基于带权重多样例学习的视觉跟踪算法

摘要

原始在线加权的多示例学习跟踪假设每个示例是独立且在包中的贡献均相同,同时为所有正样本赋予相同的权重,这不符合“包中的示例与目标位置的远近,对目标贡献程度是不一样”的事实.再加上原始算法采取单一特征无法准确和全面地表示目标包中所包含的示例,从而影响了跟踪算法的鲁棒性.针对原始算法的这些问题,提出一种基于带权重多样例学习的视觉跟踪方法.该方法同时融合多特征(HOG特征和Haar特征),在多示例学习框架下同时训练分类器,并通过样本特征相似度的比较来赋予不同的权重.对不同场景的图像序列进行实验,通过在公共测试集上与多种主流算法做对比,显示这样得到的目标外表模型对于前景和背景具有更高的区分能力.结果表明:该算法具有更高的准确性和更强的适应性,可以有效克服传统多示例学习中的分类器退化问题.%Original online weighted sample learning tracking assumes that each sample is more independent and its contribution to the package is all the same,and the same weight is given for all the positive samples.Therefore,it does not agree with the fact that the contribution degree of the target is not all the same with the distance between the target position and the sample of package.Additionally,the original algorithm cannot accurately and comprehensively represent the sample of the target package because of single feature,thus,it affects the robustness of the algorithm.Regarding the problem of original algorithm,this paper put forward a type of visual tracking method based on weighted sample learning.The method fused HOG features and Haar features at the same time,trained a classifier under the framework of learning,and gave different weights according to the similarity of the sample features.The experiments were conducted on the image sequence of different scenarios,and compared with a variety of current mainstream algorithms.Results show that the created appearance model has higher ability to distinguish between foreground and background,the algorithm has higher accuracy and stronger adaptability and can effectively overcome the traditional sample classifier degradation problems in learning.