Information Technology Reference
In-Depth Information
Refining &
match score
Normal line
p
Match scores
l
frame 1 frame 2
Fig. 2 Illustration of point profiles and refining the matching point. Refining procedure is
carried out as 1D search along the normal line on the frame 2.
Matching
Because of the aperture problem, the motion of edges can only be determined in
the direction normal to the edge. This means that the corresponding point p of
the next frame ( i +1) lies on the normal line, which is the normal at point p on the
current frame ( i ). This is useful as it restricts matching isointensity contour on the
frame ( i +1) along the edge normal. In order to enhance the intensity edge match-
ing, we add a new match score that is the residual error
e of the profiles of p
and p ′ along the edge tangent line, which is shown in Fig.2. Refining the match-
ing point p ′ on the next frame ( i +1) is thus implemented as a 1D search based on
0 ()
(
)
+ along the direction of point p 's gradient (i.e.
the normal line) instead of point p 's gradient, which is also illustrated in Fig.2.
After that, one can re-compute the match scores 012
the match score of
ep ep
()
()
0
0
eee of the points p and p in
terms of their individual intensity gradients rather than the normal line.
,,
Segmentation by Polysegment Alg.
Based on the match scores
eee in the pending areas W , we apply the
Polysegment algorithm as described in the Appendix respectively to the match
scores of 012
,,
012
eee for the layer edge detection. There are two groups here, one is
the group of layer edge points and the other is that of non-layer edge points. For
each match score, we can thus get two cluster centers
,,
(
)
μμ= = .
Moreover, to the points p œ W , there are eight cluster centers. The layer edge
points
()
i
()
i
,
()
i
,
i
0,1, 2
1
2
should
cluster
around
the
two
centers
of
(
)
(
)
cent
=
min
μ
(0)
, min
μ
(1)
, max
μ
(2)
and
cent
=
min
μ
(0)
, max
μ
(1)
, min
μ
(2)
.
1
2
The segmentation of W is obtained as follows,
Search WWH ::




Custom Search