结合深度学习和 $K$ -Means 的行道树提取及单木分割研究

针对目前城市道路场景中行道树提取方法需要设置的参数较多以及树冠点云相互重叠难以精确分割的问题，文章采用一种行道树提取与单株木分割算法。首先通过布料滤波算法从原始点云中移除地面点，并利用半径滤波滤除离群点，去除地面点和噪声点对行道树提取的影响；然后通过增加 PointNet++ 网络的点集抽象模块(set abstraction, SA) 提高模型特征提取能力，使模型更适用于行道树点云的提取，并利用改进后的网络从原始点云中提取行道树点云；最后结合密度聚类算法（density-based spatial clustering of applications with noise, DBSCAN）与 K-Means 算法对相互重叠的行道树点云进行分割，得到单株木信息。为验证该方法的有效性，以北京永昌路道路数据集进行训练测试。结果表明：改进后模型的行道树点云平均提取精度和交并比 (intersection over union, IoU) 分别提高了$9.2\%$ 和 $15.1\%$ ，达到了 $94.5\%$ 、0.916；单木分割平均精度达到了 $91.3\%$ 。

A street tree extraction and single wood segmentation algorithm is used to address the issues of multiple parameters that need to be set for street tree extraction methods in urban road scenes, as well as the difficulty in accurately segmenting overlapping tree crown point clouds. Firstly, ground points are removed from the original point clouds using cloth simulation filtering algorithm, and outliers are filtered using radius filtering to remove the influence of ground points and noise points on the extraction of street trees. Then, by adding the set abstraction (SA) module of PointNet++ network, the feature extraction ability of the model is improved, making it more suitable for extracting street tree point clouds. The improved network is used to extract street tree point clouds from the original point clouds. Finally, the density-based spatial clustering of applications with noise (DBSCAN) and$K$ -Means algorithms are combined to segment overlapping street tree point clouds and obtain single wood information. To verify the effectiveness of this method, training tests were conducted on the Yongchang Road dataset in Beijing. The experimental results show that the average extraction accuracy and intersection over union (IoU) of the im proved model for street tree point clouds increase by $9.2\%$ and $15.1\%$ , respectively, reaching $94.5\%$ and $0.916$ ; the average accuracy of single wood segmentation reaches $91.3\%$ .