A Non-invasive approach for Driver Drowsiness Detection using Convolutional Neural Networks

Abstract

Driver drowsiness has been observed as one of the most common causes for road accidents, producing nearly 40% of death and casualties. When a driver falls asleep, he starts losing control and is unable to take reflex action to avoid the accident or to reduce its impact. This necessitates the need for developing a mechanism that provides timely alerts to the driver when he is drowsy. In this paper, an efficient and non-intrusive algorithm that uses a deep convolutional neural network to analyze yawn behavior is proposed. The proposed technique is built by modifying the VGG16 architecture to include batch normalization, ReLu activation for the intermediate layers and sigmoid activation after the final dense layer. The performance of the proposed approach is verified on the YawDD dataset and is compared against VGG16, VGG19, MobileNet, and AlexNet. Experimental results show that the proposed approach outperforms the other networks in terms of accuracy.