XXXIX-B3, 2012
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Next, the back-and-front judgment using a wall surface polygon
is performed, and the image used for the wall surface matching
on the front side is selected because of the larger size of that
area on the independent rectification image. Then, the ground
was matched using only the partial voxel-images of the
puilding’s front-side. The green frame in the figure shows the
result. Moreover, in the graph of the correlation value, the
extracted roof-top was approximately 90 m, as a clear peak, and
the ground was 10 m. As a result, the polygon of a rooftop and
the ground are mapped correctly on all the independent
rectifications.
53 Division matching for wall direction compensation
The search results for the wall surface division matching are
shown in Figure 9. The example of an incorrect extraction
produced in the initial edge matching is shown on the left side
of Figure 9. Thus, even when the correct roof footprint line
segment was not obtained, the right wall surface position could
be extracted by compensating for the wall direction. Although it
was a rare example, the right side of Figure 9 shows an example
of a building where the wall surface has become an arch. Here
with each element in a division voxel-image, it appears that the
right wall surface position can be extracted by applying a
curved surface form. Next, the results of the wall matching
compensation processing on the right wall's direction angle for
six buildings (each with four planes) is shown in Figure 10.
Number of Walls |. Number of Walls
9 ) 9
8 | à
1 1 7
5 | 8
5 | 5
4 | «4
3 i 3
1: f
% | 0 4. + Scio "d^ 2 deg]
Before Compensation After Compensation
Figurel0. Results of the Wall Direction Histogram
As a result, the root-mean-square error (RMSE) of the wall
deflection angle that was at approximately 1.7 degrees at the
maximum, in comparison with the manual plotting data,
decreases to approximately 0.5 degrees when using the wall
direction compensation processing. It proves that the wall is
stabilized and an exact wall direction angle can be estimated by
this technique. Moreover, to compensate for the wall direction
angle, using the kurtosis peak of the correlation coefficient
improves the results in the case of matching processing and
contributes to improvements in the stable processing.
The results of the target building using this wall matching
technique and a comparison with the building wall corner points
obtained by the manual plotting are shown in Table 3.
Initial Position 1.28 1.37 1.73
Initial Position 0.63 0.72 0.67
Non i 0.67 0.71 1.73
Non i 0.24 0.30 0.67
With 0.44 0.61 0.69
With i E 0.17 0.20 0.27
Table 3. Results of Wall Matching (vs. Manual Plotting)
”Initial Position” in Table 3 is the position accuracy of the roof
outline by edge matching that indicates the measurement error
is 1.37 m as a maximum value, and the RMSE is approximately
0.6—0.7 m. In contrast, "Non Compensation" in Table 3 after
wall matching is reduced to approximately a 0.2-0.3 m RMSE
value. In addition, "Non Compensation" is the result from wall
matching without correction for the wall direction angle. The
error value does not change for the wall direction angle; the
error value of the horizontal position of the line vertex tended to
decrease to the maximum value and the RMSE values.
However, after wall matching in "Non Compensation", the
RMSE value remains at 0.3 m. This reduced accuracy may be
caused because the surface orientation of the roof outline that is
used for wall matching does not match against the actual
building wall. For this reason, compensation processing is
performed by the division matching for the wall direction
compensation to estimate the correct direction angle of wall.
The RMSE value of the results obtained by the wall angle
compensation, shown "With Compensation" in Table 3, is less
than 25 cm in both X and Y directions.
6. CONCLUSION
In this study, we proposed and verified an independent
rectification method that further improves the multi-image
matching method. Here the problem was to extract the surface
structure of a building using multi-view images obtained by an
aerial survey. First, we generated the voxel-image by the IR
method. Next, horizontal plane matching was conducted to
extract the rooftop and ground surface. Further, the division
matching of the wall is performed to compensate for the wall
direction. Thus, we have developed a new method to extract the
exact position of the rooftop, ground surface, and the walls
around the initial line segment of the building's footprint. In
addition, we conducted experiments to evaluate the
performance of the proposed IR method. Multi-image matching
is performed using the voxel-images generated from the aerial
images as multi-view images. The existing method of matching
was complicated or made difficult by the occlusion of the
building and the heavy distortion of the wall on the image. In
contrast, the problem is addressed using the independently
rectified images and the proposed matching method by which
the object space can be searched in various directions. Future
research includes automatic extraction to obtain the line
segments of the building's footprint when there is an inclined
slanting roof, and automated structure recognition of complex
building shapes.
REFERENCES
Christian Beder, 2004, A unified framework for the automatic
matching of points and lines in multiple oriented images, Proc.
20th ISPRS Congress, Istanbul, Turkey.
Martin Schluter, 1998, Multi-image matching in Object Space
on the Basis of a General 3-D Surface Model Instead of
common 2.5-D Surface Models and its Application for Urban
Scenes. ISPRS Com. IV Symposium on GIS - between visions
and applications.
Jianging Zhang, J.Yong Zhang, Y. and Zuxun Zhang, Z., 2005,
Multi-image matching for generation of DSM and true ortho-
image, Proc. SPIE.
Oda,K., Doihara,T., Shibasaki,R., 2004, Stereo Plane Matching
Technique. Journal of the Japan Society of Photogrammetry
and Remote Sensing, Vol.43, No.3, pp.13-21.
