Digital Signal Processing Reference
In-Depth Information
The following sections present the detection based on a TBD algorithm in detail.
Section 2 develops a model of IR image sequence that contains dim-small moving
target in the sky scene, and preprocesses the sequence by morphological filter. Sec-
tion 3 describes the major detection algorithm, where the energy of IR dim-small
moving target is accumulated without diffused by DP in section 3.1, MSHT threshold
is set to reject invalid candidate trajectory by MSHT in section 3.2 and 3.3. Section 4
is the experimental results analyses. Section 5 gives the conclusion, and related refer-
ences are listed in the last section.
2
Modeling and Preprocessing
2.1
Model of IR Image Sequence with Dim-Small Moving Target
The object will only occupy one or several pixels in the imaging plane when it is far
away from the IR imaging system, although its actual diameter may be more than ten
meters. In such situation, the amplitude of the noise signal presents to be similar with
dim-small objects and the objects is totally mixed up with noise in IR images.
Ideally, assume that the noise in the IR image is conformed to Gaussian distribu-
tion, and IR image sequence meets the principle of superposition. If IR image se-
quence consists of N frames, each of them is of the size
z
(
X
)
M
×
M
,
is the
i
X
in the
i
th frame of the IR image sequence, then the
grayscale value for the pixel
image Z is expressed as
Z
=
{(
zX
)}
∀
X
(1)
i
i
n
(
X
)
X
meets the Gaussian distribution,
Thus the noise component
in position
i
2
n
(
X
)
→
N
(
0
σ
)
2
that is
,
σ
is the variance of the background. The grayscale
i
z
(
X
)
is defined as
i
z
(
X
)
=
n
(
X
)
+
T
(
X
)
∀
X
(2)
i
i
i
i
T
(
X
)
X
.
Here
is the target component in position
i
X
,
H
X
, then
H
If
means there is a target at
means there isn't a target at
0
H
:
z
(
X
)
=
T
(
X
)
+
n
(
X
)
(3)
1
i
i
i
H
:
z
(
X
)
=
n
(
X
)
(4)
0
i
i
Besides the Gaussian noise, background noises like clouds also exist in real IR im-
ages. Suppose the background noise in position
X
as
b
(
X
)
Thus the model of IR
i
image is set as
H
:
z
(
X
)
=
b
(
X
)
+
T
(
X
)
+
n
(
X
)
(5)
1
i
i
i
i
