基于对象的激光点云数据城区树木识别方法

Identification of urban vegetation with laser point cloud data based on object

通常利用激光点云数据(LiDar)进行树木分类的方法是将点云内插生成数字地形模型(DTM),根据地物高程差值,在图像处理方法的基础上进行分割或分类。提出一种新的基于对象的LiDar数据树木识别方法,其最大特点是直接利用点云数据的三维空间关系进行分类,不需要将点云转换成二维图像进行处理,避免了转换过程中信息的丢失,提高了分类的精度。具体实现步骤:首先利用kd-trees组织点云数据,在局部邻域中利用点云位置和法线分别进行协方差分析,估算各点的空间特征变量,然后结合各点的回波次数和局部邻域中点云个数密度作为SVM分类器的输入变量,最后利用基于径向基函数的SVM方法实现点云的分类。实验结果表明:OA分类精度为91.21%,Kappa系数为85.62%。

The method,which applies LIDAR to distinguish trees,is to interpolate digital terrestrial model（DTM） and then to conduct segmentation or classification with the differences of features＇ elevation.A new method to identify trees,based on object and with LIDAR clata was put forward,thus 3D relationship among point cloud data can directly participate in classification process,and therefore there is no need to transfer 2D image from point data and information loss to a large extent can be avoid and classification accuracy can be improved.The steps are： firstly,to arrange point cloud data with kd-tree and to carry out an analysis on covariance in local neighborhood,then variables of spatial feature can be estimated;secondly,to determine input of SVM classifier with combination of echo times on points and density of point cloud in local neighborhood;lastly,to execute point cloud classification based on radial basis function with SVM.The result evidences that OA classification reaches accuracy of 91.21 % and Kappa coefficient of 85.62 %.