Point cloud scan selection for indoor floor plan generation

Abstract

Building Information Models (BIM) are becoming a standard in the construction indus- try for storing information about buildings and assets. Automatically creating BIMs has attracted a lot of attention, as it has great potential to improve efficient resource man- agement. A detailed description of the building can decrease the cost of management, heating and cooling, and restoration. For pre-existing structures design documents are typically outdated or unavailable, making BIMs challenging to acquire. The field of indoor floor plan creation has grown in recent years due to advancements in LIDAR technology. However, LIDARs create millions of points per scan, making it computationally expensive to process all of them. In order to properly create a floor it is imperative to acquire a sufficient number of scans to visualize the whole building, while simultaneously minimizing the number of scans for computational reasons. We propose a method for selecting a subset of the scans, as well as a method for clustering points into lines to be used for floor plan extraction. Our method works by clustering nearby points, creating a convex hull around them, and selecting scans based on the most area covered by the union of the hulls. The point clustering splits the pointcloud into potential lines by projecting each point along its surface normal, clustering points from the same line together. Those improvements allow for the efficient generation of floor plans for large buildings.