Abstract
The acquisition of human motion data is essential for creating interactive virtual environments, intelligent user interfaces, and computer animations. Hierarchical skeleton models are widely used to represent virtual human whose body is composed of chains of bones with interconnecting joints. In recent studies, many researchers try to estimate the parameters of the skeleton with commercial equipments. In this study, we present a markerless motion capture approach for articulated kinematic structure. This approach utilizes the manifold learning technology, ISOMAP, to deal with the situation when the occlusion of limbs and body occurs. It is based on the assumption that the nonlinearity of rigid-body kinematic motion is introduced by rotations about the joint axes. By applying Gaussian Mixture Model on the result of ISOMAP, we can easily segment each part of body, such as head, hands, legs and torso. Then we can find the end points of limbs and head. With all the information above, we fit a pre-defined human body skeleton onto the constructed volume by finding the maximum overlap between the volume and the sample points which we generated as muscle. Several experimental examples are given to demonstrate the effectiveness of proposed approach.