视频中运动目标跟踪算法的研究 硕士论文 - 图文

视频中运动目标跟踪算法的研究 中文摘要

视频中运动目标跟踪算法的研究

中文摘要

视频目标跟踪是机器视觉研究领域的关键技术之一，也是该领域的研究热点，广泛应用于机动目标跟踪、机器人研究、图像目标编码和人机接口等。而要在目标遮挡、快速移动、目标形变、光照变化、背景噪声以及有实时性要求等条件下能够实现稳健的目标跟踪是学者们关注的焦点，也是目前在实际应用过程中一个亟待解决的难题。

根据研究的实际需要，本文重点对均值偏移（Mean Shift）和粒子滤波算法进行了深入的研究。Mean Shift跟踪算法是一种基于密度梯度上升的非参数方法，通过有限次迭代运算找到目标位置，实现运动目标的跟踪。该算法计算量小，实时性较强，但不适用于PDF多峰的情况，且容易陷入局部极大，若存在物体运动过快或者是被严重遮挡等情况，往往会导致跟踪失败，且无法恢复跟踪。本文在第三章中对Mean Shift算法理论及其在目标跟踪中的应用作了详细的推导和描述，并针对基于颜色直方图特征的Mean Shift跟踪算法对光线变化、背景相似等过于敏感的问题，引入了梯度方向直方图特征，通过大量实验证明了基于梯度方向直方图特征的Mean Shift跟踪算法对光线变化和局部区域的微小形变不敏感，具有良好的跟踪性能；最后对基于梯度方向直方图特征的Mean Shift跟踪算法的优缺点做了实验分析。

粒子滤波是从上世纪90年代中后期发展起来的一种新的滤波算法, 广泛地应用于非线性、非高斯系统的估计中。其基本思想是用随机样本来描述概率分布，然后在测量的基础上，通过调节各个样本（粒子）权值的大小和样本的位置来近似实际概率分布，并以样本的均值作为系统的估计值，在短时丢失目标的情况下能自动恢复跟踪，在非线性、非高斯条件下也能实现稳定的跟踪。然而这种方法存在着粒子退化和计算量大等问题，应用在实时跟踪系统中有一定的局限性。本文在第四章中对粒子滤波算法的基本原理及其在目标跟踪中的应用作了详细的讨论，并实现了基于颜色特征的粒子滤波算法的跟踪实验。实验表明，与基于梯度方向直方图特征的Mean Shift跟踪算法相比，粒子滤波跟踪算法具有很好的抗遮挡和抗干扰性，但由于计算量大而导致实时性较差。

I

中文摘要 视频中运动目标跟踪算法的研究

理论分析和实验证明，基于梯度方向直方图特征的Mean Shift跟踪算法实时性较好，但在遇到目标遮挡或运动过快时容易丢失目标；基于粒子滤波跟踪算法虽有较强的抗遮挡能力，但存在实时性差等问题。针对这些问题，本文在第五章中提出一种融合的跟踪方法，基本思路是：正常情况下采用基于目标梯度方向直方图特征的Mean Shift跟踪算法跟踪目标，当候选目标与目标模型的相似度小于设定阈值时，自动切换到粒子滤波跟踪算法进行后续的跟踪修正。实验结果显示本文算法能有效地解决目标因遮挡或运动过快而导致的跟踪丢失问题，同时减轻了粒子的退化现象，提高了算法的实时性，并在图像对比度较低情况下也能较好的跟踪目标。

关键词：运动目标跟踪；梯度方向直方图；Mean Shift；粒子滤波；融合算法

作 者：陈家波 指导教师：赵勋杰

II

Research on the Algorithms of Moving Object Tracking in Video Abstract

Research on the Algorithms of Moving Object Tracking in Video

Abstract

Moving object tracking in video is the critical technology of the machine vision as well as the hot issue within certain research domains, which is widely applied in surveillance, robotics, object-based video coding, human machine interface, etc. However, how to accomplish the task of robust tracking under the condition of the fast motion, occlusion, object deformation, illumination variation, background clutters, real-time restriction, etc is concerned by scholars. It is also a tough problem yet to be resolved in practical application at present.

