Skin-tone-aware AI estimates breathing rate from video, cuts error by 42%
No wearables needed: a new dual-representation rPPG framework adapts to all skin tones.
A new paper on arXiv (2606.21511) introduces a skin-tone-aware dual-representation remote photoplethysmography (rPPG) framework for contactless respiratory rate estimation. Traditional rPPG methods rely on fixed green/chrominance projections designed for heart rate, missing subtle respiratory dynamics. The proposed system uses a dynamic RGB projection that adapts to skin tone, paired with a denoising network to filter non-respiratory motion from Lagrangian (motion-based) signals. A phase-independent contrastive loss forces Eulerian (color-based) and Lagrangian representations to collaboratively learn respiratory information.
Evaluated on the new RR-rPPG dataset (facial videos with Indian demographics) and the public COHFACE dataset, the framework consistently outperforms prior methods, achieving up to a 42.1% reduction in mean absolute error. This work tackles the under-explored area of rPPG-based respiratory rate estimation while addressing algorithmic bias toward lighter skin tones. The code and dataset will be released upon paper acceptance, providing a benchmark for diverse remote respiratory monitoring research.
- Dynamic RGB projection adapts to skin tone, unlike fixed green/chrominance methods that underperform on diverse tones.
- Denoising network reduces non-respiratory motion artifacts in Lagrangian (motion-based) rPPG signals.
- New RR-rPPG dataset includes Indian demographic representation, filling a gap in existing rPPG benchmarks.
Why It Matters
Enables accurate, contactless respiratory monitoring from any camera—critical for telehealth, sleep studies, and pandemic triage.