Research On Location And Recognition Algorithm Of Traffic Sign

With the further research of automatic driving and assisted driving, automatic identification technology of traffic signs has become a research hotspot. Traffic sign identification algorithm for automatic location and identification needs real-time, accurately locates and identifies traffic signs. It is robustness, precise and real-time demanding.The third chapter discusses the traffic sign rough location algorithm, which is a combination of traffic signs rough and fine location based on the color space so as to detect traffic signs. Currently common coarse location,based on the RGB color space, can not detach the traffic signs from the background quite effectively. This paper aims to improve the algorithm, making it possible to better separate signs and the background. But it cannot overcome problems of darkness and brightness the color distortion of traffic signs. This paper aims at the real-time positioning, robustness, accuracy design of two algorithms, which are:(1)Clustering segmentation based on L*a*b* color space was first applied to medical image segmentation. It is the first time that it has been applied in the traffic sign image segmentation, which achieves good results. It bears high accuracy and can extract complete traffic sighs. What's more, it has strong robustness to light shade in the environment and color distortion of traffic signs. However, the real-time is slightly not as good as the previous method.(2)The coarse segmentation of traffic signs based on YCbCr color space is also an original efficient method of detection introduced by this paper. With the advantage of YCbCr color space in image segmentation, and its combination with the colors of traffic signs themselves, the algorithm can rapidly and accurately separate the traffic-sign region. This algorithm is simple and feasible, of high accuracy, but less calculation. Furthermore, it has also a strong robustness for the light change and the color distortion of traffic signs.Comparing the accuracy and real-time of the three methods through quite a lot of experiments, this paper comes to such a conclusion: as for accuracy the second one is the best, followed by the third; as for real-time, the third one is the best with the first one the next. In general, the third method is the most suitable for the detection of traffic signs.The fine segmentation algorithm based on the regional shape is first to mark the shape of the retained region after the coarsely-segmented images, and next analyze the shape of each marked region to detect the circular area, rectangle area, triangle area, and then locate these areas, cut them out so that they are finally input as to-be-recognized traffic signs into the traffic-sign classifiers designed for the next step.Two methods of identifying traffic signs are designed here. One is the zernike moment feature matching on the base of correlation coefficient. It can identify different traffic signs by measuring the correlation coefficient between the to-be-located moment feature vector and the marked templates moment the feature vector traffic. Here the zernike moments are used as the feature vector to measure the correlation coefficient between extracted zernike moment and he standard template moment vectors of traffic signs correlation measurements. We take as the type of the to-be-detected traffic signs the templates of the category types whose correlation coefficient is the closest to one.Another method is based on BP neural network identification. After being pre-treated and standardized, the to-be-detected traffic signs, as the input vector, are input into the pre-trained neural network for identification. 200 traffic signs are used as training samples to train the BP neural network. Trained neural network has the reliability and efficiency.The identification methods used above are based on the whole traffic signs, the greater the identified area is, the more intervening information they carry. And the accuracy will decline. The structural characteristics of the traffic signs are taken into consideration. Identification of traffic signs on a separate frame inside of the image (also as internal images) can identify traffic signs. Therefore, the other approach designed here is: First train well BP network with the inside image of 200 traffic signs as BP network training samples. To-be-identified marks also extract the internal images and then are input into the BP network for identification after being standardized. When the image quality of the to-be-identified traffic signs is better, the accuracy is higher compared with the BP neural network.