基于单目视觉的实时人脸动画生成方法研究

项目名称： 基于单目视觉的实时人脸动画生成方法研究

项目编号： No.61302127

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 李月龙

作者单位： 天津工业大学

项目金额： 24万元

中文摘要： 以动画形式实时表现人脸的表情和动作变化，是颇具理论和实际意义的科研课题，在影片和游戏制作、虚拟会议场景构建等方面有广泛的应用价值。针对当前技术存在依赖专业视频采集设备、用户友好性差等弱点的问题，本项目拟研究设计一种基于普通单目视频采集设备的实时人脸动画生成方法。具体包括两部分研究内容：实时提取跟踪人脸轮廓，捕获视频中人脸的表情和动作；建立人脸轮廓关键点坐标向面部动画模型参数的映射，使所获取的面部表情和动作可以以动画形式实时表现出来。轮廓提取时，本项目将依据子空间约束和邻域内高斯分布构建人脸轮廓描述模型，并引入三维轮廓模型辅助处理人脸多姿态问题；在建立映射关系时，本项目将首先以稀疏表示系数投票的方式，提取动画模型中的关键参数，降低参数间的依赖性，再基于回归分析建立映射。本研究有助于人脸动画制作成本的降低和用户友好性的提升，同时可以有效促进人脸轮廓提取一般性方法和稀疏表示应用理论的发展。

中文关键词： 人脸动画；面部动作描述；关键点提取；信息融合；动画模型控制

英文摘要： The realtime transformation of human face expression and action into that of animation model is an interesting research topic both in theory and application, which could be widely utilized in movie and game production, virtual conference scene construction, and so on. Since most of current techniques in this field require specific video capture hardware or are not user-friendly to some extent, an ordinary single camera based realtime face animation creation methodology will be researched and designed in this project. This research is composed of two main parts: the acquisition of face expression and action through realtime face shape extraction and tracking; the construction of a mapping method that transforms face shape landmark coordinates to animation model parameters, which makes the realtime animation demonstration of acquired face expression and action possible. During shape extraction, a subspace constraint and Gaussian distribution in landmark neighborhood based face shape model is proposed and utilized, and a 3D shape model is introduced to assist in dealing with multiple face pose variations. During the construction of mapping relationship, a voting strategy of sparse representation coefficients is worked out to select most critical animation model parameters to reduce parameter dependency, and then t

英文关键词： facial animation；facial movements description；key point extraction；information fusion；animation model control