Lightweight 2D-1D CNN for Real-Time Driver Drowsiness Detection withFederated Learning on Edge Devices
Informatica · 29.06.2026 · 50 (12)
DOI: 10.31449/inf.v50i12.12530 ↗Kurz gefasst
Fachartikel in der Zeitschrift Informatica, online erschienen am 29.06.2026. Verfasst von Amina Guettas und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Driver drowsiness is one of the main leading causes of traffic crash accidents. Preventing these mishapsand limiting consequential damage and losses, whether human or material, has been the subject of manyresearch studies these recent years. Fulfilling this task requires a reliable drowsiness detection systeminstalled in the vehicle that can function on limited hardware such as edge devices with real-time executionspeed. The present study proposes a computer vision-based driver drowsiness detection system usingConvolutional Neural Networks. The traditional approach is to extract some features from the captureddriver’s face images such as PERCLOS, blink speed, or blink frequency, and determine the driver state usingmachine learning methods, predefined thresholds, or rules. However, deep learning techniques allowus to directly identify the driver state from the captured video, yet these techniques are usually computationallyintensive and unsuitable for edge computing applications. To overcome these obstacles, we proposea lightweight 2D-1D CNN model for driver drowsiness detection. We deployed a 2D-CNN model for theeyes feature extraction of both the left and right eyes of the driver from every captured frame. In addition,we applied a 1D-CNN model to capture the temporal information of the concatenated features overa defined period. We implemented and evaluated the model’s performance under both centralized learningand federated learning. Experimental results show that our approach achieves 99% accuracy on theDROZY dataset and 98.06% on subjects excluded from the training set in the CL, while in the FL it achieves95.27% on the same dataset and 77.86% on unseen subjects. Furthermore, the model runs in real time onan NVIDIA Jetson Nano, reaching 25.6 FPS. While FL offers enhanced privacy by enabling local modeltraining without sharing raw user data, it exhibits limited generalization capability. In contrast, CL provideshigher accuracy and a more robust model. Comparative analysis reveals a performance trade-off forFL, balanced by improved data privacy. This work highlights the potential of privacy-preserving, real-timedrowsiness detection systems using advanced deep learning architectures.
Zitieren
APA
Guettas, A., Ayad, S., & Kazar, O. (2026). Lightweight 2D-1D CNN for Real-Time Driver Drowsiness Detection withFederated Learning on Edge Devices. Informatica, 50(12). https://doi.org/10.31449/inf.v50i12.12530
Vancouver
Guettas A, Ayad S, Kazar O. Lightweight 2D-1D CNN for Real-Time Driver Drowsiness Detection withFederated Learning on Edge Devices. Informatica. 2026;50(12). doi:10.31449/inf.v50i12.12530
BibTeX
@article{Guettas2026,
author = {Guettas, Amina and Ayad, Soheyb and Kazar, Okba},
title = {Lightweight 2D-1D CNN for Real-Time Driver Drowsiness Detection withFederated Learning on Edge Devices},
journal = {Informatica},
year = {2026},
volume = {50},
number = {12},
doi = {10.31449/inf.v50i12.12530}
}
EUVIMED