Classifying Older Drivers’ Gaze Behaviour during Automated versus Non-Automated Driving: A Preliminary Step towards Detecting Mode Confusion

Abstract

Mode confusion occurs when the driver of an automated vehicle (AV) is confused about the active operating mode of the AV and therefore, their responsibilities as the driver. Mode confusion is a serious safety concern, especially for cohorts who are less familiar with AVs and/or who are more likely to have poorer situational awareness during automated driving such as older adults. In this article, we propose a design framework for driver state monitoring systems that can potentially be used to detect older drivers’ mode confusion by inferring drivers’ perceived AV mode using gaze behaviour data. As a proof-of-concept for an AV with two modes, the efficacy of the proposed framework is tested by applying it on a gaze behaviour dataset collected from 29 older drivers (65+) during simulated non-automated and simulated fully automated drives. The proposed framework utilizes classification models trained on features extracted from the gaze behaviour data. Among 25 features, the mRMR (maximum relevance minimum redundancy) feature ranking framework ranked our proposed feature of weighted static gaze entropy as having the highest relevance with the driver’s perceived AV modes while having the least redundancy with the rest of the selected features. An ensemble stacking model achieved the highest classification performance with an average accuracy of 73% and an average AUC score of 80%. The results indicate that gaze behaviour features can distinguish between the driving scenarios of automated and non-automated as perceived by the drivers. While the dataset does not include confirmed instances of driver’s mode confusion and therefore, the framework testing provides preliminary results towards a proof of concept, this work provides a foundational model for future studies in which actual data from confirmed mode confusions are intentionally introduced or measured. In turn, this study can inform future designs of driver state monitoring systems aimed to detect and mitigate the safety risks of driver’s mode confusions in automated vehicles.