Modeling Driver Behavior During Automated Vehicle Takeovers

Abstract

Driving crashes are a leading cause of death and injuries worldwide. Automated vehicles are expected to reduce these crashes and provide safety benefits. However, the safety of partially automated vehicles is limited by the driver ability to takeover when automation fails. Understanding factors that influence takeover performance is a critical first step in designing safer systems. Additional steps are required to integrate these factors into a design process. One method of integration is through simulation frameworks that join technology with driver models and produce safety-related predictions. Despite the considerable amount of modeling work during manual emergencies, models of automated vehicle takeover behavior are rare. This research addresses this gap by investigating the influential factors and their impacts on takeover performance, identifying promising driver models that accurately capture the impact of influential factors, and developing a comprehensive modeling framework that provides accurate predictions of driver behavior.

This work collected data from a driving simulator experiment to investigate the impact of automation design issues (e.g., silent failure) on driver performance across various transitions of control. Drivers’ takeover time and quality are explored using Bayesian regression models and a significant impact of silent failures on takeover safety, especially in critical events, is found. To capture the effects of each factor, the drivers’ reaction and control maneuver are modeled using visual looming-based models. An evidence accumulation and a piecewise linear model are proposed to predict the drivers’ braking behavior. The steering avoidance is modeled by a looming-based open-loop Gaussian model followed by a closed-loop two-point visual control model for stabilization steering. The developed braking and steering models are leveraged and a holistic algorithm is proposed that ties two parallel evidence accumulators to the developed models to account for the onset of each decision alternative as well as the drivers’ response behavior.

The development of a comprehensive and realistic model that closely matches real-life driver behaviors is vital to assess the safety-related effectiveness of automated systems following a takeover. These evaluations can guide the design of automated technologies and reduce the consequences of failures.