Database Reference
In-Depth Information
parse a sequence of movements into a set of
postures
, then representing or modeling
sequences of postures as a gesture. The topological map afforded by the SOM can
help in this regard, as the map can be used as a basis for indexing, classification, and
extraction of inherent relationships in the underlying data.
Postures are represented by a particular state of sensor values at an instant in
time. For instance, Microsoft Kinect uses skeletal tracking of joint positions. Full
motion capture technology used in film making and animation can also be used to
describe postures, which are then mapped onto the SOM. A gesture can then be
represented as a path or trajectory on the map, as traced by projecting a temporal
series of postures. Each path can be used to model a type of gesture, or transitions
between possible postures for a given gesture can be extracted. Unknown gestures
can be recognized through a matching process (template paths) as the path of an
unknown gesture is traced on the map.
Previous methods for trajectory analysis of the self-organizing maps have
attempted to use the sparse code [
343
,
351
,
352
], and the posture occurrence as
an analog to the bag-of word model [
346
,
347
,
350
]. These methods have some
limitations in their capability for temporal information analysis, since they use only
the existing of the nodes and the frequency of occurrence of the nodes (key postures)
on maps. It is evident that the transitions from posture to posture (or from one form
in space to next) preserve more temporal information about the dance sequence than
the postures (forms in space) do themselves [
340
]. In order to perform trajectory
analysis on the SSOM, in this Chapter, two methods for transition analysis of the
SSOM trajectory are presented. The first method uses transition metric and the
second method adopts the hidden Markov model (HMM) for modeling gesture on
the multiple-codebook SSOM. Based on the experimental study, the newly proposed
method appears to be very effective for recognizing human actions and outperforms
the previous methods.
Section
11.2
will look into an architecture for a dance training system in the cave
automatic virtual environment (CAVE). Section
11.3
presents the SSOM method for
the construction of posture space to explore gesture trajectory. Section
11.4
presents
the application of SSOM for the characterization of dance gesture. Section
11.5
presents trajectory analysis methods for gesture indexing and the construction
of template matching. Section
11.6
extends these template matching methods to
online recognition and gesture segmentation. Section
11.7
presents the HMMs for
transition analysis of the trajectory on the multiple-codebook SSOM.
11.2
Dance Training System
The architecture of the VR dance training system is shown in Fig.
11.1
, which
includes four components: the motion capture, gesture recognition module, assess-
ment and visual feedback module, and the CAVE. The CAVE has four stereoscopic
projectors and four corresponding screens. Driven by a graphics cluster of five
nodes, one node serves as the cluster master, while the other four drive the
Search WWH ::
Custom Search