基于l_(1/2)正则化的稀疏表示目标跟踪算法的研究  

Research on Object Tracking Algorithm of Sparse Representation Based on l_(1/2) Normalization

作　　者：贲敏[1] 邓萍[1] 王保云[1] 

机构地区：[1]南京邮电大学自动化学院,江苏南京210023

出　　处：《计算机技术与发展》2015年第1期82-86,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61271232)

摘　　要：近年来目标的稀疏表示已经广泛应用到视频跟踪中。文中提出了一种基于局部稀疏表示的鲁棒目标跟踪算法,目标的表示将局部信息考虑在内,并且做出了遮挡处理。为了在新的帧中跟踪到目标,每一个候选目标通过在线构建的过完备字典以及模板解l1/2最小化问题稀疏表示。文中用l1/2规范最小化来代替l0,而不是用l1规范最小化,通过解l1/2最小化问题,可以找到比解l1最小化更稀疏、更精确的解。此外,l1/2比l0更容易求解。目标稀疏表示后,通过在线学习的分类器将目标区分出来。实验结果表明,与现有的一些算法相比,该算法稳定性好,性能更优越。Recently sparse representation has been widely used in video tracking. In this paper,propose a robust target tracking method based on local sparse representation,considering the local information for object representation and take occlusion into account. In order to track the target in a new frame,each target candidate is sparsely represented by over-complete dictionary online constructed and target templates solving a l1/2-norm minimization problem. In this algorithm,use l1/2-norm minimization to replace l0-norm minimization in-stead of l1-norm minimization. By solving l1/2-norm minimization,can find a sparser and more accurate solution than l1-norm minimi-zation,moreover,it is much easier to be solved than l0-norm minimization. After that,a classifier is learned to distinguish the target from the background. Experimental results show that this method has good stability and the performance is superior to the current algorithms.

关 键 词：视频跟踪 稀疏表示 过完备字典 l1/2最小化 分类器 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]