北京航空航天大学学报 ›› 2017, Vol. 43 ›› Issue (7): 1364-1372.doi: 10.13700/j.bh.1001-5965.2016.0527

• 论文 • 上一篇    下一篇

基于目标性权值度量的多示例学习目标跟踪

滑维鑫1,2, 慕德俊1, 郭达伟1, 刘航1   

  1. 1. 西北工业大学 自动化学院, 西安 710072;
    2. 中国移动通信集团陕西有限公司, 西安 710074
  • 收稿日期:2016-06-20 修回日期:2016-09-01 出版日期:2017-07-20 发布日期:2016-10-17
  • 通讯作者: 慕德俊,E-mail:mudejun@nwpu.edu.cn E-mail:mudejun@nwpu.edu.cn
  • 作者简介:滑维鑫 男,博士研究生。主要研究方向:目标检测与跟踪、多目标优化;慕德俊 男,博士,教授。主要研究方向:控制理论与应用、网络信息安全。
  • 基金资助:
    国家自然科学基金(61303224,61672433)

Visual object tracking based on objectness measure with multiple instance learning

HUA Weixin1,2, MU Dejun1, GUO Dawei1, LIU Hang1   

  1. 1. College of Automation, Northwestern Polytechnical University, Xi'an 710072;
    2. Shaanxi Company, China Mobile Communications Corporation, Xi'an 710074, China
  • Received:2016-06-20 Revised:2016-09-01 Online:2017-07-20 Published:2016-10-17
  • Supported by:
    National Natural Science Foundation of China (61303224, 61672433)

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。

关键词: 多示例学习(MIL), 目标性测量, 弱分类器选择, 包概率计算, 目标跟踪

Abstract: For the problems that the multiple instance learning (MIL) tracking algorithm does not distinguish the differences of each sample when computing the bag probability and selects the weak classifiers by maximizing the log likelihood function, which reduce the performance of classifier and increase the complexity of the algorithm, this paper proposes a tracking algorithm based on objectness weighted multiple instance learning. First, the importance of each sample is measured by the objectness, which is also used to assign the weight for each instance. Then the weighted value is utilized for computing the final bag probability. In the phase of weak classifier selection, a maximized inner product between weak classifier and log likelihood function is adopted to select weak classifiers from weak classifier pool, and then these weak classifiers are combined into a strong classifier. All these strategies are beneficial for improving the tracking accuracy and reducing the computational complexity. By tracking the video sequences under different complex scenes, experimental results show that the proposed algorithm has strong robustness and high tracking accuracy compared with competing method.

Key words: multiple instance learning (MIL), objectness measure, weak classifier selection, bag probability calculation, object tracking

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发