Helmet-YOLO: A new method for real-time, high-precision helmet wearing detection

In response to the problems of low target recognition accuracy and high latency of existing helmet wearing detection algorithms under conditions of target occlusion, small targets, and low light levels, this paper optimizes the VanillaNet-6 network and combines it with the YOLOv8s general object detection algorithm to propose a Helmet-YOLO helmet wearing detection algorithm. Firstly, a small-scale GhostConv is used at the input layer of the VanillaNet-6 network to extract more detailed information from the input images at low cost. Secondly, the VanillaODBlock is proposed to enhance the network model’s focus on the current task target. Next, the spatial feature pyramid pooling structure is used to improve model computational efficiency. Lastly, the LSKBlock is added in the feature fusion network to enhance relevant background information that can serve as judgment basis in the features, thereby improving the model’s robustness. The EIoU bounding box loss function is used to optimize the detection effect of occluded targets and accelerate model convergence. Experimental results show that the improved algorithm achieves a 96.7% mAP@0.5 with a reduction in computational volume by 32.4%, model size by 33.3%, and latency by 0.2ms. This is a 0.6% improvement over the original algorithm in terms of accuracy and robustness.To help readers reproduce this work, the core code has been uploaded to https://github.com/pzy-UX/Helmet-YOLO.