【作者】 张楠；

【导师】 蔡念；

【作者基本信息】 广东工业大学 ， 信号与信息处理， 2011， 硕士

【摘要】 目标跟踪是计算机视觉领域的前沿研究课题,其本质是在图像序列中通过递推估计来确定感兴趣的具有某种显著特征(如颜色、形状、纹理、运动等)的目标位置。随着图像处理技术的飞速发展和计算机性能的不断提高,目标跟踪技术受到人们的青睐,在视频监控、视频检索、人机交互、交通监控、医疗诊断、机器人视觉导航、虚拟现实和成像制导等方面得到了广泛的应用。在众多目标跟踪算法中,基于粒子滤波的目标跟踪算法是近年来比较流行的方法。粒子滤波是处理非高斯、非线性状态估计问题的有力理论工具,也是构架目标跟踪算法的完备理论框架。该算法基于贝叶斯统计计算,首先将跟踪问题转化为非线性非高斯的贝叶斯滤波问题,然后通过对滤波问题的求解来实现对目标的跟踪。空间直方图融合了目标的颜色信息和颜色的空间分布信息,比传统的颜色直方图更具有目标鉴别能力。本文在基于粒子滤波算法的目标跟踪系统框架中,将加权样本集表示目标的状态后验概率分布；采用简单的随机漂移模型表示系统状态模型；利用目标区域的空间直方图描述目标,其中通过核概率密度估计建立目标的颜色分布,然后统计颜色分布的空间信息建立空间直方图；通过空间直方图的相似度定义来建立系统观测概率模型,最终提出一种基于空间直方图的粒子滤波目标跟踪算法。海杂波背景下的红外目标,易受到海杂波的影响,从而降低了目标检测能力,进而对目标的稳健跟踪造成很大干扰。本文将基于空间直方图的粒子滤波目标跟踪算法应用于海杂波背景下的红外目标跟踪。本文研究的算法融合了空间信息,相对于传统的灰度直方图,采用的相似度定义更加严格,因此能对海杂波背景下的红外目标进行稳健跟踪,并有效解决遮挡问题。实验结果表明,相比传统的基于灰度直方图的粒子滤波目标跟踪算法,本文提出的算法具有明显的优势,对于部分遮挡以及全遮挡等复杂的海杂波背景下红外目标跟踪具有更好的鲁棒性。更多还原

【Abstract】 Target tracking is becoming an active research topic in the areas of computer vision. The essence of target tracking is interactively searching in image sequences to validate the location of a certain object with some salient visual features (such as color, shape, texture, motion). With the rapid development of image processing and the improvement of computer, the technology of target tracking, has been widely used in video surveillance, video retrieval, human-computer interaction, traffic control, medical diagnosis, robot navigation, virtual reality and image guidance and so on.The algorithm of target tracking based on particle filter is more popular among many algorithms in recent years. Particle filter is one of the powerful tools for non-Gaussian/nonlinear state estimation problem, which is also a complete theoretical framework for target tracking. The algorithm is based on Bayesian statistical models. The tracking problem is converted into the non-Gaussian and nonlinear Bayesian filtering problem. Then target tracking can be implemented by solving the filtering problem.Spatiograms outperform the traditional color histogram, which contain color information and spatial layout of these colors for the target. In particle filters framework, the posterior distribution of the target is approximated by a set of weighted samples and a normal random drift model is utilized to describe the state model. The targets are represented by the spatiograms, which is defined by the spatial information of the color distribution estimated by kernel-based density. A similarity function of spatiograms is used as the probability model for observation. Finally, we propose a target tracking algorithm using particle filters based on spatiograms.The infrared target in sea clutter background is difficult to be tracked robustly with the target detection reduced, which is vulnerable to the impact of sea clutter. In this thesis, the target tracking algorithm using particle filters based on spatiograms is applied to infrared target tracking in sea clutter background. The algorithm which refers to spatial information using a more stringent definition of similarity than the traditional particle filters based on the gray histogram can realize to tack the infrared target in sea clutter robustly and solve the problem of occlusion effectively. Experimental results indicate that the proposed algorithm outperforms the traditional particle filters based on the gray histogram, and has better robustness under local occlusion and global occlusion of infrared target tracking in sea clutter background.更多还原