DLST: Dual-Template Co-Evolution Learning for Robust Long-Term Drone Tracking in Dynamic Environments

Jiahao Zhang, Yixin Wei, Jinli Zhang, Zongli Jiang, Peiwen Yu, Yufei Ma, Runan Jin; Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, 2025, pp. 6641-6649

To address the challenges of dynamic appearance variations, background interference, and long-term stability in drone small target tracking under complex environments, we propose a Dual Branch Learnable Siamese Tracker (DLST). The framework synergizes static and dynamically updated templates through three core modules: the Dual-Template Region Proposal Network (DT-RPN) for adaptive proposal generation via cross-correlation feature fusion, the Dual-Template R-CNN (DT-RCNN) for localization refinement guided by bidirectional template interactions, and the Dual-Template Background Suppression (DT-BS) module for distractor filtering through similarity scoring. By enabling bidirectional information flow between static and dynamic branches, DLST achieves robust adaptation to rapid appearance changes while maintaining computational efficiency. Extensive experiments on the Anti-UAV410 dataset demonstrate state-of-the-art performance, with a 68.81% State Accuracy (SA) surpassing existing methods. The framework exhibits exceptional robustness in scenarios involving dynamic backgrounds, fast motion, and occlusions, validated through comprehensive evaluations.