An interpretable stacking ensemble learning model for visual-manual distraction level classification for in-vehicle interactions

Recognizing the level of driver distraction during the execution of secondary tasks within the intelligent cockpit is crucial for ensuring a seamless interaction between human drivers and intelligent vehicle systems. To address this issue, this paper proposes a framework for recognizing driver distraction levels that integrates clustering, classification, and interpretability. First, Feature Selection with Optimal Graph(SF²SOG) is employed to identify discriminative features from the data facilitating dimensionality reduction. Following this, Agglomerative Clustering are employed to classify the gathered unlabeled data on driving distraction behaviors into three distinct categories. Additionally, a heuristic-based stacking ensemble model is introduced to identify the levels of driver distraction, using the factors that influence these levels as input parameters for classification. To improve the effectiveness of the stacking model, three diverse classifiers adaptive boosting (AdaBoost)、random forest (RF) and extreme gradient boosting (XGBoost) are selected to serve as the base models, while a relatively simple yet accurate model logistic regression (LR) is used as the meta-classifier. Finally, Shapley Additive exPlanations (SHAP) is employed for interpretability analysis. Notably, heuristic-based stacking ensemble model achieved a commendable accuracy rate of 96.25%, highlighting its significant advantage. Further analysis shows that higher maximum pupil diameter (maxPD) and mean pupil diameter (meanPD) indicate increased distraction, while greater glance frequency and lane deviation reflect reduced situational awareness and control. These findings are crucial for reducing accidents and enhancing driving safety in the era of intelligent vehicles.