Safety Helmet Wearing Detection System Based on a Two-Stage Network Model

In this study, we proposed a safety helmet-wearing detection system to identify whether workers in hazardous environments such as factories or construction sites wear safety helmets. To reduce the classification loss and maintain the availability of streaming helmet detection, we presented a two-stage network model containing YOLOv5-Small and ResNet-18. The network YOLO is used to detect and crop the regions of the images with heads; then, the cropped images will be sent to ResNet to classify. By taking advantage of the individual strengths of these two models, experimental results show that the detection system indeed provides higher precision results compared to the YOLOv5-Small-only system, especially for the cases of wearing hats but not helmets. Furthermore, we designed an additional alert mechanism for real-world applications to reduce the incidence rate of false alerts and increase the flexibility of wide deployment.