Take-Over Time Modeling and Prediction for Conditional Driving Automation

Abstract

Autonomous vehicles are designed to enhance the overall driver safety by taking the driver out of the loop. However, the autonomous vehicles that are currently available on the market still require that the driver is available to take-over control in case the automation fails (ex. Tesla AutoPilot). If the driver is not able to take-over in time and resume control of the vehicle, unsafe consequences, such as crashes, could occur. Extensive research has been done in modeling the driver behavior around these take-over events to mitigate the crash risk. However, these models use complex or unrealistic data sources (eye gaze, external environment, heart rate, video data, etc.) and are unable to handle time series data for updated predictions as new data is acquired. The main focus of this research is to develop a modeling framework for take-over events in conditional driving automation with the focus on predicting the remaining take-over time. This research also explores the impact of individual driver differences on model performance through the use of online-learning. A hidden semi-Markov model (HSMM) was proposed for modeling the take-over time and the modeling framework was assessed on data collected from a driving simulator study. The proposed model is able to accurately predict the remaining take-over time, capture the uncertainty in the model prediction, and handle variable length time series data when compared to state-of-the-art regression prediction models. By developing a framework to model the take-over time of drivers, autonomous vehicle manufacturers can mitigate the risk of vehicle control handovers as the technology continues to be developed and changed over time.