A Deep Analysis of the Existing Datasets for Traffic Light State Recognition

Traffic lights classification is a very important task that should be accomplish using computer vision techniques. RADAR or LIDAR sensors are suitable to detect traffic lights. However, they are not able to distinguish between traffic light states. This critical task embraces passengers safety during autonomous driving and it can only be solved using computer vision approaches. In this paper, a wide analysis of the state of the art regarding traffic lights classification is performed. The proposed approach is based on a ResNet architecture and it is compared against a more complex architecture (MobilNet) and also against a traditional feature-based classifier (Random Trees). Due to the importance of high quality training data, a comparative analyisis of the existing datasets related with traffic light classification is presented. The analysis take into account image patch size, number of labels and number of samples for each label.