Study On Detection Algorithm Of Degraded Traffic Sign In Natural Scenes

As an important part of an Intelligent Transportation System (ITS), Traffic Sign Recognition (TSR) system provides the driver with the position and the category of the traffic signs in the current field of vision. As the prerequisite of traffic sign recognition, accurate detection results can greatly improve the efficiency and accuracy of recognition results. The main challenges of traffic sign detection include varying illumination, partial occlusions, shadow, geometric distortions, similar background color and etc. Based on the analysis of these degraded traffic signs, further studies were performed on image enhancement, segmentation, feature extraction and classification in this thesis. The main contributions are as follows:Firstly, a threshold enhancement algorithm is proposed in the normalized RGB color space according to the varying illumination, shadow, achromatic signs and etc. Due to the high correlation among the three color components, it is difficult to find the correct thresholds in the RGB color space using empirical methods. In the normalized RGB space, the normalization of the Red, Green and Blue components with respect to the sum of the three components is used to eliminate the effects of varying illumination and shadow. For each test image, two enhanced images are produced for chromatic and achromatic traffic signs respectively. The proposed method can effectively highlight the traffic sign areas in the images and overcome the influences of varying illumination on traffic signs detection.Secondly, the Maximally Stable Extremal Regions (MSER) algorithm is used at the stage of image segmentation. The MSER algorithm can effectively deal with cases with similar background color and multiple signs clustered. But the method also brings about too many candidate regions, and its calculation is very complex. The parameters of MSER algorithm and the selection of regions are optimized in this thesis. Through the optimization algorithm, the number of candidate regions is reduced and the efficiency is improved.Thirdly, a novel classification method based on orientation gradient histogram feature and linear support vector machine is proposed. The method can effectively deal with cases with partial occlusions. The traditional HOG+SVM method calculate the score of the whole candidate region, but the occlusion would reduce the score. The method divides Histogram of Oriented Gradient (HOG) features into 16 blocks and gets the score of each block. Then the occluded block is eliminated according to its score to improve the whole score of the candidate region. The method can reduce the effect of partial occlusions of traffic sign detection and improve the detection rate.In the thesis, a traffic sign detection method has been developed based on a careful study on the characteristics of degraded traffic signs. The evaluations results show that the developed method can achieve a high accuracy and robustness for the detection of the degraded traffic signs.