Digital Signal Processing Reference
In-Depth Information
Fig. 5.9
Adaptive background penalty with occlusion reasoning
=
=
where
f
p
0 is for the static background and
f
p
1 for the moving object.
m
p
(
|
)
is the motion vector of
p
and
P
motion
m
p
f
p
is the motion log-likelihood of the
pixel associated with the label.
is a small value to avoid the division by zero.
Equation (
5.6
) indicates that the interview occlusion with small motion is more
likely to be the interobject occlusion resulting in large value of
η
α
bp
, and vice versa.
Figure
5.9
e illustrates the segmentation result with constant
α
bp
, where the inter-
view occlusions are equally penalized to be the background using the same factor,
whereas the improved result using adaptive
α
bp
is evident in Fig.
5.9
fwherethe
background penalty is changed according to the value of
α
bp
.
Depth-Assisted Object Segmentation
Given the extracted foreground regions, object segmentation is equivalent to a k-
class pixel labeling problem. By assuming that the human objects stand in the
different depth layers, a coarse labeling field as shown in Fig.
5.10
b can be obtained
by
k
-means clustering of the depth map, where the number of human hypotheses is
automatically determined as the number of continuous bins of the depth histogram.
Due to the outliers in the estimated disparity field and the resultant reconstructed
depth map, misclassifications exist in the coarse labeling field especially in the area
of intraobject occlusion. We improve the initial labeling using the
depth ordering
method that If we know the layer
L
1
is behind layer
L
2
, the occlusion region must
belong to the
L
1
.
Because of the multiple overlapped human objects, segmentation with occlu-
sion cannot be solved using bi-label graph cut. Given the initial results as shown
