Autonomous navigation systems rely on inertial navigation systems whose errors accumulate over time, leading to drift in the estimated trajectory. In the absence of GPS, terrain-aided navigation provides an alternative for correcting this drift by registering a synthetic aperture radar (SAR) image acquired by the vehicle with a georeferenced optical image. However, the physical differences between these two modalities, environmental variations, and observation degradations make this registration particularly challenging.
This thesis proposes several contributions aimed at improving the accuracy, robustness, and reliability of multimodal SAR-optical image registration. It first introduces MEOW/Europe, a large-scale, multi-year, and multi-season satellite dataset enabling model training and evaluation. It then presents a comparative study and optimization of deep learning architectures for translational registration, particularly through model hybridization. This work is extended to rigid registration with RoADNet, a novel architecture that jointly estimates translation and rotation from a dense similarity volume. Finally, the integration of conformal prediction enables the construction of sets of plausible positions with statistical guarantees, in order to quantify the reliability of the estimates and identify situations in which they should be rejected. Distillation approaches are also explored to reduce model complexity for onboard deployment.
Together, these contributions support the development of more accurate, robust, and reliable localization methods for autonomous navigation in GPS-denied environments.
