1.Intelligent Transportation Systems Research Center,Wuhan University of Technology,Wuhan 430063,China
2.State Key Laboratory of Maritime Technology and Safety,Wuhan University of Technology,Wuhan 430063,China
3.School of Transportation and Logistics Engineering,Wuhan University of Technology,Wuhan 430063,China. Corresponding author: CHEN De-shan,E-mail: dschen@whut. edu. cn
To address the problem of poor generalization and decreased accuracy when deep learning methods are utilized to handle partially overlapping point cloud registration, a point cloud registration algorithm with multi-scale feature fusion was proposed. First, feature descriptions of point clouds at different scales were performed to obtain local structural information, and features at different scales were fused through weighted summation to enhance the feature representation capability. Then, a multi-attention mechanism was used to strengthen the feature correlation between point clouds and discover more accurate correspondences. Finally, a position estimation module was employed to calculate the virtual corresponding points of the source point cloud, and singular value decomposition (SVD) was used to solve for the optimal transformation between the source point cloud and the virtual corresponding points. In the complete point cloud scenario of the ModelNet40 dataset, the mean square errors for rotation and translation are 6.3×10-2 and 1.0×10-5, respectively. In the partially overlapping point cloud scenario, the mean square errors for rotation and translation are 2.29×10-4 and 3×10-4, respectively, both of which are lower than those of the deep closest point (DCP) algorithm. Moreover, for unknown-category and noisy point clouds, the proposed method also yields lower errors than representative baseline methods such as iterative closest point (ICP), fast global registration (FGR), and DCP, further validating its effectiveness.
BeslP J, McKayN D. Method for registration of 3-D shapes[C]//Sensor Fusion IV: Control Paradigms and Data Structures.Boston:SPIE,1992: 586-606.
[2]
YangJ L, LiH D, CampbellD, et al. Go-ICP: a globally optimal solution to 3D ICP point-set registration[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(11): 2241-2254.
[3]
ChetverikovD, SvirkoD, StepanovD, et al. The trimmed iterative closest point algorithm[C]//2002 International Conference on Pattern Recognition. Quebec: IEEE, 2002: 545-548.
[4]
RusuR B, BlodowN, BeetzM. Fast point feature histograms (FPFH) for 3D registration[C]//2009 IEEE International Conference on Robotics and Automation. Rome: IEEE, 2009: 3212-3217.
[5]
ZhouQ Y, ParkJ, KoltunV. Fast global registration[C]//Computer Vision-ECCV 2016. Cham: Springer International Publishing, 2016: 766-782.
[6]
MagnussonM, LilienthalA, DuckettT. Scan registration for autonomous mining vehicles using 3D-NDT[J]. Journal of Field Robotics, 2007, 24(10): 803-827.
[7]
AigerD, MitraN J, Cohen-OrD. 4-points congruent sets for robust pairwise surface registration[C]// Special Interest Group on Computer Graphics and Interactive Techniques Conference. New York: Association for Computing Machinery,2008: 1-10.
[8]
CharlesR Q, HaoS, MoK C, et al. PointNet: deep learning on point sets for 3D classification and segmentation[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, 2017: 77-85.
[9]
AokiY, GoforthH, SrivatsanR A, et al. PointNetLK: robust & efficient point cloud registration using PointNet[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, 2020: 7156-7165.
[10]
SarodeV, LiX Q, GoforthH, et al. PCRNet: point cloud registration network using PointNet encoding[EB/OL]. (2019-08-21)[2023-10-12].
[11]
WangY, SolomonJ. Deep closest point: learning representations for point cloud registration[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul, 2020: 3522-3531.
[12]
WangY, SunY B, LiuZ W, et al. Dynamic graph CNN for learning on point clouds[J]. ACM Transactions on Graphics (TOG), 2018, 38: 1-12.
[13]
VaswaniA, ShazeerN, ParmarN, et al. Attention is all you need[C]//31st Conference on Neural Information Processing Systems (NIPS 2017). Long Beach, 2017:1-15.
[14]
WangY, SolomonJ M. PRNet: self-supervised learning for partial-to-partial registration[EB/OL]. (2019-10-27)[2021-01-03].
[15]
WuY, YaoQ L, FanX L, et al. PANet: a point-attention based multi-scale feature fusion network for point cloud registration[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 2512913.
[16]
GuoM H, CaiJ X, LiuZ N, et al. PCT: point cloud transformer[J]. Computational Visual Media, 2021, 7(2): 187-199.
[17]
WuY, HuX D, ZhangY, et al. SACF-net: skip-attention based correspondence filtering network for point cloud registration[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33(8): 3585-3595.
[18]
WuZ R, SongS R, KhoslaA, et al. 3D ShapeNets: a deep representation for volumetric shapes[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, 2015: 1912-1920.
[19]
EldarY, LindenbaumM, PoratM, et al. The farthest point strategy for progressive image sampling[J]. IEEE Transactions on Image Processing, 1997, 6(9): 1305-1315.
[20]
LiJ H, ZhangC H, XuZ Y, et al. Iterative distance-aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration[C]//Computer Vision—ECCV 2020. Cham: Springer, 2020: 378-394.
[21]
HezroniI, DroryA, GiryesR, et al. DeepBBS: deep best buddies for point cloud registration[C]//2021 International Conference on 3D Vision (3DV). London, 2022: 342-351.
[22]
ZengA, SongS R, NießnerM, et al. 3Dmatch: learning local geometric descriptors from RGB-D reconstructions[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, 2017: 199-208.