A traffic sign recognition method based on deep visual feature

With the development of modern transportation system, the technology of Intelligent Transportation System has attracted more and more interests from research and industry communities. Road traffic sign recognition is one of the most important topics in this field. The traditional methods excessively rely on image morphology, segmentation and various image feature extractions, however most of the methods cannot meet the requirements of the real applications of Intelligent Transportation. Recently, the technique of deep learning has been applied for image recognition and achieves very significant improvement. Inspired by previous work, we propose a new method base on deep visual features in our system to improve the performance of traffic sign recognition. The working process of our method: the first step is to collect and process the images, de-noised by using traditional method and performed the region of interest (ROI), then we determine the key frames in the video according to some features of ROI; for the second step, the Convolutional Neural Network is applied to extract the deep visual features of the images and then the Support Vector Machine(SVM) is employed for image region classification; the final step is to decide which one is the traffic sign according to the outputs of the SVM classification. The experimental results show that our method is robust and works effectively under complex circumstances. Compared with the traditional methods, it can input images directly and extract features independently, also, it achieves the best performance with higher accuracy rate and better real-time on some real world datasets.