The elimination of outlier points is carried out iteratively, based
on data snooping, which eventually provides reliable plane fits
for the match cloud. Now let the corresponding point in the
reference cloud define a local coordinate system in which the
plane fit is computed. Then the resulting distance d is the initial
point-to-plane distance for that location. It is used to identify
outliers in the reference cloud and, in combination, provide
reliable point/plane pairs along with their stochastics to the sub-
sequent point cloud matching.
2.2.2 Offset Derivation: Point/plane pairs from correspon-
ding info clouds are used to derive their overall three-dimen-
sional offset. This is computed in a least squares adjustment by
introducing the unknown parameters (Xo; fer, Yo ffset» Zof fset)
into each individual plane's Hessian Normal Form, aiming to
match the individual point-to-plane distances derived above.
Based on (1), the observation equations read as follows:
d= nx Xorfset + NyYoffset + nzZoffset (3)
The selection of the distance d as observable allows for weight-
ing the observation equations by the quality of the plane fit in
such a way that weaker planes with larger standard deviations
are rewarded less influence for the offset computation than
more robust plane fits. Practical tests confirm that this weight-
ing is critical for a meaningful offset computation.
Even though equation (3) is linear and, accordingly, this very
adjustment could be solved without iterating, the surface ap-
proximation by local planes has to be refined. Therefore, the
overall solution is computed iteratively, applying the resulting
offset to improve local plane locations as described above. This
iteration is carried out until the offset increment becomes in-
significant, i.e. well below the GSD.
2.3 Combined Geometric/Radiometric Approach
For point clouds from nadir-looking sensors, geometric mat-
ching alone does constrain the vertical offset component in
virtually all cases. However, horizontal correlation becomes
weak if the data lack significant height gradients in different
directions, which is obviously the case in flat terrain but also on
oriented slopes such as building roofs, especially considering
the small patch size used for the purpose of Shear Analysis.
This issue has been addressed by several algorithms that utilize
any available information on intensity to either assist the geo-
metric matching (Weik, 1997) or extend it towards a combined
solution (Maas, 2002; Akca, 2007). Intensity gradients tend to
occur more frequently than height gradients; they can provide or
complement the required information for the derivation of
planimetric offsets. Therefore, it suggests itself to use this
additional information, which is available from the info clouds,
and integrate it with the geometric point cloud matching into a
combined adjustment.
2.3.1 Adjustment Extension to Intensity: In analogy to the
height Z, the intensity or, respectively, digital number DN is an
attribute of a point’s planimetric location (X,Y). Accordingly,
the functional models for the integration of geometry and
radiometry can be written the same way, the first using height
gradients (from local X,Y,Z planes), the second intensities (local
X,Y,DN planes); the underlying idea is the well-known least
squares (image) matching. The above-described local plane fits,
applied to both heights and intensities, provide a pair of obser-
vation equations, (3) and (4), for each reference point for the
combined geometric/radiometric least squares adjustment.
104
2.3.2 Histogram Adaptation: The unavoidable difference in
viewing geometry as well as potential illumination or even
temporal differences can result in radiometric differences in the
overlapping ADS image data and, accordingly, in the corres-
ponding info clouds. Those differences are compensated as part
of the combined adjustment by applying brightness and contrast
correction terms, b and c. Similar to (3), with the heights re-
placed by corrected intensities, the radiometric observation
equations read:
don = nx pnXoffset + Ny,onYoffset + Nzpn(cDN +b) (4)
This adjustment features a total of five unknown parameters: the
three-dimensional offset as well as the required contrast and
brightness adaptation. It is obvious that the additional equations
(4) can only provide immediate constraints on the Xp rrger and
Yoffset Components. The Zorrset is determined by equation (3).
However, the planimetric offset is part of both (3) and (4); the
latter of which therefore has some (indirect) impact on the re-
sulting height offset, too.
2.3.3 Weighting between Geometry and Intensity is based
upon their group variances: If balanced properly, the relation
between these variance components before and after the ad-
justment should be identical for height and intensity observation
groups. The ratio is used to alter the weights for the intensity
equations. This adaption is iterative and, theoretically, needs to
be carried out for each adjustment computation. However, it
seems feasible to improve weighting along with the overall
iteration, considering that the main goal is to level largely dif-
ferent orders of magnitude that occur depending on terrain and
texture. Test runs have shown weight factors between 1 and
greater than 10,000, the latter in very flat regions where the
planimetric offset components are almost exclusively deter-
mined by intensity.
2.4 Offset Verification
In order to provide reliable offsets for Shear Analysis input, the
results of the combined geometric/radiometric matching are
verified as part of the computation. First of all, the roles of
corresponding info clouds are switched between reference and
match cloud, thus providing two independent offset computa-
tions that must agree within tight thresholds. The average offset
becomes the final result, given that further indicators are mea-
ningful. Those include a minimum number of point/plane pairs
in a patch, reasonable radiometric corrections as well as the
maximum number of iterations. Respective limits can depend
on a variety of parameters, predominantly terrain and imaging
configuration. Nevertheless, using rather tight general settings —
a maximum in the order of 5 to 7 iterations, based on at least
25% point/plane pairs in relation to the number of input image
points in a patch — might eliminate some correct results; but as
long as offset are attempted to be computed in a fairly dense
pattern, there will be sufficient input for the Shear Analysis.
Most important is the elimination of false positives.
3. EVALUATION ON DIFFERENT DATA SETS
The combined geometric/radiometric point cloud matching ap-
proach was initially verified with synthetic data — heights and
intensities assigned to predefined locations — and artificially
introduced offsets. Practical validation and Shear Analysis
exploration was carried out for a number of ADS blocks with
different characteristics. Eventually aiming for the replacement
of the manual QC, a crucial part of it was the verification of
auto!
men!
bine
3.1
The
by ^
verif
very
of 5
pres‘
3.1.1
Geo
fore:
aisle
imag
Nat
fore:
3.1.
City
ding
area:
This
com
poss
3.1.
Agrı
bloc
sele«
dese
Da
Ge
Ne
Tab
mat
3.2
One
sets
are
com
COIT
relia
meti
the
stan
man
sho
grot
in ti
base
this
thre
triar
pect
for 1