FU Luhua1,2， SUN Yujing1， SUN Changku1， WANG Peng1,2， ZHANG Baoshang2
（1. State Key Laboratory of Precision Measuring Technology and Instruments， Tianjin University， Tianjin 300072， China；
2. Science and Technology on Electro-Optic Control Laboratory, Luoyang Institute of Electro-Optic Equiment,
Aviation Industry Corporation of China, Luoyang 471009, China）
Abstract： The point cloud registration technique and 3D point cloud measurement technology without artificial marks are used to demonstrate a pose measurement solution in this study. The cross-source point cloud registration accuracy is relatively low due to the large density difference, especially when the measurement point cloud partially overlaps with the reference point cloud, and the typical point cloud registration procedure is susceptible to the initial position and noise. A deep learning-based end-to-end pose estimation algorithm is proposed in order to address the aforementioned issues. First, a point cloud hybrid feature coding module is created to combine the features of the source and reference point clouds through feature interaction, resulting in a richer feature representation. Second, based on the hybrid features, an overlapping mask decoding module is utilized to predict and sample all the overlapping points in the reference point cloud. Finally, using the feature of points in the overlapping area, a pose regression module is utilized to estimate the relative pose parameters of the two groups of point clouds. Experimental results show that the proposed method can significantly increase point cloud registration accuracy and has higher robustness against noise interference.
Key words： pose measurement； point cloud registration； deep learning； overlapping area
References
［1］ YANG Z, LI C. Review on vision-based pose estimation of UAV based on landmark//2nd International Conference on Frontiers of Sensors Technologies, Apr.14-16, 2017, Shenzhen, China. New York: IEEE, 2017: 453-457.
［2］ YANG J, LI H, CAMPBELL D, et al. Go-ICP: a globally optimal solution to 3D ICP point-set registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(11)： 2241-2254.
［3］ ROSEN D M, CARLONE L, BANDEIRA A S, et al. A certifiably correct algorithm for synchronization over the special euclidean group. The International Journal of Robotics Research, 2019, 38(2/3)： 95-125.
［4］ MARON H, DYM N, KEZURER I, et al. Point registration via efficient convex relaxation. ACM Transactions on Graphics, 2016, 35(4)： 73.
［5］ IZATT G, DAI H, TEDRAKE R. Globally optimal object pose estimation in point clouds with mixed-integer programming//The 18th International Symposium on Robotics Research, Online. Berlin: Springer, 2020： 695-710.
［6］ QIN H, LIU Z, TAN B. Review on deep learning rigid point cloud registration. Journal of Image and Graphics, 2022, 27(2)： 329-348.
［7］ WANG Y, SOLOMON J. Deep closest point： learning representations for point cloud registration//IEEE/CVF International Conference on Computer Vision, Seoul, Korea (South). New York: IEEE, 2019： 3522-3531.
［8］ WANG Y, SOLOMON J M. PRNet： self-supervised learning for partial-to-partial registration. In： 33rd Conference on Neural Information Processing Systems (To appear), 2019.
［9］ JANG E, GU S, POOLE B. Categorical reparameterization with Gumbel-softmax//5th International Conference on Learning Representations, Apr. 24-26, 2017, Toulon, France. 2017: 1-13.
［10］ YEW Z J, LEE G H. RPM-Net： robust point matching using learned features//IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun. 14-19, 2020, Seattle, WA, USA. New York: IEEE, 2020： 11821-11830.
［11］ SINKHORN R. A relationship between arbitrary positive matrices and doubly stochastic matrices. Canadian Journal of Mathematics, 1964, 18: 303-306.
［12］ LI J, ZHANG C, XU Z, et al. Iterative distance-aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration//16th European Conference on Computer Vision, Aug. 23-28, 2020, Glasgow, UK. Zurich: ECCV, 2020： 378-394.
［13］ AOKI Y, GOFORTH H, SRIVATSAN R A, et al. PointNetLK： robust & efficient point cloud registration using PointNet//IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun. 15-20, 2019, Long Beach, CA, USA. New York: IEEE, 2019： 7156-7165.
［14］ SARODE V, LI X, GOFORTH H, et al. PCRNet： point cloud registration network using PointNet encoding. (2019-10-04)［2022-08-28］. https://doi.org/10.48850/arXiv.1908.07906.
［15］ HELD D, THRUN S, SAVARSES S. Learning to track at 100 FPS with deep regression networks//14th European Conference on Computer Vision, Oct. 8-10, 2016, Amsterdam, The Netherlands. Zurich: ECCV, 2016： 749-765.
［16］ LI X X, SUN C K, SUN Y J, et al. Multi-constrained speckle propagation matching method combined with descriptor information. Laser & Optoelectronics Progress, 2023, 60(6)： 182-189.
基于混合特征重合区域预测的无标识点位姿测量
付鲁华1,2， 孙玉晶1， 孙长库1 ， 王鹏1,2， 张宝尚2
（1. 天津大学 精密测试技术及仪器国家重点实验室， 天津 300072； 2. 中航工业洛阳电光设备研究所 光电控制技术重点实验室， 河南 洛阳 471009）
摘要：本文提出了一种结合三维点云测量技术与点云配准算法的无标识点位姿测量方案。 针对传统点云配准算法对初始位置和噪声敏感， 目标测量点云与参考点云部分重合以及密度差异大导致的异源点云配准精度低的问题， 提出了一种基于混合特征重合区域预测的端到端位姿估计算法， 即通过深度学习的方法实现点云配准。 首先， 设计点云混合特征编码模块， 通过特征交互融合源点云与参考点云的特征， 获得更丰富的特征表示。 其次， 基于混合特征设计参考点云重合掩码解码模块， 保留参考点云中与源点云重合的所有重合点。 最后， 利用重合部分的点云特征回归两组点云的相对位姿参数， 实现了目标的位姿测量。 实验结果表明， 该算法对异源点云配准具有较好的性能， 能在保证鲁棒性的同时， 显著提升点云配准精度。
关键词：位姿测量； 点云配准； 深度学习； 重合区域
引用格式：FU Luhua， SUN Yujing， SUN Changku, et al． Pose measurement without marked points based on prediction of overlapping area with hybrid features． Journal of Measurement Science and Instrumentation， 2023， 14（3）： 253-262. DOI： 10．3969／j．issn．1674-8042．2023．03．001
[full text view]
