Information Technology Reference
In-Depth Information
To improve robustness of the descriptor against error in the location of a hypoth-
esis, the motion confidence within each hypothesis is first Gaussian weighted. In
this way, we give higher weight to the contribution of motion near the center of the
hypothesis window, whilst less emphasis is placed to further away from the center.
The Gaussian weighting mask we used in our experiments is as follows. We set it to
the window size of a hypothesis created by the static human detector, where g
(
,
)
x
y
is an anisotropic Gaussian envelope given as:
Z ex p 1
T
1
y 0 ]) Σ 1
g
(
x
,
y
)=
2 ([
x
,
y
] [
x 0 ,
([
x
,
y
] [
x 0 ,
y 0 ])
(5)
where
is the spatial coordinates of the
points within the local region, and Z is the normalization factor.
[
x 0 ,
y 0 ]
is the center point of the region,
[
x
,
y
]
Σ
is the spatial
correlation matrix. We set as 1
,where w
/
2 w 0
01
,
h are the width and height
of the hypothesis window. Examples of measuring the Gaussian weighting mask
against the corresponding bounding boxes are shown in Fig. 10. To further make
the descriptor invariant to scale changes, we normalize the hypothesis regions into
a region with of the same size of 32x32 before partitioning the motion template.
/
2 h
2.4
Pyramid Bayesian Verification
Pyramid GMM Modeling. In order to cope with variation in motion under differ-
ent viewpoints and human poses, we need to model the long-term motion distri-
bution of people. To this end, we collect a set of motion templates of human after
background modeling as in Eq. (3), denoted by q i ,
i
=
1
,
2
, ...,
n . The corresponding
motion descriptors are represented by Y i ,
n . On the other hand, we want
to incorporate the spatial constrains at different levels in the pyramid, we introduce
a spatial pyramid Gaussian mixture modeling (GMM) approach . The idea is that
for each pyramid level, we model the conditional distribution of descriptors as a
Gaussian Mixture, denoted as p
i
=
1
,
2
, ...,
, and the whole pyramid GMM model is ex-
pressed as the combination of those GMMs with different weights. Mathematically,
the model of descriptor Y can be written probabilistically as follows:
(
Y
|
l
)
p
(
Y
)=
l p
(
Y
|
l
)
p
(
l
)
(6)
In this model, l denotes the pyramid level and p
is the prior associated with each
level (the setting of its value will be explained later in this section). The conditional
distribution p
(
l
)
(
Y
|
l
)
w.r.t. l is specifically modeled as a GMM:
)=
i
p
(
Y
|
l
α
i N
(
y
(
l
)
; u i
, Σ
)
(7)
i
I
(
l
)
Where I
(
l
)
is the number of components in a GMM at level l and y
(
l
)
represents the
motion descriptor at level l , i.e., Y
(
l
=
i
,
i
0
,
1
, ...,
L
)
.
( α i ,
u i , Σ i )
are the parameters
of Gaussian mixtures.
 
Search WWH ::




Custom Search