vhich
final
(6) is
le of
re 10
irther
THPCM Eb EU ei
ion
ige
ction
ions,
mage
Kubik Kurt
Based on an analysis of the histogram of the orientations of edges, areas whose edges are not considered to be those of houses
can be eliminated. The extracted house areas are again illustrated in Figure 14. Figure 15 shows the area of dark roof
extracted by step 2 in Figure 3, while Figure 16 shows the final determination of the houses. Figure 17 is the image in which
the final extracted house areas are overlaid on the original image. This figure shows that the houses are well delineated.
Using the compound information from the analysis of image matching and 2D image segmentation, some digital elevation
points which were initially located on the tops of houses and trees have been interpolated onto the ground. Figures 18 and
20 illustrate the DEM and 3D perspective view derived directly from normal stereo image matching. A more accurate DEM
and 3D perspective view produced by the method in this paper are shown in Figures 19 and 21.
Figure 18 DEM from matching Figure 19 DEM from the Figure 20 3D perspective view Figure 21 3D perspective view
combined method from matching derived by the combined method
6 CONCLUSION
The method described in this paper combines image matching and image analysis methods, which enables the location of most
of the house and tree areas in the test images. The image segmentation and classification methods overcome the weakness of
co-occurrence matrices that is it does not consider the shapes of gray level primitives. These extracted house and tree areas
are important information for 3D terrain reconstruction and ensure that points are only measured on the natural terrain. The
method leads to more accurate determination of elevations from overlapping digital aerial images than the DSM determined
only by image matching, since it avoids errors caused by man-made or natural surface features. The method can also locate
dark roofs. The disadvantage is its inability to exactly locate the boundary of dark roofs in cases when the roof of a house is
not of regular shape. Since the classification result of co-occurrence matrices are dependent on chosen training sample and
the size of the processed image block, further research is needed to find a more reliable method for image classification. The
method will also be tested on other scales and different images.
REFERENCES
Baltsavias E., Mason S. & Stallmann D. (1995). Use of DTMs/DSMs and Orthoimages to Support Building Extraction. Automatic
Extraction of Man-made Objects from Aerial and Space Images. Birkhauser Verlag, Basel, pp 199-210.
Burns, I. and Smith, G. (1996) MeasTex Version 1.0: A Framework for Measuring the Perfomance of Texture Classification Algorithms.
University of Queensland, http://www/cssip.elec.uq.edu.au/-guy/meastex/meastex.html.
Conners, R. W., Harlow, C. A. (1980). A Theoretical Comparison of Texture Algorithms, IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. PAMI-2, No. 3, pp.204-222.
Gonzalez R. C. & Woods R. E. (1992). Digital Image Processing. Addison-Wesley, U. S. A.
Haralick, R. M., Shanmugam, K. S., and Dinstein, I. (1973) Textural Features for Image Classification. IEEE International Conference
on Systems, Man, and Cybernetics, Vol. SMC-3, No.6, pp610-621.
Henricsson O., Bignone F., Willuhn W., Ade F., & Kuebler O. (1996). Project AMOBE: Strategies, Current Status and Future Work.
International Archives of Photogrammetry and Remote Sensing, 31(3):321-330.
KB Vision System Task Reference Manual (1996). Amerinex Applied Imaging, Inc.
Tónjes R. (1996). Knowledge Based Modelling of Landscapes. International Archives of Photogrammetry and Remote Sensing,
31(3):868-873.
Wilson P. A. (1997). Rule-Based Classification of Water in Landsat MSS Images Using the Variance Filter. Photogrammetric
Engineering & Remote Sensing, 63(5):485-491.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 527
