Uncertainty-aware risk assessment and GRU-based risk level map generation for RSU-assisted vehicles

Area of Science:

• Autonomous Driving
• Artificial Intelligence
• Risk Assessment

Background:

• Collision avoidance in autonomous driving faces challenges in environmental perception and risk evaluation.
• Current methods struggle to integrate uncertainty from multimodal trajectory predictions.
• Roadside Units (RSUs) offer enhanced capabilities for comprehensive risk assessment.

Purpose of the Study:

• To propose a cooperative vehicle-infrastructure risk assessment system (CVIRAS).
• To develop an uncertainty-aware risk assessment method for multimodal predictions.
• To enhance autonomous driving safety through improved risk evaluation and trajectory prediction.

Main Methods:

• Developed an uncertainty-aware risk assessment using bivariate Gaussian distribution for collision probability.
• Introduced RLMNet with a pattern-matching attention mechanism for risk level map recognition.
• Proposed E2E-RLMNet for joint training of trajectory prediction and risk assessment.

Main Results:

• RLMNet and E2E-RLMNet demonstrated superior performance in three-class and five-class risk classification.
• The proposed methods effectively mitigated computational burdens from complex multimodal and environmental data.
• Validated on real-world NGSIM and CitySim datasets.

Conclusions:

• The CVIRAS framework significantly improves risk assessment accuracy and efficiency in autonomous driving.
• Integrating RSUs and uncertainty-aware methods enhances the reliability of collision avoidance systems.
• The proposed models offer a robust solution for complex driving scenarios.