基于集成多示例学习的Mean Shift跟踪算法  

Mean Shift Tracking Based on Ensemble Multiple Instance Learning

作　　者：罗会兰[1] 单顺勇 孔繁胜[2] 

机构地区：[1]江西理工大学信息工程学院,赣州341000 [2]浙江大学计算机科学与技术学院,杭州310027

出　　处：《计算机辅助设计与图形学学报》2015年第2期226-237,共12页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"九七三"重点基础研究发展计划项目(2010CB327900);国家自然科学基金(61105042);江西省教育厅科技项目(GJJ13421)

摘　　要：为了实现长时间稳定的对特定目标的跟踪,结合匹配型跟踪方法和决策型跟踪方法的优势,同时利用集成学习的思想构建多个强分类器,提出一种基于集成多示例学习的mean shift跟踪算法.首先在上一帧中对示例进行随机采样,构建分类器的集体,通过集成学习合成最终的分类器以确定当前帧中目标的初始位置;然后对初始位置和上一帧目标最终位置的距离与设定的阈值进行判断,决定是否采用mean shift跟踪算法对初始位置进行修订,以确定目标的最终位置.实验结果表明,该算法不但可以应对目标的形变、旋转、遮挡以及光照变化等各种复杂的情况,而且可以做到长时间的跟踪,具有较强的鲁棒性.An effective object tracking method was proposed by combining multiple instance learning and meanshift tracking. The motivation is to use the advantages of the generative model and discriminative model and useensemble learning to gain more robust tracking effect. First, instances are randomly selected to train differentclassifiers in the previous frame, and then the final integrated classifier was trained using ensemble learning toimprove the tracking accuracy. The initial position of the object was determined by using ensemble multipleinstance learning. Then mean shift tracking was used to revise the initial position by comparing the distancebetween the initial position and the object position in the previous frame to a threshold. The experimental resultshas shown that the proposed algorithm has good performance in many complicated situations, e.g. pose change,rotation, occlusion and changes of illumination, and can track successfully for a long time with strong robustness.

关 键 词：集成学习 多示例学习 meanshift跟踪 目标跟踪 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]