Deep learning-based robust optical localization for hypersonic platforms

Abstract Sensor-based localization is required for long-range platforms when GNSS can be denied. To bypass the structural limitations of the classical registration-on-reference-image framework, we offer in this paper to encode the appearance of the surrounding of the target (at all resolutions) from a stack of images of the scene into a deep network. This new framework outperforms the registration baseline in our experiments, in particular on a bimodal scene (which can or can not be snowy). This invites larger benchmarks from academic and industrial community to conclude on the applicability of this method on real use cases.