Study Of Traffic Sign Text Detection From Natural Scenes

Intelligent Transportation Systems have significant application prospect to improve traffic efficiency and reduce vehicle accidents. Traffic sign detection and recognition system are an important part of the intelligent transportation system, and it is not only a hot issue, but has made great progress. Among various traffic signs, signs with text (character traffic signs) provide rich and important road information, which is vital to unmanned vehicle and driver assistance system. However, the research on traffic sign text detection is relatively little. Because traffic signs in natural scenes are susceptible to different weather or light intensity, and there may be occlusion problem; additionally, irregular arrangement, inhomogeneous spacing and different size of characters are very common on the signs. Therefore, traffic sign text detection from natural scenes is challenging.This paper achieves a fast and accurate traffic sign text detection algorithm from natural scenes. The main work primarily consists of accurate traffic sign detection and traffic sign text detection. First, color segmentation and morphological operations are sequentially utilized to obtain candidate regions; the interior of candidate region is filled to get the contour of traffic sign, and edge orientation histogram is employed to describe the shape of traffic sign, then a linear SVM eliminates non-traffic signs and get the character traffic signs. Additionally, affine transformation is employed for the correction of tilting signs. Finally, K-means clustering algorithm in the S channel achieves image binarization, and multi-scale character extraction and character strokes merging are combined to extract text on the signs.Because there is not a uniform database in this field, our research team establishes one and completes manual labeling of all character traffic signs and texts in the database. Experiment results on a large amount of images show that the proposed method is effective.