Automatic Reconstruction of Indoor Spaces from 3D Point Clouds

Abstract

We present a new take on the unresolved challenge of automating indoor environment reconstruction from LiDAR point clouds. Utilizing point clouds as a basis for creation of BIM models yields highly accurate results and simplifies the task substantially as compared to gathering and using manual measuring methods. It is however still a time-consuming and labor intensive process for a human to draw the model using the point cloud as a mere blueprint. We therefore attempt to automate a key part of the process end to end, namely the reconstruction of polygonal room spaces. With the goal of reaching human accuracy, if not above, we attempt to automate the steps from point cloud to a 2D room-level polygonal model of a building floor.

To this end research was conducted to combine some promising methods from different studies that have previously been done in the field into a modular data pipeline. The prototype uses multiple high-performing algorithms to denoise the point cloud, accurately identify planar room dividing components and finally define the room spaces as simple polygons.

Our work shows that end-to-end automation of room space classification is indeed possible, although lack of objective measure of room divisions poses a yet unresolved challenge. For the purpose of full BIM model reconstruction, reliable room classification is a necessity. Our work shows a promising way of combining available methods into an automatic and robust indoor environment space reconstruction process with a high level of accuracy.