A deep reinforcement learning algorithm for optimizing safety and efficiency of traffic signals using traffic...

Area of Science:

• Artificial Intelligence
• Traffic Engineering
• Computer Vision

Background:

• Traditional traffic signal systems often prioritize efficiency over safety.
• Integrating real-time, cycle-level crash risk estimates into signal design is crucial but challenging.
• Existing methods lack the ability to handle non-stationary and conflict-type-specific crash probabilities.

Purpose of the Study:

• To develop and evaluate a deep reinforcement learning framework for optimizing traffic signal safety and efficiency.
• To integrate real-time crash risk estimations with traffic delay data for enhanced signal control.
• To address the need for methodologies that consider non-stationary, conflict-type-specific, and cycle-level crash probabilities.

Main Methods:

• Utilized AI-based video analytics for real-time crash risk estimation.
• Employed a Deep Q-Network (DQN) integrated with a non-stationary Extreme Value Theory model.
• Incorporated microscopic traffic simulation to extract traffic delays (waiting times).
• Trained and tested the framework on two real-world intersections in Queensland, Australia.

Main Results:

• The proposed Deep Reinforcement Learning (DRL)-based signal system reduced crash risk by up to 87.75% and delay by up to 37.91%.
• Validated rear-end crash risk estimates against observed crash frequencies using Poisson confidence bounds.
• The trained DQN model effectively dissipated traffic queues without causing excessive delays.

Conclusions:

• The proposed DRL framework significantly enhances both traffic safety and efficiency compared to traditional adaptive systems.
• A balanced weighting (around 0.5) between safety and efficiency optimizes the traffic signal control policy.
• This AI-driven approach offers a promising direction for future traffic signal design, improving upon efficiency-only optimization methods.