Machine Learning Algorithm for Target Detection on the Coastal Battlefield and Reconnaissance (COBRA) System
Navy SBIR 2015.1 - Topic N151-049
NAVSEA - Mr. Dean Putnam - [email protected]
Opens: January 15, 2015 - Closes: February 25, 2015 6:00am ET
N151-049 TITLE: Machine Learning Algorithm for Target Detection on the Coastal Battlefield and Reconnaissance (COBRA) System
TECHNOLOGY AREAS: Sensors, Electronics, Battlespace
ACQUISITION PROGRAM: PMS495, Mine Warfare Program Office, Coastal Battlefield Reconnaissance and
The technology within this topic is restricted under the International Traffic in Arms Regulation (ITAR), 22 CFR Parts 120-130, which controls the export and import of defense-related material and services, including export of sensitive technical data, or the Export Administration Regulation (EAR), 15 CFR Parts 730-774, which controls dual use items. Offerors must disclose any proposed use of foreign nationals (FNs), their country(ies) of origin, the type of visa or work permit possessed, and the statement of work (SOW) tasks intended for accomplishment by the FN(s) in accordance with section 5.4.c.(8) of the solicitation. Offerors are advised foreign nationals proposed to perform on this topic may be restricted due to the technical data under US Export Control Laws.
OBJECTIVE: Develop innovative machine learning methods for improved minefield and obstacle detection and false alarm mitigation in aerial multi-spectral imagery that incorporate ongoing operator-provided decision information and algorithm parameter optimization.
DESCRIPTION: The Coastal Battlefield and Reconnaissance (COBRA) program (Ref 1) is interested in technologies that facilitate automated target recognition (ATR) capabilities in aerial multi-spectral images for previously unseen environments and target types. Targets of interest include minefields and obstacles in various land and marine environments. Typically, ATR algorithms are developed offline (post-mission) using previously acquired test data sets. These algorithms are based on supervised learning methods (Ref 2) that incorporate data from a limited set of test fields. When data is acquired in new environments, the algorithms often must be re-optimized to have good performance in that environment, as well as maintain performance in previously seen environments. The process for performing this offline re-optimization is often costly since it requires the efforts of expert analysts to assimilate data sets, determine target truth, analyze target features, train the ATR classifiers and evaluate performance.
There is a need for innovative methods that can 1) incorporate information from new data sets into the ATR system as they are acquired, and 2) re-optimize ATR algorithms quickly across all known environments, including those of newly acquired data. Online Machine Learning (OML) algorithms (Ref 3-5) can potentially be used to "learn" in the field based on operator-provided results without affecting prior performance. The information collected online can be used to refine the prediction hypothesis (classifier) used in the ATR algorithms. In addition, the information may provide input for automated methods of optimizing ATR performance across all known data sets.
The proposed effort will develop innovative OML algorithms for ATR that can incorporate human operator decisions to optimize probability of detection and probability of false alarm performance in new environments and for new target types. These algorithms will be integrated into mission and post-mission analysis systems in which operators review acquired images. The algorithms will be implemented as object-oriented C++ code for insertion into the operator systems. Development of the online learning algorithms must be combined with identification of how the operator will interact with them to provide updated decision information. Robust optimization of the ATR algorithms may be performed post-mission, which will require the development of separate software tools for processing historical data sets. The OML algorithms and optimization tools developed in this effort will reduce program costs by minimizing the time required for optimizing ATR algorithms to perform well in unseen operational environments.
PHASE I: The small business will develop and demonstrate proof-of-concept online machine learning algorithms against existing Government Furnished Information (GFI) multi-spectral image data sets. Demonstrate the feasibility of using the algorithms to perform online machine learning and to re-optimize ATR performance quickly as new data sets are introduced. Show how these algorithms will be used to improve ATR performance in previously unseen environments beyond the performance of the system optimized on previous data sets. Prepare a plan for incorporating operator decision information into the baseline ATR algorithms.
PHASE II: Further develop and optimize Phase I algorithms and tools and implement them in C++ as object oriented classes. Implement and demonstrate the capability for incorporating operator decision information into the ATR algorithms. Develop a prototype graphical user interface for operator interaction. Demonstrate performance across a broad set of GFI imagery. Performance will be validated with government-provided target truth. The Contractor will prepare a Phase III development plan to transition the technology to Navy use. It is probable that some work under the Phase II will become classified.
PHASE III: If Phase II is successful, the company will be expected to support the Navy in transitioning the technology for Navy use. The company will incorporate the machine learning algorithms and software tools to improve performance of the COBRA Block I and II systems. The company will also support updates to the COBRA Technical Data Package (TDP) to support the Navy in transitioning the design and technology into the COBRA Production baseline for future Navy use.
PRIVATE SECTOR COMMERCIAL POTENTIAL/DUAL-USE APPLICATIONS: The technology developed here can be applied to numerous pattern recognition problems, including surveillance tasks, facial recognition, remote sensing, and Intelligence Preparation of the Operational Environment (IPOE).
2. Caruana, R., Niculescu-Mizil, A. (2006). An empirical comparison of supervised learning algorithms. Proc 23rd Int Conf Mach Learn, 161.
3. Bertsekas, D. P. (2011). "Incremental gradient, subgradient, and proximal methods for convex optimization: a survey." Optimization for Machine Learning, 85.
4. Shalev-Shwartz, S. (2011). "Online learning and online convex optimization." Foundations and Trends in Machine Learning, 4(2), 107-194.
5. Chapelle, O., Scholkopf, B., and Zien, A. Semi-Supervised Learning. MIT Press, 2006.
KEYWORDS: Online machine learning; global optimization; automated target detection; semi-supervised learning algorithms; Coastal Battlefield Reconnaissance and Analysis (COBRA); post mission analysis
Offical DoD SBIR FY-2015.1 Solicitation Site: