Digital Signal Processing Reference
In-Depth Information
14.10
MOTION ESTIMATION IN VIDEO
In this section, we study motion estimation since this technique is widely used in MPEG video
compression. A video contains a time-ordered sequence of frames. Each frame consists of image data.
When the objects in an image are still, the pixel values do not change under constant lighting
conditions. Hence, there is no motion between the frames. However, if the objects are moving, then the
pixels are moved. If we can find the motions, which are the pixel displacements, with
motion vectors
,
the frame data can be recovered from the reference frame by copying and pasting at locations specified
by the motion vector. To explore such an idea, let us look at
Figure 14.52
.
As shown in
Figure 14.52
, the reference frame is displayed first, and the next frame is the target
frame containing a moving object. The image in the target frame is divided into N
N macroblocks
(20 macroblocks). A macroblock match is searched within the search window in the reference frame to
find the closest match between a macroblock under consideration in the target frame and the mac-
roblock in the reference frame. The differences between two locations (motion vectors) for the
matched macroblocks are encoded.
The criteria for finding the best match can be chosen using the mean absolute difference (MAD)
between the reference frame and the target frame:
2
N
1
k ¼
0
N
1
l ¼
0
jTðm þ k; n þ lÞRðm þ k þ i; n þ l þ jÞj
1
N
MADði; jÞ¼
(14.27)
u ¼ i; v ¼ j for MADði; jÞ¼
minimum
;
and
p i; j p
(14.28)
There are many search methods for finding the motion vectors, including optimal, sequential, or brute
force searches, and suboptimal searches such as 2D-logarithmic and hierarchical searches. Here we
examine sequential search to understand the basic idea.
R
f
T
f
(,)
mn
Macroblock
(
,
)
mn
MV
N
N
M
acroblock
2p+1
(
m
u
nv
,
)
Searc
h
window
FIGURE 14.52
Macroblocks and motion vectors in the reference frame and target frame.
Search WWH ::
Custom Search