Respiration Rate Estimation from remote Photoplethysmography using Deep Neural Networks

Respiration Rate (RR) is one of the vital parameters and a good indicator of the health condition of a person. RR can be remotely measured from remote Photoplethysmography (rPPG), a technique to measure blood flow in blood vessels, based on the skin colour changes using RGB video cameras. Normal Photoplethysmography (PPG) measures the blood flow using contact-based methods. The research methods that extract RR from rPPG are mostly based on signal processing techniques like wavelet decomposition. The thesis work suggests an alternative way to extract RR from rPPG using deep neural networks. A convolutional autoencoder was trained to transform the rPPG signal into respiration signal and RR was estimated from the respiration signal using Fast Fourier Transform (FFT). Benchmark datasets were identified in the field of RR estimation to train and test the model. The model was trained on WESAD and PPG-DaLiA datasets together and tested on BIDMC, Capnobase, and MAHNOB-rPPG datasets. The training datasets were PPG based and testing datasets were both PPG and rPPG based datasets. The model was able to transform both PPG and rPPG signals into respiration signals accurately. The algorithm achieved good results with RRError of 3.55 BPM, 3.25 BPM, 3.9 BPM, and 3.19 BPM on the test, BIDMC, Capnobase, and MAHNOBrPPG datasets respectively.