Petra Zimmermann
3.2 Elevation - Blob extraction, coarse Roof Modelling
We use the DSM to derive coarse models of the buildings containing coarse shape and extend and the number and
direction of the main axes. Given the aspect values within a blob we can get the direction of inclination of the roofs
(Figure 8), strong change in slope and aspects indicates a ridge line (Figure 9), also, if there is a change in aspect a
break-line is expected. Ridgelines are derived by computing the second derivatives: then a ridge point lies on a local
convexity that is orthogonal to a line with no convexity or concavity, which means:
oz? 7”
Ridge afin = (5)
X dy
For further reconstruction steps each blob with its position, an average height of the recognised ridge lines and the roof
boundaries with their position and height, and the flanking regions information concerning slope and aspect is stored as
"BasicRoofModel" (Figure 10). Due to noise in the DSM and low resolution no small or weak ridges and also no details
can be recognised.
2 ridgelines
aspect and
slope values
>0
Slope = 0
Aspect = 0
[0 | No ridgeline
Figure 6: aerial Figure 7: DSM blob, Figure 8: example Figure 9: slope of a Figure 10: scheme of ridges
image projection as image of aspect of a blob selected region and boundary lines
The geometric accuracy of the positions of the ridgelines is weighted very low. The extracted blobs with their
information and ridgelines are projected back to the aerial images to get an approximate location of objects with
elevation above ground.
3.3 Grey-level Edges and Colour Edges
Long straight edges may indicate the location of man-made objects and are directly applicable for edge-based matching.
We extract edges in colour imagery, focussing on long straight edges. Edge preserving smoothing is applied to the input
image, and then the Canny gradient operation derives the gradient information, which is used by a hybrid region-and
edge based classification. In colour space -depending on the algorithm RGB, HSV or L*a*b* -we apply gradient
filtering in each colour channel to derive colour edges to get slight improvements in edge detection compared to grey-
level edge detection. To merge the information we just overlay the edges from the singe channels. Edges detected in all
channels get a detection attribute with a weight of 3, whereas edges that are detected only in one channel get a weight of
1, each time multiplied with normalised gradient strength. Edge information is stored as Polygon2D with Endpoints of
the single Lines2D, the detection attribute, the accuracy of linefitting of each Line2D and the distance and angle to the
linked Line2D in the edge polygon, whether the polygon is closed, curved or straight is also stored. In processing other
cues additional information is added e.g. colour attributes or entropy of the flanking regions.
34 Colour
Homogeneous coloured regions are assumed to belong to the same regions. We apply colour segmentation to derive
homogeneous regions in different colour spaces: Saturation from HSV space helps to discriminate between shadow and
Figure 12: segmented blob regions
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 1067
