【作者】 李石峰；

【导师】 卢湖川；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2012， 博士

【摘要】 静态图像的人体分割是指从含有人的图像中将人体轮廓分割出来,它在计算机视觉中的人体行为理解、基于内容的图像编码、背景分割与替换、人体姿势分类等多个方面都有广泛应用。静态图像的人体分割面临着巨大的挑战,包括人体姿势的多样性,衣服颜色、纹理的各异性,灯光的变换性以及背景的复杂性等多个方面。一般地说,人体轮廓分割可以分为交互式和自动式分割,本文针对这两类方法进行了较为深入的研究,主要工作和研究成果如下：首先,对于交互式人体分割,传统的图割(Graph Cuts)方法使用用户提供的前景和背景种子点来初始终端连接(Terminal Links),然后使用图的邻域特性来初始邻居连接(Neighborhood Links)。本文提出同时考虑图割中终端连接和邻居连接的方法,该方法使用完全相似图和稀疏表示两种方式通过计算超像素之间的相似性构建图结构,将终端连接和邻居连接嵌入到构建的相似图中,实现了交互式人体分割。实验表明该方法对复杂背景和参数变化具有很强的鲁棒性。其次,对于自动式人体分割,本文提出了基于模型的人体自动式分割方法。该方法根据最大后验概率原理分别对人体躯干和腿部上半肢进行建模,在实现躯干和腿部上半肢检测的同时,为图割和参考信号的独立成分分析(Independent Component Analysis with Reference, ICA-R)提供种子点和参考信号,以完成使用图割分割人体,ICA-R凸显人体的目标。该类方法可以充分利用模型的全局特性和超像素的局部特性快速地完成人体分割。再次,本文提出了层级树的自动式人体分割方法。该方法建立了邻近人体部分的模型,且根据树理论,将人体邻近部分的模型对应到树中各层结点,以此将人体姿势建模为树中从根结点到叶子结点加和。该方法可以有效避免对传统层级式方法和单个人体部分检测器的依赖性。最后,本文根据期望最大化(Expectation-Maximization, EM)算法原理,提出了基于EM算法的人体分割方法,该方法利用EM算法的迭代更新特性对图案结构(Pictorial Structure)模型得到的姿势概率图进行细化以完成人体分割。该方法可以快速地实现对任意人体的分割。更多还原

【Abstract】 Human body segmentation is to segment the human body from the static images which in-clude human body, and can be applied to many tasks of computer vision, such as human action understanding, content-based image coding, background removing, figure pose classification and so on. However, the body segmentation remains a challenging problem in computer vision. Difficulties arise from various body posture, different color and texture of clothes, various light-ing conditions and cluttered backgrounds. Generally speaking, the work can be divided into two categories:interaction method and automatic method. The main contribution is listed in the following:First, for interaction method, the traditional method uses the user labeled foreground and background seeds to initialize terminal links (t-links) and neighborhood links (n-links). This thesis proposes a novel strategy to consider the initialization of t-links and n-links simultane-ously with similarity graph. To achieve the goal, this thesis employs the complete graph and sparse coding to construct the graph for interaction method of human body. The experiments show that the proposed strategy is robust to complex background and parameter changes.Second, for automatic human body segmentation, this thesis proposes model based method. According to maximum a posteriori (MAP), this thesis proposes torso models and upper leg models to detect the torso and upper leg and provide the seeds of graph cut and reference of Independent component analysis with reference (ICA-R). This kind method can fully employ the top knowledge of models and the locality property of superpixel to finish human body seg-mentation efficiently.Third, this thesis proposes the method based on hierarchical searching tree for human body segmentation. This method proposes the model of adjacent parts, which corresponds to the node of tree according to the tree theory. As a result, the pose is modeled as the summary of nodes of a path from the tree foot to the leaf. This method can effectively avoid the dependency of part detector and the dependency of the hierarchy of hierarchical scheme.Finally, this thesis proposes the Expectation-Maximization (EM) method. This thesis pro-poses an EM based human body segmentation method according to the principle of EM. This method employs the iterate property of EM to refine the pose probability map obtained from pictorial structure model to finish the human body segmentation, which has no requirement of poses and can efficiently achieve the body segmentation.更多还原