Driver Drowsiness Detection with optimized pre-processing for eyelid-closure classification using SVM

초록

Driver’s drowsiness is of paramount importance for safety measures and can lead to fatalities if not detected and served timely. Over the years, numerous vision-based drowsiness detection algorithms have been proposed by exploiting the camera’s image/video equipped inside the car. In this paper, we have proposed a real-time, more generalized, and robust drowsiness detection system that efficiently works for unseen faces with different races and camera conditions. We have utilized Viola-Jones algorithm for face and eye detection. Most of the existing solutions for eye status (closed/open) recognition, the SVM classifier is trained on a particular dataset, which is not generalized and biased towards features of a specific race. We have proposed a gram-polynomial basis function as a pre-processing step before feature extraction. The proposed technique filters unnecessary pixels based on a gram-polynomial basis. This allows feature extractors to focus on substantial information for significant corners and edge detection inside the eyelid image, which in turn increases the accuracy of the classifier. Our proposed technique generates randomness in the pixels after decimation, which reduces the biasedness of classifier towards one class. We have analyzed different feature extractors such as LBP and HoG features and used SVM classifiers for eye status classification for drowsiness detection. We have used the MRL Eye dataset containing images of 37 different persons, and with entirely distinct lighting conditions, including infra-red images for training SVM, while for testing, we have used webcam camera with varying illumination and ten entirely different subjects containing three human races.