随着三维激光技术的发展，越来越多的应用需要用到三维激光点云数据，感知和场景理解是构建世界未知地区三维信息的关键组成部分。基于激光点云数据完成的三维模型点云精度更高，激光点云分割是模型重建的关键技术之一。地下空间是城市网络空间资源的重要组成部分，地下场景的建设也被列为地下空间规划中一个重要的部分，对其的三维数据分割是当下热点。本文首先使用LS-SVM(Least Squares Support Vector Machine)方法对地下空间点云数据进行粗分割，采用基于基点集的PointNet++网络模型实现点云数据精分割。最后利用点云分割后的数据进行地下场景模型重建。本文主要工作及研究内容如下：

（1）针对激光点云在弱光和低特征点环境的点云分割中过拟合的问题，本文提出了一种基于网格搜索的多项式核最小二乘支持向量机点云分割算法，提取点云的高程、曲率、颜色和形状特征对车辆和车库障碍物柱子等多尺度特征进行分类；为验证特征选取的有效性，采用F1 Score、召回率、查准率和评估方法的总体精确度对分割算法进行评估。结果表明：与文献中传统分类算法的柱子识别率83%和车辆识别率74%相比，基于多项式核函数的分类算法对柱子和车辆的识别率分别达到了89%和76%以上，提高了6%和2%。

（2）针对PointNet++缺乏局部上下文信息导致翻转不变性，其特征学习完全依赖于点的欧几里得坐标，本文通过应用BPS(Basic Point Setting)点云分析策略来推断编码完整点云的潜在特征。制作激光点云数据集并对其使用融合BPS的PointNet++网络分割模型，利用该网络完成对点云场景的语义分割。并对制作的数据集进行网络模型的训练，在原PointNet++网络结构上采用基于基准点集模型，与PointNet++相比，本文的改进算法平均精确度提升了0.89个百分点。与PointNet++相比，本文的改进网络模型在局部信息的提取中具有明显优势，其中平均交并比相对原始点云网络提高了约3.4%，OA相对提高了2.5%。

（3）基于对地下场景的精分割后的三维模型表面的复杂程度不同带来的重建效率的不同，本文采用了自动和半自动的点云分割策略。对地下车库场景中各物体分割后的点云数据依据不同特征进行自动和半自动重建。以地下停车场为实验目标，从获取的激光三维轮廓线中手动筛选出物体基础轮廓线，对场景中分割出来的物体如墙体、支柱、障碍物等实现模型采用区域增长算法进行自动化重建，结合小部件模型采用半自动重建方法实现地下空间的三维重建，为城市地下空间的数字化管理提供了数据支持。

With the development of 3D laser technology, more and more applications need to use 3D laser point cloud data. Perception and scene understanding are key components of constructing 3D information in unknown areas of the world. The accuracy of 3D model point cloud based on laser point cloud data is higher, and laser point cloud segmentation is one of the key technologies of model reconstruction. Underground space is an important part of urban cyberspace resources. The construction of underground scenes is also listed as an important part of underground space planning. Its three-dimensional data segmentation is a hot topic at present. In this paper, the LS-SVM (Least Squares Support Vector Machine) method is used to segment the point cloud data of underground space, and the PointNet++ network model based on base point set is used to realize the fine segmentation of point cloud data. Finally, the model of underground parking lot is reconstructed by using the data after point cloud segmentation. The main work and research contents of this paper are as follows:

(1) Aiming at the problem of overfitting of laser point cloud in point cloud segmentation in low light and low feature point environment, this paper proposes a polynomial kernel least squares support vector machine point cloud segmentation algorithm based on grid search. The elevation, curvature, color and shape features of point cloud are extracted to classify multi-scale features such as vehicles and garage obstacle pillars. In order to verify the effectiveness of feature selection, the segmentation algorithm is evaluated by F1 Score, recall rate, precision rate and the overall accuracy of the evaluation method. The results show that compared with the traditional classification algorithm in the literature, the recognition rate of the column and the vehicle is 83 % and 74 % respectively. The recognition rate of the column and the vehicle based on the polynomial kernel function is 89 % and 76 % respectively, which is improved by 6 % and 2 %.

