Abstract: For the anti-UAV task under complex backgrounds, infrared target tracking faces numerous challenges, such as multi-scale targets, occlusion, moving out of view, and background interference. To address these problems, a robust infrared UAV tracking algorithm is proposed. First, a global tracking model is designed, which adopts an efficient one-stage anchor-free framework to perform global search for handling target disappearance caused by occlusion or moving out of view, while employing a multi-level structure to track targets of varying sizes. Second, a spatio-temporal feature fusion module based on the memory network is proposed, which can enhance target discriminability by leveraging spatio-temporal features from video sequences. Then, a target enhancement and interference suppression module is introduced, which records and matches target and interference features between frames to generate weighting maps. These maps are applied to the score maps to improve the algorithm’s anti-interference capability. Finally, a dynamic hierarchical acceleration method is proposed to improve the running efficiency by removing redundant hierarchical computations. Experimental results demonstrate that the proposed algorithm achieves 92.4% and 78.7% precision and 69.4% and 56.5% success rates on the 2nd and 3rd Anti-UAV datasets, respectively, with a running speed of 26.9 fps, which significantly outperforms existing methods and achieves real-time and robust UAV tracking in complex scenarios.

Key words: infrared target tracking, anti-UAV, complex background, anti-interference, real-time tracking, deep learning