MACTrack: Spatiotemporal context propagation with motion compensation for anti-UAV tracking.

Abstract

In response to the security threats posed by unauthorized unmanned aerial vehicle (UAV) activities, robust visual tracking has become a critical component in anti-UAV systems. However, existing trackers often suffer from severe performance degradation due to agile target motion, complex dynamic backgrounds, and nonlinear motion coupling between the tracking platform and UAVs, leading to frequent tracking drift. To address these challenges, we propose MACTrack, a robust anti-UAV tracking framework that jointly integrates motion awareness with spatio-temporal context modeling. Specifically, an asymmetric spatial context module enhances semantic discrimination between the template and search region through complementary self-attention and cross-attention mechanisms, enabling accurate target localization in complex scenes. Meanwhile, a temporal context propagation module adaptively aggregates historical information via memory-guided attention, effectively alleviating performance instability caused by short-term occlusion and appearance variations. In addition, a motion compensation module explicitly aligns feature representations to mitigate tracking failures induced by platform motion. Extensive experiments on the challenging ANTI-UAV410 benchmark demonstrate that MACTrack achieves an accuracy of 88.9% and a success rate of 66.5%, outperforming a competitive baseline by 7.9% and 7.4%, respectively. These results validate the effectiveness of the proposed framework in addressing performance degradation in anti-UAV tracking scenarios.