Evaluating Driver Features for Cognitive Distraction Detection and Validation in Manual and Level 2 Automated Driving

Abstract

Objective This study aimed to investigate the impacts of feature selection on driver cognitive distraction (CD) detection and validation in real-world nonautomated and Level 2 automated driving scenarios. Background Real-time driver state monitoring is critical to promote road user safety. Method Twenty-four participants were recruited to drive a Tesla Model S in manual and Autopilot modes on the highway while engaging in the N-back task. In each driving mode, CD was classified by the random forest algorithm built on three “hand-crafted” glance features (i.e., percent road center [PRC], the standard deviation of gaze pitch, and yaw angles), or through a large number of features that were transformed from the output of a driver monitoring system (DMS) and other sensing systems. Results In manual driving, the small set of glance features was as effective as the large set of machine-generated features in terms of classification accuracy. Whereas in Level 2 automated driving, both glance and vehicle features were less sensitive to CD. The glance features also revealed that the misclassified driver state was the result of the dynamic fluctuations and individual differences of cognitive loads under CD. Conclusion Glance metrics are critical for the detection and validation of CD in on-road driving. Applications The paper suggests the practical value of human factors domain knowledge in feature selection and ground truth validation for the development of driver monitoring technologies.