Systems & Technology Research LLC — Department of Defense SBIR Phase I: AF161-133

Description

ABSTRACT: Practical automated target recognition (ATR) capabilities for synthetic aperture radar (SAR) imagery require sustainable and easily updatable target representations as well as generalizability across different SAR sensors and operating conditions.Systems & Technology Research (STR), along with Wright State University, proposes to extend ATR algorithms being developed under the AFRL Compact Automatic Target Recognition and Sustainable Environment (CASE) and DARPA Target Recognition and Adaption in Contested Environments (TRACE) programs to develop a practical SAR ATR system characterized by the following attributes:(1) Rapid target model construction using coarse grain models that capture gross target shape and scattering features.(2) Encapsulation of the scattering peak extraction to produce a sensor agnostic representation of dominant target scattering structure. (3) Robust scattering peak descriptions using pyramid match hashing codes that are learned once per target. (4) Compact target classifier encoding using Fisher vectors. (5) Low computation complexity to match encoded templates against observed data. (6) Rapid updating of target database with new models using shared visual codebooks and variable fidelity model prediction.; BENEFIT: Our SAR ATR approach provides a number of benefits for SAR exploitation systems:(1) It enables streamlined specification of new targets by developing low fidelity physical models of the targets.Target models may be built manually or though automated 3D modeling techniques if multiple photographs of the new target are available.(2) Our approach enables improved ATR performance by increasing the fidelity of the target models, but is designed to work with a broad range of target fidelities.(3) The offline target classifier generation process optimizes discriminability across mission targets of interest.(4) Since our ATR approach is based on characterizing and matching scattering mechanisms of the target at different scales, a single target representation is used to support recognition across different SAR sensors, possibly operating with different data acquisition parameters.(5) Finally, our approach for online ATR processing requires low computation and storage complexity by virtue the compact target encoding techniques.While the SAR ATR capability is targeted for tactical platforms, there are a wide range of intelligence and security applications within the DoD community as well as for homeland security.