ATION
D-scene, which
entation and 3D-
; projecting the
convenient plane
y a stereoscopic
Xo point for 3D
G)
! X and Y axes
surface using the
lues. To assign a
point (X,Y,Z) is
ions.
as are taken from
ing» the object to
Image
a
sible road area
2000) it has been
the calculation of
in appearance of
pourhoods of 3D-
e. As orthogonal
d by the surface
ce is appeared on
hood.
duced to detection
roperties on the
straight-line edges
| structures. Angle
inctions of object
form.
hardware generally
e convolutions and
| makes it quite
letection of corner
To solve the problem mentioned above the orthophoto
transformation in polar coordinate system (radial orthophoto) is
introduced. The origin of polar coordinate system for left and
right orthophoto is the appropriate camera focal plane. Point
(o, R) of radial orthophoto is calculated as (4).
Re fi-X V+ (1)
X-X
S
(4)
@ = arctan
where — Xs, Ys - denote the projection center of the camera in
road-based exterior coordinate system;
R - denotes distance from the camera focal plane to
point (X, Y);
a- denotes the angle between optical axis of camera
and line connecting point with camera.
Pixel coordinates of radial orthophoto point are calculated as:
(X, Y)-(R*sin(a)*Xs, R*cos(a)* Ys) (5)
The invisible regions in radial orthophotos have a rectangular
shape with vertical edges (see Figure 4). This important
property allows to simplify the detection algorithm. Instead of
finding 2D-corner shaped structures on differential orthophoto
image the clusters of straight-line vertical edges have to be
detected. As a result the detection problem can be reduced to
1D-case by implementing such fast, hardware supported
operation as vertical projection of image intensity.
Figure 4. Radial orthophoto image of 3D-scene on Fig.1
4. OBSTACLE DETECTION
4.1 Obtaining obstacle features
The main distinctive advantage of radial orthophoto is that it
preserves vertical edges of 3D-object in the resulted orthophoto
images. Since the vertical edges of object have a significant
brightness variance in horizontal direction it allows to consider
them as the most informative object features.
To obtain edge points we introduce a "feature" image in such a
way that obstacle edge is coded accordingly to the sign of
brightness derivative in horizontal direction and brightness
strength of the edge (Figure 5). The brightness derivative is
calculated by using operator (6), that is the convolution of
orthophoto image with special mask and addition of constant
for negative value removal. Operator (6) gives reasonable
compromise between the sharp response to edge points and
sufficient averaging of the image for noise suppression. The
brightness strength of the edge is a statistical characteristic
based on analysis of edge points distribution.
-j -1 0 1 ]
w=] —1] 0 1-]
1 (6)
-i-L -1:0 1 1|1128
20
-] -1.0 1 1
zl -i' 011
(b)
Figure 5. (a) derivative of orthophoto image in horizontal
direction with use of operator (6); (b) resulted
"feature" image
4.2 Obstacle features detection
In this work we use correlation based approach for obstacle
features detection.
Let the vertical intensity projection of the image be defined as
(7):
V(«)=2 165») 7)
where V(x) — projection value;
X, y — image coordinates;
I(x,y) — intensity of the image at point (x,y).
We consider the first derivative of such projection for left and
right orthophoto images (Figure 6-a,b). The most significant
correlation peak between projections represents the angular
disparity of 3D-object (Figure 6-c). In other words it represents
the certain distance where the maximal amount of obstacle
features are found.
—127—