Amplifying the Invisible: Enhancing Pulse Estimation with EVM and Transformers
In this project, I explored how video-based physiological monitoring could be significantly improved using a fusion of signal amplification and deep learning. I designed a hybrid system that combines Eulerian Video Magnification (EVM), 3D CNNs, and Transformer architectures to estimate pulse rates from facial videos. By amplifying subtle skin color variations and modeling long-range temporal dependencies, this method achieved a mean absolute error (MAE) of 12.47 bpm—outperforming existing approaches on Indian datasets. This work is part of a broader push to build inclusive, non-contact health monitoring tools tailored to real-world populations.
