Digital Signal Processing Reference
In-Depth Information
Figure 4.6.
Change Detection Results: for a) frame 90 of coastguard sequence, we
apply our change detection algorithm after majority motion compensation to extract
the changing regions as shown in b). Note that homogeneous regions are not marked
as changing, due to the aperture problem (of Chapter 2).
Some background objects
are marked as changing due to parallax.
techniques
of section
4.4
with
these
motion
estimates
as
our
third
component
of
E
motion
.
2.
Change Detection
As mentioned in Section 2.6, change detection can identify general
regions of change and aid in analysis (see Figure 4.6). Our change
detection has a simple camera motion compensation that uses the
majority motion vector in the frame. After compensating for camera
motion, we use a simple threshold of variance on
∂
I
to find changing
-
∂
I
regions.
Change
detection
requires
a
smaller
smoothness
neighbor-
hood than motion
estimation
algorithms
and
localizes
video
object
boundaries
more precisely.
Change
detection
also
allows
better dis-
crimination
between
object
and
background
edges.
3.
Boundary Motion Coherence
Using edge detection, we can enhance our motion estimation near
object boundaries to find motion discontinuities. Assuming motion
discontinuities coincide with edges and an edge is on an object bound-
ary, an edge is
boundary
motion coherent
(see Fig.
4.7)
iff
(4.7)
→
where
v
i
,v
o
,v
e
and
N
e
are the motions of the inside, outside and edge
regions, and the normal to edge, respectively. The interpretation
of the boundary motion coherence test can be shown in Figure 4.8,
→→
→
Search WWH ::
Custom Search