Using Multilevel Hidden Markov Models to Understand Driver Hazard Avoidance during the Takeover Process in Conditionally Automated Vehicles

Abstract

Ensuring a safe transition between the automation system and human operators is critical in conditionally automated vehicles. During the automation-to-human transition process, hazard avoidance plays an important role after human drivers regain the vehicle control. This study applies the multilevel Hidden Markov Model to understand the hazard avoidance processes in response to static road hazards as continuous processes. The three-state model—Approaching, Negotiating, and Recovering—had the best model fitness, compared to the four-state and five-state models. The trained model reaches an average of 66% accuracy rate on predicting hazard avoidance states on the testing data. The prediction performance reveals the possibility to use the hazard avoidance pattern to recognize driving behaviors. We further propose several improvements at the end to generalize our models into other scenarios, including the potential to model hazard avoidance as a basic driving skill across different levels of automation conditions.