A comparison of three subsidiary tasks used as driver drowsiness countermeasures

Two previous studies performed at Virginia Tech have shown that it is feasible to detect drowsy drivers using driving performance and physiological measures. Therefore, assuming that drowsiness can be detected, it becomes important to develop methods (countermeasures) by which drivers can regain and maintain alertness. The current study was thus undertaken in an attempt to evaluate three subsidiary tasks which differed only in regard to input modality (auditory, tactual, or visual) in terms of: 1) the degree to which they aided the driver by maintaining or restoring alertness; and 2) the degree to which the responses to these tasks could be used to detect drowsiness. Subjective measures of drowsiness were also obtained to provide an additional source of verification of level of drowsiness.

To accomplish these objectives, a total of 12 male and female driver-subjects drove a moving-base simulator continuously from 12:30 a.m. to 3:00 a.m. During this time, the subjects performed each of the subsidiary tasks for a 30-minute period; they also drove for a 30-minute period during which no subsidiary task was performed. During the simulated, nighttime, highway driving scenario, 20 driving performance, behavioral, and physiological measures were collected for each 3-minute driving interval, along with 5 subsidiary task measures and subjective alertness ratings.

The experimental results indicated that none of the three subsidiary tasks provided an effective means of maintaining driver alertness. However, the results of a second series of discriminant analyses did indicate that driver impairment due to drowsiness could be reliably detected with linear combinations of subsidiary task and driving measures. In fact, promising discriminant models for the auditory and visual tasks were identified which employed a subsidiary task response measure of the number of correct responses to the subsidiary task during each 6-minute driving interval as well as a physiological measure of the subject's heart rate variance; these models showed overall classification error percentages as low as 3% and 8%. Finally, the analyses of the subjective alertness ratings indicated that subjects' ratings were not significantly affected by either the type of subsidiary task performed or time-on-task.