SBIR-STTR Award

Computational Intelligence Approaches to Automatic Target Recognition
Award last edited on: 11/27/2002

Sponsored Program
SBIR
Awarding Agency
DOD : DARPA
Total Award Amount
$846,659
Award Phase
2
Solicitation Topic Code
SB961-039
Principal Investigator
Ronald Patton

Company Information

I-Math Associates Inc

12151 Science Drive Suite 102
Orlando, FL 32826
   (407) 737-8422
   N/A
   www.imath.com
Location: Single
Congr. District: 07
County: Orange

Phase I

Contract Number: DAAH01-96-C-R164
Start Date: 4/30/1996    Completed: 12/6/1996
Phase I year
1996
Phase I Amount
$99,000
The Computationally Intelligent ATR Architecture (CIAA) for SAR imagery will be driven by both point (scatter) and region features. I-MATH has previously developed point extraction algorithms. SAIC will suggest various region measurements, again drawing upon the current results of the MSTAR program. All of the raw features, both point and region associated, including the (x,y) hash point coordinates, will be input into the System Dynamics "e" Genetic Algorithm program, so as to produce two sets of evolved features. The first set will be tested with fuzzy logic rules (also evolved by the Genetic Algorithm) to generate an index to the Geometric Hashing. The primary index will be target pose (3D orientation) and possibly some estimates of target class. This will decisively reduce the size of the search space for the geometric hashing; hence a much smaller group of models will need to be matched. Currently, the hash matching determines which point model is closest to the point representation of the unknown live image. A number of metrics are computed, including percent of live and model points matched, live and model basis pair distances, average point mismatch distance, and roation angle dissimilarity. These metrics are applied with ad hoc thresholds and rules to determine the match. For the CIAA, we would instead use a second fuzzy logic rule set, again derived by the Genetic Algorithms, to compute more efficient match decision metrics. The CIAA represents an innovative structure for both improving ATR performance and simplifying the model search strategy. Notwithstanding this innovation, its components are all proven algorithms with which we are developing other successful ATR systems. In Phase I, we will provide detailed estimates of the CIAA's performance by benchmarking it to the geometric hashing-only algorithm. In Phase II, we will extend the CIAA evaluation by comparing its performance to other SAR ATR's, to especially include the MSTAR algorithm suite.

Phase II

Contract Number: DAAH01-97-C-R103
Start Date: 11/4/1996    Completed: 5/4/1997
Phase II year
1997
Phase II Amount
$747,659
For the past eight years, I-MATH has been a significant contributor to the development of model-based ATR systems, beginning with a SAR imagery application for the Wright Labs Target Recognition Branch. Geometric hashing algorithms and associated point model data bases of aircraft and ground vehicles were produced, with the technology transitioned to the ARAGTAP SAR ATR system. Our current Phase I DARPA SBIR for Computationally Intelligent Approaches to Automatic Target Recognition (CIAATR) innovatively addresses the problems of creating and using large data bases for model-based ATR, whereby evolutionary (genetic-like) algorithms are applied to the indexing and matching components of our geometric hashing algorithms. We are specifically employing the System Dynamics International e library, since it has proven successful on two other programs for improving a MMW radar ATR detection algorithm. These algorithms have been demonstrated using ADTS imagery of two target types in many different poses. In the current Phase I DARPA SBIR, have also shown the feasibility of a Laplacian pyramid fusion technique for combining individual polarmetric images without loss of spatial resolution but instead with enhanced target and shadow detail. In Phase II, we will further evaluate this algorithm, including combining it with a polarimetric whitening filter to reduce image speckle. In Phase II, we anticipate demonstrating our algorithms against the MSTAR imagery. During the first year of Phase II, we plan to work with six target types, and then expand to the full twelve target set in the second year. Except for the e library all of our algorithms and associated source code, and without any proprietary rights, will be shared with the entire MSTAR community. Moreover, e can be licensed for a modest price.