Driver Drowsiness Detection on the Edge
TL;DR
A comparative evaluation of several state-of-the-art deep learning models for drowsiness detection, including convolutional and transformer-based architectures, across multiple public datasets and a custom-compiled dataset reveals significant performance variability across datasets, indicating limited generalization beyond the original training domains.
Abstract
Driver drowsiness detection is a key component of modern Advanced Driver Assistance Systems (ADAS), aiming to enhance road safety through timely identification of reduced driver alertness. In this work, we present a comparative evaluation of several state-of-the-art deep learning models for drowsiness detection, including convolutional and transformer-based architectures, across multiple public datasets and a custom-compiled dataset. Our analysis focuses on cross-dataset generalization, lightweight preprocessing strategies, and suitability for on-device deployment. Experimental results reveal significant performance variability across datasets, indicating limited generalization beyond the original training domains. We further show that targeted preprocessing, including face detection and Contrast Limited Adaptive Histogram Equalization (CLAHE), consistently improves model performance without increasing training complexity. Although LiteRT conversion was explored to enable mobile deployment, inconsistent performance and incomplete model support stemming from the lack of a unified and reliable conversion framework, limited its effectiveness. These findings highlight key challenges in deploying robust, privacy-preserving, and real-time drowsiness detection systems on resource-constrained devices.