Xuemei Liu, James Biagioni, Jakob Eriksson, Yin Wang, George
Transcription
Xuemei Liu, James Biagioni, Jakob Eriksson, Yin Wang, George
Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches Xuemei Liu, James Biagioni, Jakob Eriksson, Yin Wang, George Forman, Yanmin Zhu Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches slide 2 Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches slide 3 Raw GPS traces slide 4 Inferred road map slide 5 Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches slide 6 slide 7interval Chicago shuttle data, 1 second Shanghai taxi data, 16/61 second interval slide 8 Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches slide 9 UIC campus shuttle traces slide 10 2 hours of Shanghai data slide 11 Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches slide 12 Mining Large-Scale, Sparse GPS Traces for Map Inference Comparison of Approaches slide 13 Existing approaches ‣ k-Means clustering - ‣ Kernel density estimation - ‣ Edelkamp & Schrödl (2003) Davies et al. (2006) Trace merging - Liu et al. (2012) slide 14 Why infer maps? slide 15 Road surveys slide 16 Rural/developing areas slide 17 New road construction slide 18 Road closures slide 19 Road closures slide 20 Opportunistic data collection slide 21 Existing approaches ‣ k-Means clustering - ‣ Kernel density estimation - ‣ Edelkamp & Schrödl (2003) Davies et al. (2006) Trace merging - Liu et al. (2012) slide 22 k-Means Clustering Edelkamp & Schrödl (2003) slide 23 Raw GPS traces slide 24 Drop seeds slide 25 Adjust seeds slide 26 Link seeds slide 27 Kernel Density Estimation Davies et al. (2006) slide 28 Raw GPS traces slide 29 2-D histogram slide 30 Trajectory density estimate slide 31 Thresholded image slide 32 Map extraction slide 33 Sparse GPS samples B A slide 34 Incorrect trajectory B A slide 35 Actual trajectory B X A slide 36 Current method slide 37 Current result +1 +1 +1 +1 +1 +1 slide 38 Proposed result +1 +1 slide 39 KDE variants ‣ KDE “lines” +1 +1 +1 +1 +1 +1 ‣ KDE “points” +1 +1 slide 40 Trace Merging Liu et al. (2012), “TC1” slide 41 Raw GPS traces slide 42 Segment selection slide 43 Clustering slide 44 Map extraction slide 45 Quantitative Evaluation slide 46 Ground truth map slide 47 Inferred map slide 48 Overlaid maps slide 49 Overlaid maps slide 50 True positive length 1m 1m 1m slide 51 1m 1m True positive length ≤ m? 1m 1m ≤ m? 1m ≤ m? ≤ m? 1m true positive length = # m slide 52 1m ≤ m? Evaluation metrics ||Inf erred|| = total inf erred road length (m) ||Ground T ruth|| = total ground truth road length (m) slide 53 Evaluation metrics tp = true positive length tp tp precision = , recall = ||Inf erred|| ||Ground T ruth|| precision · recall F -measure = 2 · precision + recall slide 54 Chicago Evaluation slide 55 Chicago raw GPS data slide 56 2-sec sampling interval slide 57 4-sec sampling interval slide 58 8-sec sampling interval slide 59 16-sec sampling interval slide 60 32-sec sampling interval slide 61 64-sec sampling interval slide 62 128-sec sampling interval slide 63 256-sec sampling interval slide 64 Overall comparison slide 65 K-means slide 66 KDE lines slide 67 KDE points slide 68 KDE points slide 69 TC1 slide 70 Shanghai Evaluation slide 71 Shanghai raw GPS data slide 72 Precision/recall vs. threshold slide 73 Precision/recall vs. data size slide 74 KDE lines slide 75 KDE points slide 76 TC1 slide 77 What have we learned? KDE points TC1 vs. slide 78 Tale of the tape KDE points vs. TC1 ✗ sparsity ✓ ✓ scalability ✗ ✓ intersections ✗ ✗ centerlines ✓ slide 79 Future work slide 80 Thanks! Questions?