Skip to content

Contactless vital estimation and deployment

Adapted facial pulse-estimation models to Jetson hardware and supported kiosk deployment.

Related research: glcl_rPPG

Role: Algorithm co-development and inference optimization
Contribution: Developed rPPG models, converted inference to TensorRT FP16, and integrated embedded inference with the kiosk.
Period & affiliation: 2025.05 ~ 2025.12 TIPA, CUDO Communication

01 / RESEARCH DETAILRemote pulse estimation and kiosk deployment

Challenge: Remote pulse models need physiological representations and an inference path suitable for embedded hardware.
Approach: Aligned global and local pulse spectra from the same video to learn shared physiological components without pulse labels.
Deployment: Co-developed models, converted inference to TensorRT FP16, integrated Jetson Nano and supported HEALLO field deployment.

Global and local contrastive learning of spatial pulse spectra — Global and local contrastive learning of spatial pulse spectra
View full-size figure

Manuscript Fig. 3(a): UBFC-rPPG scatter and Bland-Altman plots — Manuscript Fig. 3(a): UBFC-rPPG scatter and Bland-Altman plots
View full-size figure

Heart-rate errors reported in the GLCL manuscript
GLCL manuscript evaluation	MAE (bpm)	RMSE (bpm)
UBFC-rPPG	1.09	1.72
PURE	0.26	0.42
COHFACE	2.19	5.83

Evaluation conditions and scopeGLCL manuscript evaluation on public datasets; kiosk field performance was not measured in these experiments.