Robust Remote Heart Rate Estimation Network Based on Spatial-Temporal-Channel Learning From Facial Videos

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Non-contact heart rate estimation using video relies on subtle skin color changes.
• Current video-based methods struggle with feature extraction, spatial redundancy, and motion artifacts.

Purpose of the Study:

• To propose a novel end-to-end network, Spatial-Channel Network (STCNet), for accurate non-contact heart rate estimation.
• To address limitations in feature extraction, spatial information redundancy, and motion artifact handling in existing methods.

Main Methods:

• Developed a Spatial Attention Learning (SAL) unit to focus on relevant facial regions.
• Introduced an improved Temporal Shift Module (TSM) for enhanced long-range temporal perception.
• Designed a Temporal-Channel Learning (TCL) unit for cross-frame channel information interaction and periodic heartbeat feature extraction.
• Integrated SAL and TCL into a Feature Extraction Block (FEB) and stacked FEBs to form the STCNet architecture.

Main Results:

• STCNet demonstrated superior performance on the UBFC-rPPG and PURE datasets.
• Achieved a 0.27 bpm reduction in Mean Absolute Error (MAE) and 0.19 bpm in Root Mean Square Error (RMSE) compared to CIN-rPPG on the PURE dataset.
• Experimental results confirm STCNet's effectiveness and generalization ability, outperforming mainstream models.

Conclusions:

• The proposed STCNet effectively addresses limitations in current video-based heart rate estimation.
• STCNet offers a robust and accurate solution for non-contact physiological monitoring.
• The model shows significant improvements in accuracy and artifact handling for real-time health monitoring applications.