This tracker operates in three stages: First, objects are detected in each frame independently using the DPM object detector by Ross Girshick and Pedro Felzenszwalb. Second, all detections with a positive score are associated to detections in the next frame using appearance and the bounding box overlap. We predict objects to the next frame using a Kalman filter and associate them globally via the Hungarian method for bipartite matching. To gap occlusions and missed detections, we also associate tracklets with each other in a first stage. Similarly to the second stage the Hungarian algorithm is employed but this time based on a occlusion sensitive appearance model and regression of the bounding boxes in one tracklet from the bounding boxes in the other tracklet. The algorithm outputs all associated tracklets with a lifetime longer than three frames.
The reported running time is dominated by the object detection stage.
A. Geiger, M. Lauer, C. Wojek, C. Stiller, R. Urtasun. 3D Traffic Scene Understanding from Movable Platforms. In Pattern Analysis and Machine Intelligence (PAMI), 2014.
March 30, 2016 (4 years ago)
March 31, 2016 at 07:34:42 CET
Project page / code:
3 GHz, 1 CPU
|MOT16||33.7||0.0||76.5||55 (7.2)||411 (54.2)||5,804||112,587||38.2||92.3||1.0||2,418 (63.2)||2,252 (58.9)|