Automatic Geometric and Semantic
Reconstruction of Buildings from Images
by Extraction of 3D-Corners and their 3D-Aggregation
Overview
Overview over the different steps
The procedure for 3d-building extraction starts with the observed 2d-image descriptions.
Using a hierarchical building model the reconstruction is performed over different levels of
aggregation and uses stepwise more precise knowledge of the building to reconstruct.
The process of reconstruction is devided into local and global reasoning steps.
The locally consistent reconstruction
of object parts of one aggregation level leads to observations for
then next grouping and aggregation steps.
The complete procedure for building reconstruction can be devided into two parts:
- a. 3d-corner reconstruction
- b. 3d-corner aggregation
Building parts of type corner are the connection between image- and object space and are connecting the two reasoning steps during building extraction.
Fig 1: shows the strategy for building reconstruction. The procedure consists of three different steps:
a. 2d-aggregation until the level of feature aggregates,
b. 3d-reconstruction of feature aggregates and their interpretation as corners and
c. 3d-of corners to generic buildings
Preprocessing
As input data for building reconstruction we use digital images with multiple overlap.
The sensor orientation (exterior and interior information) is assumed to be given.
We divide the building extraction task into the two steps detection and reconstruction.
We assume the detection to be done leading to image patches, in which the objects are to be reconstructed.
In literature different procedures for detection are presented e.g. using Digital Elevation Models (cf. DEM- Analysis for Building Extraction), Color Information or GIS. Another possibility is to perform the segmentation interactively within an semiautomatic system for building extraction for example like the one presented in Semiautomatic 3D-Building Reconstruction or its further developement.
The detection is possibly followed by further preprocessing steps like noise estimation and information preserving filtering. For suppression of undesired texture influence the images may be resampled to an appropriate
ground resolution.
The first step during the analysis is the 2d-aggregation in image space as shown in the figure.
Starting from the iconic image description we derive a polymorphic, symbolic image description.
We use the system FEX which leads to extracted points, lines and regions and their mutual neighbourhood relations stored in a feature adjacency graph.
The feature extraction process may be followed by a domain-independent grouping process, but is not used in this work.
The symbolic image description enables to compare the image description with the object model.
Why to use corners for the link between image- and object space?
Following information theoretic investigation the central information for describing object contours is contained in points of extreme curvature. Such corner points in may cases are sufficient to describe polyhedral objects by connecting corners. This motivates us to perform building reconstruction
by using building components of type corner and their connections.
We propose a close interaction between geometric and semantic reconstruction
which enables to use class specific geometric, topologic and structural constraints. Corners seem to be appropriate building components for an
early integration of the semantics while reconstruction.
Especially in object space the geometric form description of objects gives
much information for the interpretation of geometic objects
as components of the aggregation hierarchy.
Thus we use the aggregation level of feature aggregates for the transition to semantic building components of type corner.
The reconstruction strategy thus is defined by 3 different processes:
- a. 2d-aggregation
until the level of feature aggregates
- b. 3d-reconstruction of feature aggregates and their interpretation as corners
- c. 3d-aggregation of corners to buildings
The interaction between geometry, topology and semantics is done by generation hypotheses and verification of hypotheses and uses the corner specific constraints in three ways:
- use of corner structure and representation:
The symbolic image information on the aggregation level of feature aggregates
is suitable for the correspondence analysis, because it is not much sensible to disturbances and structural variation of the objects.
Feature aggregates contain more characteristics than single points, lines or reagions. The enable the definition of a meaningful similarity measure during correspondence analysis. Neverthelesse feature aggregates are local enough to be insensitive to occlusion and other disturbances.
The structural representation using points, lines and regions can be compared to the b-rep as representation of polyhedral objects and thus enables the link between image and object space.
- use of the corner geometry:
The corner geometry is used for interpretation by identification of class specific constraints. Especially the 3d geomety is very meaningful for linking geometry and semantics.
- use of the corner semantics:
The semantics defines geometric, topologic and structural constraints, which serve for stabilization duroing the verification of corner hypotheses.
Further on the interpretation of corners facilitates the aggregation of corners to buildings. The geometry as well as the interpretation of the corner componentes and of their connectivity relations defines
constraints during the grouping and aggregation process.
A Detailed Overview of the procedure is given by this picture.
FL - 30.4.99