Information Technology Reference
In-Depth Information
that contain two walkers (class “1”). We use these sequences to build databases
of labeled signatures for different sets of parameters (
M
,
N
,and
L
are the pa-
rameters, that is the number of bins for the rectangle widths, the number of
bins for the rectangle heights, and the number of intra-frame features aggre-
gated in a signature respectively). For each set of parameters, we employ 5-fold
cross-validation on the corresponding database to assess the precision of the
classification according to the error rate (
E
) defined by
FP
+
FN
TP
+
TN
+
FP
+
FN
,
E
=
(3)
where TP is the number of true positives, TN the number of true negatives, FP
the number of false positives, and FN the number of false negatives.
We tested both
G
W
×
H
(
i,j,t
). The two corresponding se-
ries of results are given in Table 1 and Table 2.
G
W
+
H
(
i,j,t
)and
Table 1.
Error rates obtained for
G
W
+
H
(
i, j, t
)
E[%]
M
=
N
=2
M
=
N
=4
M
=
N
=6
M
=
N
=8
M
=
N
=10
L
=40
15.99
14.43
14.37
14.68
14.76
L
=60
9.77
8.34
8.23
8.86
8.55
L
=70
7.89
7.04
6.70
6.86
7.17
L
=80
6.98
6.05
5.68
6.19
5.91
L
=90
7.51
7.65
7.44
7.09
7.51
L
= 100
9.84
11.58
10.04
10.86
10.97
L
= 120
18.15
18.15
18.16
16.34
17.50
G
W
×
H
(
Table 2.
Error rates obtained for
i, j, t
)
E[%]
M
=
N
=2
M
=
N
=4
M
=
N
=6
M
=
N
=8
M
=
N
=10
L
=40
16.06
14.05
14.07
14.04
14.76
L
=60
9.83
8.96
9.79
10.25
10.79
L
=70
8.01
7.82
8.35
8.35
8.57
L
=80
7.12
7.45
7.5
7.26
7.64
L
=90
8.14
8.62
8.62
8.76
8.62
L
= 100
10.45
9.43
9.84
9.73
11.79
L
= 120
19.81
17.33
16.01
16.01
18.32
G
W
+
H
(
i,j,t
)
signature with
M
=
N
=6,and
L
= 80. We also observe that, for this particular
problem,
L
is the parameter with the largest variability in the result. From our
tests, the best results are obtained for a signature length
L
of 80 frames, a num-
ber that matches the average time to cross the curtain. For
M
and
N
, the choice
of a value is less critical but, from our tests, it appears that
M
=
N
= 4 is an ap-
propriate choice. We also noticed that for this particular problem,
They show that an error rate as low as 5
.
68% is reached for the
G
W
+
H
(
i,j,t
)
G
W
×
H
(
i,j,t
) while having a reduced computa-
tional cost. One explanation to this is that the shadowing effect in the lower part
of the silhouette adds more noise on
has slightly better results than
G
W
×
H
(
i,j,t
) than on
G
W
+
H
(
i,j,t
).
Search WWH ::
Custom Search