Modeling decision-making during unprotected left turns using interpretable deep learning and uncertainty...

Area of Science:

• Traffic Safety
• Human-Computer Interaction
• Autonomous Systems

Background:

• Unprotected left turns are complex driving scenarios requiring unique decision-making.
• Existing models often neglect information variability and intrinsic decision mechanisms.
• Understanding driver decision-making under uncertainty is crucial for road safety.

Purpose of the Study:

• To analyze driver decision-making in unprotected left turns through the lens of decision uncertainty.
• To explore the relationship between decision uncertainty and driving safety.
• To identify key variables influencing left-turn decisions and quantify uncertainty.

Main Methods:

• Conflict area calculation to identify interaction events.
• Transformer model and Shapley Additive Explanations to determine key decision variables.
• Jensen-Shannon divergence to quantify decision-making uncertainty.

Main Results:

• Left-turning vehicles prioritize static variables (waiting time, vehicle type); oncoming vehicles focus on dynamic variables (time to stop line, speed difference).
• Time pressure increases emphasis on lateral speed and yaw angles.
• Higher uncertainty correlates with longer negotiation times, shorter post-encroachment times, and increased likelihood of emergency braking.

Conclusions:

• Decision uncertainty is a critical factor in unprotected left-turn safety.
• Insights inform autonomous vehicle decision-making frameworks for safer navigation.
• Driver behavior in unprotected left turns is influenced by a complex interplay of static and dynamic variables.