Digital Signal Processing Reference
In-Depth Information
14.2 Motion Estimation
Happily, there are techniques to reduce this brute force
method. A larger block size can be used initially, over larger
portions of the image, to estimate the motion trend between
frames. For example, we can enlarge the block size used for
motion estimation to include macrocells on all sides, for a 48
48
pixel region. Once the motion estimation is completed, this can
be used to dramatically narrow down the number of permuta-
tions to perform the MAD or MMSE over. This motion estimation
process is normally performed using luminary (Y) data and per-
formed locally for each region. In the case of the camera panning,
large portions of an image with a common motion vector will
move, but in cases with moving objects within the video
sequence, different objects or regions will have different motion
vectors.
Note that these motion estimation issues are present only on
the encoder side. The decoder side is much simpler, as it must
recreate the image frames based upon data supplied in
compressed form. This asymmetry is common in audio, image,
video or general data compression algorithms. The encoder must
search for the redundancy present in the input data using some
iterative method, whereas the decoder does not require this
functionality.
Motion estimation vector between
middle and last frame
Figure 14.2. Motion estimation.
Once a good match is found, the macroblocks in the following
frame data can be minimized during compression by referencing
the macroblocks in previous frames. However, even a “good”
match will have some error. These errors are referred to as
