计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (18): 116-121.DOI: 10.3778/j.issn.1002-8331.1903-0085

• 模式识别与人工智能 • 上一篇    下一篇

结合低维特征和在线加权MIL的目标跟踪算法

孔凡芝,李金龙,吴冬梅   

  1. 1.浙江传媒学院 电子信息学院,杭州 310018
    2.曲阜师范大学 物理工程学院,山东 曲阜 273165
  • 出版日期:2019-09-15 发布日期:2019-09-11

Target Tracking Algorithm Based on Low-Dimensional Feature and Online Weighted MIL

KONG Fanzhi, LI Jinlong, WU Dongmei   

  1. 1.School of Electronics and Information, Communication University of Zhejiang, Hangzhou 310018, China
    2.School of Physics and Physical Engineering, Qufu Normal University, Qufu, Shandong 273165, China
  • Online:2019-09-15 Published:2019-09-11

摘要: 为了提高视频序列中目标跟踪的准确性,提出了结合低维Haar-like特征和在线加权多示例学习(OWMIL)的跟踪算法。将训练集中的图像进行剪裁,构建正负样本集。通过稀疏编码提取低维度的Haar-like特征来表示目标。通过这些正负样本的局部稀疏特征在线学习生成弱分类器集,并通过示例加权方法来促进学习过程,最终生成一个强分类器,用于测试视频中的目标跟踪。实验结果表明,该算法在旋转、光照和尺度变化等影响下取得了优异的效果。相比其他几种改进型多示例学习算法,提出的算法获得了更好的跟踪效果。

关键词: 目标跟踪, 在线加权多示例学习, Haar-like特征, 稀疏表示

Abstract: In order to improve the accuracy of target tracking in video sequences, a tracking algorithm combining low-dimensional Haar-like feature and Online Weighted Multiple Instance Learning(OWMIL) is proposed. The image in training set is clipped to construct positive and negative sample sets. The low-dimensional Haar-like feature is extracted by sparse coding to represent the target. A weak classifier set is generated by online learning the local sparse features of these positive and negative samples, and an example weighting method is used to promote the learning process. A strong classifier is generated for target tracking in test video. Experimental results show that the proposed algorithm achieves excellent results under the influence of rotation, illumination and scale change.

Key words: target tracking, online weighted multiple instance learning, Haar-like feature, sparse representation