Human motion characterization and its applications in motion analysis and synthesis

Abstract

Human motion analysis and synthesis are important but difficult problems in computer vision and computer graphics. As the primary task in these fields, human motion characterization is to extract information from human motion data and represent in appropriate models. However, the complete recovery of motion information is not always required. Statistical learning provides a tool for characterization of human motion. In this thesis we focus on statistical learning based human motion characterization methods for motion capture data in some applications of human motion analysis and synthesis. The first application is human motion classification. We treat the motion capture data as a whole and extract the discriminative features by a proposed global subspace analysis method named kernel clustering-based discriminant analysis (KCDA). KCDA works by first mapping the original motion data into another high-dimensional space, and then applying clustering-based discriminant analysis (CDA) in the transformed space to extract features for discrimination. KCDA can combine the merits of both kernel Fisher discriminant analysis (KFDA) and CDA to improve the classical Fisher discriminant analysis (FDA) in the way that one hand multiple cluster structure of the data is fully exploited and on the other hand kernel technique is imposed to get a nonlinear separation hyperplane. In particular, the framework of KCDA integrates a kernel fuzzy c-means algorithm (KFCM) to exploit the multiple cluster structure in motion data and a fuzzy cluster ensemble to improve the stability and convergence of KFCM. The second application is statistical modeling based human motion prediction. We propose a genetic optimization algorithm to learn a graphical model which characterizes the conditional independence between the body joints by representing the joints as graph nodes and the relationships between the joints as graph edges. The graphical model decomposes the high-dimensional joint probability distribution of all the body joints into a number of low-dimensional distributions over small sets of the joints. A subset of the body joints which provides predictions for the other joints is then identified from the graphical model. The associations between the body joints are learned through a set of multivariate relevance vector machines (RVM). The performance of graphical model based human motion characterization is demonstrated through experiments on motion prediction. The last application field is human motion synthesis. The difficulty in realistic human motion synthesis is due in part to the high dimensionality of the human motion. However, most dynamic human behaviors are intrinsically low-dimensional with, for example, legs and arms operating in a coordinated way. With this assumption, we propose a statistical model to represent the dynamic properties of human motion. A low-dimensional manifold is extracted through a nonlinear isometric feature mapping (Isomap), and a set of dynamic models and the transition relationships between them are learned from the manifold. Based on a nonlinear mapping function associating the manifold with the original motion space, we exploit the cyclic patterns of locomotion such as walking and running, learn dynamic models in segmented acyclic motion manifolds and synthesize new motion sequences.

Date of Award	2 Oct 2008
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Hau San WONG (Supervisor)