Research On Lightweight Algorithm Of Helmet Wearing Detection Based On Deep Learning

There are many dangerous factors in the construction site of the construction industry,which is a place where accidents occur frequently.Wearing safety helmets in the construction site can effectively reduce casualties and ensure safety of production.For the task of helmet wearing detection,human supervision is inefficient and difficult to achieve real-time control,so it is of great practical significance to use deep learning technology to realize automatic detection of helmet wearing.Aiming at the shortcomings of the existing helmet wearing detection methods,this paper proposes an improved lightweight algorithm to achieve a better trade-off between accuracy and speed.The research contents are as follows:(1)YOLOv4-tiny with good real-time performance is used as the benchmark algorithm for the helmet wearing detection task in this paper,and it is improved according to the network structure and the target distribution characteristics of the dataset.For the problem of many missed detection of small objects,the small object detection scale is added to enhance the detection performance of the algorithm for small objects.For the problem that the structure of the feature fusion part is simple which leads to the insufficient interaction between feature maps of different levels,an improved lightweight feature fusion module is proposed.The lightweight feature fusion module strengthens the information exchange of different levels of feature maps.For the problem of missed detection in dense occlusion scenes,a multi-dimensional attention is proposed,which makes full use of information of different dimensions to make the algorithm focus on the key areas of the image.In view of the conflict between classification and regression tasks,the prediction head part of the network is decoupled to alleviate the incoordination between different tasks.The experimental results show that the m AP of improved algorithm achieved92.82%,and FPS is 63 on the helmet dataset,which balances the accuracy and speed performance well,and the model parameters are only 4.79 M,which is more suitable for terminal devices and edge devices with limited resources.(2)For the task of helmet wearing detection,an improved lightweight detection algorithm based on YOLOv7-tiny is proposed,which improves the detection accuracy of the algorithm while ensuring the real-time performance of the algorithm.Firstly,the explicit visual center module is used to enhance the feature pyramid of YOLOv7-tiny to improve the algorithm’s ability to learn global information of the image and capture key areas;Secondly,in terms of loss function,Focal-EIOU is used to replace the original CIOU loss to avoid the situation that the width and height information cannot be used to guide network parameter optimization when the predicted result is consistent with the real box aspect ratio.The effectiveness of the proposed method is verified by experiments on the helmet dataset and the VOC2007 dataset.The m AP of improved algorithm achieved 94.09%,and FPS is 62 on the helmet dataset,and the model parameters are 7.07 M.The accuracy and real-time performance have reached the advanced level.At the same time,the m AP on the VOC2007 dataset has also achieved a 0.89% improvement.