5.4 Matching strategy
A strategy is followed in order to get first some reliable and
accurate match results in 2 pyramid levels. These can provide
better surface approximation, reduction of search space, and
support for weaker match features. Raw match features are
matched as dense as possible, thus making the blunder detec-
tion and correction easier, avoiding mismatches being propa-
gated to the lower levels. Due to the fact that rotations and
shifts existed in the images, plus the bases were not large
enough, matching used object space information to get some
first approximate values. The approximate 3D information
was used also by the forward intersection performed at the
end of the algorithm in the least squares adjustment.
For the extracted edgels approximate heights have been inter-
polated from the coarse surface generated from the existing
GCP’s. The interpolation of heights is done only for the tem-
plate image and by back-projecting on to the search image a
first approximation of the point is found.
Furthermore, a multi-patch approach is adopted where up to 3
passes of matching are performed with different parameters
(such as search range and patch dimensions). Larger patches
are less sensitive to noise, occlusions, multiple solutions etc.
while smaller ones are more accurate and better preserve
height discontinuities. The algorithm employs area-based
correlation in the first stage, and least squares matching at the
end for higher accuracy. Since least squares matching per-
forms slower than cross correlation, it is used in a second
stage as the last step in the quality control, where most of the
mismatches have been excluded in previous steps in the qual-
ity control. It can estimate the position with sub-pixel accu-
racy or reveal suspicious points and flag them and thus in-
crease robustness of the algorithm. Additionally, epipolar
constraints can be used with a small weighting which enables
the search area to be extended 1-2 pixels perpendicular to the
epipolar line, since the estimation of the points that lie near
the borders of the image can be less accurate.
During run time, quality measures are calculated, which play
important role in the elimination of blunders and false
matches. The cross correlation coefficient, the 2™ best simi-
larity score and its distance to 1% one, the size of search win-
dow, the change of similarity measure between the 3 patch
sizes, the change of position between patch sizes, the angle of
dominant edge direction with the epipolar line, the residuals
from 3D forward intersection, the change of final point posi-
tion from the starting position, calculated standard deviations
for x and y pixel coordinates are the quality criteria calcu-
lated. The quality criteria are combined according to the pos-
sible occurring error. E.g. in case of an occlusion, the cross-
correlation coefficient would be small and the similarity
measure would be generally decreasing from the largest to the
smaller mask. Additionally the consistency of height in the
local neighbourhood is checked. The matched points are as-
signed to error groups (e.g. occlusion, multiple solution, etc.)
depending on their quality measures. Thresholds for each
group of errors are computed from a statistical analysis of the
quality measures extracted in each given pyramid level.
5.5 Results
Constrained and unconstrained method gave similar matching
results. The constrained method was significantly faster than
the unconstrained, because of the reduction of search space.
Matching results were subsequently filtered to remove re-
maining suspect points, using a median filtering. In average
18.000 points per model were successfully matched. The
matched 3D edges were finally converted to 3D gridded
points through interpolation, with 0.02 m grid spacing. The
gridded DSM's of the normal images with best overall accu-
racy(see below) were merged through averaging in the over-
lap regions. The quantitative analysis of the DSM was done
using reference 3D points (16 GCP's), collected with topog-
raphic methods with an accuracy of 2 cm. The points were in-
terpolated from the raw matched data and compared. The dif-
ferences in height were calculated and as estimation of the
accuracy their standard was used (Table 1).
Image pairs Average std.| RMS Max abs.
dev. (in m) error
Strip 1: pair 1-2 0.04 0.05 0.12
Strip 1: pair 2-3 0.06 0.08 0.18
Strip 2: pair 1-2 0.03 0.04 0.09
Strip 2: pair 2-3 0.05 0.06 0.15
Strip 2: pair 3-4 0.09 0.1 0.18
Table 1. Accuracy of DSM’s extracted from image pairs
using constrained approach (Normal images).
Since the different DSM’s were overlapping, the ones with
the best accuracy, resulting from the comparison with the
manually extracted DSM, were selected for merging (Fig. 8).
The large difference in accuracy for pair 5-4 depends on the
orientation data (small base). The orientation was signifi-
cantly improved with the bundle adjustment but small errors
remained. Therefore from strip 1 the pair 1-2 was selected and
from strip 2 the pairs 1-2 and 2-3. The average standard de-
viation in these 3 models was 0.04.
Figure 8. Part of the DSM. Contours with an interval of 0.01
m overlaid on the image.
—444—
Fi
The mat
from the
0.06 m.
The acc
therefore
5.6 Cor
Virtual i
racy deg
sufficien
accuracy
there are
proximat
ages that
ing the n
the virtu
can be st
extractio:
