Modular and Scalable Machine Learning Framework for Infrared Search and Track

Infrared Search and Track (IRST) systems play a pivotal role in detecting and monitoring airborne targets by sensing their thermal emissions in the infrared spectrum. Due to their passive sensor capabilities, IRST systems offer a strategic edge for military operations in comparison to radar-based search and track systems, which rely on active emissions for detection and tracking. Despite this clear advantage, the effective implementation and deployment of IRST systems to airborne platforms is challenging. Firstly, one of the challenges in fielding this type of system is the massive computational expense of running common statistically-based detection and tracking algorithms in real-time. Secondly, machine learning methods cannot be directly applied off-the-shelf and trained for IRST applications. Target motion characteristics, complex backgrounds, sensor effects, and other phenomenology of the environment create a unique air scene, where target signatures can be buried in clutter and often hidden to the human eye. This limits the effectiveness of machine learning models as well as human analysts responsible for labeling data for these models. The proposed effort aims to address these challenges by, first, investigating the limitations of existing IRST capabilities, and then, developing machine learning algorithms to integrate into and complement key IRST stages. Also, to address the lack of large, high-quality datasets, a process will be developed for data curation to support the continual enhancement of an ML-augmented IRST solution. The salient aspects of the proposed solution are: (1) a Data Curation module, (2) a Spatiotemporal Feature Extraction module, (3) a Target Detection module, and (4) a Track Processing module. In Phase I, each of these modules will be developed and evaluated independently then integrated into a holistic framework for end-to-end optimization. Feasibility will be demonstrated via case studies of Air Force interest. The Phase II effort will focus on dataset expansion, algorithm optimization, extensive technology validation, and technology insertion into Air Force workflows.