Automatic Target Recognition and Tracking
Wednesday 4 June, 10:20--12:10
- Organizer: Scott VANDER WIEL
- Bell Labs
- Moderator: Scott VANDER WIEL
- Bell Labs
- Katherine Hansen SIMONSON
- Sandia National Laboratories, Albuquerque, NM, USA
A variety of discriminant methods are available
for application to pattern classification problems in which each
pattern encountered is known to come from one of the pre-defined
categories on which the classifier was trained. However, many
of these methods perform poorly when they encounter new patterns
that do not fall into any of the known categories. An
alternative approach is suggested, which uses multiple hypothesis
tests and is designed to handle previously unseen patterns
gracefully. Its application is illustrated on a data set
containing series of high-resolution range profiles from ten
different aircraft.
[Katherine Hansen Simonson, MS 0844, Sandia National Laboratories, PO Box 5800,
Albuquerque, NM 87185-0844 , kmsimon@sandia.gov]
- Anuj SRIVASTAVA
- Brown University, Providence, USA
Automated Target Recognition (ATR) is a problem of great
interest in both the military and the civilian applications.
Moving rigid targets are assumed observed via
standard remote sensors (Video, HRR, IR, SAR etc.)
and algorithms are developed to analyze the observations for
detection, tracking and recognition of objects of interest.
Taking a Bayesian approach to unify the
ATR problem, we generate
stochastic inferences on complex scenes using Jump-Diffusion algorithm.
In here,
an ergodic Markov process,
composed of continuous
diffusion paths and discrete jump transformations,
is constructed to sample from the posterior empirically
evaluating the conditional means.
Newtonian mechanics is utilized to build prior on the target trajectories.
Due to non-flat geometry of the parameters such as pose,
the formulation is extended to
matrix Lie groups.
This Bayesian approach to ATR has the added advantage in that it
provides natural metrics for
analyzing the performances of ATR algorithms.
We present Hilbert-Schmidt lower bounds for
evaluating the performance in parameters (pose, position, temperature)
estimation.
Also, the asymptotic expressions for the probabilities of
error in target detection and recognition are derived.
[Anuj Srivastava, Division of Applied Mathematics, Brown University, 182 George Street,
Providence, RI 02912, anuj@dam.brown.edu]
Back to JRC program
Back to JRC Home Page