Infrared and Laser Engineering, Volume. 50, Issue 10, 20200500(2021)

3D point cloud object detection method in view of voxel based on graph convolution network

Yiqiang Zhao... Akbar Arxidin·, Rui Chen, Yiyao Zhou and Qi Zhang |Show fewer author(s)
Author Affiliations
  • School of Microelectronics, Tianjin University, Tianjin 300072, China
  • show less
    Figures & Tables(9)
    Pipeline of graph convolution 3D object detection based on voxelization
    Schematic diagram of neighborhood of valid voxel
    Structure of designed 3D point cloud object detection network
    3D object detection result of KITTI dataset
    Comparison of P-R curve of 3D object detection
    • Table 1. Comparison of average precision of point cloud 3D object detection on KITTI validation set

      View table
      View in Article

      Table 1. Comparison of average precision of point cloud 3D object detection on KITTI validation set

      NetworksInference time/sSensorsCarPedestrianCyclist
      LiDARImageModerateHardEasyModerateHardEasyModerateHardEasy
      MV3D [10]0.3662.35%55.12%71.09%N/AN/AN/AN/AN/AN/A
      MV3D (Lidar)[10]0.24×52.73%51.31%66.77%N/AN/AN/AN/AN/AN/A
      AVOD [11]0.0865.78%58.38%73.59%31.51%26.98%38.28%44.90%38.80%60.11%
      VoxelNet [3]0.31×65.46%62.85%81.97%53.42%48.87%57.86%47.65%45.11%67.17%
      Point-GNN[6]0.6×79.47%72.29%88.33%43.77%40.14%51.92%63.48%57.08%78.60%
      VGCN (Ours)0.09×79.21%78.58%89.25%53.28%48.7460.90%71.82%68.19%85.89%
    • Table 2. Comparison of average precision of BEV object detection on KITTI validation set

      View table
      View in Article

      Table 2. Comparison of average precision of BEV object detection on KITTI validation set

      NetworksInference time/sSensorsCarPedestrianCyclist
      LiDARImageModerateHardEasyModerateHardEasyModerateHardEasy
      MV3D [10]0.3676.90%68.49%86.02%N/AN/AN/AN/AN/AN/A
      MV3D (Lidar) [10]0.24×77.00%68.94%85.82%N/AN/AN/AN/AN/AN/A
      AVOD [11]0.0885.44%77.73%86.60%35.24%33.97%42.52%47.74%46.55%63.66%
      VoxelNet [3]0.31×84.81%78.57%89.60%61.05%56.98%65.95%52.18%50.49%74.41%
      Point-GNN[6]0.6×89.17%83.90%93.11%43.77%40.14%51.92%67.28%59.67%81.17%
      VGCN (Ours)0.09×87.90%87.33%90.23%57.30%52.72%64.22%74.94%71.57%87.26%
    • Table 3. 3D and BEV object detection on KITTI test set

      View table
      View in Article

      Table 3. 3D and BEV object detection on KITTI test set

      BenchmarksModerateEasyHard
      Car(3D)77.65%84.47%73.36%
      Car(BEV)87.16%90.67%82.98%
      Cyclist(3D)62.36%78.47%55.88%
      Cyclist(BEV)67.04%81.50%59.45%
      Pedestrian(3D)37.60%45.28%34.96%
      Pedestrian(BEV)42.33%50.02%40.05%
    • Table 4. Effect of the number of graph convolutional layers on vehicle object detection

      View table
      View in Article

      Table 4. Effect of the number of graph convolutional layers on vehicle object detection

      NetworksModerateHardEasy
      VoxelNet [3]65.46%62.85%81.97%
      1-Graph conv76.32%75.18%85.85%
      2-Graph convs79.21%78.58%89.25%
      4-Graph convs80.06%75.72%86.73%
    Tools

    Get Citation

    Copy Citation Text

    Yiqiang Zhao, Akbar Arxidin·, Rui Chen, Yiyao Zhou, Qi Zhang. 3D point cloud object detection method in view of voxel based on graph convolution network[J]. Infrared and Laser Engineering, 2021, 50(10): 20200500

    Download Citation

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

    Category: Image processing

    Received: Dec. 20, 2020

    Accepted: --

    Published Online: Dec. 7, 2021

    The Author Email:

    DOI:10.3788/IRLA20200500

    Topics