Research On Methods Of Traffic Sign Recognition In Complicated Environment

Traffic signs detection and segmentation is an important research domain in ITS (Intelligent Transportation System) and part of the key components in Driving Assistance System and automatic driving technology in the future. At present, in the pace of rapid develop of computer technology and artificial intelligence technology, target detection and target recognition algorithm have been brought out and improved. However, because of the many types of traffic signs, the variability and complexity of the surrounding environment and the specific nature of the problem, it is necessary to research the algorithm of traffic signs detection and segmentation.Traffic sign data are required to be collected from both sides of the actual road. The complexity of environment and any various factors may lead to emergement of incomplete traffic sign and the existing detection algorithm can't handle the issue well. How to achieve effective feature of traffic signs is the key procedure in order to realize the well classification results of traffic sign. So far, the existence of the best have not been proved in the theory or verified by experiment. And the pursuit to robustness and effectiveness of algorithm still is a hot issue in the traffic signs classification. Research in this area has always remained constant development.This paper makes the following researches according to the above descriptions and problems.(1) Based on the research and analysis of existing algorithm of traffic signs detection, a detection algorithm is proposed base on the color probabilistic model approach. It can been used to rectify the color distortion of traffic sign collected from the natural scene, overcome the limitations of various models and improve the detection accuracy; in the analysis of shape matching templates, a local templates matching method is proposed to overcome the drawback of global templates matching method which incomplete traffic sign is not detected. In response to these aspects proposed, respectively, experiments have proved that the traffic signs detection algorithm based on the color probabilistic model and local templates is feasible and effective.(2) In the study of existing algorithm of traffic signs, the Implicit Shape Model(ISM) is proposed to solve the timeliness and lack of detection rate of existing algorithms. The main idea of ISM is to integrate recognition and segmentation into a common probabilistic framework and generate a per-pixel confidence measure specifying the area that supports a hypothesis and how much it can be trusted. It not only overcomes the shortcomings of the inefficiency of previously published detection methods of traffic signs but also has the ability to detect the incomplete traffic signs under significant partial occlusion. Finally, we present the test results which show the proposed method can efficiently detect and segment the traffic signs from the natural scenes.(3) The frequency distribution feature of traffic sign is studied after achieving the effective segment. According to the good localization characteristics of Gabor wavelet transform between spatial and frequency, as well as multi-resolution characteristics, effect which Gabor wavelet extracts object feature is focus to study, especially traffic sign object. We mainly emphasis on the edge of the process, the response of intensity and character moments of traffic signs, proposed a feature extraction based on the Gabor wavelet transform, and calculate and obtain the Gabor features of traffic signs. Finally, experiment verifies the Gabor features as the traffic sign recognition feature vector is feasible.(4) By studying the process to eliminate redundancy and reduce dimensionality of complex data, MPCA (Multilinear Principal Component Analysis), the extended form of PCA (Principal Component Analysis) in the multilinear algebra field, is studied, and theoretical basis of MPCA algorithm and iterative implementation process of algorithm are emphasized on. A new improved proposed MPCA algorithm is theoretically proved to superior to the original MPCA algorithm in data redundancy elimination, as well as processing speed. At the same time, we also compare to improved MPCA algorithm's ability to process data to PCA and 2DPCA. Finally, by comparing the experiments were to verify the improved MPCA algorithm's superiority in processing speed and eliminating data redundancy.(5) By studying the process to eliminate redundancy and reduce dimensionality of complex data based on the distinction between the various categories and between the various samples, MDA (Multilinear Discriminant Analysis), the extended form of LDA (Linear Discriminant Analysis) in the multilinear algebra field, is studied, and theoretical basis of MDA algorithm and iterative implementation process of algorithm are emphasized on. On this basis, the theoretical basis of MDA is enriched and a new iterative process is proposed to improve the speed of the processing. Then, in the in-depth analysis of MDA, a new improved proposed MDA algorithm is theoretically proved to superior to the original MDA algorithm in data redundancy elimination, as well as processing speed. At the same time, we also compare to improved MDA algorithm's ability to process data to LDA and 2DLDA. Finally, by comparing the experiments were to verify two types of improved MDA algorithm's superiority in processing speed and eliminating data redundancy.(6) By analyzing the effectiveness and differences of eliminating redundancy and reducing dimensionality between improved MPCA and improved MDA, a new feature selection method, fusion of improved MPCA and improved MDA, is advanced to realize the effective compression of feature space dimension and elimination of redundancy. Finally, experiments show that the proposed method confirms the correctness of the theoretical analysis.(7) Based on the study of existing classification algorithms, we notice that an AdaBoost algorithm, one of top 10 classification algorithms, has a very powerful classification capability and little application in the classification of traffic signs. AdaBoost classification in the application of traffic sign is presented, whose weak classifier is BP neural network and whose input feature vectors are Gabor feature, the improved MPCA feature, the improved MDA, the Integration of improved MPCA and improved MDA and Gabor feature of integration of improved MPCA and MDA, respectively. Finally, contrast experiments verify the feasibility of the proposed algorithm.