This thesis describes a new approach for a vision-based positioning system for Un-manned Aerial Vehicles using a recognition method based on known, robust geo-graphic landmarks. Landmarks are used to calculate a position estimate in a globalcoordinate frame without requiring external signals, such as GPS. Absolute systemsare of interest as they provide a redundant positioning system, allow UAVs to oper-ate when GPS-denied and can enable high-precision landings for spacecraft.The core challenge with vision-based absolute positioning is recognition of land-marks. Most abundant landmarks, such as buildings, are visually similar and dif- cult to distinguish. Previous research in the area tends to focus on matching rawaerial image data to a set of reference images. While these methods can achieveacceptable results in speci c conditions, they struggle with variations in lighting,seasonal changes and changing environments. This thesis presents a new multi-stage method that aims to solve this using a high-level matching framework wherelandmarks identi ed in an aerial image are matched to a reference database.This has led to the development of a geometric feature descriptor that encodes thetopography of landmarks. The proposed system therefore matches the arrangementof features rather than the appearance, which lets it distinguish individual landmarksin large sets (20,000+ features). Since the arrangement of landmarks often is semi-structured and ambiguous, in particular when considering man-made landmarks,a matching stage has been developed that uses a number of strategies to enablematching of individual landmarks to a full database.The results have been evaluated for two conceptual vehicles with acceptable results,highlighting the strengths of the proposed system as well as areas for improve-ment.
展开▼
