视频目标跟踪中的提升分类算法  

Boosting Classification Algorithm in Visual Object Tracking

作　　者：卞志国[1] 金立左[1] 费树岷[1] 

机构地区：[1]东南大学自动化学院,南京210096

出　　处：《数据采集与处理》2010年第6期683-688,共6页Journal of Data Acquisition and Processing

基　　金：航空科学基金(20080169003)资助项目;国家自然科学基金(60805002)资助项目

摘　　要：针对目标跟踪过程中各类图像特征分离背景和目标能力的动态变化,提出一种基于协同训练框架的在线提升分类特征选择算法。该算法采用两组特征描述目标与背景区域各像素,并分别训练一在线提升分类器对特征组进行选择,然后综合分类结果,得到最优似然图像,基于该似然图像,采用粒子滤波对目标进行跟踪并通过图像处理方法获得最佳前景分割图。该方法的主要优点是仅需对首帧图像进行训练,并在跟踪过程中通过协同训练在线更新提升分类器。同时,实验表明该算法运算速度快,并能适应环境光照变化、遮挡等恶劣条件。The ability of many kinds of features to separate the object from background is varying during tracking process,a real boosting with co-training based feature selecting is presented.Two groups of features are introduced to characterize each pixel as target or background,and online real boosting is trained to select features and classify each pixel in search area.Based on the integration of classification outcome,the particle filter is applied for tracking.Finally,the accurate state of the object is acquired by the image processing.Examples based on optical frames show that the proposed tracking framework is real-time and suitable to the partial occlusions,view-point and illumination variations.

关 键 词：目标跟踪 提升分类 特征选择 粒子滤波 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]