真实感三维人脸建模与动画技术研究
文献类型：学位论文
| 作者 | 丁宾 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2009-06-02 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 王阳生 |
| 关键词 | 个性化人脸建模与动画 形变模型 形状无关纹理 人脸动画重映射 Individual face Modeling and Animation Morphable Models Shape Free Texture Facial Animation Remapping |
| 其他题名 | Research on Realistic 3D Face Modeling and Animation |
| 学位专业 | 计算机应用技术 |
| 中文摘要 | 人脸建模与动画是计算机图形学、计算机视觉等领域中的研究热点，在三维游戏、影视制作、人机交互界面、远程呈现等方面有着重要的研究意义和应用价值。本文围绕基于图像的真实感人脸建模与动画展开研究，本文的主要工作和贡献包括： (1) 提出了一种基于单张正面人脸图像的三维人脸重建方法。首先自动检测人脸在图像中的位置，并自动定位人脸上的关键点；手动标定形变模型上对应的三维关键点(离线完成)，利用弱透视投影建立三维关键点与图像上的关键点之间的联系，通过一个最优化过程得到人脸模型的形状参数和姿态参数；通过对形状无关纹理的拟合，得到完整的人脸纹理。该方法可以快速、鲁棒地生成具有真实感的个性化人脸模型。 (2) 提出了一种基于多张非标定图像的人脸重建方法。输入多个角度的人脸图像，在每幅图像上手动标定少量关键点，并设定人脸的初始姿态，基于形变模型重建出人脸的几何形状；获取每个角度的形状无关纹理，设定合适的融合权值，将多张形状无关纹理融合为单张人脸纹理。该方法只需很少的手工交互；在纹理融合中，形状无关纹理的引入以及融合方式的改进，使得融合出的纹理图像更加平滑；对独立于视角的纹理映射的改进，可有效消除图像间的光照差异引起的接缝现象和配准误差引起的颜色错误。 (3) 提出了一种基于稀疏关键点的人脸动画重映射方法。针对Motion Capture采集的稀疏人脸动画数据，只需在目标模型上标定与Marker对应的关键点，就可以将稀疏动画数据映射到稠密人脸网格上，得到具有真实感的人脸动画。首先利用局部仿射变换对动画数据进行调整，使之与目标模型的形状和大小相匹配；然后利用凸组合方法将目标模型进行平面参数化，在平面域对关键点的运动向量进行插值，得到非关键点的运动向量。该方法可以实现稀疏动画数据的快速映射，完全满足实时动画的需要。 本文在基于图像的人脸建模与动画方面做了一些有益的尝试和探索，并取得了一些初步的成果，希望本文的结果为基于图像的人脸建模与动画的研究提供借鉴和启发。 |
| 英文摘要 | Face modeling and animation are challenging research topics in computer graphics and computer vision. Face modeling and animation techniques are valuable for both research and applications such as 3D games, movie production, human-computer interface, tele-presence, and so on. In this thesis, we focus on face modeling and animation based on images. Our contributions include: 1) We present a 3D face reconstruction method based on one frontal face image. First, the face is detected in the image and the key points on the face are localized automatically. The corresponding key points on the 3D face model are labeled manually(off-line). Weak perspective projection is used to relate the 3D key points and the key points in the image. The shape and pose parameters can be achieved with an optimization procedure. A complete face texture image can be generated by fitting the shape free texture. Realistic face models can be generated quickly and robustly with this method. 2) We present a face modeling method based on multiple uncalibrated images. A small amount of key points are labeled in the input images, and the pose of the faces are initialized manually. Then the face geometry is reconstructed based on morphable models. Shape free textures are generated from the input images. The multiple shape free textures are combined to generate one single texture image. This method needs only a small amount of manual interaction. The introduction of shape free texture and the improvements on fusion method make the fused texture image smoother. The improvements on view independent texture mapping can eliminate the seams induced by the illumination differences between the input images, and the color error induced by the registration errors. 3) We present a sparse facial animation remapping method. In order to map sparse facial animation to dense meshes, we need only to label the key points on the target model corresponding to the markers. Then we can get realistic facial animation. First, the animation data is adapted with local affine transformation in order to make the animation data consistent with the shape and size of the target model. Then the target model is parameterized with convex combination method. The motion vectors of the non-key points are achieved by interpolating those of the key points. The interpolation coefficients are computed in planar domain. Our method can achieve fast mapping of sparse facial animation data and meet the need for real-time applicatio... |
| 语种 | 中文 |
| 其他标识符 | 200518014628084 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6213]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 丁宾. 真实感三维人脸建模与动画技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2009. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。