计算机科学与探索 ›› 2016, Vol. 10 ›› Issue (7): 1035-1043.DOI: 10.3778/j.issn.1673-9418.1511060

• 人工智能与模式识别 • 上一篇    下一篇

结构稀疏表示分类目标跟踪算法

侯跃恩1+,李伟光2   

  1. 1. 嘉应学院 计算机学院,广东 梅州 514000
    2. 华南理工大学 机械与汽车工程学院,广州 510000
  • 出版日期:2016-07-01 发布日期:2016-07-01

Structured Sparse Representation Classification Object Tracking Algorithm

HOU Yue’en1+, LI Weiguang2   

  1. 1. School of Computer, Jiaying University, Meizhou, Guangdong 514000, China
    2. School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou 510000, China
  • Online:2016-07-01 Published:2016-07-01

摘要: 为提高目标跟踪算法在复杂条件下的鲁棒性和准确性,研究了一种基于贝叶斯分类的结构稀疏表示目标跟踪算法。首先通过首帧图像获得含有目标与背景模板的稀疏字典和正负样本;然后采用结构稀疏表示的思想对样本进行线性重构,获得其稀疏系数;进而设计一款贝叶斯分类器,分类器通过正负样本的稀疏系数进行训练,并对每个候选目标进行分类,获得其相似度信息;最后采用稀疏表示与增量学习结合的方法对稀疏字典进行更新。将该算法与其他4种先进算法在6组测试视频中进行比较,实验证明了该算法具有更好的性能。

关键词: 目标跟踪, 粒子滤波, 稀疏表示, 字典, 贝叶斯分类

Abstract: In order to enhance the robustness and precision of tracking algorithm in complex scenarios, this paper proposes a Bayes classification based structured sparse representation object tracking algorithm. Firstly, in the first frame, a sparse dictionary is obtained, which contains target and background templates, as well as positive and negative samples. Secondly, all samples are linearly combined based on the idea of structured sparse representation, hence the coding coefficients can be gotten. Thirdly, a kind of Bayes classifier is designed, which is trained by the coding coefficients of positive and negative samples. The classifier is able to detect the candidate target and obtain the likelihood information of them. Fourthly, the dictionary is updated by combining incremental subspace learning and sparse representation method together. Finally, the proposed tracker performs favorably against 4 state-of-the-art trackers on 6 challenging video sequences.

Key words: object tracking, particle filter, sparse representation, dictionary, Bayes classification