Abstract
Remote photoplethysmography (rPPG) enables non-contact heart rate monitoring by detecting subtle skin color changes in video. However, traditional methods rely heavily on face detection, making them vulnerable to motion, specific clinical conditions, and privacy concerns. This paper introduces HR-RST, a framework utilizing a brain-inspired Continuous Coupled Neural Network (CCNN) to track skin pixels based on their chaotic signal characteristics. By decoupling heart rate measurement from facial landmarks, the system achieves robustness against movement and extends monitoring to non-facial body parts, enhancing both clinical utility and biometric privacy.
Introduction
Remote heart rate measurement (rPPG) is a technique that extracts a pulse signal by analyzing invisible-to-the-eye color changes in skin video. While promising for non-invasive health monitoring, this technology often fails under real-world conditions such as subject movement, lighting variations, or when a face cannot be clearly identified. Traditional rPPG pipelines are inherently fragile because they depend on rigid face-tracking algorithms to define a Region of Interest (ROI).
The Limitations of Face-Centric rPPG
Most state-of-the-art rPPG methods follow a rigid sequence: detect the face, define an ROI, and extract the signal. This approach suffers from three major "fail states":
- Motion Sensitivity: Rapid head movements or tilts often cause face trackers to lose their grip, resulting in failed or inaccurate signal extraction.
- Clinical Constraints: Models trained on healthy adult faces frequently fail for infants or burn victims, whose skin textures or movement patterns differ significantly.
- Privacy Risks: The requirement for high-resolution facial video creates substantial biometric privacy concerns in medical monitoring environments.
The HR-RST Framework
The proposed Heart Rate - Robust Skin Tracking (HR-RST) framework shifts the focus from finding faces to identifying skin based on the unique "shimmer" of the pulse. This brain-inspired approach operates in three distinct phases:
1. ROI Extraction via CCNN
Instead of standard convolutional networks, HR-RST uses a Continuous Coupled Neural Network (CCNN) that mimics the primary visual cortex. The system distinguishes skin from background by analyzing signal characteristics. The CCNN encodes dynamic pixels (skin with a pulse) into chaotic signals, while static background pixels are encoded as periodic signals. By applying a Continuous Wavelet Transform (Gaussian wavelets) to the neuron outputs, the system isolates skin pixels without requiring facial landmark detection.
2. Signal Analysis
Once the skin ROI is isolated, the system processes the Green color channel, which carries the strongest pulse signal due to oxyhemoglobin absorption. A third-order Butterworth band-pass filter (0.7Hz to 2.5Hz) is applied to target the human heart rate range (42–150 bpm).
3. Pixel-wise HR Calculation
Unlike traditional methods that average the entire ROI into a single signal, HR-RST calculates the heart rate for every individual pixel using Power Spectral Density (PSD) estimation via the Welch method. The final heart rate is determined by the mode of these pixel-wise measurements. This approach provides extreme resilience against local noise, shadows, or partial occlusions.
Experimental Evaluation
The framework was validated using the VIPL-HR Database (2,378 videos) and a Custom Non-Facial Dataset featuring body parts such as the palm, forearm, and sole of the foot.
Key Findings
- Versatility: The system successfully extracted heart rate signals from non-facial areas, enabling "de-identified" monitoring where a patient's face is never recorded.
- Robustness: The CCNN-based tracker maintained skin contact during significant head rotations (pitch/yaw) where traditional face-detectors failed.
- Quantitative Performance: On the VIPL-HR dataset, the system showed consistent performance across subjects; for example, Sample p1 achieved a Mean Absolute Error (MAE) of 12.41 bpm. While deep learning models may achieve lower error in controlled settings, HR-RST offers higher interpretability and functions without massive training datasets.
Conclusion
HR-RST demonstrates that treating the pulse as a chaotic perturbation allows for robust, privacy-preserving heart rate monitoring. By moving away from face-centric detection, the system opens new possibilities for clinical monitoring in scenarios where traditional sensors or face-tracking are impractical, such as neonatal care and burn units.
Summary updated