Traffic Sign Detection And Recognition Based On Shared Convolutional Neural Network

Traffic sign detection and recognition is one of the key technologies for Advanced Driver Assistance.Accurately obtain traffic information on the road ahead can provide decision support for the driver or smart vehicle to perform actions,and reduce the probability of traffic accidents.Based on the analysis and research of traffic sign detection and recognition technology at home and abroad,based on the deep convolutional neural network architecture,this paper proposes a traffic sign detection and recognition method based on shared convolutional neural network,and makes algorithm testing and comparative analysis.The main research work of the paper is as follows:(1)Propose a traffic sign detection method based on improved Region Proposal Network(RPN).Firstly,a Residual Convolutional Neural Network(ResNet)is used to convolve the input image to obtain convolution feature maps;Secondly,in order to detect traffic signs at different scales,anchor boxes are set based on five kinds of the input image:16?16、32?32、64?64 、128?128 、256?256;Then,anchor boxes perform a convolution operation on the last layer of the convolution map to obtain a fixed dimensional feature vector;Finally,the traffic sign candidate area and the border position parameters are output through two fully connected layers.The algorithm can effectively improve the robustness of traffic sign detection at different scales.(2)Propose a traffic sign recognition method based on improved Spatial Pyramid Pooling Network(SPPNet).Firstly,different scales of traffic sign candidate regions obtained by the traffic sign detection method are mapped into the ResNet to obtain convolution feature maps of different sizes.Secondly,different sizes of the convolution feature map is pooled to a fixed-size output by the SPP layer,and a fixed-length feature vector is obtained;Finally,the traffic sign type and border position parameters are output through two fully connected layers.This algorithm can effectively solve the multi-scale input problem in traffic sign recognition tasks.(3)Implement a traffic sign detection and recognition software system.Based on the traffic sign detection and recognition algorithm proposed in this paper,a traffic sign detection and recognition software system based on Windows platform is built.The system mainly includes image acquisition and preprocessing module,network model training module and traffic sign detection and recognition module.Through the real scene data collection of Xi’an Road,the software system performance test was implemented.The test results show that the system can accurately detect the traffic sign area in front of the road and record the recognition type and time of the traffic sign.For the traffic sign detection and recognition algorithm based on shared convolutional neural network proposed in this paper,German traffic sign detection data set(GTSDB)is used for algorithm testing and performance analysis.The results show that the proposed method has better recognition results for 43 types of traffic signs in the test set,and it can reach a average recognition rate of more than 95%.By comparing and analyzing traffic sign detection and recognition method based on Faster RCNN and traffic sign detection and recognition method based on "SS+AlexNet",it is proved that the proposed method is obviously better than the other two methods on the basis of comprehensive consideration of time loss and recognition accuracy.