【作者】 张小燕；

【导师】 吴晓娟；

【作者基本信息】 山东大学 ， 信号与信息处理， 2009， 硕士

【摘要】 运动目标跟踪在自治车辆导航,机器人控制,基于运动的识别,视频压缩,基于视觉的控制,人机接口,医学成像,增强现实和视频场景监控中都具有重要的应用前景。随着应用的推广,各种新技术被应用到目标跟踪中来适应更加复杂的环境,但到目前为止,还没有一种算法对所有的情况都能适用,所以研究一种鲁棒性好、精确、高性能的运动目标跟踪方法是该研究领域所面临的一个巨大挑战。运动目标跟踪的最终目的是分析和理解其行为以及和其他目标的交互关系,运动目标的行为理解和描述引起了各国学者的高度重视,已成为最具挑战的研究方向,它是将计算机视觉由低、中层次的处理推向高层抽象思维的关键问题。本论文在回顾前人工作,研究各种跟踪算法及其应用场合的基础上,提出了一种新的跟踪算法,满足了视频监控系统对跟踪算法实时性的要求,解决了交互、碰撞、目标进出场景、多目标跟踪等问题,实现了监控系统中对多种运动目标(如人,鱼群,鸟群)的跟踪,并在数字全息干涉测量技术测量溶液浓度变化中,实现了溶液浓度变化过程的无噪声干扰跟踪重建;特征作为目标的行为描述和身份标识,是进行计算机视觉高层研究的主要信息,本文在跟踪的基础上提取了目标的关键运动特征。本文首先针对运动目标跟踪中存在的问题,提出了AMCMC(Added MarkovChain Monte Carlo)粒子滤波跟踪算法。AMCMC算法是把RJMCMC(ReversibleJump Markov Chin Monte Carlo)算法中检测进入场景和离开场景目标的部分移到马尔可夫链外,使其抽样过程相当于MCMC(Markov Chain Monte Carlo)粒子滤波器。算法的运算复杂度低,可以跟踪可变数量的运动目标,还可以定义交换运动模式来解决遮挡、碰撞等问题。通过在AMCMC算法中定义不同的观测模型和运动模式,实现了对运动人体、机动性很强的群体运动目标鱼和鸟的跟踪。通过不同运动模式不同观测模型的思想,进一步降低了运算的复杂度,达到了实时跟踪的要求。对于目标的特征提取,本文主要提取目标的运动特征。在提取目标基本尺寸特征的基础上,提取了目标的运动速度和方向,并重点提出一种在视频中利用目标尺寸的变化特点测量鸟的翅膀拍打频率的方法。通过对鸟在视频中的运动进行分析,得出鸟在图像中的尺寸变化周期与鸟的拍打频率一致的结论,采用短时傅立叶变换对尺寸信号进行时频分析,并计算信号的局部平均频率,准确测量了鸟的局部拍打频率随时间的变化。在实现溶液浓度变化过程的无噪声干扰跟踪重建时,首先介绍了跟踪重建的过程和原理,针对存在噪声的特点引入了一个似中值滤波器的后序滤波器,消除了噪声对测量结果的影响。更多还原

【Abstract】 Moving targets tracking has an important application foreground in self-government vehicle navigation, robot control, tracking-based recognition, video compression, vision-based control, man-machine interface, medical imaging, augmented reality and video monitoring. With the extended application, a variety of new technologies have been applied to target tracking in more complex environments. But so far, no algorithm is applicable to all cases, so it is a great challenge to study a robust, precise and high-performance moving targets tracking algorithm. The ultimate goal of moving target tracking is to analyze and understand their behavior and the interactive relationship with other targets. Understanding and describing of behavior of moving objects have attracted great attention of scholars, and have become the most challenging research direction. It is the key issue to push the low-level and medium-level processing of computer vision to high-level abstract thinking.Based on reviewing the previous work, and doing research in a variety of tracking algorithms and their application occasion, the thesis proposes a new tracking algorithm, which can meet the real-time requirement of video surveillance system and can solve tracking problems of interaction, collisions, targets entering and leaving scene, multi-target tracking and so on. Using the proposed algorithm, the thesis successfully tracks a wide range of moving targets (eg, human, fish, birds). Additionally, in the measurement of change of solution concentration using digital holographic interferometry, the thesis achieves noise-free tracking and reconstruction of solution concentration change. For features of targets are main information for high-level research, this thesis extracts main motion features of targets based on tracking.First, an Added Markov Chain Monte Carlo (AMCMC) particle filter algorithm is proposed. Different from the Reversible jump Markov Chain Monte Carlo (RJMCMC) algorithm, we add a new entered and left objects detection part after the MCMC particle filter which makes the sampling process like a MCMC particle filter. The AMCMC algorithm has a low computational complexity, can easily handle the variable dimension state vectors processing, and can define swap move type to resolve problems of interaction and collisions. By defining different observation models and move types in AMCMC algorithm, the thesis achieves tracking of moving people, multiple maneuverable targets fish and birds. In tracking, different observation models are used in different move types to track accurately with low computational complexity.Then, the thesis extracts main motion features of targets. The moving speed and direction of targets are gotten based on the basic size and position features. The find that the size of bird’s area in image fluctuates with wing flaps in the same period gives us inspiration for measuring the wing flap frequency by analyzing the size signal of bird’s area. Short-Time Fourier Transform (STFT) is used to analyze size signals after a smooth filter and a normalizing filter to reflect local time flap frequencies. The average frequency of every window is computed to get the quantitative local frequency.Finally, for achieving noise-free tracking and reconstruction of solution concentration change, the thesis introduces the process and theory of tracking and reconstruction firstly, and then a similar median filter is used as a post-filter to eliminate the noise on the measurement results.更多还原