【作者】 魏波；

【导师】 李在铭；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2000， 博士

【摘要】 点时空约束图像目标跟踪理论和实时实现技术研究探讨了随机多维信号在‘欠定’约束条件下，‘子信号’特征检测与提取的理论与技术。针对数字随机图像序列，研究了点时空约束条件下图像目标的检测识别与跟踪技术，并进一步研究了基于DSP的实时实现，完成了相应工程项目的研制任务。 本论文首先指出本系统的特点就是系统获得的目标信号信息仅仅是在某一时刻（时间约束）、目标上某一点（点约束）的位置信息（空间约束），本文称之为欠条件的、点时空约束图像目标跟踪系统。根据目标的点时空约束条件，建立了目标的导出前提条件，对目标图像信号做了进一步的约束，然后建立了相应的多维随机信号检测处理模型和点时空约束图像目标跟踪系统的跟踪方程，并通过探讨如何求解该系统跟踪方程，建立了系统实现的理论模型并提出了相应的各项关键技术，特别是通过将系统分为搜索捕获、正常跟踪、记忆预测等工作状态，并利用目标识别特征的短时平稳性实现了对目标长时间的稳定跟踪。 基于图像运动场间断点检测的运动图像目标分割提取技术：在复杂场景中，有效的检测域就是目标的运动特征域。我们提出了基于Markov随机场的运动边缘的检测技术，以分割提取运动目标。运动场的间断点一般是由于目标运动造成的，通过建立一个假设的、可用Markov随机场描述的运动场间断点分布的先验模型，来检测运动场间断点，并进而提取出运动目标。 基于序列图像补偿差分的运动图像目标分割检测技术：该算法首先在时间域对摄像机运动造成的全局运动进行补偿，并提出了相应的快速算法，对序列图像进行补偿后，再进行差分运算，然后在2D平面空间域对图像进行边缘提取，利用边缘图像和差分图像进行hausdorff距离匹配，以提取出准确的目标位置信息，克服目标运动造成的背景的显现和遮拦对目标准确检测的影响。这一解决方案充分利用了序列图像在时间域和空间域的相关信息，能较为准确地提取出目标区域，这一技术还可用于第三代图像编解码中的视频对象平面（VOP）的生成。 基于目标特征图像匹配的图像目标识别跟踪技术：提取目标的特征图像区域，并以此特征图像为匹配模板的图像目标匹配识别技术。为保证匹配的抗干扰特性，提出了去均值的匹配新准则；分析了图像参数对匹配的影响，给出了搜索目标特征图像的标准；最后给出了实际系统中目标初始匹配特征图像的建立和跟踪中错误匹配检测的具体算法。该技术成功地避开了在复杂 电子科技大学博士论文：摘 要场景中目标分割付固积很困难这一难题，在实际应用中得到了成功的应用。 针对实际系龌求，进行了硬件设计。讨论了系统硬件的功g骡蛐分和具体的硬件模块的设计，特别分析了点时空约束图像目梆固淙系统独特的序列图像的滑动窗瓣鹏施以及实删术。此硬件平台保证了智能、可编程、多功能点时空约束图像目标召团淙系统的实时实现。 根据系统的软件设计，提出了软件图像目标RSJGgg的概念及实现的关键技术厂一系统软件的积木式开腑构架，并给出了一个具体实现的开放性的程序框架，并进行了关键程序部件的设计。 论文最后研究了如何利用文中提出的理论与技术完成相应工程项目的设计。实际系腻相关研蛾和部队基地进行的交付试验取得了良好的结果。实际应用表明论文新提出的理论、技术及具体的实现设计是成功的。更多还原

【Abstract】 Study on the theory and real time implementation technology of image target tracking with point constraint in time-space domain mainly discusses how to extract and recognition son signal’s feature under owe-deciding constraint in random multi-dimension signal. Dealing with digital random sequence image signal, this paper studies the theory of detecting and tracking image target with point constraint in time-space domain, and the key technologies of how to implement it with DSP chip. With this research work, we have done the developing task of relevant engineenng project. This paper firstly points out that the image target tracking system that we have developed is the image target tracking system with insufficient condition and point constraint in time-space domain for the target information that the system gets only is the location information (space constraint) of some point on image target (point constraint) at some time (time constraint). This paper establishes elicitation premise of target according the point constraint in time-space domain, and this provides the possibility to detect and extract target image signal, after that this paper establishes multi-dimensions signal detecting and processing model and the tracking equations of the image target tracking system with point constraint in time-space domain. By discussing the solution of the tracking equations, it builds the theory model for system accomplishment and produces some key technologies, especially it divides working state of the system into three main states: searching and capturing state, tracking state and tracking by prediction state, and discusses how to track image target stably for a long time with stationary of the image target’s feature in short time. Moving image target detection and extraction technology based on gaps in image motion field: target is very difficult to be extracted from background in gray domain, but is easy to be detected in motion feature domain. It produces the new algorithm to extract the image target based on the detecting gaps in motion field according Markov random field (MRF) theory. Alter analyzing the sequence image signal, we can find that the gaps in motion field are made by target motion. This algorithm firstly builds a prior model that describes the motion gaps distribution according Markov random field theory, and then detects the image target by this model. Moving image target detection and extraction technology based on sequence image compensation difference: this paper produces the algorithm working in 2D+T three dimension information space to detect image target in complex dynamic scene. This algorithm compensates the global motion made by camera motion in time domain and it is quick algorithm that can be accomplished in real time system. After that this algorithm extracts image edge in 2D space domain, and then it takes edge image to match compensation difference image according hausdorff criterion. This operation can extract accurate target image by avoiding the affection of background appearance and obstruction made by target motion. This algorithm can extract accurate target image by applying the information in time domain and space domain. This algorithm not only can be applied in image target tracking system, but also plays an important role in the third generation image code technology, especially in generation of video object plane (VOP). Image target tracking technology based on feature image matching: this paper produces the algorithm that extracts feature image in first image and ta更多还原