Ontario, Canada.
GPS Tracks → Home Location Algorithm 1 Last Destination – median of last destination before 3 a.m. Median error = 60.7 meters
GPS Tracks → Home Location Algorithm 2 Weighted Median – median of all points, weighted by time spent at point (no trip segmentation required) Median error = 66.6 meters
GPS Tracks → Home Location Algorithm 3 Largest Cluster – cluster points, take median of cluster with most points Median error = 66.6 meters
GPS Tracks → Home Location Algorithm 4 Best Time – location at time with maximum probability of being home Median error = 2390.2 meters (!)
Why Not More Accurate? • GPS interval – 6 seconds and 63 meters • GPS satellite acquisition -- ≈45 seconds on cold start, time to drive 300 meters at 15 mph • Covered parking – no GPS signal • Distant parking – far from home covered parking distant parking
GPS Tracks → Identity? Windows Live Search reverse white pages lookup (free API at http://dev.live.com/livesearch/)
Identification MapPoint Web Service reverse geocoding Windows Live Search reverse white pages
Countermeasure: Add Noise original σ= 50 meters noise added Effect of added noise on address-finding rate
Countermeasure: Discretize original snap to 50 meter grid Effect of discretization on address-finding rate
Countermeasure: Cloak Home Pick a random circle center within “r” meters of home Delete all points in circle with radius “R”