以偵測特徵變形來建構逼真3D臉部動畫

Abstract

能做出表情變化的人臉模型在3Ｄ遊戲，電影，線上交談系統，網路虛擬表現和影像會議系統上極為重要。現在大部分商業用途上的方法是用3D scanner來掃瞄一個真人並做出人臉模型。但是致命的缺點是一般人幾乎不可能使用價格昂貴的3D Scanner。 在這篇論文中我們提出了一個用便宜的數位相機和電腦來做出人臉模型的方法。只用了兩張照片就可有效的生成逼真的的人臉表情動畫。這個模型模擬表情時候是用肌肉模型為基礎。臉部肌肉參數可由攝取到的image sequences來得到。 此外我們提出了一個臉部偵測的演算法。首先我們使用YCbCr來偵測出大概的人臉區域。接著利用人臉的對稱性和眼睛，嘴巴等的灰階值特性來找出我們要的特徵點。當表情變化的時候，根據特徵點的位置我們可以計算出變化量。最後根據變化量我們可以產生逼真的人臉動畫。 我們在WinsowsXP上驗證此系統的可行性。我們並調整人臉模型的多邊型數目和顏面筋的彈性係數來產生高品質的人臉動畫。最後我們得到合理的成果並且希望將來可將此技術用在手機上視訊會議系統。

Animated face models are essential to 3D games, movies, online chat, virtual presence and video conferencing. Nowadays, some commercially available tools make use of 3D laser scanners to acquire facial images. However, the drawbacks of 3D laser scanners are their costs and they are not widely used. In this thesis, we present a method using inexpensive computers and video cameras to produce face models directly from images acquired by cameras. This is an efficient approach to synthesize realistic facial expressions from only two facial images on a 3D facial muscle model. This model is capable of simulating facial dynamics through the muscle-based computation. The facial muscle parameters can be estimated from captured image sequences. Moreover, a face detection algorithm is proposed in this thesis. At first, YCbCr skin color model is used to detect the possible face area of the image. Second, we can obtain the feature points of the face by the symmetry of a face and the gray level characteristics of eyes and mouth. According to the positions of the feature points on the facial image, we can measure the quantity of transformation of the face when an expression appears. Finally, we can synthesize the realistic facial animations based on these. To prove its feasibility, we implemented the system on a Windows XP pc. We clarified conditions that could achieve high quality animations by optimizing the number of polygons that form the 3D face model and the stiffness values applied to the spring models embedded in the face model. Reasonable qualities for facial expression animations were obtained. We hope this method can be applied to video conference systems on mobile phones in the future.