Improving the Accuracy of Camera-Based Heart Rate Measurement

The heart rate (HR) measurements based on the camera (visible light) can be used to detect HR in non-contact mode, which has great application prospects both in the clinical application and home health care. However, CMOS sensors equipped with “rolling shutters”, which distinguishes different lines per frame to become light sensitive at different moments in time, and stylized dithering of image acquisition (IMAQ) time caused by different computer programs running in the background will greatly influence the accuracy of the measured HR. In this paper, we analyze the phase error caused by CMOS sensor and the system error introduced by system sampling clock jitters. According to derivation, we propose two methods, amplitude-frequency superposition and a cubic spline interpolation reconstruction method based on actual schedules, that can be widely utilized in computer vision to overcome the camera phase error and sampling time fluctuation error. Amplitude of signal is analyzed and processed in amplitude-frequency domain in the method of amplitude-frequency superposition, which ignores the signal phase. Thus it can eliminate the phase error effectively. The cubic spline interpolation reconstruction method based on actual schedules can reconstructed the non-uniform sampling of images as uniform ones, so it can eliminate the system error involved by the system clock jitters. What’s more, the properties of the methods are tested by applying them to both simulation experiments and real HR measurements. In the simulation, amplitude of measured signal is improved 4.58% relative to the amplitude measured without the method of amplitude-frequency superposition; root mean square error of signal’s frequency, detected by the cubic spline interpolation reconstruction method based on actual schedules, is reduced more than 30%. In the real HR measurements, the amplitude of HR is raised to 33.5% relatively based on amplitude-frequency superposition. And the accuracy of HR is raised to approximately 40% by the method of cubic spline interpolation reconstruction method based on actual schedules. Therefore, the simulation experiments and real HR measurement proof that we can effectively eliminate the camera phase error based on the amplitude-frequency superposition extraction method, and the cubic spline interpolation based on the timetable method can effectively reduce the random error in IMAQ due to system clock jitters. These methods can both be widely used in dynamic signal detection based on machine vision.