UIRAM: An intention-uncertainty-based risk assessment framework for interactive traffic scenarios

Area of Science:

• Autonomous Driving
• Traffic Safety
• Artificial Intelligence

Background:

• Driving behaviors are increasingly diverse, necessitating advanced methods for identifying and quantifying collision risks.
• Current risk assessment tools struggle with complex interactions and predicting uncertain future behaviors of surrounding vehicles.

Purpose of the Study:

• To develop an intention-uncertainty-based risk assessment framework (UIRAM) for interactive traffic participants.
• To enhance the accuracy and robustness of imminent collision detection in heterogeneous traffic environments.

Main Methods:

• UIRAM utilizes a 2D Gaussian conflict domain for efficient candidate selection.
• An intention prediction network fuses environmental context and interaction dynamics to generate uncertainty-aware Gaussian trajectory distributions.
• Collision probability is estimated and combined with consequence severity for a comprehensive risk score.

Main Results:

• UIRAM demonstrated up to a 0.22-point AUC improvement over baseline methods on the FLUID dataset.
• Per-vehicle uncertainty prediction and risk computation were achieved within 0.14s.
• Driver-in-the-loop simulations confirmed UIRAM's risk outputs align with human risk perception.

Conclusions:

• UIRAM effectively assesses risk in interactive traffic scenarios by incorporating intention uncertainty.
• The framework offers a promising solution for improving safety assurance in autonomous driving systems.
• UIRAM's computational efficiency and alignment with human perception highlight its practical applicability.