The Detection And Recognitiong For Traffic Signs In Natural Scenes

With the rapid growth of car ownership in today’s society, traffic congestion has become a common phenomenon of urban traffic in many cities, the resulting traffic safety problems have become increasingly serious, and caused huge losses to the society. In this context, the intelligent transportation system(Intelligent Traffic System, referred to as ITS) came into being. The traffic sign recognition system as an important part of the intelligent transportation system, has also been widely concerned.Traffic sign recognition system is mainly used in the auxiliary driving system and unmanned driving system, via identifying the traffic signs on the road, assisting drivers with driving cars or driving vehicles automatically. Because the natural scene is complex, and the recognition technology is not yet mature. How to quickly and accurately identify the traffic signs in the natural scene also needs to be deeply discussed and studied.Because the traffic sign itself has distinctive characteristics of the color and shape,these characteristics can be used to quickly locate the traffic signs and then identify them. which can greatly improve identification efficiency of the system. However, the natural scene and light conditions are complex and changeable, the traffic sign itself may exist many problems including fading, occlusion and deformation, which is the reason that the color and shape of the traffic sign correspondingly change in the images collected. To solve the above problems, this paper proposes a detection and recognition system of traffic sign based on segmentation of color- shape.Firstly, in the pre processing of image, for the adverse effects that lack of light causes to the color of traffic signs. After analyzing the advantages and disadvantages of some commonly used image enhancement algorithm, this paper adopts and improves Retinex algorithm finally, and then do some corresponding processes to remove dark light according to the brightness of the image histogram characteristics.Secondly, in the segmentation of color image, the background interference and other adverse factors, this paper puts forward a double channel weight adaptive three-component color difference method based on the three- component color difference method. The algorithm assigns different weight coefficient to a red grayscale image and a blue grayscale image, and then converted the weight coefficient to the adaptive dynamic threshold that can adapt to more light scene and background interference. Atthe same time, the OTSU threshold segmentation algorithm is used to remove the large area background interference with low saturation. The experimental results show that the color segmentation method compared to other color can better adapt to illumination in different scene. Comprehensive detection effect of this method is also better than other methods.Thirdly, in the aspect of contour and shape detection, this paper analyzes the commonly used detection algorithm of circle, triangle and rectangle. Aiming at the problem that occlusion and fouling effect the shape of traffic signs,this paper proposes the approximation algorithm based on the profile of removing depression.According to the deformation and occlusion of circular traffic signs, the constant interval least squares ellipse fitting method is proposed to improve the detection rate and detection speed on the basis of the random least squares ellipse fitting method.According to the characteristics of triangular and rectangular contour line, this paper presents a detection algorithm of combination of contour line. Experiments show that this detection algorithm improves the detection rate and detection speed for deformation and occlusion of circular traffic sign. At the same time, it enhances the detection rate for rectangular traffic sign and rectangular traffic sign, and reduces the false detection rate.Fourthly, in the feature extraction of traffic signs, this paper uses one method which is to extract the HOG features of the sample image firstly and then apply PCA technology to reducing the dimension of image HOG feature. This method improves the recognition rate and recognition speed of the algorithm. In the classification recognition algorithm, this paper selects the commonly used SVM support vector machine algorithm. The experimental results show that the feature extraction and classification algorithms have better recognition results.