Deep Learning Models for Classifying Driver Eyes: A Comparative Study

Abstract

Driving a vehicle with drowsiness is a very serious and widespread problem in society, because drowsiness has a negatively influence on the driver's reaction time. And therefore, when the level of drowsiness increases in the driver, he loses control of his vehicle. He can unexpectedly veer off the lane, colliding with an obstacle or causing a car to overturn. In this paper, we present a low-cost, non-intrusive, more accurate, and better solution for detecting driver drowsiness in real-time in real-world driving conditions, whenever the drowsiness is detected, and the system activates an audible alarm to alert the driver before he falls asleep. In the proposed method, we used the most important facial components that are considered the most effective for sleepiness. We used the Viola-Jones algorithm to detect the driver’s face and eyes area. Then we inserted the resulting image into the deep convolutional neural network to detect driver drowsiness in real-time. The purpose of this paper is to arrive at the performance of five deep learning models: AlexNet, ResNet50, GoogleNet, VGG16, and DenseNet201, which detects sleepy using RGB footage of drivers as input. The experimental results indicate that all these models produce excellent detection accuracy but DenseNet201achieves the highest detection accuracy compared with others.