請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/43554
標題: | 由使用者主導之人體動作與電腦動畫的互動 User-Guided Interaction between Human Motion and Computer Animation |
作者: | En-Wei Huang 黃恩暐 |
指導教授: | 傅立成 |
關鍵字: | 用表現驅動之動畫,動作粹取,控制映射,主動式學習, Performance-Driven Animation,Unsupervised Motion Extraction,Con- trol Mapping,Active Learning, |
出版年 : | 2011 |
學位: | 博士 |
摘要: | 人體動作與電腦動畫之間的互動在電玩遊戲、表演藝術、商業展示和博物館展覽等領域有越來越多的應用。由人體動作帶動的互動具有直覺、豐富的表現力以及促進融入感等優勢,但是為了一個互動應用的高使用度,在動作辨識及互動設計上有幾項挑戰必須被妥善處理,包括快速的反應、正確的辨識結果以及高品質的動畫。
我們提出了一個利用使用者的表現動作驅動的動畫系統以達到使用者和電腦動畫間的互動,在線上操作時,一個事先學習好的控制映射將人體姿勢對應到一組動畫系統的控制參數以驅動電腦動畫。在本論文中,我們利用locally weighted projection regression (LWPR)來建立控制映射模型, 有能力處理高維度的輸入資料, 因此非常適合高維度的人體動作,除此之外,LWPR模型能夠被遞增式地建立,因此我們能將LWPR模型的建立與我們提出的主動式學習機制整合在一起,藉由選取有用的訓練樣本逐步改進現有的控制映射。在系統離線時,我們提出了一個主動式學習的步驟,從一組連續的訓練序列挑選出有用的訓練樣本來改進現有的控制映射,一個單一的連續訓練序列包含有數個控制動作與它們之間的轉換動作,一個連續訓練序列非常類似使用者在線上操作時的實際動作,因此能幫助建立有效的控制映射。 我們利用一個以使用者動作驅動的角色動畫系統來展現我們提出的方法,實驗結果顯示我們的方法可以利用人體動作產生反應迅速與高品質的動畫。本論文的主要貢獻在提出一個有效且迅速建立利用人體動作達到細微與多樣化控制電腦動畫的方法,我們預期在電玩遊戲與表演藝術等領域可以產生新的影響。 Interaction between human motion and computer animation has more and more applications in video games, performance art, commercial demonstrations and museum exhibitions. The high degrees of freedom of human motions have some advantages in manipulating computer animations including intuition, expressiveness and improvement in the sense of immersion. However, some requirements including robust human motion recognition, quick response and high-quality animation need to be addressed for such a system. We propose a performance-driven animation system for interacting with computer animation by human motion. During online, a learned control mapping maps a human pose to the parameters of an animation model to drive the animation. The control mapping is modeled by locally weighted projection regression (LWPR) which uses a set of local linear models to approximate a non-linear function. It handles high-dimensional data well and can be learned incrementally. The first property is crucial for a performance-driven system because the representation of a human pose is high-dimensional. The second property helps us build an active learning approach upon it by selecting useful training samples to update an existed mapping. During offline, we propose an active learning approach which automatically selects samples and subsequences in a set of continuous training sequences to improve an existed mapping. A continuous training sequence containing multiple control motions and transition motions is similar to the online operation and helps create an effective mapping. The experiments of performance-driven character animation show that the proposed approach can generate responsive and high-quality animation by human motion. Our work will contribute to the applications such as performance art and video game by giving versatile and accurate interaction. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/43554 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-100-1.pdf 目前未授權公開取用 | 2.23 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。