Tracking by Detection [TBD]

Submitted on 21 Jan. 2014 12:08 by
Andreas Geiger (MPI Tübingen)

Running time:10 s
Environment:1 core @ 2.5 Ghz (Matlab + C/C++)

Method Description:
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.
See implementation.
Latex Bibtex:
author = {Andreas Geiger and Martin Lauer and Christian Wojek and Christoph Stiller and Raquel Urtasun},
title = {3D Traffic Scene Understanding from Movable Platforms},
journal = {Pattern Analysis and Machine Intelligence (PAMI)},
year = {2014}
author = {Hongyi Zhang and Andreas Geiger and Raquel Urtasun},
title = {Understanding High-Level Semantics by Modeling Traffic Patterns},
booktitle = {International Conference on Computer Vision (ICCV)},
year = {2013}

Detailed Results

From all 29 test sequences, our benchmark computes the commonly used tracking metrics CLEARMOT, MT/PT/ML, identity switches, and fragmentations [1,2]. The tables below show all of these metrics.

CAR 55.07 % 78.35 % 55.16 % 84.15 %

Benchmark recall precision F1 TP FP FN FAR #objects #trajectories
CAR 56.72 % 99.30 % 72.20 % 20023 141 15281 1.27 % 21945 1049

CAR 20.46 % 46.92 % 32.62 % 31 529

This table as LaTeX

[1] K. Bernardin, R. Stiefelhagen: Evaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics. JIVP 2008.
[2] Y. Li, C. Huang, R. Nevatia: Learning to associate: HybridBoosted multi-target tracker for crowded scene. CVPR 2009.

eXTReMe Tracker