Using Machine Learning and Multiple Physiological Measures to Estimate Drivers’ Cognitive Load During Non-Driving-Related Tasks in Automated Driving Under the Effect of Silent Automation Failures

Abstract

With the development of automated driving technology, drivers could temporarily engage in non-driving-related tasks (NDRT), such as listening to music, taking a call, and texting. However, once the automatic driving system fails, the high cognitive load of NDRT may deteriorate the driver's takeover performance, thereby threatening driving safety. Therefore, estimating the driver’s cognitive load during NDRT is crucial to traffic safety. Studies in the literature usually measured the cognitive load of NDRT based on a take-over request (TOR) paradigm in which the driving system issued a takeover request before the failure occurred. However, a recent report showed that most takeover requests were initiated by the driver rather than the system. Besides, it has also been pointed out that that currently used NDRT, such as N-back task, are not natural tasks in real driving scenarios. For another thing, it is promising to use machine learning techniques to estimate a driver’s cognitive load from physiological measures in real-time. However, it remains further explore how accurate these techniques could be and whether they can detect the variation of cognitive load in NDRT when the driving system may fail to issue a takeover request (i.e., silent automation failures).

To address these issues, this proposal aims to:

1. Within the scope of "using machine learning to estimate cognitive load from physiological measures", systematically review and summarize the types of machine learning models, types of physiological measures, and knowledge gaps in the literature;
2. Compare and analyze the classification accuracy of commonly used machine learning models and feature selection techniques in estimating the cognitive load of NDRT by using a multimodal dataset for driver cognitive load estimation (eDREAM Dataset);
3. Examine the effect of silent automation failures on driver’s cognitive load during NRDT by using the NASA-TLX method and a proposed machine learning method;
4. Develop a machine learning approach to estimate driver’s cognitive load level during nature NDRT by employing the physiological measures obtained in the N-back test.