| Project/Area Number |
05650407
|
|---|
| Research Category |
Grant-in-Aid for General Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
計測・制御工学
|
|---|
| Research Institution | KYUSHU UNIVERSITY |
|---|
Principal Investigator |
ZHA Hongbin Kyushu Univ., Dept.of CSCE Associate Professor., 工学部, 助教授 (80225680)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
OKADA Nobuhiro Kyushu Univ., Dept.of CSCE research asso, 工学部, 助手 (80224020)
NAGATA Tadashi Kyushu Univ., Dept.of CSCE Professor, 工学部, 教授 (20136542)
|
|---|
| Project Period (FY) |
1993 – 1994
|
| Project Status |
Completed (Fiscal Year 1994)
|
| Budget Amount *help |
¥2,000,000 (Direct Cost: ¥2,000,000)
Fiscal Year 1994: ¥1,200,000 (Direct Cost: ¥1,200,000)
Fiscal Year 1993: ¥800,000 (Direct Cost: ¥800,000)
|
|---|
| Keywords | Robot Vision Systems / Model-based Recognition / Automatic Model Generation / Cooperative Learning networks / Surface Curvature Consistency / Superquadric Surface fitting / 超二次曲面 |
|---|
| Research Abstract |
In the planning phase, we classified the problems relevant to the project into three main subjects
1) investigation into the model-based structures ;
2) extraction of model features and building-up of the model-base ;
3) hypothesis generation and verification using the model-base.
During the two-year period, we had gone on with the research on the three related subjects and and the results are summarized as follows.
1) Investigation into the model-base structures :
Among several well-used types of model discription, we found that the hierarchical patch-based object description is the most efficient for representing 3-D complex curved objects.
Different recognition algorithms using such a description are developed and implemented in real experiments to give support to the conclusion.
2) Extraction of model features and building-up of the model-based :
At first, we proposed surface recovery methods based on surface curvature consistency, image fitting with superquadrics, or image fitting withB=spline. A recovery algorithm for complete object reconstraction was also developed by using multiple range images. Moreever, two learning algorithm for automatically generating model-bases are designed by utilizing the backpropagation method and GRBF approximation networks, respectively.
Hypothesis generation and verification using the model-base :
We succeeded in developing an efficient model-to-image matching algorithm which uses a fast Hopfiled-style network to search reasonable interpretation of a complex scene. The algorithm takes advantage of the patch-based model description proposed by us and its applicability in real situation was shown by experiments using real range images. To deal with more complicated scenes, we also proposed a multi-vision recognition method which employ multiple cameras set at different viewpoints to overcome the difficult occlusion problem.
|
