Digital Signal Processing Reference
In-Depth Information
By varying
P
and
R
, we have LBP operators under different quantization of the
angular space and spatial resolution and multi-resolution analysis can be
accomplished by using multiple
LBP
P,
R
operators. In this work, as shown in Fig. 1,
we choose the
LBP
8, 1
.
b
(
x
i
) is represented as
LBP
(
x
i
) in LBP, we obtain the LBP
model according to Eqn. (2), the LBP histograms in target area are divided into
2
7
=128 bins. The LBP image is obtained by
LBP
8, 1
as shown in Fig.2.
24
0
1
12
4
1
79
63
0
2
(01011111)
= 95
57
33
thresholding
1
0
35
77
44
1
1
91
1
Fig. 1.
Illustration
of LBP8, 1
Fig. 2.
The LBP image processed with LBP
8, 1
.
The tracking procedure is performed via two stages. Fuse the color and LBP
features to get the weights in the PF and MS algorithm, and integration of MS and PF
to prevent degeneracy problem and reduce the required number of particles. With
respect to the two stages, how to fuse the two image features are depicted as follows.
Treated as the observation information, the Bhattacharyya distance between target
and candidate region histograms
i
c
i
d
can be calculated by
d
and
d=
1-ρ y
()
with
[
]
m
i=1
()
the Bhattacharyya coefficient
ρ y ρ py,q =
()
≡
()
p yq
. Given the
u
u
i
i
2
2
variance of Gaussian distribution σ, define
w
=
exp
(
d
) / 2
σσ
)
/
2
π
as the
(
−
c
c
i
i
2
2
weights extracted from color information, and
w
exp
(
d
) / 2
σσ
2
π
as the
=
)
/
(
−
l
l
i
weights extracted from LBP. And denote the likelihood measurement
p
as
the weights of particles in PF. In the PF algorithm, we fuse the color and LBP
information to get the particle weights by multiplying the weights:
(Z | X )
k
k
i
2
i
2
+
() ()
d
d
l
c
i i
cl
2
www
i
=
exp
2
πσ
(5)
=
{
−
} /
2
2
σ
