Compact Feature Representation and Efficient Retrieval for Human Motion Capture Data

• Main Authors: Ming-ren Su, 蘇明仁
• Other Authors: Wei-Chang Du
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/00053463282432205860

碩士 === 義守大學 === 資訊工程學系碩士班 === 97 === With the thriving of computer animation industry, fast development schedule and realistic visual effects have long been important issues which animation-making companies focus on and emphasize. Nowadays, since motion capture technique has also become an important tool in digital character animation, it is expected that more and more mocap data will be produced in the future. Within a huge and messed mocap database, it is natural to become an important research issue to develop a useful and reliable retrieval system. The main goal of this research is to develop an efficient retrieval system in which the relevant feedback with respect to user query can meet with human semantic understanding. Recently, some geometric features based on joint coordinates or angles are used to extract mocap features, and PCA is often used in reducing the dimension of mocap data. However, most of these features are quite sensitive to some transformations of motion capture data. Based on the so-called orientated relations as a static posture feature, we propose histogram based feature extraction for each clip, automatic mocap segmentation and high-dimensional indexing technique, to build the retrieval system. Experimental results are shown to demonstrate effectiveness of proposed retrieval method.