1999 Fiscal Year Final Research Report Summary
Aspect Representation of Environment
| Project/Area Number |
09680368
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Wakayama University |
|---|
Principal Investigator |
TSUJI Saburo Wakayama Univ., Faculty of Systems Engineering, Professor, システム工学部, 教授 (60029527)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
KATO Koji Wakayama Univ., Faculty of Systems Engineering, Research Associate, システム工学部, 助手 (30273874)
|
|---|
| Project Period (FY) |
1997 – 1999
|
| Keywords | computer vision / aspect / object recognition / environment model / artificial intelligence / robot vision |
|---|
| Research Abstract |
Objectives
Aspect graph has been used for representing appearances of a 3D object. This research explores to generalize "aspect" to represent appearances of a wide environment in which an intelligent agent moves and views.
1.Aspect Changes with Distance from Camera
Most of theoretical studies on the aspect graph assume orthographic projection and ideal line images of an object. However, the significant variations in distances from a viewer to visible features in the environment make the assumption of the orthographic projection. As a results, translation of the viewer causes visual events. A feature extractor finds two separated features as connected id the distance between their projections in the image is within the sensor resolution. Image quantization causes significant changes in the detected features as the camera distance varies. Moreover, the detection of the visual features become unstable while observing them from certain ranges. Representation of aspects of the environment shou
… More
ld accommodate to such difficulties.
2.Representing Aspects of Wide Environment
Since the environment covers the whole space around the viewer, we propose its aspects are represented by visual information acquired from all directions by an omnidirectional sensor, rather than by visual data in a narrow sight. Thus, the visual space is represented by a 2D floor surface which is partitioned into visual cells by the environment aspect. We can analyze aspects of appearances of simple environments covered with a small number of planar surfaces. However, the real environment is very complex and we cannot manage the huge number of its aspects, even if we could determine them. Thus, we propose a generalize concept of aspect ; partition the view space with almost similar iconic features in the omnidirectional view. An example of such iconic features is low frequency distribution of Fourier transformation of the omnidirectional image. An alternative approach is to analyze scene structures and represent the environment as their arrangements. Experiments is a real scene shows the effectiveness and robustness of the method. Less
|
