Visual Localisation for Autonomous Operation in GNSS-Denied Environments

Abstract:

Autonomous platforms operating without GNSS require an alternative source of absolute position. Cross-View Geo-Localisation (CVGL) addresses this by matching ground-level imagery against geo-referenced satellite tiles, but prevailing methods are limited by the viewpoint disparity between the two domains, an absence of explicit geo-spatial structure, precision ceilings imposed by sparse reference sampling, and computational demands incompatible with mobile deployment. This thesis presents four progressive contributions that address these limitations in turn, advancing CVGL towards a deployable metric localisation capability. The first, BEV-CV, introduces Bird’s-Eye-View transforms to CVGL, projecting ground-level features into a top-down space for direct comparison with aerial embeddings. On CVUSA and CVACT with limited-FOV road-aligned queries, BEV-CV improves orientation-aware Top-1 recall by 23% and 24% on 70◦ crops, whilst reducing embedding dimensionality by 33%. Building on this shift towards spatial structure, SpaGBOL reformulates CVGL as matching over a graph of road junctions, released alongside a novel ten-city dataset. A GNN exploits spatial correlations along graph walks, and Bearing Vector Matching filters retrievals by junction road arrangement, yielding a relative Top-1 improvement of up to 49.86% over the strongest prior baseline. To move from graph-level retrieval to metric localisation, PEnG combines city-scale CVGL with relative pose estimation from dense 3D correspondences to recover a precise 3-DoF pose. Over 31.6 km2 of Manhattan, PEnG reduces median error from 734 m to 22.77 m, a 96.9% reduction. The final contribution, TACO, reframes CVGL as one measurement source within a sensor fusion framework. A fine-grained CVGL localiser provides intermittent absolute fixes, fused with continuous IMU dead-reckoning through an Unscented Kalman Filter with adaptive triggering and confidence-weighted measurement noise. Using only a monocular camera and inertial sensor, anchored by a single GNSS reading at start-up, TACO reduces median Absolute Trajectory Error on the KITTI representative-sequence group from 97.0 m (IMU-only) to 16.3 m, a 5.9× reduction. Together, these contributions reduce domain disparity, exploit urban spatial structure, achieve metric precision via fused retrieval and pose estimation, and deliver drift-bounded localisation suitable for deployment.