A consequence-aware driving risk field framework with data-driven calibration for high-risk scenario identification

Area of Science:

• Traffic safety engineering
• Computational modeling
• Transportation systems analysis

Background:

• Current traffic risk assessment methods lack a unified representation consistent with crash consequences.
• Existing approaches struggle to integrate diverse factors like traffic interactions, road environment, and driving behavior effectively.
• A need exists for instantaneous risk scoring and high-severity crash scenario identification.

Purpose of the Study:

• To propose a Driving Risk Field (DRF) modeling framework incorporating crash consequences for unified risk representation.
• To develop an instantaneous risk score for any driving state, enabling identification and ranking of high-severity crash scenarios.
• To enhance model interpretability and internal consistency through data-driven calibration.

Main Methods:

• Developed a DRF framework integrating scenario-level risk factors and vehicle-specific characteristics.
• Implemented a parameter calibration method using real-world crash and traffic simulation data.
• Optimized key risk field parameters using a differential evolution algorithm.

Main Results:

• The DRF model captures distance decay and velocity amplification effects, with a stable composite risk metric distribution.
• Data-driven calibration improved regression metrics and enhanced the identification/ranking of high-severity crash samples compared to XGBoost.
• The model demonstrated sensitivity to weather conditions and road types, reflecting their influence on risk scores.

Conclusions:

• The proposed DRF model offers a unified, interpretable paradigm for crash-consequence-calibrated high-risk scenario scoring.
• The framework supports effective risk-scenario screening and advanced traffic safety analysis.
• Findings highlight the model's capability to reflect internal risk score influences from various scenario factors.