A Machine Learning-Based Non-Contact Respiratory Rate Monitoring Method Using an RGB Camera

Abstract

Non-contact measurement of Respiratory Rate (RR) is suitable for various clinical and home-based healthcare applications. RR monitoring can inform healthcare providers of early indicators of critical illnesses. However, the obtrusive nature of contact-based sensors for RR monitoring makes them uncomfortable for extended use and vulnerable to movement-derived noise. Hence, camera-based approaches have attracted considerable attention as they enable contact-free RR monitoring. This thesis presents an improved non-contact method for RR monitoring that leverages camera derived remote photoplethysmography (rPPG) to measure RR. Unlike previous work, the proposed method supports subject movement during monitoring. We apply Independent Component Analysis (ICA) on the RGB channels of facial videos to distinguish the source (i.e. PPG signal) from noise. We use the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) scheme to decompose the selected ICA output into its Intrinsic Mode Functions (IMFs). We propose a Machine Learning (ML) algorithm to select the IMF that best reflects the RR. We evaluated the proposed method on 200 facial videos collected from 10 subjects. Our approach decreased the RMSE by at least 39.6% compared to state-of-the-art techniques when subjects were stationary. For subjects in movement, we achieved an RMSE of 2.30 BPM (breaths/min).