检测识别跟踪分离的在线多样本视频目标跟踪  

Online Multiple Sample Video Target Tracking with Separation of Detection Recognition and Tracking

作　　者：刘国营[1] 陈秀宏[1] 庄甘霖[1] 

机构地区：[1]江南大学数字媒体学院,江苏无锡214122

出　　处：《计算机工程》2012年第23期194-197,共4页Computer Engineering

基　　金：中央高校基本科研业务费专项基金资助项目(JUSRP211A70)

摘　　要：在视频目标跟踪中,经常出现"漂移"现象,且学习算法需要离线训练。为此,提出一种检测识别跟踪分离的在线多样本视频目标跟踪方法。利用多样本学习方法解决学习算法在更新过程中的内在不确定性,使用优于在线半监督的boosting方法解决"漂移"现象。实验结果表明,该方法鲁棒性较好,可以有效解决"漂移"现象,并能实时地完成在线跟踪。There is the drift phenomenon in video target tracking,and learning algorithm needs offline training.In order to solve this problem,a online multiple sample video target tracking method for separation of detection,recognition and tracking is proposed.In this method,multiple sample approach tries to resolve the inherent ambiguities of many practical learning in the process of update instance.And the beyond semi-supervised learning tends to settle the ＂drifting＂ problem.Experimental results show that this method is more robust,can effectively solve the ＂drifting＂ problem,and track the target real-time online.

关 键 词：多样本 分类器 在线学习 目标跟踪 BOOSTING方法 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]