Identifying factors affecting driver takeover time and crash risk during the automated driving takeover process

This study aims to develop prediction models of driver takeover time and crash risks during the automated driving takeover process. A driving simulator experiment was conducted to collect vehicle trajectory and driver behavior data. The random-parameter duration model was first built to model driver takeover time. Results indicated that young drivers, novice drivers, takeover request lead time, and traffic volume had varying impacts on takeover time due to the unobserved heterogeneity. Then, an explainable machine learning model was utilized to predict and explore various predictors' impacts on takeover crashes. Validation results revealed that the developed model provided satisfactory accuracy in predicting crashes. SHAP was used to interpret the estimated results by examining contributory factors' main effects and interactive effects on crash risks. Takeover crash risk is positively correlated with vehicle speed, takeover time, maximum lateral acceleration, traffic volume, and tasks of watching videos and playing games. Additionally, takeover request lead time and the maximum longitudinal deceleration were found to affect crash risks negatively. Research findings shed insights into modeling takeover time and predicting crash risks during the takeover process, and highlight the importance of considering the heterogeneity of drivers when designing automated driving systems (ADS) to improve driver takeover performance.