Visual Landmark Selection and Recognition for Autonomous Unmanned Aerial Vehicle Navigation An image analysis framework based on local scale-invariant features

Reliable localisation is an essential component of a successful autonomous vehicle system. Using GPS makes the vehicle dependent on an external system. Possible technical and political problems make it impossible to guarantee continuous accessability to this system. Simultaneous Localisation and Map Building (SLAM) is the process of concurrently building a feature based map of the environment and using this map to estimate the locations of the vehicle and the landmark positions. For a SLAM system to function satisfactorily, it needs reliable relative observations of landmarks (features) as input. This thesis describes a framework for selecting, tracking and recognizing landmarks from infrared image data. The purpose of the framework is to provide landmark observations to a SLAM system which is used for navigation of an autonomous unmanned aerial vehicle (UAV). The algorithms used for landmark selection, tracking and recognition are based on a landmark representation consisting of a set of local features, referred to as keypoints. The keypoints are extracted from a scale-space framework and described using local image gradients, making them invariant to scale and rotation respectively. This information makes it possible to individually match new keypoints against a landmark representation database. Furthermore, the keypoints are partially invariant to variations in other parameters, such as illumination or 3D-viewpoint. The landmark representation is continuously updated as new observations of the landmark are made. Results show that selected landmarks can be successfully recognized from a relatively wide range of 3D-viewpoints, relative to the viewpoint from which the landmark has been observed. The method has also proven to be fairly resistant to the presence of additive white Gaussian noise. Visuell igenkänning och urval av landmärken för navigering av autonoma luftburna ygfarkoster Ett bildanalysramverk baserat på lokala skalinvarianta kännetecken