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.

Tese de mestrado, Engenharia Biomédica e Biofísica , 2024, Universidade de Lisboa, Faculdade de Ciências