• Opto-Electronic Engineering
  • Vol. 51, Issue 5, 240051 (2024)
Runmei Zhang1,2,3,4, Yufei Xiao1, Zhennan Jia1, Zhong Chen1,2..., Zihua Chen1,2, Bin Yuan1,2,3,4, Weiwei Cao4 and Weiwei Song3,*|Show fewer author(s)
Author Affiliations
  • 1School of Mechanical and Electrical Engineering, Anhui Jianzhu University, Hefei, Anhui 230601, China
  • 2Key Laboratory of Intelligent Manufacturing of Construction Machinery, Hefei, Anhui 230601, China
  • 3Anhui Simulation Design and Modern Manufacturing Engineering Technology Research Center, Huangshan, Anhui 242700, China
  • 4Key Laboratory of Civil Aviation Flight Technology and Flight Safety, Guanghan, Sichuan 618300, China
  • show less
    DOI: 10.12086/oee.2024.240051 Cite this Article
    Runmei Zhang, Yufei Xiao, Zhennan Jia, Zhong Chen, Zihua Chen, Bin Yuan, Weiwei Cao, Weiwei Song. Improved YOLOv7 algorithm for target detection in complex environments from UAV perspective[J]. Opto-Electronic Engineering, 2024, 51(5): 240051 Copy Citation Text show less
    SimAM attention module
    Fig. 1. SimAM attention module
    SSG-YOLOv7 overall structure
    Fig. 2. SSG-YOLOv7 overall structure
    GhostConv structure
    Fig. 3. GhostConv structure
    Comparison of (a) NMS and (b) soft NMS detection effect sample chart
    Fig. 4. Comparison of (a) NMS and (b) soft NMS detection effect sample chart
    Data augmentation comparison of two kinds of datasets
    Fig. 5. Data augmentation comparison of two kinds of datasets
    Visual comparison of YOLOv7 and SSG-YOLOv7 detection effect
    Fig. 6. Visual comparison of YOLOv7 and SSG-YOLOv7 detection effect
    特征图尺寸感受野锚框
    20x20Big[33,49],[63,73]
    40x40Medium[14,35],[27,23]
    80x80Small[20,8,8,15,14]
    160x160Tiny[2,5,4,11]
    Table 1. Anchor frame size generated by K-means++
    模块类型Parameters/MGFLOPS
    SPPCSPC模块12.816.2
    SG-SPPCSPC模块3.694.9
    Table 2. Comparison of SPPCSPC and SG-SPPCSPC
    ModelK-means++SimAMSG-SPPCSPCSoft NMSVis_mAP@0.5/%RSOD_mAP@0.5/%Parameters/MFPSGFLOPs
    A40.8995.6037.682106.5
    B44.1596.9137.682106.5
    C46.4097.2237.687107.2
    D48.6197.9128.59395.9
    E51.34(+10.45)98.27(+2.67)28.59395.9
    Table 3. Results of ablation experiments
    Model原始VisDrone增强后VisDrone原始RSOD增强后RSOD
    YOLOv736.7640.8992.0195.60
    SSG-YOLOv742.6351.3493.8298.27
    Table 4. Comparison of mAP(%) before and after data enhancement
    MethodVisdrone_mAP@0.5 /%Visdrone_mAP@0.5:0.95 /%RSOD_mAP@0.5 /%RSOD_mAP@0.5:0.95 /%FPSParameters/M
    Faster R-CNN[6]20.08.9185.654.143137.10
    SSD[23]10.25.187.452.624926.29
    YOLOv5s27.415.694.059.51267.28
    YOLOv5m32.018.895.266.49821.38
    YOLOv5l36.521.595.168.37547.10
    YOLOv7[13]40.824.095.669.68237.62
    YOLOv8s43.125.094.163.016011.17
    YOLOv8m39.622.894.168.712225.90
    YOLOv8l43.725.196.068.99843.69
    本文算法51.329.298.370.09328.49
    Table 5. Comparison of experimental results
    Runmei Zhang, Yufei Xiao, Zhennan Jia, Zhong Chen, Zihua Chen, Bin Yuan, Weiwei Cao, Weiwei Song. Improved YOLOv7 algorithm for target detection in complex environments from UAV perspective[J]. Opto-Electronic Engineering, 2024, 51(5): 240051
    Download Citation