摘要
为解决目标跟踪中的光照、位姿变化及遮挡问题,提出分块多示例学习算法。该算法将目标图像分块,对每块图像片应用多示例学习算法。跟踪过程中结合样本所有图像片的综合分类器分数和每块图像片的分类器分数检测并判断跟踪过程中的光照、位姿变化,及遮挡问题。针对不同的跟踪状态,采用分类器学习率自适应调整策略,以避免"过更新"及更新不及时的问题。最后将所提方法在典型视频序列中进行验证,并与其他多示例学习算法进行对比实验。实验结果表明所提方法跟踪性能稳定,实时性强,解决了跟踪过程中的光照、位姿变化及遮挡等问题。
To deal with the problems of variations on illumination or pose,and serious partial occlusion,patch based multiple instance learning(MIL) algorithm was proposed.The algorithm divided an object into many blocks.Then,the online MIL algorithm was applied on each block.In the tracking process,the problems of variations on illumination or pose and partial occlusion problem were measured and distinguished by the global classification score and local classification score of each block.Furthermore,the learning rate was tuned for each case to avoid the problems of over-updating and less-updating.Finally,the proposed method was compared with other state-the-art algorithms on several classical videos.The experiment results illustrated that the proposed method performed well especially in case of variations on illumination or pose,and partial occlusion,and realized real-time object tracking.
作者
张华
杨岑玉
ZHANG Hua;YANG Cen-yu(Department of Electrical and Electronic Engineering,Shijiazhuang University of Applied Technology,Shijiazhuang 050081, China;Global Energy Interconnection Research Institute,Beijing 102209,China)
出处
《控制工程》
CSCD
北大核心
2019年第1期23-28,共6页
Control Engineering of China
基金
河北省科技厅科技支撑项目(18210329D)
关键词
目标跟踪
遮挡
分块多示例学习算法
强分类器
Object tracking
occlusion
patch based multiple instance learning
strong classifier
