On Road Vehicle Activity GPS Data and Privacy

Vetri Venthan ElangoDr. Randall GuenslerSchool of Civil and Environmental EngineeringGeorgia Institute of TechnologyOn Road Vehicle Activity GPS Data and Privacy

OverviewIntroductionBackgroundDataMethodologyCase StudyConclusions and Future Work2IntroductionThe use of GPS devices in travel behavior studies continues to increase in frequency and depthGPS devices can provide accurate and detailed spatial and temporal dataHigh resolution GPS data are useful in studyingTravel behaviorDriver behaviorSafetyEmissionsEtc.

3ObjectiveHigh resolution GPS data have a significant potential to compromise privacyAggressive driving (speed/acceleration) - lawsuitsHome departure and arrival - home securityLocations visitedMethodological goalPost-process high resolution GPS data and ensure privacy of participantsRetain enough detailed data to be useful to various research communities4Ensure participants cannot be identified4GPS-based Travel Study ExamplesLexington travel data demonstration 1997GPS for personal travel surveysGeorgia Tech comprehensive electronic travel monitoring system 1999-2000The Rtt-Fart Borlnge GPS study 1999-2001California Statewide Household Travel Survey GPS study, 2001Commute Atlanta Study, 2004-2006University of Minnesota, I-35 Bridge use study, 2008-2009Cobb County School District, GPS-based anti-idling study 2010 - ongoing55Travel Survey Methods using GPSActive handheld systems replicating traditional travel-diariesGT comprehensive electronic travel monitoring system studyLongitudinal studies with vehicle-based GPS systems installed in participant vehiclesCommute Atlanta StudyA hybrid of longitudinal passive GPS data collection coupled with intermittent online travel surveysUniversity of Minnesota, I-35 Bridge Crossings Study66Commute Atlanta StudyInstrumented vehicle research collecting high-resolution GPS (2004-2006)Assess the effects of converting operating costs into variable per-mile driving costs Approximately 500 vehicles in 270 householdsMore than 1.8 million vehicle tripsBaseline data from 270 householdsApproximately 100 households in the pricing study7Proposed Commute Atlanta Dataset for Public accessTravel Diary Data (trip-level travel data)95 households that had complete data for the study periodSummary of trips including origin/destination TAZs, distance, duration, date, time etc.Second-by-second data withheldOnroad Vehicle Activity Data (second-by-second)Approximately 175 households that did not have complete data for study periodTrip summary data (distance, duration, etc.) withheld88On Road Vehicle Activity DataCharacteristicsSecond by second vehicle speed and position dataData are tied to specific roadwaysFHWA highway functional classification, number of lanes, lane width, etc.UsesSafety studiesDriver behaviorEmissions analysis9AttributesGPS dataLatitude and Longitude, speed, heading, date and timeNumber of satellites, position quality information, etc.Roadway CharacteristicsHPMS attributes Georgia Tech Household Classification Group (income, vehicle ownership, and household size)Vehicle Characteristics Fuel Type, Engine Type, Body Type and Model Year GroupDriver CharacteristicsAge group and Gender

10Privacy ConcernsHigh resolution GPS data can identify participantsHome, and work locationsShopping, recreational and social preferencesDriver risk parametersHigh resolution GPS data + Vehicle Characteristics + Driver Characteristics, will yield the identity of individual participantsRecent news about public access to location dataIphone location trackingTomTom selling GPS data to Police11Driver Risk Parameters are Significant in defining crash risk groups over demographic variables11Identifying Home LocationsCharacteristicsMost frequent trip endUsually the last trip end of the dayMethodologyPool the data based on vehicle/driver characteristics, location, and date-time from vehicle activity datasetsIdentify frequent trip ends by time of daySpatially analyze the most frequent trip ends and the last trips in each day

12Data Filtering Techniques IFilter using buffer around the home locationCenter of the buffer is the home locationFilter using a polygon around the home locationPolygon centroid is the home location

13Insert a graphic in the blank space13Data Filtering Techniques IIFilter using a random polygon with its centroid away from homeDefine random centroid from householdRandom distance (minimum 500 ft and maximum 750 ft) from the household in a random directionGenerate a random six sided polygon Vertices at a minimum distance of 0.5 miles and maximum distance of 0.75 miles from the centroidAll GPS data that are within this polygon are trimmed

14Remind audience No trip distance, and duration information provided with this dataset14Case StudyOne month of activity data for 2 householdsOnroad vehicle activity data were filtered using a randomly generated polygon for each householdPost-analysis of the filtered dataset (detective work)Identify home location using filtered data with a trip ends algorithmSpatially identify clipped trip endpoints Find the centroid of these last known pointsDo a Network Analysis of clipped trip endpoints15Eliminate all records that fall within the random polygon around the household.Because a the randomized polygon employed to clip the data cannot be predictedCreate Service Area from each trip end point. Intersecting Service Area contains home location

15Household 1All GPS Data near Home

16Household 1On-Road Vehicle Activity Data

17Household 1Actual Filter Polygon

More than 1000 parcel centroids within filter Polygon18Household 1Home Location Estimate from Dataset

Estimated Home Location19

Household 1Home Location from Spatial AnalysisEstimated Home Location20Household 1Network Analysis137 parcel centroids within Intersect Area21

Household 2All GPS Data near Home

22Household 2On-Road Vehicle Data

23Household 2Actual Filter Polygon

More than 300 parcel centroids within filter Polygon24Household 2Home Location Estimate

Estimated Home Location25Household 2Home Location from Spatial Analysis

Estimated Home Location26Household 2Network Analyst22 parcel centroids within Intersect Area27

Access to Vehicle Registration DataImproves MatchingHousehold zip code = 30306Household has exactly 2 vehicles registeredBoth vehicles are SUVsOne vehicle is 1995-1999 model yearOne vehicle is 2000-2004 model year28Note that insurance company records also have vehicle registration and demographic dataSo do credit reporting bureaus28Query StatisticsRegistered Vehicles: 16,071SUVs: 3,718HH with 2 SUVs: 418HH with 2 SUVs and specified model year groups: 19

2919 Potential Household Locationsamong 16,000 residences

30Zip Code30306ConclusionsHigh-resolution GPS data need to be filtered before data are shared to prevent loss of privacyA filtering method that uses a random polygon around the home location was appliedHowever, using network and available on-road vehicle activity data, home locations can be identifiedUsing vehicle registration data and other data sources make it even easier to identify householdsLiability associated with participant privacy protection lies with data collector31Under Federal law31Ongoing ActivitiesContinue work to develop post-processing methods that ensure travel diary data and high resolution GPS data do not compromise participant privacyUnfiltered data remain available for research only at Georgia Tech through 2011Other researchers can visit Tech and do their research in partnership

32You can mention that a fellowship program is forthcoming32Questions?Vetri Venthan Elango vetri@gatech.edu

Randall Guensler randall.guensler@ce.gatech.edu33