Study of Online Driver Distraction Analysis using ECG-Dynamics

Abstract

Majority of the road fatalities occur due to a common cause of human error while driving. Distracted driving is one of the most important contributors to road disaster, because it involves temporary suspension of driver’s vigilance while driving. This hypo-vigilance can occur through variety of ways such as talking on cell-phone, texting, conversing with passenger, etc. In order to minimize threats happening across the hypo-vigilance through driver distraction, it becomes highly essential to characterize and identify distraction. During the last decade, many research investigations were conducted on driver state estimation. Particularly, Electroencephalography (EEG), camera-based systems, and behavioral data analysis. Although those systems achieved high empirical performances, there are serious roll block to adopt them practically such as privacy issues, detection latency, or intrusiveness. In this study, we investigate continuous Electrocardiogram (ECG) signals to monitor physiological changes during normal vs. distracted driving in an on-road recording experiment. ECG-based driver state detection is particularly of interest due to its being easy to wear/embed, reliable and minimally intrusive recording technology, and its high signal to noise ratio recording. In this paper, we generated a set of ECG-based measures in order to characterize and identify common pre-defined distracted scenarios. Our aim is to provide an empirical approach for accurate analysis of driver distraction. In this study we introduced distraction by 1) hand-held phone conversation, 2) driver conversation with a passenger next to him, and 3) driver texting on phone while driving. Our effort primarily focuses on the efficient characterization of distraction while driving via localizing R-R interval series based on temporal features as well as spectral features. In addition to this, we further investigated different short window sizes on the ECG recording stream for real-time predictive ability of the extracted features through state of the art predictive algorithms. Our experimental analysis demonstrated ~92% average predictive accuracy of driver distraction identification in near real-time. In the later part of this study, we also achieved the secondary workload estimation while driving by introducing wavelet as a filter bank approach, this method performed significantly well, yields an open door on spectral analysis in greater depth.