Re-detection and distractor association from a global perspective: A long-term tracking system

Great success has been achieved in short-term visual object tracking in recent years. However, long-term tracking which is closer to practical applications still lacks enough attention, especially for the target disappearances and reappearances which are common phenomena and important properties in long-term tracking due to full occlusion and out-of-view. In this paper, we propose a long-term tracking framework based on the cooperation of a short-term local tracker and the high-quality global re-detection and distractor association mechanisms. The former mechanism provides target candidates for target relocation, and the latter utilizes the Kalman Filter and Hungarian algorithm with motion cues to build distractor tracklets which attempts to avoid incorrect target relocation after the target disappears. We apply the tracking framework on the state-of-the-art algorithm STARK and evaluate the performance on the popular long-term benchmarks VOTLT-2020 and TLP. The experimental results show the effectiveness of the proposed framework and our tracker performs well compared with other competing algorithms.

Automated summary

Great success has been achieved in short-term visual object tracking, but long-term tracking is still lacking attention, especially for target disappearances and reappearances. This paper proposes a long-term tracking framework that combines a short-term local tracker and a high-quality global re-detection and distractor association mechanisms. The framework effectively handles target relocation and avoids incorrect target relocation after disappearance. Experimental results on benchmark datasets show that the proposed framework performs well compared to other competing algorithms.