Triple Loss based Satellite Image Localisation for Aerial Platforms

Abstract

We present a vision-based technique for aerial platform localisation using satellite imagery. Our approach applies a modified VGG16 network in conjunction with a triplet loss to encode aerial views as discriminative scene embeddings. The platform is localised by comparing the encodding of its current view with a database of pre-encoded embeddings using a cosine similarity metric. Recent image-based localisation research has shown potential for such learned embeddings, however, to ensure reliable matching they require dense sampling of views of the environment, thereby limiting their operational area. In contrast, the combination of our proposed architecture in conjunction with the triplet loss shows robustness over greater spatial shifts, reducing the need for dense sampling. We demonstrate these improvements through comparison with a state-of-the-art approach using simulated ground truth sequences derived from a real-world satellite dataset covering a 1.5km × 1km region in Karslruhe.