Modeling Driver-Automation Interaction using A Naturalistic Multimodal Driving Dataset

Abstract

This paper explores a holistic approach to understanding the impacts of on-road factors on continuous driver-automation interaction in naturalistic environments. We processed and synchronized CAN Bus signals, GPS coordinates, and high-definition videos collected from two drivers in 98 trips (~97 hours) involving Tesla Autopilot engagement collected over three years as part of the MIT Advanced Vehicle Technology naturalistic driving study. Bayesian generalized linear models were trained with the synchronized data and revealed that steering wheel rotations and speed changes, even to a small extent, cause Autopilot disengagement. Road types and experience were also associated with drivers’ probability of using driver assistance, while speed and surrounding vehicles had little impact (when traveling in stable states). This data-driven approach enables a more comprehensive understanding of driver-automation interaction to enhance safety.