Extraction of Semantic Floor Plans from 3D Point Cloud Maps

3D mapping is increasingly important for mobile robotics in general and for safety, security and rescue robotics (SSRR) in particular as complex environments must but captured in this domain. But it is hard to visualize 3D data in a simple way, e.g. to print maps for first responders, or to use it in standard robotics algorithms, e.g., for path planning. This paper describes a new approach to extract standard planar maps from large scale 3D maps in a very fast manner. In doing so, the approach can detect multiple floors, e.g., in a multi-story building or in a pancake collapse, and segment the 3D map accordingly. To each floor or level, a planar map is assigned, which is augmented by semantic information, especially with respect to traversability. Experiments are presented that are based on 3D maps generated in the large scale environments of USARsim, a high fidelity robot simulator. It is shown that the approach is very fast. The total processing of a complete 3D map takes just a few hundred milliseconds, leading to a proper extraction of floor plans to each of which semantic maps are assigned.