A method for feature extraction of traffic sign detection and the system for real world scene

Traffic sign detection is the significant step before recognizing the class of traffic signs. In the detection, most studies rely on region of interest (ROI) from color information. In practice, however, there is no way to cover the various conditions such as illumination effects or weather conditions. To overcome the problem, this work uses the ROI-free detection by the supervised learning in which the predictor trains the positive examples of traffic sign image and negative examples of non traffic sign image. The proposed method is robust to illumination effects although it searches the traffic sign over the input scene. Because the real world scene often contains occluded or overlapped traffic signs, it is required that the detection algorithm should handle the cases. In this work, we introduce a novel feature extraction method inspired by vision perception theory developed in biological system and by power spectrum in frequency domain. The method was combined with support vector classifier. The proposed method showed accurate classification results (99.32%, 5 fold cross validation) over combined image sets of positive and negative traffic signs samples. Finally, we compared the detection ability of the proposed method and a previous work using ROI on real-world traffic scenes.