【作者】 刘崇文；

【导师】 何中市；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 本文介绍了作者对运动目标跟踪中的一些算法的实现和研究,主要包括粒子群优化与粒子滤波运动目标跟踪方法的结合,仿射坐标在特征提取中的应用,分段仿射模型在双侧学习跟踪方法的应用,使用聚类方法搜索目标等。本论文的目的是研究寻找简单高效的运动目标跟踪算法和设计适用的体系结构,让计算理论、算法和系统相集合,使理论成果走向实际应用。本论文的主要内容列举如下:本文按照运动目标跟踪方法的处理路线来介绍运动目标跟踪方法的各个模块。首先介绍了图像预处理的相关工作和特征提取与选择;接着是本文的中心工作内容,提出了两个运动目标跟踪算法,然后通过实验对本文提出的算法和其他同类算法进行比较。本文中心内容列举如下:第一,根据粒子滤波目标跟踪方法,结合粒子群优化的思想,研究提出了一种高效的运动目标跟踪方法。该方法能降低基于粒子滤波的运动目标跟踪方法在迭代一定次数后会出现粒子权重聚集的粒子退化速度。因为粒子聚集后需要重采样来分散粒子权重,会使得学习间断或采样与跟踪目标的实际特征出现偏差。实验结果表明本文的方法与粒子滤波方法相比,在目标跟踪过程中的重采样次数更少,目标跟踪效果更好。第二,根据双侧学习运动目标跟踪方法,结合分段仿射目标跟踪方法中仿射变换参数的思想,提出了一种新的目标跟踪方法,实验表明改进后的算法有较好的效果。在提取特征的时侯,引入仿射坐标参数,使得新特征能很好的应对目标旋转的情况。通过重新定义双侧学习方法中的均值和协方差等重要参数,对双侧学习的目标跟踪方法进行了改进,使这些参数适合运动目标跟踪。第三,在实验中发现,使用聚类方法寻找目标侯选区域,与同类方法相比时间开销较低且效果更好。分段仿射双侧学习目标跟踪方法中不足是寻找目标位置需要逐个对区域进行匹配,搜索速度慢,而采用聚类的方法把分割区域聚集为侯选区域,处理速度更快。采用聚类方法在一定程度上平衡了跟踪效果和程序运行速度,总的看来在大部分情况下可以减少运算次数且获得较好的跟踪结果。在本文的最后,设计实验验证了本文所提出的方法的方法。实验结果表明,本文所提出的算法效果较好,具有一定的应用前景。更多还原

【Abstract】 This paper introduces some implementations and research on moving target tracking algorithms, including the particle filtering method based on particle swarm optimization, apply affine coordinates in the feature extraction and the piecewise affine method in bilateral learning moving targets tracking method, using clustering method to search the target. The purpose of this paper is to search one simple and efficient algorithm and design applicable architecture, combine the theory, the algorithm and the architecture, and change the theoretical results to practical application.This paper follows the processing of the moving target tracking method to introduce each module of moving target tracking method. First it introduces the related works about image preprocessing and feature extraction and selection followed by the main tasks of this paper and certain contents, proposing two moving target tracking algorithms, then doing experiment on the proposed algorithms and other algorithms for comparison. The central elements are listed below.First, based on particle filter target tracking, combined with the idea of particle swarm optimization, study and propose an efficient moving object tracking method. This method can reduce the weight degradation rate of particle aggregation which occurs in the particle filter moving target tracking method after a certain number of iterations. Because of it needs to re-sample weights to spread the particles when the weight of particles move together, which makes a learning interrupt and sampling deviation from the objectives actual characteristics. The experimental results show that compared with the particle filter moving target tracking method, the method proposed in this paper can reduce frequency of re-sampling, and tracking better.Second, based on bilateral learning moving target tracking method, combined with the thinking of parameter affine transform appear in piecewise affine moving target tracking method, bring one new moving target tracking method, and the experiments show the new method has better results. Introducing affine transformation parameters in future extraction, makes the new features can take a good deal of the situation of target rotation. By redefining the bilateral learning methods, the mean and covariance, and other important parameters, improvement is made on the bilateral learning moving target tracking method, and makes these parameters more suitable to moving target tracking. Third, in experiments, compard with other methods, the clustering method shows time complexity low and better experimental results in the search target area. Using piecewise affine bilateral learning moving traget tracking method to find the target location needs to match the region one by one, searching slowly, but using clustering method to collect segmental areas to candidate areas, it process faster. Clustering method balances the method’s performance and processing speed, totally to obtain better tracking results by reducing operating time in most situations.In the last part of this paper, experiments are designed to verify the methods and algorithms proposed in this paper. Experiment results show that the proposed methods and algorithms are better, and have good prospects.更多还原