Eulerian Magnification and Transformers for Enhanced rPPG-Based Pulse Rate Estimation in Indian Populations

This research focuses on pulse rate prediction using video-based remote Photoplethysmography (rPPG) techniques applied to an Indian dataset. In this study, we developed a hybrid model that integrates 3D Convolutional Neural Networks (3D CNN), Transformer architecture, and Eulerian Video Magnification (EVM) to enhance pulse rate estimation from facial video data. EVM amplifies subtle color variations in facial regions, while the combination of the 3D CNN and Transformer network captures spatiotemporal patterns from video frames. We collected one-minute facial videos from 40 Indian participants using smartphone cameras, along with pulse and respiratory rate measurements obtained via a fingertip pulse oximeter. Experimental results demonstrate that integrating EVM with Transformers improves performance metrics, achieving lower Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The proposed model outperformed existing supervised and unsupervised methods, achieving an MAE of 12.47 bpm. This work contributes to advancing video-based PPG methods tailored to the Indian population.