Traffic Sign Detection And Recognition Based Multi-feature Fusion

Road traffic signs is the road facilities that used to pass the road traffic information and traffic rules to the driver with combination of different colors, shapes and patterns, therefore traffic signs have significant color, shape and scale feature. This thesis taking full advantage of the above features of traffic signs, researches automatic detection and recognition in natural scene. This thesis presents a traffic sign detection and recognition based on multiple features fusion, using a variety of features to achieve synergy traffic sign detection and recognition of traffic signs, with good accuracy and robustness.Traffic sign detection stage, according to the specific color feature of traffic signs(red, blue), this thesis uses color invariants to establish the appropriate color Mixtures of Gaussians probability model. And then calculate the probability of a particular color belongs to a traffic sign for each image pixel by the model, to get the corresponding color probability map. Then the color probability map is transfer to grayscale, in this grayscale, the brighter the area is related to the specific color area. This thesis use maximum stability extremal region algorithm(MSER) to looking grayscale gradation stability in the grayscale, if the picture has traffic sign, then most likely located in the stable region, and use the scale feature of traffic sign to filter these regions to get the final candidate detection regions. Then depending on the shape characteristics of traffic signs, use the histogram of oriented gradients(HOG) feature training support vector machine(SVM) classifier, using the classifier to classify candidate detection area to determine whether the region is the presence of a traffic sign.In traffic sign recognition stage, according to the results of a traffic sign detection, the shape, color and category belongs(warning, prohibition, instructions) of the traffic sign is known, and thus in the identification phase, the main task is to identify differences in the pattern of traffic signs inside. In this thesis, we use traffic signs grayscale image as training samples, use Contrast Limited Adaptive Histogram Equalization algorithm(CLAHE) process samples, eliminating the effect of light on the picture, according to the pattern characteristic of the major categories of traffic signs, we train different recognition network, resulting in better recognition results.Algorithm used in this thesis show through experimental verification is feasible and effective in the detection and classification of traffic sign.