DynaTrack-Air: Dynamic Template Fusion with Retrospective Consistency Verification for Robust UAV Tracking

Transformer-based object tracking algorithms have achieved remarkable breakthroughs. However, under the dynamic aerial-view scenarios typical of unmanned aerial vehicles (UAVs), these algorithms must strike a balance between tracking accuracy and computational efficiency. Studies show that conventional Large Models are difficult to deploy on resource-constrained UAV platforms due to their high computational complexity, while existing lightweight solutions still suffer from two major technical bottlenecks: feature drift in dynamic scenes and long-term error accumulation. To address these challenges, this paper introduces DynaTrack-Air, a novel framework with two key innovations: (1) a periodic retrospective verification module that enhances temporal and spatial consistency to effectively suppress error accumulation, and (2) a lightweight dynamic fusion module designed for efficient and adaptive template updating. The proposed approach is plug-and-play and can be seamlessly integrated into various Transformer-based trackers. Experimental results demonstrate that DynaTrack-Air achieves state-of-the-art performance on multiple authoritative UAV tracking benchmarks, with a precision of 83.7% and a success rate of 63.2%. It exhibits robust performance in challenging scenarios such as complete occlusion and cluttered backgrounds. Moreover, the algorithm maintains a real-time inference speed of 114 FPS, offering an efficient and reliable solution for real-time UAV visual tracking.