Drowsiness detection using remote photoplethysmography signals with machine learning techniques

Name: Panupong Sunkom

LCSH: Kasetsart University -- Theses. M.Eng. (Agtificial Intelligence and Internet of Things) 2023

Classification :.LCCS: QA76.87

LCSH: Kasetsart University. -- Department of Electrical Engineering -- Theses

LCSH: Neural networks (Computer science)

LCSH: Drowsiness

LCSH: Machine learning

LCSH: Plethysmography

LCSH: Time-domain analysis

LCSH: Biosensors

Abstract: A drowsiness detection system based on the neural network (NN) model is introduced in order to classify the subject's condition into alert and drowsy states utilizing the heart rate (HR) and the ultra-short-term pulse rate variability (PRV) in various domains, including time, frequency, and nonlinear. These features are extracted from the remote photoplethysmography (rPPG) signal obtained from the human’s facial video and will be used as inputs for the NN model. The rPPG signal is extracted from the region of interest (ROI) located on the forehead. However, the hair may be partially covering the skin, and changes in ambient light may cause spurious frequencies, affecting the quality of the rPPG signal, HR, and the ultrashort-term PRV. Because of these reasons, the hair detection algorithm and the majority vote in subintervals are proposed to improve the HR and ultra-short-term PRV before providing them to the NN model. The experiments were conducted on the public dataset, UBFC-RPPG. The results indicate that utilizing the proposed method improves the mean absolute error (MAE) of the HR, ultra-short-term PRV, and mean absolute percentage error (MAPE) of the ultra-short-term PRV. The AlertDrowsy (AD) dataset is used to extract the improved HR and ultra-short-term PRV for training and testing the NN model. The results showed that, when the NN model is trained with 70% and tested with 30% of the improved HR and ultra-short-term PRV, the accuracy of the NN model is 92%, with 92.64% precision, 91.11% recall, 92.87% specificity, and 91.87% F1 score. Furthermore, when the ten-fold crossvalidation is applied to the NN model, the model achieves an accuracy of 94.66 ± 0.76%, 94.69 ± 1.16% of precision, 94.65 ± 1.61% of recall, 94.67 ± 1.24% of specificity, and 94.66 ± 0.78% of the F1 score.

Kasetsart University. Office of the University Library

Address: Bangkok

Email: tdckulib@ku.ac.th

Name: Denchai Worasawate

Role: Thesis Advisor

Name: Chadchai Srisurangkul

Role: Thesis CO-Advisor

Name: Minoru Nakayama

Role: Thesis CO-Advisor

Created: 2023

Modified: 2025-07-23

Issued: 2025-07-23

วิทยานิพนธ์/Thesis

application/pdf

URL: https://www.lib.ku.ac.th/KUthesis/2566/panupong-sun-all.pdf

CallNumber: QA76.87 .P37

eng

DegreeName: Master of Engineering

Level: Master's Degree

Descipline: Artificial Intelligence and Internet of Things

Grantor: Kasetsart University

©copyrights Kasetsart University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	panupong-sun-all.pdf	4.76 MB

ใช้เวลา

0.032206 วินาที