Anti-UAV Net for Target Detection of Small Unmanned Aerial Vehicles

With the widespread adoption of unmanned aerial vehicles (UAVs), safety concerns have increasingly emerged. To mitigate these security issues, the accurate and rapid detection and identification of UAV targets—particularly smaller UAVs in complex backgrounds—has become an urgent challenge that requires immediate attention. Traditional object detection algorithms often struggle with accuracy and robustness due to various factors, including the small size of targets, complex backgrounds, and varying lighting conditions. To address these challenges, we propose a dual-modal small UAV detection network, termed Anti-UAV Net (AU-net), which utilizes adaptive fusion of infrared and visible images. Our model is built on the classic single-stage object detection framework YOLOv8n to satisfy real-time performance demands. We have designed an infrared-visible dual-modal feature extraction backbone and integrated the Simple Attention Module (SimAM) to facilitate adaptive weighted fusion of dual-modal feature maps. Additionally, we developed a small object detection layer specifically tailored for small UAV targets, enhancing the resolution of the final feature maps and improving the network’s effectiveness in detecting small UAVs. Experimental results show that the AU-net algorithm significantly improves the mean Average Precision (mAP) from 76.70% to 97.68% on the self-constructed Anti-UAV Dual-Modal Dataset (AU-dataset) compared to single-modal detection methods. The detection speed achieves 76 FPS, meeting the requirements for real-time processing.