Reliable Inference with Missing,
Masking, Malfunctioning or Malicious Sensors
Sponsor: IARPA
This project is concerned with how to reliably
make inferences concerning physical phenomena given practical sensor
impairments. The goal is construction of a general theoretical framework that
can apply to a variety of domains. The initial applications are cognitive
radios, wireless health, and utility monitoring.
The focus of Prof. Pottie’s
group is on the wireless health application, with the main impairments being
noise and drift, improper sensor placement, and missing sensors. One research
thrust is to easily go back and forth between data from visual motion capture
systems and inertial sensors, since the former are costly but accurate, while
the latter are portable and low cost in themselves but as noted above it is
difficult to establish ground truth. If carefully modeled, it would be possible
to simulate a variety of sensor positions on the body, in order to develop more
robust algorithms without the need for costly experiments to acquire large
quantities of ground truth. Along similar lines, the motions themselves do not
usually need all the observations to be recorded in order for motion
trajectories to be reconstructed. Thus motion models can also lead to data
compression.
These experimental activities support
the main goal of understanding how to best construct a multi-level modeling
framework on Bayesian principles so that the best quality of inference can be
made, given the observations available. Abstraction levels in this case include
context of the motions (where, when), activities, sensor pose and placement,
and motion trajectories for particular limbs. Particular algorithms are
associated for making inferences at each of these levels, and also for
iterating among decisions at different levels to improve the inference quality.
For more information on this project,
see http://www.loris.ee.ucla.edu/rim4s