骨架优化下的地面激光树木点云重建方法

本文针对地基式激光雷达扫描点云的分布特征,提出一种基于骨架优化的三维树木重建方法。首先从原始点云中提取出树木的“类主干点”,利用无向图数据结构存储和组织点云,然后基于“主干点”约束下的最小生成树算法生成单株树木的初始骨架,最后通过“密度调整”和“树枝平滑”优化操作,建立高逼真度的三维树木模型。本文提出的树木几何重建算法对点云数据缺失和点云的密度不敏感,对试验区不同的树种建模具有较高的鲁棒性。另外,本文通过对不同密度的点云和不同抽稀层次下的点云进行效率测试,发现结合类主干点表达和点云抽稀能共同确保本文算法可以重建大场景TLS植被点云数据。

Individual Tree Modeling from Terrestrial Laser Scanning Point Clouds via Skeleton-based Optimization

In this paper, we propose a methodology to reconstruct the individual-tree models from Terrestrial Laser Scanning (TLS) point clouds via skeleton-based optimization. The proposed method is a data-driven method, and in theory, it can generate any geometric shapes of different tree species. Mathematically, the salient points reflected from trunks and remarkable branches have been successfully recognized by using statistical analysis method. Then, we organize the raw points via an undirected graph, from which the initial skeleton of individual tree is created by using the Minimum Spanning Tree (MST) algorithm under the constraint of salient points. The initial skeleton tree models are further enhanced and refined through a series of optimizations, i.e., point density adjustment and branch smoothing. The tree skeletal structure is inflated into a tree model by simultaneous combination of a robust cylinder fitting method and allometric model. The tree leaves are finally properly added into the tree models, thereby enhancing the photorealistic representation of the geometric tree models. Various experiments on different tree species captured at Nanjing Forestry University show that the proposed methodology is insensitive to the point density and data missing, and meanwhile can generate meaningful and accurate individual geometric tree models. In addition, it is to be found that our modeling tree algorithm based on salient points and sampling strategy can reach the optimal computational efficiency, compared to that only using the salient points or the original laser scanning tree points as inputs. This enhancement in the efficiency can significantly expand our current individual tree modeling into the reconstruction of large-scale scanning scenes.