Refined YOLOv5s for UAV-based real-time object detection with improved accuracy

In response to the challenges of detecting small targets and the limitations of computational resources in real-time target detection using drones, this paper proposes an optimized lightweight YOLOv5s method. This approach introduces BiFPN for bidirectional feature fusion and designs an MS-VoV-GSCSP module to enhance multi-scale feature extraction capabilities. To improve the detection effect on small targets, a P2 detection layer is added, and a Slim-Neck structure is adopted to reduce the number of model parameters and lower computational complexity. Experimental results show that this method increases mAP@50 by 5.2% and mAP@50:95 by 3.0% on the VisDrone2019 dataset, reduces the number of parameters from 7.05M to 5.36M, and boosts inference speed by 19 FPS. This method achieves a better balance between accuracy and lightweight design, making it suitable for real-time target detection tasks with drones.