The Research On Classification And Recognition Of Traffic Signs Based On Convolution Neural Network

The role of road traffic signs is to guide the driver to drive safely.However,the driver's negligence to the surrounding environment,blindly go through red lights,blindly overspeed,and other disregard of road traffic regulations behaviors often occur and are the major causes of traffic accidents.With the development of big data,GPU and Internet technologies,advanced driver assistance systems and unmanned system have become a key technology for solving various traffic problems caused by human factors.Traffic sign recognition is part of the context of ADAS and perceived environmental aspects of unmanned technology,real-time and accurate identification of road traffic signs and feedback of identified traffic signs to drivers and driverless decision-making terminals are of great significance for improving traffic safety.However,at present,the existing traffic sign recognition technology has either low recognition accurac y or poor real-time performance.Compared with mature systems such as lane departure warning and frontal collision warning,the present traffic sign recognition technology is hard to meet the demand of vehicle application.The perfect combination of GPU parallel processing and deep convolutional neural network realizes the unification of accuracy and real-time,which can effectively improve the real-time and accuracy of traffic sign recognition.Therefore,a method of traffic sign recognition based on convo lution neural network is proposed in this paper.A convolutional neural network model for classification and recognition of traffic signs in China is developed.The model extracts image features by convolutional neural network and learns features to realize real-time and accurate traffic sign classification and recognition.The main contents of this paper are as follows:1)Design a neural network model for traffic sign classification in China.By drawing on the advantages of networks such as Le Net-5,Alex-Net and VGG Net,a network for traffic sign classification in China was designed and the model was pre-trained using German and Belgian traffic sign datasets.Then,by adopting the method of migration training,Using the collected traffic sign datasets to train the network model,the optimal network model for traffic sign classification in China is obtained.2)Design a neural network model of traffic sign recognition and location.Based on the YOLOv2 target detection method,the traffic sign recognition network model is designed and the traffic sign recognition network model is initialized with the trained traffic sign classification network weights.Then,the network model is trained with a large number of marked traffic sign data sets.The method of onlin e cycle learning and automatic trimming of traffic signs is put forward,and the network is trained cyclically to improve the recognition accuracy of the network.3)to achieve the development of embedded platform model transplant.The trained network model was transplanted to the Jetson TX2 embedded platform and tested on real roads over 100 km on rural and urban roads to verify the real-time performance,accuracy and stability of the model.The research shows that the traffic sign recognition model based o n convolutional neural network has good accuracy.The running speed of Jetson TX2 platform can reach 33 fps,which has high real-time performance and can meet the needs of vehicle application.