Research On Data-driven Virtual Human Motion Analysis And Synthesis

The human motion capture data can record the detailed information of human motion and can be widely used in computer animation,virtual reality,etc.With the further growth of human motion capture data,how to realize motion analysis and virtual motion synthesis for human motion capture data has become a hot research topic.Based on the key technologies of data-dirven virtual motion analysis and motion synthesis,this paper focuses on three aspects:human motion behavioral segmentation?motion behavioral templates extraction and annotation and virtual human motion synthesis to be in-depth research.The primary investigation contents of this paper includes the following aspects:In the aspect of data-driven motion behavioral segmentation,this paper proposed two behavioral segmentation methods.(1)Behavioral segmentation based on graph cut:To construct undirected and weighted graph for human motion capture data and the problem of behavioral segmentation is transformed into graph cut problem;however,traditional graph cut problem is NP hard,by analyzing the relationship between graph cut and spectral clustering,applying t-nearest and Nystrom method to realize behavioral segmentation;furthermore,the energy function based on the dynamic programming principle is constructed to optimize behavioral segmentation result.(2)Behavioral seg-mentation based on split EM(Expectation Maximization)to estimate GMM(Gaussian Mixture Model):assuming human motion capture data for the same motion behavior conforming to a Gaussian model and modeling;applying split EM algorithm to estimate the parameters of GMM and realizing behavioral segmentation;the energy function based on maximum similarity is constructed to improve the accuracy of behavioral seg-mentaion.In the aspect of data-driven motion behavioral templates extraction and annotation,hierarchical templates extraction and annotation based on the hierarchical template are proposed.The human motion capture data,which is belonging to single motion behavior,is modeled by Hidden Markov Model.The parent hierarchical template based on motion type and the child hierarchical template based on motion style are constructed respectively.The DTW(Dynamic Timing Warping)method is used to mark the parent hierarchical template of motion type and the CTW(Canonical Time Warping)method is used to mark the child hierarchical template of motion style.In the aspect of data-driven virtual human motion synthesis,this paper proposed a method of virtual human motion synthesis based on behavioral annotation.Constructing undirected and weighted motion graph for the behavioral annotation human motion capture data.Floyld algorithm is used to calculate shortest path between any nodes in this motion graph.Finally,using the cross interpolation method to interpolate motion segments on the shortest path and realizing motion synthesis.