Traffic Sign Recognition Research And Application Based On Convolutional Neural Network

Traffic sign contains a lot of road traffic information, providing warnings, instructions and other auxiliary information for the drivers. It’s playing an important role for relieving the driver’s driving stress, reducing road traffic pressure and accident rates. If the drivers pay all their attention to identity and discover the traffic and make the correct response, it will inevitably increase the burden on the driver to accelerate fatigue, may lead to serious accidents. Therefore, the safe and reliable traffic sign recognition has received more and more favor from automobile manufacturers and drivers. However, the traffic environment is complex in real world, such as light intensity, bad weather conditions, partial occlusion, perspective tilt and so on, which makes the study of traffic sign recognition system is facing many difficulties, and the actual applications are far from mature stage. Based on research studies at home and abroad, we make studies focusing on the convolutional neural network in traffic sign recognition.Convolutional neural network is good at processing two-dimensional images of the location of translation, scaling, tilting or other forms of distortion. It has been successfully used in image recognition, voice recognition and traffic sign recognition. However, due to the deep structure of convolutional neural network, it’s relatively time-consuming for model training and pattern recognition, which is a serious flaw for traffic sign recognition system requiring a higher real-time performance. Therefore, this paper presents a fast convolutional neural network to solve the problem of traffic sign recognition, compared with the conventional convolutional neural network, to extract the same number in the characteristics of the case, can significantly reduce the running time.The main work of this paper include the following aspects:(1) Proposed a fast convolution neural network algorithm for the traffic sign detection problem. Firstly, we transform the original image into the gray scale image by using support vector machines, then use the template matching method to find the Region Of Interest (ROI), and send the ROIs to the fast convolutional neural network for classification. At end, test our detection algorithm on German Traffic Sign Detection Benchmark (GTSDB) data set. Experiment results show that the algorithm has high detection accuracy, able to adapt to a variety of adverse conditions of poor lighting, occlusion, and rotation.(2) To process traffic sign recognition problem, propose a hierarchical classification algorithm based convolutional neural network. Firstly, the traffic sign is divided into several crude categories, then the characteristics of identity for all types of pretreatment, post-treatment to identify subdivisions to draw the final result. Finally, testing the proposed recognition algorithm on German Traffic Sign Recognition Benchmark (GTSRB) data set. Experiment results show that the algorithm not only has higher classification accuracy, but also has a faster processing speed, more applicable for real-time traffic sign recognition system.