【摘要】：基于表演驅(qū)動的人臉動畫生成技術(shù)是一項根據(jù)表演者面部表演驅(qū)動虛擬人物做出相同頭部姿態(tài)和面部表情的技術(shù),可廣泛地應(yīng)用于影視制作、人機交互、游戲制作、遠程會議、醫(yī)療輔助等領(lǐng)域,近年來一直是計算機視覺和計算機圖形學(xué)的研究熱點。目前,一些成熟的系統(tǒng)都有比較苛刻的使用條件,有的需要復(fù)雜昂貴的采集設(shè)備作為支撐,有的需要高性能計算機完成高額運算,難以推廣到普通消費者中使用。本文旨在搭建一個實時、精確、消費級的表演驅(qū)動人臉動畫生成系統(tǒng),利用普通單目攝像頭捕捉人臉表情,并將捕捉到的信息重定向到動畫角色中。圍繞系統(tǒng)的開發(fā)中的兩大模塊——人臉表情捕捉和人臉動畫合成,本文主要工作內(nèi)容和創(chuàng)新性成果如下:(1)對基于優(yōu)化方法的快速三維人臉建模算法提出了改進。為了去除用于合成的人臉模型張量中的冗余信息,提出用高階奇異值分解(HOSVD)對張量進行壓縮。針對手動標定誤差影響殘差物理含義的問題,根據(jù)不同人臉部位的標定特點,在能量方程中建立不同的殘差模塊。實驗證明本文提出的改進有效提高了算法的精確度。(2)提出了基于隨機森林特征提取的三維表情捕捉算法,直接從二維圖像中得到三維人臉形狀,解決了現(xiàn)有系統(tǒng)中捕捉設(shè)備復(fù)雜或計算量大的問題。該算法中利用隨機森林提取聯(lián)合局部二值特征并用于線性回歸運算,提高了系統(tǒng)的精度。采用三角重心坐標表達的形狀索引NPD(Normalized Pixel Difference)特征作為森林中決策樹的分裂依據(jù),提高了對不同人臉姿態(tài)的魯棒性。(3)針對不同驅(qū)動目標的人臉三維結(jié)構(gòu),提出了基于三角網(wǎng)格形變傳遞的動畫驅(qū)動生成算法和基于Blendshape的動畫驅(qū)動生成算法,提高了動畫驅(qū)動的效率和真實度。并提出了增量權(quán)重法增強動畫的穩(wěn)定性,運用基于拉格朗日插值法提高動畫幀率,有效改善了現(xiàn)有基于表演驅(qū)動的人臉動畫生成系統(tǒng)中常見的抖動和幀率不足的問題。
[Abstract]:Performance-driven face animation generation technology is a technology that drives virtual characters to make the same head posture and facial expression according to performers' facial performance. It can be widely used in film and television production, human-computer interaction, game production, remote conference. In recent years, medical assistance has been a hot research topic in computer vision and computer graphics. At present, some mature systems have harsh use conditions, some need complex and expensive acquisition equipment as support, some need high performance computer to complete high operation, so it is difficult to be popularized to ordinary consumers. The purpose of this paper is to build a real-time, accurate and consumer performance-driven face animation generation system, use ordinary monocular camera to capture facial expressions, and redirect the captured information to animated characters. Around the two modules of the development of the system, facial expression capture and face animation synthesis, the main contents and innovative results of this paper are as follows: (1) an improved fast 3D face modeling algorithm based on optimization method is proposed. In order to remove redundant information from Zhang Liang, a face model used for synthesis, Zhang Liang is compressed by high order odd value decomposition (HOSVD). In order to solve the problem that manual calibration error affects the physical meaning of residual error, according to the calibration characteristics of different face parts, different residual modules are established in the energy equation. Experiments show that the improved algorithm effectively improves the accuracy of the algorithm. (2) A 3D expression capture algorithm based on random forest feature extraction is proposed to obtain 3D face shape directly from two-dimensional image. The problem of complex capture equipment or large amount of computation in the existing system is solved. In this algorithm, random forest extraction combined with local binary features is used and applied to linear regression operation, which improves the accuracy of the system. The shape index NPD (Normalized Pixel Difference) feature represented by triangular barycenter coordinates is used as the split basis of decision tree in forest, which improves the robustness to different face postures. (3) 3D face structure for different driving targets. The animation drive generation algorithm based on triangular grid deformation transfer and the animation drive generation algorithm based on Blendshape are proposed, which improves the efficiency and reality of animation drive. The incremental weight method is proposed to enhance the stability of animation, and the Lagrangian interpolation method is used to improve the frame rate of animation, which effectively improves the common jitter and frame rate shortage in the existing performance-driven face animation generation system.
【學(xué)位授予單位】：浙江大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41
，

本文編號：2475537