基于小光斑机载LiDAR数据的单木三维分割

发布时间：2018-06-06 23:01

本文选题：机载LiDAR + 单木三维分割　；参考：《北京林业大学》2016年硕士论文

【摘要】：从小光斑机载LiDAR点云中获取结构完整的单木数据,在林业研究中具有广泛的应用。近十几年来,如何从海量LiDAR点云数据中提取单木信息,一直是国内外学者研究的重点。本文提出了一种从离散的小光斑机载LiDAR豹椐中提取单木点云的三维分割方法。首先将滤波后的LiDAR点云栅格化生成树冠极大值模型,应用传统的分水岭算法对其进行粗糙分割,并将分割区域内的高程值局部最大值点作为树梢标记；而后在各粗糙分割区域内通过高度阈值过滤潜在的树干反射点,并通过随机抽样一致算法,将水平聚类后潜在的树干点云拟合树干,以实现对树干的识别与标记；最后以树干识别中标记的树梢位置作为初始聚类中心,将滤波后的点云高程值按比例缩小,进行迭代的k均值聚类,实现了单木点云数据的三维分割。本文以美国华盛顿州西部国会森林的蓝岭地区为研究区域,LiDAR数据由搭载在直升机上的Saab TopEye LiDAR系统所采集,脉冲密度为每米4次且最多记录4次回波。研究结果证明,新的三维单木分割方法能够识别位于较低森林层次的矮小树木,弥补了基于树冠高程模型分割方法的局限,有效改善了在对郁闭度较高的森林中相邻树木以及位于树冠高程模型之下低矮树木的识别与分割。与直接应用分水岭算法分割相比,运用结合树干检测的分水岭方法和改进的k均值聚类方法,单木分割的精度得到明显提高。
[Abstract]:It is widely used in forestry research to obtain single tree data with complete structure from airborne LiDAR point cloud with small spot. In recent years, how to extract single wood information from massive LiDAR point cloud data has been the focus of domestic and foreign scholars. In this paper, a 3D segmentation method for extracting single tree point cloud from discrete airborne LiDAR leopard with small spot is proposed. Firstly, the filtered LiDAR point cloud rasterization is used to generate the tree crown maximum model, and the traditional watershed algorithm is used to segment it rough, and the local maximum value of elevation in the segmented region is used as the tree tip mark. Then the potential tree trunk reflection points are filtered through the height threshold in each rough segmentation area, and the potential tree trunk point cloud after horizontal clustering is fitted to the trunk by random sampling consistent algorithm to realize the recognition and marking of the trunk. Finally, the tree tip position marked in tree trunk recognition is taken as the initial clustering center, the filtered point cloud height value is reduced proportionally, and the k-means clustering is carried out iteratively to realize the 3D segmentation of single tree point cloud data. In this paper, the Lidar data are collected from the Blue Ridge region of the Congressional Forest in western Washington State of the United States. The pulse density is 4 times per meter and the maximum number of echoes recorded is 4 times per meter. The results show that the new method can identify the small trees at the lower forest level, which makes up for the limitation of the method based on the crown elevation model. It can effectively improve the recognition and segmentation of the adjacent trees in the forest with higher canopy density and the low trees located under the crown elevation model. Compared with direct watershed segmentation, the accuracy of single tree segmentation is improved obviously by using watershed method combined with tree trunk detection and improved k-means clustering method.
【学位授予单位】：北京林业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP79

【相似文献】