A deep reinforcement learning algorithm for optimizing safety and efficiency of traffic signals using traffic conflict technique and artificial intelligence-based video analytics.

The advancements in computer vision have opened doors to estimate crash risks in real-time and revisit traffic signal systems to optimize safety and efficiency within a unified framework. While the efficiency reward can be based on operational characteristics, there remains a critical need for methodologies that integrate non-stationary, conflict-type-specific, and cycle-level crash probability estimates into traffic signal design. This study proposes a deep reinforcement learning technique to optimize the safety and efficiency of traffic signals using AI-based video analytics. Specifically, the proposed framework builds on a Deep Q-Network (DQN) that integrates real-time cycle-level crash risks, estimated from a non-stationary Extreme Value Theory model, with traffic delays (waiting times), extracted from a microscopic traffic simulation platform. The proposed framework is trained and tested on two isolated signalized intersections in Queensland, Australia, and compared with the observed adaptive traffic signal control system. The estimated rear-end crash risks from the developed non-stationary extreme value model were validated by using crash frequency estimates against the Poisson confidence bounds of observed crashes. The developed model was utilized in an integrated Deep Reinforcement Learning-based framework to optimize the safety and efficiency of traffic signals. Compared to the observed adaptive traffic signal control system, the proposed Deep Reinforcement Learning-based signal system has been found to reduce crash risk and delay by 69.95% and 33.14% at the Gold Coast Rd - Hope Island Rd intersection, and 87.75% and 37.91% at the Granard Rd - Beaudesert Rd intersection. The trained DQN model has also been found to consistently dissipate the queues without causing excessive delays in any direction. The safety and efficiency weights of around 0.5 have been found to form the optimal policy for traffic signal optimization. The proposed traffic signal design framework has the potential to enhance both safety and efficiency, offering improvements beyond approaches that focus solely on efficiency optimization.