• Optics and Precision Engineering
  • Vol. 31, Issue 20, 3021 (2023)
Dandan HUANG1, Han GAO1, Zhi LIU1,2,*, Lintao YU1, and Huiji WANG1
Author Affiliations
  • 1School of Electronics and In formation Engineering, Changchun University of Science and Technology, Changchun30022, China
  • 2National and Local Joint Engineering Research Center of Space Photoelectric Technology, Changchun University of Science and Technology, Changchun1300, China
  • show less
    DOI: 10.37188/OPE.20233120.3021 Cite this Article
    Dandan HUANG, Han GAO, Zhi LIU, Lintao YU, Huiji WANG. Lightweight target detection network for UAV platforms[J]. Optics and Precision Engineering, 2023, 31(20): 3021 Copy Citation Text show less
    YOLOV5's Backbone layer various structural diagrams
    Fig. 1. YOLOV5's Backbone layer various structural diagrams
    FasterNet network structure
    Fig. 2. FasterNet network structure
    PConv working principle
    Fig. 3. PConv working principle
    Comparison of convolutional variants
    Fig. 4. Comparison of convolutional variants
    Improved YOLOV5 model
    Fig. 5. Improved YOLOV5 model
    C3_FN network structure
    Fig. 6. C3_FN network structure
    VisDrone 2019 Data set target statistical diagram
    Fig. 7. VisDrone 2019 Data set target statistical diagram
    Effect of ultra -parameter on YOLOV5 model
    Fig. 8. Effect of ultra -parameter on YOLOV5 model
    Object detection schematic diagram of this model
    Fig. 9. Object detection schematic diagram of this model
    Exeriment diagram on Jetson Nano
    Fig. 10. Exeriment diagram on Jetson Nano
    检测分支锚框设定
    P2(1,4), (2,9), (5,6)
    P3(5,13), (10,10), (8,20)
    P4(19,17), (15,31), (34,42)
    P5(30,61), (62,45), (59,119)
    Table 1. Setting value of the anchor frame of each detection branch
    包含的检测分支mAP0.5/%
    P3,P4,P540.8
    P2,P4,P541.2
    P2,P3,P540.7
    P2,P3,P440.8
    P2,P3,P4,P544.7
    Table 2. Different detection branches comparison results
    总训练轮数预测边界框mAP0.5/%
    50 epochsCIOU39.6
    NWD38.8
    100 epochsCIOU40.8
    NWD41.1
    Table 3. Different predictive boundary box comparison results
    方法mAP0.5/%Parameters/MFLOPs/GInference time/ms
    C340.844.0180.670.8
    C3_Ghost33.419.663.365.8
    C3_FN40.929.970.836.2
    Table 4. Different network structure performance comparison
    方法

    mAP0.5/

    %

    mAP0.5-0.95/

    %

    Inference time/ms
    Light-RCNN2939.523.252.1
    Cascade-RCNN3037.822.5657.3
    RetinaNet3131.620.148.6
    YOLOv5l40.823.976.8
    TPH-YOLOv546.227.555.1
    YOLOv742.123.448.9
    OURS47.628.745.9
    Table 5. VisDrone test data set experiment results
    方法mAP0.5/%Parameters/MFLOPs/GInference time/ms
    Baseline40.844.0108.476.8
    +P2分支44.748.1186.688.8
    +NWD与IOU混合47.248.1186.689.7
    +C3_FN47.632.8121.645.9
    Table 6. VisDrone test data set discipline experiment
    模型加速前平均推理时间加速后平均推理时间
    YOLOv5l563153
    OURS34284
    Table 7. Comparison of inference time before and after TensorRT acceleration
    Dandan HUANG, Han GAO, Zhi LIU, Lintao YU, Huiji WANG. Lightweight target detection network for UAV platforms[J]. Optics and Precision Engineering, 2023, 31(20): 3021
    Download Citation