AO-AutoTrack: Anti-occlusion Real-Time UAV Tracking Based on Spatio-temporal Context

Hongyu Chu, Kuisheng Liao, Yanhua Shao, Xiaoqiang Zhang, Yanying Mei, Yadong Wu

Abstract: Correlation filter (CF)-based methods have shown extraordinary performance in visual object tracking for unmanned aerial vehicle (UAV) applications, but target occlusion and loss is still an urgent problem to be solved. Aiming at this problem, most of the existing discriminant correlation filter (DCF)-based trackers try to combine with object detection, thereby significantly improving the anti-occlusion performance of the visual object tracking algorithm. However, the significantly increases the computational complexity and limits the speed of the algorithm. This work proposes a novel anti-occlusion target tracking algorithm using the target’s historical information, i.e., the AO-AutoTrack tracker. Specifically, after the target is occluded or lost, the target information contained in the spatio-temporal context is used to predict the subsequent position of the target, and the historical filter is employed to re-detect the target. Extensive experiments on two UAV benchmarks have proven the superiority of our method and its excellent anti-occlusion performance.