Research On Construction Site Safety Helmet-Wearing Detection Algorithm Based On Deep Learning

In the field safety management,it is very important to supervise the safety helmetwearing of construction personnel.In recent years,some scholars have done some research on the detection and recognition of safety helmet,but they mainly use traditional target detection methods,which have low detection accuracy and slow speed.With the development of deep learning technology,this technology has been widely used in target detection tasks,and achieved good results.Therefore,based on deep learning technology,this thesis carries out research on safety helmet-wearing detection,realizes rapid and accurate safety helmet-wearing detection for construction site workers,provides technical support for site safety supervision and law enforcement,and has significance for safety management of construction site.The work and achievements of this thesis are as follows:(1)A data enhancement method based on image pyramid is proposed,and a set of safety helmet-wearing detection dataset is constructed.Aiming at the problem of no publicly available safety helmet-wearing detection dataset,this thesis constructs a set of safety helmet-wearing detection dataset based on site monitoring video data.In the process of dataset construction,a data enhancement method based on image pyramid is proposed.This method can increase the number of effective targets in the safety helmet-wearing detection dataset,and solve the problem that the number of effective targets in the safety helmet-wearing detection dataset is too small.Experiments show that the proposed data enhancement method can improve the detection effect of the model to a certain extent.(2)Based on the target clustering method,the generation scheme of target recommendation box in the process of safety helmet-wearing detection is optimized.Firstly,the size distribution of the image target in the safety helmet-wearing detection dataset is analyzed.It is found that the size of the target recommendation box used in YOLO v3 is not suitable for the safety helmet-wearing detection scene in this thesis.Then,aiming at the safety helmet-wearing detection dataset,the target recommendation box clustering method based on target clustering is studied,and the target recommendation box in the safety helmet-wearing detection process is optimized.Experiments show that the detection accuracy of the model can be improved by using the target recommendation box generated by the optimization scheme in the safety helmet-wearing detection.(3)A multi-scale feature network structure based on the separation of location and category prediction is designed,based on this network,a multi-scale detection training strategy is proposed.Firstly,this thesis introduces the current mainstream safety helmet-wearing detection algorithm based on deep learning,including YOLO v3.It uses the same feature map in position and category prediction.However,in position prediction and category prediction,the algorithm has different requirements on the nature of feature map.Therefore,taking YOLO v3 network structure as the main body,a location and category prediction based classification is designed.Based on the network structure,the influence of replacing full connection layer with convolution on the detection performance of the whole helmet-wearing is analyzed,and a multi-scale detection training strategy is proposed.Experiments show that the algorithm improves the detection effect of different resolution images after multi-scale training using the network model.Experiments show that the improved algorithm proposed in this thesis achieves 94.13% of the mean Average Precision in the task of safety helmet-wearing detection,which is superior to other mainstream algorithms,and the detection speed reaches 55 fps,which is approximately the same as YOLO v3 algorithm which can meet the realtime detection requirements.