Talks and Poster Presentations (with Proceedings-Entry):

L. Eysn, C. Ressl, A. Grafl, M. Hollaus, W. Mücke, F. Morsdorf, N. Pfeifer:
"Extraction of 3D tree models based on equirectangular projections of terrestrial laser scanning data";
Poster: SilviLaser 2012, Vancouver, British Columbia, Canada; 2012-09-16 - 2012-09-19; in: "SilviLaser 2012 - Conference Proceedings", (2012), Paper ID SL2012-091, 1 pages.

English abstract:
Extracting 3D tree models based on high-density terrestrial laser scanning (TLS) point clouds
with automatic, semi-automatic or manual methods is a challenging task as trees are complex,
individual objects. Various publications in this research field show the demand for tree
reconstruction methods based on TLS data. Data sets acquired with current TLS devices allow
detailed reconstruction of tree stems and branches, which are a fundamental input for e.g. stem
volume assessment or setting up virtual forest scenes. For dense forests, a completely automated
reconstruction of a tree is often limited by occlusions and data gaps, as well as varying point
density. Beside automatic extraction methods based purely on the point cloud (e.g. region
growing algorithms), tree models can be semi-automatically created by e.g. local cylinder fitting.
This task is challenging and time consuming because the interpreter has to navigate through
dense point clouds and the selection of subsets for cylinder fitting can be tricky.
In the presented approach a semi-automatic method for extracting coniferous and deciduous tree
models based on projected 2D maps of the TLS point cloud is described. Equirectangular
projections (EP) based on the observation angles of the scan are created, thereby displaying the
distance (range map, RM) and intensity information (intensity map, IM) detected by the
scanning device. The so-called tree structure elements (i.e. stems and branches) are clearly
interpretable in the IM and RM. These easily navigable maps provide a good basis for extracting
trees by digitizing the axis of the structure elements and assigning their respective local
diameter which were measured in the 2D maps. EP derived from multiple scan positions around
the trees are used to complete occluded sections. Erroneous measurements, arising from moving
tree parts (e.g. branches affected by wind), or by imperfections in the relative orientations of the
scans, are overcome because the extraction of the tree structure is performed using single maps
instead of a merged point cloud of individual scans. The digitized 2D skeletons are transformed
to 3D space and furthermore extruded to 3D models. Additionally to the modelling process a
classification of the point cloud into two classes (stem/branches, needles/leaves) is performed
using a voxel approach. The tree models are a fundamental input for this classification.
The method is applied to a dense TLS dataset acquired in a managed forest in Tharandt,
Germany. About 34 scans were carried out during the data acquisition to measure approximately
90 spruce and fir trees with minimal occlusions. The results demonstrate the feasibility of
extracting tree models semi-automatically based on 2D maps with a very high degree of
completeness. In comparison to other approaches, the number of reconstructed trees is higher
(by factor 3) than the number of scans. The quality assessment was based on a comparison of
the point cloud and the cylinder modelīs 3D views and quantitative evaluation. Visual
assessment showed deviations in the order of the measurement accuracy and tree surface
irregularity. The extracted tree models are used to set up a virtual forest scene for radiative
transfer modelling

Electronic version of the publication:

Created from the Publication Database of the Vienna University of Technology.