Traffic Sign Detection And Recognition Based On Convolutional Neural Networks

Traffic sign detection and recognition play an important role of the intelligent vehicle and driver assistance system.In many cases,the driver will miss the traffic sign information,because the driver lacks the attention to the traffic sign and the traffic sign is blocked by the occlusion.Therefore,the automatic recognition system of traffic sign can reduce the occurrence of traffic accidents.Traffic sign detection and recognition is not a simple process.A wide-angle camera is usually installed at the top of the vehicle to capture traffic sign and other relevant visual features.Many factors affect the quality of the traffic sign image,such as the speed of vehicles,changes of light,climate changes and so on.Traditional computer vision and machine learning methods are widely used in traffic sign detection and recognition.But these methods are quickly replaced by methods based on deep learning.In recent years,compared with other traditional learning methods,the deep convolutional neural network has great advantages.With the development of deep learning algorithm,the realization of stable and high performance,from an effective deep learning perspective,deep learning can solve the problem of traffic sign recognition better.On the basis of a large number of literature research and summarize,and according to the existing shortcomings of the traffic sign detection and recognition algorithm based on convolution neural network,this paper proposes a traffic sign detection algorithm based on single shot multiBox detector,and a traffic sign recognition algorithm based on compressed sensing domain and cross-connected convolutional neural network.The main content is as follows:(1)In the existing traffic sign detection methods,the detection accuracy of the fully convolutional neural network method is not high,and the regions with CNN features based on region proposal network is slow to detect.To solve these problems,in the paper,single shot multiBox detector(SSD)is used to detect traffic sign for the first time.SSD is one of the best object detection algorithms with high precision and fast operation speed.SSD is based on the network structure of a forward propagating convolutional neural network,producing a series of fixed-size boundary boxes and the score of the possibility of object instances in each bounding box.Finally,non-maximum suppression(NMS)outputs the final detection results.(2)In many computer visualization applications,sometimes it is not possible to reconstruct the target object successfully,so that the correct proportion of the image can not be determined.Because the reconstruction methods of compressed sensing have the following disadvantages:the high cost of computation;undesirability of the reconstruction result in low measurement;unsure some parameters such as sparsity.Therefore,the compressed sensing domain gets developed.This paper applies the compressed sensing domain to the recognition of traffic sign.The images of traffic sign are converted to the compressed sensing domain by using the measurement matrix without reconstruction process.And then it is used as the input of the traffic sign classifier.(3)The research shows that it is beneficial to improve the recognition performance of the visual system by taking advantage of high level features and low level features.However,a disadvantage of traditional convolutional neural network is difficult to effectively use the low level features and high level features to construct better classifiers.According this problem,based on the traditional convolutional neural nerwork,this paper used a cross-connected convolutional neural network model with 9-layer structure by introducing the idea of cross-layer connectioa The aim is to effectively integrate low level features and high level features to construct better classifier.The above proposed algorithms were verified on the well-known GTSDB dataset and GTSRB dataset,and compared with the current mainstream traffic sign detection and recognition algorithms.The results demonstrate that the algorithms proposed in this paper can effectively improve the detection accuracy and recognition accuracy of traffic sign,and successfully enhance the robustness of traffic sign detection and recognition.