Enhancing Visual Ego-Localisation through Cross-View Image Registration

Abstract

In the dynamic landscape of modern transportation and navigation, the need for a robust ego-localization solution has become increasingly critical. This precision is essential not only for navigation efficiency but also for ensuring the safety and reliability of automated systems. Conventional GPS-based solutions often fall short, particularly in urban environments where signal obstructions and multipath effects can significantly compromise accuracy. This project addresses the growing demand for resilient ego-localization solutions designed to improve spatial awareness and navigation accuracy. The initiative involves the preparation of a new dataset, outlining our methodology. Our ego-localization approach include ground-view image transformation via a Bird's-Eye-View Mapping module, the use of rough GPS coordinates for accessing aerial imagery, integration of deep learning-based keypoint matching framework, and final image registration. In testing various components of our architecture, we report a mean improvement of 57.08% and a median improvement of 73.10% in localization accuracy, compared to GPS-based localization.