基于逐点前进法的改进型点云配准方法

李茂月; 许圣博; 孟令强; 刘志诚

doi:10.37188/CO.2023-0166

基于逐点前进法的改进型点云配准方法

doi: 10.37188/CO.2023-0166

哈尔滨理工大学先进制造智能化技术教育部重点实验室, 黑龙江哈尔滨150080

基金项目: 国家自然科学基金（No. 51975169），黑龙江省自然科学基金（No. LH2022E085）资助项目

详细信息

中图分类号: TP394.1
计量
- 文章访问数: 114
- HTML全文浏览量: 59
- PDF下载量: 20
- 被引次数: 0
出版历程
- 网络出版日期: 2023-11-07

An improved point cloud registration method based on the point-by-point forward method

The Key Laboratory of Advanced Manufacturing and Intelligent Technology of the Ministry of Education, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, China

Funds: Supported by National Natural Science Foundation of China (No. 51975169), Natural Science Foundation of Heilongjiang Province (No. LH2022E085)

More Information

Corresponding author: lmy0500@163.com

摘要

摘要:
点云配准是获取三维点云模型空间姿态的关键步骤，为了进一步提高点云配准的效率和准确性，提出了一种基于逐点前进法特征点提取的改进型点云配准方法。首先，利用逐点前进法快速提取点云特征点，在保留点云模型特征的同时大幅精简点云数量。然后，通过使用法向量约束改进的KN-4PCS算法进行粗配准，以实现源点云与目标点云的初步配准。最后，使用双向Kd-tree优化的LM-ICP算法完成精配准。实验结果表明，本文方法具有较高的精度和效率，同时具有较好的鲁棒性，在斯坦福大学开放点云数据配准实验中，其平均误差较SAC-IA+ICP算法减少了约70.2％，较NDT+ICP算法减少了约49.6％，配准耗时分别减少约86.2%和81.9%，同时在引入不同程度的高斯噪声后仍能保持较高的精度和较低的耗时。在真实室内物体点云配准实验中，其平均配准误差为0.0742 mm，算法耗时平均为0.572 s。通过斯坦福开放数据与真实室内场景物体点云数据对比分析表明，本方法能够有效地提高点云配准的效率、准确性和鲁棒性，为基于点云的室内目标识别与位姿估计奠定了良好的基础。
- 点云配准 /
- KN-4PCS /
- 双向Kd-tree /
- LM-ICP
Abstract:
To improve both the efficiency and accuracy of point cloud registration, this study proposed an improved method based on point-by-point advance feature point extraction. Firstly, the point-by-point advance method extracts point cloud feature points rapidly, and greatly reduces the number of point clouds, while retaining the characteristics of the point cloud model. The KN-4PCS algorithm, using normal vector constraints, conducts a preliminary registration of the source and target point cloud. Finally, the fine registration is achieved with the two-way Kd-tree optimized LM-ICP algorithm. In the open point cloud data registration experiment of Stanford University, the average error is reduced by about 70.2% compared with the SAC-IA+ICP algorithm, and the registration time is reduced by about 86.2% and 81.9%, respectively. The algorithm maintains high accuracy and low time consumption even with varying degrees of Gaussian noise. In the point cloud registration experiment of indoor objects, the average registration error was measured to be 0.0742 mm with an average algorithm time of 0.572 s. The comparison and analysis of Stanford open data and real indoor scene object point cloud data shows that this method can effectively improve the efficiency, accuracy, and robustness of point cloud registration. Furthermore, this study establishes a strong foundation for indoor target recognition and pose estimation through the point cloud.
- point cloud registration /
- KN-4PCS /
- bidirectional Kd-tree /
- LM-ICP

HTML全文

图 1 共面四点集的仿射不变量

Figure 1. Affine invariant ratio for congruent 4-points

下载: 全尺寸图片幻灯片

图 2 仿射不变四点集提取示意图

Figure 2. Diagram of affine invariant congruent 4-points extraction

下载: 全尺寸图片幻灯片

图 3 双向Kd-tree 点对搜索示意图

Figure 3. Diagram of bidirectional Kd-tree point pair search

下载: 全尺寸图片幻灯片

图 4 点云配准方法总流程图

Figure 4. Flowchart of the point cloud registration algorithm

下载: 全尺寸图片幻灯片

图 5 点云数据采集系统

Figure 5. Point cloud data acquisition system

下载: 全尺寸图片幻灯片

图 6 斯坦福点云配准结果

Figure 6. Registration results of Stanford point cloud data

下载: 全尺寸图片幻灯片

图 7 高斯噪声σ=0.003时斯坦福点云的配准结果

Figure 7. Registration results of Stanford point cloud data at gaussian noise of σ= 0.003

