Research On Traffic Sign Recognition Algorithm

Traffic sign recognition (TSR) is an important part of the Intelligent Transportation Systems (ITS). It has important application in unmanned driving and driver assistance systems. It’s very challenging to recognize traffic signs fast and exactly due to various lighting and weather conditions, shape distortion, occlusion, camera angle changes and other interference factors in the complex natural scenes.The traffic sign recognition approach based on image color and shape matching is simple and fast, but it is difficult to achieve the requirements of application because of its low recognition rate. The Learning-based approach effectively improves the recognition accuracy through image feature extraction, classifier design and training. However, further research of feature effectiveness, recognition speed, etc. is also needed. According to the significant shape and internal structure characteristics of traffic signs, an effective traffic sign recognition algorithm based on multi-feature fusion is proposed. The experiment results show that the proposed algorithm has high recognition accuracy and fast recognition speed.The main achievements are described as following:1. As the traffic images are obtained from natural scenes, the pre-processing of original images can reduce the impact of various shape, size and illumination, enhance image quality and reduce computational complexity of system. The main operations include gray and gray-normalization, size-normalization and segmentation of the region of interest (ROI).2. To reduce the influence of cluttered background and other redundant information, this paper obtains effective feature description of traffic signs by feature extraction. According to the local edge, contour, texture and other characteristics of traffic signs, three kinds of feature descriptor:Gabor filter, HOG and LBP are researched. And though the fusion of three features, the proposed algorithm compensates for the disadvantage that a single feature is difficult to comprehensively describe traffic signs and improves system recognition rate.3. The Support Vector Machine theory is researched, and the pros and cons of SVM with different kernel function are analyzed. As linear SVM has the advantage of fast and exactly recognizing the objects with large category and high feature dimension, this paper uses one-against-rest linear SVM classifier to classify traffic signs. 4. Through experiments on large dataset, the recognition performance of a single feature and the fusion of multiple features is compared and analyzed. The results showed that:the three features indicate complementariness and their fusion significantly improves the recognition accuracy. Based on the fusion of three features and linear SVM classifier, the algorithm obtains a high accuracy of98.65% with a low computational complexity.