Game Development Reference
In-Depth Information
Another class of common used background modeling methods is based on non-
parametric techniques, as in Elgammal et al. ( 2000 ), Thirde and Jones ( 2004 ). The
most common method is the median and mean value models, as Eqs. ( 8.1 ) and ( 8.2 )
shown, where B
(
x
,
y
)
is the background model value at
(
x
,
y
)
position and L is the
training frame number
t
1
L
B t (
x
,
y
) =
1 (
I i (
x
,
y
))
(8.1)
i
=
t
L
+
B t (
x
,
y
) ={
I i (
x
,
y
) |
i
=
t
L
+
1
t
}
.
(8.2)
median
Mean shift-based methods are also nonparametric ones (Liu et al. 2007a ), com-
prising steps of: (1) representative points' selection; (2) mean shift procedure to
get the candidate models in Step 2; (3) deriving the candidate model by a merge
procedure in Step 3; and (4) obtaining the reliable background model represented
by a frame in Step 4. Elgammal et al. ( 2000 ) builds a nonparametric background
model by kernel density estimation. For each pixel, observed intensity values are
kept for estimating the underlying probability density function and new intensity
probability values can be calculated by this function. The model is robust and can
handle situations where the background of the scene is cluttered and not completely
static but contains small motions which are introduced by the moving tree branches
and bushes. The optimized methods include (Lo and Velastin 2001 ; Cucchiara et al.
2003 ). The approach proposed by Oliver et al. ( 2000 ) is also based on an eigenvalue
decomposition, but this time applied to the whole image instead of blocks. Such an
extended spatial domain can extensively explore spatial correlation and avoid the
tiling effect of block partitioning.
8.3.2 Background Models for Video Coding
As mentioned before, the surveillance videos usually have large and almost static
background, it is worthwhile to specially develop coding methods that consider this
characteristic to significantly improve the coding efficiency. One reasonable solution
is to segment the foreground objects and the background, and then encode them sep-
arately (Wang and Adelson 1994 ; Chai and Ngan 1997 ). To better segment the fore-
ground objects and the background, background modeling methods are researched in
depth. Thesemethods can be roughly classified into parametric ones and nonparamet-
ric ones. With background modeling techniques, many powerful coding schemes are
proposed for surveillance videos captured by stationary cameras (Liu et al. 2007b ;
Vetro et al. 2003 ; Venkatesh Babu and Makur 2006 ; Hakeem et al. 2005 ). How-
ever, these coding schemes rely on accurate foreground/background segmentation
besides the reduction of additional coding bits for the description of segmentation
 
Search WWH ::




Custom Search