Enhancing YOLOv5 for UAV target detection in complex aerial scenarios

With the increasing use of UAV technology in military target tracking, traffic monitoring, and disaster observation, the demand for effective target detection algorithms has grown. However, challenges like occlusion, overlap, and scale variation in aerial images hinder existing methods. To address these, this paper improves the YOLOv5 model by introducing an SPD-conv module to reduce fine-grained information loss and inefficient feature learning, and a joint downsampling module in the neck network to prevent small target feature loss during fusion. Experimental results show significant improvements in accuracy and stability, offering a reliable solution for UAV-based target detection in military and agricultural fields.