You are here

Transfer Learning For Video Analysis of Infrared Datasets

Description:

TECHNOLOGY AREA(S): Electronics 

OBJECTIVE: Develop and demonstrate techniques for machine learning and Aided Target Recognition (AiTR) on infrared (IR) and other limited size datasets from ground-to-ground sensors. 

DESCRIPTION: Great progress has been made in recent years in the analysis of visible band imagery and full motion video—specifically the ability to detect and classify objects of interest in imagery and video. Progress has also been achieved in the classification of human activities in visible band full motion video. However, this recent success has been strongly reliant on massive amounts of training data. The Army has clear interest in the ability to analyze and discover threats in video and imagery, but military applications are typically lacking in the amount of data (including data of militarily significant target types) to properly implement modern learning techniques such as deep learning. This deficiency of data is even more apparent in the IR domain. What is needed is a set of techniques and algorithms which can exploit highly trained and effective learning models from large visible and Civilian datasets (or artificially constructed datasets) by transferring the models to perform detection and analysis on similar but smaller militarily significant IR and other data. This technique is generally called transfer learning. In addition, effective transfer learning would enable the rapid adjustment of trained IR models to new target types and environments (“learning on the fly”). This effort aims at overcoming limitations listed above and making IR AiTR an effective fieldable system. This effort directly supports Army Modernization Priority: Next Generation Combat Vehicle (NGCV)—benefitting the automation associated with the NGCV through improved algorithm performance. This effort will enable NGCV sensors to rapidly determine external threats and alleviate operator fatigue via automation of surveillance and navigational functions. 

PHASE I: Show proof-of-concept for transfer learning algorithms for target and threat detection in full motion IR video and IR imagery. Show proof-of-concept for algorithms to greatly increase classification effectiveness (high probability of correct classification with minimal false alarms). Integrate algorithms into a comprehensive algorithm suite. Test algorithms on existing data. Demonstrate feasibility of technique in IR video sequences. Distribute demonstration code to Government for independent verification. Successful testing at the end of Phase 1 must show a level of algorithmic achievement such that potential Phase 2 development demands few fundamental breakthroughs but would be a natural continuation and development of Phase 1 activity. 

PHASE II: Complete primary algorithmic development. Complete implementation of algorithms. Test completed algorithms on government controlled data. System must achieve 90% classification rate with less than 5% false alarms. Principle deliverables are the algorithms. Documented algorithms will be fully deliverable to government in order to demonstrate and further test system capability. Successful testing at end of Phase 2 must show level of algorithmic achievement such that potential Phase 3 algorithmic development demands no major breakthroughs but would be a natural continuation and development of Phase 2 activity. 

PHASE III: The topic enables the Army’s Next Generation Combat Vehicle modernization priority, addresses PEO IEW&S and PEO GCS needs, and supports technology development efforts occurring in 63710/K70 which will become 633462/BG1. Complete final algorithmic development. Complete final software system implementation of algorithms. Test completed algorithms on government controlled data. System must achieve 90% classification rate with less than 5% false alarms. Documented algorithms (along with system software) will be fully deliverable to government in order to demonstrate and further test system capability. Applications of the system will be in NVESD Multi-Function Display Program, vehicle navigation packages, and AiTR systems. Civilian applications will be in night surveillance, crowd monitoring, navigation aids, and devices requiring rapid adaptation to new environments. 

REFERENCES: 

1: M. W. Berry, M. Browne, A. N. Langville, V. P. Pauca, and R. J. Plemmons, "Algorithms and Applications for Approximate Nonnegative Matrix Factorization,"

2:  Computational Statistics and Data Analysis, vol. 52,no. 1, pp.155-173,Sep. 2006

3:  Machine Learning: Proceedings of the International Workshop (ML92)

4:  Held in Aberdeen, Scotland, on 1-3 July 1992 ADA255362)

KEYWORDS: Deep Learning, Aided Target Recognition, Transfer Learning, Neural Networks, Infrared Video 

US Flag An Official Website of the United States Government