Thomas Vögtle
(a) (b) (c) (d)
Figure 8: Segmentation of roof planes: (a) L-shaped hip roof, (b) flat roof, (c) complex shape, (d) hip roof;
upper row: vegetation reduced laser scanner dDEM of buildings (see section 3.2)
lower row: resulting segmentation of roof planes
Before performing this intersection of two neighbouring planes it has to be verified, if there is a significant elevation
difference at the edge of these two planes. To avoid disturbances single rejected pixels around this edge (caused by
dormers, chimney etc.) are eliminated by a simple majority decision out of the N8 neighbourhood. In the case of
significant elevation differences, a vertical plane has to be introduced. The position of these vertical planes can be
estimated by an adjustment procedure based on the respective edge gradients in laser DEM. Hence, the intersection
process provides two contour lines of the same x, y position but different elevation levels. The resulting contour lines
and corner points built up a 3D wire frame model of the building (CAD model). This could be subsequently processed
for various purposes, e.g. numerical applications, statistics or visualization.
NV
bii
Figure 9: Reconstructed buildings as wireframe models in vertical and perspective view;
test building 1 (left side) and 2 (right side), compare Figure 8 (a) and (d)
The results of determination of contour lines and corner points are shown in Figure 9 as wireframe models in vertical as
well as in perspective view. It is obvious that the geometry of the buildings has been reconstructed correctly and
completely according to the result of the segmentation process (see section 4.1). At corner points where more than three
detected planes intersect, ambiguous solutions occur. In this case an adjustment has to be performed regarding the
probabilities of each plane. These probabilities have been already been estimated during segmentation process, derived
from the membership probability of each pixel to the respective plane. On the other hand, this fact enables the
estimation of the interior accuracy of the presented method. First experiences with building modelling in our test areas
have resulted in an accuracy of about +/- 0.2m to 0.3m in position and about +/- 0.05m to 0.10m in elevation.
932 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.