基于近邻几何特征的TLS林分点云分类研究
|
摘要
【目的】在地面激光雷达点云分类任务中多存在特征维度较高的问题,然而当点云数量较多,分类任务中构造较高维度的特征往往需要较多的计算成本和运行内存。为了解决这一问题,本研究提出用近邻点构造5个几何特征训练成熟分类器,以期在将林分点云分为地面、树干与枝叶3个类别的同时达到降低特征维度的目的。【方法】在构造特征的过程中采用近邻值为140的快速KDtree搜索近邻点,获得近邻点后利用其计算协方差矩阵特征值、法向量、曲率、方差和最大高程差构造5个几何特征训练分类器。为了检验本研究构造的特征在林分点云分类中的稳定性,分类器分别采用随机森林和xgboost做比较研究。本研究的实验数据均来自地面激光雷达扫描获得的单站蒙古栎人工林点云数据。【结果】使用随机森林和xgboost分类器训练的模型在测试集中正确估计样本数量和样本总量的比值分别为0.932 1和0.936 3。这两个分类器在地面、树干和枝叶这3个类别中的查准率达到0.97、0.93、和0.91以上,且在这3个类别中的分类结果中xgboost较随机森林均有千分级的优势。【结论】结果表明本研究构造的特征能够完成林分点云分类任务,在保证点云分类准确率的基础上,既减少了特征维度,又有助于提高特征计算效率,具有较高的稳定性。本研究的分类结果可为林分参数反演和生物量估计等研究奠定基础。
[Objective] Most researches on TLS point cloud classification always calculate high dimensional features. But the higher the dimensional features calculated, the more the calculating consumption and running memory needed. So to solve the problem, we designed five geometric features of nearby points to train existing classifier. And then it was used to label the forest point clouds into ground, stem and leaf.[Method] The 140 neighbors gotten by fast KDtree were used to compute the five features, including eigen values of covariance matrix, normal vector, curvature, variance and maximum distance of elevation. And the classifier could be trained with all of them. In order to check the stability of the five features in forest point cloud classification, both random forest and xgboost were introduced. The data in this research were all obtained from Mongolian oak plantation by TLS. [Result] In test sets, the ratios between correct prediction samples and total samples were 0.932 1 and 0.936 3 with random forest and xgboost. Both classifier precision in ground,stem and leaf reached 0.97, 0.93 and 0.91 or more. And compared with random forest,the xgboost's performance in the three categories had millesimal advantage. [Conclusion] On the basis of ensuring the accuracy of point cloud classification, the five features are not only with less dimensions but also helpful to enhance the efficiency of feature computation. It shows they can deal with point cloud classification problem in forest well and have high stability. And the classification result in this research will be helpful for forest parameter extraction and biomass estimation.
[Objective] Most researches on TLS point cloud classification always calculate high dimensional features. But the higher the dimensional features calculated, the more the calculating consumption and running memory needed. So to solve the problem, we designed five geometric features of nearby points to train existing classifier. And then it was used to label the forest point clouds into ground, stem and leaf.[Method] The 140 neighbors gotten by fast KDtree were used to compute the five features, including eigen values of covariance matrix, normal vector, curvature, variance and maximum distance of elevation. And the classifier could be trained with all of them. In order to check the stability of the five features in forest point cloud classification, both random forest and xgboost were introduced. The data in this research were all obtained from Mongolian oak plantation by TLS. [Result] In test sets, the ratios between correct prediction samples and total samples were 0.932 1 and 0.936 3 with random forest and xgboost. Both classifier precision in ground,stem and leaf reached 0.97, 0.93 and 0.91 or more. And compared with random forest,the xgboost's performance in the three categories had millesimal advantage. [Conclusion] On the basis of ensuring the accuracy of point cloud classification, the five features are not only with less dimensions but also helpful to enhance the efficiency of feature computation. It shows they can deal with point cloud classification problem in forest well and have high stability. And the classification result in this research will be helpful for forest parameter extraction and biomass estimation.
引文
[1]李丹,庞勇,岳彩荣,等.基于TLS数据的单木胸径和树高提取研究[J].北京林业大学学报,2012,34(4):79-86.Li D,Pang Y,Yue C R,et al.Extraction of individual tree DBHand height based on terrestrial laser scanner data[J].Journal of Beijing Forest University,2012,34(4):79-86.
[2]Heinzel J,Huber M O.Detecting tree stems from volumetric TLSdata in forest environments with rich understory[J].Remote Sensing,2017,9(9):1-17.
[3]Xia S,Wang C,Pan F,et al.Detecting stems in dense and homogeneous forest using single-scan TLS[J].Forests,2015,6(11):3923-3945.
