A Maximum Likelihood Formulation To Exploit Heart Rate Variability for Robust Heart Rate Estimation From Facial Video

Area of Science:

• Biomedical Engineering
• Signal Processing
• Computer Vision

Background:

• Estimating heart rate (HR) from videos is challenging.
• Current methods often use independent component analysis (ICA) on facial color profiles.
• These methods may not ensure realistic heart rate variability (HRV).

Purpose of the Study:

• To propose a novel maximum likelihood formulation for HR estimation from videos.
• To introduce a new scheme, termed vICA, that optimizes HR trajectory selection.
• To ensure realistic HRV by selecting source signals that align with spectral peaks and temporal smoothness.

Main Methods:

• Developed a maximum likelihood formulation for HR estimation.
• Utilized dynamic programming, similar to Viterbi decoding, for efficient optimization.
• Compared the proposed vICA scheme against traditional ICA and sparse spectral peak tracking (SSPT) methods.
• Conducted experiments using videos from 15 subjects recorded on two smartphone types at varying distances.

Main Results:

• vICA demonstrated superior performance compared to baseline methods.
• Significant improvements in mean absolute error (MAE) were observed across different phone types and distances.
• For Samsung Galaxy, MAE improvements were -8.69%, 52.77%, and 8.00% at 6 inches, 1 foot, and 2 feet, respectively.
• For iPhone, MAE improvements were 12.13%, 13.59%, and 18.34% at the same distances.

Conclusions:

• The proposed vICA scheme offers more accurate HR estimation from videos.
• Incorporating HRV smoothness into HR trajectory prediction enhances accuracy.
• vICA represents a significant advancement in non-contact HR monitoring using video.