| Project/Area Number |
16500616
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Educational technology
|
|---|
| Research Institution | Kinki University |
|---|
Principal Investigator |
TANAKA Kazumoto Kinki University, School of Engineering, Lecturer, 工学部, 講師 (60351657)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
KUROSE Yoshinobu Kinki University, School of Engineering, Professor, 工学部, 教授 (00043802)
|
|---|
| Project Period (FY) |
2004 – 2005
|
| Project Status |
Completed (Fiscal Year 2005)
|
| Budget Amount *help |
¥1,700,000 (Direct Cost: ¥1,700,000)
Fiscal Year 2005: ¥200,000 (Direct Cost: ¥200,000)
Fiscal Year 2004: ¥1,500,000 (Direct Cost: ¥1,500,000)
|
|---|
| Keywords | Sign Language / Key Feature / Feature Extraction / Image Distortion / Self Information / Path-Searching Method / Human Memory Processes / Memory Strategy / 歪み補正 / 形状関数 / 短期記憶 / カラートラッキング / オプティカルフロー / 固有空間法 / フーリエ記述子 / 調動 / 音韻 / 文字列照合アルゴリズム |
|---|
| Research Abstract |
1.A Method for Extraction and Modeling of Key Features in Sign Language Words
(1)Feature Extraction from Sign Language Video
We have developed a method for analyzing hand motions and finger shapes by color-tracking of colored fingerstalls, thumbstalls and wrist bands that are attached to a sign language speaker. The method extracts principal component flow vectors as key features using eigenspace method. Next, our method cancels harmonic contents of flow vectors by using fourier descriptor. In addition, the method can correct image distortion occurred by filming with wide-range lens.
(2)Feature Modeling and Key Feature Extraction Based on Self Information
We expressed sign features by motion phonemes and finger shape phonemes. Next, we have proposed a method that extracts key features based on self information of each feature within sign feature database.
2.Feature Matching Method
We classify motion errors of sign language learners into two levels in terms of the error seriousness, and propose an feature matching method for each level by using path-searching. Our method employs extended standard motion features that include error features deformed slightly. Owing to the employment, our detection system extracts motion errors and determines their level simultaneously for realizing real-time error feedback.
3.Leaning Control Method
We propose a memory strategy by phased learning of key features in terms of feature importance and human memory processes. By an experiment of learning sign words, we show that our method is effective in long-term memorizing of the words.
|
