To address the problem of low detection accuracy and poor robustness of the traditional helmet wearing recognition algorithm, we propose a deep learning-based helmet wearing detection method. The method is based on the YOLOv3 detection algorithm, and its network structure and loss function are improved. Firstly, the poor detection of small targets by the original YOLOv3 algorithm is compensated by adding feature maps. Then, on the basis of adding feature maps, the K-means clustering algorithm is used to cluster the collected helmet datasets and select the appropriate a priori anchor frames. Finally, GIoU Loss is adopted as the boundary frame loss, and Focal Loss is added to the loss function to reduce the errors due to positive and negative sample imbalance. Experimental results show that, compared with the YOLOv3 detection algorithm, the improved algorithm improves the average accuracy by 3.47% and the accuracy of helmet identification by 4.23%, which is advanced and effective in helmet identification.