Traffic Signal Lights Recognition

Accurate traffic signal lights recognition has not only played an invaluable role in the intelligent transportation systems, unmanned technology, but also makes it possible to drive a car for the color-blind. So far, there is not much research on traffic signal light recognition technology, and most of existing algorithms are characterized by color and shape features of traffic signal lights. As we all know, High-quality images are generally the requirement for the algorithms based on image processing, which are not suitable for the traffic signal lights recognition in the natural environment. Therefore, in order to cater for a variety of natural environment, the machine learning method are proposed. Overall, this paper carried out the work from the following aspects:Firstly, we propose a S channel that is more clear than the original RGB image. The S channel is achieved by transforming the original RGB image. Compared with the traditional CIE Lab, HSV, and HIS color space, the calculation of the transformation to S channel which is done in linear time could be saved to a great extent. More importantly, each sub-channel could not be affected by changes in illumination because they are linearly independent. Candidate regions of traffic lights are extracted by Maximally Stable Extremal Regions and represented by SIFT feature vectors. Then the SVM classifier is used to recognize the candidate regions. Finally the Camshift algorithm based on S channel can be used to track traffic signal lights, And a decision program is designed to correct the recognition results, which will further improve the traffic signal lights recognition accuracy.A lot of comparative experiments with other different methods were carried out on our databases and the traffic signal lights video sequence database and our own video traffic signal light image database of Robotics Centre of MinesParisTech to verify the performance of the proposed method. We have tried some features, such as Hog, Gabor, LBP, SIFT, Transformed color sift, Rgsift, and some classifiers, such as SVM, Adaboost, LR classifier and so on. The experimental results show that our method is effective in the real-time traffic signal light recognition under a variety of natural environment.