Respiratory Rate Estimation from Thermal Video Data Using Spatio-Temporal Deep Learning

Area of Science:

• Biomedical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Remote health monitoring requires privacy-preserving methods.
• Respiration rate (RR) estimation is crucial for health assessment.
• Thermal imaging offers a non-intrusive data source for physiological monitoring.

Purpose of the Study:

• To develop an end-to-end deep learning model for accurate RR estimation using thermal video data.
• To introduce a novel loss function addressing phase shifts in respiration measurement.
• To evaluate the model's performance across various conditions, including face mask usage.

Main Methods:

• Utilized a detection transformer (DeTr) for facial region identification.
• Employed 3D convolutional neural networks and bi-directional LSTMs for respiratory signal extraction.
• Introduced a novel loss function combining negative maximum cross-correlation and absolute frequency peak difference.

Main Results:

• The proposed method achieved an average error of 1.6 breaths per minute.
• Outperformed existing RR estimation models across four tested conditions (sitting/standing, with/without mask).
• Demonstrated state-of-the-art accuracy for RR estimation from thermal video.

Conclusions:

• The developed deep learning approach offers highly accurate and privacy-preserving RR estimation.
• The method is suitable for real-time applications in remote health monitoring.
• Thermal video analysis presents a promising avenue for non-contact physiological monitoring.