Non-Intrusive Drowsiness Detection for Accident Prevention Using a Novel CNN

Abstract

Drowsiness detection problem is not only complex, but also very important for accident prevention. In this paper, we propose a non-intrusive drowsiness detection method using the right eye and mouth. Face detection is performed using HOG + SVM methodand facial features are segmented using 8 landmarks obtained by an ensemble of regression trees and classified using a novel convolutional neural network that we call Dozy-Net. Then, drowsiness detection is carried out using three behavioral parameters: PERCLOS, blink frequency, and yawning duration. Two state-of-the-art and one self-constructed dataset were used to train, test, and compare Dozy-Net’s performance against other six state-of-the-art convolutional neural networks, being Dozy-Net significantly faster. Drowsiness detection model was tested on a real-life dataset performing 75.8% accuracy and an average speed of 21.51 FPS. Compared to other existing models, our proposal has the advantage of having been tested in conditions similar to those to be expected in a real environment.