【作者】 俞俊；

【导师】 庄越挺；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2009， 博士

【摘要】 通过图像、视频、3D模型和运动捕获数据等可视媒体进行数字角色生成的研究已成为热点。数字角色由纹理、形状和行为三方面要素构成,并来源于图像、视频、3D模型和运动捕获数据等可视媒体。因此,对可视媒体的融合处理,充分发挥各种可视媒体的优势,提高数字角色生成的效率和质量,成为结合图形学、计算机视觉、图像/视频信号处理等领域的一个新兴研究方向。本文提出的可视媒体融合框架,将图像、视频、运动捕获数据和三维模型这四类能被人类视觉所感知,能通过相应的硬件设备制作获取,能被编辑、压缩/解压及存储的媒体定义为可视媒体,将计算机数字角色动画中出现的各类研究从可视媒体融合的角度系统分类。为了提高二维数字角色的制作效率及质量,引入可视媒体融合理论,提出了基于可视媒体融合的二维数字角色生成新思路。提出具有鲁棒性及高抗噪性的多风格二维数字角色提取算法。基于二维数字角色图像中的灰度信息,计算图像中每帧的梯度并检测出角色的轮廓,对轮廓采用像素填充算法获得侧影并提取。其中轮廓提取是逐一计算每个像素点在灰度上的二阶导数,并通过高斯.拉普拉斯算子(Laplacian of Gaussian,简称LOG)近似得到,当某一点的二阶导数为零时,则该点被定为边界的一部份。边界图像的噪音,采用卷积算法去除。提出基于低维流形空间的自适应二维数字角色融合算法。采用全局流形降维技术(ISOMap),依据距离矩阵确定当前结点与周围邻居结点之间的距离为最短路径,由此构建无向图;基于无向图计算任意两点之间的最短路径,得到最短路径图;采用多元尺度分析(Multidimensional scaling,简称MDS)算法构建低维空间。基于该低维内嵌空间,由用户在空间中任意指定两点作为起点和终点,基于两点之间最短路径所包含的卡通帧合成二维数字角色动画。提出基于扩散激发图模型(Constrained Spreading Activation Network,简称CSAN)的二维数字角色融合算法。分析不同特征(包括欧式边界距离,霍氏边界距离,运动距离,颜色直方图距离及分块子区域颜色距离)对二维数字角色动画平滑度的影响,引入模糊关联度计算特征权重;精确计算任意两帧之间的相似度并构建扩散激发图模型;基于该模型,依据用户的轨迹输入,生成二维数字角色动画。提出基于背景计算的交互式二维数字角色自动生成框架。基于二维数字角色库,构建扩散激发图模型;由用户在二维场景图像中任意绘制轨迹并自动生成新的二维数字角色动画;通过灭线计算得出二维数字角色在场景中的比例透视关系,将场景与动画相融合。该方法使用户直接面对场景绘制轨迹,并控制二维数字角色在场景中的运动,其灵活性及易用性适合用户的操作习惯。更多还原

【Abstract】 In recent years, digital roles have been widely used in various applications such as virtual reality, video games, animation films and sport simulations. The generation of digital roles, which utilizes several visual media, including image, video, 3D model and motion capture data, is a cross research field of visual media fusion. In the computer animation world, the role is composed of the texture, shape and behavior, which are generally from these visual media. Dealing with the visual media fusion is becoming an emergent research topic in the fields of computer graphics, computer vision and image/signal processing.This paper defines the image, video, motion capture data and 3D model to be the visual media, which can be perceived by the human perception and be able to be edited and stored. Based on the definition, we propose a visual media fusion framework, which can be used to classify varied kinds of computer animation from the visual media fusion’s point of view.This paper proposes a robust and anti-noise segmentation algorithm based on Laplacian of Gaussian edge detection to extract the data from the image precisely. In this algorithm, the standard deviationαis coordinated to extract the character’s edge, and a convolution process is carried out to remove the noise and extract the complete edge for the characters. We utilize the fulfilling algorithm to extract the whole character.This paper proposes an adaptive 2D digital role fusion algorithm in the lower-dimensional space. It adopts the Isomap as a lower-dimensional reduction method. According to the adjacency matrix, an undirected graph can be built. Based on this undirected graph, a shortest path graph can be constructed, and then by using the Multidimensional scaling, a lower-dimensional embedding space is built. Based on this space, the cartoonists can specify two points and construct the shortest path between them. Those points along the path can generate the new 2D digital role animation.This paper proposes a controllable 2D digital role fusion algorithm based on constrained spreading activation network. Initially, the features’ effects on similarity including Euclidean edge distance, Hausdorff edge distance, motion distance, Color histogram distance and Sub-region color distance are analyzed and a fuzzy approach is adopted to assign the weights for each feature. After calculating the similarity precisely, the constrained spreading activation network is constructed.This paper proposes an automatic 2D digital role generation framework based on the constrained spreading activation network. The new animation can be generated according to the path drawn in the background image. After calculating the vanishing line in the background, the new generated animation can be applied into the background according to the scaling factor. This framework is convenient and flexible for the users.更多还原