DeepTLR: A single deep convolutional network for detection and classification of traffic lights

Reliable real-time detection of traffic lights is a major concern for the task of autonomous driving. As deep convolutional networks have proven to be a powerful tool in visual object detection, we propose DeepTLR, a camera-based system for real-time detection and classification of traffic lights. Detection and state classification are realized using a single deep convolutional network. DeepTLR does not use any prior knowledge about traffic light locations. Also the detection is executed frame by frame without using temporal information. It is able to detect traffic lights on the whole camera image without any presegmentation. This is achieved by classifying each fine-grained pixel region of the input image and performing a bounding box regression on regions of each class. We show that our algorithm is able to run on frame-rates required for real-time applications while reaching notable results.