Using a distracted driver's behavior to inform the timing of alerts in a semi-autonomous car

Abstract

The time taken for a distracted driver to return to vigilance in the task of driving a semi-autonomous car is dependent on how engaged the driver is in other activities - such as finishing a text, phone call, etc. In a pilot study of multitasking, we found that auditory alerts that are given in fixed intervals (for example, every 2 seconds) tend to be ignored by or annoy the driver. Our method learns the statistical nature of the engagement of the driver to the particulars of the task using Markov Renewal Processes (MRP) and we use these statistics to optimize the timing of auditory alerts to make semi-autonomous driving safer and enjoyable. The virtue of the method also allows it to continually learn the driver's behavior at all times, even after the initial training period. Therefore, our method alerts the driver in a way that minimizes annoyance and increases effectiveness. More importantly however, it learns to alert each driver in an individualized manner, tailored to the driver's unique behavioral patterns learned over time.