X-iPPGNet: A novel one stage deep learning architecture based on depthwise separable convolutions for video-based...

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Contactless pulse rate estimation is crucial for long-term health monitoring.
• Existing imaging photoplethysmography (iPPG) algorithms struggle in real-world, unconstrained scenarios like head movements and varying environmental conditions.
• There is a need for robust and accurate non-invasive pulse rate monitoring techniques.

Purpose of the Study:

• To propose a novel end-to-end spatio-temporal network, X-iPPGNet, for instantaneous pulse rate estimation directly from facial video recordings.
• To develop a method that learns iPPG from scratch, without prior knowledge or explicit blood volume pulse signal extraction.
• To improve the robustness and accuracy of contactless pulse rate estimation in challenging conditions.

Main Methods:

• Developed X-iPPGNet, an end-to-end spatio-temporal deep learning network inspired by the Xception architecture.
• Utilized color channel decoupling to enhance photoplethysmographic information capture and reduce computational load.
• Enabled pulse rate prediction from a short 2-second time window, beneficial for fluctuating heart rates.

Main Results:

• X-iPPGNet demonstrated high performance across various conditions, including head motions, facial expressions, and different skin tones.
• The model significantly outperformed state-of-the-art methods on three benchmark datasets (MMSE-HR, UBFC-rPPG, MAHNOB-HCI).
• Achieved Mean Absolute Errors (MAE) of 4.10, 4.99, and 3.17 on the respective datasets, with high correlation coefficients (r).

Conclusions:

• X-iPPGNet offers a robust and accurate solution for contactless pulse rate estimation from facial videos.
• The proposed method effectively handles unconstrained scenarios, making it suitable for real-world applications.
• This approach advances non-invasive health monitoring by providing reliable pulse rate measurements without physical contact.