High-Fidelity Machine Learning Techniques for Driver Drowsiness Detection

Abstract

It is devastating that daily, there is an ample number of car crashes that cause damage to automobiles, onboard passengers get injured, and others tend to lose their lives. Road crashes are fast rising across the globe and have drawn many road safety commissions and concerned individuals to discuss ways to reduce this menacing situation drastically. With the introduction of artificial intelligence and technological advancement, the government and state commissions have beckoned on the various universities and research institutions to develop methods to curb the rise of automobile crashes. Some causes of these crashes include drunk driving and drowsiness, the latter is most prevalent as it occurs to all and sundry. Drowsiness detection can be categorized into three main techniques; behavioral-based, vehicular-based, and physiological-based. In this research, the behavioral-based approach was studied, with significant consideration being the cost of implementation, execution time, and accuracy. Three machine learning (ML) classifiers were considered: Support Vector Machine (SVM), Naïve Bayes (NB), and Random Forest (RF). A dataset of 1448 images was used for training and testing these classifiers: 70% for training and 30% for testing. Random Forest classifier gave the best accuracy of (92.41%) compared to SVM (90.34%) and Naïve Bayes (69.43%). A deep neural network (VGG16) was used to classify drowsiness, and this gave a high accuracy of 97.20%, which outperformed the traditional machine learning models.