Unified Local and Global Transformer for Infrared Small UAV Tracking

Long-term tracking of infrared small unmanned aerial vehicle (UAV) poses substantial challenges, including dynamic backgrounds, clutter interference, and target occlusion. This study introduces an innovative method that integrates a local tracking network with a global search strategy to effectively address these issues. Firstly, a background motion estimation module is proposed to mitigate dynamic background interference by aligning consecutive frames through motion state assessment. Secondly, a full-transformer local tracking network is developed to suppress background clutter. It enhances feature representation using Spectformer as the backbone and leverages cross-attention mechanisms to robustly handle clutter. Finally, a global search strategy featuring a large-scale search module is designed to address target occlusion. This module provides reliable local search regions for the tracking network when occlusion occurs. Extensive experiments on infrared drone datasets validate that the proposed method outperforms state-of-the-art approaches, achieving high success rates, high precision, and high real-time processing at 45 FPS under specific configurations.