of feature points include geometric and radiometric properties.
The geometric properties of a feature point include its location
in the image, the number of edges which intersect at the point
and their spatial directions, angles between the edges and the
locations of points which connect with the current point. The
radiometric properties include the average intensity value
between two edges.
The feature relations include topological relations between
edges and between points and edges. Intersecting edges are
arranged clockwise based on their spatial orientation. Their
relation can be represented by left and right while the relation
between a point and an edge can be described by connect-to.
3.2 Matching of Roof Points
After feature points are extracted from both vertical and oblique
images and their properties and relations are determined, they
are matched to generate 3D feature points. Due to the nature of
oblique images, the background around a feature point is quite
different in different images. Therefore a feature-based image
matching method instead of an area-based image matching is
used to find the corresponding feature points on the overlapping
images. Each feature point is represented as the intersection of
two or more than two edges. Corresponding feature point in the
overlapping images should be the intersection of corresponding
edges if they are visible in the images and are extracted. Thus,
the criteria used in the image matching are the spatial directions
of edges associated with the feature point and the gray value
between two edges. The point with minimum difference of edge
direction and gray value is treated as the corresponding point.
Since image matching is done among the vertical and four
oblique images with different view directions, usually a feature
point can be matched on more than three images. Once a point
is matched successfully, its coordinates in ground coordinate
system are computed by the least square adjustment method.
After the ground coordinates are calculated, the residuals of the
point can be computed on all the images and the points with
large residuals are treated as unreliable point and removed, and
the matched points on the remaining images are used to
. compute the point's coordinates again. Finally, the point’s
elevation is checked against the terrain's elevation around the
point and the point is kept if the elevation difference is larger
than the given threshold. One example of matched roof points
is shown in Figure 3.
(a) (b)
Figure 3. Matched roof points (a) matched roof points on nadir
image (b) matched roof points on oblique image with North
view
333 Grouping of Roof Points
In grouping, the generated 3D feature points belonging to the
same roof facets are grouped together to create roof facets. The
grouping starts with any feature point and one edge associated
with the point as point a, and edge /, shown in Figure 4. It
proceeds to the next point using the relation connect-to and
finds the next edge by relation right (a; and /). This process is
repeated until it reaches at the starting point. Once all feature
points of the facet are found, a plane is fitted to the points by
the least squares adjustment. If the facet has more than three
roof points, the elevation of every point will be checked against
the fitted plane. The elevation of a point will be corrected if the
difference between the fitted surface and the point is larger than
the given threshold. In this way, all roof facets are created.
(a adi e 03
in
f^
az
aj lj
Figure 4. Grouping of feature points
4. TESTS RESULTS
To test the developed approach, a number of images have been
tested. The test images are color image data captured by
Pictometry's imaging system. Each group of test images have
one nadir image and four oblique images with different view
directions, i.e. north, south, east and west. The nadir images
have a GSD of 10 cm (4") while the GSD of oblique images is
between 10 cm and 13 cm within the image. The areas covered
by the test images are typical residential area and most
buildings in the areas are two story residential buildings. Some
of the test results are listed in Table 1.
No. of Type of Number of Number of
building building roof facets extracted roof
facets
1 Hip 4 4
2 Complex 10 10
3 Complex 11 11
4 Complex 15 13
5 Complex 30 25
Tablel. Results of building extraction
The first test building in the table is a residential building with
hip roof as shown in Figure 5. This is a typical simple
residential house. The reconstruction of this type of building is
relatively easy since they have very simple roof structure. The
only issue with this building is that there is a deck on one side
of the building and there are some decorations around the
building. This needs to be dealt with carefully. The result shows
that four roof facets were extracted successfully.
The second test building is a two story residential building with
combined hip and gable roof structure. The roof has ten roof
facets and each has a reasonable size. Most facets have simple
rej