Prediction of Future Collision Risk for Truck Drivers Using the Time-Series Autonomic Nerve Function

Abstract

To prevent fatigue-related collisions among drivers, objective assessments of their physiological states are essential. A simple and quantitative evaluation method for fatigue involves the use of autonomic nerve function (ANF) indices obtained from heart rate variability analysis, which provides an understanding of the health state reflected by the activity of the autonomic nervous system. However, predicting the occurrence of crashes using physiological data is challenging. This is primarily owing to the complexity of the relationship between drivers and accident risk. In most previous studies, the targets of driving situations have been often limited, or the prediction targets have been set as the driver's internal state rather than the accident. In this paper, we propose a novel collision risk prediction model using ANF and several simple external information types, which can be extracted from standard in-vehicle sensors without limiting the driving scene. Our experiments using actual truck drivers' data reveal that the proposed model can achieve collision risk prediction for the following 30 min with an accuracy of 74.9% recall and 0.79 AUC. Furthermore, we discover that simple external information obtained based on the vehicle speed significantly contributes to the prediction accuracy. As our prediction method only requires commonly equipped sensors as the sources of external information, this method is expected to be easily implemented not only for truck driving but also for general vehicle driving, where crashes are often likely.