Research On Helmet Wearing Detection Algorithm Based On Improved YOLOv4

According to data from the Ministry of Emergency Management of the Peoples’ s Republic of China,67.95% of construction accidents are caused by workers not wearing safety helmets.Traditional manual inspection methods no longer meet the needs of modern construction safety management,making it essential to develop an advanced safety helmet detection algorithm.This article takes the safety helmet wearing situation in various environments as the research object and improves the YOLOv4 object detection algorithm as the baseline,aiming to design a more accurate,smaller parameter,and more robust safety helmet detection algorithm.The specific work and algorithm innovation are as follows:For the current situation that the safety helmet wearing detection dataset is small and unrepresentative,the existing safety helmet wearing dataset is integrated and manually annotated with relevant images from other scenes for expansion.In addition,due to the large number of small objects in the safety helmet dataset,publicly available and representative small object datasets from other fields are collected and integrated and validated for algorithm generalizability,and guidance suggestions are provided.The proposed multi-scale feature fusion helmet wear detection algorithm based on YOLOv4: YOLO-P.For the detection of helmet wear with different target sizes and mainly small targets,the network structure is improved to adapt to small target detection and the detection head anchor assignment is optimized to fully extract the features in the main C2 layer,which can maximize the information of small and medium targets in the shallow layer.Then,for the complex background of the field and the serious phenomenon of mutual occlusion between targets,we propose a multi-scale feature fusion structure based on PANet: E-PAN module,which effectively filters out the complex background information through the residuals between close upsampling and the same scale,fully fuses the features and highlights the foreground information.It is verified that the AP50 and recall rate of YOLO-P in helmet wearing detection are higher than those of advanced algorithms such as YOLOv7.The lightweight algorithm YOLO-PL is proposed in response to the high number of parameters in the YOLO-P algorithm,which is difficult to deploy in practice.Finally,the Swish activation function is introduced into the feature fusion structure,and finally,the number of parameters is reduced by 36.8%,the computation is reduced by 33.1%,and the inference speed is improved compared with the YOLO-P algorithm,while maintaining the same accuracy of YOLO-P and YOLO-PL.In this paper,various experiments are designed for validation as well as visual analysis to demonstrate various performance improvements of YOLO-PL in the field of helmet detection as well as other small target detection areas.