Comparative Analysis of Machine Learning and Deep Learning Models for Real-Time Driver Drowsiness Detection

Drowsy driving is among the major factors leading to accidents on roads, especially in scenarios where long hours of traveling or working are involved. The current paper performs a comparative analysis between traditional machine learning (ML) techniques and lightweight deep learning (DL) methods concerning driver drowsiness detection using the YawDD Dataset (a yawning detection dataset). This Dataset consists of 322 videos from the mirror (rear-view mirror camera position) subset and 29 videos from its Dash (dashboard) subset, recorded from 107 drivers of diverse ages, gender and ethnicities. The videos are presented in various levels of illumination and facial occlusions (glasses and sunglasses) as well as poses. Frames were extracted from the videos at regular intervals resulting in thousands of labelled images. This study evaluates two types of algorithms based on their performance through several parameters such as accuracy, precision, recall, and F1 score while simultaneously considering the efficiency of computation measured by inference time, CPU usage, memory, and frames per second (FPS). According to the findings, while DL models like the convolutional neural network based on EfficientNet and TinyCNN offer better classification performance with accuracy levels surpassing 93%, these models exhibit inferior inference rates. Consequently, these DL models cannot be employed in real-time situations since they require hardware accelerators. On the other hand, traditional ML models, mainly the combination of Random Forest and Support Vector Machine, offer a good balance between accuracy and efficiency. Logistic regression provides the best processing rate but with limited accuracy. Conclusively, the results imply that traditional ML models are still applicable for real-time tasks in resource-constrained environments compared to DL models. However, DL models are recommended for implementations where hardware accelerators are available.