Constrained Independent Vector Extraction Of Quasi-Periodic Signals From Multiple Data Sets

Quasi-periodic signals are important signals that appear commonly in communication,radar,astronomy,economics and biomedical signal processing.In these fields,there is a strong demand to extract quasi-periodic components from instantaneously mixed signals in a single data set or multiple data sets.Existing blind source separation(BSS)methods can demix quasi-periodic source signals from observations in a single signal set.However,the source extraction of quasiperiodic signals from multiple correlated signal sets still requires further investigation.In this thesis,we introduce a constrained independent vector extraction(CIVE)to solve the problem of joint blind source extraction for quasi-periodic signals,and successfully apply the method to the task of video-based contactless heart rate extraction.The main work in this thesis is as follows:Firstly,the CIVE method is proposed to jointly extract the quasi-periodic signals from multisets of instantaneously mixed signals.The method transforms the original problem into a constrained optimization problem by taking the mutual information and the negentropy of the target source component vector(SCV)as the cost function and the autocorrelation function as the constraint.The principle is that the maximization of mutual information is used to ensure the dependence within the SCV,the maximization of negative entropy is used to ensure the independence of the SC V,and the autocorrelation function is used to determine the quasi-periodicity.The proposed CIVE can be used to only extract the target SCV,which does not rely on initial guesses of the demixing vectors,and is not limited by the probability distribution assumption.In the experiments,the comparison results with other methods indicate the effectiveness,stability and applicability of the proposed method for extracting quasi-periodic SCV.Secondly,considering the pulsatile signal is quasi-periodic,the CIVE algorithm is applied to extract heart rate from videos based on remote photoplethysmography(rPPG).The proposed method first divides the face in the captured video into several patch-shape regions of interest(ROIs),each of which is a set containing the observed signals.The optimal patch ROIs are then selected and the RGB signals are obtained in each sub-ROI,which have shared heart rate information.Finally,the pulsatile signals are extracted using the CIVE.The method is demonstrated to be accurate and effective by comparing it with other methods in the public COHFACE database.This thesis verifies the effectiveness of the proposed method for joint blind source extraction of quasi-periodic signals from both the theoretical and application levels.This method also provides an efficient and reliable new way for solving related problems in other fields.