ForestSAR-YOLO: real-time detection of small human targets in UAV imagery

Human detection in UAV imagery is required for aerial search operations in natural environments where targets occupy only a small number of pixels and appear within heterogeneous backgrounds. This work presents ForestSAR-YOLO, a detection framework designed for real-time localization of humans in nadir UAV imagery captured over forest and mixed terrain. The architecture preserves high-resolution spatial features through a P2 feature level (stride 4), combines multiscale feature fusion, and applies channel attention to suppress vegetation-induced background responses. An anchor-free detection head with uncertainty-aware localization predicts bounding boxes and confidence scores in a single forward pass, enabling low-latency inference during continuous UAV monitoring. The model was trained on UAV imagery collected during search exercises conducted by the Emergency Situations Ministry of the Krasnoyarsk Krai. Targets typically occupy fewer than 20 pixels in height at flight altitudes of 50-120 m, which creates challenges for conventional detection models due to strong feature downsampling and background clutter. The proposed detector achieves precision of 0.98, recall of 0.97, mAP[Formula: see text] of 0.98, and mAP[Formula: see text] of 0.83 while maintaining an inference speed of 68 FPS. Comparison with several object detection architectures demonstrates higher detection accuracy for small human targets while preserving processing speed compatible with real-time aerial search operations.