RhythmNet: End-to-end Heart Rate Estimation from Face via Spatial-temporal Representation

Area of Science:

• Physiological signal monitoring
• Computer vision
• Machine learning

Background:

• Traditional heart rate (HR) monitoring requires contact, causing discomfort.
• Existing remote HR estimation methods struggle in unconstrained environments.
• Limited large-scale datasets hinder deep learning for remote HR estimation.

Purpose of the Study:

• To develop an end-to-end deep learning model for accurate remote HR estimation.
• To create a large-scale, diverse dataset for training and evaluating HR estimation models.
• To improve HR estimation robustness in real-world, less-constrained scenarios.

Main Methods:

• Proposed RhythmNet, an end-to-end convolutional neural network.
• Utilized spatial-temporal representations from facial regions of interest.
• Incorporated Gated Recurrent Unit (GRU) for temporal HR dynamics.
• Introduced VIPL-HR, a large-scale multi-modal HR database.

Main Results:

• RhythmNet demonstrated superior performance compared to state-of-the-art methods.
• The VIPL-HR database includes diverse variations (head movement, illumination).
• The model achieved efficient and accurate HR measurements.

Conclusions:

• RhythmNet offers a robust solution for remote HR estimation.
• The VIPL-HR database facilitates advancements in unconstrained HR monitoring.
• This work advances non-contact physiological signal measurement.