Unsupervised object extraction for semantic segmentation of bridge scenes
Philipp Tornow
(Prof. Dr.-Ing. Volker Rodehorst, PD Dr. Andreas Jakoby, Paul Debus, Jan Frederick Eick)
An increasing interest in 3D models of existing bridges leads to pressing need to automate the creation of such models. Current research achieves efficient acquisition techniques running partially automated. However, the extraction of the bridge model is still mostly done manually using current software. This thesis proposes a pipeline for fully automated bridge extraction from infrastructure scenes. The process is divided in preprocessing steps, a detection of the bridge slab and filtering methods.
The preprocessing simplifies and augments point cloud data, the slab detection extracts the position as well as the dimensions of the bridge and different filtering steps remove 3D data that are not part of the bridge. Furthermore, together with first steps for semantic part segmentation, achieves this work a subdivision of the
extracted bridge in deck, slab and beams. The evaluation and implementation demonstrates the performance of the approach and outlines the problem of downsample point clouds by methods which produce varying point distances across a testing dataset.