Behavioral decision-making and safety verification approaches for autonomous driving system in extreme scenarios

Autonomous vehicles are crucial for improving traffic efficiency and reducing accidents, yet the complexity of driving scenarios and behavioral uncertainty pose challenges for decision-making. Recent research integrates virtual simulation with decision algorithms to enhance system intelligence and performance. Nonetheless, the potential hazards associated with extreme weather conditions are often overlooked. To mitigate this issue, this paper proposes a Bayesian network decision-making model based on hazard probability inference. The model enables the driver assistance system to take over the control of the vehicle in extreme scenarios and dynamically adjust decision strategies based on the potential hazard values under multivariate data. First, safety elements of sporadic hazardous scenarios are extracted using the Accidental and Catastrophic Automatic Driving Scenario Modeling Language and used as nodes to construct a Bayesian network for inferring potential driving hazards. Second, a Bayesian decision-making model is designed based on the semantic hierarchy of the autonomous driving system domain ontology, aiming to derive the optimal driving behavior for the current vehicle in extreme scenarios. The safety of these decisions is verified using the UPPAAL-SMC statistical model checker. Finally, the model’s validity is confirmed through a real-world autonomous vehicle accident, with results indicating more rational decisions and improved safety performance.