Safety Helmet Detection Based on Optimized YOLOv5

Whether employees wear safety helmets is an important safety issue in power related work scenarios, and various safety issues can be avoided by monitoring this situation. However, traditional target detection methods are vulnerable to interference due to the weather, light, personnel density, location of surveillance cameras and other problems in the working environment, and the recognition and detection effect of such small targets is not very good. Therefore, this paper uses the high-precision YOLOv5 (You Only Look Once) as the target detection framework, and modifies its backbone network to improve its ability in small target recognition. The original backbone structure is cut and compressed, and the SwinT (Swin Transformer) modules are added to improve the overall recognition accuracy based on its powerful small target recognition ability. At the same time, SE (Squeeze and Excitation) and CBAM (Convolutional Block Attention Module) modules are added to further improve the recognition accuracy of the entire network. Finally, experiments are conducted on the SHWD (Safety Helmet Wearing Dataset) dataset. The experimental results show that compared to the network before modification, the accuracy of the optimized YOLO structure proposed in this paper is significantly improved on the validation dataset, with an average recognition accuracy of 93%.