ISPRS Commission III, Vol.34, Part 3A »Photogrammetric Computer Vision“, Graz, 2002
By tracking the paired image and object points that have
contributed to the peak in the accumulator array for the final
iteration, the correspondence problem is solved. The resulting
matches between the object and image space features are used
in a simultaneous least-squares adjustment to solve for the EOP.
It has to be mentioned that the obtained correspondences are for
low-level objects (i.e. between points). The following section
explains the consistency check we implemented to identify the
high level correspondence (i.e. between linear features) in
addition to highlighting discrepancies (changes) between object
and image space features.
3.2 Feature to Feature Correspondence and Change
Detection
So far, we established the following:
e EOP of the image under consideration, and
e Point-to-point correspondences between object and image
space linear features.
Now, we will proceed by performing a consistency check
between these features using the feature labels. The consistency
check has four steps:
Step1: Feature to feature correspondence
We check the label of the features containing the matched
object and image space points. Considering the frequency of the
matched labels, one can establish the correspondence between
the image and object space features.
Step 2: Object to image space projection of non-matched
object points
Using the estimated EOP and the ground coordinates of non-
matched object points, one can compute the corresponding
image coordinates. The standard deviation of the computed
image coordinates can be estimated using error propagation.
Step 3: Distance computation
The closest distance, as well as the associated standard
deviation, between the projected image points in step 2 and the
closest points along the corresponding image space features is
computed. One should note that the image to object feature
correspondence is already established in step 1.
Step 4: Blunder and change detection
If the distance is greater than a predefined threshold (e.g. three
times the associated standard deviation), we label these points
as either blunders or changes between object and image space
features. Single occurrences of non-matching points are
identified as blunders while successive occurrences of the non-
matching points are labelled as change (discrepancies).
Figure 2 is a schematic drawing for illustrating the concept of
the consistency check. In this figure, points i; to ij, are the
projected data points along a linear feature from the object
space into the image space, while points /, to /,7 are image data
points along the corresponding linear feature in the image
space. Consider points ij, ig, ij, iy and iy to be correctly
matched with points jj, ji, J13» Jis And j17, respectively; while
points i, is, is, is and io do not have matching entities in the
image space. Instead, their closest points in the second data set
along the corresponding linear feature are points Js, js, j7, jo and
Jie, respectively. In order to distinguish between the consistent
changes and blunders, non-matching points along the linear
feature are segmented and labelled. From this analysis, the pair
(io, ig) Will be considered as one label and the pairs (is, j3) to
(is, jo) will be considered as another label. The former label will
be considered as blunder because it has only one change pair,
while the latter will be highlighted as a consistent change. For
consistent changes, the longitudinal distance along the linear
feature as well as the average lateral distance will be computed
as the change attributes, Figure 2.
Consistent Change
Blunder
js Js
Average
Dlaterak----
distance
VER E.
B
S
—
=
Longitudinal distance
À
A
Figure 2: Consistency check between the object and image data
points. The rectangular points with labels 7; to i; are
the projected object space points along a linear feature
into the image space, while the crosses with labels 7;
to ji; are the points along the corresponding linear
feature in the image space.
4. EXPERIMENTS/RESULTS
Experiments have been conducted using real data. To carry out
the outlined methodology in the previous section, one should
have:
e A sequence of 3-D points along the ground control
features.
e A sequence of 2-D points along the image features.
e The interior orientation parameters (IOP) of the
camera.
Once again, the suggested algorithm does not require full
correspondence between the object and image space features.
The main requirement is having enough common features
between the two data sets. The input data and the results are
presented in the following paragraphs.
In the area covered by the aerial image, there exist a number of
major and secondary roads. The object space roads, represented
as a sequence of 3-D points, were extracted from a
photogrammetric stereo model containing the image under
consideration. Two data sets in the object space with different
number of roads had been digitised. A 2-D view of the 3-D
road network can be seen in Figures 4-a and 4-b. To complete
the data set, a 2-D point sequence along the image road network
must be extracted. In a digital environment, the extraction
process can be established by applying a dedicated operator
(e.g. Canny or any other operator for road network extraction).
In this work however, 2-D image features have been manually
digitised, Figure 3-c. Another data set in the image space had
been obtained by introducing digitisation errors (Figure 3-d) to
check the ability of the suggested system to detect those
changes. By combining different data sets from the object and
image space, we conducted four experiments, Table 2.
From Table 2, one can see that image space has much more data
available than the object space. After carrying out the
experiments, matched points were used to estimate the EOP in a
simultaneous least-squares adjustment. The estimated EOP are
listed in Table 3, together with the their initial (approximate)
values. These values can be obtained from navigation data.
However, very rough knowledge about the initial values of EOP
A- 153