Artificial intelligence for traffic signal control based solely on video images

Learning-based traffic control algorithms have recently been explored as an alternative to existing traffic control logics. The reinforcement learning (RL) algorithm is being spotlighted in the field of adaptive traffic signal control. However, no report has described the implementation of an RL-based algorithm in an actual intersection. Most previous RL studies adopted conventional traffic parameters, such as delays and queue lengths to represent a traffic state, which cannot be exactly measured on-site in real time. Furthermore, the traffic parameters cannot fully account for the complexity of an actual traffic state. The present study suggests a novel artificial intelligence that uses only video images of an intersection to represent its traffic state rather than using handcrafted features. In simulation experiments using a real intersection, consecutive aerial video frames fully addressed the traffic state of an independent four-legged intersection, and an image-based RL model outperformed both the actual operation of fixed signals and a fully actuated operation.