Cross-view localization in GNSS-denied environments

Abstract

Accurate localization in GNSS-denied environments is a critical challenge for autonomous systems. This thesis addresses this problem by developing and evaluating a low-cost cross-view geolocalization system that relies solely on a monocular camera and an IMU, focusing on the correction of unbounded drift from dead reckoning. A probabilistic framework based on a particle filter is proposed to fuse a kinematic motion model with visual observations. The core contribution is a novel "Six-Cone Directional Model" which serves as the measurement model by translating sparse, real-time semantic object detections (’Tree’, ’Building’) from a YOLOv8-based perception pipeline into a robust likelihood score against a pre-built semantic grid map.

A comprehensive evaluation demonstrates that the system successfully mitigates IMU drift under ideal conditions, producing a statistically consistent final position estimate with a 6.40 m error. Subsequent stress tests successfully identified the system’s operational boundaries, revealing that performance is critically dependent on the quality and density of visual observations, with failures occurring due to path ambiguity in night conditions. Furthermore, generalization tests in new geographic locations (Kiruna, Sweden and Çeşme, Türkiye) revealed the perception front-end to be the primary limiting factor. The analysis diagnosed key failure modes, including a domain gap, where the detector failed to recognize different-looking objects, and a fundamental semantic mismatch, where the pre-defined object classes were not relevant to the new environment. This work validates the viability of a sparse semantic approach for localization and provides a rigorous analysis of its real-world limitations, underscoring the importance of an adaptable perception system for true generalizability.