Heart Rate Measurement Based on 3D Central Difference Convolution with Attention Mechanism

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) offers non-contact heart rate monitoring using video.
• Existing rPPG methods often struggle with effectively utilizing spatiotemporal video features for accurate signal extraction.

Purpose of the Study:

• To propose a novel deep learning network, CDCA-rPPGNet, for enhanced rPPG signal extraction.
• To improve the accuracy of non-contact heart rate measurement by effectively integrating spatial and temporal features.
• To address the limitations of existing methods in capturing subtle video signal variations.

Main Methods:

• Development of a 3D central difference convolutional network (CDCA-rPPGNet) incorporating an attention mechanism.
• Implementation of facial landmark detection for region of interest (ROI) extraction and stitching.
• Utilizing central difference convolution to enhance spatiotemporal feature representation and time-difference information capture.

Main Results:

• The CDCA-rPPGNet achieved Mean Absolute Error (MAE) of 0.46 bpm and Root Mean Square Error (RMSE) of 0.90 bpm on the PURE dataset.
• On the UBFC-rPPG dataset, the method obtained MAE of 0.60 bpm and RMSE of 1.38 bpm.
• Pearson's correlation coefficient (R) reached 0.99 on both datasets, demonstrating high accuracy.

Conclusions:

• The proposed CDCA-rPPGNet effectively captures subtle color changes and spatiotemporal features for accurate rPPG signal recovery.
• The ROI extraction method provides high-quality input, and the attention module enhances feature extraction capabilities.
• The experimental results validate the effectiveness of the proposed approach for non-contact heart rate measurement.