EnTeR-Track: Efficient UAV tracking via entropy-guided pruning and reversible token recovery

Vision Transformers have achieved strong performance in visual tracking, but their quadratic computational complexity limits real-time deployment on resource-constrained Unmanned Aerial Vehicles (UAVs). Token pruning offers a promising solution, yet existing methods face three key challenges: unreliable token importance estimation in early layers due to weak semantic representations, irreversible information loss caused by hard pruning operations, and static pruning policies that cannot adapt to varying scene complexity. To address these issues, we propose EnTeR-Track, an entropy-guided token management framework designed for efficient UAV tracking. Our method follows a closed-loop prune-compensate-adapt paradigm. Specifically, we introduce attention entropy as an uncertainty-aware saliency measure to enable reliable early-layer token selection. To alleviate irreversible information loss, we design a lightweight reversible estimator that approximates cross-layer feature evolution for pruned tokens. Furthermore, an adaptive threshold predictor dynamically adjusts token sparsity according to scene complexity, enabling flexible computation allocation. Extensive experiments on UAV benchmarks show that EnTeR-Track achieves a strong balance between accuracy and efficiency, reaching 67.5% AUC on UAV123 while running at 257 FPS on GPU and 119 FPS on CPU. Our code will be publicly available at https://github.com/zha-z/EnTeR-Track/.