MaKAN-Mixer: Channel Interaction-Based Mamba Method for rPPG Extraction.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) enables non-contact heart rate monitoring via facial video analysis.
• Challenges in rPPG include variable lighting, motion artifacts, and capturing spatio-temporal dynamics.
• Existing methods struggle with long-term dependencies and complex physiological signal extraction.

Purpose of the Study:

• To introduce MaKAN-Mixer, an end-to-end network for robust and accurate rPPG signal extraction.
• To improve rPPG performance in complex environments and under challenging conditions.
• To enhance heart rate estimation by effectively modeling spatio-temporal features and long-term dependencies.

Main Methods:

• Integration of Hybrid of Eulerian Video Magnification and Temporal Shift Module Amplification (HETA) for signal amplification.
• Development of Mamba-KAN Fusion Module (MKFM) for efficient long-term dependency modeling and channel fusion.
• Utilizing a KAN Feedforward Neural Network (KFN) for capturing complex physiological patterns.

Main Results:

• MaKAN-Mixer demonstrates superior performance on four benchmark datasets in both intra- and cross-dataset testing.
• Achieved significant reduction in Root Mean Square Error (RMSE) compared to the state-of-the-art.
• Exhibited exceptional robustness in challenging scenarios, including compressed video and complex environments.

Conclusions:

• MaKAN-Mixer offers a significant advancement in non-contact heart rate monitoring technology.
• The proposed network architecture effectively addresses limitations in current rPPG methods.
• Results underscore the potential for accurate, real-world rPPG monitoring applications.