Research Article
A Real-Time Deep Learning-Based Driver Drowsiness Detection System Using Computer Vision
@INPROCEEDINGS{10.4108/eai.19-12-2025.2366262, author={Arya Zade and Tamanna Vaikkath and Arnav Padole and Akhil Gupta and Monali Gulhane and Firdous Sadaf Mohammad Ismail and Nitin Rakesh}, title={A Real-Time Deep Learning-Based Driver Drowsiness Detection System Using Computer Vision}, proceedings={Proceedings of the 1st International Conference on Computational Techniques in Data Science, IICTDS 2025, 19-20 December 2025, Chandigarh, India}, publisher={EAI}, proceedings_a={IICTDS}, year={2026}, month={11}, keywords={Driver Drowsiness Detection Computer Vi- sion Convolutional Neural Network TensorFlow Lite FastAPI Streamlit PERCLOS Real-Time Monitoring}, doi={10.4108/eai.19-12-2025.2366262} }- Arya Zade
Tamanna Vaikkath
Arnav Padole
Akhil Gupta
Monali Gulhane
Firdous Sadaf Mohammad Ismail
Nitin Rakesh
Year: 2026
A Real-Time Deep Learning-Based Driver Drowsiness Detection System Using Computer Vision
IICTDS
EAI
DOI: 10.4108/eai.19-12-2025.2366262
Abstract
This paper presents a practical, real-time, software- only Driver Drowsiness Detection System that leverages computer vision and a lightweight convolutional neural network (CNN) to detect prolonged eye closure (drowsiness) from a webcam feed. The prototype integrates a Streamlit frontend (live webcam capture and ROI extraction using Haar cascades), a FastAPI backend (receives eye-region crops and returns predictions), and a TensorFlow Lite (TFLite) CNN model optimized for on-device inference. The dataset comprises ∼4000 labeled eye images; the model undergoes augmentation and is converted to TFLite for deployment. In testing, the TFLite-optimized system achieves real-time performance (up to 18 FPS on consumer hardware) and strong classification results (overall test accuracy ≈95.8%). The proposed architecture provides a low-cost, deployable solution for Advanced Driver Assistance Systems (ADAS) and can be extended with multimodal signals for enhanced robustness.