Game Development Reference
In-Depth Information
background model-based fast-and-efficient transcoding (FET) platform with
AVC/H.264 high profile. They transcode the input AVC/H.264 streams at 1,000kbps
for eight sequences (crossroad-sd, overbridge-sd, office-sd, bank-sd, crossroad-cif,
overbridge-cif, snowroad-cif, and snowgate-cif) to the output streams at bit rates
of 64, 128, 256, and 512kbps. The four methods are the Gaussian Mixed Models
(Haque et al.
2008b
) using 1 or 5 models for each pixel (GMM-1 or GMM-5), the
Mean Shift (namely MS) proposed in Liu et al. (
2007a
), and the popularly used
Gaussian running average (RA). For background modeling in surveillance and con-
ference video transcoding, as is referred in Piccardi's (Piccardi
2004
), performance,
memory cost, and running time are the same important factors. The calculations for
their memory cost in each pixel position are listed as follows. (1) RA: one current
pixel with type of char and one float-precision mean value for each pixel should be
buffered. (2) GMM-X: besides the buffered input pixel, a GMMmodel is required to
be buffered. The model is composed of double-precision mean value, variance, and
weight. Moreover, an 8-bit value should be stored to count the number of matched
points for each GMM model. (3) MS: Mean shift-based algorithms usually buffer
all the training frames and very few additional temporal variables are used for the
clustering and sorting operations.
To maintain or improve background quality, an ideal solution for background
modeling is to calculate the mean value of all the purely background pixels in the
training frames. However, it is very difficult in recent years to exactly justify which
pixels belong to the background. Physically, background equals to the most fre-
quently appearing content. This inspires FET to utilize a novel segment-and-weight-
based running average (SWRA) to approximately calculate background by paying
larger weight on the frequently appearing values in the averaging process. Because
SWRA is based on a running average procedure, there will not be large memory
cost and computational complexity. Generally, SWRA divides the pixels at a posi-
tion in the training frames into temporal segments with their own mean values and
weights, and then calculates the running and weighted average result on the mean
values of the segments. In the process, pixels in the same segment have the same
background/foreground property and the long segments have larger weight. This
method ignores the foreground/background property of each segment, so foreground
recognition is avoided. Meanwhile, low memory cost and no delay modeling are
guaranteed.
In detail, SWRA models a background value of pixels at position
(
x
,
y
)
by fol-
lowing five steps:
1.
Initialization
: Initialize background model value AVG and its weight
W
for the
following weighted average procedure to 0, and then create first segment. Length
of the first segment
L
equals to 0 and its mean value
a
vg
=
0. The model value
before the current segment avg' is also set 0.
2.
Calculate the threshold for segmenting
: Supposing
2
is the
μ
is the mean value,
σ
mean square error, the probability of
|
f
(
x
)
−
μ
|
>
2
σ
in normal distribution f(x)
is less than 4%. So we use 2
as the threshold
th
to temporally segment a pixel
in training frames. The threshold
th
is initialized to 14 and updated by two times
σ
