Safety helmet wearing recognition based on improved YOLOv4 algorithm

In safety helmet wearing detection, traditional methods have problems such as low detection accuracy due to small object size. In this paper, we propose an improved helmet wearing detection algorithm YOLOv4-P. Firstly, we use k-means clustering algorithm to readjust the prior bounding box parameters to improve the matching between the prior bounding box and the object, and secondly, we introduce the Pyramid Split Attention (PSA) model to further process the multi-scale feature information. The network structure of YOLOv4 is improved by adding a layer of network features to improve the detection accuracy through feature fusion. Experimental results show that the improved model YOLOv4-P has an average accuracy improvement of 2.15% over the YOLOv4 algorithm on the SHWD helmet target detection dataset, satisfying the accuracy of the helmet wearing detection task.