Mitigating Drowsiness: Linking Detection to Mitigation

Drowsy driving contributes towards up to 24% of crashes and near crashes observed; 886 fatal crashes per year can be attributed to drowsy, fatigued or sleeping drivers. Drowsiness mitigation technology is composed of a detection algorithm and a mitigation component. This paper is primarily concerned with the latter, specifically for a driving simulation study about mitigating drowsy driving. The study is part of National Highway Traffic Safety Administration's (NHTSA’s) Driver Monitoring of Inattention and Impairment using Vehicle Equipment (DrIIVE) program. The detection algorithm incorporates time series probabilistic estimation using a Hidden Markov Model, so a drowsiness prediction at any time is dependent on a previous history of observations. Two mitigation methods are designed for testing in the simulation study. One is a three stage audio/visual alert that requires a driver response through a button press. The second is a binary haptic alert that uses a vibrating seat. Additionally, each mitigation will include three varying levels of sensitivity: a nominal model, an over-sensitive model, and an under-sensitive model. These variations will expose drivers to different numbers of false alarms while also potentially missing episodes of drowsiness. Various parameters in the detection algorithm were tested and the vote thresholds of two Random Forest models were selected for variation. It was observed how these parameters affected the output of the detection and mitigation system using previously collected drowsy driving data. Three specific levels were chosen as candidates for the experiment. It is hoped that the study will answer questions about how effective a mitigation system is at changing driving performance, whether drivers willfully ignore the mitigation, and how many alerts are too many.