Classification of driver fatigue in conditionally automated driving using physiological signals and machine learning

Abstract

Conditionally automated vehicles (Level 3 SAE) are emerging on the roads, but long periods without engaging in a non-driving-related task can reduce drivers’ vigilance. This study aims to determine whether driver fatigue can be accurately predicted using physiological signals and machine learning (ML) techniques in such a context. 63 young drivers completed two separate conditional automated drives of 30 min each, in either a rural or a urban area. Half of them had been mildly sleep-deprived the previous night (slept less than six hours). Electrocardiogram (ECG), electrodermal activity (EDA), and respiration were collected, along with subjective measures of sleepiness and affective state. Using ML, sleep deprivation, driving environment, and sleepiness could be predicted from physiological features with an accuracy of 99%, 85%, and 73% respectively. Signal segmentation increased model accuracy, and EDA features were the most predictive. The differences between the results obtained from statistical analyses of sleepiness measures and the accuracy achieved by ML models are discussed. The results of this empirical study indicate that even mild sleep deprivation affects the physiological state of drivers, which can have serious consequences when combined with long periods of inactivity. Car manufacturers and researchers should take this into account when designing intelligent systems capable of providing drivers with appropriate warnings before a critical situation arises.

Keywords

• Driver state
• Driving environment
• Sleepiness
• Fatigue
• Machine learning
• Physiological state