Michel Morgan
For the building in figure (3-a), the direction of the extracted building is the same as the direction of the building in the
ground truth data but the dimensions of each rectangle differ from the ground truth values. The two extracted faces of
the building slightly overlap within the pixel size. This building is on the edge of the ground truth data and it was not
known whether the building in the ground truth data is trimmed at the upper part of the figure. Moreover, the ground
truth data does not have face information. Therefore, only the lower left and lower right coordinates of the building in
the figure are used for accuracy estimation. The planemetric and the height accuracy are 2.25m and 0.16m respectively,
The accuracy achieved for the height (0.16m) is very good compared to the standard deviation of elevation from laser
scanning (0.10m). The planemetric accuracy (2.25m) is very low in comparison to the pixel size (0.577m). The results
of building detection (the gray pixels in figure (3-a)) are better than the results after vectorisation (the dotted lines in the
same figure). As shown in the same figure the extracted faces are smaller than the face segments in raster format. Area
constraint for each face may be used as a constraint in the rectangle extraction process (hopefully to extract the correct
rectangle dimensions). Moreover, the rectangle extraction could be constrained to the direction of the intersection line
between the two planes of the faces.
For the second building, the extracted rectangle of the upper face in the figure (3-b) is rotated by 4 degrees with respect
to the ground truth data. Again because there is no face information in the ground truth data, only the upper right and
upper left corners of that face will be checked. The planemetric and the height accuracy are 0.55m and 2.47m
respectively. The achieved planimetric accuracy (0.55m) is very good and it is within the pixel size (0.577m). The very
large error in height is caused by the erroneous plane parameters of the faces. The segmentation of the building segment
into sub-segments (faces) is too sensitive to noise because the criteria for segmentation are applied for pairs of adjacent
pixels. Therefore, the criteria for segmentation must not be applied only for each two adjacent pixels while doing region
growing, but also for the pixel and the growing sub-segment. The second face in the building is not rectangular as can
be seen in the figure, therefore the rectangle extraction results in a rotated rectangle with different dimensions than that
in the ground truth data. Constraining rectangle orientation to the direction of the intersection between the faces could
lead to better results for the second face.
For the building shown in figure (3-c), the boundaries of the extracted faces are generalized because only one line
fitting is done for each face adjacent to the background. For the four main building corners, the planemetric and the
height accuracy are 0.73m and 1.30m respectively. The planimetric accuracy (0.73m) is 1.3 times the pixel size
(0.577m). The reason behind achieving a planimetric error larger than the pixel size is the generalization of the building
outlines. The low accuracy in height is obtained because of the erroneous plane parameters of the face planes due to the
face segmentation process. Moreover, the low planimetric accuracy results in low height accuracy having sloping faces,
Model refining/recovery can be done to reconstruct the building (hopefully) in order to improve the results.
5 | CONCLUSIONS AND RECOMMENDATIONS
In this research a procedure for building detection and roof extraction from a DSM obtained by laser scanning is
investigated. The procedure is meant to work for all terrain and many building types. Empirical testing, however, was
only done in two areas of low terrain variability. Multi level roofs are not considered in the procedure, which could be
modified to consider them. The procedure shows promising results for the building detection part and for some roof
faces in the extraction part. We have only considered planar faces, therefore, curved faces can not be detected as one
face. Prior knowledge about buildings and terrain type is required for the developed building detection and extraction.
More experiments are recommended, including test areas of high terrain variability. More calibration studies on the
threshold values should be done. In the suggested procedure, building reconstruction is dependent on the results of the
building detection. Moreover, building pixels are not allowed to be reclassified into non-building pixels and vice versa.
Therefore, vegetation which is adjacent to a building will be detected and constructed as a part of the building. Analysis
of the subsegments inside a large segment (or object), in the same way as the classification of objects into buildings or
vegetation, has to be done in order to single out the vegetation and reconstruct the building using only the roof faces.
Some adaptations of the segmentation of the face segments should be done in order to reduce the effect the noise in tht
segmentation process. More work in primitive matching and extraction has to be done as well. More constraints can be
added to the primitive extraction such as the face area and the intersection line between the adjacent faces. In the used
procedure the laser data are resampled into a regular grid, which can lead to geometric errors. It has to be investigated il
the procedure could be modified to directly use the original irregularly distributed data. Once the procedure is refined it
would also be of interest to study its performance on DMSs originating from other data sources.
622 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
ACKN
We wi
Works
REFE
Hug, C
Made (
Hug, C
Vol. 32
Kilian,
Archivi
Maas,
Worksl
Tempfl
Weidne
Proceec
Weidne
Models