Traffic Sign Recognition Based On Deep Convolutional Neural Networks

Traffic sign recognition is one of the most important parts in advanced intelligent transportation system and auxiliary driving system, which is a hot research topic at home and abroad in computer vision and patter recognition. Regarding the German traffic sign image GTSRB dataset as the research object, this dissertation concentrate on new recognition methods for traffic signs, which combines the convolutional neural network (CNN) with support vector machine (SVM).The innovative work mainly includes:(1)SAE?RBM?DBN and CNN in deep learning are discussed first. Then the convolution neural network LeNet-5 model is analyzed, since the model can extract multi-scale features from digital images. Finally, the principle of SVM classifier is described, which constructs the optimal hyperplane to gain a good classification effect and has all kinds of applications.(2)A simple image preprocessing method is designed for GTSRB dataset, which is collected in real world. First, the original image are cropped to reduce the disturbance from the background; Second, after the image was converted into a grayscale image, it is enhanced and normalized, which made the useful characteristics more obvious. Experimental results show that the proposed method can effectively improve the image quality and be helpful to subsequent image recognition.(3)A traffic sign recognition algorithm based on the 2-level improved LeNet-5 model is proposed, which combines convolutional neural networks with SVM. With the consideration of the requirement of real-time recognition, the traditional network structure of LeNet-5 model is improved first. Next, a 2-level improved LeNet-5 model is trained with GTSRB dataset, where the first level categorized traffic signs to 6 categories with the improved LeNet-5 model, and the second level improved LeNet-5 model provide with the final category. Experimental results show that the proposed algorithm could provide with a correct recognition ratio 91.76%, since the multi-scale features could be fully analyzed with 2-level improved LeNet-5 model.