Combat identification with sequential observations, rejection option, and out-of-library targets.

摘要

Combat target identification (CID) is the process by which detected objects are characterized pursuant to military action. Errors in CID such as mis-labeling targets and non-targets carry significant costs. Fusing data from multiple sources and allowing a rejection, or non-declare, option can improve CID error rates. This research extends a mathematical framework that selects the optimal sensor ensemble and fusion method across multiple decision thresholds subject to warfighter constraints. The formulation includes treatment of exemplars from target classes on which the CID system classifiers are not trained (out-of-library classes), and it enables the warfighter to optimize a CID system without explicit enumeration of classifier error costs.; A time-series classifier design methodology is developed and applied, resulting in a multi-variate Gaussian hidden Markov model (HMM) with a specially constructed hidden state space. The extended CID framework is used to compete the HMM-based CID system against a template-based CID system. The assessment uses a real world synthetic aperture radar (SAR) data collection comprised of ten in-library target classes and five out-of-library target classes. The framework evaluates competing classifier systems that use multiple fusion methods, including neural network fusion and label fusion, varied prior probabilities of targets and non-targets, varied correlation between multiple sensor looks, and varied levels of target pose estimation error. Also, an on-line target pose estimator is developed using principal component analysis of masked target SAR images. This estimator validates experimental assumptions on target pose prior to classification.; The CID system assessment using the extended framework reveals larger feasible operating regions for the HMM-based classifier across experimental settings. In some cases the HMM-based classifier yields a feasible region that is 25% of the threshold operating space versus 1% for the template-based classifier. Similar performance results are obtained for rule-based label fusion and the more complex neural network fusion and are explained by the new ability to independently set classifier thresholds with the label fusion method.