Research On Traffic Sign Recognition Based On Deep Learning

Traffic sign recognition is an important research direction in the field of intelligent driving.Providing accurate and fast road traffic sign information for vehicle drivers will help reduce the traffic accident rate and improve the operation efficiency of urban road traffic.In recent years,artificial intelligence technology has developed rapidly,and deep learning algorithm represented by convolutional neural network has made certain progress in the field of target recognition,with relatively high target recognition accuracy and stable detection performance.Therefore,YOLOv3,a representative of convolutional neural network algorithm,is selected in this paper to study traffic sign recognition.The main work content is as follows:First,the basic principles of convolutional neural network algorithm and relevant technologies used in model training are analyzed in detail,including neural network model,convolution operation,gradient descent algorithm,random inactivation,as well as the principles of typical convolutional neural network RCNN and YOLO algorithm.Secondly,the traffic sign recognition method based on YOLOv3 is studied,and TT100 k Chinese traffic sign data set is selected.In view of that YOLOv3 cannot fully extract the shallow feature information of images,thus affecting the identification of small-size traffic sign targets,this paper improves the YOLOv3 model,proposes the improved model Yolov3-improve,increases the convolutional layer by 1*1,and uses the four-fold down-sampling feature map to predict the target information.Due to the imbalance of positive and negative samples in the training samples,in this paper,Focal loss cross entropy function was used in the loss function of Yolov3-improve to further improve the model classification ability.The experimental results show that the average accuracy(m AP)of the YOLOv3-improve model is 3.7% higher than that of YOLOv3,indicating that the improved structure of YOLOv3 improves the classification ability of the model for smallsize traffic signs.After the use of Focal Loss function in Yolov3-improve,the m AP value of the model increased by 1.7%,indicating that the Focal loss function has a certain effect.Finally,the model inference acceleration method was studied,the principle of integrating BN layer and convolutional layer was analyzed,and the BN layer fusion experiment was carried out for the YOLOv3-improve +Focal Loss model.The experimental results showed that when detecting a single image,the time of YOLOv3-improve +Focal Loss after integrating BN layer was improved by 0.02 s.