Efficiently Computing and/or Compensating for Object Variability in Automatic Target Recognition (ATR) Applications

There are numerous targets of interest to the Air Force that may be manifested in many different forms. For example, many ground targets may be configured with or without particular subcomponents, and such subcomponents may also be situated in different locations on a given target. Moreover, the same target may appear differently to a radar as a function of specific articulations even when the same components are present. Target variability poses a significant challenge to current ATR systems, which may be addressed from two different perspectives: (i) the ATR algorithm may be designed to address such target variation by diminishing the importance of target details that are susceptible to change, or (ii) by yielding a database of sufficient diversity such that a trained algorithm is capable of addressing the variations in actual data. In the proposed Phase I research, SI will investigate both (i) and (ii), examining their relative merits. The issues associated with choosing between these two options, or perhaps merging components of each, motivate the research proposed here for Phase I. In Phase II we will pursue the appropriate balance of these two approaches based on AFRL input and results from Phase I.