Information Technology Reference
In-Depth Information
the convergence is smaller too. As to the wavelet coefficients
dk
at time
k
,
( )
need to be estimated. We adopt an iterative predic-
tion method, the iterative relationship is given as following:
{(
pk
+
1), (
pk
+
2), , (
pk i
+
)}
pk
(
+=
)
f dk d k
(
( ),
(
−
),
,
d k i
(
−+
))
pk
(
+=
2)
f pk
(
(
+
1),
dk
( ),
,
dk i
(
−+
2))
(7)
......
(
pk i
+ =
)
f pk i
(
(
+−
1),
pk i
(
+−
2),
,
dk
( ))
fx
is the LMS prediction operator.
With wavelet packet decomposition, original video traffic with complex properties
of long-range and short-range dependence can be transformed into a sequence in wave-
let domain with short correlation. On this basis, by utilizing LMS algorithm to
achieve approximate prediction of the wavelet coefficients after the wavelet packet
transform, we can realize the real-time video traffic prediction algorithm, which is
described as followings:
Where
()
─
Step 1: conduct wavelet packet decomposition on each group of video frames
acquired and output sequence of wavelet coefficients;
─
Step 2: use LMS algorithm to predict the next set of wavelet coefficients in the
wavelet domain with the new acquired wavelet coefficients;
─
Step 3: conduct inverse wavelet transform with predicted wavelet coefficients, then
the prediction of video traffic in the next time window is realized;
─
Step 4: every time new video frame traffic is acquired, traffic data should be rec-
orded, then repeat step 1, 2, 3.
4
Simulatioin and Analysis
In this paper, experimental video “StarWars” and “News” are chosen from the
MPEG4 video trace database of Berlin university. These Video adopt MPEG-I
compression standard , with QCIF format, frame rate set to 30fps, and quantitative
parameters is fixed at 10, 14 and 18.
First of all, NMSE (Normalized Mean Squared Error ) is introduced to evaluate
performance of our algorithms. NMSE is defined as follows:
11
1
[
]
[
]
(8)
2
2
NMSE
=
xt
()
−
x t
() =
xt
()
−
x t
()
Var xt
( ())
σ
N
N
2
Where
xt
is the actual frames at time
t
,
xt
is the prediction of
()
xt
. N is the num-
()
()
ber of predict test,
is the variance of the observed sequence, 1024 frames are
randomly extracted from “Star Wars” and “News”. Prediction is conducted on follow-
up continuous 200 frames, and the prediction performance is studied.
σ
2
In order to reduce the prediction time and avoid constantly modify of the model, we
adopt an iterative prediction method and video traffic signal is decomposed with
