Database Reference
In-Depth Information
The system recognized the actions performed by five dancers using HMMs which
were trained by 50 % of the data samples. Table
11.8
shows the recognition rate
obtained by HMMs compared to others. The average for the 40 gesture classes was
77.3 %, which is significantly improved from the template matching methods (which
performed at 42.4 %).
It was observed that the HMM parameters critically depended on the selection
of training patterns. The performance was unstable when the number of training
patterns was small. In comparison, the larger number of training patterns increased
the recognition rate to: 80.5 %, 85.7 % and 89.4 % when 60 %, 70 %, and 75 % of
samples were used for training, respectively. It was also observed that increasing
the number of trainings [
N
i
in Eqs. (
11.8
) and (
11.10
)] for the template matching
methods (i.e., PO, PSC, PT, and PTSC), has little effect on their recognition
performance.
11.8
Summary
The first part of the chapter presents a new framework and implementation for
the real-time capture, assessment and visualization of ballet dance movements
performed by a student in an instructional, virtual reality (VR) setting. Using
joint positional features, a spherical self-organizing map is trained to quantize
over the space of postures exhibited in typical ballet formations. Projections of
posture sequences onto this space are used to form gesture trajectories, used to
form templates in a library of predetermined dance movements to be used as an
instructional set. Two different histogram models are considered in describing a
gesture trajectory specific to a given gesture class (posture occurrence and posture
transitions). The histogram approach to both of the descriptors offers flexibility
and generalization across instances of movement recorded from a candidate user:
recognition for which, due to the natural variation of the human when repeating
movements and the sensor noise introduced by the Kinect, can be a challenging
task. The recognition evaluation was extended to the online case, where a dance
consisting of continuous gestures is segmented online using a Bayesian formulation
of the recognizer. This formulation shows much promise, effectively delineating a
student's dance movement into constituent gestural units.
In the second part of the chapter, the Hidden Markov Model (HMM) method
is adopted to analyze the sequential data of gesture trajectory on a spherical self-
organizing map This method addresses the temporal information of human motion
and aids in improving recognition accuracy. The experimental result of isolated
gesture recognition using the standard motion capture database shows that the
current method provides significant improvement in recognition accuracy. This
recognizer will be highly important in assessing dance gestures in a completed
virtual reality dance training system.
Search WWH ::
Custom Search