Research On Key Technologies Of Traffic Sign Detection And Recognition

As an important part of Intelligent Transportation System(ITS),traffic sign recognition system has broad prospects for development and application in unmanned and driver assistance systems.In the complex and varied road environment,traffic signs are susceptible to many factors such as light and weather changes,defacement and motion blur,which makes it difficult for existing traffic sign detection and recognition algorithms to meet real-time and accuracy requirements.Therefore,this thesis focuses on the traffic sign detection and recognition algorithms in natural scenes,the main works are as follows:Using an enhanced algorithm based on discrete wavelet transform and multi-scale Retinex to preprocess the traffic sign image.Multi-scale Retinex adjustment is performed in the low frequency sub-band of wavelet decomposition to enhance the contrast of the image;And use multi-scale detail boosting method on high frequency sub-band to enhance detail information.Which significantly improves visual effects of the image,solves the problem of low contrast and uneven brightness of traffic sign images caused by illumination and weather changes,and improves the accuracy and robustness of image segmentation.According to the color characteristics of traffic signs,the RGB image is processed though the color enhancement method to obtain the color enhanced grayscale image of each channel.And then set dynamic threshold based on the pixel value to segment the image to get traffic sign candidate regions,which improves the defect of traditional segmentation methods such as poor adaptability and poor segmentation effect.Finally,training the SVM classifier though the HOG feature to accurately locate the candidate regions,which further improves the accuracy of the detection.Using three-layer convolution-pooling structure,and increase the number of C1 and C3 convolution kernels to 20 and 40 respectively,and adjust the convolution kernel size of C3 layer to 3×3 to improve the Lenet-5 convolutional neural network,in order to fully extract the details of traffic signs,enhance the recognition ability of the network;Secondly,use the ReLU activation function instead of the Sigmoid function to accelerate the convergence of the network;Finally,the SVM classifier is used to classify at the output layer to further improve the recognition accuracy.Which solves the problems of low recognition rate and long time-consuming of traditional traffic sign recognition algorithm.The experimental results show that the improved algorithm is superior to the traditional algorithm in recognition speed and accuracy.The recognition accuracy reaches 97.88%,which is 4.75% and 8.34% higher than that of traditional Lenet-5 and SVM recognition algorithm respectively.