下载: 全尺寸图片幻灯片

图 8 不同高斯噪声时各算法的配准误差

Figure 8. Registration error of each algorithm with different gaussian noise

下载: 全尺寸图片幻灯片

图 9 不同高斯噪声时各算法的配准耗时

Figure 9. The registration time consumed by each algorithm under different gaussian noise

下载: 全尺寸图片幻灯片

图 10 真实室内场景物体点云配准结果

Figure 10. Object point cloud data registration results in indoor scene

下载: 全尺寸图片幻灯片

表 1 深度相机参数

Table 1. Deep camera parameters

	参数名称	数值
	工作范围	0.6-8 w
深度	精度	lm: ±3 mm
	视场角（FOV)	H58.4 × V45.5°
	分辨率@帧率	640 × 480@30 fps
	视场角（FOV)	H66.1° × 740.2°
RGB	分辨率@帧率	640 × 480@30 fps
	UVC	支持

下载: 导出CSV

表 2 斯坦福点云配准定量分析结果

Table 2. Quantitative analysis results of Stanford point cloud data registration

Model	SAC-IA+ICP		NDT+ICP		Ours
Model	RMSE/mm	Time/s	RMSE/mm	Time/s	RMSE/mm	Time/s
Armadillo	0.0288	3.21	0.0168	2.19	0.00730	0.437
Dragon	0.0363	3.77	0.0218	3.25	0.0125	0.528

下载: 导出CSV

表 3 真实室内场景物体点云配准定量分析结果

Table 3. Quantitative analysis results of object point cloud data registration in indoor scene

模型	源点云数/个	源点云特征点数/个	目标点云数/个	目标点云特征点数/个	RMSE/mm	平均误差/mm	耗时/s	平均耗时/s
Chair	8549	3702	10431	4847	0.0686		0.544
Kettle	3713	1359	3879	1449	0.0633	0.0742	0.487	0.572
Mannequin	45292	16098	46507	16462	0.0907		0.686

下载: 导出CSV

参考文献(20)

