Dynamic Heart Rate Measurements from Video Sequences using Canonical Component Analysis

Dynamic heart rate computation from facial images obtained from video sequences has random artifacts and noises. A novel method is formulated by assuming that two video durations will contain the heart rate signals that are strongly correlated to each other while the random artifacts and noises are not correlated to each other. Canonical Component Analysis (CCA) is used to recover the heart-rate signals by determining the maximum correlation of the two video durations. The identified heart signal is then passed to a bandpass filter (0.8 - 4Hz) followed by Fast Fourier Transform to obtain the heart rate. Two experiments related to increasing and decreasing heart rate variations are carried out to determine the effectiveness of the proposed method. Eight subjects participated in each experiment, where their facial images were captured for a minute while they were cycling. Their heart rates varied from 83 to 153 beats per minute (BPM). The results of the proposed method are compared to a method using independent component analysis (ICA). The root mean square errors (RMSE) for the proposed method and ICA based-method that used 5-second video duration for the first and second experiments are 3.70 BPM and 2.33 BPM and 14.36 BPM and 9.72 BPM, respectively.