CHILL

Remote photoplethysmography (rPPG) offers a non-invasive means of estimating heart rate in telemedicine settings. Yet its reliability remains uncertain due to the limited diversity and ecological validity of existing benchmark datasets. In this work, we systematically investigate the robustness of rPPG methods under challenging conditions, specifically low illumination and elevated heart rates. We introduce the CHILL dataset, which comprises video and PPG signals collected from 45 participants across two lighting conditions (bright and dark) and exercise-induced heart rates ranging from 54 to 141 beats per minute. We assessed eight rPPG algorithms, including four signal processing-based and four deep learning-based approaches, across three datasets: the newly collected CHILL dataset and two widely used public benchmarks, PURE and COHFACE. Within-dataset analysis on the CHILL dataset revealed that many existing rPPG methods struggle under challenging conditions. Five of the eight methods experienced a statistically significant decline in performance at elevated heart rates. In contrast, low illumination had a comparatively smaller impact. Cross-dataset analysis further revealed that several deep learning models failed to generalize effectively to the CHILL dataset. Among the models that did generalize, many still showed a significant decline in performance under elevated heart rate conditions, regardless of the training dataset. These findings highlight a critical limitation in current rPPG algorithms, namely their susceptibility to high heart rates. Our evaluation of rPPG methods on the CHILL dataset underscores the need for more robust approaches to enable accurate, non-invasive physiological monitoring in real-world digital health environments.

Verbatim fromthe publisher's DOI record· Publisher licence: https://creativecommons.org/licenses/by/4.0