融合视觉先验与背景信息的视频目标跟踪方法  

Video Object Tracking Via Visual Prior and Context Information

作　　者：管皓[1] 薛向阳[1] 安志勇[1,2] 

机构地区：[1]复旦大学计算机科学技术学院上海市智能信息处理重点实验室,上海200433 [2]山东工商学院计算机科学与技术学院,山东烟台264005

出　　处：《小型微型计算机系统》2016年第9期2074-2078,共5页Journal of Chinese Computer Systems

基　　金：上海市科委项目(15511104402)资助;山东省自然科学基金项目(ZR2013FL018)资助

摘　　要：针对传统目标跟踪算法特征表示单一、背景信息利用不足而造成跟踪鲁棒性较差的问题,提出一种融合视觉先验及背景信息的跟踪算法.首先通过离线非监督学习的方式获取一般性目标物体的视觉先验知识,以视觉字典的方式加以存储.在线跟踪时,利用视觉字典,通过卷积运算获取跟踪目标及其周围局部背景的外观特征,并通过相关滤波算法来预测位移变化.实验结果表明,本文算法能够有效对抗光照变化、局部遮挡等噪声干扰,具备较高的跟踪精度与鲁棒性.Due to the simplicity of features and lack of background information, traditional tracking algorithms are less robust to noi- ses. This paper proposes a tracking method which integrates visual prior and background information to tackle the problems. General visual prior is got through unsupervised learning off-line on large labeled image dataset and stored as visual prior dictionary. During on-line tracking, the learned visual prior can be used as convolutional filters to extract appearance features of the object being tracked and its local background. Correlation filter is used to predict the new location of the target. Experimental results show that the pro- posed tracking algorithm is both accurate and robust to deal with illumination changes and partial occlusions.

关 键 词：目标跟踪 视频分析 在线学习 计算机视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]