Previous Abstract Return to Session E9

Session E9: PNT for Autonomous Systems

Multi-Camera VBN for Autonomous Low-Altitude GPS-Denied Missions
Mallory Nation, Zach Barbeau, Velislav Stamenov, Matthew Castleberry, Bryce Karlins, and Kevin Betts, Leidos, Inc.
Location: Ballroom B
Alternate Number 2

Low-altitude trajectories pose a challenge to vision-based navigation (VBN) map matching because there are rarely enough unique features in the camera’s field of view to match with reference imagery. Leidos developed a modular, hardware-and-software solution that leverages information from multiple cameras to enable map matching at very low altitudes. In this multi-camera approach, features are extracted from each image, aggregated into a single frame, refined with a geometric consistency check, and then used as measurements in the Leidos filter. Leidos designed and manufactured a standalone, multi-camera VBN hardware suite to enable test and integration on a variety of aircraft. During a live event that included ten flights at altitudes under 75 meters (250 feet) AGL, Leidos showcased stable and precise positioning while GPS-denied, consistently achieving 3D position errors of under 20 meters with accurate uncertainty bounding.

Background and Objectives:
Leidos has a long legacy of using its georegistration software to precisely locate aerial vehicles while GPS-denied. The core component is “map matching,” which involves projecting a live camera image to the ground, querying a reference image from a georeferenced database, and matching features between the two images. The matched features undergo multiple validation steps to remove outliers and ensure accuracy. A critical step is the geometric consistency check, which removes outliers with an iterative affine transform. Map matching is reliable when there is sufficient content in the observed ground area. When flying at low altitudes, such as below 75 meters (250 feet), the ground content is limited; it is common to find only a few features and therefore be unable to return a measurement to the filter.

Methodology and Key Innovations:
The Leidos team hypothesized that VBN accuracy at low altitudes could be significantly enhanced by combining features from multiple cameras. While a single image may lack sufficient features, multiple cameras that captured different perspectives simultaneously would provide the necessary feature density. The geometric consistency check could then be run on the aggregated feature set.
Leidos georegistration software was modified in several major ways. First, the core georegistration application was updated to accept multiple input camera streams. Next, the georegistration pipeline was modified to run a subset of its standard checks on each camera stream, returning a larger number of partially-validated features. Then, a top-level feature aggregation step was added that transforms the matched features from each camera into a unified coordinate frame and runs the geometric consistency check on the aggregated features to remove outliers via the iterative affine transform solver. Finally, if there are sufficient inliers present after all checks, the features are used to generate measurements for the Leidos filter. Leidos’ rapid wide area search (RWAS) capability, which is used to recover from long periods of drift caused by low content terrain, such as flying over water or barren fields, was modified to alternate between camera streams in search of a recovery.
Alongside the software updates, a standalone, multi-camera VBN hardware suite was developed to be easily integrated on a variety of aircraft. The hardware suite consists of a mission computer box and a camera pod. The former houses a Jetson Orin NX mission computer, Lepton carrier board, Iridium STL receiver, and GPS receiver. The latter houses a Microstrain CV7 IMU and four FLIR Boson 320 cameras with 92-degree horizontal field of view lenses, each pointing up 35 degrees from the pod’s vertical axis and offset 90 degrees from each other around the vertical axis to maximize the viewed ground area. The Boson cameras are externally triggered so that they capture images simultaneously. The mission computer is designed to be internally mounted in the fuselage and the camera pod to be Picatinny rail mounted underneath the fuselage or wing, so that the hardware is easily integrated with standard systems.

Results, Conclusions, and Significance:
Multi-camera georegistration was demonstrated at an event in November of 2025 where the VBN box was mounted on a Group 1 uncrewed fixed wing platform. A total of ten flights were conducted in representative environments at altitudes under 75 meters AGL. GPS was used to initialize the system, cut once the aircraft was in the air, and re-enabled for landing. While GPS-denied, multi-camera VBN consistently achieved 3D position errors below 20 meters with well-bounded two sigma uncertainties. On several of the flights, georegistration measurements were artificially blocked from going into the filter for several minutes to induce drift, and the system successfully recovered using RWAS. Additionally, two of these flights were closed-loop, where the VBN solution fed to the platform autopilot; these flights were stable and achieved below 20 meters RMSE 3D position error.
Leidos’ modular, hardware-and-software, multi-camera VBN solution has shown precise, reliable navigation performance at altitudes below 75 meters (250 feet) AGL. This technology will enable safe, low-altitude, autonomous missions in support of key objectives for national defense.

Previous Abstract Return to Session E9