Respiration Rate Estimation from Thermal Video of Masked and Unmasked Individuals Using Tensor Decomposition

Abstract

Respiration rate (RR) measurement is a crucial part of health monitoring. Long-term monitoring requires a non-contact solution to avoid inconveniencing patients and to promote compliance. This study examines the use of thermal video for RR estimation since it is more privacy-preserving than visible light cameras. In this study, tensor decomposition is implemented on the whole frame of thermal camera video, avoiding the need to define a specific region of interest (ROI). The proposed method automatically finds a loading vector of the decomposed tensor from which a time series respiratory signal may be reconstructed. The mean absolute difference between the estimated breaths per minute (BPM) from tensor decomposition and the ground truth measurement was used to evaluate the model. Data were recorded from 22 subjects with and without a mask while sitting or standing using a thermal camera; concurrent recordings with a respiratory effort belt provided gold standard RR. Results show that when subjects are seated and wearing a face mask, tensor decomposition can detect RR with an average error of 0.3 BPM over the 22 subjects. In the worst scenario (standing without a mask) the average error is 8.0 BPM. This study shows that tensor decomposition can be a useful tool for automatic non-contact RR estimation without using any pre-defined ROI, particularly for seated subjects wearing a face mask.