Traffic Sign Recognition And Its Implementation Based On Deep Learning

Traffic sign recognition is a key technology in advanced driving assistance and autonomous driving.At present,the main research thought for this problem is to obtain natural scene images through cameras installed on vehicles,then the traffic signs in the scenes can be detected and recognized in real time through image processing,pattern recognition and other technologies.Due to the complexity and changefulness of the real traffic environments,as well as the high accuracy and real-time performance demand in application,the study of traffic sign recognition has great theory and practical significance.The traditional way,which uses sliding windows together with manual features for traffic signs recognition,cannot meet the needs of intelligent driving system in the aspects of accuracy and real-time.In recent years,the development of deep learning algorithms such as Convolution Neural Network(CNN)and Region Proposal have provided new possibilities for the traffic sign recognition technology.This thesis applies the object detection idea,which is based on deep learning,to traffic sign recognition,and further implements the traffic sign recognition based on the intelligent vehicle platform.The specific works mainly include the following steps:(1)The traffic sign detection algorithm is studied in the way of deep learning based on the GPU server platform.First,Faster Region-based Convolutional Neural Networks(Faster R-CNN)is used to detect traffic signs,and it can implement the end-to-end training through Region Proposal Network(RPN)sharing full image convolution features with Fast R-CNN detection network.Then,the deep ResNet is applied to deepen the network and the Region-based Fully Convolutional Networks(R-FCN)algorithm is used to carry out the full convolution calculation of the whole image,so as to improve the traffic sign detection performance.Finally,the algorithm contrast experiment is implemented on the TT100K traffic sign data set.The results,which come from the traffic sign detection based on R-FCN,show that the recall rate is 98.1%,the accuracy rate is 98.7%and the mAP is 93.5%.The performance is better than other traffic sign detection algorithms used in this thesis.(2)The traffic sign recognition system is implemented based on the intelligent vehicle platform.First,the intelligent vehicle experiment platform is built,including the design of Jetson TX1 embedded platform and the creation of R-FCN embedded platform environment.Next,Labelling is used to mark the data collected from the real roads and to build a traffic sign data set based on the intelligent vehicle platform;The R-FCN further training model optimization is done by model fine-tuning,parameter debugging and other methods.Finally,the traffic sign recognition test is carried out inoff-line videos under different driving scenarios on the Jetson TX1 embedded platform.The results show that the recall rate is 93.8%and the accuracy rate is 94.1%.The recognition rate is 2.6 frames per second.The experiment verifies the reliability of the traffic sign recognition system based on the intelligent vehicle platform.