Signal quality-guided adaptive denoising for robust remote photoplethysmography

This study presents a Signal Quality-Guided Adaptive Denoising (SQAD) framework for improving the robustness of remote photoplethysmography (rPPG) under nonstationary artifacts. The main idea is to use signal quality estimation as an explicit control variable for rPPG denoising. After lightweight motion compensation, a calibrated support vector machine estimates probabilistic signal quality scores for each facial region using rPPG-specific features. These scores are then used to select reliable regions for ROI-level spatial fusion and to adapt the Kalman gain during frequency-domain spectral tracking. Experiments on Vicar-PPG2, ECG-Fitness, and UBFC-rPPG demonstrated that the proposed framework outperformed existing noise reduction methods across multiple pulse extraction algorithms, with the best within-dataset mean absolute errors of 1.66, 9.05, and 1.24 bpm, respectively. The results suggest that SQAD provides a lightweight and interpretable strategy for robust camera-based heart rate monitoring.