【作者】 白亮；

【导师】 老松杨；

【作者基本信息】 国防科学技术大学 ， 军队指挥学， 2008， 博士

【摘要】 信息化与全球化时代视频情报大量涌现并在战略决策中发挥重要作用,研究如何从大量视频情报中获取有价值信息已成为必然,而其核心在于分析获取视频情报包含的语义内容。语义鸿沟的存在使得视频情报语义内容分析面临巨大困难,严重制约了视频情报应用。本文的研究旨在解决上述问题。本文首先建立了视频情报分析体系,指明了视频情报分析需要解决的核心问题——视频情报语义内容分析,进而提出本体支持的视频情报语义内容分析框架。重点研究了该框架下的视频情报低层语义内容抽取、视频情报高层语义内容分析等关键问题,并设计与实现了本体支持的视频情报分析平台(VIAPO, Video Intelligence Analysis Platform using Ontology)。论文的主要贡献体现在以下几个方面:一、提出了视频情报分析的概念体系和技术体系。在概念体系中明确了视频情报分析的概念、任务和层次结构;在技术体系中提出了视频情报分析技术的体系结构以及关键技术。二、提出了本体支持的视频情报语义内容分析框架。定义了视频情报感知概念、元概念、高层概念(元概念和高层概念统称为视频情报概念),将视频情报内容抽象为上述概念以及概念间关系的集合;指明了视频情报感知概念、视频情报概念及其关系的抽象与本体理论的本质联系,提出了本体支持的视频情报知识基础构建方法;提出了紧密结合领域知识、分层跨域语义鸿沟的视频情报分析方法。三、提出了基于颜色空间互信息度量和PetriNet模型的镜头探测方法,提高了渐变镜头的探测准确率;提出基于机器学习的视频情报感知概念探测方法。视频情报感知概念的探测需要对大量高维低层感知特征样本数据进行自动分析处理,从中发现有意义的模式,机器学习是解决这一类问题的有效方法。本文分别采用支持向量机、条件随机域、高斯混合模型等机器学习方法来分类识别重要的音频概念、视觉对象概念和运动类型概念,提高了视频情报感知概念的探测准确率。四、提出了本体支持的视频情报高层语义分析方法。视频情报高层语义分析包括两个方面:视频情报概念探测和视频情报检索。针对以往基于内容的方法的缺陷,提出了本体支持的元概念探测方法,在感知概念探测的基础上,融合低层感知特征和上下文语义信息探测元概念。区别于以往基于内容的方法以及简单线性加权的融合模型,本文提出了基于贝叶斯网络模型的高层概念探测方法,通过贝叶斯网络建模高层概念与低层概念的关联以探测高层概念,提高了视频情报概念探测的性能。针对视频情报检索个性化的需求,提出了基于概念合成PetriNet的视频情报查询描述模型,通过PetriNet模型描述概念之间的时序关系,自定义的建模用户查询语义,满足了用户个性化的视频情报检索需求。五、设计实现了本体支持的视频情报分析平台VIAPO,验证了本体支持的视频情报语义内容分析框架和相关方法的有效性,以及平台在情报分析中的应用效果,为视频情报分析提出了一条可行的解决思路。综上所述,本文提出了视频情报分析体系以及本体支持的视频情报语义内容分析框架,深入研究了视频情报语义内容分析技术,完整的实现了视频情报从低层语义抽取到高层语义概念探测的全过程,有效的解决了视频情报分析面临的语义鸿沟难题。本文的研究不仅为视频情报分析建立了一定的理论和实践基础,同时也将对视频语义内容分析技术产生积极影响。更多还原

【Abstract】 Public video intelligence is emerging as one kind of important resources for analyzing international relations and making strategic decisions. The rapid increase in the available amount of video data is creating a growing demand for efficient methods for understanding and managing it at the semantic level. One of the major challenges facing video semantic content analysis and the related applications is the so-called "the Semantic Gap" between the rich high-level semantics that users desire and the shallowness of the low-level features that the automatic algorithms can extract from the media. In this thesis, we systematically explore the problem of modeling and managing semantics of public video intelligence.Firstly, an architecture for video intelligence analysis is proposed. And video semantic content analysis is shown to be the core for video intelligence analysis. Secondly, a general framework for video semantic content analysis is presented based on ontology. Within this framework, methods of low-level semantic extraction and high-level semantic analysis are developed for video analysis. Finally, the above framework and methods are validated by designing and implementing a Video Intelligence Analysis Platform using Ontology (VIAPO). The main contributions of the thesis are as follows:We propose an architecture for video intelligence analysis, consisting of concept architecture and technique architecture. Concepts and the hierarchy of video analysis are defined within the concept architecture. And key techniques implementing video analysis are illustrated within the technique architecture.We suggest a novel unified framework for video semantic content analysis using ontology. Perception Concept, Meta Concept and High-level Concept are defined. Video semantic content are modeled with the above concepts and the relationships between them. Moreover, the construction of video intelligence knowledge base is proposed using ontology. And we propose a hierarchical approach for bridging the semantic gap combining domain knowledge.We address the methods of detecting Perception Concepts using machine learning techniques. In order to detect Perception Concepts, it is necessary to process high-dimensioned low level features automatically and discover meaningful patterns from the large amount of video data. Three methods are proposed to detect Perception Concepts comprehensively, which are composed of Audio Concepts detection based on Support Vector Machine、Visual Object Concept detection based on Conditional Random Field and Motion_Type Concepts detection based Gaussian Mixture Model.We develop an approach for high-level semantic analysis using ontology, which consists of concept detection in video intelligence and video intelligence retrieval. Meta Concept detection using ontology is proposed to overcome the drawbacks of traditional content-based methods. Based on Perception Concepts detection, Meta Concepts are detected combined with low-level features and context information. With the results of Meta Concepts detection, a novel method for high-level concept detection is proposed using Bayesian Net, which models the relations between low-level concepts and high-level concepts. With the demand of customizing video intelligence retrieval in mind, we propose a query description model based on Perception Concepts and video concepts composite PetriNet. The temporal relationships between the concepts interested by user are modeled by PetriNet, which supports the customization of video intelligence retrieval.We design and implement a Video Intelligence Analysis Platform using Ontology, which gives a sound support to the above framework and methods of video semantic content analysis.In conclusion, this thesis provides an in-depth investigation into the architecture of video intelligence analysis, the framework of video semantic content analysis and methods for bridging the semantic gap. This research is the foundation of video intelligence analysis, theoretically and practically. And it also improves the technology of video semantic content analysis.更多还原