Research On Traffic Sign Detection And Recognition Method Based On Deep Learning

Traffic sign detection and recognition has become a key technology in autonomous driving and Advanced Driving Assistance Systems(ADAS).This technology can be generally divided into traffic sign detection and traffic sign recognition.Although it has been studied for many years,it still faces many challenges.In this regard,this thesis designs a network structure of "Detection + Localization Refinement + Classification" through in-depth research on traffic sign detection and recognition methods based on deep learning,and verifies its effectiveness through experiments.The specific work of this thesis is as follows:(1)To solve the problem that traffic signs are difficult to detect due to their small size,this thesis proposes two deep convolutional networks based on YOLOv3.One is a detection network based on multi-scale features,which can extract multi-scale features by upsampling top-level features.Then the feature fusion network fuses different scale features to improve the detection effect of small-scale targets.The other is a detection network based on multiple receptive fields.This network uses Darknet-53 as the framework of the backbone network,improves the detection accuracy for small-scale targets by increasing the downsampling rate,and introduces dilated convolution to adapt to multi-scale targets.Experimental results show that the proposed two detection networks have higher detection accuracy and detection speed,and their performance is better than other detection methods.(2)In view of the problem that the deviation of the bounding box of the detection network prediction,this thesis proposes a localization refinement method based on UNet.Firstly,the bounding box is expanded to include the whole traffic sign.Then,the U-Net network is used to segment the traffic sign in the bounding box and obtain the corresponding binary image.Finally,the more accurate bounding box is determined by extracting the traffic sign contour in the binary image.By analyzing the overall performance of “Detection + Localization Refinement + Classification” in GTSDB dataset,it is verified that this method can effectively improve the overall performance of the network.(3)To further improve the performance of traffic signs recognition methods,a traffic sign recognition method based on spatial transformation network(STN)is proposed in this thesis.This method achieves fast and accurate classification of traffic signs by combining STN with classification network based on asymmetric convolution.Experiments on the GTSRB dataset show that the proposed recognition method achieves a classification accuracy of 99.52%.