Project Abstract:
We will investigate methods of exploiting the availability of different
types of signals such as video, sound and physiological signals
for improving the performance of classifiers and for enabling
semi-automatic adaptation to the environment.
One key requirement for monitoring systems such as the elderly
health monitoring system is to reduce the occurrences of false
positives as users will turn off systems that “cry wolf” too often.
The use of multiple types of signals can reduce false positives by“explaining away” some of the false indicators e.g. a video showing
a person walking around while his heart rate sensor measures zero
heart rate probably means that the heart rate sensor is not
functioning properly rather than a heart failure. One challenge is
hence to develop models and learning methods that can increase
specificity while maintaining sensitivity by exploiting multiple types
of signals. The difficulty of this problem is compounded by the fact
that the events that we want to detect are rare. We expect that
most of the measurements using the different signals would be
done independently (without the other sensors being present) with
only a small amount of combined measurements available. Another
requirement for adaptive systems would be to adapt to the new
environment without (or with very little) explicit feedback from
humans. A system that is tailored to a particular individual or
environment will be able to outperform a system that is designed
to operate robustly across different individuals or environments.
By exploiting the independence of errors in classifiers based on
the different signals, it may be possible to adapt the system “blind”
(without explicit human feedback) or in a semi-supervised (with
very little human feedback) manner. The main sources of signals
available will be video streams, physiological signals and audio
signals. Processing of video streams will require tracking of human
subjects using algorithms from the field of computer vision. We
intend to develop machine learning algorithms that will be able to
take into account evidence from multiple sources of information
and incorporate them in a meaningful way to achieve the objectives
of a monitoring system. |