Cryptography Reference
In-Depth Information
user wants to retrieve the motion contributed by part of a video object, the
trajectory-based representation scheme can not provide a correct retrieval re-
sult. We therefore propose the use of the motion vectors originally embedded
in an MPEG bitstream to construct motion flows in a single shot. Although
the motion vectors do not always correspond to the real motion of objects in a
video when compared with the optical flow, they are relatively easy to derive.
Fig. 9.10.
A current motion vector and the four possible previous motion vectors
that could be linked to it.
Before constructing the motion flows, we must do some pre-processing
steps in order to compensate for camera motion. First, we produce a motion
vector field between the last P-frame in the current GOP and the I-frame in
the next GOP by using the B-frames between the P- and I-frame, as proposed
in [26]. This yields a forward reference motion vector field between any two
consecutive anchor-frames (I- and P- frames). It is well known that a motion
vector field usually comprises camera motion, object motion, and noise. We
assume that the global motion in a video is contributed primarily by camera
motion. Thus, we use the following four-parameter global motion model, which
is fast and valid for most videos [26], to estimate the camera motion from the
motion vector field,
−−−−→
MV
cam
=
zoom rotate
−rotate zoom
x
y
pan
tilt
•
+
.
(9.13)
Once the four parameters have been estimated, we can find the degree
of correspondence between each pixel in the current frame and its counter-
