You are here

Machine Learning Algorithms for Infrared Search and Track Applications




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 the Announcement. 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 machine learning algorithms that can be implemented on low cost, size, weight and power processing hardware to aid detection and tracking processing for infrared search and track (IRST) applications.


DESCRIPTION: The United States Air Force needs an extended range passive air-to-air surveillance capability for contested environments where solutions based on active emissions and/or radar returns may not be available or are ineffective.   Detection and tracking algorithms have demonstrated great capability when paired with IRST sensors. However, common statistically-based detection and tracking algorithms are computationally expensive and require large compute resources to operate real-time, leading to compromises in execution methodologies and performance. In addition, required processing resources limit the ability to deploy such IRST systems on platforms with stringent cost, size, weight and power (C-SWaP) constraints and/or may significantly reduce platform endurance and associated mission effectiveness. As such, alternative algorithms must be developed with similar, if not improved, performance but requiring significantly less computational resources.   Machine learning (ML) algorithms offer a possible solution to this challenge. ML algorithms have been explored and developed for various detection and tracking applications [1 - 5]. However, they have not specifically been developed for use in IRST applications with very low contrast targets imbedded in diverse background clutter including sensor-induced artifacts. In this application, the targets are unresolved with their signatures and motion characteristics differing significantly from other tracking scenarios. Here unresolved does not mean the system generates single-pixel targets, but rather the spatial shape is dictated by the impulse response of the imaging system and sampling at the focal plane array. In addition, the specific type of sensor implementation for IRST may dictate methodologies employed. Ultimately, an optimal algorithm/processing solution might be a combination of a conventional approaches with ML techniques applied to a specific aspect of the problem.  In order to be effective, robust and generalizable to a variety of environments and different IRST sensor instances, the ML methods should not rely solely on training data collected by the respective sensors themselves. The effort should not be based on blind application of numerous ML methods and evaluating the results. It should instead focus on the entirety of the detection and tracking process and determine where and how ML should be specifically applied. This effort should explore and demonstrate the ability to train the ML algorithm using properly simulated target signatures and clutter plus noise and interference effects, and achieve comparable detection performance to baseline algorithms. In addition, the ML approach should take into account of limitations of truth data for real world IRST data that could be used for the ML training and should be able to overcome this limitation.  For this effort, the government will provide 1) Limited real-world IRST data with relevant targets and truth for testing and validation 2) Modeled target signatures 3) IRST sensor characteristics  It is expected the Offeror will incorporate physical phenomenology, radiometry, and realistic focal plane characteristics within the structure of the algorithm. The Offeror must also demonstrate knowledge of conventional IRST algorithms and processing products in order to understand the problem space.  Offerors must have the ability to process and store classified data up to Secret//Collateral.


PHASE I: Develop a modular machine learning architecture optimized within a detection and track processing framework for IRST. Clearly identify the areas where ML would apply or integrate into a detection and track processing pipeline.


PHASE II: Develop and refine the architecture in described in Phase 1. Demonstrate the ability to train the ML algorithm using a combination of synthetic and real-world data. Apply the ML-enhanced algorithm to real-world government furnished data with relevant targets and associated truth. Compare detection and false track performance of the ML-enhanced algorithm with baseline algorithms. The ML-enhanced algorithm performance should be evaluated against the truth data and should achieve a specified threshold of False Positive rate (or False Alarm) and False Negative rate. Compare computation time/resources via demonstration and/or timing studies of the ML-enhanced algorithm with baseline algorithms.


PHASE III DUAL USE APPLICATIONS: Implement the ML-enhanced IRST detection and tracking algorithm in a ruggedized low SWaP processor to meet the provided platform requirements including Open Mission Systems (OMS), integrate with IRST system(s), and demonstrate performance and capability through mountaintop and/or flight testing.



  1. Hu, Y, Xiao, M, Zhang, and K, Wang, "Aerial Infrared Target Tracking in Complex Background Based on Combined Tracking and Detecting", Mathematical Problems in Engineering, vol. 2019, Article ID 2419579, 17 pages, 2019;
  2. Wang, T, Qin, R, Chen, Y, et al., “A reinforcement learning approach for UAV target searching and tracking.” Multimed Tools Appl, 78, 4347–4364 (2019);
  3. Kim, S “Analysis of small infrared target features and learning-based false detection removal for infrared search and track,” Pattern Analysis and Applications, 17, 883-900 (2014);
  4. Kim, S and Lee, J, “Small Infrared Target Detection by Region-Adaptive Clutter Rejection for Sea-Based Infrared Search and Track,” Sensors (Special issue Detection and Tracking of Targets in Forward-Looking Infrared Imagery), 14(7), 13210-13242 (2014);
  5. Ryu, J and Kim, S, “Small infrared target detection by data-driven proposal and deep learning-based classification,” Proc. SPIE, Vol. 10624, Infrared Technology and Applications XLIV, 106241J (2018).


KEYWORDS: Machine Learning; AI/ML; Infrared Detection and Tracking; Infrared Search and Track; IRST sensor; low SWaP; real-time processing; RT processing; low contrast targets; electro-optical/infrared; EO/IR; passive EO/IR

US Flag An Official Website of the United States Government