Visual localization of a Rover with limited visual observations

This thesis investigates visual localisation for planetary rover navigation in Mars-like environments. It focuses on geolocation and evaluates how modern deep learning-based feature matching influences pose graph optimisation (PGO) under challenging conditions such as low texture, repetitive terrain, strong illumination changes, and strict computational constraints. These challenges require combining visual perception with inertial motion and wheel odometry to mitigate drift in visual estimates. In this framework, feature-based visual constraints are extracted from a limited set of stereo image pairs, so each patchwise visual update is sparse yet highly informative for estimating the global trajectory. Visual constraints are fused with inertial motion estimates within a PGO formulation, where robot poses are modeled as nodes and odometry, IMU, and vision measurements as edges, yielding an effective optimisation pipeline. The thesis also analyzes how feature matching errors, such as too few inliers, poor spatial coverage, or high outlier rates, propagate through the back-end, degrading error accumulation, loop closure consistency, and overall optimisation stability. To this end, classical handcrafted methods such as SIFT and ASIFT are compared against a deep learning-based front-end using SuperPoint and SuperGlue, highlighting the trade-off between more robust correspondences and increased computational cost. The proposed method is tested both on stereo image data acquired with the ExoMars testing rover (ExoTeR) in the ALTEC Mars Terrain Simulator (MTS), a representative Martian- analogue environment, and on the DLR Morocco dataset, a desert Mars-analogue dataset used to validate the approach and study generalization across trajectories and environmental conditions. To address domain shift, the work investigates training and fine-tuning SuperPoint and SuperGlue on Mars-like imagery from the Morocco dataset, and evaluates the resulting changes in matching performance and pose estimation. Finally, the thesis outlines its main limitations and suggests future research directions, including more efficient learned architectures for resource-constrained platforms, improved constraint-selection strategies, more robust outlier rejection, and tighter integration of visual matching with additional sensing modalities.