Open Collections

Abstract

Adaptive traffic signal control (ATSC) strategies hold significant promise for enhancing the efficiency and safety of signalized intersections, particularly in the era of connected vehicles (CVs) that provide real-time data on vehicle positions and trajectories. While numerous ATSC algorithms have been developed to optimize traffic efficiency, focusing on minimizing delays or maximizing throughput, they usually ignore traffic safety. This limitation is primarily due to a lack of tools capable of real-time safety evaluation of signalized intersections. However, recent research has introduced innovative tools and methodologies that enable real-time safety assessments, paving the way for more comprehensive ATSC systems that balance efficiency and safety objectives. Building on this foundation, this thesis advances the real-time optimization of safety and mobility in ATSC systems within a connected vehicle environment. First, a novel ATSC algorithm for real-time safety optimization was developed. The algorithm utilizes a traditional reinforcement learning (RL) approach (i.e., Q-learning) and recently developed extreme-value-theory (EVT) real-time crash prediction models to establish the safety rewards. In these models, a real-time safety measure, namely the return level of a cycle (RLC), is used to represent the crash risk at each signal cycle using dynamic traffic parameters. The algorithm was validated using real-world traffic video data from two signalized intersections in British Columbia. The algorithm outperformed the traditional actuated signal control (ASC) systems and proved effective even under low market penetration rates of CVs. Second, a multi-objective self-learning ATSC algorithm was developed using RL. Trained in a microsimulation environment and validated with real-world traffic data, the algorithm optimizes crash risk and mobility in real time using CV data. The algorithm demonstrated superior safety and mobility performance compared to traditional ASC, showcasing its practicality and scalability for real-world deployment. Lastly, two multi-agent, decentralized, and centralized deep RL ATSC algorithms were introduced for corridor-level safety improvement. The safety reward was represented by using extreme EVT real-time crash risk prediction models. The effectiveness was then verified using real-world data for a corridor of three signalized intersections. Findings demonstrate that the introduced safety-oriented signal control optimization algorithm can substantially improve traffic safety while maintaining or even enhancing traffic mobility.