基于机载激光雷达点云的交互式树木分割与建模方法研究

机载激光雷达扫描(ALS)系统可大规模获取地表树木点云,有助于较高精度树木结构参数提取和景观层面的几何重建,然而树木复杂的拓扑结构和树种的多样性给树木的准确分割与建模带来挑战.传统基于点云的自动树木分割和建模算法虽然效率高,但存在分割误差较大、建模鲁棒性较差等问题,难以满足深度学习大背景下用户对树木分割与建模结果进行精...

Interactive tree segmentation and modeling from ALS point clouds

The airborne lidar scanning (ALS) system provides the possibility of acquiring large-scale tree point cloudsonly from a single scan, which helps to achieve the structural parameter extraction of trees and geometricreconstruction at the landscape level with higher accuracy. However, it remains a challenge to accurately segment andmodel trees from ALS point clouds, due to the diversity of tree species and the complex topology of trees. Althoughthe traditional point cloud-based automatic tree segmentation and modeling algorithm are efficient, there remain suchproblems as great segmentation errors and lack of robustness for the modeling algorithm, making it difficult to meet users’ need for the precise annotation of tree segmentation and of modeling results in deep learning. In order to solvethe problems of automatic segmentation and modeling caused by the sparse and incomplete ALS point clouds, aninteractive hierarchical segmentation method was proposed based on height mapping. The proposed method canextract a single tree point cloud from sparse point clouds, and then utilize the improved space colonization algorithm(SCA) to model trees by interactively adjusting the parameters of constraint angle, kill distance, and influence radius.Experimental results show that the proposed interactive segmentation algorithm can avoid the false segmentationarising from the minimum spanning tree and Normalized-Cut algorithm. The proposed interactive modeling algorithmcan robustly generate tree models from the sparse and incomplete ALS point clouds, as well as preserving the featuresof the original point cloud by selecting the appropriate combination of SCA parameters.