Urban vegetation inventory at city-scale using terrestrial light detection and ranging (LiDAR) point clouds is very challenging due to the large quantity of points, varying local density, and occlusion effects, leading to missing features and incompleteness of data. This paper proposes a novel method, named Point Non-Local Means (PointNLM) network, which incorporates the supervoxel-based and point-wise for automatic semantic segmentation of vegetation from large scale complex scene point clouds. PointNLM captures the long-range relationship between groups of points via a non-local branch cascaded three times to describe sharp geometric features. Simultaneously, a local branch processes the position of scattered feature points and captures the low and high level features. Finally, we proposed a fusion layer of neighborhood max-pooling method to concatenate the long-range features, low level features and high level features for segmenting the trees. The proposed architecture was evaluated on three datasets, including two open access datasets of Semantic3D and Paris-Lille-3D, and an in-house dataset acquired by a commercial mobile LiDAR system. Experimental results indicated that the proposed method provides an efficient and robust result for vegetation segmentation, achieving an Intersection over Union (IoU) of 94.4%, F₁-score of 92.7% and overall accuracy of 96.3%, respectively.

由于点数量多、局部密度不同和遮挡效应，使用地面光检测和测距 (LiDAR) 点云在城市范围内进行城市植被清查非常具有挑战性，导致特征缺失和数据不完整。本文提出了一种新方法，称为点非局部均值（PointNLM）网络，它结合了基于超体素和逐点的自动语义分割从大规模复杂场景点云中的植被。PointNLM 通过级联 3 次的非局部分支来捕捉点组之间的长程关系，以描述尖锐的几何特征。同时，局部分支处理分散特征点的位置并捕获低层和高层特征。最后，我们提出了一个邻域最大池化方法的融合层来连接远程特征、低级特征和高级特征来分割树。所提出的架构在三个数据集上进行了评估，包括 Semantic3D 和 Paris-Lille-3D 的两个开放访问数据集，以及由商业移动 LiDAR 系统获取的内部数据集。实验结果表明，所提出的方法为植被分割提供了有效且稳健的结果，实现了 94.4% 的联合交集（IoU），F_{1 -}分数分别为 92.7% 和总体准确度为 96.3%。