29 March 2025 in Knowledge-Based Systems
21 March 2025 in IEEE Signal Processing Letters
Visual object tracking utilizing adaptive computation presents challenges stemming from the complexities of modeling intricate motion patterns and achieving computational efficiency. While recent transformer-based trackers have shown promising results, they struggle to effectively capture varying motion dynamics and often waste computation on less informative regions, leading to degraded performance under fast motion and occlusion. In this letter, we present AdaMoT, an innovative motion-aware transformer framework featuring three lightweight modules that integrate adaptive attention and motion estimation: a Lightweight Adaptive Motion Estimation (LAME) module that guides transformer attention through motion pattern modeling, a Saliency-based Hard Attention Sampling (SHAS) module that reduces computation by 60% through focusing on motioncritical regions, and an Adaptive ViT Attention Head Adjustment (AVAHA) module that dynamically allocates attention heads based on motion complexity. Our framework uniquely integrates motion estimation with transformer attention through a shared feature space, achieving robust tracking with minimal overhead. Comprehensive testing indicate that AdaMoT attains superior performance on various demanding benchmarks (75.1% AO on GOT-10k, 84.9% AUC on TrackingNet, 72.9% AUC on LaSOT) while maintaining real-time speed (32.1 FPS) with only 4% FLOPs increase.
4 April 2025 in Neurocomputing
21 January 2026 in Knowledge-Based Systems
11 February 2026 in SSRN Electronic Journal
Robust visual tracking under severe occlusions remains challenging due to the lack of explicit occlusion detection mechanisms and adaptive feature compensation strategies in existing methods. To address this problem, we propose EntropyTrack, a unified transformer-based framework that exploits attention entropy for occlusion-aware tracking. Our approach introduces an Attention Entropy Occlusion Detector that quantifies attention distribution uncertainty to explicitly identify occlusion states with spatial consistency regularization. Building upon this detection, an Adaptive Feature Compensation module performs intelligent feature repair through dynamic fusion of spatial inpainting and temporal information based on detected occlusion patterns. Furthermore, a Multi-Scale Contrastive Learning module enhances feature discriminability via attention-guided sampling across multiple scales for robust target-background separation. Extensive experiments on seven challenging benchmarks including LaSOT, GOT-10k, TrackingNet, LaSOText, VOT2020, TNL2K, and OTB100 demonstrate that EntropyTrack achieves state-of-the-art performance while maintaining real-time efficiency at 42 fps. The synergistic integration of entropy-driven occlusion detection, adaptive compensation, and multi-scale contrastive learning provides a principled solution for robust visual tracking under challenging conditions.
28 August 2025 in Neurocomputing