利用图语法的地理视频流智能解析

发布时间：2018-10-23 15:13

【摘要】：视频GIS是当前地理信息科学的研究热点之一。传感器和计算机视觉技术的快速发展，以及多样终端、异构网络和海量数据的涌现，为视频GIS带来了新的机遇和挑战。如何实现地理视频解析过程的自动化和智能化，以适配复杂的应用环境，是视频GIS亟待解决的问题。目前，视频解析常用的数据驱动方法能够改善基于模型的方法难以解决的多事例、多样性和多模态等问题，并能有效挖掘信息、学习知识。但是，其提取的特征仅局限于底层特征，难以反映高层语义，地理视频解析中存在的“语义鸿沟”仍待解决。同时，视频运动要素的行为通常与地理环境密切相关，考虑空间约束可增强其行为理解的准确性。为此，本文基于视频运动要素完备性定义，采用数值化方法描述地理空间中视频运动要素的相互作用关系。构建以随机图动态分析模型和演化规则为基础，以地理视频内容的结构化描述为核心，以实现地理视频流解析的自动化、智能化为目标的方法体系。具体工作主要包括： (1)综合分析地理视频解析所涉及的地理视频编码、视频智能解析和基于边的随机图三个关键技术，讨论地理空间认知并给出地理视频空间认知地图。 (2)准确定义视频运动要素相关概念，引入单向时间维划分广义、狭义地理空间距离。在此基础上，分析视频运动要素相互作用关系的动态特性及其数值计算方法。基于上下文相关随机图语法，建立稀疏随机图动态分析模型，详细描述融合时空语义信息的随机图动态演化过程。 (3)引入地理空间约束进行视频场景区域分割，给出基于单帧的视频运动要素空间关系定性表示方法。描述视频运动要素空间关系的连续变化过程，利用随机图演化规则，，建立可全局观察和分析的随机图动态演化模型。详细分析可结构化表达地理视频内容变化过程的SRG地理视频特征文件，并给出地理视频内容可视化解析模型。 (4)以视频监控数据集为例，对本文提出的思路进行实例验证，初步实现地理视频智能解析。实践结果表明，本文提出的方法能动态、直观地描述地理视频流中运动要素的空间关系，为地理视频场景的语义描述与智能解析提供一种新的思路。
[Abstract]:Video GIS is one of the research hotspots in geographic information science. The rapid development of sensor and computer vision technology, as well as the emergence of multiple terminals, heterogeneous networks and massive data, bring new opportunities and challenges to video GIS. How to realize the automation and intelligence of geographic video parsing process to adapt to the complex application environment is the urgent problem of video GIS. At present, the commonly used data-driven methods for video parsing can improve the multi-case, diversity and multi-modal problems that are difficult to solve by model-based methods, and can effectively mine information and learn knowledge. However, the extracted features are limited to the underlying features, which are difficult to reflect the high-level semantics. The "semantic gap" in geographic video parsing still needs to be solved. At the same time, the behavior of video motion elements is usually closely related to the geographical environment, and spatial constraints can enhance the accuracy of their behavior understanding. Therefore, based on the definition of the completeness of video motion elements, a numerical method is used to describe the interaction of video motion elements in geographic space. Based on the stochastic graph dynamic analysis model and evolution rules, this paper constructs a method system based on structured description of geographic video content, which aims at automating and intelligently analyzing geographic video stream. The main work includes: (1) synthetically analyzing the three key technologies of geographic video coding, intelligent video parsing and edge-based random graph. This paper discusses geospatial cognition and gives geographic video spatial cognitive map. (2) the concept of video motion elements is defined accurately, and the generalized division of one-way temporal dimension and narrow geographical space distance are introduced. On this basis, the dynamic characteristics of the interaction of video motion elements and its numerical calculation method are analyzed. Based on context-dependent random graph syntax, a sparse random graph dynamic analysis model is established to describe the dynamic evolution process of random graph with temporal and spatial semantic information in detail. (3) Geo-spatial constraints are introduced to segment video scene region. A qualitative representation method of spatial relationship of video motion elements based on single frame is presented. This paper describes the continuous changing process of spatial relation of video motion elements, and establishes a dynamic evolution model of random graph which can be observed and analyzed globally by using the evolution rule of random graph. The SRG geographic video feature file which can structurally express the changing process of geographic video content is analyzed in detail, and a visual analysis model of geographic video content is given. (4) taking video surveillance data set as an example, The method proposed in this paper is verified by an example, and the intelligent analysis of geographic video is preliminarily realized. The practical results show that the proposed method can dynamically and intuitively describe the spatial relationship of motion elements in geographic video streams and provide a new way of thinking for semantic description and intelligent analysis of geographic video scenes.
【学位授予单位】：重庆邮电大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：P208

【参考文献】