Risk quantification and prediction of non-driving-related tasks on drivers' critical intervention behavior in autonomous driving scenarios

Abstract

For autonomous driving, drivers’ intervention may be required when vehicles fail or are in a dilemma to detect emergent and unprogrammed events. In such situations, non-driving related tasks may have a great impact on the safety of drivers’ critical intervention behavior thus leading to traffic accidents. Therefore, exploring the impacts of non-driving-related tasks on drivers’ critical intervention behavior, quantifying and predicting the corresponding risks have become important. In this paper, driving simulation experiments are carried out to obtain the vehicle driving state data and visual behavior information of drivers during the autonomous driving scenarios that require critical interventions. To construct the risk quantification model for drivers’ critical intervention behavior, the fuzzy comprehensive evaluation method and the criteria importance though intercriteria correlation (CRITIC) weighting method are employed. Then, for risk prediction, a model is constructed based on the visual behavior information before the occurrences of intervention. Multivariate logistic regression (MLR) and support vector machine are compared. The results show that non-driving tasks significantly postpone driver's critical intervention responses, increasing crash risks of the driving. For prediction, SVM performs better than the MLR in terms of metrics including the precision, the recall, and the overall accuracy. This paper examines the risks during situations requiring drivers’ critical intervention, associated with different non-driving tasks, which has remained much unexplored in the previous research. The methodology of this paper can be applied to smart vehicle systems in alerting vehicles for take-over reactions, with recognizing and predicting potential risks.