基于特征融合的林下环境点云分割
Point cloud segmentation algorithm based on feature fusion used for understory environments
-
摘要: 针对林下环境几何特征的复杂性,以及基于边检测、表面增长和聚类分割方法存在的效率低、分割不足及过度分割等问题,提出了一种基于特征融合的点云分割方法。采用地面激光扫描仪FARO在北京林业大学选择样本区域进行扫描,对扫描得到的数据进行采样点剔除及滤波,得到由1166302个点组成的林下环境点云数据,主要包括林木、地面、石块、人4类目标。综合利用点云法向量信息和激光反射强度信息可实现点云分割。其中,点云激光反射强度可直接从扫描得到的点云数据中获取;法向量可根据点云数据的三维坐标信息,通过对点云数据建立kd-tree数据结构,执行k-邻域搜索,并基于PlanePCA算法计算得到。将点云法向量和激光反射强度2方面的特征优势进行融合,计算中心点与邻域点的综合相异度,并判断其是否在阈值范围内,最终实现点云分割。比较基于特征融合、法向量和激光反射强度3种聚类分割方法得到的分割结果可知,基于特征融合的聚类分割方法能较好地保留数据特征,且分割完整度明显优于其他2种方法。Abstract: Aimed at the complexity of geometric features of understory environments and the deficiency of edge detection based method, region growing based method and clustering feature based method, we propose a new point cloud segmentation algorithm based on feature fusion. The 3D data set acquired from Beijing Forestry University using a FARO laser scanner consists of 1166302 points after removing outliers and filtering. The data set has four targets, i.e., tree, ground, stone and person. Point cloud segmentation can be achieved via fusing normal vector and laser reflection intensity of each point. The laser reflection intensity values can be obtained from point cloud data set directly, and normal vector should be calculated based on the Plane PCA algorithm. Also, it is necessary to create kd-tree data structure and perform k-NN search during the calculation of normal vector. Segmentation is realized after fusing the advantages of normal vector and laser reflection intensity and calculating synthetical difference degree between query points and neighborhood points. Comparing the segmentation results from point cloud segmentation algorithms based on feature fusion, normal vector and laser reflection intensity, the method based on feature fusion overcomes the problem of data deficiency that the other two methods suffer.