TY - JOUR TI - Camera based detection of the onset of cognitive fatigue DO - https://doi.org/doi:10.7282/T3959MH6 PY - 2017 AB - The onset of cognitive fatigue is associated with a period of transient, subconscious decrease in maximal cognitive ability, typically influencing decision making. The ability to visually detect this early stage of fatigue can help prevent numerous workplace hazards where top cognitive performance is of utmost importance. In this work, we developed a camera-based system that utilizes visual symptoms of fatigue to estimate its early stage. For the first part of the work, in collaboration with the Magnetoencephalography Lab at MIT, we conducted a 3-hour long, fatigue inducing experiment on 13 test subjects and collected synchronous camera (visual) and Magnetoencephalography MEG (brain) data. We extracted 8 eyelids and 6 head-movement related features to train binary classifiers like Support Vector Machine, k-Nearest Neighbor, Random Forest and Artificial Neural Network to distinguish between “Non-Fatigue” (early stage) and “Fatigue” (late stage), achieving test accuracy of 89%, 90%, 95% and 98% respectively. We propose a temporal sliding window technique of using these binary classifiers for detecting a gradual change in the level of fatigue. We observed a progressive increment in detection of “Fatigue” class inside this window as it moves towards the later stages of the experiment time-line. For validation, we compared our model’s results with fatigue-induced brain signals from the MEG data, namely the alpha band (8-12 Hz) power. Regressing alpha power on camera-based features yielded an average r-squared value of 0.6. For the second part of the work, we conducted a similar experiment at the Laboratory for Computational Brain at Rutgers. We recorded synchronous camera and Electroencephalography (EEG) data for a 90-minute long experiment conducted on 4 test subjects. For this experiment, the fatigue-inducing task involved was made adaptive to the cognitive abilities of the test subject, aiming to make the subject tired in a shorter amount of time. We repeat our analysis for the camera based and the brain data based fatigue detection models. We obtained similar progressive increment in the "Fatigue" class for all subjects. By regressing the alpha power from EEG data on the visual features, we obtained average r-squared up to 0.8 for fatigue-induced brain regions. We validate our camera model based on the EEG indicator of fatigue. Our results demonstrate promise in terms of using a vision-guided fatigue estimation model for designing a real-time fatigue detection system. KW - Computer Science KW - Machine learning KW - Mental fatigue LA - eng ER -