On the Automatic Reconstruction of Building Information Models from Uninterpreted 3D Models

1
On the Automatic Reconstruction of Building
Information Models from Uninterpreted 3D Models

• Thomas H. Kolbe
• Director of the
• Institute for Geodesy and Geoinformation Science
• Berlin University of Technology
• Joint work with Claus Nagel Alexandra Stadler
• kolbe nagel stadler_at_igg.tu-berlin.de
• Academic Track of Geoweb 2009 Conference,
  Vancouver

2
Building Information Models IFC

• Building Information Model (BIM)
• digital representation of the physical and
  functional characteristics of a constructed site
  or facility
• comprehensive information source on a facility
  aiming at collaborative usage
• intended to be used along the entire lifecycle of
  a facility
• key feature models have well-defined semantics
• Industry Foundation Classes (IFC)
• ISO standard for semantic building models
• diverse crafts/themes incl. billing of material
  and costs
• supported 3D geometry types CSG, Sweep, B-Rep,
  etc.

3
BIM Application 1 Energy Assessment

• Image ThermoRender, Nemetschek North America

4
BIM Application 2 Space Management
Image Space planning created with Onuma Planning
System
5
BIM Application 3 Structural Analysis

• Image Autodesk Robot Structural Analysis Brochure

6
Problem Statement

• BIM models are typically prepared for newly
  planned buildings only
• But applications should also be usable with
  existing buildings
• Acquisition method required for BIM models for
  existing buildings
• manual acquisition is expensive ? automation
  required
• Challenges
• What are appropriate data sources?
• Which problems have to be faced and how could
  they be overcome concerning the interpretation /
  reconstruction?

7
Starting Point 3D Geometry / Visual Models

• Preprocessed sensor data from LIDAR /
  Photogrammetry, i.e. point clouds or surface
  patches
• Visual models / surface based models (polygon
  soups)
• From CAD or computer graphics
• Characteristics of input data
• Pure geometry (and radiometry)
• Geometry can be unstructured or structured
  according to visualization purposes it can also
  be incomplete
• Topological errors (permeations, overshoots,
  undershoots)
• No semantic information

Photogrammetric models
Airborne laser scan models
CAD and planning models
Visualization models
8
Goal Reconstruction of BIM models

• Reconstructed BIM models
• explain (most of) the observed geometrical
  entities in the best way
• are composed of fully classified and attributed
  entities like Walls, Slabs, Roofs, Spaces, etc.
• thus, they are semantically rich and structured
• semantics follow the IFC standard
• have volumetric, parametric geometries (CSG)
• required in order to make the models editable by
  CAD tools
• should make hypotheses about 3D components with
  respect to invisibility / unobservability

9
IFC Example
10
Surface-based modeling of Interior Space
11
Two-stage Reconstruction Process

• Automatic reconstruction of BIM models
• from 3D geometry models faces a high level of
  complexity
• Unstructured, uninterpreted geometry ? semantic
  classification
• handled in Stage 1graphics model ?
  semantically enriched boundary model
• Accumulative ? generative modelling paradigm
• handled in Stage 2semantically enriched
  boundary model ? building information model with
  volumetric, parametric components

12
Urban Models and their Applications
13
CityGML

• OGC Standard for virtual 3D city models
• spatial and
• thematic disaggregation / semantic modeling
• LOD4 Building model including interior space
• Modeling paradigm BRep explicit semantics
• Close to photogrammetric / lidar observations
• In fact, closer than IFC models using CSG

14
Stage 1 Graphics model ? CityGML

• classification stage
• Purely geometric graphics models (e.g., KML) are
  converted to semantically enriched boundary
  models (e.g., CityGML)

15
Underneath the surface

• Visual models are explicitly built for
  visualisation? only visible parts are trustworthy

16
Strategies for Geometry Handling (I)
17
Strategies for Geometry Handling (II)
A Keep original geometry

• Geometry remains unchanged
• Merely attach semantic information to polygons
• Transform polygon soup into structured geometry
  aggregates
• No effect on coordinate values
• New geometry is generated according to target
  model and fitted to observations
• May result in topological changes (e.g. closing
  of volumes)
• New geometry is generated according to target
  model and fitted to observations
• After geometric/semantic structuring keep
  original geometry in final result

B Structure geometry
C Replace geometry
D Additional requirements on the target model
18
Level-of-Detail (LOD) Concept

• How to decide on the appropriate target model
  LOD?

