AI-Enabled Efficient Traffic Scheduling for Autonomous Vehicles

Traffic congestion not only wastes valuable time but also exacerbates greenhouse gas emissions. We envision a future city where smart traffic management systems are capable of eliminating traffic jams entirely through coordinated traffic planning, with connected autonomous vehicles interleaving each other without collisions. The state-of-the-art automatic traffic plannings focus on improving traffic throughput of single intersections, however, these approaches may not be optimal for multiple intersections. To the best of our knowledge, there is currently no algorithm addressing large-scale irregular multi-intersection road networks for overall traffic optimization. In this work, we propose UrbanTrafficNet, one of the first AI assisted large-scale traffic management plan generation mechanism. The UrbanTrafficNet employs a unique algorithm to encode dynamic vehicle status and complex road network across multiple intersections. This information is then fed into novel deep learning-based models which generate detailed traffic plans for individual vehicles. The UrbanTrafficNet is universally applicable to any complicated layouts of urban road networks and achieves near-maximum traffic throughput. Compared to traditional traffic light scheduling, UrbanTrafficNet reduces stop-and-go percentages by up to 80%, resulting in smoother traffic flow and decrease in greenhouse gas emissions. Extensive real-world experiments validate the effectiveness and efficiency of our proposed system, showcasing its practicality in addressing traffic congestion and environmental concerns.