A carregar...
Publicação
Remote Heart Rate Estimation Leveraging Eulerian Video Magnification
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 4.52 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Introduction: Measuring physiological signals like Heart Rate, Respiration Rate, and Blood Pressure is essential for health assessment but often requires physical contact, which is a limitation emphasized
by situations such as the SARS-CoV-2 pandemic. This constraint has spurred the development of remote
monitoring solutions, with Remote Photoplethysmography (rPPG) being a prominent method. Via captured video, rPPG detects subtle color changes in the skin due to blood flow, though it is sensitive to
noise like motion and lighting variations. To enhance robustness, the study investigates the use of Eulerian Video Maggnification (EVM), which amplifies color changes for better explainability and signal
processing.
Objectives: The study aimed to develop a robust remote Heart Rate estimation model using EVM
and deep learning. It also aimed to create a public dataset of facial videos, supporting further research in
this field.
Methods: The dataset included videos with Electrocardiogram as ground truth and additional subject/environmental details. The EVM technique followed Wu’s original methodology, emphasizing color
magnification to highlight subtle skin changes [1]. The deep learning model, incorporates 1D CNNs and
LSTM layers for effective temporal pattern recognition. Performance was assessed using MAE, MAPE,
and RMSE metrics.
Results: EVM showed potential but suffered from performance drops under noise like lighting
changes, particularly at higher heart rates. The forehead region yielded better results than cheeks. The
proposed model achieved a significant performance improvement over baselines, with an MAE of 4.57±
0.87 bpm and overall better performance when trained on the forehead region. Cross-region training
improved results but with variable success.
Conclusion: EVM and rPPG remain sensitive to noise and require either controlled conditions
or adaptations for practical use. The deep learning model demonstrated improvements over baseline
methods but was affected data variability, highlighting the need for further refinement.
Descrição
Tese de mestrado, Engenharia Biomédica e Biofísica , 2024, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Fotopletismografia Remota (rPPG) Magnificação Euleriana de Vídeo (EVM) Estimação do Batimento Cardíaco Aprendizagem Profunda Teses de mestrado - 2024