基于目标性权值度量的多示例学习目标跟踪

发布时间：2018-05-16 08:31

本文选题：多示例学习(MIL) + 目标性测量　；参考：《北京航空航天大学学报》2017年07期

【摘要】：针对多示例学习(MIL)跟踪算法在包概率计算过程中对示例样本不加以区分导致分类器性能下降,及采用最大化似然函数选择相应的弱分类构造强分类增加了算法复杂度的问题,提出了一种基于目标性权值学习的多示例目标跟踪算法,该算法利用目标性测量每个示例样本对包概率的重要性,根据其目标性测量结果对每个正示例样本赋予相应的权值,从而判别性地计算包概率,提高跟踪精度。同时在弱分类器选择过程中,采用最大化弱分类器与似然函数概率内积的方法从弱分类器池中选择弱分器构造强分类器,减少算法的计算复杂度。通过对不同复杂场景下视频序列的跟踪,实验结果表明,本文所提出的目标性权值学习的多示例目标跟踪算法优于其对比算法,表现出较好的跟踪精度和鲁棒性能。
[Abstract]:For multi-example learning MIL) tracking algorithm, the lack of discrimination of sample samples in the calculation of packet probability results in the performance degradation of classifier, and the use of maximum likelihood function to select the corresponding weak classification to construct strong classification increases the complexity of the algorithm. In this paper, a multi-example target tracking algorithm based on objectivity-weight learning is proposed. The algorithm uses objectiveness to measure the importance of each sample to packet probability, and assigns corresponding weights to each positive sample according to its target measurement results. Thus the packet probability is calculated discriminatively and the tracking accuracy is improved. At the same time, in the process of weak classifier selection, the method of maximizing the probability inner product of weak classifier and likelihood function is used to construct a strong classifier from the weak classifier pool to reduce the computational complexity of the algorithm. By tracking video sequences in different complex scenes, the experimental results show that the target weight learning multi-example target tracking algorithm proposed in this paper is superior to its contrast algorithm, and shows better tracking accuracy and robust performance.
【作者单位】：西北工业大学自动化学院;中国移动通信集团陕西有限公司;
【基金】：国家自然科学基金(61303224,61672433)~~
【分类号】：TP391.41

【参考文献】