Aerial-DEM geolocalization for GPS-denied UAS navigation

Abstract

Accelerated by the proliferation of small, affordable, and lightweight electronically scanning radar systems as well as advances in Unmanned Aircraft System (UAS) technology, Geo-Registered Radar Returns data are becoming an incredible source for geolocalization in GPS-denied UAS navigation. Most existing approaches match aerial images to pre-stored Digital Elevation Models (DEMs) through 3D terrain reconstruction or GPU-based terrain rendering techniques. However, these reconstruction or rendering processes are themselves error-prone and time-consuming, which further decrease UAS navigation accuracy. In this work, we propose a novel geolocalization approach by directly matching aerial images to DEMs. Inspired by success of deep learning in face recognition/verification, we develop a triplet-ranking network to embed aerial images and DEMs into the same low-dimensional feature space, where matching Aerial-DEM are near one another and mismatched Aerial-DEM are far apart. To create large-scale training dataset, we design an efficient terrain generation approach using per-pixel displacement mapping technique. This approach augments aerial datasets by simulating visual appearances of terrain under different lighting conditions. Experiments are conducted to show the effectiveness of our deep network in finding matches between aerial images and DEMs.