Traffic Sign Recognition Based On Feature Fusion And Dictionary Learning

With the development of the economy,the improvement on living standard and the number of vehicles is increasing,the traffic congestion and safety problems become more serious.Therefore,the research and application of intelligent transportation system is highly valued by the industrial world and the academic world.As one of the core technologies in ITS,traffic sign recognition has been the focus on research.However,the research results have not yet reached maturity,This is due to in the complexity of the real scene,the visual recognition of traffic signs is affected by the weather conditions,motion blur,other objects occlusion,the color degradation,rotation tilt and so on.Therefore,traffic sign recognition is still a challenging task.The main research works in this paper are summarized as follows:(1)Due to the influence of the various factors,the traffic signs vary in different aspects such as the size,brightness and shape etc.Therefore,it is necessary to preprocess the traffic sign images before the feature extraction.This paper mainly deals with it from three aspects:?Grayscale and grayscale normalization;?Segmentation of the region of interest;?Scale normalization.(2)According to various characteristics of the traffic signs such as the local edge,the overall contour,texture and so on,HOG and GIST are studied and the parameters of each feature are optimized.Since a single feature is difficult to fully describe the characteristics of traffic signs,so a multi-feature fusion strategy is used in this thesis,but the inappropriate fusion algorithm will make the expression ability of the fusion feature even worse.In view of GCCA(the generalized Canonical Correlation Analysis)has been used in the face recognition and achieved good results,which is applied to the field of traffic sign recognition of this thesis.The experimental results show,the fusion of HOG and GIST is more beneficial to classification than single feature.(3)In view of the experimental results,there are some redundancy in the expression of the fusion feature for the traffic signs of similar structure,especially,the speed limited class.So the fusion feature is optimized by using the dictionary learning sparse coding in this thesis,which uses K-SVD algorithm for dictionary learning and Orthogonal matching pursuit algorithm for sparse coding.The experimental results shows that the optimized feature is more expressive for the traffic signs of similar structure.(4)The classification results from single feature,fusion feature and the optimized feature are analyzed by the comparative experiments on the GTSRB(German Traffic Sign Benchmarks).The experimental results show that the optimized feature with the linear SVM can achieve better results.