BIM-compliant reconstruction of 3D geometries and semantic information for building modeling

Overview

The project BIM-compliant building reconstruction is an explorative subproject in the flagship project i_city. The intelligent city of the future can only be created on the basis of high-quality data. Especially building and city models are at the centre of such a development. For precise and three-dimensional representation of buildings, Building Information Modeling is becoming increasingly popular in the planning phase of new buildings. However, existing buildings are usually not modelled as BIM-models. In this project a possible concept for such a BIM-compliant building survey is proposed.

Research Question

Various state-of-the-art methods known from geodesy are available for recording buildings and interior geometries. However, their workflows are very time-consuming, either when measuring or processing for such highly complex three-dimensional BIM-models. In addition, the semantic information that is important for the BIM-models is not determined automatically, but has to be noted down by hand and transferred to the BIM-model manually. Automatic extraction of this component from the measurement data is not yet available. In order to enable simple modelling, a new procedure towards automation of BIM-compliant building reconstruction is being developed in this project. This is illustrated by the implementation in a demonstrator application.

Research Methods

Based on an investigation of suitable recording procedures, a concept for recording and processing was developed. The combination of mobile laser scanners and image-based photogrammetry form the basis for automated processing.

The central component of this is - in addition to the geometric information of the point cloud - especially the semantic information contained in the images. Based on these, an automatic and pixel-level extraction of the object type is carried out using deep learning methods. This forms the foundation for the assignment of the points of the photogrammetric point cloud to clear categories. This forms the foundation for labeling each point in photogrammetric point cloud using the segmented images.

The inherent geometric and semantic information of the point cloud is the basis for BIM-compliant modeling.

 

Further steps of the concept:

- Combination with mobile laser scanning point cloud

- Extraction of objects and their properties

- Extraction of further semantic information from the data

Results

  • The concept was implemented in a demonstrator
    • High-quality trained neural network for semantic segmentation of indoor spaces
    • Classified point cloud by projecting the segmented images based on position and rotation
    • Automated post-processing of the point cloud based on semantic information.
    • Basis for further work towards a BIM model was established
  • A BIM model was manually derived from a comparison point cloud acquired with laser scanning
  • Linking between interior and exterior areas is possible. 
  • An approach to automation has been developed and validated
     

Conclusio

  • Photogrammetry and deep learning methods complement each other and make high-quality use of the information available in the images
  • A combination with mobile laser scanning enables the generation of a total point cloud for the complete modeling of interior and exterior areas
  • Further automation of BIM modeling of existing buildings based on the geometric and semantic object and component information extracted here is possible
ManagementProf. Dr. Eberhard Gülch, Prof. Dr. Michael Hahn
FundingFederal Ministry of Education and Research (BMBF)
Call for proposalStrong universities of applied sciences - impulse for the region (FH-Impuls)
Duration01.08.2017 - 31.08.2021
(Start of project work: 01.11.2017)

 

Team