【作者】 卢晓鹏；

【导师】 邹谋炎；

【作者基本信息】 中国科学院研究生院（电子学研究所） ， 信号与信息处理， 2007， 博士

【摘要】 视频目标跟踪研究是视觉领域的一个重要课题，受到国内外学者的普遍关注，具有广泛的应用前景。本论文致力于视频序列中目标跟踪技术的研究，创新点有以下几个方面：1．针对传统质心位移算法使用单一颜色直方图作为特征描述目标，在跟踪时不能很好地适应目标在相邻帧间移动缓慢和目标与邻域颜色特征相似的情况。本文提出了基于核密度估计相关度量的目标跟踪方法，对参考目标模板和当前帧目标模板进行描述时，在诸如颜色、梯度等目标像点的特征区间内融入了目标像点的空间位置信息。与基于单一颜色直方图所得Bhattacharyya系数比较，本文提出的核密度相关度量的收敛梯度更陡，算法进行目标跟踪时速度更快更稳健，对目标跟踪的位置更加精确。2．提出了基于高斯混合模型的类EM跟踪方法。算法选用GMM对目标区域内的像点参数建模，以多个高斯分布函数的加权和来逼近目标像点的二维空间位置分布概率函数。同时，将模型中的各个高斯函数的权值与基于颜色特征的目标相似度量联系起来，将目标跟踪问题转化为视频帧之间目标参数的EM估计。每次迭代中，估计目标位置的同时可以得到目标大小的协方差矩阵参数，有效地解决了目标区域的自适应。3．质心位移算法具有简单快速的优点，但是在目标局部或全部被遮挡的情况下，往往无法得到满意的跟踪效果；与此同时，粒子滤波算法虽具有抗遮挡的优点，但计算过程中需要大量的粒子样本来近似描述目标的状态，使得算法非常费时。混合跟踪算法引入自适应粒子样本数选择的采样策略，保证跟踪精度同时减少了跟踪总体时间花费，有效提高了系统的实时性，更好满足工程实用。4．多目标视觉目标跟踪相比于单目标跟踪问题，充满了背景嘈杂、目标与目标之间以及目标与背景之间遮挡等许多不确定因素，故而面临着更多困难。本文提出了基于颜色分布的联合多目标概率分布粒子滤波器，以运动学先验概率作为重要性概率密度函数，根据粒子样本中代表不同目标的划分之间的欧式距离判定划分之间是否关联，并据此在独立划分和关联划分两种粒子滤波器之间进行切换。仿真和实际实验证明，该滤波器能够同时有效地对多个目标进行跟踪。更多还原

【Abstract】 Tracking visual objects in image sequences is an important topic in the field ofcomputer vision. This problem has received wide attention and it has a wide range ofapplication in different fields. In this dissertation, we discuss the problem of objecttracking in image sequences. The innovative points of the dissertation can be listed asfollows.In the traditional mean shift algorithm, only the color histogram is used fordescribing the feature of an object. The dissimilarity among the reference targets andthe target candidates is expressed by the metric derived from the Bhattacharyyacoefficients. The traditional mean shift procedure is used to find the real location ofthe object through looking for the regional minimum of the distance functioniteratively. However, in the case if the object moves slowly from frame to frame in theconsecutive image sequence or if the color histogram of the target is similar to thecolor histogram of the neighborhood, the algorithm does not work well. To overcomethe difficulty, a concept on the correlation of kernel density estimation of trackingregion and a new generalized distance is proposed. Experiments manifest thataccuracy and robustness of the tracking algorithm can be improved.We suggest an EM-like algorithm for object tracking based on the Gaussianmixture model (GMM). The algorithm uses the weighted summation of several Gaussfunctions to approach the probability function of position’s distribution in twodimensional space. In the meanwhile, the GMM is related with the similarity measureof targets on the basis of their color features. Then, by converting the object trackingalgorithm into the EM estimation of object’s parameters from frame to frame, the algorithm estimates the location of the object and updates the variance matrix,simultaneously, in each iterative step. Compared with the traditional Mean shiftalgorithm, the new algorithm presented in this paper can track the object more rapidlyand accurately, and update the object window adaptively.The traditional mean shift method does not work well when the target gets anocclusion. The tracking algorithm on the basis of particle filtering has the ability tooverpass the occlusion. Unfortunately, its performance relies heavily on the number ofthe used particles. This involves a large number of computations and therefore isdifficult to be implemented in real time. To settle the problem, this article brings abouta hybrid algorithm by combining the mean shift and the particle filter trackingtechnique on the basis of the color histogram distribution. By adopting the strategythat the number of particles is adaptively determined, it amalgamates the virtues of thetwo techniques. As a result, the computational cost can be reduced and the trackingperformance can be ensured simultaneously. The experimental results show that theproposed method is effective and robust.This article proposes the joint multi-target probability density particle filteringbased on the color distribution to address the problem of tracking multiple movingtargets with difficult conditions such as target crossing, target occlusion andbackground confusion. Here we use the kinematic prior as the importance probabilityfunction. Through comparing the Euclidian distance of the partitions that denotingdifferent targets, we can switch the procedure from the coupled partition particlefiltering to the independent partition particle filtering and inversely. The simulationand experimental results show that the proposed algorithm can track multiple objectseffectively.更多还原