(2) PointNet++ lacks local context information, which leads to flip invariance. Its feature learning depends entirely on the Euclidean coordinates of points. In this paper, the BPS (Basic Point Setting) point cloud analysis strategy is applied to infer the potential features of the encoded complete point cloud. The laser point cloud data set is made and the PointNet++ network segmentation model fused with BPS is used to complete the semantic segmentation of the point cloud scene. The network model is trained on the data set, and the benchmark point set model is adopted on the original PointNet++ network structure. Compared with PointNet++, the average accuracy of the improved algorithm in this paper is improved by 0.89 percentage points. Compared with PointNet++, the improved network model in this paper has obvious advantages in the extraction of local information. The average intersection over union is about 3.4 % higher than that of the original point cloud network, and OA is relatively increased by 2.5 %.

(3) Based on the different reconstruction efficiency caused by the different complexity of the 3D model surface after the fine segmentation of the underground scene, this paper adopts the automatic and semi-automatic point cloud segmentation strategy. The segmented point cloud data of each object in the underground garage scene are automatically and semi-automatically reconstructed according to different characteristics. Taking the underground parking lot as the experimental target, the basic contour line of the object is manually selected from the obtained laser three-dimensional contour line. The model of the objects such as walls, pillars and obstacles segmented from the scene is automatically reconstructed by the region growth algorithm. Combined with the small component model, the semi-automatic reconstruction method is used to realize the three-dimensional reconstruction of the underground space, which provides data support for the digital management of urban underground space.

[1] 李德仁, 姚远, 邵振峰. 智慧城市中的大数据[J]. 武汉大学学报·信息科学版, 2014, 39(6): 631-640.

[2] Ying W, Ewert D, Schilberg D, et al. Edge Extraction by Merging the 3D Point Cloud and 2D Image Data[J]. Automation, Communication and Cybernetics in Science and Engineering, 2014: 773-785.

[3] Artime Ríos E M, Suárez Sánchez A, Sánchez Lasheras F, et al. Genetic Algorithm based on Support Vector Machines for Computer Vision Syndrome Classification in Health Personnel[J]. Neural Computing and Applications, 2018: 381-390.

[4] Yu L, Li X, Fu C W, et al. PU-Net: Point Cloud Upsampling Network[C]//The 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City: UT, 2018: 2790-2799.

[5] 赵江洪, 孙铭悦, 王殷瑞, 等. 三维点云孔洞修复方法综述[J]. 测绘科学, 2021, 46(01): 114-123.

[6] Axelsson P. Processing of laser scanner data—algorithms and applications[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 1999, 54(2-3): 138-147.

[7] Axelsson P. DEM generation from laser scanner data using adaptive TIN models[J]. International Archives of Photogrammetry & Remote Sensing, 2000, 33(4): 110-117.

[8] Cang P, Yu Z. Research on 3D building information extraction and image post-processing based on vehicle LIDAR[J]. EURASIP Journal on Image and Video Processing, 2018, 2018(1): 247-263.

[9] 赵传. 航空影像辅助机载LiDAR点云的建筑物三维模型重建技术研究 [D]. 郑州: 战略支援部队信息工程大学, 2020.

[10] Atik M E, Duran Z, Seker D Z. Machine Learning-Based Supervised Classification of Point Clouds Using Multiscale Geometric Features[J]. ISPRS International Journal of Geo-Information, 2021: 1154-1158.

[11] Kilian J, Haala N, Englich M. Capture and evaluation of airborne laser scanner data[J]. International Archives of Photogrammetry and Remote Sensing, 1996, 31: 383-388.

[12] Chernyshev V V. Potentialities of Grid Search Method[J]. Acta Crystallographica Section A: Foundations and Advances, 2000: S132-S132.

[13] 吴军, 刘荣, 郭宁, 等. SVM加权学习下的机载LiDAR数据多元分类研究[J]. 武汉大学学报(信息科学版), 2013, 38(01): 1-6.

[14] 李海伦, 黎荣, 丁国富, 等. 应用遗传模糊聚类实现点云数据区域分割[J]. 计算机应用研究, 2012, 29(05): 1974-1976.

[15] 李茂民, 邹臣嵩. 基于改进遗传算法的阈值图像分割方法[J]. 软件工程, 2022, 25(01): 37-40.

[16] Zhao Z, Song Y, Cui F, et al. Point Cloud Features-Based Kernel SVM for Human-Vehicle Classification in Millimeter Wave Radar[C]//The IEEE Access, 2020: 26012-26021.

[17] Gevaert C, Persello C, Vosselman G. Optimizing Multiple Kernel Learning for the Classification of UAV Data[J]. Remote Sensing, 8(12): 1025.

