BEV View Multi-object Detection Algorithm Based on Improved YOLOv8

With the rapid development of the Chinese economy and automobile market, the road traffic environment has become increasingly complex, resulting in prominent traffic safety issues. Consequently, target detection tasks for urban road supervision have become a research focus. Traditional target detection methods, using the Front of View (FOV) perspective, suffer from long-distance target occlusion and limited field of view. To tackle this challenge, the paper introduces the YOLOv8-mAF algorithm, which is based on the YOLOv8m model, utilizing data collected from the Bird's Eye View (BEV) perspective by drones as data samples. This approach resolves occlusion issues at the planar scale and incorporates the Coordinate Attention mechanism and the C2f_Faster lightweight module to enhance detection capabilities for small-scale targets while streamlining the model architecture. Experiments on the VisDrone2019-DET public dataset demonstrate that YOLOv8-mAF achieves a recall rate of 42.4%, outperforming YOLOv8m by 0.2%; mAP@50 reaches 42.9%, a 0.9% improvement over YOLOv8m; mAP@50-95 reaches 25.8%, a 0.7% improvement over YOLOv8m. Additionally, the model is more lightweight, achieving a detection frame rate of 120FPS, meeting real-time detection requirements, and enhancing the accuracy and robustness of multi-object detection under BEV.