Terrain relative navigation based on deep feature template matching and visual odometry

Abstract

The main hurdle for terrain relative navigation systems is the incongruity of visual features between a patch of a satellite reference map and a view from an onboard UAV camera. Images are taken during different time of year, under different weather, vegetation and lighting conditions, with different angles of observation. This work proposes the usage of deep feature template matching, where features are extracted during unsupervised training using a triplet loss. It provides semantic understanding, agnostic to terrain transformations. In order to overcome struggling to navigate over featureless terrains, the work proposes additional usage of visual odometry with the procedure of sticking to the map after encountering enough features, with the procedure of hypothesizing over possible locations. Passing a fragment of the reference map through the trained feature extractor, applying an entropy filter and then a path-finding algorithm allows planning a flying path over areas rich of features relevant for navigation.