Many real-time Virtual
Reality applications use head-mounted displays. Detecting and processing of
user head movements are among the critical technical obstacles of such
applications. Today's systems exploit head-mounted inertial sensors in order to
update the virtual scene. The drawback of this approach is that the head
rotations are reported only after the actual movement took place. It could
cause latencies in the VR system that result in an unnatural feel,
disorientation, and simulation sickness in addition to errors in
fitting/matching of virtual and real objects.
In this work we present
a method for reducing the latencies by anticipating future head motions based
on electromyography (EMG) signals originating from the neck muscles. This
anticipation is performed for "nominal" subject in ideal laboratory
conditions.
To adapt this basic
model to a particular subject in general real-life conditions, a new in-line
learning model is proposed. This model constantly learns deviations of the
predicted velocity from the real velocity and acts to correct the basic
predictions.
In order to implement
the in-line model we have developed a new heuristic method, which we call
“in-line SVR”, based on incremental learning with SVM. This method has been
compared to a number of standard pattern recognition techniques, like Neural
Networks and SVR, and has demonstrated superior performance achieving 65-75%
improvement over the basic model.