[4]Wang D,Kankare V,Puttonen E,et al.Reconstructing stem cross section shapes from terrestrial laser scanning[J].IEEEGeoscience and Remote Sensing Letters,2017,14(2):272-276.
[5]Kelbe D,Romanczyk P,Van Aardt J,et al.Reconstruction of 3Dtree stem models from low-cost terrestrial laser scanner data[C/OL]//Proceedings of SPIE8731,Laser Radar Technology and Applications XVIII.Baltimore:SPIE,2013:873106[2018-06-06].http://128.197.168.195/wp-content/uploads/2016/07/kelbe.pdf.
[6]邢万里,邢艳秋,黄杨,等.基于体元逐层聚类的TLS点云数据单木分割算法[J].中南林业科技大学学报,2017,37(12):58-64.Xing W L,Xing Y Q,Huang Y,et al.Individual tree segmentation of TLS point cloud data based on clustering of voxels layer by layer[J].Journal of Central South University of Forestry&Technology,2017,37(12):58-64.
[7]Rusu R B,Marton Z C,Blodow N,et al.Persistent point feature histograms for 3D point clouds[C/OL]//Proceedings of the 10th International Conference on Intelligent Autonomous Systems.Baden-Baden:IAS,2008:119-128[2018-05-12].http://www.doc88.com/p-9943174653805.html.
[8]Rusu R B,Blodow N,Beetz M.Fast point feature histograms(FPFH)for 3D registration[C]//Proceedings of 2009 IEEEInternational Conference on Robotics and Automation.Kobe:IEEE,2009:3212-3217.
[9]Tombari F,Salti S,Stefano L D.Unique signatures of histograms for local surface description[M]//Daniilidis K,Maragos P,Paragios N.Computer vision:ECCV 2010.Berlin:Springer,2010:356-369.
[10]Lari Z,Habib A.Alternative methodologies for the estimation of local point density index:moving towards adaptive LiDAR data processing[C]//Proceedings of International Archives of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Melbourne:ISPRS,2012:127-132.
[11]Weinmann M,Jutzi B,Mallet C.Feature relevance assessment for the semantic interpretation of 3D point cloud data[C]//Proceedings of ISPRS Annals of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Antalya:ISPRS,2013:313-318.
[12]Jutzi B,Gross H.Nearest neighbor classification on laser point clouds to gain object structures from buildings[J].The International Archives of the Photogrammetry,Remote Sensing and Spatial Information Sciences,2009,38(1):4-7.
[13]West K F,Webb B N,Lersch J R.Context-driven automated target detection in 3D data[C]//Proceedings of SPIE 5426,Automatic Target Recognition XIV.Orlando:SPIE,2004:133-143.
[14]Weinmann M,Urban S,Hinz S,et al.Distinctive 2D and 3Dfeatures for automated large-scale scene analysis in urban areas[J].Computers&Graphics,2015,49:47-57.
[15]Hackel T,Wegner J D,Schindler K.Fast semantic segmentation of 3D point clouds with strongly varying density[C]//Proceedings of ISPRS Annals of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Prague:ISPRS,2016:177-184.
[16]DemantkéJ,Vallet B,Paparoditis N.Streamed vertical rectangle detection in terrestrial laser scans for facade database production[C]//Proceedings of ISPRS Annals of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Melboume:ISPRS,2012:99-104.
[17]Becker C,Rosinskaya E,H?ni N,et al.Classification of aerial photogrammetric 3D point clouds[J].Photogrammetric Engineering&Remote Sensing,2018,84(5):287-295.
[18]Friedman J H,Bentley J L,Finkel R A.An algorithm for finding best matches in logarithmic expected time[J].ACM Transactions on Mathematical Software,1977,3(3):209-226.
[19]Muja M,Lowe D G.Fast approximate nearest neighbors with automatic algorithm configuration[C]//Proceedings of International Conference on Computer Vision Theory and Application.Lisboa:VISSAPP,2009:331-340.
[20]Elseberg J,Borrmann D,Nüchter A.One billion points in the cloud:an octree for efficient processing of 3D laser scans[J].ISPRS Journal of Photogrammetry and Remote Sensing,2013,76:76-88.
[21]Breiman L.Random forests[J].Machine Learning,2001,45(1):5-32.
[22]Chen T Q,Guestrin C.Xgboost:a scalable tree boosting system[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.San Francisco:ACM,2016:785-794.
[23]Ye J,Chow J H,Chen J,et al.Stochastic gradient boosted distributed decision trees[C]//Proceedings of the 18th ACMConference on Information and Knowledge Management.Hong Kong:ACM,2009:2061-2064.
[2]Heinzel J,Huber M O.Detecting tree stems from volumetric TLSdata in forest environments with rich understory[J].Remote Sensing,2017,9(9):1-17.
[3]Xia S,Wang C,Pan F,et al.Detecting stems in dense and homogeneous forest using single-scan TLS[J].Forests,2015,6(11):3923-3945.
