Detection and Recognition of Alert Traffic Signs

Traffic signs provide drivers important information for safety and efficient navigation. Automatic detection and recognition of traffic signs inevitably become more popular. In this paper, an efficient algorithm/platform is presented to achieve automatic alert traffic signs detection and recognition. Histogram of Gradient (HOG) is adapted to extract features and an over-complete set of 1680 features is designed. A cascade classifier for each sign is trained and built with Support Vector Machine (SVM) as the single stage classifier. To encode the color information, features from different layers in RGB space are combined into a single vector as the feature descriptor. Furthermore, color segmentation is performed to reduce the search regions and a specially designed integral image is used to extract features in a look-up manner. Experimental results show that our system can achieve invariance to illumination, scale and pose. The smallest detectable size is 14!14. The average detection rate is around 98% and the false positive rate is around 1.6%. The processing time for a typical 640!480 image is around 7-9 seconds.