利用立體視覺的技術作基於知識的人臉表情追蹤及模型建構

Abstract

本論文提出了一個利用基於知識追蹤及建構表情變化模型的立體視覺系統。在本研究中，我們利用影像處理的技術來抽取安置在人臉部特徵點的位置和計算其相對的運動參數〈含頭部整體的剛體運動及特徵點的自體運動〉，並且利用電腦圖學的技術來重現人臉部表情的立體面貌。本系統使用了三支攝影機來建立一套立體電腦視覺系統，藉由特徵點在攝影機間的對映關係，來計算特徵點在三度空間的座標，並且利用現在和先前所計算出特徵點在三度空間的座標，來求得頭部整體的剛體運動及特徵點的自體運動的速度、加速度，進而用以預測特徵點下一刻在影像中的位置。我們還利用預測的結果來加速搜尋在下一時刻影像上相對應的特徵點。最後，我們利用以一個肌肉為主的臉部模型來重現臉部表情變化。實驗結果證明了以上的方法確實是可行及有效。

In this study, a stereo vision system for knowledge-based tracking and modeling of facial expressions is proposed. Image processing techniques are used to extract feature points on a human face and calculate related facial parameters, including the head pose caused by rigid motion and the displacements of feature points caused by local motion. And computer graphics techniques are used to re-animate the facial expressions. Three cameras are used in the system, and feature-point correspondences are employed to calculate the 3-D coordinates of the feature points. The velocity and acceleration of the head pose caused by rigid motion and the displacements of feature points caused by local motion are calculated from the 3-D coordinates of the feature points in the current and the previous image frames. The locations of feature points in the next frame are predicted by the use of the motion parameters, with help from the knowledge of the geometric properties of the feature points on the face. The knowledge-based prediction result is used to speed up search of the corresponding feature points in the next image frame. Finally, a muscle-based face model is used to re-animate the facial expressions. Experimental results show the feasibility and practicability of the proposed approaches.