Laser & Optoelectronics Progress, Volume. 62, Issue 10, 1012001(2025)

Enhancing PointPillars Three-Dimensional Object Detection with Density Clustering and Dual Attention Mechanisms

Qingxin Yang1、*, Deming Kong1, Jing Chen2, Xiaowei Li3, and Yue Shen4
Author Affiliations
  • 1School of Electrical Engineering, Yanshan University, Qinhuangdao 066004, Hebei , China
  • 2School of Information Science and Engineering, Yanshan University, Qinhuangdao 066004, Hebei , China
  • 3School of Information, Beijing Wuzi University, Beijing 101149, China
  • 4Hebei YSUSOFT Co. Ltd., Qinhuangdao 066000, Hebei , China
  • show less
    Figures & Tables(14)
    PointPillars basic steps
    Overall architecture of column feature extraction network
    Overall architecture of 2D feature extraction network
    Flow chart of DBSCAN-based point cloud data
    Self attention mechanism
    Cross attention mechanism
    Attention mechanism based pillar feature extraction module
    Car and cyclist classification error
    Car and cyclist position error
    KITTI test set visualization results. (a) Test set 107; (b) test set 181
    • Table 1. Variation of network accuracy with different minimum point parameters

      View table

      Table 1. Variation of network accuracy with different minimum point parameters

      Modelcarcyclist
      easymoderatehardeasymoderatehard
      PointPillars82.4871.8467.3869.4555.4751.10
      DBSCAN-PointPillars(NMinPts=2)84.3673.1369.6269.6155.2251.09
      DBSCAN-PointPillars(NMinPts=3)83.0273.8069.7766.4352.7648.96
      DBSCAN-PointPillars(NMinPts=4)82.2471.5868.1168.0353.6450.19
    • Table 2. AOS of different modules on the KITTI validation set

      View table

      Table 2. AOS of different modules on the KITTI validation set

      Modelcarcyclist
      easymoderatehardeasymoderatehard
      PointPillars92.7088.6787.3573.4258.7054.73
      DBSCAN-PointPillars94.6389.0487.5078.9564.7060.47
      SACA-PointPillars94.5390.3587.7979.4764.8261.01
      PointPillars++95.1790.7388.0979.5565.2361.10
    • Table 3. Detection accuracy of PointPillars with different attention mechanisms on the KITTI validation set

      View table

      Table 3. Detection accuracy of PointPillars with different attention mechanisms on the KITTI validation set

      Modelcarcyclist
      easymoderatehardeasymoderatehard
      PointPillars82.4871.8467.3869.4555.4751.10
      SA-PointPillars84.3174.7670.7273.2455.9352.65
      SACA-PointPillars84.1474.8871.9274.4157.0453.24
    • Table 4. Results of ablation experiments with DBSCAN and SACA

      View table

      Table 4. Results of ablation experiments with DBSCAN and SACA

      ModleDBSCANSACAcarcyclist
      easymoderatehardeasymoderatehard
      PointPillars82.4871.8467.3869.4555.4751.10
      Experiment 184.3673.1369.6269.6155.2251.09
      Experiment 284.1474.8871.9274.4157.0453.24
      PointPillars++84.8975.3272.2575.7156.8752.74
    Tools

    Get Citation

    Copy Citation Text

    Qingxin Yang, Deming Kong, Jing Chen, Xiaowei Li, Yue Shen. Enhancing PointPillars Three-Dimensional Object Detection with Density Clustering and Dual Attention Mechanisms[J]. Laser & Optoelectronics Progress, 2025, 62(10): 1012001

    Download Citation

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

    Category: Instrumentation, Measurement and Metrology

    Received: Feb. 22, 2024

    Accepted: Apr. 26, 2024

    Published Online: Apr. 23, 2025

    The Author Email:

    DOI:10.3788/LOP240732

    CSTR:32186.14.LOP240732

    Topics