![10FrameworkPost-ProcessingSegment[i].P(Bike) = Segment[i].P(Bike) * P(Bike|Car)Segment[i].P(Walk) = Segment[i].P(Walk) * P(Walk|Car)](https://image.slidesharecdn.com/learningtransportationmodefromrawgpsdataforgeographicapplicationsontheweb-100318230506-phpapp01/85/Learning-Transportation-Mode-From-Raw-Gps-Data-For-Geographic-Applications-On-The-Web-10-320.jpg)
Learning Transportation Mode From Raw Gps Data For Geographic Applications On The Web
- 1. 1Learning Transportation Mode from Raw GPS Data for Geographic Applications on the WebYu Zheng, Like Liu, Xing Xie, WWW’08Microsoft Research AsiaAdvisor: Chia-Hui ChangPresenter: Teng-Kai FanDate: 2010-03-19
- 3. 3BackgroundPercentage of GPS-enabled handset among mobile phone(Gartner Dataqueste: Forecast: GPS-enabled device 2004-2011)
- 4. 4IntroductionWhat we do: Infer transportation modes from users’ GPS logsGPS logInfer model
- 5. 5IntroductionMotivationDifferentiate GPS trajectory of different transportation modesLearning knowledge from raw GPS dataenable people to absorb more knowledge from others’ life experienceTrigger people’s memory about their pastUnderstand people’s life patternUnderstanding user behaviorContext-aware computingModeling traffic conditionDiscover social pattern…DifficultyA trajectory may contain more than two kinds of transportation modesPure velocity-based method may suffer from congestion
- 6. 6IntroductionDistribution of mean velocity (m/s) of different transportation modesDistribution of maximum velocity (m/s) of different transportation modes
- 7. 7IntroductionContributionsWe proposeA change point-based segmentation methodAn inference model based on supervised learningA post-processing algorithm based on conditional probabilitySignificanceA step toward mining knowledge from raw GPS data for geographic applications on the WebA step toward understanding user behavior based on GPS dataEvaluation resultsLarge-scale data collected by 45 people over a period of 6 monthsAlmost 70 percent accuracy
- 8. 8FrameworkPreliminarya place where people changetheir transportation modes
- 9. 9FrameworkInference strategydivide the GPS track into trips and then partition each trip into segments by change points
- 10. 10FrameworkPost-ProcessingSegment[i].P(Bike) = Segment[i].P(Bike) * P(Bike|Car)Segment[i].P(Walk) = Segment[i].P(Walk) * P(Walk|Car)
- 12. 12MethodologyCommonsense knowledge from real worldTypically, people need to walk before transferring transportation modesTypically, people need to stop and then go when transferring modesWalk should be a transition between different transportation modesTransition matrix of transportation modes
- 13. 13MethodologyChange point-based Segmentation AlgorithmStep 1: using a loose upper bound of velocity (Vt) and acceleration (at) to distinguish all possible Walk Points, non-Walk Points. Step 2: merge short segment (the length less than a thredshold) composed by consecutive Walk Points or non-Walk pointsStep 3: merge consecutive Uncertain Segment (less than 50 meters) to non-Walk Segment.Step 4: end point of each Walk Segment are potential change points
- 16. length
- 17. mean velocity
- 22. 16ExperimentEvaluation methodPrecision of inference a segment Accuracy by Length Accuracy by DurationChange PointPrecision of change pointRecall of change pointN: the total number of the segments after beingpartitioned by a segmentation method.m: # of segments our approach correctly predicted
- 23. 17Experiment: ResultInference performance Inferring accuracy of transportation mode over change point-based segmentation method
- 24. 18ExperimentInference performance of change pointRecall of change point using change point based segmentation methodPrecision of change point using change point based segmentation method
- 25. 19Experiment: ResultComparison of different segmentation methods using Decision Tree
- 26. 20Experiment: ResultComparison of inference results of CRF over different segmentation methods
- 27. 21ConclusionSegmentation methodInference methodSVMChange Point basedBayesian NetUniform Duration basedDecision TreeUniform Length basedCRF
- 28. 22Future workIdentify more valuable featuresLocation-constraint conditional probabilityImproving prediction performance of CRF-based approach