Fusion of Geometric Attention and Multi-Scale Feature Network for Point Cloud Registration

doi:10.3778/j.issn.1002-8331.2303-0301

Abstract

Abstract: The different geometric features of point clouds affect the difficulty of point cloud registration. However, most point clouds have partial overlap, geometric features are disturbed by noise, and there are some indistinguishable features on the point cloud surface, which makes it more difficult to extract representative geometric features. Combining the above reasons, this paper proposes a point cloud registration network GMNet that fuses geometric attention and multi-scale features, using geometric Transformer to extract geometric features and encode point cloud point pairs with distances and angles to make it more robust under low overlap. Using multi-scale feature architecture to aggregate abundant semantic information at different scales to improve the accuracy of point cloud registration. Finally the features are selected by consistent voting algorithm with the appropriate domain size. The experimental results show that the overall registration accuracy of GMNet is higher, and the registration recall is improved to 93.4% and 76.0% for 3DMatch and 3DLoMatch datasets respectively. The relative rotation error and relative translation error are reduced to 6.2 cm and 0.26° on the KITTI dataset, respectively. The geometric Transformer used in this method extracts representative geometric features and combines multi-scale features to learn different levels of geometric information in the point cloud, which effectively improves the point cloud registration accuracy.

Key words: point cloud registration, geometric Transformer, multi-scale feature, consistent voting algorithm

摘要： 点云不同的几何特征影响着点云配准的难度。然而，大多数点云存在部分重叠，几何特征被噪声干扰，并且点云表面有部分不可区分的特征，这使得提取具有代表性的几何特征更加困难。结合以上原因提出一种融合几何注意力和多尺度特征的点云配准网络GMNet，利用几何Transformer提取几何特征并对点云点对距离和角度编码，使其在低重叠情况下更具鲁棒性；并使用多尺度特征架构聚合不同尺度上丰富的语义信息，提高点云配准的准确率；最后特征通过一致决策算法选择具有适当邻域大小的特征。GMNet分别在室内数据集3DMatch、3DLoMatch和室外数据集KITTI上进行实验，实验结果表明GMNet的整体配准精度较高，在3DMatch数据集和3DLoMatch数据集上配准召回率分别提升到93.4%和76.0%，在KITTI数据集上相对旋转误差和相对平移误差分别降低到6.2?cm和0.26°。该方法使用的几何Transformer提取有代表性几何特征，联合多尺度特征学习点云中的不同层次几何信息，有效提升点云配准精度。

关键词: 点云配准, 几何Transformer, 多尺度特征, 一致决策算法

DU Jiajin, BAI Zhengyao, LIU Xuheng, LI Zekai, XIAO Xiao, YOU Yilin. Fusion of Geometric Attention and Multi-Scale Feature Network for Point Cloud Registration[J]. Computer Engineering and Applications, 2024, 60(12): 234-244.

杜佳锦, 柏正尧, 刘旭珩, 李泽锴, 肖霄, 尤逸琳. 融合几何注意力和多尺度特征点云配准网络[J]. 计算机工程与应用, 2024, 60(12): 234-244.

