Improved YOLOv5-based Safety Wear Detection Algorithm for Railroad Construction Personnel

ZHU Chunyuan; ZHENG Guanghai; BAI Xuepeng

doi:10.3969/j.issn.1672-9730.2025.08.011

Abstract：Aiming at the problem that railroad workers do not wear helmets and reflective undershirts in the inspection process in a standardized way,which is prone to safety accidents,thsi paper proposes a safety wear detection algorithm for railroad construc-tion workers based on the improved YOLOv5s to solve the problem of target leakage and misdetection in the detection of small tar-gets,and to reduce the rate of accidents.Firstly,the super-large scale detection layer is eliminated on the basis of increasing the small scale detection layer,so as to enhance the detection capability of small targets and reduce the network complexity.Secondly,the weighted bidirectional feature pyramid network BiFPN is borrowed in the Neck part,and the PANet in the original model is par-tially improved,and an unweighted BiFPN structure is adopted to strengthen the feature fusion capability of the model,and then the SPPFCSPC to replace the SPPF module,which is fused with the returned feature layer of Backbone through multi-scale spatial pyra-mid pooling and convolution operations to improve the sensory field and feature expression ability of the model,and to strengthen the feature fusion ability for small targets.Finally,by replacing the decoupled detector head,the model convergence speed is accel-erated,and the detection accuracy is improved.Experiments show that the detection accuracy of this paper's algorithm is improved by 3.0％compared with the original algorithm,which reduces the number of omissions and false detections in small target detection.