Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 2, 256(2023)

Single-objective tracking algorithm based on Siamese networks

Zong-cheng MIAO1,2、*, Shi-yan GAO1, Ze-min HE1, and Yuan OU3
Author Affiliations
  • 1Xi'an Key Laboratory of Advanced Photo-electronics Materials and Energy Conversion Device,Xijing University,Xi'an 710123,China
  • 2School of Artificial Intelligence,Optics and Electronics,Northwestern Polytechnical University,Xi'an 710072,China
  • 3Institute of System Engineering,Academy of Military Sciences,Beijing 100039,China
  • show less
    Figures & Tables(11)
    Effects of visual object tracking under complex appearance changes.(a)Ambient lighting changes;(b)Fast camera motion;(c)Complete occlusion;(d)Noise interference;(e)Non-rigid shape deformation;(f)Plane outer object rotation and pose. Changes in the appearance of objects caused by these factors can cause tracking performance to degrade or even fail[5].
    Network structure of SiamFC[16]
    Network structure of SiamRPN[29]
    Network structure of SiamRPN++[41]
    Network structure of SiamAttn[48]
    Network structure of ATOM [50]
    Comparison of the represent algorithms of the three methods on the LaSOT dataset. The larger values indicate better performance.
    • Table 1. Comparison based on full-convolutional Siamese network algorithms

      View table
      View in Article

      Table 1. Comparison based on full-convolutional Siamese network algorithms

      跟踪器准确性鲁棒性EAO
      SiamFC0.500.590.19
      SA-Siam0.500.460.24
      CFNet0.430.480.10
      SiamBM0.570.480.32
      DensSiam0.540.350.25
      StructSiam0.530.380.26
      Triloss0.500.540.21
    • Table 2. Comparison of algorithms based on the introduction of regression Siamese networks

      View table
      View in Article

      Table 2. Comparison of algorithms based on the introduction of regression Siamese networks

      追踪器准确性鲁棒性EAO
      SiamRPN0.490.460.24
      GOTURN0.510.200.21
      SPM-Tracker0.580.300.34
      DaSiamRPN0.560.340.33
      C-RPN0.550.320.29
      Ocean0.600.170.47
      SiamFC++0.460.180.40
      SiamCAR0.480.200.41
      SiamBAN0.570.130.45
      SiamDW0.520.410.30
      SiamRPN++0.600.230.42
      SiamMask0.590.460.29
    • Table 3. Comparison of Siamese network algorithms based on online updates

      View table
      View in Article

      Table 3. Comparison of Siamese network algorithms based on online updates

      追踪器准确性鲁棒性EAO
      MDNet0.540.230.26
      RT-MDNet0.540.340.27
      DSiam0.540.280.28
      SiamAttn0.680.140.54
      UpdateNet0.610.210.48
      ATOM0.590.200.40
      ROAM0.500.230.33
      DIMP0.600.150.44
      PrDIMP0.620.170.44
    • Table 4. Three Siamese network methods represent algorithmic comparisons

      View table
      View in Article

      Table 4. Three Siamese network methods represent algorithmic comparisons

      代表算法工作机制与优缺点作用场景
      SiamFC在孪生网络用相似性学习问题替换目标跟踪过程。优点是速度在实时性方面大幅度提高,对速度和精度有很好的平衡;缺点是目标出现遮挡、形变等情况时跟踪性能下降,且缺乏尺度估计。
      StructSiam使用局部结构模式,把相似性学习问题替换为局部特征块学习问题。优点是在实时运行时跟踪精度和速度方面均优于SiamFC。基于全卷积神经孪生网络的目标跟踪算法与基于引入回归的目标跟踪算法二者在目标跟踪时都不需要对模板进行更新,可以进行稳定持久的跟踪,但是性能容易受到形变、背景变化等外界因素影响。
      SiamRPN在孪生网络中引入区域建议模块,利用回归等两个分支进行目标追踪。优点是通过边界框回归和区域建议网络进行目标尺度预测,从而提高性能。缺点是在模型识别方面仍有相对较低的泛化能力,难以处理与目标对象外观相似的干扰物。
      DaSiamRPN设计干扰器感知模块,引入一种简单局部到全局搜索策略进行目标跟踪。优点是利用干扰感知特征学习方案显著提高了网络的判别能力,缺点是缺乏足够强大的判别能力。
      UpdateNet设计了UpdateNet神经网络,其可以集成到所有Siamese跟踪器中。优点是提出的更新方法显著提高了跟踪器在标准线性更新(或根本不更新)方面的性能,缺点是更新时产生误差并且会逐渐积累。基于在线更新的目标跟踪算法可以随时对跟踪过程中出现的目标形变,背景变化等外界因素做出反应,但是误差会随之累积。
    Tools

    Get Citation

    Copy Citation Text

    Zong-cheng MIAO, Shi-yan GAO, Ze-min HE, Yuan OU. Single-objective tracking algorithm based on Siamese networks[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(2): 256

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research Articles

    Received: Jun. 4, 2022

    Accepted: --

    Published Online: Feb. 20, 2023

    The Author Email: Zong-cheng MIAO (miaozongcheng@nwpu.edu.cn)

    DOI:10.37188/CJLCD.2022-0186

    Topics