Deep Convolutional Traffic Light Recognition for Automated Driving

Abstract

Robust traffic light detection and state recognition is of crucial importance on the path to automated vehicles. However, the mere classification of the signaled states does not suffice at complex multi-lane intersections. Rather, a complete understanding of the intersection, but at least the recognition of additional information (like arrows displayed on the traffic lights) is necessary. In this work, we developed a unified deep convolutional traffic light recognition system on the basis of the Faster R-CNN architecture, which is able to not only detect traffic lights and classify their state, but also distinguish their type (circle, straight, left, and right). An in-depth analysis of its performance on the large and diverse DriveU Traffic Light Dataset shows an overall detection performance of 0.92 Average Precision for traffic lights of width greater than 8 px. Additionally, other kinds of traffic lights, e.g. pedestrian lights, have been identified as main cause of false positives. Moreover, we evaluated the usefulness of the developed system to assess the traffic light states for all present driving directions revealing inconsistencies among multiple detections in single images.