Abstract
[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]


Selecting Representative Shapes for Interpolation to Describe Shapes of a Human Body in Motion


As the computer graphics technology is developed with the progress of computers, it has become easy to construct a three-dimensional world on a computer. Consequently, It becomes more necessary to represent various kinds of three-dimensional shapes, including those of human bodies, on computers. Representation of human bodies plays an important role not only in creating computer-synthesized animations in the screen, but also in realizing interface agents and avatars, which are employed for human-computer communication in a virtual space.

Since a human body has many joints, it takes various shapes in different postures. The human body model that is most widely used copes with this variation by its articulated structure constituted by rigid parts with flexible joints. But a real human body consists of non-rigid parts, and their shapes change lithely for different postures. Thus it is difficult to represent shape variations of a human body appropriately with the articulated human body model.

In this paper, a human body is represented as an elastic object. This method of modeling employs a representation based on instances which are acquired by observing the shapes of a real human body in various postures. To reduce the number of those instances in this approach, each shape that is not included in the set of instances is represented using interpolation between the instances similar to the shape, whereas the observed shapes which cannot be represented by the interpolation are added to the instances as ``representative shapes''.

In order to describe each instance, we use a patch model, which represents a surface of a three-dimensional object by many polygons connected with each other. Because of its high degree of freedom to represent different shapes, it can describe the shapes of various kinds of objects including a human body with a small amount of error, and fast graphics processing environments are available to display it with a computer. We prepare the initial patch model, which represents the general shape of a human body, and deform this model to each observed shape of a human body. Since correspondence between the patches of different instances is determined uniquely as the results of this deformation, interpolation between instances can be realized easily. In this method, silhouette images are given as observation of a human body. The shape of a human body is three-dimensional, but a special device such as lasor range censor is required to get three-dimensional data of a human body. On the other hand, silhouette images can be easily obtained by an ordinary camera.

Selecting representative instances to describe the shapes of a human body in motion requires a criterion for evaluating each instance. In our approach, representative instances are supplemented if interpolated shapes are not evaluated sufficiently similar to the corresponding observed shapes, comparing the interpolated shape and the observed result of a real human body. The criterion for this evaluation consists of two factors. One is difference from an observed human body shape, and the other is smoothness of the change of the described shapes in a serious of postures that constitute a single motion.

We applied this method to represent the shapes of human arm during its bending motion. As the result, human body shapes described by this method were more realistic than those described by the articulated body model.


Go back to Thesis Page