Digital Signal Processing Reference
In-Depth Information
Figure 4.5.
Iterations of multiresolution optical flow for progressive motion estim-
tion a) At a low resolution, we do a optical flow algorithm over resolution block b)
we split the low resolution motion information block and use it as initial conditions
for the next resolution, c) we find the motion vectors at the higher resolution. We
repeat the process and increase resolution until we reach pixel-level resolutions.
(4.6)
where the projected motion field is
v
(
x, y, t
)=
→
〈
v
x
(
x, y, t
),
v
y
(
x, y, t
)
〉
and
g
motion
is a thresholding function.
In our system,
E
motion
consists
of three extraction techniques:
1.
Multiresolution Optical Flow
We use a multiresolution version of the original Horn and Schunck
algorithm. Displacements of more than 2 or 3 pixels can cause the
updating scheme of the original Horn and Schunck algorithm to fail.
A multi-resolution pyramidal scheme can progressively estimate the
motions at increasing resolutions (see Fig. 4.5) and use the motion es-
timates of the previous resolution to guide calculations of the current
resolution. The motion field may be calculated in three successive
passes at resolutions of
,
and used as a initial conditions for the motion field of higher reso-
lution. This process allows the optical flow to expand its range of
maximum displacement from approximately 3 to 12 pixels.
Motion clustering of this optical flow motion field can localize video
objects and approximate the video object shape. As mentioned in
Section 2.4, this optical flow algorithm is inaccurate near the object
boundaries, although it gives good qualitative estimates of an ob-
ject with respect to its motion. We use the optical flow to initialize
our optimization and localize our objects. However, for precise de-
termination of the object boundary, we combine our image analysis
-
1
1
-
and 1, expanded to the next resolution,
Search WWH ::
Custom Search