Research On Traffic Sign Recognition Method Based On Improved YOLOv4

With the advancement of science and technology,ownership of cars daily increments,people in the choice way to travel all the way to the car as the first choice.In order to improve people’s driving comfort and safety,a large number of researchers have done a lot of research on the auxiliary driving functions of cars.According to the survey,in the cause of traffic accidents in the collection,we found the vast majority of accidents are due to the driver failed to notice traffic signs information leading to unsafe incidents.So,traffic signs and information have a particularly important position on the road,which not only deliver environmental information to drive traffic,it can also regulate drivers safe driving.As a technical extension of target recognition technology for applications,the development of traffic sign recognition technology is the optimized product of target recognition algorithms.In recent years,target detection algorithms based on deep learning technology have improved a lot,and scholars at home and abroad have been relatively mature in this regard.For example,the YOLO algorithm based on the representative of the convolutional neural network series has undergone four major improvements until today’s YOLOv4.The algorithm is mainly for target detection technology and has a certain degree of efficiency,but there is still a great improvement in accuracy.In this paper,YOLOv4 improved algorithm research based on the traffic sign recognition system.(1)In this paper,the convolution neural network algorithm and its derivatives YOLOv4,its network structure,algorithm principle to make an in-depth study.In view of the two problems of YOLOv4,namely training efficiency and test accuracy,corresponding researches and improvements have been made.(2)In this paper,three times are trained on the TT100 K Chinese traffic sign data set creation model,and three corresponding weights are obtained,and the test results are performed on performance experiments and the results are obtained.Improvements based on gradient descent,self-confrontation training methods,and batch normalization have a better optimization effect on training efficiency.Clustering prior frame and the non-maximal suppression may be improved method has improved the accuracy of the test.Two of the non-maxima suppression method DIo U-NMS and Soft-NMS test results also make a comparison,the extent of the performance quite.(3)After herein achieve an improved method and obtained YOLOv4 result,the improved algorithm for the design of a Web system according to the visual side B / S network architecture,and the results show.