[18] 陈景年. 一种适于多分类问题的支持向量机加速方法[J]. 计算机科学, 2022, 49(S1): 297-300.

[19] Seyoun, Park, Xiaohu, et al. Shape and Appearance Repair for Incomplete Point Surfaces[J]. Korean Journal of Computational Design & Engineering, 2007: 1260-1267.

[20] 杨春雨, 张鑫. 煤矿机器人环境感知与路径规划关键技术[J]. 煤炭学报, 2022, 47(07): 2844-2872.

[21] 蒋腾平, 杨必胜, 周雨舟, 等. 道路点云场景双层卷积语义分割[J]. 武汉大学学报(信息科学版), 2020, 45(12): 1942-1948.

[22] Qi C R, Su H, Mo K, et al. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation[C]//The 2017 IEEE Conference on Computer Vision and Pattern Recognition. 2017: 652-660.

[23] Qi C R, Yi L, Su H, et al. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space[J]. Computer Science, 2017, 30: 327-332.

[24] Newman E L, Varley T F, Parakkattu V K, et al. Revealing the Dynamics of Neural Information Processing with Multivariate Information Decomposition[J]. Entropy, 2022, 24(7): 930-935.

[25] Wang, Lei, Wu, et al. Semantic segmentation of large-scale point clouds based on dilated nearest neighbors graph[J]. Complex & Intelligent Systems, 2022, 8(5): 3833-3845.

[26] Chen C, Qian S, Fang Q, et al. HAPGN: Hierarchical Attentive Pooling Graph Network for Point Cloud Segmentation[C]//IEEE Transactions on Multimedia, 2020: 2335-2346.

[27] Gyeongsik M, Ju Yong C, Kyoung Mu L. V2V-PoseNet: Voxel-to-Voxel Prediction Network for Accurate 3D Hand and Human Pose Estimation from a Single Depth Map[J]. arXiv - CS - Computer Vision and Pattern Recognition, 2018: 5079-5088.

[28] Gernot R, Ali Osman U, Andreas G. OctNet: Learning Deep 3D Representations at High Resolutions[J]. arXiv - CS - Computer Vision and Pattern Recognition, 2017: 3577-3586.

[29] Ochmann S, Vock R, Klein R. Automatic reconstruction of fully volumetric 3D building models from oriented point clouds[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 151: 251-262.

[30] Zhirong W, Shuran S, Aditya K, et al. 3D ShapeNets: A Deep Representation for Volumetric Shapes[J]. arXiv - CS - Computer Vision and Pattern Recognition, 2015: 1912-1920.

[31] Kripasindhu S, Basavaraj H, Kiran V, et al. Learning 3D Shapes as Multi-Layered Height-maps using 2D Convolutional Networks[J]. arXiv - CS - Computer Vision and Pattern Recognition, 2018: 71-86.

[32] Kanezaki A, Matsushita Y, Nishida Y. RotationNet for Joint Object Categorization and Unsupervised Pose Estimation from Multi-view Images[C]//IEEE Transactions on Pattern Analysis and Machine Intelligence 2020, 43(1): 269-283.

[33] Wang L, Zhan Y. A conceptual peer review model for arXiv and other preprint databases[J]. Learned Publishing, 2019, 32(3): 213-219.

[34] 杨柳, 刘启亮, 袁浩涛. 城市激光点云语义分割典型方法对比研究[J]. 地理与地理信息科学, 2021, 37(01): 17-25.

[35] Cheng W, Yang S, Zhou M, et al. Road Mapping and Localization Using Sparse Semantic Visual Features[C]//IEEE Robotics and Automation Letters, 2021, 6(4): 8118-8125.

[36] David P, Hengyuan Z, Qinru L, et al. Probabilistic Semantic Mapping for Urban Autonomous Driving Applications[J]. arXiv - CS - Computer Vision and Pattern Recognition, 2020: 2059-2064.

[37] Long R, Rauch C, Zhang T, et al. RigidFusion: Robot Localisation and Mapping in Environments With Large Dynamic Rigid Objects[C]//IEEE Robotics and Automation Letters, 2021, 6(2): 3703-3710.

[38] Muzhi H, Zeyu Z, Ziyuan J, et al. Reconstructing Interactive 3D Scenes by Panoptic Mapping and CAD Model Alignments[J]. arXiv - CS - Robotics, 2021: 12199-12206.

