EVIDENCE ACCRUAL FOR MODEL BASED VISION

SUCCESSFULLY APPLYING MODEL-BASED REASONING TO REAL WORLD VISION PROBLEMS (E.G. AUTOMATIC TARGET RECOGNITION ?ATR?) REQUIRES ACCURATE REPRESENTATION OF COMPLEX RELATIONSHIPS AMONG THINGS IN THE WORLD AND OBSERVABLE IMAGE FEATURES. EVIDENCE ACCRUAL TECHNIQUES MUST BE ABLE TO OPERATE OVER THESE RELATIONSHIPS WITH LITTLE LOSS IN INFORMATION. CURRENT TECHNIQUES SUCH AS BAYESIAN NETWORKS SHOW PROMISE, BUT HAVE SOME LIMITATIONS IN THE NATURE OF THE RELATIONSHIPS THEY CAN EASILY REPRESENT. BINARY CONSTRAINT SYSTEMS SEEM TO BE A STRONG CANDIDATE FOR SUPPLEMENTING THIS OR OTHER EVIDENCE ACCRUAL SCHEMES. THE PROPOSED EFFORT WILL INVESTIGATE AND DEVELOP TECHNIQUES FOR PARAMETER ESTIMATION AND EVIDENCE ACCRUAL USING BINARY CONSTRAINT SYSTEMS. THIS DEVELOPMENT WILL TAKE PLACE ON THE SARES TESTBED.