Dual-Stream Fusion for Advanced Spatio-Temporal Analysis in Remote Photoplethysmography

Remote photoplethysmography (rPPG) enables contactless heart rate (HR) estimation by detecting subtle skin color variations. However, existing methods struggle with illumination changes, motion artifacts, and head movement, reducing accuracy. To address these challenges, we propose a dual-stream framework that combines local spatial features with global temporal dynamics. The main stream enhances signal robustness through multi-scale feature extraction and attention mechanisms, while the auxiliary stream leverages adaptive color space transformations and a convolutional LSTM network to capture long-range temporal dependencies. Evaluations on the VIPL-HR dataset and real-time monitoring scenarios show that our approach significantly reduces mean absolute error (MAE) and root mean square error (RMSE) compared to state-of-the-art methods, demonstrating its effectiveness for practical applications.