BDC-YOLOv5: a helmet detection model employs improved YOLOv5

Automatic helmet-wearing detection is an effective way to prevent head injuries for construction site workers. However, the current helmet detection algorithms still need to improve, such as low accuracy of small target recognition and poor adaptability to complex scenes. This paper proposes possible improvements to YOLOv5 and calls it BiFPN Detection CBAM YOLOv5(BDC-YOLOv5). Regarding the problem of error detection, two modifications are offered. First, an additional detection layer is introduced on YOLOv5, and more detection heads are used to detect targets of different scales, thus improving the detection ability of the model in complex scenarios. Then, the Bidirectional Feature Pyramid Network (BiFPN) is introduced, and the shallow semantic features can be fused better by adding jump connections, which significantly reduces the detection error rate of the model. In terms of reducing the missed detection rate, the Convolutional Block Attention Module (CBAM) was added to the original YOLOv5, thus making the model more focused on all helpful information. Finally, the BiFPN, the additional detection layer, and the CBAM module are combined simultaneously in YOLOv5, which reduces the model’s false detection and missed detection rate while improving the detection ability of small-scale objects. Experimental results on the public dataset Safety-Helmet-Wearing-Dataset(SHWD) show a mean average precision (Map) improvement of 2.6%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document} compared to the original YOLOv5, which reflects the significant improvement in the target monitoring capability of the model. To demonstrate the performance of the proposed BDC-YOLOv5 model, a series of comparative experiments with other mainstream algorithms are carried out.