计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (4): 212-220.DOI: 10.3778/j.issn.1002-8331.2009-0184

• 模式识别与人工智能 • 上一篇    下一篇

基于孪生网络融合多模板的目标跟踪算法

杨哲,孙力帆,付主木,张金锦,常家顺   

  1. 河南科技大学 信息工程学院,河南 洛阳 471023
  • 出版日期:2022-02-15 发布日期:2022-02-15

Object Tracking Algorithm Based on Siamese Network with Multiple Templates

YANG Zhe, SUN Lifan, FU Zhumu, ZHANG Jinjin, CHANG Jiashun   

  1. School of Information Engineering, Henan University of Science and Technology, Luoyang, Henan 471023, China
  • Online:2022-02-15 Published:2022-02-15

摘要: 基于全卷积孪生网络的视频目标跟踪算法由于在跟踪过程中使用单一模板,在运动目标外观发生变化时容易出现跟踪漂移并导致精度下降。因此,提出了一种基于孪生网络融合多模板的目标跟踪算法。该算法可在特征级上建立模板库,并使用平均峰值相关能量和模板相似度来保证模板库中各个模板的有效性,从而对多个响应图进行融合以获得更高的跟踪精度。OTB2015和VOT2016数据集上的测试结果表明,在运动目标外观发生变化的复杂环境下,所提算法不但具有较快的跟踪速度,而且相比现有的其他算法能取得更为优异的跟踪性能。

关键词: 目标跟踪, 孪生网络, 模板更新, 平移等变性, 模板库

Abstract: The object tracking algorithms based on fully-convolutional siamese network use single template in the tracking phase. It causes the tracking drift when the appearance of moving target changes, which decreases tracking accuracy. In view of this, an object tracking algorithm based on siamese network with multiple templates is proposed, which builds the template library at the feature level. Here, the average peak-to correlation energy and template similarity are used to ensure the effectiveness of each template in the template library. Multiple response maps are fused to improve the tracking accuracy. The experiment results on OTB2015 and VOT2016 datasets demonstrate that the proposed algorithm not only gets fast tracking speed, but also achieves better tracking performance in comparison with other mainstream algorithms in the complex environment where the appearance changes.

Key words: object tracking, siamese network, template update, translation equivariance, template library