Research On Video Detection Model Of Helmet Wearing Based On Convolutional Neural Network

As an essential branch of artificial intelligence,image recognition technology has been diffusely utilized in many fields such as industry,agriculture,and public security.In order to ensure safe production and reduce accidents rate caused by workers without wearing helmets,it has significant engineering value to apply image recognition technology to the real-time inspection of helmet wearing and to raise the recognition accuracy.The convolutional neural network is used to detect whether the staff wears a helmet in this paper.The main work is as follows:(1)Aiming at the problems of small targets and difficult detection in the process of helmet wearing detection,this paper proposes an YOLOv4 algorithm based on ASPP(Atrous Spatial Pyramid Pooling)and multi-scale feature fusion.A detection layer is added to the algorithm to obtain shallower information,and the information of smaller targets is obtained through a lager size feature map;ASPP is used to replace SPP(Spatial Pyramid Pooling),and the atrous convolution with different expansion rates is introduced to obtain a larger receptive field and reduce the loss of small target information.At the same time,the K-means++ clustering algorithm is utilized to obtain12 anchor boxes that are more suitable for helmet wearing recognition,which is are evenly distributed on the four detection layers.The experimental results show that the detection accuracy,the detection speed,and the testing effect of the new algorithm are significantly improved.Compared with the original network,the MAP(average mean accuracy)is increased by 2.7%,and the missed detection rate is reduced by 43.9%.(2)By observing the convergence curve of the loss function generated by the improved network during the training process,it is found that there is a problem of oscillation and slow convergence.In this paper,the loss function CIOU(Complete Intersection Over Union)in the algorithm is analyzed for this problem,it is found that the parameter v in its formula only reflects the difference in aspect ratio,which hinders the effective optimization of the similarity of the model,and then affects the training results.Therefore,this paper uses the Focal-EIOU loss function to replace the original loss function to solve the above problems.Among them,EIOU(Efficient Intersection over Union)can save the complete features of the loss,and Focal focuses on high-quality anchor boxes,which can solve the problem of sample imbalance.The experimental results show that the new loss function improves the curve oscillation caused by the small number of batch samples and the unbalanced sample quality.The MAP value is increased by 0.6%,and the missed detection rate is reduced by 21.8%.(3)When using the above improved algorithm to detect continuous objects in a video stream,the problem of missing detection of a certain frame,false detection or continuous missed detection of multiple frames may occur due to uncertain factors of the environment.Aiming at this problem,this paper proposes a confidence correction algorithm based on time correlation.The algorithm uses exponential smoothing method to predict and correct the confidence of missed targets and suppress the confidence of mis-detected targets.The experimental results show that compared with the traditional detection model,the MAP value is increased by 1.2%,and the missed detection rate is reduced by 48%.In this paper,the improved algorithm based on(1)(2)points is compared with the original algorithm on the mixed data set.The results show a lower missed detection rate and a higher accuracy rate for small target detection.On this basis,combine with point(3),the missed detection rate is significantly reduced.Compared with the original detection model,the improvement in this paper reduces the missed detection rate by 77.2%,which significantly improves the accuracy of video stream target detection for helmet wearing.