Information Technology Reference
In-Depth Information
The foreground likelihood
p
F
is calculated by using low background likelihood
pixels on the input image. To extract changes even when the changes are small, our
method employs a histogram foreground likelihood model as follows:
N
B
+
C
p
F
(
i
r
)=
(5)
N
T
+
K
×
C
where
N
B
is the histogram bin value which includes
I
r
,
N
T
is the total histogram bin
value,
K
is the number of bins and
C
is a constant value (in our implementation,
C
1). If
N
T
is less than the fixed threshold
N
T
min
,
P
F
is set to the constant value
P
const
(in our implementation,
N
T
min
=
=
1
255
3
).
The smoothing term
S
rs
uses edge subtraction value of the input image and the
background image, to handle both of increase / decrease of edge value caused by
object placement / removal:
150 and
P
const
=
2
i
r
−
i
s
||
2
e
−
β
|||
i
r
−
i
s
||
−||
|
S
rs
=
(6)
1
β
=
<
)
>
(7)
i
r
−
i
s
||
2
(
||
i
r
−
i
s
||
2
+
||
2
2
where
β
is a parameter that balances the color contrast,
||·||
is the
L
2
norm and
< · >
is the expectation operator.
3.2
Pixel-Level State Detection by Layered Background Model
After the proposed method extracts changed pixels from the input images, then the
method categorizes the changed pixels into the foreground state and the removed-
layer state. To categorize pixels into these three state, background state (unchanged),
foreground state and removed-layer state, we employ the layered background model.
Fig. 3 depicts an overview of our layered background model. The layered back-
ground model contains two background models: the base background and the lay-
ered background. The base background records static backgrounds (e.g. furniture),
and the layered background overlays placed objects on the base background. The
method generates the base background when object movement detection starts (in
this chapter, we call the state before the object movement detection starts as “initial
state”). The method inserts the detected object into the layered background when
the method detects object placement, and delete the detected object from the lay-
ered background when the method detects object removal.
Pixel categorization by the layered background model is performed as follows.
First, the method compares the input image to the layered background, and extracts
changed pixels by the region-level background method. Next, the method compares
the changed pixels in the input image to the base background. if the pixel is chang-
ing from the layered background but not changing from the base background, the
pixel is changing after object placement but is not changing before object place-
ment, so it represents “something removed”. So, the method classifies the pixel as
the removed-layer state. On the other hand, the pixel is changing from both of the
Search WWH ::
Custom Search