[1]	杨鹏程, 杨朝, 孟杰, 等. 基于法向量和面状指数特征的文物点云棱界配准方法[J]. 中国光学(中英文),2023,16(3):654-662. doi: 10.37188/CO.2022-0156 YANG P CH, YANG ZH, MENG J, et al. Aligning method for point cloud prism boundaries of cultural relics based on normal vector and faceted index features[J]. Chinese Optics, 2023, 16(3): 654-662. (in Chinese). doi: 10.37188/CO.2022-0156
[2]	ZOU ZH, LANG H, LOU Y X, et al. Plane-based global registration for pavement 3D reconstruction using hybrid solid-state LiDAR point cloud[J]. Automation in Construction, 2023, 152: 104907. doi: 10.1016/j.autcon.2023.104907
[3]	CAO T, LIU Y P, DAI J J, et al. Local scanned point-cloud registration of aeroengine pipeline system based on axial constraints[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 2527011. doi: 10.1109/TIM.2023.3256460
[4]	DENG Y H, CHEN G H, LIU X L, et al. 3D pose recognition of small special-shaped sheet metal with multi-objective overlapping[J]. Electronics, 2023, 12(12): 2613. doi: 10.3390/electronics12122613
[5]	SHI X Y, PENG J J, LI J P, et al. The iterative closest point registration algorithm based on the normal distribution transformation[J]. Procedia Computer Science, 2019, 147: 181-190. doi: 10.1016/j.procs.2019.01.219
[6]	AIGER D, MITRA N J, COHEN-OR D. 4-points congruent sets for robust pairwise surface registration[J]. ACM Transactions on Graphics, 2008, 27(3): 1-10.
[7]	傅瑶, 陈鹏, 郭贵松, 等. 基于4PCS和SICP的点云配准方法在钢轨磨耗计算中的应用[J]. 电子测量与仪器学报,2022,36(12):210-218. doi: 10.13382/j.jemi.B2105036 FU Y, CHEN P, GUO G S, et al. Application of the point cloud registration method based on 4PCS and SICP in rail wear calculation[J]. Journal of Electronic Measurement and Instrumentation, 2022, 36(12): 210-218. (in Chinese). doi: 10.13382/j.jemi.B2105036
[8]	THEILER P W, WEGNER J D, SCHINDLER K. Keypoint-based 4-points congruent sets–automated marker-less registration of laser scans[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2014, 96: 149-163. doi: 10.1016/j.isprsjprs.2014.06.015
[9]	TAO W Y, LIU J B, XU D, et al. Automatic registration of point clouds by combining local shape descriptor and G4PCS algorithm[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 16: 6339-6351. doi: 10.1109/JSTARS.2023.3293409
[10]	BESL P J, MCKAY N D. Method for registration of 3-D shapes[J]. Proceedings of SPIE, 1992, 1611: 586-606. doi: 10.1117/12.57955
[11]	林森, 张强. 应用邻域点信息描述与匹配的点云配准[J]. 光学精密工程,2022,30(8):984-997. doi: 10.37188/OPE.20223008.0984 LIN S, ZHANG Q. Point cloud registration using neighborhood point information description and matching[J]. Optics and Precision Engineering, 2022, 30(8): 984-997. (in Chinese). doi: 10.37188/OPE.20223008.0984
[12]	YAO Z W, ZHAO Q X, LI X F, et al. Point cloud registration algorithm based on curvature feature similarity[J]. Measurement, 2021, 177: 109274. doi: 10.1016/j.measurement.2021.109274
[13]	LV CH L, LIN W S, ZHAO B Q. KSS-ICP: point cloud registration based on Kendall shape space[J]. IEEE Transactions on Image Processing, 2023, 32: 1681-1693. doi: 10.1109/TIP.2023.3251021
[14]	LIU Y SH, CHEN X D, WANG K F, et al. A Chebyshev metamodel based BnB approach to efficiently search global optimum for 3D ICP point set registration[J]. Computer Aided Geometric Design, 2023, 101: 102178. doi: 10.1016/j.cagd.2023.102178
[15]	XIE L F, ZHU Y Y, YIN M, et al. Self-feature-based point cloud registration method with a novel convolutional Siamese point net for optical measurement of blade profile[J]. Mechanical Systems and Signal Processing, 2022, 178: 109243. doi: 10.1016/j.ymssp.2022.109243
[16]	ZHANG ZH Y, SUN J D, DAI Y CH, et al. Self-supervised rigid transformation equivariance for accurate 3D point cloud registration[J]. Pattern Recognition, 2022, 130: 108784. doi: 10.1016/j.patcog.2022.108784
[17]	何宽, 孙瑞, 官云兰, 等. 基于逐点前进法的点云数据精简[J]. 测绘通报,2022(9):167-169. HE K, SUN R, GUAN Y L, et al. Point cloud data simplification based on point-by-point advancing method[J]. Bulletin of Surveying and Mapping, 2022(9): 167-169. (in Chinese).
[18]	李茂月, 赵伟翔, 马康盛, 等. 结构光检测点云精简与重构参数自动调节方法[J]. 仪器仪表学报,2022,43(8):122-130. LI M Y, ZHAO W X, MA K SH, et al. Point cloud simplification and reconstruction parameters’ automatic adjustment method of structured light detection[J]. Chinese Journal of Scientific Instrument, 2022, 43(8): 122-130. (in Chinese).
[19]	FITZGIBBON A W. Robust registration of 2D and 3D point sets[J]. Image and Vision Computing, 2003, 21(13-14): 1145-1153. doi: 10.1016/j.imavis.2003.09.004
[20]	王宾, 刘林, 侯榆青, 等. 应用改进迭代最近点方法的三维心脏点云配准[J]. 光学精密工程,2020,28(2):474-484. WANG B, LIU L, HOU Y Q, et al. Three-dimensional cardiac point cloud registration by improved iterative closest point method[J]. Optics and Precision Engineering, 2020, 28(2): 474-484. (in Chinese).