Digital Signal Processing Reference
In-Depth Information
Fig. 2.
Affine Particle Sampling
4
Random KD-Forest Observation Matching
After the paritcle candidate
X
, and the corresponding image patch
I
are given, the
following critical step is to model and estimate
p
IX
.
(
|
)
t
t
4.1
Modeling the Observation Estimation by Nearest-Neighbor-Based
Appearance Similarity
We assume that an appearance model is constructed by estimating the similarity of a
certain target candidate to the target observation in previous frame. Thus we could
model the observation estimation based on a certain target candidate
X
is
t
proportional to its likelihood to the target:
1
(5)
p
(|
IX
)
∝
t
t
ˆ
(,
D
XI
)+
ε
t
t
-1
ˆ
(,
where
D
XI
represents the similarity measurement between the observation of
the state candidate
)
t
t
-1
is a small constant to
prevent denominator of the right side in (5) from being zero. The similarity
computation can be computed conducted by nearest neighbor image matching.
A nearest neighbor matching problem could be simply defined as follows: given a
set of points
X
and the target observation
I
, and
ε
t
-1
P={p ,p ,..., p
n
in a vector space
X
, they would be preprocessed in
such a way that given a new query point
q
∈
X
, find the points in P that are nearest
to
q
can be performed efficiently. In this paper, since the image patches are directly
12
