Computer Vision-Based Estimation of Respiration Signals

Abstract

Respiration movement and respiration rate have been used to monitor breathing status for diagnosis and fitness purposes. From a given video sequence of a person facing a camera, this system here automatically detects and tracks a region of interest (ROI) on the chest, using the Kanade-Lucas-Tomasi method, after applying the Viola-Jones algorithm and identifying the Harris-Stephens features for tracking the ROI across frames. The displacement in the vertical direction of the ROI (frame by frame) was used to estimate the respiration movement, after low-pass filtering at a proper cutoff frequency. Finally, the respiration rate is estimated from the respiration movement signal by a root MUSIC-based estimator. For the 12 video files provided, we obtained respiration movement signals with correlation indexes respect to the corresponding ‘references’ of above 90%, in most cases, and respiration rate signals with normalized root mean-squared errors with respect to the inst_freq around 10%. A global ranking index of around 0.8 was consistently obtained. Computer vision algorithms are well-suited for estimating respiration movement and respiration rate signals from frontal video sequences.