【作者】 张鹏林；

【导师】 关泽群； 王新洲；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2005， 博士

【摘要】 在发达国家视频监视系统一贯被广泛应用于银行、电力、交通、安全、仓储、建筑以及军事设施等领域的安全防范和现场记录报警系统。因此,基于视频时变序列图像的运动分析成为计算机视觉和图像处理领域中的重要研究分支。近年来,随着国民经济的快速发展和综合国力逐步提高,智能化监视系统在军事、银行、交通等关键行业的预警、安全防范和现代化管理中的需求和应用愈加广泛。作为视觉监视系统重要技术支持的时变序列图像运动分析自然就成为计算机视觉的重要研究领域。尤其是随着各国城市化进程的加速,城市中机动车辆的拥有量逐年递增,而城市交通基础设施的建设却相对滞后的情况下,城市交通压力大大增加。在现有交通资源条件下尽可能的为人们提供一个更加舒适的交通环境成为全球各大城市普遍关注的问题。实现交通的智能化利现代化管理成为达到这一目标的重要途径,因此,交通视频监视时变序列图像的运动分析便引起了计算机视觉和图像处理领域研究人员的广泛研究兴趣。 面对城市交通压力逐年增加的严酷现实,城市交通的科学化、智能化管理成为当前各大城市交通管理中的普遍需求,随着信息技术的发展和信息化进程的加快,我国各主要城市开始广泛采用信息技术手段来提高城市交通管理的现代化水平,并为城市交通的现代化管理提供科学的数据支持,为此,一些城市广泛采用用于交通参数量测的传感器,如磁循环探测器等。但由于磁循环探测器固有的弱点,如仅能进行交通流量的量测,另外由于其安装在地面下容易被损坏等。致使其应用受到很大的局限性。然而,基于视觉的视频监视系统不仅具有不易被损坏,而还可以量测多种交通参数等特点而深受城市交通管理和研究部门的青睐,因此,获取的大量交通监视数据的自动化快速分析和理解便成了人们主要关注的问题,如何自动、科学的从大量的视频监视数据中提取运动信息成为管理者的广泛需求,同时也成了研究者的研究热点,本文就是基于这样的背景提出的,目的是通过本文的研究建立一套科学的交通视频监视数据的分析方法和技术手段,从而为交通管理提供科学的数据支持。 动态视频序列图像分析是计算机视觉和图像分析领域的一贯研究热点,论文在认真的分析了该领域的国内外研究现状的基础上,确立了本文研究的主要目标是以背景减的算法思想为基础研究复杂场景背景模型的准确重建方法,通过准确的背景模型重建来达到运动物体的准确分割,并对研究运动物体的追踪方法。在这样的研究目标的指导下通过详细分析国内外大量的已有研究成果,总结了视频序列图像运动分析的基本问题,提出了背景减运动物体提取的一般过程和研究框架。 运动物体提取是基于视觉的城市交通视频监视序列图像运动分析的第一个阶段,也是一个非常关键阶段。准确的运动物体提取是物体跟踪、分类、识别和运动参数提取的基础。多年来国内外研究人员进行了大量的研究,提出许多针对不同场景的运动物体提取方法。尽管如此,复杂场景的运动物体提取问题始终未能很好的解决,本文以城市交通复杂场景为研究对象,以背景减运动物体提取思想为基本出发点,以准确的背景模型重建为研究的指导心想,更多还原

