SINGLE-CHANNEL REAL-TIME DROWSINESS DETECTION BASED ON ELECTROENCEPHALOGRAPHY
Drowsiness is considered as a major risk factor in workplace injuries and fatalities as much as alcohol. Drowsiness-related accidents tend to be catastrophic. The need of a reliable drowsiness detection system is arising today, as drowsiness is considered as a major cause for many accidents in different sectors. In this thesis, we propose a real-time drowsiness detection system based on a single-channel electroencephalography (EEG). Towards that goal, we introduced three main contribution proposed in this thesis: (1) a real-time drowsiness detection algorithm based on EEG suitable for portable applications with low computational complexity; (2) several novel algorithms to train classifiers that can be implemented on chip with low-power fixed-point arithmetic with extremely small word length; (3) an instantaneous drowsiness detection system suitable for short-time windows of single-channel EEG signal. The proposed real-time drowsiness detection algorithm adopts a cumulative counter to extract important features from 8 different frequency bands. Our experimental results demonstrate that the proposed algorithm is capable of detecting drowsiness with superior accuracy (83.36%) over the conventional method (70.62%). The proposed fixed-point algorithms incorporate the non-idealities (i.e., rounding and overflow) associated with fixed-point arithmetic into the offline training process so that the resulting classifiers are robust to these non-idealities. Our numerical experiments demonstrate that the proposed methods are able to achieve up to 1.67x reduction in the word length compared to the conventional approaches without surrendering any classification accuracy. The instantaneous drowsiness detection algorithm proposed in this work is based on Convolutional Neural Network (CNN). Our experimental results demonstrate that our CNN-based drowsiness detection system is capable of detecting drowsiness in short-time windows (five seconds) with higher accuracy (84.8%) compared to conventional methods (71.0%) and the counter-based method (77.2%). Finally, we briefly discuss few possible research tasks for the future: (1) wearable derive for industrial workers, (2) fixed-point implementation for CNN, and (3) multimode data fusion.
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 United States License.
Rights for Collection: Duke Dissertations