• Acta Photonica Sinica
  • Vol. 52, Issue 9, 0912002 (2023)
Weili LIU1、2, Deli ZHU1、2、*, Huahao LUO1、2, and Yi LI3
Author Affiliations
  • 1School of Computer and Information Science,Chongqing Normal University,Chongqing 401331,China
  • 2Chongqing Digital Agricultural Service Engineering Technology Research Center,Chongqing 401331,China
  • 3Information Center of Chongqing Academy of Animal Husbandry,Chongqing 401331,China
  • show less
    DOI: 10.3788/gzxb20235209.0912002 Cite this Article
    Weili LIU, Deli ZHU, Huahao LUO, Yi LI. 3D Object Detection with Fusion Point Attention Mechanism in LiDAR Point Cloud[J]. Acta Photonica Sinica, 2023, 52(9): 0912002 Copy Citation Text show less
    Overall framework of PointPillars algorithm
    Fig. 1. Overall framework of PointPillars algorithm
    The overall framework of the improved PointPillars algorithm
    Fig. 2. The overall framework of the improved PointPillars algorithm
    Structure of point-wise spatial attention module
    Fig. 3. Structure of point-wise spatial attention module
    2D backbone network structure
    Fig. 4. 2D backbone network structure
    CSPNet,BottleNeck network structure
    Fig. 5. CSPNet,BottleNeck network structure
    Comparison of the visualization results of PointPillars and the algorithm in this paper
    Fig. 6. Comparison of the visualization results of PointPillars and the algorithm in this paper
    LayerRepeatKernel sizeStrideOutput channels
    Conv2d13×3264
    Stage1Conv131×1132
    Conv21×1132
    BottleNeck

    1×1

    3×3

    32

    32

    Conv31×1164
    Conv2d13×32128
    Stage2Conv151×1164
    Conv21×1164
    BottleNeck

    1×1

    3×3

    64

    64

    Conv31×11128
    Conv2d13×32256
    Stage3Conv151×11128
    Conv21×11128
    BottleNeck

    1×1

    3×3

    128

    128

    Conv31×11256
    Table 1. CSPNet network structure of this paper
    EasyModerateHard
    Min height of bounding box40 pixels25 pixels25 pixels
    Max blocking levelFully visiblePartially obscuredHard to see
    Maximum cut-off15%30%50%
    Table 2. Data division in three scenarios
    Experimental environmentConfiguration
    Operating systemUbantu 16.04
    ProcessorIntel Xeon Silver 411
    Memory64 GB
    Video cardNVIDIA TITAN V
    Deep learning frameworkPytorch 1.5
    Development languagePython 3.7
    Table 3. Experimental environment configuration
    Method/R40Car-3D(IoU=0.7)Car-BEV(IoU=0.7)
    EasyModerateHardEasyModerateHard
    F-PointNets2582.1969.7960..5991.1784.6774.77
    VoxelNet1087.9375.3773.2189.3579.2677.39
    SECOND1283.3472.5565.8289.3983.7778.59
    PointPillars1386.2976.7773.9291.8988.0787.02
    TANet2684.3975.9468.8275.7059.4452.53
    SegVoxelNet2786.0476.1270.7691.6286.3783.04
    PointRCNN886.9675.6470.7092.1387.3982.72
    Part-A22887.8178.4973.5191.7087.7984.61
    Ours88.5279.0276.2292.6388.5387.16
    Table 4. Comparison of AP for different methods(%)
    MethodReasoning speed/(frame·s-1
    F-PointNets250.169
    VoxelNet100.033
    SECOND120.380
    3DSSD90.04
    TANet260.035
    SegVoxelNet270.04
    PointRCNN80.067
    Part-A2280.08
    SA-SSD300.04
    Ours0.037 2
    Table 5. Inference speed comparison among different methods
    MethodCar-3D(IoU=0.7)
    EasyModerateHard
    PointPillars86.2976.7773.92
    PPPA87.7278.2375.13
    PPCSP87.8378.3075.60
    PPCSP+PPPA88.5279.0276.22
    Table 6. Average precision of 3D detection for ablation experiments in the KITTI test set(%)
    MethodCar-BEV(IoU=0.7)
    EasyModerateHard
    PointPillars91.8988.0787.02
    PPPA92.5688.6087.24
    PPCSP92.1388.0286.68
    PPCSP+PPPA92.6388.5387.16
    Table 7. Average precision of detection in the BEV scenario of the KITTI test focused ablation experiment(%)
    Weili LIU, Deli ZHU, Huahao LUO, Yi LI. 3D Object Detection with Fusion Point Attention Mechanism in LiDAR Point Cloud[J]. Acta Photonica Sinica, 2023, 52(9): 0912002
    Download Citation