Impact of motion artifacts on video-based non-intrusive heart rate measurement

Measuring vital signs such as heart rate using a camera has the potential to enable better health monitoring for subjects at risk and as such enhance their quality of life. Applications could include driver monitoring via in-dash camera, critical function operator monitoring at work, or remote health monitoring via a webcam. For such a system to be feasible however, it needs to be work well in realistic scenarios where the subject does not sit completely still in front of a camera. Motion artifacts, if not taken into account when designing the system, yield inaccurate results and potentially create false alarms. In this paper, we start with a popular algorithm for extracting heart rate from video based on spatial and temporal filtering, quantify how key parameters used in the algorithm affect its performance in situations when the subject is not sitting still, analyze in detail the performance of the filtering approach in videos with motion, identify issues, and propose approaches to overcome those limitations. The paper shows that the use of wider filters and more levels in the Gaussian pyramid lead to a better performance when the subject is moving, but that the motion artifacts dominate the extracted signal.