Information Technology Reference
In-Depth Information
Trajectory direction histogram is a constrain for raw classification. Correlation
distance between TDH is calculated as below
I
(
H
a
(
I
)
H
a
)(
H
b
(
I
)
H
b
)
−
−
d
(
H
a
,H
b
)=
I
(
H
a
(
I
)
(7)
H
a
)
2
I
(
H
b
(
I
)
H
b
)
2
−
−
N
J
H
k
=
1
H
k
(
J
)
(8)
The smaller the value of
d
is the lower similarity between TDH is.
The number of trajectories is
n
, and the set of Trajectory Direction Histogram
for
n
trajectories and Trajectory Direction Histogram distance between every two
trajectories are
H
=
{
H
1
,H
2
,...,H
n
}
,
D
=
{
d
(
H
i
,H
j
)
|
i
=
j
;
i,j
=1
,
2
,...n
}
.
For raw classification using TDH similarity, a modified k-means method is
adopted. Generally, the first step of k-means cluster is to confirm the
k
class
centroid stochastically, but those centroid may be close to each other, which
is not favorable for the further classification. So we'd better choose the ones
between which the distance value
d
are the smallest in set
D
to be the
k
centroid.
And a threshold
d
min
is settled. If the maximum distance value
d
between a
trajectory and each class centroid is bigger than
d
min
, the trajectory can be the
member of the class with the highest similarity. Otherwise the trajectory can be
the centroid of a new class.
Class centroid is defined as
m
Cid
=
1
m
H
i
;
(9)
i
=1
where
H
i
∈
Class,i
=1
,
2
,...,m
.And
m
is the number of trajectories belong to
Class
.
The more specific algorithm is shown in Algorithm 2.
S
1
,S
2
,...,S
n
}
Fine Classification.
Segments in main sequence
S
n
=
also have
position and acceleration features. Through the new main sequence, we can pick
out main points set
P
L
=
{
{
P
0
,P
1
,P
2
,...,P
n
}
. Main points are calculated as
below
P
i
=
P
i−
1
+(
S
i
)
E
(
x,y
)
−
(
S
i
)
S
(
x,y
)
(10)
where,
P
0
=(0
,
0).
The raw classification above have divided trajectories into
k
classes. We need
to subdivide them by using more specific position and acceleration features to
ensure an effective classification result. This time, position distance
d
p
and ac-
celeration distance
d
a
between trajectories are defined as below
((
P
a
)
k
−
n
d
p
=
1
n
(
P
b
)
k
)
2
+((
P
a
)
k
−
(
P
b
)
k
)
2
(11)
k
=0
Search WWH ::
Custom Search