Research On Detection And Recognition Algorithm Of Traffic Sign Based On Deep Learning

With the increase of vehicles and the complexity of road environment,ITS has been favored by many researchers.Traffic sign detection and recognition,as a key issue in ITS,plays an important role in the implementation of the whole system.Due to the increasingly complex road environment,especially the influence of external factors such as light,weather and so on,traditional methods are difficult to meet the actual needs.Aiming at the practical problems of traffic sign detection and recognition,combined with deep learning and image processing knowledge,the research on road traffic sign detection and recognition is carried out.A traffic sign detection and recognition algorithm based on deep learning is designed and implemented.On this basis,the design of a prototype system for traffic sign detection and recognition is completed.Compared with traditional methods,the detection accuracy and robustness are greatly improved.Follows are the main research contents:This paper introduces the general classification of traffic signs,some common data sets,the general methods of solving the problem of traffic sign detection and recognition,and deep learning framework Caffe.At the same time,this paper analyses the principle of deep learning,activation function and common basic network.Traditional traffic sign detection methods have some problems,such as poor robustness,large amount of calculation and time-consuming detection.In order to solve the shortcomings of traditional methods,a traffic sign detection algorithm based on Faster RCNN is proposed in combination with deep learning.The principle of Faster R-CNN target detection is analyzed,and the working process of network feature extraction,RPN network and detection network is emphatically analyzed.The original data of GTSDB are preprocessed and expanded to effectively improve the training effect of the network.Optimize the internal structure of Faster R-CNN,adjust the proportion of anchors and detection threshold,and use ZF as the pre-training model to detect traffic signs with smaller pixel value.The appropriate learning rate,batch size and momentum coefficient are selected by comparing several groups of parameters.The validity of this method is proved by the loss curve and the test experiments in different environments designed on GTSDB data set in Germany.The migration test on Swedish STSD dataset shows good generalization ability.To solve the problem of traffic sign recognition in natural environment,a CNN recognition method based on image clustering is proposed.The image clustering algorithm is used to optimize the traffic sign data set,which effectively improves the overall quality of the samples and the training effect of the model.This method can be generalized to other data optimization problems.The data after clustering are pre-processed multiple times to ensure the training effect.The working principle and parameter updating process of CNN are analyzed.By optimizing the internal parameters of CNN,a new nine-layer CNN structure is constructed.The final recognition model is obtained after several iterations.Input the pictures to CNN model to realize automatic recognition.The image feature map is extracted,and the effectiveness of the algorithm is verified by loss and precision curve.The actual performance of the algorithm is verified on the German Traffic Sign Data Set GTSRB and Belgium Traffic Sign Data Set Belgium TSC.The comparison experiments show that the proposed method has the characteristics of fast recognition speed and high accuracy.By analyzing the practical requirements of traffic sign detection and recognition,a system implementation method to solve the problem of traffic sign detection and recognition is presented.Two common methods of traffic sign detection and recognition system in road scene are analyzed,and the actual demand of the system is analyzed.The overall design framework of the system is given from the aspects of software and hardware platform,system environment and the functions that the system should have.Starting from the function of the system,the prototype system of traffic sign detection and recognition is designed and implemented,and its interface and various functions are analyzed.It has the characteristics of simple operation and convenient use.The overall workflow and performance evaluation index of the system are given,and a new data set is made to test the performance of the system.The results show that the system has good robustness and recognition accuracy,and has obvious advantages over traditional feature description methods and infrastructure interconnection methods.