基于核协同表示的快速目标跟踪算法  

Fast object tracking algorithm via kernel collaborative presentation

作　　者：王海军[1,2] 葛红娟[1] 张圣燕[2] 

机构地区：[1]南京航空航天大学民航学院,江苏南京211106 [2]滨州学院山东省高校航空信息技术重点实验室,山东滨州256603

出　　处：《浙江大学学报（工学版）》2017年第2期399-407,共9页Journal of Zhejiang University：Engineering Science

基　　金：山东省自然科学基金资助项目(ZR2015FL009);山东省重点研发计划资助项目(2016GGX101023);山东省高等学校科技计划资助项目(J16LN02);滨州市科技发展计划资助项目(2013ZC0103);滨州学院科研基金资助项目(BZXYG1524;BZXYG1318)

摘　　要：针对复杂环境下的单目标跟踪问题,提出一种采用核协同表示的快速目标跟踪算法.在粒子滤波的框架下,在字典矩阵中引入方块矩阵建模跟踪过程中可能出现的遮挡,然后将字典矩阵和候选样本分别映射到高维空间,建立候选样本和字典矩阵在高维空间的线性表示目标跟踪模型,同时采用L2正则化减弱传统方法对系数稀疏性的要求,有效地降低关键步骤的计算复杂度.实验结果表明,该方法能够克服遮挡、光照变化、尺度变化、运动模糊等影响跟踪性能的因素,具有较高的平均覆盖率和较低平均中心点误差,能够实现快速鲁棒的跟踪.A fast object tracking algorithm via kernel collaborative presentation was proposed to solve the problem of single object tracking in complex scenes.In the framework of particle filter,square matrix was introduced into the dictionary matrix to model the occlusion appeared in the object tracking.Then the dictionary matrix and candidates were mapped to the high dimensional space,separately;the linear representation model between the candidates and dictionary matrix in high space was established.L2 regularization was used to reduce the requirement of the traditional method for the sparsity of coefficients,which can effectively reduce the computational complexity of the key steps.Experimental results show that the proposed algorithm can effectively overcome the influence of occlusion,illumination change,scale change and motion blur,with higher average overlap rate and lower average center location error.

关 键 词：目标跟踪 稀疏表示 协同表示 字典矩阵 方块矩阵 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]