Occlusion detection in visual tracking: A new framework and a new benchmark

Abstract

Occlusion remains being a challenge in visual object tracking. The robustness to occlusion is critical for tracking algorithms, though not much attention has been paid to it. In this paper, we first propose an occlusion detection framework which calculates the proportion of the target that is occluded, hence to decide whether to update the model of target. This framework can be integrated with existing tracking algorithms to increase their robustness to occlusion. Then we introduce a new benchmark which contains sequences where occlusion is the main difficulty. The sequences are chosen from public benchmarks and are fully annotated. The proposed framework is combined with several standard trackers and evaluated on the new benchmark. The experimental results show that our framework can improve the tracking performance, with explicit incorporation of occlusion detection.