Motion Database Retrieval with Application to Gesture Recognition in a Virtual Reality Dance Training System - Multimedia Database Retrieval: Technology and Applications - page 306

Table 11.1 Isolated gesture

Ballet gesture
Label

dataset

Description # instances ( # frames )

G 1 1st position ₒ 2nd position 8 (56-96)

G 2 2nd position ₒ 3rd position 10 (57-77)

G 3 3rd position ₒ 4th position 8 (59-75)

G 4 4th position ₒ 5th position 10 (48-81)

G 5 5th position ₒ 6th position 10 (43-80)

1st position 10 (41-88)
Note : First position (posture 1): both arms lifted in front; second
position (posture 2): both arms open; third position (posture 3):
left arm in front and right arm lifted up; fourth position (posture
4): left arm open to the side and right arm lifted up; fifth position
(posture 5): both arms lifted up; sixth position (posture 6): both
arms put down

G 6 6th position ₒ

Table 11.1 summaries the dataset used for training the SSOM. The dataset
included a set of six isolated gestures (i.e. each gesture G 1 - G 6 was recorded
individually, independent of any sequence of other movement/gestures). These
gestures were the six basic positions of ballet dance [ 354 ]. Gesture G 1 was the dance
gesture moving from the first position to the 2nd position, where the 1st position was
the posture with both arms lifted in front, and the 2nd position was the posture with
both arms open. All gestures G 1 - G 6 are defined in Table 11.1 .
In the experiments, observations are made as to how the variability in repeated
gestures maps into posture space. Figure 11.5 shows a series of mappings of gesture
instances (columns) per gesture type (rows). A visualization of the SSOM and
associated gesture trajectories shows that even differences in frame length and
duration of the gesture (variations of up to 40 % difference in frame length) do
not appear to impact the consistency with which the gesture maps onto posture
space. All gestures appear to trace quite characteristic and repeatable paths on the
unit sphere. The start (solid blue marker) and end points (solid red marker) of the
trajectories are also shown. Although gesture G 5 and G 6 are quite similar in terms of
the postures traced, there is quite a clear difference in the direction of the trajectory.
It is clear from these mappings that the paths traced for different gestures
are quite unique from one another, which is expected to translate into better
discrimination between trajectories (and therefore, gestures). The consistency of
the mapping indicates some stability in the representation of gestures, and suggests
sufficient overlap should exist when generating histogram templates.

11.5 Trajectory Analysis

( ) , ( ... ... )

A posture can be represented in time series as x whichisaunitofa

gesture element, where T is the time length of the gesture. In this way, the gesture G i
can be represented by G i

t 1 t T

=( ( ) ,..., ( ) ,..., ( ))

. As shown in Fig. 11.5 , each
input vector x is quantized by the set of SSOM weight vectors. For convenience, we

x i 1 x i t x i T