• Optics and Precision Engineering
  • Vol. 30, Issue 16, 1988 (2022)
Wen HAO1,2,*, Wenjing ZHANG1,2, Wei LIANG1,2, Zhaolin XIAO1,2, and Haiyan JIN1,2
Author Affiliations
  • 1School of Computer Science and Engineering, Xi'an University of Technology, Xi'an70048, China
  • 2Shaanxi Key Laboratory for Network Computing and Security Technology, Xi’an710048, China
  • show less
    DOI: 10.37188/OPE.20223016.1988 Cite this Article
    Wen HAO, Wenjing ZHANG, Wei LIANG, Zhaolin XIAO, Haiyan JIN. Scene recognition for 3D point clouds: a review[J]. Optics and Precision Engineering, 2022, 30(16): 1988 Copy Citation Text show less
    Images of the same scene at different times and illumination in the dataset Oxford RobotCar[6]
    Fig. 1. Images of the same scene at different times and illumination in the dataset Oxford RobotCar6
    Classification of scene recognition algorithms for point clouds
    Fig. 2. Classification of scene recognition algorithms for point clouds
    Flowchart of SegMatch algorithm
    Fig. 3. Flowchart of SegMatch algorithm
    Flowchart of Seed algorithm
    Fig. 4. Flowchart of Seed algorithm
    Chronological overview of scene recognition for point clouds
    Fig. 5. Chronological overview of scene recognition for point clouds
    Network architecture of PointNetVLAD[58]
    Fig. 6. Network architecture of PointNetVLAD58
    Network architecture of LPD-Net[19]
    Fig. 7. Network architecture of LPD-Net19
    Network architecture of Semantic Graph[61]
    Fig. 8. Network architecture of Semantic Graph61
    网络模型年份网络主干结构关键技术数据
    PointNetVLAD582018PointNet,NetVLAD转换网络T-Net、多层感知机、对称函数Oxford Robotcar
    PCAN baseline162019PointNet,NetVLAD转换网络T-Net、多层感知机、对称函数、SAG层Oxford Robotcar
    DAGC baseline592020DGCNN, NetVLAD双注意力模块、EdgeConvOxford Robotcar
    SOE-Net172021PointSift, NetVLADPointOE模块Oxford Robotcar
    AttDLNet182021RangeNet++注意力模块KITTI
    ARIConv622021DenseNet注意旋转不变卷积Oxford Robotcar
    Lpd-Net192019DGCNN, NetVLAD十维几何特征计算、转换网络、动态图网络Oxford Robotcar
    SRNet602020Static Graph Convolution (SGC), NetVLADSGC模块、三层空间注意力模块Oxford Robotcar
    SemGraph612020RangeNet++,DGCNNEdgeConv、图相似性匹配模块KITTI
    EPC-Net632021EPCNet, Grouped VLAD多层ProxyConvOxford Robotcar
    MinkLoc3D242021Feature Pyramid Network architecture局部特征提取网络、广义均值池Oxford Robotcar
    DH3D642020PointNet, NetVLADFlexConv、挤压和激励模块Oxford RobotCar
    TransLoc3D252021External Transformer, NetVLAD自适应感受野模块, 3D稀疏卷积模块Oxford Robotcar
    SVT-Net262021Sparse Voxel Transformers基于原子的稀疏体素变换器、基于聚类的稀疏体素变换器Oxford Robotcar
    Table 1. Network models based on learning to obtain features
    数据集传感器移动平台变化场景相机

    IMU

    频率/Hz

    数据总量
    Oxford RobotCar[6]2017SICK LMS-151车辆不同季节、光照、动态目标遮挡、建筑物改造等综合变化与干扰室外3单目1×1223.15TB
    KITTI odometry[70-71]2013Velodyne HDL-64E车辆室外2双目1×10180 GB
    North Campus Long Term (NCLT)[72]2016Velodyne HDL-32ESegway机器人不同季节、光照、植被等综合变化校园(室内、室外)

    6单目(全向)

    4单目

    1×100

    1×200

    2.95 TB
    MulRan[73]2020

    Ouster OS1-64

    Navtech CIR204-H

    车辆不同时间段会议中心、校园、高速公路、河边道路--387 GB
    Ford[74]2011Velodyne HDL-64E车辆福特研究院、密歇根州迪尔伯恩市中心1单目1×100100 GB
    SEU-FX[75]2019速腾聚创 RS-32车辆不同天气、时间、光照条件城市道路、校园场景双目1×30-
    Table 2. Dataset for scene recognition of point cloud
    ModelNetwork parameter quantity/MBRuntime per frame/ms
    PointNetVLAD5819.7815
    PCAN1620.4255
    Lpd-Net1919.8126
    Minkloc3D241.121
    Table 3. Network parameter quantity and runtime of different scene recognition models
    DescriptorSize
    SHOT33352
    USC341 960
    FPFH3533
    Gestalt3D13130
    NBLD141 408
    ISHOT481 344
    Table 4. 3D local descriptor dimension
    MethodsAverage recall @1%
    OxfordU.S.R.A.B.D.
    PointNetVLAD5880.31%72.63%60.27%65.3%
    PCAN baseline1683.81%79.05%71.18%66.82%
    DAGC baseline5987.49%83.49%75.68%71.21%
    SOE-Net1796.4%93.17%91.47%88.45%
    SRNet6094.56%94.33%89.23%83.49%
    Lpd-net1994.92%96%90.46%89.14%
    EPC-Net6394.74%96.52%88.58%84.92%
    MinkLoc3D2497.9%95%91.2%88.5%
    TransLoc3D2598.5%94.9%91.5%88.4%
    SVT-Net2697.8%96.5%92.7%90.7%
    Table 5. Scene recognition results based on deep learning
    Methods000205060708Mean
    M2DP150.8360.7810.7720.8960.8610.1690.719
    ScanContext440.9370.8580.9550.9980.9220.8110.914
    Locus300.9830.7620.9810.9921.00.9310.942
    PointNetVLAD580.8820.7910.7340.9530.7670.1290.709
    SemGraph610.9600.8590.8970.9440.9840.7830.904
    Table 6. F1 max scores on the KITTI dataset
    Wen HAO, Wenjing ZHANG, Wei LIANG, Zhaolin XIAO, Haiyan JIN. Scene recognition for 3D point clouds: a review[J]. Optics and Precision Engineering, 2022, 30(16): 1988
    Download Citation