Image-Based Learning to Measure the Stopped Delay in an Approach of a Signalized Intersection

Traffic delays are inevitable when evaluating the performance of a signalized intersection, but these delays cannot be directly measured in the field based on existing spot detectors. Traffic-light controllers have adopted a reinforcement learning (RL) algorithm, which is currently prevalent in the field of study and requires real-time measurement of traffic delays to derive the state and reward for each time period. No RL-based study, however, has provided a robust way to measure traffic delays. In order to bridge the gap, we devised a convolutional neural network (CNN) to directly measure traffic delays from video footage in an end-to-end manner. The proposed methodology proved superior to both a state-of-the-art vision technology and an analytic formula that has widely been used to estimate delays. Furthermore, a robust method to secure labeled data without human input was suggested based on a cycle-consistent adversarial network (CycleGAN).