The team from Great Bay University has presented PhysMamba, an innovative AI framework for measuring heart rate and related physiological signals through facial videos. This important innovation in remote photoplethysmography (rPPG) provides a contactless method for health monitoring, which opens new opportunities for immediate medical and wellness applications.
PhysMamba sets itself apart from earlier rPPG methodologies that mostly relied on Convolutional Neural Networks (CNNs) and Transformers. These traditional measurement approaches often found it difficult to accurately capture the essential long-range temporal dependencies crucial for heart rate measurement, especially when dealing with longer video sequences. PhysMamba resolves these barriers by introducing a state-of-the-art Temporal Difference Mamba (TD-Mamba) block alongside a dual-stream SlowFast architecture. By doing this, the model effectively processes short-term and long-range temporal features, thereby increasing its accuracy in detecting precise physiological signals. You can read the paper here.
Through a series of detailed experiments on benchmark datasets, including PURE, UBFC-rPPG, and MMPD, PhysMamba showed impressive advancements compared to current models. This resulted in decreased error rates and heightened accuracy in heart rate estimation. Significantly surpassing typical CNN and Transformer models, the innovative framework was particularly effective in real-world situations affected by lighting variations and facial movements.
This new version of an AI model, embraced by CCBR 2024, is a crucial development in noninvasive physiological monitoring. The research team has released the code for PhysMamba on GitHub, granting opportunities for additional research and development in this exciting domain of computer vision and health technology.
