欢迎访问《兵工学报》官方网站,今天是 分享到:

兵工学报 ›› 2021, Vol. 42 ›› Issue (9): 1940-1950.doi: 10.3969/j.issn.1000-1093.2021.09.015

• 论文 • 上一篇    下一篇

融合时空特性的孪生网络视觉跟踪

姜珊, 底晓强, 韩成   

  1. (长春理工大学 计算机科学技术学院, 吉林 长春 130022)
  • 上线日期:2021-10-20
  • 作者简介:姜珊(1991—), 女, 博士研究生。 E-mail: sakina_js@163.com
  • 基金资助:
    国家自然科学基金青年基金项目(61702051、61602058)

Siamese Network for Visual Tracking with Temporal-spatial Property

JIANG Shan, DI Xiaoqiang, HAN Cheng   

  1. (College of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China)
  • Online:2021-10-20

摘要: 针对现有跟踪算法处理快速运动和相似目标干扰挑战时精度欠佳的问题,提出一种基于时空注意力机制的孪生网络跟踪算法。设计时间注意力模块,利用视频初始帧作为参考,依照多幅历史参考帧的贡献程度,自适应地为其赋予权重并进行融合,构建时效性较强的多帧融合模板;结合空间注意力模块,通过非局部操作增强算法对跟踪图像的整体感知能力,进而提升算法的判别能力;在网络训练阶段,利用Focal Loss函数训练网络,以平衡正负样本的比例,提高算法判别困难样本的能力。仿真实验采用标准数据集OTB2015和VOT2016测试算法性能,并与近年来的12种优秀算法即ECO算法、DSST算法、HDT算法、CFNet算法、KCF算法、SRDCF算法、SiamFC算法、DCFNet算法、MEEM算法、SiamVGG算法、BACF算法、ANT算法进行对比。结果表明,融合时空特性的孪生网络跟踪算法可以很好地应对快速运动和相似目标干扰挑战,并有效提升基准算法的性能。

关键词: 目标跟踪, 孪生网络, 时间注意力, 空间注意力

Abstract: A siamese network with temporal-spatial attention mechanism is proposed to tackle the problem of poor accuracy in dealing with fast motion and background clutter in the current algorithms. A temporal attention module is designed, the features of multi-frames are fused according to the contribution of each reference frame,and the weights are assigned adaptively to construct an effective temporal fused template. The spatial attention module is adopted to percept the whole tracking image by non-local operation,which can improve the discriminative ability of the network. During the training stage, the Focal Loss is utilized to train the network model,which can balance the proportion of positive and negative samples,and improve the ability of model to distinguish the hard samples. Several simanation experiments were conducted on the OTB2015 and VOT2016 benchmark to evaluate the performance of the proposed algorithm,and compared it with the state-of-the-art algorithms, i.e., ECO, DSST, HDT, CFNet, KCF, SRDCF, SiamFC, DCFnet, MEEM, SiamVGG, BACF and ANT algorithms. The experimental results demonstrate that the proposed siamese network tracking model with temporal-spatial attention property can well handle the fast motion and background clutter situation,and effectively improve the performance of baseline algorithm.

Key words: objecttracking, siamesenetwork, temporalattention, spatialattention

中图分类号: