Traffic Sign Detection And Recognition Algorithm Based On Convolutional Neural Network

At present,the number and types of vehicles in our country are increasing rapidly.While causing traffic congestion,the incidence of traffic accidents is also high.Therefore,how to alleviate traffic pressure and ensure travel safety has become an urgent problem to be solved in China.As one of the key technologies in the field of automobile assisted driving,traffic sign recognition can effectively alleviate driving fatigue and ensure driving safety.Aiming at the problem of traffic sign detection and recognition,this paper studies the convolutional neural network model using deep learning for traffic sign detection and recognition.The specific research work is as follows:(1)For the problem of image detection of traffic signs,the Faster R-CNN network model is studied.Because the Faster R-CNN network uses the VGG16 network in the feature extraction stage,the shallow and deep features of the target cannot be merged.The Res Net50 network is studied as the feature extraction network of the Faster R-CNN model using transfer learning.Due to a certain degree of deformation problem in the target area when the car camera is framing,the traffic sign images collected from reality are all rectangles with an aspect ratio close to 1: 1.Modify the area and recommend that the network obtain the candidate area’s aspect ratio In order to ensure that enough positive samples are obtained during the training process,reducing the difficulty of training due to the large gap between the number of positive and negative samples increases the training difficulty of the border regression parameters.Aiming at the situation of single target area and multi-target area in the real picture,the data in GTSDB was used for verification.The improved algorithm achieved a detection accuracy of 87.2% and a recall rate of 91.4%.(2)In the recognition stage,research and propose a Res Net network traffic sign recognition method that introduces Inception structure to improve the recognition accuracy to more than 99%.In actual traffic sign recognition application scenarios,it is often necessary to take pictures at a far place,resulting in the target area in the picture is generally small,and its feature structure will change from dominant to recessive in the propagation of deep neural networks.The residual structure combines the shallow and deep features of the input image,which can deepen the neural network structure and improve the recognition accuracy ofsmall-size target images.The Inception structure can expand the width of the neural network.Its essential feature is to train the neural network model to find the smallest structure that can fit the target feature,reducing human participation in control,that is,the advance setting of hyperparameters.To a certain extent,the connection of convolutional layers is consistent with the idea and effect of the fusion of shallow and deep features with the residual structure.Using the GTSRB data set,by comparing the classic Alex Net network,VGG16 network,etc.,the recognition accuracy of the recognition algorithm proposed in this paper has reached99.1%,the improvement effect is obvious and meet the real-time requirements.(3)Using MATLAB to design and develop a traffic sign detection and recognition software system.The system mainly includes image preprocessing module,convolutional neural network module,Faster R-CNN network module and help module.You can perform edge detection and other preprocessing operations on the real image containing traffic signs,you can also build and train your own convolutional neural network model to realize the traffic sign recognition function,and you can also use the improved Faster R-CNN network to the traffic signs in the picture.Perform positioning and identification.