Machine Learning Inspired Vision-based Drowsiness Detection using Eye and Body Motion Features

Drowsiness-a state before the onset of sleep- resulting from insufficient s leep i s recognized a s a g lobal problem due to associated health and safety risks for the individuals involved in activities requiring constant attention. Therefore, several computer vision-based non-invasive techniques have been proposed for the timely detection of drowsiness. However, these methods are generally based on drowsy behavior indicators like yawning and excessive eye blinking. Moreover, the results are generally reported for databases with very few subjects or acted drowsy data. This paper proposes a drowsiness detection technique based on hybrid features using comprehensive and challenging real drowsy data. Primarily, eye state and body motion analysis is performed to determine drowsiness. Towards ameliorating this, the eye region is selected from each frame using facial landmarks and is described using a histogram of oriented gradients (HoG) descriptors. For body motion description, frame difference is computed and parameterized using HoG descriptors. Then, the hybrid feature set, i.e., the combination of eye and body motion features, is subjected to dimensionality reduction through principal component analysis. Finally, SVM is trained and tested on the hybrid feature set to detect drowsiness. The detection accuracy of 90% is achieved through our proposed technique.