• Laser & Optoelectronics Progress
  • Vol. 57, Issue 6, 061002 (2020)
Zhen Peng1、2, Yuanjian Lü1、2, Chao Qu1、2, and Dahu Zhu1、2、*
Author Affiliations
  • 1Hubei Key Laboratory of Advanced Automotive Components Technology, Wuhan University of Technology, Wuhan, Hubei 430070, China
  • 2Hubei Collaborative Innovation Center for Automotive Components Technology, Wuhan University of Technology, Wuhan, Hubei 430070, China
  • show less
    DOI: 10.3788/LOP57.061002 Cite this Article Set citation alerts
    Zhen Peng, Yuanjian Lü, Chao Qu, Dahu Zhu. Accurate Registration of 3D Point Clouds Based on Keypoint Extraction and Improved Iterative Closest Point Algorithm[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061002 Copy Citation Text show less
    Flowchart of registration of point clouds P and Q
    Fig. 1. Flowchart of registration of point clouds P and Q
    Keypoint extraction. (a)Voxel grid filtering; (b) extracting keypoints using normal distance
    Fig. 2. Keypoint extraction. (a)Voxel grid filtering; (b) extracting keypoints using normal distance
    Distribution of keypoints under different parameters. (a) a0=0.3 mm, r=1.0 mm, thr=10%, m=5, the number of keypoints is 658; (b) a0=0.4 mm, r=2.0 mm, thr=10%, m=5, the number of keypoints is 597; (c) a0=0.4 mm, r=2.0 mm, thr=10%, m=10, the number of keypoints is 364
    Fig. 3. Distribution of keypoints under different parameters. (a) a0=0.3 mm, r=1.0 mm, thr=10%, m=5, the number of keypoints is 658; (b) a0=0.4 mm, r=2.0 mm, thr=10%, m=5, the number of keypoints is 597; (c) a0=0.4 mm, r=2.0 mm, thr=10%, m=10, the number of keypoints is 364
    Different nearest point models. (a) “Point to point” model; (b) “point to triangle plane” model
    Fig. 4. Different nearest point models. (a) “Point to point” model; (b) “point to triangle plane” model
    Coarse registration of model point clouds. (a) Feature matching; (b) correct correspondences after improved RANSAC; (c) results of coarse registration
    Fig. 5. Coarse registration of model point clouds. (a) Feature matching; (b) correct correspondences after improved RANSAC; (c) results of coarse registration
    Coarse registration of building point clouds. (a) Feature matching; (b) correct correspondences after improved RANSAC; (c) results of coarse registration
    Fig. 6. Coarse registration of building point clouds. (a) Feature matching; (b) correct correspondences after improved RANSAC; (c) results of coarse registration
    Fine registration of model point clouds. (a) Results of fine registration by proposed method; (b) chromatographic comparison of point clouds distance deviation under fine registration; (c) registration error comparison of fine registration among different methods
    Fig. 7. Fine registration of model point clouds. (a) Results of fine registration by proposed method; (b) chromatographic comparison of point clouds distance deviation under fine registration; (c) registration error comparison of fine registration among different methods
    Fine registration of building point clouds. (a) Results of fine registration by proposed method; (b) chromatographic comparison of point clouds distance deviation under fine registration; (c) registration error comparison of fine registration among different methods
    Fig. 8. Fine registration of building point clouds. (a) Results of fine registration by proposed method; (b) chromatographic comparison of point clouds distance deviation under fine registration; (c) registration error comparison of fine registration among different methods
    Registration experiment comparison of Gaussian noise point clouds under different methods. (a) Bunny; (b) happy; (c) armadillo
    Fig. 9. Registration experiment comparison of Gaussian noise point clouds under different methods. (a) Bunny; (b) happy; (c) armadillo
    Registration results of different point clouds with Gaussian noise σ=0.02 in proposed method. (a) Bunny; (b) happy; (c) armadillo
    Fig. 10. Registration results of different point clouds with Gaussian noise σ=0.02 in proposed method. (a) Bunny; (b) happy; (c) armadillo
    DatasetSize ofpoint cloudNumberof keypointsNumber ofcorrespondencesNumber of correctcorrespondensesRMS/mm
    Happy024Happy0487558269158433381102750.44
    Dragon120Dragon144218332353038241180531.04
    Armadillo15Armadillo45322082481340537197640.96
    Table 1. Coarse registration results of model point clouds
    MethodHappyArmadilloDragon
    Time /sRMS /mmTime /sRMS /mmTime /sRMS /mm
    Uniform+FPFH+SAC-IA42.902.066.401.644.481.14
    NARF+FPFH+SAC-IA28.672.1816.401.768.431.91
    ISS+FPFH+SAC-IA17.802.539.091.1510.171.02
    KFPCS6.701.293.281.032.711.09
    Proposed method1.230.441.800.960.611.04
    Table 2. Comparison of coarse registration results of model point clouds by different methods
    DatasetSize ofpoint cloudNumber ofkeypointsNumber ofcorrespondencesNumber of correctcorrespondencesRMS / (10-2 m)
    Dagstuhl000Dagstuhl0018135981360453404113462.07
    Hokuyo_0Hokuyo_1370261370277269532835651231.82
    Table 3. Coarse registration results of building point clouds
    MethodDagstuhlHokuyo
    Time /sRMS /mTime /sRMS /m
    Uniform+FPFH+SAC-IA27.200.041593.800.0267
    NARF+FPFH+SAC-IA4.330.036063.700.0439
    ISS+FPFH+SAC-IA12.040.023977.400.0206
    KFPCS5.230.024929.700.0279
    Proposed method0.720.020712.090.0182
    Table 4. Comparison of coarse registration results of building point clouds by different methods
    MethodHappyArmadilloDragon
    Time /sRMS /mmTime /sRMS /mmTime /sRMS /mm
    Standard ICP18.900.0827.600.1705.800.230
    GICP32.230.05515.100.09111.500.167
    LM-ICP20.240.0796.970.15010.170.183
    NDT5.300.0872.350.1501.690.180
    Proposed method11.20.0536.470.0844.700.173
    Table 5. Comparison of fine registration results of model point clouds under different methods
    MethodDagstuhlHokuyo
    Time /sRMS /(10-3 m)Time /sRMS /(10-3 m)
    Standard ICP15.704.77105.62.82
    GICP23.974.31121.72.43
    LM-ICP60.604.52203.64.56
    NDT10.675.7349.62.57
    Proposed method13.603.5870.91.61
    Table 6. Comparison of fine registration results of building point clouds by different methods
    Zhen Peng, Yuanjian Lü, Chao Qu, Dahu Zhu. Accurate Registration of 3D Point Clouds Based on Keypoint Extraction and Improved Iterative Closest Point Algorithm[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061002
    Download Citation