Safety helmet detection method based on YOLO v4

Abstract

In the actual production process, safety accidents caused by workers not wearing safety helmets often occur. In order to reduce safety accidents caused by wearing helmets, a helmet detection method based on improved YOLO v4 is proposed. By collecting a self-made data set of on-site construction site video, using the K-means algorithm to cluster the data set in getting appropriately a priori frame dimensional center and obtaining more targeted edge information. Subsequently, a multi-scale training strategy is used in the network training process to improve the adaptability of the model from different scales of detection. The experimental results show that, in the helmet wearing detection task, the model mAP value reached 92.89%, the detection speed reached 15f/s, and its detection accuracy and detection speed were improved compared with YOLO v4, which satisfies the real-time requirements of the helmet detection task.