Enhanced Association With Supervoxels in Multiple Hypothesis Tracking

Abstract

Remarkable progress has been made in the field of multi-object tracking. Although tracking-by-detection has recently became one of the most popular frameworks, it still has one main drawback: this approach relies heavily on the quality of detection. Thus, the missing detections caused by partial occlusion usually lead to fragment problem. To address this problem, this paper introduces supervoxels to represent objects with partial occlusion, even for missing detections. We first extract superpixels of the foreground, and then our proposed supervoxel consists of spatial-temporal sequences of superpixels. The supervoxels represent tracklets at the image level, so it is robust for initial detection. Then, we incorporate supervoxels into multiple hypotheses tracking by considering the enhanced association with supervoxels (EAS). Moreover, we propose a detection refinement method based on EAS. As our approach allows us to handle partial occlusion problems, we achieve remarkable results in crowded scenes. Finally, our experiments on both MOT15 and MOT16 benchmarks show that our EAS is competitive with the state-of-the-art trackers.