动态系统分段模型及其在视频运动模式挖掘中的应用

Motion pattern detection and mining in video sequence is currently one of the most active areas of high-level vision analysis and understanding, largely because motion information play an important role in describing the semantic content of video data. In this proposal, a new theoretical research framework is put forward based on dynamical system segment model(DSSM), focusing on how to learn the basic human motion patterns and how these basic patterns fit the human actions. The main content of this proposal covers four parts: (1)how to mine and describe the basic human motion patterns; (2)how to construct and optimize the basic motion pattern dictionary; (3)how to learn the human action model based on the motion patterns; (4)how to recognize and understand complex human behaviors. Our framework shows promise compared with the traditional ones in two aspects. First, we propose describing the action segments by dynamical system, while these segments are learned by using the segment model. Such segments can be noted as Dynamical System Words, which means they are the human motion primitives. By mining the human action database, we can get all the human motion patterns and thus form the motion pattern dictionary, which noted as Dynamical System Dictionary. Second, we propose a non-parametric model HDP-DSSM to describe the human action distribution and semantics by combining the Hierarchical Dirichlet Process(HDP) and DSSM. The HDP model has continuous and infinite state space. It can discover unseen states and identify the new class information when new observations are presented. So it is very suitable for modeling state-varying time series data like video motion sequence. By combining the HDP and DSSM, we can learn and mine the basic patterns and the pattern distribution in an unified framework. After that, we can realize the recognition and understanding of complex human behaviors.

视频序列中运动模式的检测与挖掘是高层视觉信息分析与理解的一个重要研究内容。本项目围绕人体行为基本运动模式以及基本运动模式分布规律如何挖掘的问题，提出建立一种全新的基于动态系统分段模型的视频人体运动模式挖掘的理论研究框架。基于该研究框架，本项目拟从人体行为基本运动模式的挖掘与描述、基本运动模式数据库的构造与优化、基本运动模式分布规律的学习与挖掘以及复杂人体行为的识别与理解等方面开展一系列创新性的研究工作。在基本运动模式的表达上，通过引入分段模型，将人体运动特征数据序列连续地分割成若干序列片段，并用动态系统对这些序列片段进行描述。在基本运动模式分布规律的表达上，提出一种非参数化的HDP-DSSM模型来描述人体行为的构造规则和主题语义。最后将两者相结合，实现对视频序列中复杂人体行为的识别与理解。

视频序列中运动模式的检测与挖掘是高层视觉信息分析与理解的一个重要研究内容。本项目围绕人体行为基本运动模式以及基本运动模式分布规律如何挖掘的问题，提出建立一种全新的基于线性动态系统的视频人体运动模式挖掘的理论研究框架。基于该研究框架，我们从人体行为基本运动模式的挖掘与描述、基本运动模式数据库的构造与优化、基本运动模式分布规律的学习与挖掘以及复杂人体行为的识别与理解等方面开展一系列创新性的研究工作。截止到目前为止，我们从线性动态系统的角度，提出了一种全新的能够同时刻画视频表观信息和运动信息的特征提取方法。在此基础上，我们将Fisher vector方法推广到流形空间，提出了一种流形切空间的tangent Fisher vector方法。最后，我们构建了视频内容分类与识别领域的大规模数据库，并将我们的工作在该数据平台上进行了评估与测试，我们的方法均达到了国际领先的性能。这一系列工作我们已提交并发表在国际顶级刊物和会议TPAMI、PR、ICIP上。