• Optics and Precision Engineering
  • Vol. 31, Issue 3, 393 (2023)
Dakuan DU1, Jianfeng SUN1, Yuanxue DING1, Peng JIANG2,*, and Hailong ZHANG1
Author Affiliations
  • 1National Key Laboratory of Science and Technology on Tunable Laser, Institute of Opto-Electronic, Harbin Institute of Technology, Harbin5000, China
  • 2Science and Technology on Complex System Control and Intelligent Agent Cooperation Laboratory, Beijing100074, China
  • show less
    DOI: 10.37188/OPE.20233103.0393 Cite this Article
    Dakuan DU, Jianfeng SUN, Yuanxue DING, Peng JIANG, Hailong ZHANG. Small object detection based on GM-APD lidar data fusion[J]. Optics and Precision Engineering, 2023, 31(3): 393 Copy Citation Text show less
    Framework of target detection network
    Fig. 1. Framework of target detection network
    Structure of FPN
    Fig. 2. Structure of FPN
    Structure of CBAM
    Fig. 3. Structure of CBAM
    Structure of RFB-s
    Fig. 4. Structure of RFB-s
    Structure of DGCNN
    Fig. 5. Structure of DGCNN
    Reconstructed intensity image and range image
    Fig. 6. Reconstructed intensity image and range image
    Comparison of detection results on lidar data set
    Fig. 7. Comparison of detection results on lidar data set
    Comparison of results with or without secondary detection
    Fig. 8. Comparison of results with or without secondary detection
    Number of imagesImage sizeAverage number ofobjects per imageProportion of small objectsMinimum number of pixels occupied by object
    1 60064×641199.1%6
    Table 1. Details of data set
    Configuration itemValue
    SystemUbuntu18.04
    GPUGeForce RTX 2080Ti
    CPUIntel Core i7-8700 CPU @3.20 GHz
    Table 2. Software and hardware environment
    MethodParmsAP50:95AP50AP75FPS
    Faster RCNN108M47.592.446.226
    YOLO102M39.884.131.945
    YOLOv3235M46.394.945.753
    YOLOv521.2M54.196.055.355
    SSD100M53.095.553.341
    Our method163M54.796.556.420
    Table 3. Precision of different detection networks on lidar intensity image data set
    MethodParmsAP50:95AP50AP75FPS
    Faster RCNN108M47.692.747.326
    YOLO102M39.984.732.545
    YOLOv3235M46.795.245.953
    YOLOv521.2M54.596.355.355
    SSD100M53.195.753.341
    Our method (Stage1)163M54.896.756.220
    Our method (Stage1,2)181M56.798.857.317
    Table 4. Precision of different detection networks based on intensity and range information
    MethodRFBCBAMAP50∶95AP50AP75
    Mod_1Mod_2Mod_3Mod_4Mod_4
    FPN53.095.552.2
    Im_FPN154.396.355.4
    Im_FPN254.796.556.4
    Im_FPN353.896.154.9
    Im_FPN454.796.254.8
    Im_FPN553.895.854.3
    Im_FPN653.996.053.2
    Table 5. Precision of different methods in ablation experiment
    InputParmsEpochsAP50∶95AP50AP75
    3D point clouds13.7M~8054.396.754.2
    4D point clouds13.8M~3056.798.857.3
    Table 6. Comparison of detection accuracy of 3D point clouds and 4D point clouds
    Dakuan DU, Jianfeng SUN, Yuanxue DING, Peng JIANG, Hailong ZHANG. Small object detection based on GM-APD lidar data fusion[J]. Optics and Precision Engineering, 2023, 31(3): 393
    Download Citation