Localization of unmanned aerial vehicles using Terrain classification from aerial images

Abstract

In this paper we investigate the benefit of terrain classification for selflocalization of a flying robot. The key idea is to use aerial images, which are already available from online databases such as GoogleMaps™, as reference map and to match images taken with a downward looking camera with this map. Using different terrain classes as features, we can make sure that our method is invariant to lighting/weather changes as well as seasonal variations or minor changes in the environment. A particle filter is used to register the query image with parts of the map. The proposed method has shown to work on image data from both simulated and real flights.