A hierarchical intelligent traffic control method for an unsignalized world

Abstract

As the penetration rate of robot vehicles (RV) increases, a new paradigm of controlling the traffic flow has emerged, which uses robot vehicles to control the traffic flow instead of the signal system. Though some algorithms have made some progress on this new paradigm, controlling large-scale unsignalized traffic remains challenging. In this thesis, we study hierarchical intelligent traffic control methods under unsignalized conditions. We divide the unsignalized traffic scenarios into several levels: intersectional scenarios, city-scale scenarios, highway scenarios, and unstructured scenarios. According to different levels, we propose an algorithm to deal with different challenges. The thesis first presents a decentralized reinforcement learning-based method for controlling hybrid traffic at unsignalized intersections. The evaluation results demonstrate that the presented method increases traffic efficiency compared to traffic signal control. Even using only 5% RVs, the presented method can still prevent congestion inside the intersection. To further scale up the traffic scenario into a city-scale map, the thesis proposes an intelligent coordination system for city-scale unsignalized traffic control, which consists of a novel route planning algorithm considering RV shortage. The proposed system can mitigate the imbalance of RV distribution and maintain the minimal RV penetration rate. For the highway scenario, the thesis introduces an intelligent truck system for highway transportation, which aims to mitigate the reality gap of truck simulations. Besides, for unstructured scenarios, the thesis presents a two-stage reinforcement learning multi-UAV collision avoidance approach. However, collision avoidance of multi-UAV is a more challenging task compared to ground vehicles. The validation experiments show that the presented approach can generate time-efficient and collision-free paths under imperfect sensing.

In summary, this thesis presents four methods for traffic control under unsignalized conditions, covering four different levels of traffic scenarios. These four methods can be integrated into a hierarchical intelligent traffic control method for an unsignalized world.