Opto-Electronic Engineering, Volume. 51, Issue 12, 240244-1(2024)

THI-YOLO: Improved non-motorized drivers helmet detection of YOLOv8

Guangling Sun1,2、* and Xinbo Wang1
Author Affiliations
  • 1School of Electronic and Information Engineering, Anhui Jianzhu University, Hefei, Anhui 230601, China
  • 2Anhui Interational Joint Research Center for Intelligent Perception and High-dimensional Modeling of Ancient Buildings, Anhui Jianzhu University, Hefei, Anhui 230601, China
  • show less
    Figures & Tables(24)
    THI-YOLO model
    Structure of GSConv module
    CBAM structure
    Channel attention submodule
    Spatial attention submodule
    Diagram of the C2f_BC structure
    Diagram of the Bottleneck_GC structure
    Diagram of the MP-Parnet structure
    Labelimg script interface
    Scatterplots of length and width distribution of different categories
    Comparison of box_loss before and after improvement
    Display of qualitative experiment results. (a) Cases of object leakage; (b) Cases of object misdetections; (c) Comparison of detection effects for dense scenes of drivers
    • Table 1. Number of target after Helmet dataset processing

      View table
      View in Article

      Table 1. Number of target after Helmet dataset processing

      ClassDriverHelmetNo helmet
      Target number902777433184
    • Table 2. Number of target after TWHD dataset processing

      View table
      View in Article

      Table 2. Number of target after TWHD dataset processing

      ClassTwo_wheelerHelmetWithout_helmet
      Target number13790117426895
    • Table 3. Configuration of experimental parameters

      View table
      View in Article

      Table 3. Configuration of experimental parameters

      Key parameterParameter value
      Epoch200
      lr00.01
      Imgsz640
      Batch8
      Momentum0.937
    • Table 4. mAP50 values at different thresholds

      View table
      View in Article

      Table 4. mAP50 values at different thresholds

      du
      10.90.80.70.6
      00.8420.8400.8400.8410.842
      0.10.8420.8420.8380.8400.842
      0.20.8400.8400.8390.8420.840
      0.30.8410.8420.8430.8420.838
      0.40.8390.8410.8420.8380.840
    • Table 5. mAP50∶95 values at different thresholds

      View table
      View in Article

      Table 5. mAP50∶95 values at different thresholds

      du
      10.90.80.70.6
      00.6810.6820.6840.6810.681
      0.10.6840.6850.6820.6820.682
      0.20.6820.6840.6860.6860.679
      0.30.6850.6850.6900.6860.678
      0.40.6830.6820.6880.6800.679
    • Table 6. Comparison of mAP before and after improvement

      View table
      View in Article

      Table 6. Comparison of mAP before and after improvement

      AlgorithmPre-improvement lossPost-improvement loss
      mAP50mAP50∶95mAP50mAP50∶95
      YOLOv5n0.8300.6520.832(+0.2%)0.656(+0.4%)
      YOLOv8n0.8390.6860.843(+0.4%)0.690(+0.4%)
    • Table 7. Comparison of GSConv and Conv

      View table
      View in Article

      Table 7. Comparison of GSConv and Conv

      Convolutional typemAP50mAP50∶95Params/106
      Conv0.8430.6903.15
      GSConv0.849(+0.6%)0.693(+0.3%)2.93(−7.0%)
    • Table 8. Introducing CBAM at different convolutional layers

      View table
      View in Article

      Table 8. Introducing CBAM at different convolutional layers

      ModulemAP50mAP50∶95Params/106
      C2f+GSConv0.8490.6932.93
      C2f_B10.8490.6913.02
      C2f_B20.853(+0.4%)0.696(+0.3%)3.02 (+3.1%)
    • Table 9. Comparison of mAP among different attentions

      View table
      View in Article

      Table 9. Comparison of mAP among different attentions

      No.ModuleParams/106mAP50mAP50∶95
      1C2f+GSConv2.930.8490.693
      21+ECA2.930.8480.694
      31+SE2.980.8490.695
      41+GAM3.470.8530.697
      51+EMA2.940.8520.692
      61+CBAM3.020.8530.696
    • Table 10. Adding MP-Parnet to different network necks

      View table
      View in Article

      Table 10. Adding MP-Parnet to different network necks

      ModulePre-ImprovementPost-Improvement
      mAP50mAP50∶95mAP50mAP50∶95
      YOLOv5n0.8320.6560.840(+0.8%)0.674(+1.8%)
      YOLOv6n0.8220.6630.841(+1.9%)0.671(+0.8%)
      YOLOv8n0.8430.6900.856(+1.3%)0.698(+0.8%)
      YOLOv10n0.8370.6850.844(+0.7%)0.690(+0.5%)
      RT-DETR (r18)0.8020.6510.829(+2.7%)0.674(+2.3%)
    • Table 11. Comparison of ablation experiments

      View table
      View in Article

      Table 11. Comparison of ablation experiments

      No.YOLOv8nFocaler-CIoUC2f_BCMP-ParnetHelmetTWHDParams/106GFLOPs
      mAP50mAP50∶95mAP50mAP50∶95
      10.8390.6860.7250.4373.158.7
      20.8430.6910.7300.4403.158.7
      30.8530.6960.7410.4543.028.2
      40.8560.6980.7390.4523.319.0
      50.8610.7050.7430.4563.128.4
    • Table 12. Comparison with other models

      View table
      View in Article

      Table 12. Comparison with other models

      ModuleHelmetTWHDParams/106GFLOPs
      mAP50mAP50: 95mAP50mAP50: 95
      YOLOv5n0.8320.6560.7190.4302.57.1
      YOLOv5s0.8480.6910.7240.4469.1123.8
      YOLOv6n0.8220.6630.7130.4324.2311.8
      YOLOv8s0.8590.7000.7380.45311.1228.4
      YOLOv8n-ghost0.8340.6550.7200.4291.715.0
      YOLOv10n0.8370.6850.7290.4412.698.2
      RT-DETR (r18)0.8020.6510.6960.41719.957.0
      Ours0.8610.7050.7430.4563.128.4
    Tools

    Get Citation

    Copy Citation Text

    Guangling Sun, Xinbo Wang. THI-YOLO: Improved non-motorized drivers helmet detection of YOLOv8[J]. Opto-Electronic Engineering, 2024, 51(12): 240244-1

    Download Citation

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

    Category: Article

    Received: Oct. 18, 2024

    Accepted: Nov. 16, 2024

    Published Online: Feb. 21, 2025

    The Author Email:

    DOI:10.12086/oee.2024.240244

    Topics