LSTM-based Heart Rate Estimation from Facial Video Images

The heart rate provides important information about a person’s health and emotional state. The processing of photoplethysmographic signals obtained from videos is so far strongly dependent on temporal filtering operations. In this work, we propose a custom long short-term memory (LSTM) architecture that is able to learn the processing of such signals and the associated specific perturbations in order to produce very accurate estimates of the human heart rate from video images. Our trained network was tested on three different datasets and delivered very promising results for each of them. When comparing the performance with other state-ofthe-art methods, our network was able to outperform all the others.