AUTOMATIC RECONSTRUCTION OF SKELETAL STRUCTURES FROM TLS FOREST SCENES

Abstract. Detailed information about 3D forest structure is vital in forest science for analyzing the spatio-temporal development of plants as well as precise harvest forecasting in forest industry. Up to now, the majority of methods focus on complete structural reconstruction of trees from multiple scans, which might not be a suitable starting point with respect to modeling forest scenes over larger areas. For this reason, we propose a strategy to obtain skeletal structures of trees from single scans following a divide-and-conquer approach well-known from computer science. First, we split the range image into components representing surface patches and trace each component’s boundary, which is essential for our skeleton retrieval method. Therefore, we propose an extension to standard boundary tracing that takes a component’s interior depth discontinuities into account. Second, each component is processed separately: A 2D skeleton is obtained via the Voronoi Diagram and refined. Subsequently, a component is segmented into subsets of points depending on their proximity to skeleton nodes. Afterwards, the skeleton is split into paths and a Principal Curve is computed from each path’s point cluster in order to retrieve the component shape as a set of 3D polygonal lines. Our method retrieves the intricate branching structure of components representing tree crown parts reliably and efficiently. The results are fitting with respect to a visual inspection. At present, results are fragments of tree skeletons, but it opens an attractive perspective to complete tree skeletons from skeletal parts, which we currently regard as future work.