Study On Traffic Panel Layout Understanding From Natural Scenes

As vehicle quantity and population density rise constantly in cities, traffic jam is increasingly severe. Playing an important role in improving traffic efficiency, reducing traffic accidents, etc., intelligent traffic system has been an active research topic, in which traffic sign detection and recognition is one of key issues. At present, traffic sign recognition is mainly concentrated on the classification of traffic sign and the research of traffic panels, which contain characters and symbols, is still a relatively remote area. However, traffic panels provide rich road information, which is vital to traffic safety and traffic efficiency. Therefore, the understanding of traffic panels has significant application prospects and practical values.Traffic panel detection and location in natural scenes, text detection and location on traffic panels and symbol detection and recognition on traffic panels are studied in this thesis, to realize the goal of layout understanding of traffic panels in natural scenes. The main works in this thesis include three parts as follows:(1) Traffic panel detection and location from natural scenes. As blue is the predominant background of traffic panels, color is the most significant feature of traffic panels. As a consequence, traffic panels are located coarsely by color information and located precisely based on histogram of oriented gradient and support vector machines in this thesis.(2) Text detection and location on traffic panels. Hough transform is used for geometric correction of traffic panel images at first. Then, OTSU threshold is improved for binarization and text detection is completed by connected component analysis based on geometric information of text regions. My given block-division based OTSU threshold can offer the reliable guarantee for text detection due to good binarization effect.(3) Symbol detection and recognition on traffic panels. Traffic panels contain not only text information but also symbol information, which provides driving rules and is significant for unmanned driving and driver assistant system. Consequently, symbol detection and recognition on traffic panels are studied in this thesis. Symbols are detected based on connected component analysis after preprocessing, including distortion correction, etc. Symbols are recognized using my given method based on a model named bag of spatial visual words and a combination of support vector machines and random forests.For testing the performance of algorithm, this dissertation collects 3023 traffic panel images in Chinese cities, of which 1015 images are selected randomly for testing our proposed algorithm. The detection precision of traffic panel is 96.89%, the detection precision of text is 98.47% and the detection precision of symbol is 93.40%. The recognition precision of 27 symbols, which are divided into 5 categories, is 94.29%. Experimental results shows the effectiveness of our proposed algorithm.