Abstract:
|
Accurate prediction of driver risk is challenging due to crash rarity and individual driver variability. The rapid development of in-vehicle data collection technology provides great potential to improve driver risk prediction through kinematics data. In this study, we propose a decision-adjusted approach to develop the optimal kinematic-information-incorporated driver risk prediction model. Specifically, we tune the threshold values for elevated longitudinal and lateral acceleration and other model parameters through a decision-based objective function. The merits of the proposed method are validated through the SHRP2 naturalistic driving data. Our decision-adjusted model improves the prediction precision by 43%-10% relatively compared to the baseline model, under the decision rules of identifying the riskiest 1%-20% drivers. It also outperforms general model selection rule such as AUC, especially in imbalanced data scenarios. The study confirms that using kinematic information can improve individual driver risk prediction, and the optimal thresholds of deceleration and lateral acceleration really depends on the decision rule.
|