Predicting Takeover Performance in Conditionally Automated Driving

Abstract

In conditionally automated driving, drivers decoupled from operational control of the vehicle have difficulty taking over control when requested. To address this challenge, we conducted a human-in-the-loop experiment wherein the drivers needed to take over control from an automated vehicle. We collected drivers’ physiological data and data from the driving environment, and based on which developed random forest models for predicting drivers’ takeover performance in real time. Drivers’ subjective ratings of their takeover performance were treated as the ground truth. The best random forest model had an accuracy of 70.2% and an F1-score of 70.1%. We also discussed the implications on the design of an adaptive in-vehicle alert system.