一种结合粒子滤波和张量子空间的目标跟踪算法  

An Adaptive Object Tracking Algorithm Combined with Particle Filter and Tensor Subspace

作　　者：温静[1] 李洁[1] 高新波[1] 

机构地区：[1]西安电子科技大学电子工程学院,西安710071

出　　处：《光子学报》2010年第6期1047-1052,共6页Acta Photonica Sinica

基　　金：国家自然科学基金(60702061);教育部长江学者和创新团队支持计划(IRT0645);深圳大学ATR国防科技重点实验室开放基金;总装备部预研基金(9140A06050107DZ0113)资助

摘　　要：由于传统的子空间方法易于丢失图像目标的二维特性,为此本文提出了一种新颖的自适应目标跟踪算法,通过张量的方式建立目标的外观模型——张量子空间,利用在线学习的方法更新其外观模型,同时,利用目标仿射运动的先验信息,通过粒子滤波自适应地跟踪运动目标,并将获得的最优目标观测作为新数据反馈回子空间更新.此外,为了保证子空间更新能获得精确且紧致的目标子空间表达,引入动态部分函数滤除样本野点.实验结果表明,本文提出的自适应目标跟踪方法具有较强的鲁棒性,对于存在姿态变化、短时遮挡和光照变化等情况下均可有效地跟踪目标.In order to overcome the disadvantage that traditional subspace methods usually lose the two- dimensional information of the objects in image, a novel adaptive object tracking method is proposed. The appearance of the object in tensor subspace is modeled and the object model is updated with online learning method. The object is tracked by using particle filter and the prior of affine motion, and the optimal observation is feeded back to the tensor subspace updating. Moreover,DPF is introduced into the subspace updating to reject outliers so as to keep the object subspace precise and compact. The proposed method is able to track targets effectively and robustly under pose variation, short-time occlusion and large lighting and so on in the experiments.

关 键 词：模式识别 目标跟踪 张量子空间 在线学习 粒子滤波 仿射运动 DPF 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]