ceeds in a goal-directed manner. A control unit, con-
taining a priority ordered queue of processing elements,
is added to the blackboard system. Further details con-
cerning the dataflow of the BPl-System are described
in previous papers [Lütjen, 1986; Stilla, 1995]
4 OBJECT MODEL
In order to explain the 3D-reconstruction we chose a
simple model for the recognition of a house. In many
aerial images containing houses only their roofs can be
recognized. Thus the houses are actually described by
their roofs. This paper only takes houses with simple
gabled roofs (ROOF) into account. It is assumed that
significant parts of a roof can be described as rectan-
gles in the scene and parallelograms in the image. This
applies for many houses in aerial images.
À simple rewriting system for a roof is determined by
Groor = ({R},{R, PF, P,U, A}, {L}, {P1, Pa, P3, Pa, Ps })-
According to Groor and starting with the primitive
objects L, the partial objects A, U, P, F, R are com-
posed using the productions P;-P; (see Tab. 2), with
object R representing the target object ROOF.
B: Objects X, Y Qm opectZ
P, : LAL OD hms aid
B: AAA 5. = U
Py: UAL Q up
Pa: (UVP)A (UVP) © — F
Ps : FATF (5 2 R
(D angle-shaped, 45° < a < 135°
@ u-shaped
@ parallelogram-shaped
@ corresponding in 3D
© building an egde in 3D, 90° < «y « 170?
Tab. 2: Table of productions
A production net for the object ROOF is depicted in
Fig. 3. The analysis distinguishes between a 2D- and
3D-analysis. First, the 2D-analysis is carried out inde-
pendently in different images.
4.1 2D-Analysis
Starting with the object primitives LINE, the objects
ANGLE, U.STRUCTURE and PARALLELOGRAM can
be built up by applying the productions (P,-P3). Ob-
jects ANGLE are built up by connections in pairs of ob-
jects LINE (Pi). The constellation of objects LINE
enclosing an angle a can be L-shaped or T-shaped
(Tab. 2 (D). If two objects ANGLE form a struc-
ture like an open parallelogram they are combined to
an object U. STRUCTURE (P5). An object PARAL-
LELOGRAM can be built up if objects U. STRUCTURE
and LINE are compatible (P3). In the subsequent 3D-
analysis 2D-objects form the primitives.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Fig. 3: Production net
4.2 3D-Analysis
The 3D-analysis attempts to find in two different im-
ages pairs of 2D-objects (U_STRUCTURE or PARAL-
LELOGRAM) which are projections of the same 3D sur-
face. This is done by selecting pairs and examining rays
originating at the centre of the projection and by pass-
ing through the vertices of the 2D-objects. The rays
result from inverse mapping of the projection. On ideal
conditions rays running through corresponding object
vertices of two images will intersect in the 3D-space.
Due to image noise, processing errors, and inaccurate
camera parameters the rays generally do not intersect.
Hence, the minimal distance between the rays is calcu-
lated. The 2D-objects will be called not correspond-
ing if this distance between the rays of pairs of ver-
tices is greater than a given threshold. Additionally, a
model-based plausibility check ist carried out (e.g. re-
garding the position: object points must not be under
the earth’s surface).
If 2D-objects of different images correspond, the ob-
ject ROOF_AREA is generated (P4). If objects
ROOF_AREA are oriented in a way that the surface
normals enclose an angle vy that lies within a certain
angle interval and if they are located in a way that the
vertices are neighbouring, then a target object ROOF is
generated (P5). The angle interval is assumed to be
known (Tab. 2 ©). After the 3D-analysis is complete
the best objects ROOF are selected.
5 PREPROCESSING
In the preprocessing stage a symbolic description of the
scanned aerial image is created. The preprocessing is
carried out in four steps.
(1) At first the greylevel image (Fig. 4a) is transferred
ida Mid
T vo
CX RS
—