Techniques For Human Motion Analysis Based On 3D Capture Data

With the popularity of Motion Capture Systems since 1990s, Computer Animation has made great progress, and has been widely used in film, game, education and military. Problems, however, occur concerning how to produce animation efficiently and effectively using motion capture data. Under this background, motion capture data based animation has been a hot topic in Computer Graphics recently. The work of this thesis is to explore new approaches to human motion analysis based on motion capture database, and presents the following algorithms, including extraction of motion features, dimensionality reduction of motion features, motion database index, motion recognition and retrieval based on machine learning, motion edit and synthesis under manifold subspace. The thesis is organized as follows:1 Extraction of motion features, chapter 3 first proposes 2D geometric features to describe the local joint geometric relationship of motion data, and then temporal-spatial features are defined in this chapter, which describes 3D space relationship of each joint. 2D geometric features has far-reaching consequences regarding efficiency, flexibility and automation in view of indexing, content-base retrieval and time alignment of motion capture data. For 3D features, each joint's features can represent a part of the whole motion independently. So we can process₃D features of each joint separately.2 Dimensionality reduction and motion segment, since the dimension of motion feature extracted is very high, chapter 4 use PCA(principle component analysis) and a generalization of ISOMAP to do this. Based subspace generated by PCA and ISOMAP, two novel method are proposed to get distinct movements from continuous long Mocap sequence efficiency, which is convenient for manipulating in animation authoring system and retrieval based on keyword or content.3 Index for large-scale Mocap database and machine learning, in chapter 5, two novel index approaches are presented for motion retrieval that reduces the number of costly distance computations for similar measure. In chapter 6, multiple-instance data-driven decision tree is automatically constructed to reflect the influence of each point during the comparison of motion similarity and a novel approach is presented for motion recognition and retrieval based on Ensemble HMM learning.4 Motion synthesis and Editing in subspaces, in chapter 7, linear PCA method and nonlinear ISOMAP are used to map original styled human motions into subspaces, which can reduce computational complexity while reserving the intrinsic properties of original data. Motion style is edited based on quantitative analysis of motion style in PCA linear subspace. For nonlinear of human motion, this chapter also learn a decomposable generative model that explicitly decomposes the content parameters and style parameters of motion's manifold in nonlinear subspace, which can edit motion style in nonlinear subspace.5 Implementation of the algorithms. Chapter 8 describes how the methodologies and the algorithms construct a system that runs in the large-scale Mocap database.