Facial Video-based Contactless Blood Pressure Estimation Using End-to-End Graph Convolutional Network
Poster presentation
Research approach
Divided facial skin regions into five-second clips and converted RGB to YUV. A two-layer GCN estimates blood pressure from a sample-similarity graph of skin-color changes without a separate pulse-extraction stage.
View full-size figure
Results
Transductive evaluation on VV-Medium with 250 participants and a subject-independent 70:30 split. Skin-color inputs yielded SBP/DBP MAE of 17.66/10.01 mmHg; rPPG inputs yielded 19.68/11.67 mmHg.
| Input | SBP | DBP |
|---|---|---|
| Contact PPG | 17.37 | 10.02 |
| rPPG | 19.68 | 11.67 |
| Skin color | 17.66 | 10.01 |