Infrared and Laser Engineering, Volume. 54, Issue 2, 20240373(2025)

Sparse multiple hypothesis matching and model lightweighting for infrared multi-object tracking

Changqi XU1, Haoxian WANG1,2,3, Jun WANG1,2,3, and Zhiquan ZHOU1,2,3
Author Affiliations
  • 1Department of Information Science and Engineering, Harbin Institute of Technology, Weihai 264209, China
  • 2Shandong Provincial Key Laboratory of Marine Electronic Information and Intelligent Unmanned Systems, Weihai 264209, China
  • 3Key Laboratory of Cross-Domain Synergy and Comprehensive Support for Unmanned Marine Systems, Ministry of Industry and Information Technology, Weihai 264209, China
  • show less
    Figures & Tables(16)
    Solve the cost matrix set
    Flow of SMH algorithm
    A set of cost matrices is obtained by grouping based on pseudo-depth information
    Modeling of hypothesis trees
    The question of the trajectory trees is converted to the question of maximum clique
    (a) The CBS module is not split; (b) The CBS module is divided into two parts
    Channel reduction and LAMP pruning
    The general flowchart of the algorithm
    (a) Impact of threshold \begin{document}$\Delta $\end{document} on MOTA; (b) Impact of threshold \begin{document}$\Delta $\end{document} on IDF1; (c) Impact of threshold \begin{document}$\Delta $\end{document} on HOTA
    (a) BoT-SORT tracking results; (b) SMH tracking results
    (a) Images with small objects; (b) Feature maps of small objects before pruning in YOLOv8s; (c) Feature maps of small objects after pruning in YOLOv8s; (d) Images with large objects; (e) Feature maps of large objects before pruning in YOLOv8s; (f) Feature maps of large objects after pruning in YOLOv8s
    Comparison of the number of channels before and after pruning
    • Table 1. The algorithm in this article SMH is compared with other algorithms

      View table
      View in Article

      Table 1. The algorithm in this article SMH is compared with other algorithms

      MOTAMOTPHOTADetAAssAIDF1↑
      MAA[21]91.7%91.8%86.6%84.1%89.8%92.8%
      StrongSORT[22]91.6%91.9%86.7%84.1%89.8%92.8%
      C-BIoU[7]91.3%91.8%86.5%83.8%89.9%92.7%
      ByteTrack[6]91.2%91.9%86.4%83.7%89.8%92.5%
      BoT-SORT[23]91.4%92.1%86.8%84.2%90.0%92.6%
      SparseTrack[8]90.5%91.7%85.8%83.0%89.3%91.8%
      SMH91.8%92.1%87.0%84.4%90.1%92.9%
    • Table 2. Speed comparison between SMH and MHT

      View table
      View in Article

      Table 2. Speed comparison between SMH and MHT

      Stage1/ms↓Stage2/ms↓All/ms↓
      MHT1.183.674.85
      SMH0.923.634.55
    • Table 3. Comparison of YOLOv8s and YOLOv8s-prune

      View table
      View in Article

      Table 3. Comparison of YOLOv8s and YOLOv8s-prune

      Params/106FLOPs/109mAP50mAP50-95FPS/s−1
      YOLOv8s11.128.498.0%80.5%774.4
      YOLOv8s-prune4.818.997.8%80.6%923.0
    • Table 4. Ablation experiment

      View table
      View in Article

      Table 4. Ablation experiment

      MOTAMOTPIDF1↑HOTA
      YOLOv8s + Vanilla91.3%91.8%92.6%86.5%
      YOLOv8s + MHT91.7%91.8%92.8%86.6%
      YOLOv8s + SMH91.8%92.1%92.9%87.0%
      YOLOv8s-prune + SMH91.8%91.8%93.2%86.9%
    Tools

    Get Citation

    Copy Citation Text

    Changqi XU, Haoxian WANG, Jun WANG, Zhiquan ZHOU. Sparse multiple hypothesis matching and model lightweighting for infrared multi-object tracking[J]. Infrared and Laser Engineering, 2025, 54(2): 20240373

    Download Citation

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

    Category:

    Received: Nov. 13, 2024

    Accepted: --

    Published Online: Mar. 14, 2025

    The Author Email:

    DOI:10.3788/IRLA20240373

    Topics