Scalable GPS-Denied Autonomy for Scan Eagle and Other UAV: Machine Learning for Monocam GPS-Denied Vision Autonomy

Rhoman Aerospace is delivering a no-emit, GPS-denied, vision-based UAV navigation solution to USSOCOM. The current solution uses stereo cameras (two cameras separated by a baseline distance) to measure the relative distance to visually-tracked objects during flight. The proposed SBIR adapts this Alt-PNT & navigation solution to work on a single cameraand deploys the solution on a Boeing/InSitu Scan Eagle UAV. The solution builds off of current work, adds cutting edge NVIDIA machine-vision algorithms, and enables the solution to work on a single camera, making the solution much more scalable across DoD UAV and modular ISR/Targeting payloads.The proposed solution uses a novel machine-learning ōgeo-tunedö visual inertial odometry (VIO) system & object recognition algorithms to refine the current stereo-camera based system, to deploy the solution on a mono-camera, greatly increasing the deployability of the solution across DoD UAV.ĀThe solution is built around end user & stakeholder feedback from in person events and is designed to address immediate warfighter needs while enabling the future GPS/Electronic-denied battlefield.Ā The solution allows UAV to perform autonomous missions in GPS denied environments, with no detectable signal (passive cams, no LIDAR), and no preexisting spatial maps, using ōvisual-breadcrumbö feature-navigation. The solution visually detects objects in the surroundings of the UAV, tracks movement of the UAV relative to these objects to keep it moving in the desired direction, and creates feature-based areas maps that provide ōvisual-breadcrumbsö to allow the UAV to always get back to launch. The current solution requires stereo-cameras to measure distances, but the Scan Eagle has a single camera; monocam version will also work on ISR payloads, which adds significant scalability to the solution. In order to adapt the solution to the Scan Eagle, the component that measures distances to detected objects must be adjusted to work with a monocam instead of a stereo-cam (which required two cameras) and the added variance of distance measurements must be corrected. This is accomplished in two ways: Deployment of NVIDIA object detection algorithms allow us to estimate the distance of objects using the known sizes of objects and bounding-box size, and aĀmachine-learning geo-tuned Visual Inertial Odemetry system is added to the solution for increased precision. Ā