多支持向量机在线联合的运动目标跟踪算法  

Moving object tracking algorithm based on on-line ensemble SVMs

作　　者：张谢华[1,2] 张申[1] 田敏[3] 

机构地区：[1]中国矿业大学信息与电气工程学院,江苏徐州221008 [2]徐州师范大学现代教育技术中心,江苏徐州221116 [3]同济大学电子与信息工程学院宽带无线通信与多媒体实验室,上海201804

出　　处：《计算机工程与应用》2009年第36期180-184,共5页Computer Engineering and Applications

摘　　要：依据二元分类的思想,提出了一种新的基于多支持向量机在线联合的运动目标跟踪算法。首先选择线性支持向量机作为分类器最大限度地将目标和背景区分开来,对线性支持向量机进行简单高效的在线更新,采用支持向量自动记录运动目标"关键帧"的信息。然后通过Adaboost算法为每个线性支持向量机分别赋以不同的权重,进行在线联合获得强分类器。实验结果表明,该算法具有较强的鲁棒性,尤其在目标变化过于激烈的情况下能够实现较为稳定的跟踪。Considering the tracker as a binary classification problem,a novel moving object tracking algorithm is presented,which is based on on-line ensemble SVMs.First of all,linear SVMs is choosed as the classifiers to distinguish the target from the background.A simple yet effective way is used for on-line updating linear SVMs,where useful ＂Key Frames＂of target are automatically selected as support vectors.Then,each linear SVM is separately given different weight through the Adaboost algorithm and is on-line ensembled to get a strong classifier.Experimental results show the robustness of the proposed algorithm,especially under large appearance change during tracking.

关 键 词：运动目标跟踪 线性支持向量机 在线更新 支持向量 ADABOOST 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]