YOLO-Helmet: A Novel Algorithm for Detecting Dense Small Safety Helmets in Construction Scenes

Safety helmet wearing is an effective measure for reducing construction safety accidents. However, the current algorithms for detecting helmet-wearing face several challenges, including high missed detection rates and low accuracy in detecting dense small safety helmets. Therefore, this paper proposes a novel algorithm called YOLO-Helmet. Firstly, in order to solve the problem of difficult detection due to the small area of the helmet in the image, a small size detection layer was extended to improve the detection sensitivity of the network to small size targets. Secondly, in order to reduce the influence of occlusion on the accuracy of helmet detection, the C-ELAN module was constructed, and the receptive field is expanded by deformable convolution to provide rich contextual feature information for coordinate attention, so as to improve the accuracy of the network for the discrimination of target position information. Thirdly, CIoU was combined with NWD to reduce the sensitivity of position deviation while retaining the excellent classification ability of CIoU. Finally, in order to facilitate the model deployment, the VoV-DG module based on GSConv was constructed in the neck. The experimental results show that the YOLO-Helmet algorithm achieved an average detection accuracy of 93.1% on the SHWD dataset and is more suitable for the identification of dense small helmets in construction scenes than other mainstream algorithms.