联合混合范数约束和增量非负矩阵分解的目标跟踪  

Object Tracking via Jointing Mixed Norm Constrained and Incremental Non-negative Matrix Factorization

作　　者：陈芸[1,2] 董西伟[2] 荆晓远[2] 

机构地区：[1]江苏信息职业技术学院物联网工程系,江苏无锡214153 [2]南京邮电大学自动化学院,南京210003

出　　处：《计算机工程》2015年第12期260-264,共5页Computer Engineering

基　　金：国家自然科学基金资助项目(61272273;61170305);江苏省普通高校研究生科研创新计划基金资助项目(CXLX13_465);江苏省高等职业院校国内高级访问学者计划基金资助项目(2014FX034)

摘　　要：针对当前目标跟踪算法因光照变化、部分遮挡、姿态变化以及背景杂乱等因素引起的跟踪漂移问题,联合混合范数约束和增量非负矩阵分解,提出一种目标跟踪算法。通过对目标的非负矩阵分解获得其局部结构信息,有效应对局部遮挡,同时达到降维目的。通过稀疏描述下的混合范数约束进一步抑制外界环境的干扰,并利用加速近似梯度算法迭代求解优化问题。为更好地满足实时精准跟踪的需求,应用遮挡检测及在线更新策略读取跟踪目标位置。在粒子滤波跟踪框架中的实验结果显示,相比IVT、多示例学习、Frag和L1APG跟踪算法,该算法的鲁棒性更好。Aiming at the problems of the tracking drift tracking caused by illumination changes,partial occlusions,pose changes and background clutter and other factors,an object tracking algorithm of jointing mixed norm constraints and Incremental Non-negative Matrix Factorization（INMF）is presented.The local structure information of the target is gained by the non negative matrix factorization,which is effective to deal with the partial occlusion,and achieves the purpose of reducing the dimension.The interference of the external environment is further suppressed by the mixed norm of sparse description,and the optimization problem is solved iteratively by using the accelerated approximate gradient algorithm.In order to better meet the demand of realtime and accurate tracking,the occlusion detection and online update policy are presented to read the location of tracking object.The algorithm puts into particle filter tracking framework for robust target tracking algorithm is proposed.Experimental results show that the proposed algorithm performs favorably against IVT,multi-instance learning,Frag and L1 APG tracking algorithms.

关 键 词：增量非负矩阵分解 混合范数 稀疏表示 目标跟踪 粒子滤波 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]