MultiView can share a single screen or OverView may be
disabled.
During measurement, 3D coordinates are obtained through
triangulations. Currently, the operator specifies a point in the
first image, the points in other images needed for the
triangulation can be obtained either through manual
interactions or through the automated image matching. Each
time the first point is chosen, the corresponding epipolar lines
in other displayed images are shown to give the operator a
reference. Up to six points can be used to perform the
triangulation using the least squares for an optimal solution.
Tests have shown that the positional accuracy of object points
is up to 30 cm within a 50m object range.
For the purpose of automation, image matching techniques are
used for measuring conjugate image points. To this end, two
area-based single-point matching methods have been
implemented: cross correlation and least squares matching. In
the cross correlation process, a small window is used to find
the maximum correlation coefficients along the corresponding
epipolar line(s). Therefore, two-dimensional searching is
replaced by one-dimensional operations in the images. This
significantly reduces the amount of computational time. Since
geo-referencing errors and other errors may exist, the search
is performed in three adjacent lines centered on each epipolar
line. This feature improves the matching reliability and
accuracy. In most situations, the object distance is within a
certain range. The searching range is, therefore, reduced in
order to minimize computation time. Furthermore, the pair of
target and search windows are normalized to have the same
mean gray level. This technique does not require precise
initial approximation. The result of the cross correlation
algorithm is used as the initial approximation to perform the
least squares matching (i.e. refinement). Following this
procedure, a high precision sub-pixel accuracy can be
achieved. The least squares matching is very appropriate for
performing multi-image (i. e. more than two) matching.
In close-range photogrammetry, due to the consideration on
the wide coverage of objects and high accuracy in 3D
determination, the cameras have considerably large
differences in orientation. This in turn results in large
difference of geometry in stereo images because of the short
distance between cameras and objects. This effect is very
troublesome to area-based matching techniques. As such,
feature based matching has to be employed in this case.
Image intensity and its changes (edges) are dependable factors
for algorithms to determine 3D positions. In the current
approach, area-based and feature-based matching are
combined to take the full advantage of epipolar geometric
constraints and multiple images.
Although automation tools based on image matching are not
dependable tools for all objects for all situation at this point,
they are reasonably reliable in some situations such as the
extraction of street lines.
4. OPTIMAL ACQUISITION OF 3D OBJECT
COORDINATES
3D object coordinates can be uniquely intersected by two
georeferenced digital images that overlap the object. The
precision of the coordinates depends on not only the quality of
georeferenced digital images but also the space geometry
formed by the target point and the two camera exposure
stations. According to an error analysis, if only the
measurement errors of two image coordinates are taken into
consideration, the optimal object coordinates will be obtained
by using two image stations i and j respectively, with the
following criterion (Li et al, 1995):
RMS* = (D*; +D*)/ sin'( C)
+ H/ (sin^(A; ) *- sin? (B;)) =minimun,
Where, D; is the distance between camera station i and the
target P, D; is the distance between camera station j and the
target P, Cj is the space angle at the intersection of lines
iP and P Hj is the distance between point P and baseline
Us A; is the space angle at the intersection of line iP and ij,
and A; is the space angle at the intersection of line jP and ij :
Point matching techniques are used to automatically search all
images that cover the desired target from a sequence of
images. If approximate coordinates X, Y and Z of the target are
known, the 3D coordinates can be projected onto all selected
images. A sequential matching technique will find out precise
position of the target in each image within constrained
searching areas. Otherwise, the target will be identified as
invisible in the image.
In order to obtain approximate 3D object coordinates X, Y and
Z, the operator points at the target in two images. The
approximate coordinates X, Y and Z can be calculated. The
system will then automatically search all visible images and
records. Finally, according to the optimal criterion, the two
optimal images (i and j) will be determined. The image
coordinates (x; yj, and (x, yj are modified automatically by
point matching techniques and are used to intersect the optimal
3D object coordinates X, Y and Z. The coordinates are further
employed to form GIS entities and are recorded in the
database.
5. AUTOMATION OF INFORMATION EXTRACTION
The increasing need for the automation of information
extraction from mobile mapping systems poses a significant
challenge in the current research and commercialization of
mobile mapping technologies, due to the large volume of
captured images needing to be processed. As mentioned
above, the obtained terrestrial images exhibit large scale
variations and geometrical discrepancies across images, which
have to be accounted for in the automatic process of
information extraction. This increases the complexity of the
task substantially. In the current system development, an
object-directed and user-guided strategy has been proposed
and applied to the system design of the automated or semi-
automated functions of information extraction. The automatic
extraction algorithms are designed to be tailored to the specific
234
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
objec
and
extra:
scene
an O
algor
The 1
envir
the S!
objec
the si
is des
meas
5.1R
Road
very
inforr
paran
globa
cente:
imple
marki
recon
image
cente:
3D sl
of ro
deteri
order
object
mode
snake
defor
defor
such
action
3D pc
exteri
desire
algori
cente1
sterec
smool
of the
assum
combi
mode]
in wt
model
3D rc
that t
where
S2R
It is t
bound
Howe
smoot
match
point
