Traffic lights recognition based on PCANet

Traffic lights recognition is important to make intelligent vehicles safe. Most of existing means to detect and recognize traffic lights focus on color, size and shape of traffic lights, which are great affected by weather and illumination conditions. In this work, we utilize deep learning and SVM classifiers to recognize traffic lights for varying illumination conditions. More specifically, a PCA Network (PCANet) is used to extract features from a set of traffic light images to train SVM to classify the traffic lights. We have analyzed and compared the features extracted by PCANet with HoG´s. The features have excellent characteristic to maintain features of traffic lights. In our expect, experiment results present us a satisfied recognition rate, and the algorithm can provide robust and efficient traffic lights information to support the intelligent vehicle.