Deep Feature UAV Localization in Urban Areas and Agricultural Fields and Forests

Sammanfattning: The reliance on GPS for Unmanned Aerial Vehicle (UAV) localization limits the areas of application to places with a stable GPS signal. The emergence of deep learning in computer vision has made deep learning methods for visual UAV navigation a promising candidate for autonomous GPS denied localization. These method locate using images taken by a mounted camera on the UAV. Most works in the field evaluate localization ability in urban environments dense with artificial structures. This thesis analyses the localization ability of one such method over agricultural fields and forests in comparison to urban areas to investigate whether such systems rely on artificial structure or if they can function in a general environment. The localization technique is based on the deep feature Lucas-Kanade algorithm and uses convolutional neural network extracted feature representations of images taken by the UAV and satellite images to place the UAV within the satellite image for a position estimate. A network interpretation method is also applied to the problem to investigate whether it can help explain what causes the potential differences in localization accuracy between the areas. The investigation finds that the localization method is applicable in both forests and agricultural fields and pinpoints other factors than the prevalence of artificial structure that are more important for accurate localization. Further, a potential improvement to the algorithm is proposed that is shown to notably improve localization accuracy in certain conditions. It is based on obtaining a second position estimate by reversing the optimization direction and choosing the better of the two based on a loss function.