Shape classification guided method for automated extraction of urban trees from terrestrial laser scanning point clouds
Shape classification guided method for automated extraction of urban trees from terrestrial laser scanning point clouds
复制标题
从地面激光扫描点云中自动提取城市树木的形状分类引导方法
DOI：
10.1007/s11042-021-11328-7
复制
发表时间：
2021-08-19
影响因子：
3.6
通讯作者：
Wang, Yinghui
中科院分区：
文献类型：
--
作者：
Ning, Xiaojuan;Tian, Ge;Wang, Yinghui
树木个体的准确检测和提取是当前研究的热点之一，在车辆导航、树木建模、树木生长监测、城市绿色量估算等方面有着广泛的应用。与单个树木提取相关的困难是城市场景中杂乱的点云中与其他对象的遮挡，这抑制了单个树木的自动提取。在本文中，我们提出了一个全面的框架，可用于从地面扫描的户外场景中提取单个树木。在我们的框架中，一个自底向上的方法通过形状引导分类来选择候选树冠和树干，并提出了一种新的三阶段形状合并规则，包括定位，过滤和匹配（LFM），以生成一个完整的个体树。所提出的方法的主要优点是，它是独立的数据的质量和不同的形状。对支持向量机分类方法和随机森林分类方法在准确率评估上进行了对比实验。该框架的有效性进行了测试，在奥克兰户外MLS数据集的点云在五个街道场景。五个测试点的结果实现了高于97%的树木检测率，总体准确率约为98%，两个程序的完成质量为96%。未检测到的树木通常是稀疏的，这是由于点云数据中存在遮挡;大多数错误分类发生在与树木相邻且与树干高度相同的人造支柱上。与现有方法的对比实验表明了该方法的有效性。
Accurate detection and extraction of individual trees is one of hottest topics, which can be widely used in vehicles navigation, tree modeling, tree growth monitoring and urban green quantity estimation. The difficulty associated with individual trees extraction is the occlusion with other objects in cluttered point clouds of urban scenes, which inhibits the automatic extraction of individual trees. In this paper, we present a comprehensive framework that can be used to extract individual trees from terrestrial scanned outdoor scene. In our framework, a bottom-up method by shape-guided classification is achieved to select the candidate tree crowns and tree trunks, and a novel three-stage shape merging rule containing localization, filtering, and matching (LFM) are proposed to generate a complete individual tree. The primary advantage of the proposed method is that it is independent of the quality of data and different shapes. We made comparison experiments of classification methods of support vector machine and random forest on the accuracy assessment. The effectiveness of the proposed framework was tested in five street scenarios in point clouds from Oakland outdoor MLS dataset. The results for the five test sites achieved tree detection rates higher than 97%; the overall accuracy was approximately 98%, and the completion quality of both procedures was 96%. Non-detected trees are always sparse which come from occlusions in the point cloud data; most misclassifications occurred in man-made pillars adjacent to trees and have the same height with tree trunk. Comparison experiments to the existing methods are made to illustrate the effectiveness of our method.