A STOCHASTIC APPROACH TO AUTOMATED RECONSTRUCTION OF 3D MODELS OF INTERIOR SPACES FROM POINT CLOUDS
Automated reconstruction of 3D interior models has recently been a topic of intensive research due to its wide range of applications in Architecture, Engineering, and Construction. However, generation of the 3D models from LiDAR data and/or RGB-D data is challenged by not only the complexity of building geometries, but also the presence of clutters and the inevitable defects of the input data. In this paper, we propose a stochastic approach for automatic reconstruction of 3D models of interior spaces from point clouds, which is applicable to both Manhattan and non-Manhattan world buildings. The building interior is first partitioned into a set of 3D shapes as an arrangement of permanent structures. An optimization process is then applied to search for the most probable model as the optimal configuration of the 3D shapes using the reversible jump Markov Chain Monte Carlo (rjMCMC) sampling with the Metropolis-Hastings algorithm. This optimization is not based only on the input data, but also takes into account the intermediate stages of the model during the modelling process. Consequently, it enhances the robustness of the proposed approach to inaccuracy and incompleteness of the point cloud. The feasibility of the proposed approach is evaluated on a synthetic and an ISPRS benchmark dataset.