Comparison of Blind Source Separation Techniques for Respiration Rate Estimation from Depth Video

Non-contact respiration rate estimation can form an important part of health screening in public spaces or clinical settings. Depth cameras can provide contactless information related to changes in chest and/or abdomen volume during respiration; however, most existing approaches require manual definition of the region of interest (ROI). In this study, respiration rate estimation was examined using depth camera data without defining an ROI. Instead, multiple time-varying signals were extracted from the depth video by computing spatial averages across an arbitrary grid of sub-images. Blind source separation (BSS) methods, specifically canonical correlation analysis (CCA) and independent component analysis (ICA), were applied to these signals and the sources most relevant to respiration rate estimation were identified automatically. In this study, CCA achieved a mean absolute error of 2.89 breaths per minute from eight free-standing participants, whereas ICA resulted in larger errors. Based on these findings, it can be concluded that the respiratory sources had low correlation with, but were not independent of, non-respiratory sources in this depth camera dataset. Therefore CCA resulted in better performance compared to ICA.