FastBVP-Net: a lightweight pulse simulation network for measuring heart rhythm via facial videos

Remote photoplethysmography (rPPG) is a useful camera-based health motioning method that can estimate the heart rate (HR) and heart rate variability (HRV) from facial videos. Many well-established deep learning models can provide highly accurate and robust results in heart rhythm measurement using rPPG. However, most previous models usually require enormous computational resources and a 30-second facial video, which significantly limits their applications in real-world scenarios. Hence, we propose a lightweight pulse simulation network named FastBVP-Net to measure heart rhythm via facial videos. In the FastBVP-Net, we designed a multi-frequency mode signal fusion mechanism to suppress the noise components and get stable blood volume pulse (BVP) signals quickly. Moreover, we developed an oversampling training strategy to solve the unbalanced distribution of HR in the dataset. Finally, we estimate the HR and HRV based on BVP signals derived by the FastBVP-Net. Comprehensive experiments are conducted on three benchmark datasets, our approach achieves competitive performance both on HR and HRV estimation from 30-second facial videos and HR estimation from 15-second facial videos comparing to the state-of-the-art methods. We also achieve promising results of computational speed and the number of parameters, demonstrating proposed network is a lightweight method.