面向大型集装箱港区三维重建的无人机点云切片SLAM

胡钊政; 左志航; 许聪; 陶倩文; 刘超; 孟杰

doi:10.11999/JEIT251112

面向大型集装箱港区三维重建的无人机点云切片SLAM

doi: 10.11999/JEIT251112 cstr: 32379.14.JEIT251112

胡钊政¹,
左志航^{1, 2},
许聪¹,
陶倩文³,
刘超⁴,
孟杰^1, ,

1.
武汉理工大学智能交通系统研究中心武汉 430070
2.
武汉理工大学信息工程学院武汉 430063
3.
南京林业大学汽车与交通工程学院南京 210037
4.
中国工程物理研究院机械制造工艺研究所绵阳 621022

基金项目: 国家自然科学基金项目(52472453)，武汉市科技局科技成果转化项目(2024030803010173)

详细信息

作者简介:
胡钊政：男，教授，研究方向为3D计算机视觉、智能网联汽车、机器人定位与导航、智能车路协同

左志航：男，硕士生，研究方向为激光SLAM定位、移动机器人等

许聪：男，博士生，研究方向为机器人定位与感知、BEV多模态大模型等

陶倩文：女，副教授，研究方向为3D计算机视觉、机器人定位与导航等

刘超：男，助理研究员，研究方向为特种机器人，机器人导航定位等

孟杰：男，副研究员，研究方向为移动机器人、自动驾驶、自主导航定位等

通讯作者:
孟杰　mengjie09@whut.edu.cn

中图分类号: TN249; TP242
计量
- 文章访问数: 19
- HTML全文浏览量: 9
- PDF下载量: 3
- 被引次数: 0
出版历程
- 收稿日期: 2025-10-22
- 修回日期: 2026-04-29
- 录用日期: 2026-05-12
- 网络出版日期: 2026-05-31

A Point Cloud Slice-based UAV SLAM for 3D Reconstruction of Large Container Port Areas

HU Zhaozheng¹,
ZUO Zhihang^{1, 2},
XU Cong¹,
TAO Qianwen³,
LIU Chao⁴,
MENG Jie^{1
, ,}

1.
Intelligent Transport System Research Center, Wuhan University of Technology, Wuhan 430070, China
2.
School of Information Engineering, Wuhan University of Technology, Wuhan 430063, China
3.
College of Automobile and Traffic Engineering, Nanjing Forestry University, Nanjing 210037, China
4.
Institute of Mechanical Manufacturing Technology, China Academy of Engineering Physics, Mianyang 621900, China

Funds: National Natural Science Foundation of China (52472453), Wuhan Science and Technology Achievement Transformation Project (2024030803010173)

