作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2013, Vol. 39 ›› Issue (3): 213-217,222. doi: 10.3969/j.issn.1000-3428.2013.03.042

• 人工智能及识别技术 • 上一篇    下一篇

基于稀疏表达的多示例学习目标追踪算法

苏巧平1,刘 原2,卜英乔3,黄 河4   

  1. (1. 安徽新华学院电子通信工程学院,合肥 230088;2. 安徽医学高等专科学校医学技术系,合肥 230000; 3. 解放军电子工程学院,合肥 230000;4. 中国科学院合肥智能机械研究所,合肥 230000)
  • 收稿日期:2012-07-06 出版日期:2013-03-15 发布日期:2013-03-13
  • 作者简介:苏巧平(1980-),女,讲师、硕士,主研方向:目标跟踪算法,图像处理;刘 原,讲师、硕士;卜英乔,助理研究员、硕士;黄 河,博士
  • 基金资助:
    国家自然科学基金资助项目(31171456);安徽省教育厅自然科学基金资助项目(kj2011z156)

Multiple Instance Learning Target Tracking Algorithm Based on Sparse Representation

SU Qiao-ping 1, LIU Yuan 2, BO Ying-qiao 3, HUANG He 4   

  1. (1. College of Electronics and Communications Engineering, Anhui Xinhua University, Hefei 230088, China; 2. Department of Medical Technology, Anhui Medical College, Hefei 230000, China; 3. Electronic Engineering Institute of PLA, Hefei 230000, China; 4. Institute of Intelligent Machines, Chinese Academy of Sciences, Hefei 230000, China)
  • Received:2012-07-06 Online:2013-03-15 Published:2013-03-13

摘要: 追踪目标在经历较大姿势变化时,会导致追踪目标偏移甚至丢失。为此,提出一种基于稀疏表达的多示例学习目标追踪算法。联合多示例学习与稀疏表达方法,将目标物体的局部稀疏编码作为多示例学习的训练数据,通过学习正负样本的局部稀疏编码获得一个多示例学习的分类器,分类的结果与粒子滤波框架相结合,估计目标在整个视频序列中的运动状态。实验结果表明,该算法稳定性较好,与增量学习追踪算法、范式学习追踪算法和多示例学习追踪算法相比,其中心位置误差率减少30%以上。

关键词: 目标追踪, 多示例学习, 稀疏表达, 分类器, 粒子滤波, 数据字典

Abstract: To solve the difficulty when objects undergo large pose change most existed visual tracking algorithms tend to drift away the target or even fail in tracking it, this paper proposes a Multiple Instance Learning(MIL) target tracking algorithm based on sparse representation. This algorithm is to model the appearance of an object by local sparse codes which are formed as training data for the MIL framework. Within MIL framework, the proposed algorithm learns the sparse codes for a MIL classifier. Results from the trained classifier are input into a particle filter framework to sequentially estimate the target state over time for visual tracking. Experimental results show that this algorithm stability is good, and can decrease the center position error rate by 30% compared with IVT algorithm, L1 algorithm, MIL tracking algorithm.

Key words: object tracking, Multiple Instance Learning(MIL), sparse representation, classifier, particle filtering, data dictionary

中图分类号: