The Research On Traffic Sign Detection And Recognition Based On Convolutional Neural Network

Autonomous driving is a current research craze.The detection and recognition of traffic signs is an important part of the automatic driving system.It must transmit timely and accurately some important road traffic information to the driving syst em to guide the car to make a correct driving judgment,so as to avoid traffic accidents and achieve safe driving.However,the existing traffic sign detection and recognition technologies are either low in accuracy or poor in real-time,which is difficult to meet the needs of real vehicle-mounted applications.Therefore,this paper uses convolutional neural network to study the detection and identification of traffic signs.The research content is as follows:The traditional recommended candidate area detection algorithm,which takes a long time to generate candidate areas and has low quality,fails to meet the real-time and accurate requirements of traffic sign detection.Therefore,this paper adopts a detection algorithm based on Faster RCNN for traffic sign detection.Faster RCNN used RPN network to quickly generate high-quality candidate regions,and shared convolution features between RPN network and Fast RCNN network through alternative training,so as to achieve fast and effective traffic sign detection.In order to give consideration to the detection of small-size traffic signs,32×32 and 64×64 are added on the basis of the original three anchor box parameters of 128×128,256×256 and 512×512.Finally,Faster RCNN was achieved by using VGG16 as the basic network on German Traffic Sign Detection Benchmark(GTSDB),its mAP value reached 89.47%,and the average speed of detecting an image after GPU acceleration was only 0.41 s,which was nearly real-time.In addition,this paper designed an improved convolutional neural network recognition algorithm for traffic sign recognition.Firstly,a network model suitable for traffic sign recognition is designed by combining the advantages of traditional networks such as LeNet5 and VGGNet and fully considering the factors of the size of convolution kernel and the number of convolution layers that affect the network structure performance.Then,Parameteric Rectified Linear Units(PReLU),Exponential Linear Units(ELU)and Batch Normalization(BN)are introduced to form a variety of improvement schemes,and through a large number of experiments,it is found that the performance of the network model with the introduction of ELU alone is better than that with the introduction of ReLU or PReLU alone,but the combination BN and PReLU is not as effective as the combination of BN with ReLU or PReLU,and the combination of BN and PReLU is the best among the major improvement schemes.Finally,the combination of BN and PReLU is adopted to accelerate the network training speed and improve the accuracy of identification,further improve the network model of traffic signs,and make the identification rate of traffic signs reach 98.17%.