摘要

摘要: 在大型集装箱港区堆场环境中，大量重复性语义特征以及部分退化场景导致无人机在大面积场景下难以实现高效可靠的三维重建。为此，该文提出一种基于无人机点云切片SLAM的大型集装箱港区三维重建方法，基于地面约束与点云密度梯度变化自适应提取多层切片点云，高效精准获取堆场语义信息，并基于同层切片点云匹配有效改善了里程计以及回环检测的精度。首先提出一种面向快速特征提取的点云切片方法，通过快速提取主方向并将点云划分为多层切片，高效获取多层语义点云。其次基于集装箱堆场场景特点进一步优化切片提取方法，基于重力方向简化场景主平面提取过程，并通过点云梯度变化自适应获取各层集装箱所在高程区间，构建多层切片点云。然后构建基于切片点云的递进式自适应激光里程计，利用高程切片自适应判别退化场景，同时在层间利用增量式迭代策略实现切片融合匹配，从而提升激光里程计精度、效率与稳定性。此外，设计融合激光点云切片信息的因子图优化方法，通过对多层切片点云匹配结果进行融合投票，筛除错误结果并减少大量重复结构对回环检测的影响，并利用切片因子来构建因子图边，从而提升全局优化水平，实现高效稳定的三维重建。最后，通过Carla仿真场景以及武汉某大型集装箱港区的实际场景测试，证实了该方法的可行性和有效性。
- 无人机 /
- 大型集装箱港区 /
- 多层切片匹配 /
- 里程计 /
- 回环检测
Abstract: Objective With the continuous advancement of port intelligence, the demand for digital management in container port areas is increasingly growing. In large container yard scenarios, 3D reconstruction of the yard environment can be achieved by utilizing drone Simultaneous Localization and Mapping (SLAM) technology. However, container port areas contain an abundance of repetitive semantic structural information, where traditional semantic matching methods suffer from low efficiency and poor accuracy. Furthermore, during the 3D reconstruction process conducted by drones over container port areas, the lanes between yards present large feature-sparse regions, which can easily lead to odometry degradation. Additionally, the extensive presence of repetitive scene features also interferes with loop closure detection. To address these issues, this paper proposes a slicing method for rapid feature extraction, which is further optimized based on the characteristics of the container yard scenario. A UAV point cloud slicing SLAM method tailored for large-scale container port 3D reconstruction is introduced, enabling high-precision 3D reconstruction. Methods To address point cloud semantic extraction, this paper proposes a point cloud slicing method for rapid feature extraction, which quickly extracts the principal direction and divides the point cloud into multiple layers to efficiently obtain multi-layer semantic point clouds. The slicing method is further optimized based on the characteristics of the container yard scenario: the principal plane extraction is simplified using the direction of gravity, and the elevation range of each container layer is adaptively obtained through point cloud gradient changes to construct multi-layer sliced point clouds. Subsequently, a progressive adaptive LiDAR odometry based on sliced point clouds is constructed, which adaptively identifies degraded scenarios using elevation slices and employs an incremental iterative strategy for inter-layer slice fusion matching, thereby improving the accuracy, efficiency, and stability of the LiDAR odometry. In addition, a factor graph optimization method that fuses information from sliced point clouds is designed. By performing fusion voting on the matching results of multi-layer sliced point clouds, erroneous results are filtered out and the impact of repetitive structures on loop closure detection is reduced; slice factors are then used to construct factor graph edges, enhancing global optimization and achieving efficient and stable 3D reconstruction. Results and Discussions The feasibility and effectiveness of the proposed method are verified through testing in Carla simulations and real-world scenarios at a large container port in Wuhan. Results are as follows: First, through comparative analysis with three algorithms—RANSAC, Region Growth, and 3DG_SEG—the efficiency and accuracy of the proposed semantic extraction algorithm are demonstrated. Furthermore, by comparing mapping trajectories with two renowned open-source LiDAR algorithms, FAST-LIO2 and Faster-LIO, the superiority of the proposed odometry method is proven. Finally, comparisons of speed and confidence level are conducted with six algorithms: ICP, NDT, GICP, Fast_GICP, Scan Context+ICP, and Quatro. Simultaneously, the loop closure detection module from LIO-SAM is integrated into FAST-LIO2, and the Scan Context module into Faster-LIO. The mapping trajectories are then compared with that of the proposed algorithm, validating the effectiveness of the proposed loop closure detection algorithm. The proposed method achieves high 3D reconstruction accuracy; therefore, it is suitable for practical application in operational processes. Conclusions The proposed method uses an efficient point cloud slicing technique and a multi-layer slice matching mechanism. Points within the same elevation range form a slice point cloud (Slice), and the segmentation process is called slice generation. This enables efficient and robust 3D reconstruction in large-scale scenes with repetitive features.First, the LiDAR point cloud is aligned to the Z-axis using IMU-derived gravity direction. A sliding window records density gradient changes to adaptively determine each layer’s elevation range. This simplifies slicing and reduces the impact of non-standard containers or ground height variations on semantic extraction.Multi-layer slice data are then integrated into the odometry module to detect degenerate scenarios. Under normal conditions, progressive slice matching initializes pose estimation; otherwise, IMU-based iterative Kalman filtering is used.Finally, fusion voting removes outliers from multi-layer slice matching results. The best match initializes loop closure for global container point cloud registration, enabling dual-stage loop closure detection and slice factor construction. Integrating slice point cloud information into factor graph optimization unifies coordinates and achieves efficient, robust 3D reconstruction.
- UAV /
- Large Container Port Areas /
- Multi-layer Slice Matching /
- Odometry /
- Loop Closure Detection

