Heart Rate Estimation in Intense Exercise Videos using Temporal Superpixels

Heart Rate Estimation in Intense Exercise Videos using Temporal Superpixels

Marwade, Anwesh (TU Delft Electrical Engineering, Mathematics and Computer Science)

van Gemert, J.C. (mentor)

Delft University of Technology

2021-07-26

Estimation of physiological signals like heart rate using remote photoplethysmography (rPPG) provides non-contact health monitoring with important applications in remote healthcare, patient interaction, and sport related activities. Existing rPPG approaches can robustly measure heart rate from facial videos under some degree of motion through face tracking and alignment. However, accurate face detection is not always feasible as the subject's face might be occluded or even outside the camera frame. This is especially the case for unconstrained settings like exercise and sports. Here we present IBIS-CNN, a novel approach to rPPG based heart rate estimation using spatio-temporal superpixels, which improves on existing models by eliminating the requirement for a visible face and face tracking. Experiments conducted on two publicly available rPPG datasets in addition to a self-collected dataset show that IBIS-CNN outperforms the existing rPPG methods which have difficulty in estimating heart rate in these unconstrained settings. We present experiments to analyse the issues with detection/alignment based methods, highlighting the promising potential for using temporal superpixels for rPPG based HR estimation using a convolutional neural network in IBIS-CNN.

Heart Rate estimation
Temporal Superpixels
Machine Learning
Physiology
Computer Vision

http://resolver.tudelft.nl/uuid:51c5e7c9-4c64-494d-acdd-96d3982cab0c

Student theses

master thesis

© 2021 Anwesh Marwade