請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/53935
標題: | 基於連續蒙特卡羅法之近身搏擊動作生成 Synthesizing Close Combat Using Sequential Monte Carlo |
作者: | I Chiang 蔣易 |
指導教授: | 歐陽明 |
關鍵字: | 動畫,動作生成,動作規劃,連續蒙特卡羅法,粒子濾波器,最佳化, animation,motion synthesis,motion planning,sequential Monte Carlo,particle filter,optimization, |
出版年 : | 2015 |
學位: | 碩士 |
摘要: | 本論文提出一基於連續蒙特卡羅法之近身搏擊動作生成,我們是使用物理模擬的方法並不依靠動態捕捉資料,人物的行為可以被分類為攻擊和防禦並且可以被寫成一個目標函數。在攻擊模式中,人物致力於打擊對手身體的重要部位,目標的身體部位是根據詠春拳法的中線理論而定義,另一方面,防禦模式的主要目的是要阻擋對手的出拳以避免造成自身重要部位的傷害。我們運用連續蒙特卡羅取樣的方法去找到最佳的控制策略,欲求解的最佳化變數是各個關節角度的運動軌跡,在我們系統中的每個樣本含有一個座標和一個目標函數值。我們方法起初會均勻地在參數空間中產生一組樣本點,接著我們進行修整只保留下一些分數較高的樣本點並由他們建立一棵k-d樹。隨後,我們使用適應性重要性抽樣從舊的樣本點中取出新的樣本點,由於每個樣本點都有一個高斯分布以之為中心,k-d樹可被視為一個高斯分布的總和,新的樣本是由隨機產生的變數根據逆高斯分布得到該座標。我們將新的樣本點代入物理引擎得到每個剛體的位置、速度和碰撞等資訊,最後依照這些物理模擬的數據,我們可以進行目標函數評估來決定新樣本點的分數,新樣本會被加入k-d樹中直到達到一定的數量後再重複上述的步驟,藉由演化後我們的結果可以呈現一些簡單的攻擊、防禦和互動等動作。 This thesis presents a method to synthesize close combat using sequential Monte Carlo. We perform physics-based simulation without motion capture data. The behavior of the character can be classified into attack and defense and is formulated as an objective function. In the attack mode, character aims to hit the critical body regions of the opponent. On the other hand, the main goal of defense is to block the opponent’s blow in order to prevent injuries to critical regions. These principles are designed according to fundamental theory of Chinese martial arts. We use a kd-tree sequential Monte Carlo sampler to find the optimal joint angle trajectories for each character. Each sample in our system contains a coordinate and an objective function value. For each iteration, pruning is performed to keep a few samples with the highest objective function values. Then, a kd-tree is constructed based on those remaining samples. Next, adaptive important sampling is applied to draw new samples from the old ones. Subsequently, we will feed the new generated sample into physics engine to get positions, velocities and contacts of rigid bodies. Finally, by using the information from the physics simulation, we are able to make evaluation on the objective function to determine the scores of the new samples. These samples are added to the kd-tree until a budget is reached, then we will repeat the above steps. Through the evolution, our results can show some simple attack, defense and interaction movements. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/53935 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-104-1.pdf 目前未授權公開取用 | 1.08 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。