摘要： 多示例学习是一种处理包分类问题的新型学习模式,传统基于多示例学习的目标跟踪算法在自适应获取正包时受到无益或有害示例的干扰,不能很好地提取目标的鉴别性特征。为此,设计基于核密度估计的示例选择方法,剔除训练集中的无益示例或有害示例,提高多示例学习算法的有效性,并在此基础上提出一种基于示例选择的目标跟踪改进算法,针对负示例占多数的情况建立核密度估计函数来精简正包中的示例,使用精简后的样本数据进行训练学习,最终实现对目标的实时跟踪。实验结果表明,该算法在光照变化、目标部分遮挡及形体变化等情形下都具有较好的稳健性。

关键词: 多示例学习, 有害示例, 核密度估计, 示例选择, 稳健性, 目标跟踪

Abstract: Multiple Instance Learning(MIL) is a new learning paradigm that deals with the problem of bags classification.The object tracking algorithm based on MIL is often interfered by useless or harmful instances when adaptively choose positive bags,and can not extract good discriminative feature,and this will affect the robustness of tracking algorithm.Therefore,this paper proposes an instance selection method based on kernel density estimation to improve the efficiency of MIL algorithm,which eliminates useless or harmful instances in training set,and then on this basis it provides a robust object tracking algorithm based on instance selection.Firstly,the algorithm builds up kernel density estimation function,uses it to optimize instances of positive bag for the majority of them are negative instances,and then uses the optimized training data to training-and-learning,and finally achieves real-time object tracking.Experimental results demonstrate that when the object suffers from the illumination changes,object partial occlusions,appearance changes and some other situations,the proposed object tracking algorithm is more robust.

Key words: Multiple Instance Learning(MIL), harmful instance, kernel density estimation, instance selection, robustness, object tracking

中图分类号: 

• TP181