Real-Time Driver Fatigue Detection Using Automated Eye-State Annotation and Transformerbased Deep Learning Models

View at publisher Summarise with VIRAT
Authors: A Hepziba, A Jenefa, Bijolin Edwin, P Santhiya, Ebenezer V, Stewart Kirubakaran S
Cited by: 0
References: 12

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accidents caused by driver fatigue are a public health issue that can be prevented, but still, they impose a heavy burden in terms of the social and economic costs to society. The detection of driver sleepiness using vision-based monitoring with regular cameras has evolved to become a promising noninvasive methodology. We present here a complete pipeline that merges the automated data labeling and leading-edge deep learning models for the detection of driver fatigue via the incabin video. To begin with, via Google's MediaPipe Face Mesh, we automatically pinpoint eye landmarks and calculate the eye aspect ratio (EAR), which results in the creation of a vast annotated dataset of “eyes open” vs. “eyes closed” images with very minimal human intervention. After that, we proceed to train and evaluate various models, including (1) YOLOv8m, a speedy single-stage CNN detector that has been finetuned for eye-state detection, (2) RT-DETR-L, a transformer-based end-to-end object detector that has been modified for driver monitoring, (3) DINOv3-ViT, a self-supervised Vision Transformer with 7 billion parameters, operating in the frozen mode, and a linear head for image-level “Alert/Drowsy” classification, and (4) a small custom CNN used as a baseline. Tests conducted on our datasets indicate that the DINOv3 classifier reaches the highest image-level accuracy (~ 96.1%) thanks to its huge pre-training. For accurate eye-state detection, the fine-tuned RT-DETR-L detector turned out to be the best performer: after prolonged training (AdamW optimizer, 200 epochs), it reached the mean average precision (mAP) of 0.971 (IoU 0.5: 0.95) on the eye-state task. In contrast, a similarly tuned YOLOv8m model provided slightly lower mAP on the same dataset. The optimized RT-DETR-L operates in real-time, in line with claims that RT-DETR architectures can easily balance speed and accuracy. First, we combined the best model with a demo app, which was developed in Streamlit, that takes webcam frames, predicts, and sends alerts through N8n webhook to notify if the driver is drowsy. Our contributions are, in brief: a pioneering Mediapipe-based labeling pipeline that automatizes data processing; an exhaustive comparison of three approaches CNN, transformer detector, and ViT-based methods for drowsiness detection -; and the realization of a realtime system prototype with the capability to provide alert based on fatigue detection. Which could also help to a certain extent in avoiding accidents caused by fatigue.

Fields of Study & Subjects

Sleep and Work-Related Fatigue Advanced Neural Network Applications Face recognition and analysis

Publication details

Journal / source
International Conference on Communication, Computing and Electronics Systems (ICCCES)
Pages
601–606
Publication date
21 January 2026
Language
EN