Collaborative Spatiotemporal Anchors and Key Feature Enhancement for UAV Tracking

Target tracking in unmanned aerial vehicle (UAV) platforms faces significant challenges due to high maneuverability-induced appearance variations, including abrupt scale changes, viewpoint shifts, and non-rigid deformations. Existing trackers suffer from two critical limitations: (1) fixed-interval sampling strategies fail to capture nonlinear state transitions, and (2) error accumulation in dynamic reference updating degrades long-term robustness. To address these issues, we propose a novel dynamic spatio-temporal perception framework with error suppression. Our approach introduces two core innovations: First, the Spatio-Temporal Anchors Module employs the Wasserstein-1 distance to quantify feature distribution evolution, enabling geodesic-equidistant sampling of representative reference frames that uniformly cover target transition trajectories. This distribution-aware mechanism adaptively balances update density during stable phases and critical transitions. Second, the Key Feature Enhancement Module conducts attention-driven fusion of candidate regions and historical references, dynamically propagating spatially salient features—identified through attention response analysis—via a cascaded architecture to mitigate error accumulation. Extensive evaluations on UAV123, UAVTrack112L, DTB70, and LaSOT benchmarks demonstrate state-of-the-art performance, with notable improvements in occlusion scenarios and deformation resistance, confirming its practical viability for aerial observation systems.