Database Reference
In-Depth Information
Table 11.5
Gesture
recognition results averaged
over the 30 gestures defined
in Table
11.2
Average recognition accuracy
Testing data L1 L2 HI
Teacher
PO 77.7 74.3 77.7
PSC 58.0 61.3 57.7
PT 96.0 79.3 96.0
PTSC 83.0 84.3 83.3
Student
PO 66.7 66.3 66.7
PSC 54.7 56.0 54.7
PT 88.3 73.3 88.3
PTSC 79.7 83.0 76.7
These include the reversals of the gestures
dataset can reach 96 %. However, the system has a poorer performance at about 88 %
for recognition of the Student dataset. This may be because the dance sequences
performed by the student may be inconsistent, as compared to the teacher.
11.6
Online Gesture Recognition and Segmentation
In order to perform matching between an incoming gesture and known templates,
the incoming set of postures is projected onto the SSOM to extract the unknown
posture sequence
S
. This projection is conducted
online as the student is performing a set of moves. The task of recognition is non-
trivial, due to the differing lengths of gestures (across classes), and the differing
speeds with which they may be enacted (by the student/teacher). In order to address
this, an online probabilistic framework demonstrated in [
352
] is adopted. The
standard Bayesian framework is utilised for progressively estimating an updated
posterior probability
P
=(
u
1
,...,
u
t
,...,
u
T
)
,
t
∈
[
1
,
T
]
K
gesture classes. The
likelihood is computed at each unit of time by considering the single posture
triggered on the map, and whether or not it occurred in each gesture template. In
[
352
], the likelihood
P
(
k
|
S
)
for each of the
k
=
1
,...,
was computed as the ratio of the existence of the current
posture in gesture class
k
, to the total number of different postures in class
k
.Inthe
current work, the likelihood is reframed as a histogram intersection [Eq. (
11.17
)],
between a progressively growing sequence
S
(inclusive of postures from time
t
0
to
t
), which may be described as a histogram of either: PS, PO, PT or PTSC (defined
in Sect.
11.5
), versus the corresponding templates for each gesture class.
Let
h
s
be the input histogram for the current sample at time
t
, and
h
k
to be the
reference template for the class
k
. We thus define (for time
t
), the posterior
P
t
(
(
S
|
k
)
k
|
h
s
)
,
likelihood
P
t
(
h
s
|
k
)
, and prior probabilities
P
t
(
k
)
according to the following:
P
t
(
h
s
|
k
)
P
t
(
k
)
P
t
(
h
s
|
k
)
P
t
(
k
)
P
t
(
k
|
h
s
)=
=
(11.14)
P
t
(
h
s
)
∑
P
t
(
h
s
|
k
)
P
t
(
k
)
Search WWH ::
Custom Search