Digital Signal Processing Reference
In-Depth Information
Fig. 5.8
Multiview segmentation of initial frame:
First & Second row
: Reading sequence;
Third
& Fourth row
: Calling sequence;
top
: input image;
bottom
: superimposed mask
section, we discuss the proposed algorithm that first extract the foreground as mul-
tiple overlapped human objects, and then segment them into individual objects.
Adaptive Background Penalty with Occlusion Reasoning
When the segmentation starts from the initial frame with overlapped objects, not all
parts of the objects in the target view are also visible in other reference views, as
shown in Fig.
5.9
a-c. The interview occlusions displayed in Fig.
5.9
d exist not only
in the transition between the object and the background (interobject occlusion), but
also in the interior of the object (intraobject occlusion). In order to distinguish the
interobject occlusion from intraobject occlusion, we can assign more background
energy on the interobject occlusion rather than on the intraobject occlusion, adaptive
penalization of the occlusion region is adopted that the penalty factor
α
bp
in (
5.4
)is
assigned according to the motion property using the following equations:
P
motion
(
m
p
|
f
p
=
0
)
α
bp
(
p
)=
P
motion
(
m
p
|
f
p
)=
log
h
(
m
p
|
f
p
)
,
(5.6)
P
motion
(
m
p
|
f
p
=
1
)+
η
