Research On Road Traffic Sign Recognition Method Based On Deep Learning

With the rise of assistance and driverless driving,traffic environment perception is becoming more and more important.Road traffic sign recognition is an important research content of traffic environment perception.With the maturity of computer vision technology and wide application in the field of transportation,the use of object recognition technology to provide information such as the location and type of traffic signs for assistance and driverless driving is conducive to the intelligent development of traffic and the information management of traffic signs.Traffic sign recognition methods based on traditional features and shallow neural networks are difficult to meet the needs of recognition accuracy.Therefore,deep neural networks are used to extract traffic sign features to achieve efficient,real-time detection,classification and recognition of traffic signs.In this paper,traffic sign detection is based on feature extraction and candidate regions.In order to improve the accuracy and real-time performance of traffic sign classification,a lightweight deep learning model is used to implement traffic sign classification.Considering the real-time requirements,a deep neural network with integrated recognition is used for traffic sign detection and classification,and provide traffic sign information support for intelligent connected vehicle and traffic management.First,the characteristics of traffic sign images and basic characteristics of traffic signs in natural environments are analyzed.And the attributes of mainstream traffic sign classification and recognition data sets at home and abroad are summarized.At the same time,the method of deep learning is explained,and the similarities and differences between deep learning and traditional computer vision technology and machine learning are analyzed.Then,based on the distinctive characteristics of traffic signs such as color,shape,texture,etc.,the analysis of different traffic sign image processing techniques to eliminate interference factors such as light,smog and other factors.Different color space,HOG,U-LBP and other methods are used to extract the significant features of traffic signs.At the same time,the effects of different feature points of traffic signs are compared and analyzed.Image processing methods and feature extraction methods are provided for subsequent traffic sign detection,classification and recognition..Secondly,based on the HSV-HOG fusion feature traffic sign candidate area is extracted to detect the location of the traffic sign.At the same time,based on the traditional CNN,the spatial pyramid and BN methods are introduced to propose an improved SPPN-CNN classification model,and the Softmax classifier is used to achieve traffic sign classification.Based on the traditional Yolov3 network,the FCM clustering method is used to improve the default a priors anchor boxes size of Yolov3.The improved Darknet53 network,which combines convolution layers and batch normalization layers,is used to extract multi-scale features of traffic signs.At the same time,the GIOU is used to solve the problem of incomplete detection of IOU.These methods are used to construct an improved Darknet53-Yolov3 high-performance traffic sign recognition model.Finally,a deep learning platform based on Tensorflow-GPU was built,and the Chongqing traffic sign recognition data set CQTSDB was self-built from the aspects of traffic sign image acquisition and image annotation.The fusion feature is combined with MSER to extract candidate regions,and the HSV-HOG-SVM method is used to detect traffic signs.At the same time,based on the GTSRB data set,the traffic sign classification model based on improved SPPN-CNN was trained and verified,and comparison experiments were carried out from three aspects of image processing,model parameters and model structure,and the performance of the improved model was tested.Then,based on LISA,GTSDB and self-made CQTSDB data sets,the improvement of the training process and test recognition effect of the Darknet53-Yolov3 model are discussed.Experiments show that the improved model has excellent performance,high recall,accuracy,strong robustness and generalization,and can realize real-time and efficient recognition of road traffic signs in different traffic environments.