Conventional and deep learning methods in heart rate estimation from RGB face videos

Area of Science:

• Biomedical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Traditional vital signs monitoring often requires physical attachments, limiting patient comfort and mobility.
• Contactless monitoring offers a promising alternative, leveraging advanced technologies for remote physiological assessment.
• Extracting heart rate (HR) from facial cues is a key area within contactless monitoring.

Purpose of the Study:

• To provide a comprehensive review of state-of-the-art methods for contactless heart rate estimation.
• To evaluate both conventional signal processing and deep learning approaches for HR extraction from RGB face videos.
• To identify the limitations of current deep learning methods and challenges with less-controlled video datasets.

Main Methods:

• Review of established signal processing techniques for remote photoplethysmography (rPPG).
• Analysis of deep learning architectures, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), for HR estimation.
• Evaluation of methods based on their performance with diverse and less-controlled facial video datasets.

Main Results:

• Both conventional and deep learning methods show potential for accurate HR estimation.
• Deep learning methods, while powerful, face challenges related to data variability and generalizability.
• The availability and quality of training datasets significantly impact the performance of contactless HR monitoring systems.

Conclusions:

• Contactless HR monitoring is a rapidly evolving field with significant potential for healthcare applications.
• Further research is needed to address the limitations of deep learning methods, particularly concerning robustness and adaptability to varied conditions.
• Understanding the benefits and drawbacks of different rPPG extraction techniques is crucial for future advancements.