A non-contact heart rate estimation method using video magnification and neural networks

Heart rate (HR) monitoring is a significant task in many medical, sports and aged care in assisted living applications, among other disciplines. In the literature, several works have reported effectiveness in addressing the measurement of HR using contact sensors such as adhesive or dry electro-conductive electrodes. However, there are several issues associated with contact sensors like portability problems, skin irritation, discomfort and body movement constraints. In this regard, this paper presents a non-contact HR estimation method using vision-based methods and neural networks. This work uses a bio-inspired Eulerian motion magnification approach to highlight the blood irrigation process of the cardiac pulse, which is later inputted to a feed-forward neural network trained to estimate the HR. For experimental analysis, we compare two magnification procedures, based on Gaussian and Hermite decomposition, over video recordings collected from the wrists of five subjects. Results show that the Hermite-based magnification method is robust under noise analysis (4.24 bpm of root mean squared-error in the worst case scenario). Furthermore, our results demonstrate that the Hermite-based method is competitive in the state-of-the-art (1.86 bpm in average of root mean squared-error) and can be implemented using a single camera for contactless HR estimation.