A Quantitative Prioritization Assessment Method for Autonomous Vehicle ODD Extension Based on Interpretable Machine Learning

To solve the path ambiguity and decision subjectivity in operational design domain (ODD) extension, a crash data-driven quantitative prioritization assessment framework is investigated. Data support and decision bases for safety capability iterations of autonomous driving systems in complex environments are provided by this framework. To address the long-tail distribution defect of severe crash samples, random forest and extreme gradient boosting (XGBoost) classification models are constructed. The synthetic minority over-sampling technique (SMOTE) is introduced to optimize the sample distribution. The dataset is rebalanced at a 1:20 classification weight ratio to ensure the identification capability for critical minority samples like severe crashes. To break the "black box" limitation of traditional models, the Shapley additive explanations (SHAP) method is employed to deconstruct internal decision-making mechanisms. Furthermore, the nonlinear effect thresholds of environmental feature variables on crash severity are quantified. A marginal effect model for ODD extension is established based on the control variates method and principal component analysis (PCA). By simulating the boundary expansion of a single ODD variable, the improvement in theoretical crash coverage under various extension strategies is calculated. The results indicate that the identification performance of classification models for critical severe crash samples is significantly enhanced after SMOTE processing. While an 83.4% overall accuracy is maintained, the sensitivity of the random forest model is increased from 17.8% to 90.5%. The area under the receiver operating characteristic (ROC) curve (AUC) is increased to 0.922, and high-risk features are captured more precisely. The highest marginal safety benefit among candidate dimensions is yielded by extending to unstructured road grades, as revealed by quantitative data. Theoretical coverage of ordinary and severe crashes is significantly increased by 15.7% and 13.0%, respectively, when expanding ODD from structured highways to state and county roads. Vehicle collision rates under emergency braking in extended scenarios like low road friction and low illumination are verified by CARLA simulation experiments. It is observed that collision rates are increased from 5.0% in the baseline group to 45.0% and 20.0%, respectively. Meanwhile, the average time to collision (TTC) is shortened by 1.3 seconds. The physical safety performance degradation trend is highly consistent with the high-risk feature ranking derived from the quantitative assessment model. It is proven by these quantitative indicators that the priority ODD extension direction with the greatest safety value is reliably identified by the proposed framework.