In terms of the actual needs, this paper has deeply looked into the algorithms of mean shift and particle filter among many object tracking algorithms. The algorithm of mean shift is a kind of non-parametric methods based upon the climbing gradient. It realizes the aim of object tracking through iteration. The obvious mean shift is superior in less amount of calculation, simpler and easier to utilize, so it can meet the need of real-time tracking. But it fails in tracking fast moving targets and recovering a track after a severe occlusion. The third chapter gives the deduction and the description of the algorithm of mean shift theory and its use in object tracking in detail. Meanwhile, the tracking performance of the mean shift algorithm, based on color histogram, declines rapidly when the object is similar to the background color or the illumination changes. This paper focuses on the mean shift tracking algorithm on the basis of histograms of oriented gradients. The experimental results show that the above method is not sensitive enough to illumination changes or the partial small deformation；Finally, the experimental analysis of drawbacks of the algorithm of mean shift is stated in the third chapter.

The method of particle filters has been developed since the 1990s as a new filter calculation and applied successfully in a variety of nonlinear and non-Gaussian filtering estimations. The fundamental principle of this technique is to describe the probability densities by sets of random samples, which allows online, real-time estimation of nonlinear, non-Gaussian dynamic systems. However, two common problems of the particle

III

Abstract Research on the Algorithms of Moving Object Tracking in Video

filter-technique are the degeneracy phenomenon and the huge computational cost. Thus, those problems will be the bottlenecks to the application of particle filter in real-time tracking systems. The algorithm of particle filter theory and its utilization in object tracking are fully discussed in the forth chapter. The experimental results show that the algorithm of particle filter is much more remarkable in anti-shelter and anti-inference, in comparison with the algorithm of mean shift, in addition to its worse real-time due to the huge amount of calculation.

Since object tracking algorithm based on histograms of oriented gradients always loses objects under the condition of occlusion or fast motion and particle filter tracking algorithm costs plenty of time in calculation, the object tracking algorithm based on histograms of oriented gradients integrated with the particle filter is proposed in the fifth chapter. Under normal circumstances, the object is tracked by the algorithm based on histograms of oriented gradient, when the conformability of candidate object is less than the threshold. The tracking result would be verified by particle filter algorithm. The experiments demonstrates that the algorithm effectively resolves the problem of the object-loss under the condition of occlusion or fast motion and the algorithm comes out with a better real time quality, which also overcomes the particle-degeneracy problem and achieves good tracking results under low-contrast conditions.

Key words: Object tracking; Histograms of oriented gradients; Mean Shift; Particle filter; Fusion algorithm

Written by: Jiabo Chen Supervised by: Xunjie Zhao

IV

目 录

中文摘要 ............................................................................................................................... I ABSTRACT ....................................................................................................................... III 第一章 绪 论 ....................................................................................................................... 1 1.1 论文研究背景及意义 ................................................................................................. 1 1.2 目标跟踪技术研究的现状及面临的难题 ................................................................. 2 1.2.1 研究现状.............................................................................................................. 2 1.2.2 研究面临的难题.................................................................................................. 3 1.3 本文的研究内容及章节安排 ..................................................................................... 4 第二章 运动目标跟踪处理的相关理论 ............................................................................. 6 2.1 引言 ............................................................................................................................. 6 2.2 目标检测 ..................................................................................................................... 6 2.3 目标跟踪 ..................................................................................................................... 8 2.4 本章小结 ..................................................................................................................... 9 第三章 MEAN SHIFT算法的理论基础以及在目标跟踪中的应用 ............................. 10 3.1 引言 ........................................................................................................................... 10 3.2 MEAN SHIFT跟踪算法 ............................................................................................... 10 3.2.1 无参密度估计理论............................................................................................ 11 3.2.2 Mean Shift算法原理 .......................................................................................... 13 3.2.3 基于Mean Shift的目标跟踪算法 .................................................................... 17 3.2.4 算法实现及结果分析........................................................................................ 19 3.3 基于梯度方向直方图特征的MEAN SHIFT目标跟踪算法 ..................................... 21 3.3.1 梯度方向直方图特征........................................................................................ 21 3.3.2 基于梯度方向直方图特征的Mean Shift目标跟踪算法 ................................ 23

百度搜索“77cn”或“免费范文网”即可找到本站免费阅读全部范文。收藏本站方便下次阅读，免费范文网，提供经典小说综合文库视频中运动目标跟踪算法的研究 硕士论文 - 图文在线全文阅读。