The Research On Deep Learning Algorithm Of Helmet Wearing Detection Under Video Surveillance

In recent years,more than 60% of the casualties in the safety accidents in China’s construction industry are caused by the failure to wear safety helmets.Wearing safety helmet is an important guarantee for life and production safety in various engineering environments,and it has become a necessary measure to supervise workers to wear safety helmet.The existing helmet wearing manual supervision way is far behind the current industry 4.0 information,intelligent production mode.How to realize intelligent helmet wearing detection has become a hot issue to be solved.Based on the cutting-edge deep learning target detection technology,this paper proposes an intelligent detection scheme for helmet wearing in the field of computer graphics through the research and improvement of relevant algorithms.Through a comparative study of recent deep learning target detection models,this paper selects YOLOv4 algorithm as the main research object to study the deep learning algorithm for helmet wearing detection under video surveillance.Based on YOLOv4,the network structure of the algorithm is improved,which includes the following research contents:First,aiming at the problem that the detection accuracy of small targets is not high in the current helmet wearing detection example,this paper proposes a method by adding feature fusion channels through feature enhancement.By extracting deep and shallow features and using spatial pyramid pooling and fusion to build the model,the spatial location information of small targets such as hard hats in the large-scale feature map is retained,and the classification and location of small target detection is enhanced.Through the comparison of experiments,it is proved that the detection accuracy of the improved method is improved greatly.Secondly,aiming at the problems of large number of parameters and long training time in the training model of YOLOv4 and our algorithm,which limit the further play of the algorithm in complex scenes,an improved model is proposed,which changes the Backbone network of CSPDarknet into a deeply separated convolution network.For each input channel,a single convolution kernel is used for deep convolution to obtain the depth of the number of input channels.Then,a 1×1convolution(i.e.point-by-point convolution)is used to combine the outputs in the deep convolution linearly.Compared with the traditional convolution,the number of parameters is reduced,so for the scene with occlusion,occlusion can be reduced.The proposed model achieves good performance in recognition accuracy and performance consumption,and the detection speed can reach 20 FPS on NVIDIA 1080 Ti,which improves the accuracy by 1.12% at the same speed.Thirdly,aiming at the problem that the training model is not accurate enough in the test set,the idea of ensemble learning is used to further improve the algorithm’s accuracy.By redividing the data set,several detection models are trained in the form of batch data augmentation.Bagging algorithm is used to solve the problem of trained fitting,which can save the training time and further improve the detection accuracy.Finally,through a large number of experimental verification,experimental results show that the proposed security detection method based on YOLOv4 and depth-separable convolution and the safety helmet detection method based on depth-shallow feature fusion in complex environment have an average detection accuracy of 91.99%,which is 3.05% higher than the existing algorithm using YOLOv4 model.It can meet the practical application requirements of helmet detection in most working scenarios.