Real-time Driver Drowsiness Detection using CNN, MediaPipe, and ML classifiers

Driver Drowsiness is among the primary reasons for road accidents. It poses a potential threat to the drivers, passengers and individuals involved in traffic. To address this critical issue, this paper explores the efficacy of driver drowsiness detection using deep learning (DL) and machine learning (ML) techniques to develop robust systems that enhance road safety. Transfer learning is performed on the robust pretrained Convolutional Neural Network (CNN) models- VGG16 and ResNet50 on image dataset. ResNet50 outperformed VGG16 achieving 95% of validation accuracy when trained with a similar model architecture. As an alternative approach, the study explores the efficacy of traditional classifiers on facial features extracted from the same image dataset used for transfer learning. Feature extraction is performed using MediaPipe Face Land marker model. A total of fifty-two (52) facial numerical features-such as eye blink rate, eye openness, eye squint, and jaw openness-are extracted from driver's images. These features are then trained and evaluated using traditional classifiers such as Random Forest, Gaussian Naive Bayes, Support Vector Classification (SVC), Decision Tree, XgBoost and a Multi-Layer Perceptron (MLP) neural network with and without Principal Component Analysis (PCA). Random Forest achieved test accuracy of 97%, the highest among all the models when trained without performing PCA. The study showcases the influence of data format, model architecture, and hyper parameter tuning on model performance. This work also emphasizes the potential for developing a robust drowsiness detection system by adapting trained models to classify video frames in real time. Such a system holds promises for significantly improving road safety and addressing the critical issue of drowsy driving.