Utilizing ML Algorithms to Track and Identify UAS Threats

Unmanned aerial systems (UASs) present a growing threat in the US defense arena as both low-cost intelligence gathering and payload delivery capabilities for adversaries, posing security risks to both military and civilian assets. Detecting, tracking, and identifying UAS threats has been accomplished with conventional imaging and radar techniques, but each technology has limitations. Passive imaging systems have excellent spatial resolution; however, their system complexity and cost are driven by day/night operation requirements that need different sensor systems for solar reflection and/or thermal emission as the primary detection method. Radar based solutions have broad spatial detection coverage but poor resolution for precise target location and identification. Active lidar has the potential to fill the gap where other methodologies fall short by providing day/night operation at high spatial resolution. Many lidar systems exist with performance capabilities dependent on design application. For example, the automotive industry has been using low-SWAP lidar for spatial awareness with machine learning (ML). ML is an excellent method for object identification but relies heavily on man-in-the-loop model training. Variation in real-world target scenes have a strong influence on the performance of these algorithms. Different scenarios present different types of in-scene clutter or detection noise that can negatively impact the reliability of models. It is unclear as to which lidar system would perform best when integrated into a counter UAS system with ML automated target identification. Spectral Sciences Inc.’s (SSI) is a leader in electro-optical performance analysis for remote sensing, measurement simulation, and target detection strategies. In Phase I, we propose design trade studies on different lidar system specifications that will maximize UAS detection range, accuracy, and sensitivity, while identifying a feasible low-SWAP hardware configuration coupled with ML algorithms for a SOCOM counter UAS (C-UAS) system. We will reduce the complexity of the training data for ML algorithms and pre-process measured lidar data to improve ML identification accuracy and precision by utilizing our advanced clutter suppression signal processing techniques. SSI's state of the art radiation transport modeling capabilities will address the well-established need for synthetic data during ML training reducing the risk and time for costly measurements and data labeling while increasing the variety of real-world measurement scenarios needed for a reliable ML based C-UAS system. In Phase II, these methods will be combined to develop a lidar system with automated ML acquisition of UAS profiles, locations, and threat assessments.