Study Of Traffic Sign Detection And Recognition Based On Deep Learning

In recent years,motor vehicles which mainly are private cars have gradually became the main transportation tools for people.The huge burden for the transportation system and many safety problems are brought.Traffic signs are important road safety facilities,so the traffic sign detection and recognition become a key technique for intelligent traffic system which has gained more and more attention for researchers.Consequently,deep learning,machine learning,image processing and other related technologies are used to realize the following research.In the traditional traffic sign detection and recognition algorithms,complex artificial features are often designed to recognize the traffic sign.The recognition accuracy is limited by these features.In this thesis,A traffic sign detection and recognition method based on convolutional features is proposed.Firstly,The image is preprocessed by using saliency method.Then the maximally stable extremal region algorithm is used to extract the candidate regions,Geometric constraints based on the shape of traffic signs are set to select candidate regions.Finally,a multi-column convolutional neural network is built for feature extraction,and an ensemble classifier based on naive bayes classifiers is used to classify the candidate regions by these convolutional features.The experimental results show that the application of the convolutional neural network achieves a higher recognition accuracy compared with traditional methods.In the road scene,difficult challenges for traffic sign detection and recognition can be presented due to the variations in illumination or the complex background.The previous method can not complete the detection and recognition effectively.For this problem,a traffic sign detection and recognition method based on conditional random field is presented.Firstly,The prior color feature map and color probability map are extracted from input image.Then a conditional random field model is trained to fuse these feature maps.Candidate regions are extracted from the fused feature map.At last,candidate regions are classified by a multiscale convolutional neural network.Experimental results demonstrate that the proposed method can improve the performance of traffic sign detection.In the previous methods,the algorithm structures especially in the candidate regions proposal stage are too complicated,which results in the time consuming problem.So a traffic sign detection and recognition method based on eye-movement information and fully convolutional network is proposed.The fully convolutional network is trained by the guidance of the visual attention map which obtained from eye-movement information extracted by Eye Link.Candidate regions are segmented by the trained fully convolutional network,which improves the effectiveness of region extraction.Then the candidate regions are classified by a trained convolutional neural network.Experimental results show that the proposed method can significantly reduce the time cost compared with the previous methods.