Database Reference
In-Depth Information
Table 11.6
Continuous gesture dataset
Ballet dance
# instances ( # frames )
Postures (gesture sequence)
Teacher
Student
Rest
ₒ
1st position
ₒ
2nd position
ₒ
3rd position
ₒ
4th position
1 (281)
1 (273)
ₒ
5th position
ₒ
Rest, (
G
6
,
G
1
,
G
2
,
G
3
,
G
4
,
G
5
)
The results for the Student's performance are also quite satisfactory, as the person
performing the movements is different from the Teacher, and more so, their ability
to repeat the correct movement is somewhat limited. Nevertheless, with some minor
noise, the selection of gesture class appears to follow the actual sequence (i.e.,
recognition accuracy of 100 %). When confusion does occur, nearby postures are
selected for a relatively brief period before switching back to the correct gesture.
11.7
Trajectory Analysis on the Multicodebook SSOM
Using Hidden Markov Model
The application of HMM for pattern recognition is well established [
337
]. The rel-
evant work in [
338
] has demonstrated the application of HMM for hand gesture
recognition, by learning the coding symbols of the self-organizing feature map
(SOM). In the current work, the HMM is adopted and the proposed system has
different aspects from the previous works. The self-organizing method is imple-
mented with the SSOM in the current work instead of SOM, and the multicodebook
is designed for the SSOM in order to minimize the vector quantization errors. It is
argued that a random error in the detection of the joint's positions in the postures of
gesture has an effect on the performance of the recognition stage. This error is due
to the variance in input posture sequence, e.g., sensor noise, inexact repetitions, etc.
Figure
11.8
illustrates the application of HMMs and multicodebook SSOM to
build the isolated gesture recognizer. There is a set of
K
gestures to be recognized
and each gesture is modeled by a distinct HMM. The vector quantization is
implemented by the multicodebooks, whereas the system in Fig.
11.3
uses a single
codebook. This design is motivated by the study in [
353
], which suggest that it is not
possible to create a universal codebook (efficient for each data class to be encoded).
In addition, although a low distortion can be achieved by a large sized codebook,
this leads to problems in implementing HMMs with a large number of parameters.
According to Eq. (
11.5
), let the SSOM indices obtained by the quantization of
x
and
x
u
, where
+
ʔ
x
be respectively described by
Q
(
x
)=
u
and
Q
(
x
+
ʔ
x
)=
ʔ
x
is small, the posture coordinates with
relatively small variance are mapped to the same node of the SSOM so that
u
x
is the random error. In the case where
ʔ
u
.
This error will not affect the recognition since it is compensated during the SSOM
clustering process. On the other hand, when
=
u
, the introduced error
x
will affect the trajectory
Tr
. The result of this error can be observed from Fig.
11.5
.
ʔ
x
>>
:
u
=
ʔ
Search WWH ::
Custom Search