Model-Based ATR: Algorithms Based on Reduced Target Models, Learning and Probing

Two new model based ATR algorithms are presented. The algorithms employ economic target models, recent mathematical developments in the extraction and reduction of target silhouettes from noisy images, and on-line probing for clutter reduction and target classification. The target models used employ a small, carefully selected set of characteristic viewpoints and local features. The design of the algorithms are described and their performance evaluated using synthetic FLIR data. The two algorithms are: the Probing Algorithm and the 4-Way Distance Algorithm. Both algorithms have an early phase which is primarily statistical, and a late phase which is primarily geometric, although the two phases are tightly coupled, with frequent revisitation of the image data used to increase decision confidence. In the Probing Algorithm novel on-line probing is performed using stored target geometric models to efficiently implement a generalized maximum likelihood principle. The Probing Algorithm self-adapts to find the most appropriate target location, range and viewpoint geometry and then performs decision-making based on selected maximum likelihood estimates and on-line computed, optimal probing sets. The 4-Way Distance Algorithm performs decision making based on a classification tree logic utilizing matching scores computed on four components of the target silhouette.