Driver Drowsiness Estimation Based on Hybrid Feature Extraction and Light weighted Dense Convolutional Network

Abstract

Researchers propose a fully automated method of detecting drowsiness using driving images with a focus on fatigue driving detection. Kanade – Lucas – Tomasi - ViolaJones (KLT-ViolaJones) is used to locate feature points and detect faces in the proposed algorithm and feature points are used to extract the region of interest (ROI). In order to determine the status of the eye from the ROI images, Histogram oriented Gradient (HoG) is used. Two parameters with which fatigue can be detected are percentage of eyelid closure over pupil over time (PERCLOS) ratio and Eyes Aspect Ratios (EAR). Experimental results demonstrate that the proposed Light Weighted Dense Convolution Network (Li-DenseNet) can detect drowsiness levels in drivers using the National Tsing Hua University Driver Drowsiness Detection dataset (NTHU-DDD). The proposed algorithm Li-DenseNet outperforms other CNN-based methods, AlexNet, VGG, RNN, and ResNet showing accuracy, sensitivity, specificity, precision and F1-Score rates of 98.44%, 91.5%,92.3%,98.2 and 97.02%, respectively.