![](https://csdnimg.cn/release/download_crawler_static/19900819/bg3.jpg)
tion. Notably, it is not easy to realize a search for an object
from its meaning or to perform a summarized operation.
Several designs of spatiotemporal databases have
been proposed, including the Minimum Bounding Box
(MBB), which represents an object by using circumscribed
squares [4, 8]; data maneuvering based on MBB [9]; repre-
sentation of spatial relationship of an object by using icon
mapping [10]; and a realization of Ref. 9 by using a spa-
tiotemporal theory [11]. In limited applications (modeling
or data maneuvering), the MBB can handle data at high
speed, since Allens temporal intervals theory [12] with an
expansion can be applied to the spatial region. However,
this method cannot model fine parts of an object. To com-
pensate this, a large amount of additional computation is
required. Also, a space approximated by using this method
does not have a realistic representation.
Therefore, it is necessary to construct a new spatial-
temporal database with efficient storing of scenes, good
visual representation of scenes, an efficient data searching
capability, and an economical operating time.
3. Construction of Database
3.1. Method of modeling
A simplified model based on the mining of an object
has been introduced in the proposed method, as shown in
Fig. 2, so that a mediator which corresponds to the sample
data and their meaning is formed by semiautomated fitting
of sample data. The mediator accurately represents the
shape of each object in a scene with a small number of
polygons. The scene database is constructed by applying a
virtual CG space mapping [13].
The real input data have various kinds and levels. This
project aims to store their shapes and time variations. When
they involve human behavior [14], their movements might
be affected by their environments and apparatus around
them. Therefore, 3D images (taken by video cameras),
external shapes of objects (taken by a 3D scanner), and their
texture data alone were used for sample data from the scene.
These sample data contain no meaning data such as the
name of a part of the human body. Data about an object can
be described only when a human adds a meaning to it.
However, there are too many data (represented by pixels or
polygons) to add meanings manually. To add such mean-
ings semiautomatically and to store and search them, the
authors propose the following idea:
Shape of the fundamental model. Shape data which
can represent the general shape of an object and its charac-
teristic parts by means of a few polygon data are called the
shape of the fundamental model. A human head with
eyes, nose, mouth, and ears, or a hammer with a metal block
and a wooden handle are examples. A 12-vertex prism with
nine layers is used for the fundamental model of a human
head in this paper, with additional data of the second cross
section (from the top) representing the eye level, and the
second lowest cross section representing the level of the
lower jaw. Referring to the cylinder in Fig. 2, it is possible
to add vertices representing positions of two eyes and
mouth. Specific data for each component can be unified by
adding their common data, and by using the meaning of the
fundamental model shape.
Fig. 2. Scene data acquisition.
64