SiamSTA: Spatio-Temporal Attention Based Siamese Tracker for Tracking UAVs

Bo Huang, Junjie Chen, Tingfa Xu, Ying Wang, Shenwang Jiang, Yuncheng Wang, Lei Wang, Jianan Li; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 1204-1212

With the growing threat of unmanned aerial vehicle (UAV) intrusion, anti-UAV techniques are becoming increasingly demanding. Object tracking, especially in thermal infrared (TIR) videos, though provides a promising solution, struggles with challenges like small scale and fast movement that commonly occur in anti-UAV scenarios. To mitigate this, we propose a simple yet effective spatio-temporal attention based Siamese network, dubbed SiamSTA, to track UAV robustly by performing reliable local tracking and wide-range re-detection alternatively. Concretely, tracking is carried out by posing spatial and temporal constraints on generating candidate proposals within local neighborhoods, hence eliminating background distractors to better perceive small targets. Complementarily, in case of target lost from local regions due to fast movement, a three-stage re-detection mechanism is introduced to re-detect targets from a global view by exploiting valuable motion cues through a correlation filter based on change detection. Finally, a state-aware switching policy is adopted to adaptively integrate local tracking and global re-detection and take their complementary strengths for robust tracking. Extensive experiments on the 1st and 2nd anti-UAV datasets well demonstrate the superiority of SiamSTA over other competing counterparts. Notably, SiamSTA is the foundation of the 1st-place winning entry in the 2nd Anti-UAV Challenge.