Article Preview
Top1. Introduction
In the process of sport and training, a proper exercise physiological load is conducive to improve the competitive level of athletes. However, excessive exercise physiological load may cause athletes to be injured or fatigue. When the athlete’s body is engaged in a specific sport or training, the exercise physiological load refers to the amount of changes in the internal functional state or physiological parameters. Thus, the exercise physiological load can be measured by the physiological parameters. The previous methods to monitor physiological parameters are contact by using wearable devices or electrodes on body (Miramontes 2017). It is inconvenient for athletes during sport and training.
Image photoplethysmography (iPPG) (McDuff 2017) is a non-contact physiological parameter detection technology. The principle of iPPG is that the color of the skin changes slightly due to the perfusion of superficial subcutaneous blood vessels. It can obtain the blood volume pulse (BVP) (Gao 2020) from the video of surface part of the human body. The heart rate (HR) and respiratory rate (RR) can be estimated by the time interval between two consecutive peaks (or valleys) of the blood volume pulse signal. Other physiological parameters, such as human blood oxygen, blood pressure, heart rate variability, and vascular microcirculation can be further analyzed to monitor the human health status. However, the human pulse wave signal is very weak. The blood volume pulse from iPPG signal is susceptible to the interference which is caused by low frequency baseline drift and high frequency noise. The low frequency baseline drift are generated by motion noise, while the high frequency noise is produced by light changes. The main part of human blood volume pulse signal lies in the part between low frequency and high frequency.
Lei et al. (Ghodratigohar 2019) used Empirical Mode Decomposition (EMD) to decompose the BVP signal into a series of Intrinsic Mode Functions (IMF) (Li 2017) with different frequency components, and finally obtained the heart rate through frequency domain analysis. The EMD discards the IMF component with high frequency signal to remove the noises in original signal. The high frequent does not contain effective information. The denoising can be implemented by reconstructing with the remaining IMF component. However, the remaining IMF component may still contain some noises which cannot be removed by the EMD, while the useful information in high frequent component is discarded. The Variational Mode Decomposition (VMD) (Lian 2018) is another signal processing method in which the number of modal components 
can be set the optimal one according to the mean of the instantaneous frequency of the modal components. The VMD can adaptively decompose the effective components according to each center frequency which has better accuracy and stability for feature extraction in low frequent bands. The VMD can overcome the weakness in EMD, empirical wavelet transform, such as modal mixing, difficultly eliminating additional noise and determining the number of eigenmode functions. Compared with previous methods, the VMD are more robust to the noises. However, the VMD cannot still remove all noises in IMF components.