【Abstract】 In developed countries, video surveillance system is widely used in safeguarding, real-time recording and alerting systems in banks, power systems, traffic and transportation, security, warehouse, architecture and military equipments. In that case, the analysis of moving objects based on video time serial images becomes a very important branch of Computer Vision (CV) Research. In recent years, as a result of the rapid progress of our national economy and advance of general national power, the desirous and use of intellectual surveillance systems is more and more wide in protecting, secure guarding and modern management in key industries such as military, banking and traffic controlling. As an important technique support of vision surveillance system, the analysis of time serial images naturally becomes an important field in the CV domain. Especially as the acceleration of urbanization, the quantity of motor vehicles is increasing year by year, but the infrastructure of urban traffic is comparatively lagged behind, so the pressure of urban traffic system has growing badly. How to provide a traffic environment as comfortable as possible becomes an import problem in metropolitans all over the world. The implement of traffic-intelligence and modern traffic management becomes a good solution for this problem. In that case, the movement analysis of time serial images from traffic video monitors has widely interested many researchers of CV and image processing.With the increasing traffic problems year by year, the management by scientific and intelligentizing of urban traffic has become the universal requirement of various main urban traffic management, with the expediting of information technology and information progress, each main city in our Country has began using information technology to enhance the modernization level of urban traffic management, and provide scientific data auspice for modern urban traffic management, thus some cities begin using sensor for measuring in traffic parameter widely, such as detector of magnetic circle and so on. However, because of the connatural weakness of detector of magnetic circle such as just carry through in measure of traffic flux, and it would be destroyed easily when installed on the ground, and so on. So the application of it is limited. Nevertheless, video surveillance system base on vision can’t be destroyed easily and it could measure multi-kinds of traffic parameter, because of that, it loved by the departments of traffic management and research. So acquiring automatization in analyzing and comprehending of a great deal of traffic surveillance data is the main issue for people. How to pick up dynamic information from a great deal of traffic surveillance data automatically and scientificly became administrators’ broad requirement, and it also became researchers’ hotspot in research, this article is brought forward in the background, its goal is to build up a suit of scientific analyzing method and technique instrument in acquiring traffic video surveillance data, so as to provide scientific dataauspice for traffic management.Analyzing serial dynamic images is one of the right research hotspots in the area of computer vision and images analyzing, base on analyzing the status quo of research inside and overseas in this area, this article’s main goal is researching the exact rebuilding method of complex scene and background base on arithmetic of decreasing the background, to exactly divide the dynamic objects by rebuilding the exact background models, and pursue the dynamic objects. Minutely analyzing a great deal of research results inside and overseas in the direction of research goal, this article summarize the basic issue of analyzing of serial dynamic images, and bring forward a generic progress and research frame about picking up dynamic objects in background decreasing.The detection of moving objects is the first stage of vision-based urban traffic video monitor time serial images, which is also the key stage. A precise detection of moving objects is the basic of object tracking, classifying, recognition and parameter acquiring, in which home and abroad researchers has done a lot in recent years, and many methods have been put forward. Although, the detection of moving objects in a complex scene has never been done well. This thesis objects in complex urban traffic scene, starting from the "background minus moving objects" method, and the theme is precise background model reconstruction, which tries to achieve the goal of precise movement segmentation by precise background portraying. As a result, this article is based on classic Kalman Theories, has investigated the application methods of Kalman Filter theories in background model reconstruction of complex scene, and analyzed the advantages and disadvantages of Kalman linear optimal filter theory applied in background model reconstruction of complex scene in different viewpoint; since the disadvantage of Kalman Filter in background model reconstruction of complex scenes, we proposed a new algorithm for complex scene time series image background model reconstruction, which derives from the theory that background signal energy varies by time, and is supported by the theory of Lebesgue measurement and integral theory. What’s more, the validation of algorithm is verified by a lot of experiments.Dynamic division is the important phase of dynamic analyzing, in the progress of dynamic division, the choose of dividing threshold is a important and necessarily calculated issue, to solve this problem, this article bring forward the theory of weighed enhancing mean threshold for decreasing background images, consequently it solve the problems of choosing dynamic threshold automatically in dynamic division, it decrease the infection due to manual interference in choosing threshold, and validate the validity of this method by experimenting.Multi-object tracking is an important stage of urban traffic video image series movement analysis after moving object detection and before movement parameter acquiring. Most of existing research on this topic is achieved on the basic of the feature of objects. This article sponsors the tracking method based on the object’s centroid after the analysis of research object’s actual feature.So, the multi-object search in moving variable slide window (VSW) algorithms by continuous modifying the size and center of window is put forward to achieve the segmentation and centroid measurement, which is verified to be valid by lots of experiments. Consequently it provide new solving artifice for pursuing multi-objects.Pursuing dynamic objects is one of the main phases in dynamic analyzing, it is the first phase in picking up dynamic parameters after picking up dynamic objects and dividing multi-objects, it also is the analyse basic of other dynamic parameters(such as moving speed, the estimate of dynamic locus, the information of the transformation of cars). In total, there are two main dynamic Pursuing methods that one base on character and the other base on light flow . The one that base on light flow have some weakness in dynamic analyzing in complex sense as we mention in the chapter in this article, so, this article mainly base on the theory that base on character, it brings forward the thought of pursuing mobiles that base on the center of mass, that is pursuing mobiles’ Dynamic progress by pursuing the center of mass. Moreover pursuing the center of mass is by Kalman Filter theories, experiments indicate that the thought has nice purpose whether in the perfect sense inside or in actual complex traffic sense outside. Moreover the study of this article indicate that the pursuing thought that base on the center of mass also has definite applied value for pursuing multi-objects.Additionally, it is valuable in practice to extract motion objects based on features. Finally, a proposal that corners features of rigid objects are used has been discussed in this paper. Since traditional matching methods with correlation coefficients often keep low precision, a novel approach has been presented in this paper, the approach performs local Walsh transformation of corners features of rigid objects from neighbor frame images, then match between objects with feature of local entropy. Experiments show that high precision of matching has been achieved using the proposed method.To evaluate accuracy and efficiency of the proposals in this paper, a prototype experimental system has been developed using oriented object programming.Summarily, a general frame to extract motion objects from complex background has been suggested by analyzing previous studies. Accordingly, some algorithms have been proposed specifically in city communication based on traditional theories of motion analysis, i.e. reconstruction of background model in complex scene, dynamic threshold selected automatically, segmentation of multi-objects and measurement of quality center, objects trace based on kalman filter and quality center. Finally, advantages and disadvantages of the proposed methods in this papers have been given, as well as further research.更多还原