[Back]

S. Mikolka-Flöry:
"Comparison of Selected Segmentation Algorithms of 3D Point Clouds";
Supervisor: N. Pfeifer, J. Otepka; Department für Geodäsie und Geoinformation, 2019; final examination: 2019-10-10.

Segmentation is an important step in the processing pipeline of 3D point clouds. It can be used to identify and extract individual objects or build the base for a subsequent classi cation. Within this work two methods, so far mainly used in the eld of computer vision, are extended to the segmentation of 3D point clouds.
The rst method, based on the minimum spanning tree of the neighborhood connectivity graph, shows promising properties enabling segmentation on an object-based level. The second method, SLIC (Super Linear Iterative Clustering), is designed to create a strong oversegmentation which can be used to reduce the amount of data in the rst place.
To evaluate both approaches a new metric, termed completeness, is introduced. In contrast to other commonly used metrics like the number of segments or mean segment size, completeness measures the quality of the segmentation with respect to individual objects. The investigated datasets represent di erent scenes (urban/rural), show di erent point densities and contain objects of di erent size, shape and color.
All conducted tests show that both approaches are suitable for the segmentation of 3D point clouds. While the graph-based method improves segmentation in inhomogeneous regions, SLIC is an useful option for point clouds with higher point densities. Particularly their combination seems to be an interesting option for the segmentation of more complex scenes.

https://repositum.tuwien.ac.at/urn:nbn:at:at-ubtuw:1-129874