Optics and Precision Engineering, Volume. 31, Issue 6, 860(2023)

Multi-object pedestrian tracking method based on improved high resolution neural network

Hongying ZHANG*, Pengyi HE, and Xiaowen PENG
Author Affiliations
  • College of Electronic Information and Automation, Civil Aviation University of China, Tianjin300300, China
  • show less
    Figures & Tables(15)
    Network structure
    Structure diagram of bottleneck(left) and bottle2neck(right)
    Schematic diagram of ECA
    Structural contrast diagram of basicblock(left) and 2ECA-basicblock(right)
    Tracking flow chart
    Visualization of tensors in the middle layer of the network
    Tracking results on Eth-PedCross2 (frames 46, 57 and 69 from left to right)
    Tracking results on MOT17-02-DPM (frames 1,50,100 and 150 from left to right)
    Tracking results on MOT20-04 (frames 0,10,50 and 100 from left to right)
    • Table 1. Comparison test results between bottleneck and bottle2neck

      View table
      View in Article

      Table 1. Comparison test results between bottleneck and bottle2neck

      Layer1组成结构MOTAIDF1MOTP
      bottleneck71.774.678.6
      bottle2neck71.975.078.0
    • Table 2. Comparison results of parameters and detection performance indexes between ECA and CBAM

      View table
      View in Article

      Table 2. Comparison results of parameters and detection performance indexes between ECA and CBAM

      ECA添加位置MOTAIDF1MOTP
      stem65.266.578.7
      stage69.870.778.3
      stem+stage70.870.578.5
    • Table 3. Comparison results of tests with different numbers of ECA modules added inside basicblock

      View table
      View in Article

      Table 3. Comparison results of tests with different numbers of ECA modules added inside basicblock

      ECA模块添加数目MOTAIDF1MOTP
      ECA-basicblock70.870.578.5
      2ECA-basicblock71.871.278.0
    • Table 4. Partial weight parameters of the network

      View table
      View in Article

      Table 4. Partial weight parameters of the network

      Layer nameWeight
      Conv164×3×3×3
      ECA1.conv1×1×3
      Conv264×64×3×3
      ECA2.conv1×1×3
      Layer1

      [(64×64×1×1),(64×64×3×3),(64×256×1×1)]

      [(256×64×1×1),(64×64×3×3),(64×256×1×1)]×3

      ECA3.conv1×1×3
      ..........
      Last layer64×270×3×3,bias=64
      hmhm.0(256×64×3×3,bias=64) hm.2(1×256×1×1,bias=1)
      whwh.0(256×64×3×3,bias=64) wh.2(2×256×1×1,bias=2)
      idid.0(256×64×3×3,bias=64)id.2(128×256×1×1,bias=128)
      regreg.0(256×64×3×3,bias=64) reg.2(2×256×1×1,bias=2)
    • Table 5. Test results of the proposed algorithm and FairMOT

      View table
      View in Article

      Table 5. Test results of the proposed algorithm and FairMOT

      DatasetMethodMOTA↑IDF1↑IDS↓FN↓FP↓FPS↑
      2DMOT15FairMOT71.774.61366 1001 84918.31
      Ours71.871.21055 9292 0229.05
      MOT20FairMOT12.817.34 4221 098 26162 43414.69
      Ours13.618.04 7471 088 28662 7278.18
      MOT17FairMOT75.176.72 23855 09226 44216.23
      Ours77.777.21 53355 03818 5798.96
    • Table 6. Test results of ours compared with other algorithms

      View table
      View in Article

      Table 6. Test results of ours compared with other algorithms

      DatasetTrackerMOTAIDF1IDSFNFPTime elapsed
      MOT17_trainTube_TK79.575.13 57056 8508 6015 316.88
      CSTrack75.976.91 96258 94720 1781 009.53
      TransCenter70.167.52 01794 9793 80215 948.00
      Fair(HRNetV2)75.176.72 23855 09226 442982.79
      Ours77.777.21 53355 03818 5791 780.20
    Tools

    Get Citation

    Copy Citation Text

    Hongying ZHANG, Pengyi HE, Xiaowen PENG. Multi-object pedestrian tracking method based on improved high resolution neural network[J]. Optics and Precision Engineering, 2023, 31(6): 860

    Download Citation

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

    Category: Information Sciences

    Received: May. 26, 2022

    Accepted: --

    Published Online: Apr. 4, 2023

    The Author Email: Hongying ZHANG (carole_zhang0716@163.com)

    DOI:10.37188/OPE.20233106.0860

    Topics