Automatic Target Recognition Characterization using Causal Models

This proposal seeks to design a prototype evaluation framework for characterizing the difficulty of Automatic Target Recognition (ATR) Algorithms in Electro-Optical (EO) sensors. Our proposed framework provides the ability to express relationships betweenoperating scenarios and algorithm performance. The ATR operating scenarios are parameterized by Extended Operating Conditions (EOCs), or conditions beyond those used during algorithm training and development. The result of this research will be a set ofqualitative EOCs for four different sensor modalities: LADAR, PMMW, FLIR and CCD. Charles River Analytics has a strong background in developing algorithms for each of these sensors. In addition, multi-sensor datasets have already been collected and existat our disposal for this project.A part of the prototype evaluation framework will allow the construction of causal models. These models will indicate which EOCs are most likely to cause problems for ATR algorithms. Finally, a set of experiments will be designed to assess the correctnessof the developed models. Using ATR algorithms developed under previous contracts, baseline performance on existing datasets will be reported. These experiments will also be used to support the correctness of the evaluation framework and model.The empirical evaluation of Automatic Target Recognition algorithms has direct benefits to many military applications. Spin-off benefits (both government and commercial) of an evaluation framework include tools for automated test case generation, automateddata collection criteria and measured performance evaluation. The potential also exists for the development of data mining applications for algorithm evaluation.