Helmet detection models based on deep learning

Deep learning-based helmet detection is a common aspect of industrial safety production. The existing hard hat data set covers a large number of open-air scenes, and lacks low-light scenes such as underground mines and tunnel projects. Therefore, it cannot meet the detection requirements under low light. To develop a dataset better suited for this purpose, we first collected multiple images in harsh light to supplement the data. Then, we analyzed and compared the network structure and model features of the current phase-based helmet detectors, including the YOLO series, SSD, RetinaNet, CenterNet, and EfficientDet. Experimental results show that the YOLO series algorithm is the most accurate model, and its detection model YOLOv7 also shows the best performance, accuracy score is 94.57%, a recall rate is 94.83%, and mean average precision (mAP) is 97.7%. The model has good applicability and robustness for helmet detection in different scenarios.