Low-altitude small aerial vehicle detection based on cross-domain feature enhancement

Abstract

Small unmanned aerial vehicles (UAVs) pose increasing challenges to low-altitude security because they are often small, slow, and easily confused with complex backgrounds. This paper proposes a visible-infrared mixed-modal anti-UAV detection method based on YOLO26n. Since visible and infrared images in Anti-UAV300 are not strictly aligned in time and space, image-level fusion is avoided; instead, both modalities are converted into a unified YOLO detection format and used for mixed-modal training. To improve the localization of distant small UAVs, a P2 high-resolution detection head is added to the original multi-scale detection structure. In addition, an efficient channel attention (ECA) module is inserted before the multi-scale detection heads to enhance target-related channels and suppress background interference. Experiments on Anti-UAV300 show that the proposed YOLO26n + P2 + ECA model achieves 99.049% mAP@0.5 and 55.807% mAP@0.5:0.95, improving the baseline by 1.309 and 2.751 percentage points, respectively. The recall increases from 94.659% to 97.179%, indicating that the proposed design reduces missed detections while improving small-object localization in challenging low-altitude scenes.