[39] Victor A, Timothy P O, Arnaud C, et al. Lidar-Monocular Surface Reconstruction Using Line Segments[J]. arXiv - CS - Computer Vision and Pattern Recognition, 2021: 5631-5637.

[40] Wang X, Xu L, Sun H, et al. On-Road Vehicle Detection and Tracking Using MMW Radar and Monovision Fusion[C]//IEEE Transactions on Intelligent Transportation Systems,2021, 17(7): 2075-2084.

[41] Macher H, Landes T, Grussenmeyer P, et al. Boundaries Extraction from Segmented Point Clouds as Input for Historical Building Information Modelling[J]. International journal of heritage in the digital era, 2014, 3(4): 669-682.

[42] 林伟. 城市地下管廊点云语义分割与三维重建方法研究 [D]. 昆明: 昆明理工大学, 2021.

[43] Liu X, Chen Y, Cheng L, et al. Airborne Laser scanning Point Clouds Filtering Method based on the Construction of Virtual Ground Seed Points[J]. Journal of Applied Remote Sensing, 2017, 11(1): 016032-016043.

[44] Zhang K, Chen S-C, Whitman D, et al. A Progressive Morphological Filter for Removing Nonground Measurements from Airborne LIDAR Data[C]//IEEE Transactions on Geoscience and Remote Sensing, 2021, 41(4): 872-882.

[45] 释小松, 程英蕾, 赵中阳, 等. 基于三角网滤波和支持向量机的点云分类算法[J]. 激光与光电子学进展, 2019, 56(16): 32-40.

[46] Pingel T J, Clarke K C, McBride W A. An improved simple morphological filter for the terrain classification of airborne LIDAR data[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2013, 77: 21-30.

[47] Sun Y, Ding S, Zhang Z, et al. An improved grid search algorithm to optimize SVR for prediction[J]. Soft Computing, 2021, 25: 5633-5644.

[48] 蒋卫涛, 李民, 姚雄, 等. 基于最小二乘支持向量机的太阳能集热效率预测[J]. 工业加热, 2020, 49(06): 5-9.

[49] 黄新波, 蒋卫涛, 朱永灿, 等. 基于时间序列和支持向量机的变压器故障预测[J]. 高电压技术, 2020, 46(07): 2530-2538.

[50] Petro L, Pavlo L. Grid Search, Random Search, Genetic Algorithm: A Big Comparison for NAS[J]. arXiv - CS - Machine Learning, 2019, 1912-1923.

[51] Bengio Y. Learning Deep Architectures for AI[J]. Foundations & Trends in Machine Learning, 2009, 2(1): 121-127.

[52] Guo Z, Feng C C. Using multi-scale and hierarchical deep convolutional features for 3D semantic classification of TLS point clouds[J]. International Journal of Geographical Information Science, 2018: 1-20.

[53] Munoz D, Bagnell J A, Vandapel N, et al. Contextual classification with functional Max-Margin Markov Networks[C]//IEEE Conference on Computer Vision and Pattern Recognition, 2009: 975-982.

[54] Su H, Maji S, Kalogerakis E, et al. Multi-view Convolutional Neural Networks for 3D Shape Recognition[C]//IEEE International Conference on Computer Vision, 2015: 945-953.

[55] Hackel T, Savinov N, Ladicky L, et al. Semantic3D.net: A new Large-scale Point Cloud Classification Benchmark[J]. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences, 2017: 1704-1727.

[56] 李晓宇, 陈伟, 杨维, 等. 基于超像素特征与SVM分类的人员安全帽分割方法[J]. 煤炭学报, 2021, 46(06): 2009-2022.

[57] Azim A, Thierry M, Devy M M, et al. 3D Perception of Outdoor and Dynamic Environment using Laser Scanner[J]. Universite de Grenoble I-Joseph Fourier, 2013: 2532-2547.

[58] Qi C R, Su H, Niebner M, et al. Volumetric and Multi-View CNNs for Object Classification on 3D Data[C]//IEEE Conference on Computer Vision and Pattern Recognition, 2016: 5648-5656.

[59] Maturana D, Scherer S. VoxNet: A 3D Convolutional Neural Network for real-time object recognition[C]//IEEE/RSJ International Conference on Intelligent Robots and Systems, 2015: 922-928.

[60] Wang Z, Zhang L, Li R, et al. A Deep Neural Network With Spatial Pooling (DNNSP) for 3-D Point Cloud Classification[C]//IEEE Transactions on Geoscience and Remote Sensing, 2018, 56(8): 4594-4604.

