Visible-Infrared Feature Fusion Network with Adaptive Attention for Robust Human Detection in Post-Disaster UAV Surveillance

Post-disaster person detection from aerial platforms faces significant challenges, including adverse illumination, extreme imbalance between foreground and background, and small target sizes. We proposed REM-YOLO, an enhanced YOLO11- based framework that integrates visible and thermal infrared imagery for all-weather emergency response. The architecture introduces four key innovations: early-stage multispectral fusion to capture complementary spectral information, C2PSA DEA modules that leverage multi-scale attention mechanisms for improved feature extraction, the PP-HGNetv2 backbone for high-quality representation with reduced computational cost, and Adaptive Threshold Focal Loss (ATFLoss) to effectively address class imbalance through dynamic reweighting. On the LLVIP dataset, REM-YOLO achieves 97.2% mAP@50 and 66.8% mAP@50:95, outperforming the baseline YOLO11 by 7.8% and 14.8%, while reducing computational complexity by 20% and parameters by 24%. These advances make REM-YOLO a robust, lightweight solution for UAV-based search-and-rescue operations, offering vital support for emergency post-disaster response systems