References

[1] CHIANG K W, TSAI G J, LI Y H, et al. Development of LiDAR-based UAV system for environment reconstruction[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(10): 1790-1794.
[2] 邱鑫, 郑明魁, 郭里婷, 等. 基于图卷积网络的三维点云配准研究[J]. 广播电视网络, 2022, 29(10): 124-128.
QIU X, ZHENG M K, GUO L T, et al. Research on 3D point cloud registration based on graph convolutional network[J]. Radio and Television Networks, 2022, 29(10): 124-128.
[3] 李建微, 占家旺. 三维点云配准方法研究进展[J]. 中国图象图形学报, 2022, 27(2): 349-367.
LI J W, ZHAN J W. Review on 3D point cloud registration method[J]. Journal of Image and Graphics, 2022, 27(2): 349-367.
[4] 秦红星, 刘镇涛, 谭博元. 深度学习刚性点云配准前沿进展[J]. 中国图象图形学报, 2022, 27(2): 329-348.
QIN H X, LIU Z T, TAN B Y. Research on deep learning rigid point cloud registration[J]. Journal of Image and Graphics, 2022, 27(2): 329-348.
[5] HEBERT M, JOHNSON A. Using spin images for efficient object recognition in cluttered 3D scenes[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1999, 21(5): 433-449.
[6] DENG H, BIRDAL T, ILIC S. Ppfnet: global context aware local features for robust 3d point matching[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, June 18-22, 2018. Piscataway, NJ: IEEE, 2018: 195-205.
[7] CHOY C, PARK J, KOLTUN V. Fully convolutional geometric features[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, October 27-November 2, 2019. Piscataway, NJ: IEEE, 2019: 8958-8966.
[8] BAI X Y, LUO Z X, ZHOU L, et al. D3feat: Joint learning of dense detection and description of 3d local features[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, June 14-19, 2020. Piscataway, NJ: IEEE, 2020: 6359-6367.
[9] HUANG S Y, GOJCIC Z, USVYATSOV M, et al. Predator: registration of 3D point clouds with low overlap[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 19-25, 2021. Piscataway, NJ: IEEE, 2021: 4267-4276.
[10] YU H, LI F, SALEH M, et al. Cofinet: reliable coarse-to-fine correspondences for robust pointcloud registration[C]//Advances in Neural Information Processing Systems, 2021: 23872-23884.
[11] THOMAS H, QI C R, DESCHAUD J E, et al. Kpconv: flexible and deformable convolution for point clouds[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Long Beach, June 16-20, 2019. Piscataway, NJ: IEEE, 2019: 6411-6420.
[12] WANG Y, SOLOMON J M. Deep closest point: learning representations for point cloud registration[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Long Beach, June 16-20, 2019. Piscataway, NJ: IEEE, 2019: 3523-3532.
[13] QIN Z, YU H, WANG C J, et al. Geometric transformer for fast and robust point cloud registration[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, June 19-24, 2022. Piscataway, NJ: IEEE, 2022: 11143-11152.
[14] GAO S H, CHENG M M, ZHAO K, et al. Res2net: a new multi-scale backbone architecture[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 43(2): 652-662.
[15] QIU S, ANWAR S, BARNES N. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 19-25, 2021. Piscataway, NJ: IEEE, 2021: 1757-1767.
[16] QIN C, YOU H X, WANG L C, et al. PointDAN: a multi-scale 3D domain adaption network for point cloud representation[C]//Advances in Neural Information Processing Systems, 2019: 7192-7203.
[17] KUANG H W, WANG B, AN J, et al. Voxel-FPN: multi-scale voxel feature aggregation for 3D object detection from LIDAR point clouds[J]. Sensors, 2020, 20(3): 704.
[18] LU F, CHEN G, LIU Y L, et al. Hregnet: a hierarchical network for large-scale outdoor lidar point cloud registration[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, June 19-25, 2021. Piscataway, NJ: IEEE, 2021: 16014-16023.
[19] HORACHE S, DESCHAUD J E, GOULETTE F. 3D point cloud registration with multi-scale architecture and unsupervised transfer learning[C]//2021 International Conference on 3D Vision (3DV), London, December 1-3, 2021: 1351-1361.
[20] YANG J C, ZHANG Q, NI B B, et al. Modeling point clouds with self-attention and gumbel subset sampling[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, June 16-20, 2019. Piscataway, NJ: IEEE, 2019: 3323-3332.
[21] YEW Z J, GIM H L. REGTR: end-to-end point cloud correspondences with Transformers[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, June 19-24, 2022. Piscataway, NJ: IEEE, 2022: 6677-6686.
[22] LI Y, HARADA T. Lepard: learning partial point cloud matching in rigid and deformable scenes[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, June 19-24, 2022. Piscataway, NJ: IEEE, 2022: 5554-5564.
[23] VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems, 2017.
[24] GOJCIC Z, ZHOU C F, WEGNER J D, et al. The perfect match: 3D point cloud matching with smoothed densities[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, June 16-20, 2019. Piscataway, NJ: IEEE, 2019: 5545-5554.
[25] AO S, HU Q Y, YANG B, et al. Spinnet: learning a general surface descriptor for 3D point cloud registration[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 19-25, 2021. Piscataway, NJ: IEEE, 2021: 11753-11762.
[26] WANG H P, LIU Y, DONG Z, et al. You only hypothesize once: point cloud registration with rotation-equivariant descriptors[C]//Proceedings of the 30th ACM International Conference on Multimedia, 2022: 1630-1641.
[27] YEW Z J, LEE G H. 3dfeat-net: weakly supervised local 3D features for point cloud registration[C]//Proceedings of the European Conference on Computer Vision (ECCV), Munich, September 8-14, 2018. Berlin: Springer, 2018: 607-623.