Opto-Electronic Engineering, Volume. 51, Issue 6, 240055-1(2024)

A traffic sign recognition method based on improved YOLOv5

Liguo Qu1...2,*, Xin Zhang1, Zibao Lu1, Yuling Liu1 and Guohao Chen3 |Show fewer author(s)
Author Affiliations
  • 1School of Physics and Electronic Information, Anhui Normal University, Wuhu, Anhui 241002, China
  • 2Anhui Provincial Engineering Research Center for Information Fusion and Control of Intelligent Robots, Wuhu, Anhui 241002, China
  • 3Wuhan Mingke Rail Transit Equipment Co., Ltd., Wuhan, Hubei 430074, China
  • show less
    Figures & Tables(18)
    YOLOv5 network structure
    Fog image
    Structure diagram of C3 and PC3
    Conv and PConv
    Distribution of label aspect ratio
    YOLOv5 default path and EFPN path
    Removing the path EFPN' from the large object detection layer
    CBAM attention mechanism
    Improved network structure
    Distribution of traffic signs after expansion
    Comparison of accuracy effects of traffic signs before and after improvement
    Comparison of missed detection effects of traffic signs before and after improvement
    Comparison of false detection effects of traffic signs before and after improvement
    • Table 1. Comparison of EFPN and EFPN' structures

      View table
      View in Article

      Table 1. Comparison of EFPN and EFPN' structures

      改进方法精确率P/%Params/M
      pnw32ph4
      模型说明:YOLOv5+PC3+EFPN表示YOLOv5中C3中的普通卷积用轻量化部分卷积PConv替换,构成PC3替换掉原C3结构,检测头用EFPN替换;YOLOv5+PC3+EFPN'表示YOLOv5中C3中的普通卷积用轻量化部分卷积PConv替换,构成PC3替换掉原C3结构,检测头用EFPN'替换。
      YOLOv5+PC3+EFPN0.930.700.764.96
      YOLOv5+PC3+ EFPN '0.960.960.983.80
    • Table 2. YOLOv5 default anchor box size

      View table
      View in Article

      Table 2. YOLOv5 default anchor box size

      检测尺度Anchor1Anchor2Anchor3
      小尺寸[10,13][16,30][33,23]
      中尺寸[30,61][62,45][59,119]
      大尺寸[116,90][156,198][373,326]
    • Table 3. Results of K-means clustering algorithm

      View table
      View in Article

      Table 3. Results of K-means clustering algorithm

      检测尺度Anchor1Anchor2Anchor3
      小尺寸[5,5][6,7][8,9]
      中尺寸[9,14][9,14][14,15]
      大尺寸[19,20][19,20][25,26]
    • Table 4. Results of ablation experiment

      View table
      View in Article

      Table 4. Results of ablation experiment

      编号模型PRmAP0.5FPSL/msParams/M
      模型说明:FOG代表扩充雾化数据集TT100K-FOG;PC3代表使用更加轻量的PConv构建PC3特征提取模块来取代YOLOv5骨干和颈部网络中的C3模块;EFPN代表采用延伸的特征金字塔结构,替代YOLOv5中检测头;EFPN'代表在EFPN结构中删除大目标检测层后,替代YOLOv5中检测头;Focal-EIoU代表采用Focal-EloU取代YOLOv5默认函数CIoU;CBAM代表在YOLOv5主干网络中嵌入空间和通道注意力模块。
      0YOLOv50.8420.8240.861145.76.97.10
      1YOLOv5+FOG0.8930.8380.870145.76.97.10
      2YOLOv5+FOG+PC30.8530.7680.840166.76.04.87
      3YOLOv5+FOG+PC3+EFPN0.8620.7590.842135.17.44.96
      4YOLOv5+FOG+PC3+EFPN'0.8760.7820.854161.36.23.80
      5YOLOv5+FOG+PC3+EFPN'+Focal-EIoU0.9060.7900.860161.36.23.80
      6YOLOv5+FOG+PC3+EFPN'+Focal-EIoU+CBAM0.9170.8530.899151.56.73.95
    • Table 5. Performance comparison with other algorithms

      View table
      View in Article

      Table 5. Performance comparison with other algorithms

      模型平台主干网类型P/%mAP0.5 /%FPS
      Faster R-CNNMMDetectionResNet50Anchor-based71.979.957.7
      YOLOv4DarknetDarknetAnchor-based58.782.280.9
      YOLOv5YOLOv5DarknetAnchor-based84.286.1145.7
      YOLOXMMDetectionDarknetAnchor-free72.679.793.6
      YOLOv6YOLOv6EfficientRepAnchor-free77.781.1162.8
      YOLOv7YOLOv7E-ELANAnchor-based72.077.4130.2
      YOLOv8YOLOv8DarknetAnchor-free87.783.7171.4
      OursYOLOv5DarknetAnchor-based91.789.9151.5
    Tools

    Get Citation

    Copy Citation Text

    Liguo Qu, Xin Zhang, Zibao Lu, Yuling Liu, Guohao Chen. A traffic sign recognition method based on improved YOLOv5[J]. Opto-Electronic Engineering, 2024, 51(6): 240055-1

    Download Citation

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

    Category:

    Received: Mar. 7, 2024

    Accepted: Jun. 4, 2024

    Published Online: Oct. 21, 2024

    The Author Email: Qu Liguo (曲立国)

    DOI:10.12086/oee.2024.240055

    Topics