Indoor point cloud aggregation and simplification based on parsimony of description

Abstract This paper presents a framework for automatically generating interpretable LoD2 indoor 3D models from unstructured point clouds. The method treats indoor modeling as a cartographic generalization problem, introducing a 3D corpus of geospatial feature classes that represents indoor environments as parsimonious compositions of geometric primitives organized by spatial grammar and compatible with IndoorGML. A sparse convolutional neural network simplifies, and aggregates point clouds by learning recurring geometric primitives and their spatial relationships, reducing data complexity while preserving essential structural information. Experiments on terrestrial LiDAR, BLK360, and RGB-D datasets achieved up to 97% simplification and 84% planar surface segmentation accuracy, demonstrating robustness across sensors and point densities. The generated LoD2 models are suitable for applications such as smart cities, digital twins, and indoor navigation. Current limitations include low detection accuracy for doors and windows (around 30%) and reduced performance in environments containing curved geometries.