HTML全文

图 1 Slice-SLAM系统框图

下载: 全尺寸图片幻灯片

图 2 切片原理示意图

下载: 全尺寸图片幻灯片

图 3 集装箱堆场切片点云提取

下载: 全尺寸图片幻灯片

图 4 点云密度梯度结构示例

下载: 全尺寸图片幻灯片

图 5 里程计示意图

下载: 全尺寸图片幻灯片

图 6 回环检测示意图

下载: 全尺寸图片幻灯片

图 7 因子图示意图

下载: 全尺寸图片幻灯片

图 8 集装箱港区仿真场景

下载: 全尺寸图片幻灯片

图 9 真实实验环境与平台

下载: 全尺寸图片幻灯片

图 10 仿真场景下各算法轨迹误差对比

下载: 全尺寸图片幻灯片

图 11 语义提取结果

下载: 全尺寸图片幻灯片

图 12 真实场景下各算法轨迹误差

下载: 全尺寸图片幻灯片

图 13 局部放大对比图

下载: 全尺寸图片幻灯片

表 1 实验参数表

参数	$ \boldsymbol{\lambda } $	$ \alpha $	$ \beta $	$ {\delta }_{t} $	$ {\delta }_{\theta } $	飞行高度	$ \tau $
数值	0.85	0.6	0.4	0.02	0.017	100米	0.04

下载: 导出CSV

表 4 语义提取运行效率

方法	单帧帧率(FPS)	五帧帧率(FPS)	十帧帧率(FPS)
RANSAC	9.215	8.456	7.450
Region Growth	1.993	0.722	0.345
3DG_SEG	3.425	1.276	0.309
Ours	206.267	95.950	32.660

下载: 导出CSV

表 2 语义提取运行效率

方法	单帧帧率(FPS)	三帧帧率(FPS)	五帧帧率(FPS)
RANSAC	14.388	11.990	10.010
Region Growth	5.376	1.001	0.529
3DG_SEG	4.274	1.458	0.740
Ours	211	112.104	62.893

下载: 导出CSV

表 3 轨迹精度对比

场景	指标	方法
场景	指标	FAST-LIO2	Faster-LIO	Slice-SLAM (无回环)	FAST-LIO-SAM	Faster-LIO-SC	Slice-SLAM (有回环)
仿真场景	MAXE(m)	29.043	15.884	13.025	6.487	3.502	3.173
	MAE(m)	2.625	1.999	1.672	1.846	1.472	1.401
	RSME(m)	3.842	1.445	1.131	0.730	0.656	0.632
	HME(m)	1.181	0.646	0.530	0.264	0.143	0.129
真实场景	MAXE(m)	35.694	26.014	17.234	28.008	23.501	10.812
	MAE(m)	13.598	8.808	6.268	10.261	8.137	5.215
	RSME(m)	7.185	6.581	3.753	5.010	5.855	2.443
	HME(m)	0.972	0.708	0.469	0.763	0.639	0.294

下载: 导出CSV

表 5 回环检测结果对比

指标		方法
指标		ICP	NDT	GICP	Fast_GICP	Quatro	Scan_Context + ICP	Ours
0～2 m	运行时间(s)	0.817271	0.735771	0.364556	0.336739	0.166818	0.422534	0.245286
0～2 m	置信分数	0.113421	0.513051	0.107523	0.115033	2.470698	0.111862	0.108105
4～6 m	运行时间(s)	0.760850	0.961706	0.379740	0.374925	0.169581	0.453751	0.250143
4～6 m	置信分数	0.129462	0.452106	0.117008	0.112414	2.587779	0.116497	0.110242
8～10 m	运行时间(s)	0.783931	0.863472	0.448796	0.333281	0.174932	0.465473	0.264625
8～10 m	置信分数	0.128523	0.581624	0.114952	0.114824	2.909427	0.114505	0.113278

下载: 导出CSV

参考文献(28)