[61] Landrieu L, Simonovsky M. Large-scale Point Cloud Semantic Segmentation with Superpoint Graphs[J]. arXiv e-prints, 2017: 4558-4567.

[62] Vallet B, Bredif M, Serna A, et al. TerraMobilita/iQmulus urban point cloud analysis benchmark[J]. Computers & Graphics, 2015, 49: 126-33.

[63] Dai Y, Gong J, Li Y, et al. Building segmentation and outline extraction from UAV image-derived point clouds by a line growing algorithm[J]. International Journal of Digital Earth, 2017, 10(11): 1077-1097.

[64] Yi L, Kim V G, Ceylan D, et al. A Scalable Active Framework for Region Annotation in 3D Shape Collections[J]. ACM Transactions on Graphics (ToG), 2016, 35(6): 1-12.

[65] Li Y, Bu R, Sun M, et al. PointCNN: Convolution on x-transformed points[J]. Advances in neural information processing systems, 2018, 31: 225-243.

[66] 赵鹏程. 室内外场景三维建模与语义分割关键技术研究 [D]. 武汉: 武汉大学, 2020.

[67] Wu Z, Song S, Khosla A, et al. 3D ShapeNets: A Deep Representation for Volumetric Shapes[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2015: 1912-1920.

[68] 詹旭. 点云数据预处理之关键算法研究 [D]. 四川: 西南科技大学, 2021.

[69] 张强, 李朝奎, 李俊晓, 等. 一种改进的基于法矢方向调整的平面点云分割方法[J]. 地理与地理信息科学, 2015, 31(01): 45-48.

[70] 熊艳, 高仁强, 徐战亚. 机载LiDAR点云数据降维与分类的随机森林方法[J]. 测绘学报, 2018, 47(04): 508-518.

[71] 陈朋, 谭晔汶, 李亮. 地面三维激光扫描建筑物点云特征线提取[J]. 激光杂志, 2016, 37(03): 9-11.

[72] Zuowei H, Feng L, Guangwei H. Improved Method for LiDAR Point Cloud Data Filtering Based on Hierarchical Pseudo-Grid[J]. Acta Optica Sinica, 2017, 37(8), 0828004.

[73] 骆磊, 马荣贵, 马园. 车窗三维点云数据修复算法[J]. 计算机技术与发展, 2020, 30(07): 6-11.

[74] 林豪, 江竹, 李树彬. 基于机器学习的城市快速路速度-密度关系模型[J]. 西安科技大学学报, 2020, 40(06): 109-116.

[75] 宋涣. 深度学习点云语义分割辅助稠密建图研究 [D]. 南京: 东南大学, 2019.

[76] 朱滨. 历史建筑点云模型虚拟几何修复研究 [D]. 江西: 江西理工大学, 2021.

[77] Li J, Hu B, Noland T L. Classification of tree species based on structural features derived from high density LiDAR data[J]. Agricultural and Forest Meteorology, 2013: 104-114.

[78] Friedman S, Stamos I. Online Facade Reconstruction from Dominant Frequencies in Structured Point Clouds[C]//Computer Vision and Pattern Recognition Workshops, 2012: 1-8.

[79] Xiao Y, Hu S, Xiao S, et al. A Fast Statistical Method of Tree Information from 3D Laser Point Clouds[J]. Chinese Journal of Lasers, 2018, 45(5): 472-487.

[80] Chauve A L, Labatut P, Pons J P. Robust Piecewise-planar 3D Reconstruction and Completion from Large-scale Unstructured Point Data[C]//The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, San Francisco: CA, 2020: 13-18.

[81] Xiao C, Zheng W, Miao Y, et al. A Unified Method for Appearance and Geometry Completion of Point Set Surfaces[J]. Visual Computer, 2007, 23(6): 433-443.

[82] Axelsson P. DEM Generation from Laser Scanner Data using Adaptive TIN Models[J]. International archives of photogrammetry and remote sensing, 2000, 33(4): 110-117.

[83] Sithole G, Vosselman G. Experimental Comparison of Filter algorithms for bare-Earth Extraction from Airborne Laser Scanning Point Clouds[J]. ISPRS journal of photogrammetry and remote sensing, 2004, 59(2): 85-101.

[84] Pauly M, Mitra N J, Giesen J, et al. Example-Based 3D Scan Completion[J]. Eurographics Association, 2005: 23-35.