ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision*, Graz, 2002
can extract 3D surface patches from laser points. Such
patches may be planar or higher order surfaces, depend-
ing on the scene. In built-up areas, usually many planar
surface patches exist, corresponding to man-made objects.
After surface patches have been extracted, a grouping pro-
cess establishes spatial relationships. This is followed by
forming hypotheses as to which patches may belong to the
same object. Adjacent patches are then intersected if their
surface normals are different enough to guarantee a geo-
metrical meaningful solution. In fact, if the adjacency hy-
pothesis is correct then the intersection is a 3D boundary of
an object. Lee (20022) treats the steps of extracting and
grouping patches as a perceptual organization problem.
Table 3: Multi-stage feature extraction from LIDAR and
aerial images.
LIDAR aerial imagery
raw data 3D point cloud
pixels
ng ng
Let us now examine the features that can be extracted from
images. The first extraction level comprises edges. They
correspond to rapid changes in grey levels in the direction
across the edges. Most of the time, such changes are the
result of sudden changes in the reflection properties of the
surface. Examples include shadows and markings. More
importantly, boundaries of objects also cause edges in the
images because the two faces of a boundary have different
reflection properties too. Hence we argue that some of the
2D edges obtained from aerial imagery correspond to 3D
edges obtained from laser points. That is, edges are poten-
tially sensor invariant features that are useful for solving the
registration problem. Note that the 2D edges in one image
can be matched with conjugate edges in other images. It
is then possible to obtain 3D features in model space by
performing a relative orientation with linear features.
Are 3D surface patches also sensor invariant? They cer-
tainly correspond to some physical entitities in object space,
for example a roof plane, face of a building, or a parking lot.
Surface patches are first order features that can be extracted
from laser point clouds relatively easily. However, it is much
more difficult to determine them from images. One way to
determine planar surfaces from images is to test if spatially
related 3D edges are lying in one plane. Surface patches
then can also be considered sensor invariant features.
3.3 Referencing aerial images to LIDAR data
From the discussion in the previous section we conclude
that 2D edges in images, 3D edges in models, and 3D sur-
face patches are desirable features for referencing aerial
images with LIDAR. Table 4 lists three combinations of sen-
sor invariant features that can be used to solve the fusion
problem. As pointed out earlier, we consider this first step as
the problem of determining the exterior orientation of aerial
imagery. Extracted features from LIDAR data serve as con-
trol information.
Table 4: Sensor invariant features for fusing aerial imagery
with LIDAR.
LIDAR aerial imagery Method
3D edges 2D edges SPR, AT
3D edges 3D edges ABSOR, AT
3D patches 3D patches ABSOR, AT
Orientation based on 2D image edges and 3D LIDAR
edges The first entry in Table 4 pairs 2D edges, extracted
in individual images, with 3D edges established from LIDAR
points. This is the classical problem of block adjustment,
except that our fusion problem deals with linear features
rather than points. Another distinct difference is the number
of control features. In urban areas we can expect many
control lines that have been determined from LIDAR data. It
is quite conceivable to orient every image individually by
the process of single photo resecting (SPR), that is, the
problem can be solved without tie features. This offers the
advantage that no image matching is necessary—a most
desirable situation in view of automating the fusion process.
Several researchers in photogrammetry and computer vi-
sion have proposed the use of linear features in form of
straight lines for pose estimation. Most solutions are based
on the coplanarity model. Here, every point measured on a
straight line in image space gives rise to a condition equa-
tion in that the point is forced to lie on the plane defined by
the perspective center and the control line in obiect space,
see, e.g. Habib et al. (2000). The solutions mainly differ in
how 3D straight lines are represented.
Although straight lines are likely to be the dominant linear
features in our fusion problem, it is desirable to general-
ize the approach and include free-form curves. Zalmanson
(2000) presents a solution to this problem for frame cam-
eras. In contrast to the coplanarity model, the author em-
ploys a modified collinearity model that is based on a para-
metric representation of analytical curves. Thus, straight
lines and higher-order curves are treated within the same
representational framework.
With the recent emergence of digital line cameras it is nec-
essary to solve the pose estimation problem for dynamic
sensors. The traditional approach is a combination of di-
rect orientation and interpolation of orientation parameters
for every line. This does not solve our fusion problem be-
cause no correspondence between extracted features from
LIDAR and imagery is used—hence no explicit quality con-
trol of the sensor alignment is possible. In Lee, Y. (2002b)
the author presents a solution of estimating the pose for line
cameras by using linear features. In this unique approach
every sensor line is oriented individually, without the need
for navigation data (GPS/INS).
Orientation based on 3D model edges and 3D LIDAR
edges We add fusion with 3D model edges more for the
purpose of completeness than practical significance. In con-
trast to the previous method, edges must be matched be-
tween images to obtain 3D model edges. In general, image
matching, especially in urban areas, is considered difficult.
We should bear in mind, however, that in our fusion prob-
A - 313