Multi-view Navigation and Hazard Avoidance for Autonomous Aircraft

nbsp;With this SBIR, KEF Robotics will scale its attachable aerial autonomy subsystem to support a wide variety of camera configurations beyond rigid baseline stereo. This subsystem imbues small UAS with the capabilities of GPS-denied navigation and hazard detection and avoidance. With this added scalability, our subsystem will be able to fly on a wider range of aircraft, including fixed-wing aircraft with cameras on the wings and quot;air taxisquot; with cameras distributed over the body of the vehicle. These new capabilities will be critical in the first and last 50 feet of takeoff and landing, especially in emergency scenarios such as GPS or propulsion failure.To enable these features, KEF will explore two objectives: (1) non-rigid sparse visual state estimation and (2) dense stereo-vision under changing extrinsics. Sparse visual state estimation is the current gold-standard for GPS denied SLAM, but currently works primarily on monocular or rigid, parallel-boresight stereo cameras. To achieve this, we will implement online extrinsics calibration methods and multi-camera feature initialization and test them with simulated datasets. Dense stereo-vision is commonly used for hazard detection and avoidance, but again almost exclusively uses rigid, parallel-boresight stereo cameras. KEF will test a number of dense stereo methods under uncertain calibration and develop techniques for arbitrary camera configurations.