Facial Video-based Non-contact Physiological Signal Measurement and Applications
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
Auditorium F101, Faculty of Biochemistry and Molecular Medicine (Aapistie 7)
Topic of the dissertation
Facial Video-based Non-contact Physiological Signal Measurement and Applications
Doctoral candidate
Master of science Zhaodong Sun
Faculty and unit
University of Oulu Graduate School, Faculty of Information Technology and Electrical Engineering, Center for Machine Vision and Signal Analysis
Subject of study
Computer science and engineering
Opponent
Assistant Professor Ehsan Adeli, Stanford University
Custos
Docent Xiaobai Li, University of Oulu
Facial Video-based Non-contact Physiological Signal Measurement and Applications
Traditional physiological signal acquisition often necessitates contact-based or invasive approaches, which are not always practical or accessible. In contrast, remote photoplethysmography (rPPG) leverages facial videos to capture physiological signals through the subtle color variations in the face corresponding to blood volume changes. This thesis focuses on two aspects of rPPG: rPPG measurement methods and the applications of rPPG.
For the measurement methods, this thesis introduces innovative unsupervised and weakly supervised rPPG techniques to address label issues that are frequently encountered in rPPG datasets, e.g., missing and misaligned ground truth signals. The proposed approaches employ contrastive learning and integrate rPPG-specific prior knowledge, enabling the model to discern rPPG-relevant features. The proposed methods demonstrate performance on a par with or superior to existing supervised rPPG signal measurement techniques.
Regarding applications, this thesis harnesses rPPG for non-contact atrial fibrillation detection by analyzing facial videos and formulates a novel loss function aimed at enhancing heart rate variability (HRV) measurement through the identification of systolic peaks. The method achieves high atrial fibrillation detection accuracy (exceeding 95%). Additionally, the thesis validates the potential of using rPPG-derived HRV features from facial videos during virtual meetings to estimate stress levels. In the realm of security, the uniqueness of individual rPPG signal morphology is validated, enabling its use in biometric authentication. Furthermore, an algorithm is proposed to alter rPPG signals within facial videos, thus preventing their unauthorized extraction and safeguarding physiological privacy.
In summary, the thesis significantly contributes to the rPPG field by presenting innovative methods and validating their applications, thereby enhancing the prospects for non-contact physiological measurement and offering new avenues for advancements in telehealth, biometric security, and overall human well-being.
For the measurement methods, this thesis introduces innovative unsupervised and weakly supervised rPPG techniques to address label issues that are frequently encountered in rPPG datasets, e.g., missing and misaligned ground truth signals. The proposed approaches employ contrastive learning and integrate rPPG-specific prior knowledge, enabling the model to discern rPPG-relevant features. The proposed methods demonstrate performance on a par with or superior to existing supervised rPPG signal measurement techniques.
Regarding applications, this thesis harnesses rPPG for non-contact atrial fibrillation detection by analyzing facial videos and formulates a novel loss function aimed at enhancing heart rate variability (HRV) measurement through the identification of systolic peaks. The method achieves high atrial fibrillation detection accuracy (exceeding 95%). Additionally, the thesis validates the potential of using rPPG-derived HRV features from facial videos during virtual meetings to estimate stress levels. In the realm of security, the uniqueness of individual rPPG signal morphology is validated, enabling its use in biometric authentication. Furthermore, an algorithm is proposed to alter rPPG signals within facial videos, thus preventing their unauthorized extraction and safeguarding physiological privacy.
In summary, the thesis significantly contributes to the rPPG field by presenting innovative methods and validating their applications, thereby enhancing the prospects for non-contact physiological measurement and offering new avenues for advancements in telehealth, biometric security, and overall human well-being.
Last updated: 9.8.2024