Robust Infrared UAV Detection Method Based on Spatiotemporal Transformer in Thermal Crossover Dynamic Scenarios

Infrared Unmanned Aerial Vehicle (UAV) detection remains challenging due to thermal crossover interference, small target characteristics, and dynamic background clutter. Existing methods, such as Siamese trackers and single-frame detectors, often struggle to balance spatial discriminability and temporal continuity, leading to degraded robustness in complex environments. To address these issues, we propose IRST-DETR, an end-to-end video detection framework that integrates spatiotemporal features for robust UAV detection. IRST-DETR combines a spatial branch with a detail-enriched pyramid module (DEPM) for fine-grained feature extraction and a temporal branch using Video Swin Transformer to model motion dynamics. Hierarchical fusion modules, including the Temporal Gate Fusion Module (TGFM) and the Spatial-Temporal Attention Fusion Module (STAFM), adaptively merge spatiotemporal information, enhancing robustness against thermal noise. To rigorously evaluate performance in challenging conditions, we construct a specialized evaluation subset from Anti-UAV410, focusing on cluttered infrared scenes with dynamic thermal interference. Extensive experiments on Anti-UAV300, Anti-UAV410, and the challenging subset demonstrate that IRST-DETR achieves state-of-the-art performance. It outperforms Siamese trackers by 2.8% in state accuracy and surpasses single-frame detectors by over 9% in complex scenarios, with ablation studies confirming its efficiency. This work advances practical infrared UAV detection in real-world environments. Code and models are available at: https://github.com/henryHu-manboo/irst_detr.