[1]	PARASKEVAS A, MADAS M, ZEIMPEKIS V, et al. Smart ports in industry 4.0: A systematic literature review[J]. Logistics, 2024, 8(1): 28. doi: 10.3390/logistics8010028.
[2]	SVISHCHEV N, LINO P, MAIONE G, et al. A comprehensive survey of advanced SLAM techniques[C]. E3S Web of Conferences, Almaty, Kazakhstan, 2024: 05004. doi: 10.1051/e3sconf/202454105004. (查阅网上资料,未找到本条文献出版地信息,请确认).
[3]	胡钊政, 王书恒, 孟杰, 等. 基于路径似然模型与HMM序列匹配定位的地铁隧道三维重建[J]. 电子与信息学报, 2025, 47(7): 2273–2284. doi: 10.11999/JEIT241122. HU Zhaozheng, WANG Shuheng, MENG Jie, et al. 3D reconstruction of metro tunnel based on path likelihood model and HMM sequence matching localization[J]. Journal of Electronics & Information Technology, 2025, 47(7): 2273–2284. doi: 10.11999/JEIT241122.
[4]	AMJAD M, SAHIN ALI M, YAO Shouwen, et al. Self and target locating with cooperation of heterogeneous unmanned vehicles in the denial environment[J]. IEEE Access, 2025, 13: 64699–64718. doi: 10.1109/ACCESS.2025.3558873.
[5]	CHEN Shoubin, ZHOU Baoding, JIANG Changhui, et al. A LiDAR/visual SLAM backend with loop closure detection and graph optimization[J]. Remote Sensing, 2021, 13(14): 2720. doi: 10.3390/rs13142720.
[6]	GAO Ruizhen, LI Yuang, LI Baihua, et al. FELC-SLAM: Feature extraction and loop closure optimized LiDAR SLAM system[J]. Measurement Science and Technology, 2024, 35(11): 115112. doi: 10.1088/1361-6501/ad6e0c.
[7]	余浩扬, 李艳生, 肖凌励, 等. 面向动态环境的巡检机器人轻量级语义视觉SLAM框架[J]. 电子与信息学报, 2025, 47(10): 3979–3992. doi: 10.11999/JEIT250301. YU Haoyang, LI Yansheng, XIAO Lingli, et al. A lightweight semantic visual simultaneous localization and mapping framework for inspection robots in dynamic environments[J]. Journal of Electronics & Information Technology, 2025, 47(10): 3979–3992. doi: 10.11999/JEIT250301.
[8]	CHOI S, CHAE H W, JEUNG Y, et al. Fast and versatile feature-based LiDAR odometry via efficient local quadratic surface approximation[J]. IEEE Robotics and Automation Letters, 2023, 8(2): 640–647. doi: 10.1109/LRA.2022.3227875.
[9]	QIAO Zhijian, YU Zehuan, JIANG Binqian, et al. G3Reg: Pyramid graph-based global registration using Gaussian ellipsoid model[J]. IEEE Transactions on Automation Science and Engineering, 2025, 22: 3416–3432. doi: 10.1109/TASE.2024.3394519.
[10]	YAN Shaocheng, SHI Pengcheng, and LI Jiayuan. ML-SemReg: Boosting point cloud registration with multi-level semantic consistency[C]. Proceedings of 18th European Conference on Computer Vision – ECCV 2024, Milan, Italy, 2024: 19–37. doi: 10.1007/978-3-031-72784-9_2.
[11]	ZHANG R, SONG Jingwei, GAO Xiang, et al. RKHS-BA: A robust correspondence-free multi-view bundle adjustment framework for semantic point clouds[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(12): 10882–10899. doi: 10.1109/TPAMI.2025.3593521.
[12]	XU Yajun, ARAI S, LIU Diyi, et al. FPCC: Fast point cloud clustering-based instance segmentation for industrial bin-picking[J]. Neurocomputing, 2022, 494: 255–268. doi: 10.1016/j.neucom.2022.04.023.
[13]	QIAN Zhentian, FU Jie, and XIAO Jing. Towards accurate loop closure detection in semantic SLAM with 3D semantic covisibility graphs[J]. IEEE Robotics and Automation Letters, 2022, 7(2): 2455–2462. doi: 10.1109/LRA.2022.3145066.
