An enhanced YOLOv5 model with multi-strategy improvements for detecting helmet-wearing on electric vehicle riders in complex scenes

In urban transportation, the weak safety awareness of electric vehicle riders and the frequent occurrence of accidents due to not wearing helmets make the research on automated detection of helmet-wearing status urgent. This paper builds an enhanced detection model based on YOLOv5s. Firstly, images are obtained through web crawlers, on-site shooting, and surveillance frame extraction. After removing invalid images, a dataset is built by manual annotation using LabelImg. Secondly, to address issues such as sample imbalance, difficulty in detecting small targets, and target occlusion, four improvement strategies are proposed: Mosaic9 data augmentation, adding a P2 feature detection layer, integrating the CBAM attention mechanism, and adopting DIoU-NMS for optimization. Finally, the optimal strategy combination is selected through ablation experiments and compared with other models. The results show that the improved model's average precision and recall are improved by 3.76% and 2.32% respectively compared to the baseline model, providing theoretical support for intelligent transportation management and contributing to the intelligent safety supervision of non-motorized vehicles.