Digital Signal Processing Reference
In-Depth Information
In which
T
is the image acquisition time interval.
3.2
The Measurement Model of Target
The measurement model [4] at time k can be defined as formula (5) shown below:
zk
(
+=
)
h X k ek
(
( ), ( ))
,
k
=
1, 2, 3...
(5)
k
In which
()
zk
∈
is observation vector at time keach observation measurement
are formed from its location and intensitythat is
n
x
and
is the
zk
() (, ,)
=
xyI
ek
()
∈
R
n
observation noise.
There are have less information inside the tracking window that the detection algo-
rithm dependent on, and resulting in the true measurements are missed or mixed with
false measurements, so that the localization of the target is often inaccurate and in-
complete. The current detection algorithm does not fully consider the IR target's in-
tensity and its distribution characteristic that the detected measurements included
false or missed measurements. In order to improve the tracking accuracy of dim mov-
ing target, this paper used spatial and time domain information of target and back-
ground pixels, and put forward a new detecting algorithm.
Because of atmosphere's heat diffusion phenomenon, the point targets in images
obtained by the infrared sensors appeared in the shape of circular or elliptic, and its
grayscale values are follow two-dimensional Gaussian distribution which defined as
formula (6) shown below
1
(
μμ
−
)
(
νν
−
)
2
2
s
(,)
μν
=
s
exp
−
+
0
0
(6)
2
a
a
max
2
2
μ
ν
In which,
and
are the target's intensity distribution area
is Gaussian
2
a
μ
2
a
ν
(,
)
μ
ν
00
distribution center,
μ−<
and
−<
are the diffusion levels of target intensity,
|
|
a
|
vv
|
a
0
0
v
while
a
→
the point spread function tends to
δ
function. Based on the target gray
level value and its distribution characteristics, the proposed algorithm is as follows:
0
the image sequences are putted into the system, and a window is opened
at the predicted position, and the frame differences are taken in time domain, adap-
tive threshold detection in spatial domain are taken in that tracking window respec-
tively. We can see from the paper [3], the residual noise of the differenced image
will obey Gaussian distribution with a mean of 0The time-domain differential
frame operator is given in formula (7),
Step1
|
Ixy I xy
(,)
−
(,)|
>
3
σ
(7)
k
k
−
1
1
Where,
is the variance of the residual noise in the tracking windowWithin the
tracking window
σ
1
according to the local advantages of target intensity
make an