[14]	LI Lin, XIN Kong, ZHAO Xiangrui, et al. SA-LOAM: Semantic-aided LiDAR SLAM with loop closure[C]. 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China, 2021: 7627–7634. doi: 10.1109/ICRA48506.2021.9560884.
[15]	MA Junyi, ZHANG Jun, XU Jintao, et al. OverlapTransformer: An efficient and yaw-angle-invariant transformer network for LiDAR-based place recognition[J]. IEEE Robotics and Automation Letters, 2022, 7(3): 6958–6965. doi: 10.1109/LRA.2022.3178797.
[16]	LEE D, JUNG M, YANG W, et al. Lidar odometry survey: Recent advancements and remaining challenges[J]. Intelligent Service Robotics, 2024, 17(2): 95–118. doi: 10.1007/s11370-024-00515-8.
[17]	KAESS M, JOHANNSSON H, ROBERTS R, et al. iSAM2: Incremental smoothing and mapping with fluid relinearization and incremental variable reordering[C]. 2011 IEEE International Conference on Robotics and Automation, Shanghai, China, 2011: 3281–3288. doi: 10.1109/ICRA.2011.5979641.
[18]	FISCHLER M A and BOLLES R C. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography[J]. Communications of the ACM, 1981, 24(6): 381–395. doi: 10.1145/358669.358692.
[19]	RABBANI T, VAN DEN HEUVEL F, and VOSSELMANN G. Segmentation of point clouds using smoothness constraint[J]. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, 2006, 36(5): 248–253. (查阅网上资料, 未找到本条文献信息, 请确认).
[20]	ZERMAS D, IZZAT I, and PAPANIKOLOPOULOS N. Fast segmentation of 3D point clouds: A paradigm on LiDAR data for autonomous vehicle applications[C]. 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore, Singapore, 2017: 5067–5073. doi: 10.1109/ICRA.2017.7989591.
[21]	XU Wei, CAI Yixi, HE Dongjiao, et al. FAST-LIO2: Fast direct LiDAR-inertial odometry[J]. IEEE Transactions on Robotics, 2022, 38(4): 2053–2073. doi: 10.1109/TRO.2022.3141876.
[22]	BAI Chunge, XIAO T, CHEN Yajie, et al. Faster-LIO: Lightweight tightly coupled LiDAR-inertial odometry using parallel sparse incremental voxels[J]. IEEE Robotics and Automation Letters, 2022, 7(2): 4861–4868. doi: 10.1109/LRA.2022.3152830.
[23]	KIM G and KIM A. Scan Context: Egocentric spatial descriptor for place recognition within 3D point cloud map[C]. 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Madrid, Spain, 2018: 4802–4809. doi: 10.1109/IROS.2018.8593953.
[24]	LIM H, KIM B, KIM D, et al. Quatro++: Robust global registration exploiting ground segmentation for loop closing in LiDAR SLAM[J]. The International Journal of Robotics Research, 2024, 43(5): 685–715. doi: 10.1177/02783649231207654.
[25]	LOW K L. Linear least-squares optimization for point-to-plane ICP surface registration[R]. 2004. (查阅网上资料, 未找到本条文献报告编号信息, 请确认).
[26]	SRINARA S, LEE C M, TSAI S, et al. Performance analysis of 3D NDT scan matching for autonomous vehicles using INS/GNSS/3D LiDAR-SLAM integration scheme[C]. 2021 IEEE International Symposium on Inertial Sensors and Systems (INERTIAL), Kailua-Kona, USA, 2021: 1–4. doi: 10.1109/INERTIAL51137.2021.9430476.
[27]	SEGAL A V, HAEHNELl D, and THRUN S. Generalized-ICP[J]. Robotics: Science and Systems, 2009, 2(4): 1–8. doi: 10.7551/mitpress/8727.001.0001. (查阅网上资料,未找到本条文献信息,请确认).
[28]	KOIDE K, YOKOZUKA M, OISHI S, et al. Voxelized GICP for fast and accurate 3D point cloud registration[C]. 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China, 2021: 11054–11059. doi: 10.1109/ICRA48506.2021.9560835.