[4]Wang D,Kankare V,Puttonen E,et al.Reconstructing stem cross section shapes from terrestrial laser scanning[J].IEEEGeoscience and Remote Sensing Letters,2017,14(2):272-276.
[5]Kelbe D,Romanczyk P,Van Aardt J,et al.Reconstruction of 3Dtree stem models from low-cost terrestrial laser scanner data[C/OL]//Proceedings of SPIE8731,Laser Radar Technology and Applications XVIII.Baltimore:SPIE,2013:873106[2018-06-06].http://128.197.168.195/wp-content/uploads/2016/07/kelbe.pdf.
[6]邢万里,邢艳秋,黄杨,等.基于体元逐层聚类的TLS点云数据单木分割算法[J].中南林业科技大学学报,2017,37(12):58-64.Xing W L,Xing Y Q,Huang Y,et al.Individual tree segmentation of TLS point cloud data based on clustering of voxels layer by layer[J].Journal of Central South University of Forestry&Technology,2017,37(12):58-64.
[7]Rusu R B,Marton Z C,Blodow N,et al.Persistent point feature histograms for 3D point clouds[C/OL]//Proceedings of the 10th International Conference on Intelligent Autonomous Systems.Baden-Baden:IAS,2008:119-128[2018-05-12].http://www.doc88.com/p-9943174653805.html.
[8]Rusu R B,Blodow N,Beetz M.Fast point feature histograms(FPFH)for 3D registration[C]//Proceedings of 2009 IEEEInternational Conference on Robotics and Automation.Kobe:IEEE,2009:3212-3217.
[9]Tombari F,Salti S,Stefano L D.Unique signatures of histograms for local surface description[M]//Daniilidis K,Maragos P,Paragios N.Computer vision:ECCV 2010.Berlin:Springer,2010:356-369.
[10]Lari Z,Habib A.Alternative methodologies for the estimation of local point density index:moving towards adaptive LiDAR data processing[C]//Proceedings of International Archives of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Melbourne:ISPRS,2012:127-132.
[11]Weinmann M,Jutzi B,Mallet C.Feature relevance assessment for the semantic interpretation of 3D point cloud data[C]//Proceedings of ISPRS Annals of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Antalya:ISPRS,2013:313-318.
[12]Jutzi B,Gross H.Nearest neighbor classification on laser point clouds to gain object structures from buildings[J].The International Archives of the Photogrammetry,Remote Sensing and Spatial Information Sciences,2009,38(1):4-7.
[13]West K F,Webb B N,Lersch J R.Context-driven automated target detection in 3D data[C]//Proceedings of SPIE 5426,Automatic Target Recognition XIV.Orlando:SPIE,2004:133-143.
[14]Weinmann M,Urban S,Hinz S,et al.Distinctive 2D and 3Dfeatures for automated large-scale scene analysis in urban areas[J].Computers&Graphics,2015,49:47-57.
[15]Hackel T,Wegner J D,Schindler K.Fast semantic segmentation of 3D point clouds with strongly varying density[C]//Proceedings of ISPRS Annals of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Prague:ISPRS,2016:177-184.
[16]DemantkéJ,Vallet B,Paparoditis N.Streamed vertical rectangle detection in terrestrial laser scans for facade database production[C]//Proceedings of ISPRS Annals of the Photogrammetry,Remote Sensing and Spatial Information Sciences.Melboume:ISPRS,2012:99-104.
[17]Becker C,Rosinskaya E,H?ni N,et al.Classification of aerial photogrammetric 3D point clouds[J].Photogrammetric Engineering&Remote Sensing,2018,84(5):287-295.
[18]Friedman J H,Bentley J L,Finkel R A.An algorithm for finding best matches in logarithmic expected time[J].ACM Transactions on Mathematical Software,1977,3(3):209-226.
[19]Muja M,Lowe D G.Fast approximate nearest neighbors with automatic algorithm configuration[C]//Proceedings of International Conference on Computer Vision Theory and Application.Lisboa:VISSAPP,2009:331-340.
[20]Elseberg J,Borrmann D,Nüchter A.One billion points in the cloud:an octree for efficient processing of 3D laser scans[J].ISPRS Journal of Photogrammetry and Remote Sensing,2013,76:76-88.
[21]Breiman L.Random forests[J].Machine Learning,2001,45(1):5-32.
[22]Chen T Q,Guestrin C.Xgboost:a scalable tree boosting system[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.San Francisco:ACM,2016:785-794.
[23]Ye J,Chow J H,Chen J,et al.Stochastic gradient boosted distributed decision trees[C]//Proceedings of the 18th ACMConference on Information and Knowledge Management.Hong Kong:ACM,2009:2061-2064.