【作者】 文志强；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 计算机应用技术， 2008， 博士

【摘要】 视频目标跟踪问题引起广泛关注是由于它能够应用于民用和军事的许多领域,如视频监控、机器人导航、人机接口、图像压缩等,因此对目标跟踪的研究具有非常重要的意义。论文围绕目标跟踪技术这个课题,针对摄像机运动的运动目标跟踪的难点问题,着重研究目标跟踪的初始化问题和如何提高目标跟踪性能的难题,提出新的理论和方法,并将这些方法应用于机器人的视觉导航领域,为目标跟踪及其应用提供理论和技术支持。论文主要做了以下四个方面的工作。(1)针对具有对称高斯核的均值偏移(mean shift)算法,对mean shift算法的表达形式,研究现状进行了研究,从理论上证明了mean shift算法的收敛性,分析了mean shift算法的迭代步长和收敛速度,而且证明了mean shift算法具有线性收敛速度,并提出了一种mean shift算法的加速方法。(2)在对概率框架下目标检测方法进行研究的基础上,提出一种最大后验概率的运动目标检测方法。根据条件随机场模型和马尔可夫随机场模型建立了一个最大后验概率框架。考虑到传统方法融入的特征信息不够,检测目标的准确度不高的缺陷,结合Gibbs势能模型和邻域信息,在概率模型中充分融入了各种时域、空域和边缘特征,以便消除更多噪声及提高检测运动目标的准确度。(3)通过对基于mean shift的目标跟踪方法的研究及影响跟踪性能因素的分析,考虑到目标模型中背景像素的影响,提出一种构造模糊核直方图的方法,并针对模糊核直方图,分别给出构造模糊隶属度函数的两种策略,即比率策略和差分策略,分别对这两种策略进行了比较和分析。其次考虑到巴氏系数Taylor公式近似展开的误差,通过对这种误差的分析,将巴氏系数的优化问题转化为求解约束优化问题,从而提出一种适合于快速运动目标的跟踪方法,而且证明了该方法的收敛性,分析了跟踪性能和相关参数的影响。(4)考虑到基于粒子滤波的目标跟踪方法的缺点,从提高粒子滤波的鲁棒性角度出发,通过建立运动信息的表达模型和自适应选择模型,提出一种自适应模型选择的粒子滤波方法。然后建立一个自适应选择模型,在跟踪精度不准确时使用mean shift算法对粒子进行优化,以改善粒子性能,另外根据粒子的质量来获取跟踪结果,从而达到提高目标跟踪性能的目的。最后通过机器人平台实验来验证所提出方法的有效性。更多还原

【Abstract】 For the technology of object tracking based on video can be used in fields of civilian and military affairs etc, such as video surveillance, robot navigation, human-computer interface and image compression, so it is quite important to study the technology of object tracking. Thus, this dissertation focuses on two key problems, namely initialization of object tracking and improvement on object tracking performance under the complex environments. Some new theories and methods are presented and are applied in field of robot navigation based on video. These fruits will contribute to further theoretical study and extensive application for technology of object tracking. Several aspects around these key problems are studied in this dissertation, and main contribution and work are described as follows:(1) For mean shift algorithm with Gaussian kernel function, its expression and its review is given and a convergence theorem with its rigorous proof is provided. Moreover, iterative step size along the gradient direction in mean shift is proved to be less than the optimum size and mean shift algorithm converges at a linear rate. Lastly, a fast mean shift (FMS) algorithm is presented, and the experimental results show that FMS can reduce the iteration number.(2) An approach based on maximum a posteriori is presented for moving object detection in complex video scenes. Firstly, maximum a posteriori framework is created according to conditional random field model and Markov random field model. For lack of feature information and low accuracy of detected object in traditional method, based on temporal Gibbs potential energy model and neighboring information, dependencies of consecutive label fields, spatial dependencies within each label field and edge features are merged into this framework by kinds of probability models.(3) By analysis on performance of object tracking based on mean shift, an object tracking method based on fuzzy kernel histogram is presented to reduce the localization error of object tracking. Two strategies are given for the fuzzy membership function to build the fuzzy kernel histogram. They are respectively ratio strategy and difference strategy. The experimental results of these two strategies are given and the advantage and disadvantage of these two strategies are discussed. Furthermore, errors of Bhattacharyya coefficient and its influence on object tracking are studied in this paper. Based the analysis on errors of Bhattacharyya coefficient, the optimization problem of Bhattacharyya coefficient is transformed into a constrained optimization problem, so an improved object tracking method is presented. In addition, the convergence of the improved object tracking method is proved.(4) For improving the robustness of particle filter, an adaptive particle filter is presented by creating the expression model of moving information. Furthermore, while tracking precision is low, mean shift algorithm is selectively used to optimize particles for improving the quality of particle. Tracking results is obtained according to the quality of particle.更多还原