节点文献
地理场景协同的多摄像机目标跟踪研究
Object Tracking Methods Based on Geographic Scene and Multi-camera
【作者】 张兴国;
【导师】 刘学军;
【作者基本信息】 南京师范大学 , 地图学与地理信息系统, 2014, 博士
【摘要】 地理场景具有高动态性、多尺度性和不确定性等特点,研究动态目标的实时感知方法,快速智能感知地理场景中动态目标时空特征,基于此探索动态目标的行为规律,已成为当前学术界和政府管理部门亟待解决的问题。目前,视频监控系统以其高清实时、功能智能、价格低廉等优势在安全领域发挥着越来越重要的作用。然而,现有基于监控视频的动态目标跟踪局限于二维图像空间,无法感知其在真实地理场景中的时空特征,对于具有大量视频监控设备的区域,不能实现多摄像机对动态目标的协同跟踪。鉴于此,本文以监控视频与地理场景的协同分析为手段,针对地理场景下动态目标状态的感知这一科学问题,发展地理场景约束下的动态目标时空信息提取与分析方法,取得了以下研究成果:(1)提出了多平面约束下的监控视频与2D地理空间数据几何互映射模型。现有监控视频至地理空间的映射模型假定地面为单一高程,当监控区域存在多个不同的高程平面时,现有方法需重新求解互映射矩阵,过程较繁琐。本文通过对比分析摄影测量学与计算机视觉的相机模型,构建了适用于固定摄像机与PTZ摄像机的监控视频与2D地理空间数据几何互映射模型,该模型具有各参数物理意义明确、理论严密、灵活性强等特点。(2)提出了基于深度的监控视频与3D地理空间数据几何互映射模型。从三维空间坐标到图像坐标的转换可基于针孔成像模型来实现,但从图像坐标到三维空间坐标的转换,现有方法主要通过视线与3D模型相交的方式来实现,其计算量大且过程繁琐。本文基于3D地理空间数据可视化过程中的缓存深度值构建了监控视频与3D地理空间数据的互映射模型。与传统方法相比,本模型具有过程简洁,效率更高,可实现监控视频与3D地理空间数据间实时同步的动态互映射。(3)提出了一种监控视频与地理空间数据的半自动互映射方法。本文系统分析了传统的基于单应的几何互映射方法,并进行了不确定性分析。鉴于地理空间数据精度日益提高,本文基于结构化地理场景的约束,设定灭点相似性、特征线相似性两个指标作为匹配的依据,探讨了监控视频与2D/3D地理空间数据视图的半自动匹配方法。(4)设计了一种面向监控视频的前景目标时空信息提取方法。基于监控视频与地理空间数据互映射模型,提出了地理场景中监控视频前景目标时空信息的提取方法,包括目标方位信息、几何信息、目标轨迹、前景图像等,基于面向对象思想建立了其对应的时空数据模型,实现了对前景目标数据的管理与GIS集成。(5)构建了一种基于路网约束的盲区目标轨迹估计模型。监控摄像机通常布设在相对重要的位置,具有独立性、分散性等特点,完全基于监控视频无法感知监控目标的连续运动轨迹。现有多摄像机协同下的目标连续跟踪,多利用监控视频场景间具有重叠条件下开展研究,不适合大场景中目标的连续跟踪。本文以场景中动态目标的时空信息及基础地理信息为数据基础,提出了路网(及目标行为规则)约束下的动态目标连续跟踪方法。(6)研发了地理场景中多摄像机协同的动态目标连续跟踪原型系统。基于以上研究成果,设计并开发了动态目标连续跟踪原型系统,该系统具有监控视频与地理空间数据的互映射、动态目标时空信息提取、动态目标跟踪及可视化、大场景下监控目标轨迹估算等功能。
【Abstract】 With the development of society and economy, human activities become more and more frequent. This forms a more dynamic, multi-scale, uncertain and complex system. So, fast perception and monitoring of dynamic objects has become an urgent problem in regions of academia and government administration.The surveillance video is a kind of real-time and high-definition data source and contains a wealth of spatial information and attribute, which plays an important role not only in security but also in GIS. At present, the integration of surveillance videos and geospatial data are focused on static, one-way and interactive operation. It is difficult to meet the needs of the dynamic, two-way, automatic mapping between the surveillance videos and2D/3D geospatial data. And object tracking algorithms are mainly focused on the single surveillance video and it is difficult to form the continuous trajectory of the object in a large area which results in the failure of the understanding of the dynamic object’s behavior.It is necessary to develop an efficient and macroscopic analysis technology with the integration of intelligent video analysis and spatial analysis to meet the requirements for the whole view and behavior understanding of the dynamic object. Therefore, we combine videos with2D/3D geospatial data and propose some methods of object information extracting and analysis in geographical scene. The contributions are as follows:(1) The mutual mapping modal between surveillance videos and2D geospatial data based on multi-planar constrains is proposed. Current2D mutual mapping model assume the ground surface has the same elevation. When there are multiple different elevations, the existing methods need to respectively solve the matrix and the process is cumbersome. Through the comparative analysis of the differences in photogrammetry and computer vision camera model, a new mutual mapping modal is proposed and the parameters have clearly physical meaning. The modal is not only applicable to fixed cameras, but also dynamic cameras, such as PTZ cameras. Based on the modal, we analyze the characteristics of the mapping deviation under undulate ground and the resolution of surveillance videos.(2) The mutual mapping modal between surveillance videos and3D geospatial data based on depth buffer is proposed. Based on pin-hole camera modal, the3D coordinates can be converted to the corresponding image coordinates. But the transformation from the image coordinates to3D coordinates is difficult. The existing method is mainly to calculate the intersection point between the sight line and3D scene. We propose a new modal which is based on the depth buffer when3D data display in3D GIS. Through the modal, the surveillance video and the corresponding3D GIS view can achieve real-time synchronization. Based on the modal, we analyze the characteristics of the resolution of3D surveillance videos and dynamic mapping.(3) A semi-automatic mapping method between surveillance video and geospatial data is proposed. The mutual mapping uncertainty was discussed through the analysis of the homographic matrix method. In order to achieve the automatic mapping between some surveillance videos and3D GIS views, a new method is proposed, which is based on the features of vanishing points and lines.(4) A dynamic object information extracting method is proposed based on geographic scene and the corresponding data model is designed. The information contains the location and direction, geometric size, trajectories, foreground image under geospatial reference. The data modal is based on object-oriented method and is easily integrated with GIS.(5) An object trajectory estimation model is proposed. In order to get the whole trajectories in a large area where was not all covered by surveillance cameras, we combine geospatial data with the movement parameters, geometric size and color information of the object to track the object. On the one hand, through the road and the object behavior regulars, the object location can be estimated. On the other hand, the similarities including the location probability, the feature similarly of geometric size and movement parameters, images similarity can be calculated.(6) We design and develop a system for object continuous tracking. This system contains functions of the mapping between videos and geospatial data, object tracking, information extracting and trajectory estimation.
【Key words】 GIS; video; geographic scene; mapping; dynamic object tracking;