Research On Traffic Sign Recognition Based On Deep Learning

Most of the traffic accidents are caused by drivers’ non-standard driving,such as drunk driving,fatigue driving,wrong operation and so on,which threaten People’s Daily travel safety.With the rise of artificial intelligence,the emergence of autonomous driving technology,for the operation of vehicles in the process of regulation,can effectively avoid traffic accidents caused by human operation.One of the key technologies of automatic driving is the accurate detection and recognition of traffic signs.However,traffic signs have many categories,different shapes and colors,so it is difficult to accurately detect and identify them.In this paper,the target detection algorithm in deep learning is adopted to solve the problem of detection and recognition of traffic signs,and the network structure of the model is analyzed and improved to achieve a better recognition effect.By comparing the functions and efficiency of mainstream target detection algorithms in target detection,this thesis chooses the YOLOv3 model as the baseline model.First of all,the feature fusion of the FPN of YOLOv3 model will result in an increase in the number of feature channels in the feature map,large dimensions,large computation,high model complexity and low efficiency.In addition,the importance of different fusion features is different,so three weighted ADD fusion methods are proposed,and the weighted ADD fusion method with better performance and normalized Re LU restriction is selected through experiments.Secondly,the weighted confidence loss function is proposed to solve the imbalance of the German traffic sign data set,and the weight value with better performance is selected through experiments.Finally,the improved YOLOv3 model was trained and tested on the German traffic sign dataset,and the m AP was increased from 87.4% to 98.45%,an improvement of 11.05%.At the same time,the detection speed of traffic signs is up to32 fps.It can be concluded that the method in this thesis can not only improve the accuracy of the model,but also have good real-time performance.To further research,the application of improved model,designed the traffic sign recognition system based on improved YOLOv3 model,drive the vehicle to the German traffic signs of real images and video,which can identify the different categories under the German traffic signs of real test achieved good recognition effect and proves the effectiveness of the design of traffic sign recognition system.