Driver state prediction from vehicle signals: An evaluation of segmentation approaches

Abstract

Modern vehicles typically are equipped with assistance systems to support drivers in staying vigilant. To assess the driver state, such systems usually split characteristic vehicle signals into smaller segments which are subsequently fed into algorithms to identify irregularities in driver behavior. In this paper, we compare four different approaches for vehicle signal segmentation to predict driver impairment on a dataset from a drunk driving study (n=31). First, we evaluate two static approaches which segment vehicle signals based on fixed time and distance lengths. Intuitively, such approaches are straightforward to implement and provide segments with a specific frequency. Next, we analyze two dynamic approaches that segment vehicle signals based on pre-defined thresholds and well-defined maneuvers. Although more sophisticated to define, the more specific characteristics of driving situations can potentially improve a driver state prediction model. Finally, we train machine learning models for drunk driving detection on vehicle signals segmented by these four approaches. The maneuver-based approach detects impaired driving with a balanced accuracy of 68.73%, thereby outperforming timebased (67.20%), distance-based (65.66%), and threshold-based (61.53%) approaches in comparable settings. Therefore, our findings indicate that incorporating the driving context benefits the prediction of driver states. I. INTRODUCTIONIn recent years, with the advent of driver assistance systems, research in reliably predicting driver states has increased significantly. The aim is not only to enhance the comfort of drivers [1], but especially to improve road safety [2] by preventing accidents related to driver impairment. Advances in in-vehicle computing capacities and sensor technologies further support these endeavors. Besides camera-based technologies [3], the focus is on privacypreserving and non-intrusive approaches [4], [5]. As outlined in recent work, a large number of in-vehicle signals can be accessed via the controller area network (CAN-bus), which allows insights into the driving behavior [6], [7]. The research field around gathering driver insights from in-vehicle signals is extensive. It ranges from driver identification [4], [5], [8], over driving style recognition [9]-[11] to