Digital Signal Processing Reference
In-Depth Information
Fig. 5.15
Objects tracking after six frames: (
a
) input image, (
b
) mask by only motion compensa-
tion and uncertainty analysis, (
c
) improved result by motion occlusion with layer transition
six frames of initial frame using only basic strategy in Sect.
5.3.3.1
. In the object
mask shown in Fig.
5.15
b, the new uncovered regions across different object layers
are lost or mislabeled, resulting from the prediction and segmentation errors accu-
mulating frame after frame.
Motion Occlusion As Layer Transition
Tracking the focused candidate regions by only motion compensation and uncer-
tainty analysis with objects' overlapping introduces errors because of the motion
occlusion even in the newly exposed regions. To handle this problem, we model
the motion occlusion as layer transition, since the emergence of occlusion is always
accompanied by label transition between different object layers. We now discuss
two distinct classes of layer transitions for the occluded pixels corresponding to
background to be covered
and
uncovered new regions
.
Background to Be Covered
If the pixel in the previous frame is labeled as background layer (
f
t
−
1
=
0), it will
p
only transit to a certain foreground object in the current frame.
The determination of the object index is formulated as a Bayesian maximum a
posteriori (MAP) problem:
f
t
p
=
(
|
)
,
arg
max
P
f
p
x
p
(5.7)
f
p
∈
F
foreground
=
{
1
,
2
,...,
N
}
where
f
t
p
is the label of pixel
p
in the current frame at time instance
t
.
F
foreground
is
the foreground label set and
N
is the number of objects. According to the Bayesian
rule, the posterior probability
P
that an observation of pixel
x
p
belonging to
an object can be decomposed into a joint likelihood function
P
(
f
p
|
x
p
)
(
x
p
|
f
p
)
and a prior
P
(
f
p
)
is given as:
P
(
f
p
|
x
p
)
∝
P
(
x
p
|
f
p
)
P
(
f
p
)
.
(5.8)
