Digital Signal Processing Reference
In-Depth Information
Where
∈−
re-pr
esents the pixel that belongs to skin region but not model silhouette and its proba
bility belongs to background,
background
and
skin
j
,
background
p
∈−
R
R
p R R
∈
−
p
R
R
i
image
projection
projection
image
i
image
projection
∈ −
represents the pixel bel-ongs to
model silhouette but not skin region and its probability belongs to ski-n. As Fig.
2 right displayed. To increase computational efficiency, we calculateits
log-likelih
ood instead of multiplication operation.
skin
j
p R R
projection
image
n
m
(4)
non skin
−
skin
log
p
=
log
p
+
log
p
color
i
j
i
=
1
j
=
1
Fig. 2.
The left represents the hand model silhouette got from hand model projection. The right
represents the overlap of hand model silhouette and the separated binary hand image,
is represented by red color,
is represented by blue color. The
RR
−
R
−
R
image
projection
projection
image
green region is the overlapping region.
3.2
Edge Feature-Based Similarity Measurement Function
Edge feature is an another important feature. As it is relatively robust to lighting
change, many tracking system use it to build similarity measurement function. We also
use this feature to measure the similarity between hand image and hand model. There
are many edge detection operators, here we choose canndy operator because it has a
smooth process to reduce the influence of image noise.
(a) (b)
(c)
(d)
Fig. 3.
(a) represent the origin image.(b) is the resulted edge image got from canndy detec-
tion.(c) is the hand model contour got from hand model projection.(d) is the distance image got
from edge image.
The resulted edge image is displayed in Fig. 3(b). On the other hand, we can get
hand model contour from hand model projection, Fig. 3(c) displays the hand model
contour. After getting the edge and contour, we measure their similarities by caculating
their chamfer distance [5]. There have points set
N
ii
and points set
N
ii
,
Aa
=
=
{}
a
Bb
=
=
{}
b
1
1
the chamer distance from
A
to
B
is defined as:
