Research Papers

Non-Intrusive Monitoring of Drowsiness Using Eye Movement and Blinking

Filename 3C_4_Zandi_Paper.pdf
Filesize 864 KB
Version 1
Date added June 26, 2017
Downloaded 3 times/fois
Category 2017 CARSP XXVII Toronto
Tags Research and Evaluation, Session 3C
Author/Auteur Ali Shahidi-Zandi
Stream/Volet Research and Evaluation

Slidedeck Presentation

3C_4_Zandi

Abstract

Due to life style and work demands, people are becoming more sleep-deprived than ever before, which can have a negative impact on health, safety and performance, and even lead to deadly consequences. According to the National Highway Traffic Safety Administration (NHTSA), drowsy driving caused about 72,000 crashes in 2013 on U.S. roadways, resulting in 800 fatalities and 44,000 injuries. NHTSA also reported that nearly 3% of all crash fatalities in 2014 involved drowsy driving. The goal of this study was to investigate the characteristics of eye movement and blinking as a correlate of drowsiness, and their relationships to behavioural vigilance. Ultimately, this research would lead to development of non-intrusive techniques for real-time assessment of the state of vigilance in drivers, as a crucial step towards managing fatigue and improving road safety. Eye tracking data was obtained from 15 subjects (age 22.9±3.3 years; 11 female) using an infra-red-based system at the Brain and Mind Sleep Research Laboratory, Western University, Canada. Each subject participated in two psychomotor vigilance task (PVT) sessions, under different sleep conditions required for the prior night: normal sleep (NS) and sleep restriction (SR). Each session included 6 episodes of 10-min PVT with a total of 600 stimuli-response trials, where the reaction times to visual stimuli were used as the baseline. From a 10-second window immediately preceding each stimulus, a set of features from eye tracking data was extracted, including statistical measures, frequency, percentage and/or duration of gaze coordinates, fixations, saccades and blinks. These feature vectors were then used to identify drowsiness by adopting two separate machine learning approaches: parametric modelling of the desired state of vigilance (e.g. Gaussian mixture models - GMMs) and classification (e.g. support vector machine - SVM). We evaluated the performance of the eye-tracking-based features in identification of drowsiness for each subject in every session (NS or SR). Parametric modelling of the state of alertness using GMM resulted in low normalized root-mean-square errors for both NS and SR sessions (median of less than 0.2), while outperforming a random estimator. On the other hand, binary classification of the state of vigilance using SVM revealed a high median accuracy (sensitivity-specificity) of 85% (84%-89%) and 83% (90%-88%) for NS and SR sessions, respectively, across all subjects. PVT reaction time as an objective measure of vigilance was used as the baseline. In addition to serving as an effective experiment to induce fatigue, the prolonged vigilance task adopted in this study provided the opportunity to continuously measure the level of vigilance without interrupting subjects, resulting in a baseline time series and consequently more accurate evaluation. Overall, the results of this study verify the potential of the extracted eye tracking features (together with the proposed machine learning methodologies) for reliable and non-intrusive assessment of drowsiness. Results of this study suggest a high correspondence between features extracted from eye tracking and reaction time, during a sustained vigilance task. Further investigations, under various levels of fatigue and time of day, will be required to assess the performance of the proposed methodologies.

Ali Shahidi-Zandi