Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 2, 192(2024)

Advances in twin network research in visual tracking technology

Zemin HE1, Juntao ZENG1, Baoxi YUAN1, Dejian LIANG2, and Zongcheng MIAO3、*
Author Affiliations
  • 1Technological Institute of Materials & Energy Science,Xijing University,Xi'an 710123,China
  • 2Beijing Xinghang Electromechanical Equipment Co. Ltd.,Beijing 100074,China
  • 3School of Artificial Intelligence,Optics and Electronics,Northwestern Polytechnical University,Xi'an 710072,China
  • show less
    Figures & Tables(11)
    Evolution of single target tracking
    SiamDA network structure of SiamDA
    Network structure of SBAN
    Network structure of SiamRPN
    Network structure of SiamCAR
    Network structure of SiamET
    Comparison of the representative algorithms in the three methods done on the OTB-2015 dataset.(a)Success rate plot;(b)Accuracy plot.
    • Table 1. Comparison of twin network algorithms based on attention mechanism

      View table
      View in Article

      Table 1. Comparison of twin network algorithms based on attention mechanism

      跟踪器准确性鲁棒性EAOFPS
      SiamDA0.5230.3250.31280
      TASNet0.4960.3860.25740
      SA-Siam0.5000.4590.23650
      TA-Siam0.6180.4720.30480
      SiamAtt0.5940.2150.41740
      SAGT0.5980.2030.44225
      SiamATL0.5140.5490.24725
      SBAN0.6140.2010.43540
    • Table 2. Comparison of twin network algorithms based on hyperparametric inference

      View table
      View in Article

      Table 2. Comparison of twin network algorithms based on hyperparametric inference

      跟踪器准确性鲁棒性EAOFPS
      SiamFC0.5030.5850.18886
      SiamRPN0.5860.2760.383160
      SiamRPN++0.6000.2340.41435
      DaSiamRPN0.5690.3370.326160
      SiamGan0.5920.1860.42535
      SiamBAN0.5970.1780.45240
      Siamese-ORPN0.6100.1430.49985
    • Table 3. Comparison of template-based update twin network algorithms

      View table
      View in Article

      Table 3. Comparison of template-based update twin network algorithms

      跟踪器准确性鲁棒性EAOFPS
      GradNet0.5070.3750.24780
      SiamET0.5940.3660.33131
      SiamRAL0.5610.2170.30227
      DSiam0.5400.2800.28045
      SiamRTU0.6030.2150.42320
      SiamAttn0.6800.1400.53733
      UpdateNet0.6100.2060.48130
    • Table 4. Comparisons of the represent algorithms of three twin network methods

      View table
      View in Article

      Table 4. Comparisons of the represent algorithms of three twin network methods

      代表算法工作原理优点和缺点作用场景
      SiamDA该算法骨干网是一个用于特征提取的孪生网络,引入了双重注意模块,两个注意模块相互训练,共同补充,以提高目标表示能力,从而增强模型的辨别能力优点是运用了注意力机制来突出目标区域并抑制背景信息,效果好,可以并行训练,速度快。缺点是局部信息的获取不如传统的CNN强。基于注意力机制和超参数推理的目标跟踪算法在跟踪目标时都未对模板进行更新,可以对环境进行长时间的稳定跟踪,但遇到形变、遮挡时性能会下降。
      SiamRPN该算法骨干网是改进的AlexNet的孪生网络,引入了区域建议网络,包括分类分支和回归分支。分类分支用来区分目标和背景,回归分支用来对候选区域进行微调。优点是通过区域建议网络来预测目标尺度,增强了目标的判别性,提高了运行速度和精度。缺点是跟踪过程容易出现飘逸,对模型识别泛化能力较低。
      GradNet该算法以SiameseFC的网络为主干,引入了梯度引导网络。网络包括初始嵌入、梯度计算和模板更新,作用是利用梯度中的判别信息更新当前帧的模板,并且为了更好地利用梯度信息,引入模板泛化训练来避免过拟合。优点是提出了利用梯度中的信息来进行模板更新,提高了跟踪器在标准线性更新的性能。缺点是进行模板更新时会产生误差并且会逐渐累积,且运行速度较慢。基于模板更新的目标跟踪算法能够应对跟踪过程中形变、遮挡等复杂环境的出现,但误差会相应增加。
    Tools

    Get Citation

    Copy Citation Text

    Zemin HE, Juntao ZENG, Baoxi YUAN, Dejian LIANG, Zongcheng MIAO. Advances in twin network research in visual tracking technology[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(2): 192

    Download Citation

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

    Category: Research Articles

    Received: Mar. 29, 2023

    Accepted: --

    Published Online: Apr. 24, 2024

    The Author Email: Zongcheng MIAO (miaozongcheng@nwpu.edu.cn)

    DOI:10.37188/CJLCD.2023-0113

    Topics