Combating the Impact of Video Compression on Non-Contact Vital Sign Measurement Using Supervised Learning

Abstract

Imaging photoplethysmography (iPPG) and imaging ballistocardiography (iBCG) are popular approaches for unobtrusive camera-based measurement of vital signs. These involve recovering pulse signals from very subtle variations in video pixel intensities, which are easily corrupted by noise. Therefore, while the signal might be easy to obtain from high quality uncompressed videos, the signal-to-noise ratio drops linearly with video bit-rate. Uncompressed videos require large amounts of storage making them prohibitive to store, stream and transfer in large quantities. By learning compression specific models we show that supervised learning can be used to increase the signal-to-noise ratio (SNR) of pulse signals and reduce the mean absolute error (MAE) of heart rate estimates extracted from temporally compressed videos. We perform a systematic evaluation of the performance of our algorithm showing that the network trained on compressed videos consistently outperforms the model trained on the original less compressed compressed videos, both on videos with and without significant head motions. We found improvements in SNR of up to 8 dB and MAE of 6 BPM.