Multi Sensor Information Integration and Automatic Understanding
Final technical rept.
SIGNAL INNOVATIONS GROUP INC DURHAM NC
Pagination or Media Count:
This program addresses Automatic Image Understanding and Automatic Integration of Disparate Sources of Information. The techniques are particularly focused on asymmetric warfare, urban warfare, guerrilla warfare, and portbase security, for which automatic integration of disparate sources is particularly important, typically with very limited if any a priori training data. Concerning automatic image understanding, we are principally considering image sequences video. The approaches utilize the new field of semi-supervised learning. Specifically, most existing Automatic Target Recognition ATR approaches are supervised, in the sense that they require an a priori training set of labeled data DL. The set DL is composed of example signatures features and their associated identity label. These data are typically employed to design a classifier, with the hope that the labeled training set DL is well matched statistically to the unlabeled data DU to which the ATR algorithm is applied. Such supervised algorithms are vitiated by the inherent differences in training and testing data DL and DU, respectively found in practice. In addition, in conventional techniques the classifier is applied to each element of DU, one at a time, without accounting for the cumulative contextual information inherent to DU. The semisupervised algorithms employed here ameliorate the limitations of conventional approaches by performing learning based on all available data, both labeled and unlabeled. By explicitly employing DU in design of the classifier, the algorithm automatically accounts for context and for changing sensing conditions. The performance of the semi-supervised classifier is directly related to the features extracted from the imagery and video. LMMFC has done extensive research and testing on target detection algorithms based upon Quadratic Correlation Filtering QCF theory which project the image data onto a subspace which is optimal for discriminating targets versus
- Target Direction, Range and Position Finding