【目的】针对行道树树冠在线检测问题,研究基于移动激光扫描(mobile laser scanning,MLS)的行道树树冠点云逐点检测方法,构建能够在线、快速、准确检测出行道树树冠点云的高性能树冠检测器,为行道树对靶施药提供基础数据。【方法】应用搭载一个2D激光雷达(light detection and ranging,LiDAR)的MLS系统实时采集街道轮廓线测量数据,从中提取点云三维坐标、一次回波强度和回波次数等属性;构建点云半径为δ的球域搜索方法,实现点云邻域在线快速查询;从待识别点δ球域中提取宽度、深度、高度、维度、密度、次数和强度7类点云局部特征;采用监督学习算法融合点云局部特征、训练树冠检测器,预测待识别点的类别。采集一段长137 m街道的点云数据,开展了邻域搜索方法、监督学习算法、点云局部特征和树冠逐点检测器4个对比实验。【结果】构建的δ球域搜索方法的搜索时间为k-D树法的10.90%;在神经网络(neural network,NN)、支持向量机(support vector machine,SVM)、Boosting和随机森林(random forest,RF)4种监督学习算法中,RF算法得到的树冠检测器分类精度最好;与单类特征相比,组合特征具有更好的泛化性能;本研究方法设计的树冠逐点检测器在检测精度和效率上均优于已有方法,球域半径δ在0.1~1.0 m范围内变化时,测试集F₁分数≥97.74%。【结论】提出的方法能够从实时采集的MLS点云数据中快速、准确地检测出行道树树冠点云,为行道树对靶施药提供数据支撑。

【Objective】 To address the problem of online detection of street-tree crowns, this study proposes a pointwise detection method for the point cloud of street-tree crowns based on mobile laser scanning (MLS), and generates a high-performance tree-crown detector capable of online, fast and accurate detection of the point cloud of street-tree crowns to provide basic data for street-tree-targeted spraying. 【Method】 The MLS system equipped with a 2D LiDAR (light detection and ranging) was used to collect the street contour measurement data in real time, and five attributes such as three-dimensional coordinates, primary echo intensity and number of echoes were extracted; a δ spherical neighborhood search method was built to realize online and fast query of point cloud neighborhood; seven kinds of local features of point cloud such as width, depth, height, dimensionality, density, number of echoes and echo intensity were extracted from the spherical domain of the point to be identified; the supervised learning algorithm was used to fuse the local features of point cloud and train a crown detector to predict the category of the point to be identified. 【Result】 The point cloud data of a 137 m long street were collected, and four comparative experiments of neighborhood search methods, supervised learning algorithms, local features of point clouds, and pointwise crown detectors were conducted. The experimental results show that the query time of the proposed δ spherical neighborhood search method is 10.90% of that of k-D tree; among the four supervised learning algorithms of neural network (NN), support vector machine (SVM), Boosting and random forest (RF), the classification performance of the crown detector trained by RF is the best; compared with single kind of features, combined features have better generalization performance; the pointwise crown detector designed by this paper is obviously superior to the existing method in terms of detection accuracy and efficiency. When the radius δ changes within the range of 0.1-1.0 m, the F₁ score on the test set is ≥ 97.74%. 【Conclusion】 The proposed method can quickly and accurately detect street-tree crown point clouds from MLS point-cloud data collected in real time and provide a spray prescription map for targeted spraying applications on street trees.