A new integrated collision risk assessment methodology for autonomous vehicles

Area of Science:

• Autonomous Driving Systems
• Artificial Intelligence
• Traffic Safety

Background:

• Real-time risk assessment for autonomous vehicles is challenging due to complex contextual and circumferential factors.
• Existing interaction-aware motion models often assume vehicle communication and are computationally intensive.
• Current methods struggle with large, dynamic datasets in complex environments like dense urban areas.

Purpose of the Study:

• To develop a novel methodology for real-time risk assessment in autonomous driving.
• To integrate network-level and vehicle-based risk estimates using Dynamic Bayesian Networks (DBN).
• To improve the prediction of collision probabilities in dynamic traffic scenarios.

Main Methods:

• Developed a joint framework combining interaction-aware motion models and Dynamic Bayesian Networks (DBN).
• Utilized machine learning classifiers for real-time network-level collision prediction.
• Incorporated collision predictions into the DBN for integrated real-time risk assessment.

Main Results:

• The integrated approach demonstrated an enhancement of up to 10% in interaction-aware models during collision-prone traffic.
• Machine learning classifiers effectively predicted network-level collisions in real-time.
• The methodology successfully predicted collision probabilities in complex, dynamic traffic scenarios.

Conclusions:

• A well-calibrated collision prediction classifier is crucial for enhancing autonomous vehicle risk perception.
• The proposed integrated risk assessment methodology offers improved performance in complex traffic environments.
• This approach addresses limitations of current methods in handling large sequential data and computational demands.