Modeling interactive crash avoidance behaviors: A multi-agent state-space transformer-enhanced reinforcement learning...

Area of Science:

• Traffic Safety Research
• Artificial Intelligence in Transportation
• Human-Machine Interaction Modeling

Background:

• Vehicle-pedestrian interactions at urban intersections are critical for traffic safety.
• Near-miss scenarios present complex decision-making challenges for road users.
• Existing models struggle to capture the dynamic and interactive nature of these events.

Purpose of the Study:

• To model the interactive crash avoidance behavior of vehicles and pedestrians in near-miss scenarios.
• To develop a framework capable of learning from rare, safety-critical events.
• To enhance the realism of traffic safety simulations.

Main Methods:

• Utilized high-resolution trajectory data from unmanned aerial vehicles (UAVs).
• Proposed a multi-agent state-space Transformer enhanced deep deterministic policy gradient (MA-SST-DDPG) framework.
• Integrated state-space models for temporal dependencies and Transformers for feature prioritization.

Main Results:

• The MA-SST-DDPG framework effectively simulated realistic evasive behaviors in near-miss scenarios.
• Model demonstrated superior performance and generalizability across datasets.
• Higher speeds increased conflict rates; yielding behavior depended on relative speeds.

Conclusions:

• The developed framework accurately replicates real-world near-miss dynamics.
• Findings provide insights into factors influencing vehicle-pedestrian interactions and crash avoidance.
• Enables development of advanced safety-aware simulations for proactive crash prevention.