Search

Digital twins for buildings can significantly reduce building operation costs. However, existing methods for constructing geometric digital twins fail to model the complex geometry of indoor environments. To address this problem, this paper proposes a novel method for digitizing building geometry with arbitrary shapes of spaces by detecting empty regions in point clouds and then expanding them to occupy the entire indoor space. The detected spaces are then used to detect structural objects and transition between spaces, such as doors, without assuming their geometric properties. The method reconstructs the volumetric representation of individual spaces, detects walls, windows and doors between them and splits the point cloud data (PCD) into point clusters of individual spaces from large-scale cluttered PCDs of complex environments. We conduct extensive experiments on Stanford 3D Indoor Spaces data set (S3DIS) and TUMCMS data sets and show that the proposed method outperforms existing methods for digitizing Manhattan-world buildings. In contrast to existing approaches, the method allows digitizing buildings with arbitrarily shaped spaces, including complex layouts, nonflat, nonvertical walls, and nonflat, nonhorizontal floors and ceilings. [ABSTRACT FROM AUTHOR]

The paper presents a novel method for automatically creating semantic digital models for buildings in Manhattan environments from dense point cloud data. Unlike previous approaches, which rely solely on data-driven methods, our method integrates artificial intelligence with domain engineering knowledge to overcome challenges in indoor point cloud processing and geometry representation in complex layouts. A feature-based Decision Tree classifier extracts main building elements, utilized in a knowledge-based algorithm for 3D space parsing. On this basis, an optimization process generates parameterized floor plans, used to finally create volumetric digital models. The method was validated on datasets from the Technical University of Munich and Stanford University, achieving a mean accuracy of approximately 0.08 m for model placement and 0.06 m for estimating element parameters, which highlights its effectiveness for generating a building's semantic digital model. This approach underscores the potential of AI integration in digital twinning workflows for more automated solutions. [Display omitted] • Designing a hybrid bottom-up, top-down approach for automatic building digital model creation. • Align the capabilities of AI in scene understanding with domain knowledge to overcome challenges in indoor 3D reconstruction. • Developing a knowledge-based method for 3D space parsing in largescale indoor point cloud. • Reconstruction of the parametric building's floor plan using point cloud with low sensitivity to poor data quality. • Creation of dynamic parametric building digital models (LOD200) for flexible facility management and space analysis. [ABSTRACT FROM AUTHOR]