Automatic Geometric and Semantic

Reconstruction of Buildings from Images

by Extraction of 3D-Corners and their 3D-Aggregation

Felicitas Lang ,Institute for Photogrammetry, Bonn University

3D-Corner Reconstruction (Corner EXtraction CoEX)

Overview:

The corner reconstruction results in locally consistent 3d reconstructions of building specific object parts. They can be used as derived observations for the subsequent 3d aggregation of building components. 3d corners are suitable for the aggregation to buildings as they can be used for formulating geometric as well as semantic constraints for the higher aggregation levels. The reconstruction and interpretation of the image data and the derived symbolic image description is performed by using the hypothesize and verify paradigm. Thus the process for deriving corners can be devided into three steps, which are performded sequentially:

Extraction of a symbolic image description: 2d aggregation of the iconic raster level to the feature level of the hierarchy. We use a suitable feature extraction which derives a polymoprphic image description consisting of points, lines and regions and the corresponding feature adjacency relations. The feature adjacencies leads to a grouping of features and gives a basis for aggregating features to objects of a higher level of our modeling hierarchy.
Generation of Corner Hypotheses: by deriving the 3d geometry of feture aggregates and their class membership. The transition to 3d is performed by multi-image correspndence of selected 2d feature aggregates, which are selecte by analysing the feature adjacency graph. The class membership is established by interpretation of the 3d geometry of the 3d feature aggregates and leads to corner hypotheses. During the building reconstruction they are used for integrating object knowledge.
Verification of corner hypothese: During the verification the corner hypotheses are tested and evaluated by backprojecting the corner instance into the images and newly reconstructiong their geometry. In contrast to the corner generation step the maximum available corner knowledge on the level of the feature aggregates can be introduced during the selection of suitable observations for the reconstruction of the corner.

Generation of Corner Hypotheses:

Selection of 2d vertices:
correspondence analysis of 2d vertices:

Fig 1: shows the epipolar constraint for the corner points of a corresponding vertice pair.
Deriving 3d vertices:

Fig 2: shows two examples of possible corresponding vertex pair. Structural constraints have to be fulfilled during the relational matching of the vertex components while establishing the correspondence of the components and define restrictions during the matching.
Interpretation of 3d vertices as corners:

Fig 3: shows different corner subclasses, that are sufficient for describing the building types flat roof, prismatic roof, gable roof, hip roof and crippled hip roof.

Verifcation of Corner Hypotheses:

Fig 4: shows the parameters of a saddleback roof corner. 5 parameters are sufficient for describing the corner geometry (3 parameters for the corner point coordinates, 2 angles for orientation and roof slope).

hr> FL - 30.4.99