Graphics Reference
In-Depth Information
To avoid false correspondences or inaccurate disparities, the value of
d
c
computed
according to (
1.110
) is only accepted if the four following conditions are fulfilled.
1. The
i
-th point of both contour segments is on the same epipolar line, respectively,
for all values of
i
:
v
(w
(i)
v
(w
( j)
−
L
S
/
2
+
i)
−
L
S
/
2
+
i)
¯
1
=¯
2
for
i
=
1
,...,L
S
.
(1.111)
1
2
2. The average absolute slopes of both contour segments are above a threshold
θ
slope
:
v
(w
(i)
v
(w
(i)
L
S
−
L
S
/
2
+
i)
−
L
S
/
2
+
i
−
1
)
1
L
S
¯
1
−¯
1
1
1
>θ
slope
(1.112)
u
(w
(i)
u
(w
(i)
−
L
S
/
2
+
i)
−
L
S
/
2
+
i
−
1
)
i
=
2
¯
1
−¯
1
1
1
and analogously for the left image.
3. The absolute cross-correlation coefficient of the contour segments (Heisele,
1998
) exceeds a threshold
θ
corr
:
L
S
i
1
(
s
(i)
(
s
(i)
2
−
s
1
)
·
−
s
2
)
1
−|
λ
1
−
λ
2
|
=
1
L
S
i
=
>θ
corr
,
(1.113)
2
L
S
i
=
λ
1
+
λ
2
s
(i)
1
s
(i)
2
2
1
|
−
s
1
|
1
|
−
s
2
|
where
λ
1
=
L
S
and
λ
2
=
L
S
s
(i)
1
s
(i
−
1
)
1
s
(i)
2
s
(i
−
1
)
2
2
|
−
|
2
|
−
|
denote the length
i
=
i
=
(i)
the corresponding contour seg-
ment averages. This condition ensures that only contour segments which run
approximately parallel are matched.
4. The intensity gradient in the difference image exceeds a threshold
θ
grad
:
∇
g
¯
(i)
of the contour segments and
s
1
and
s
2
1
>θ
grad
.
u
(w
(i
1
)
1
v
(w
(i
1
)
¯
,
(1.114)
This condition ensures that only contour pixels are accepted whose position is
characterised by intensity changes, i.e. transitions from the background to an ob-
ject, rather than by pixel noise in nearly uniform areas of the difference image. It
is assumed that the surfaces of the objects that may enter the scene, e.g. persons,
are arced rather than flat, such that an intensity difference between object and
background is still perceivable due to shading even if the corresponding reflec-
tivities are identical. This shading contrast determines the value of
θ
grad
.
The segmentation obtained with binary threshold
θ
0
often does not yield contours
around the objects in the scene, but around secondary structures such as shadows
indirectly caused by the objects (Fig.
1.18
b). Hence, for the bounding box of each
blob segmented with binary threshold
θ
0
, the described procedure is repeated with
N
T
different adaptive threshold values
θ
(q)
1
,...,N
T
which are derived from
the intensity histogram of the corresponding bounding box. The values of the adap-
tive thresholds may be chosen to be equally spaced between
θ
0
and the maximum
possible grey value, or to lie within a suitable range around the minimum of the in-
tensity histogram, which is in many cases a binary threshold value that can separate
,
q
=
a
Search WWH ::
Custom Search