Edge Feature and Optical Flow Terrain Aid for GNSS-Denied Airborne Visual Navigation

Conventional autonomous aircraft navigation relies on the use of a Global Navigation Satellite System (GNSS). This risks system disorientation due to GNSS failure from atmospheric events, jamming or satellite destruction. Other systems which do not require GNSS predominantly rely on active sensors, preventing stealth operations. They also add large power, cost and weight penalties to the airframe. In contrast, human pilots are capable of navigation without active sensors or GNSS, by Visual Flight Rules. Visually distinct landmarks are identified and tracked to localise the vehicle. However, a human pilot adds substantial vehicle design constraints and limits mission and performance characteristics. This thesis outlines the development and implementation of a computer vision based autonomous navigation system. The emulation of human visual navigation techniques allow localisation from curve and surface based landmarks. Simultaneous Localisation and Mapping (SLAM) is employed in six degrees of freedom, using a spline model of edge features. Image segmentation algorithms are designed to determine visually distinct and recognisable water edge features in aerial video frames. Google Earth imagery is condensed into a spline feature database via image segmentation, with curvature based data associations used to provide a visual terrain aided navigation solution. Digital terrain elevation maps are employed to account for changing river and dam water levels, or alternatively the estimation of this information if unavailable. Visual terrain aided navigation is also demonstrated through a passive optical-flow based terrain contour matching system. The outlined visual SLAM system is shown to greatly enhance navigation accuracy in unknown environments, using real flight test data. Contour based terrain associations are shown to further refine this accuracy. Terrain assistance with curve based edge associations shows localisation accuracies of under 20 metres, in real-time