Automatic LOD recognition

• Conclusions about target model LOD according to
  input model granularity
• E.g., no window setoffs or molded roof structures
  ? LOD2
• User specifies target model LOD
• Attention input model may not fulfill
  requirements of the chosen LOD
• Specification of one basic target LOD (automatic
  recognition or user input)
• Build LOD series covering all lower LODs
  (probably using generalization)
• Explicit linkage between multiple LOD
  representations

User input
Build a LOD series
19
Stage 2 CityGML ? IFC

• Reconstruction of component-based volume model
  from a surface model
• Instantiation and rule-based combination of
  volumetric building objects (walls, roofs, )
  which most likely explain the input model
• CityGML model is seen as the observation, IFC
  model will be the interpretation result
• Key aspect Semantic information as a priori
  knowledge
• Both CityGML and IFC provide semantic models of
  the built environment
• Allows for reducing the search space of potential
  IFC elements
• Complexity results from the fact that
• CityGML and IFC follow different modeling
  paradigms
• Building components are only observable in parts
  or not observable ? typ. only observable parts
  are contained in input 3D model
• From each component two or more surfaces may be
  observable
• Represented as individual semantic entities in
  CityGML

20
Differing Modeling Paradigms
BIM (e.g., IFC) Constructive Solid Geometry
3D GIS (e.g., CityGML) Boundary Representation
Volumetric, parametric primitives representing
the structural components of buildings
Accumulation of observable surfaces of
topographic features
21
Matching between CityGML and IFC Entities

• Generation of IFC element hypotheses from CityGML
  entities
• Semantic information as a priori knowledge
• Evaluation of geometric-topological relations
  between CityGML entities

• n CityGML entities may represent one IFC element

• n CityGML entities may result in m competing IFC
  elements

• Further 11 and 1m relations possible ? High
  combinatorial complexity

22
Instantiation of IFC Elements (I)

• Instantiation of CSG primitives which best fit
  the spatial properties of all matched CityGML
  entities
• Man-made objects often deviate from the idealized
  CSG shape
• Parameter estimation has to obey contextual
  constraints
• Unary usually impair the best fit of a single
  element
• Mutual aim at aligning elements ? affect
  parameters of many elements
• Conversion of B-Rep to CSG in general is
  ambiguous
• Building components are only observable in parts
• CSG primitives cannot be derived from closed
  volumes

• Competing hypotheses
• Requires additional a priori knowledge /
  assumptions

23
Instantiation of IFC Elements (II)

• Purely geometric-topological constraints on IFC
  primitives cannot prevent unreasonable element
  hypotheses
• E.g., IfcRoof elements at the bottom of a
  building
• Both CityGML and IFC do not explicitly qualify
  objects and inter-object relations in order to
  ensure sensible configurations
• Based on UML, XSD, and EXPRESS
• Focus on generic notion of objects and
  associations
• Reconstruction requires a framework providing
  enhanced model expressiveness
• Physical, functional, semantic / logical object
  constraints
• Rules for structural valid element configurations
• What makes a valid building?

24
Interpretation Strategy (I)

• How to express / formalize knowledge about 3D
  building models?
• CityGML and IFC data models do not provide
  (formal) constraints on object instances
• A more expressive formal representation is
  required specifying how complex objects are
  aggregated in a logically / semantically sound
  way
• ? Formal Grammars are becoming applied
• increasingly often for this purpose
• Formal grammars originate from computational
  linguistics
• Definition of different classes of grammars by N.
  Chomsky later extended by D. Knuth (attributed
  grammars)

25
Example of a Formal Grammar (in EBNF)
26
Interpretation Strategy (II)

• Requirements on the grammar
• Words / Objects have attributes (attribute
  grammar)
• Geometric Shapes (shape grammar / split grammar)
• Stochastical aspects (a priori probabilities)
• Combination of all grammar types is required
• Further requirements
• Need of an evaluation / objective function
• In order to determine the best interpretation
  from all possible hypotheses
• meaningful definition of best interpretation
• Using probability theory the most likely model
  under the given data
• Avoid overfittings

27
Conclusions

• Reconstruction of BIM models is a specific
  instance of the general 3D object recognition
  problem
• What makes it especially difficult?
• Gap between observed surfaces and volumes to be
  reconstructed (BRep ? CSG ambiguities)
• High structural and semantic complexity of BIM
  models
• Uncertainty, unobservability, and errorneous
  observations
• Definition of an objective function / measure to
  compare the appropriateness (probability?) of BIM
  model hypotheses
• Which aspects help in this process?
• Two-stage strategy allows for a step-wise
  interpretation and extraction of semantic
  information (divide-and-conquer)
• IFC and CityGML are target models with
  well-defined semantics