Ego-Localization Using Object Matching in Aerial and Car-View Images

Abstract

In the dynamic realm of modern transportation and navigation, the need for a robust ego-localization solution has become increasingly crucial. As advanced driver assistance systems, autonomous vehicles, and smart city infrastructure continue to integrate, accurately determining a vehicles real-time position has emerged as a pivotal challenge. This precision is not only vital for navigation efficiency but is also a cornerstone for ensuring the safety and reliability of automated systems. Conventional GPS-based solutions often fall short, especially in urban settings where signal obstructions and multipath effects can compromise accuracy. This project addresses the escalating demand for resilient ego-localization solutions aimed at improving spatial awareness and navigation accuracy. The initiative involves the creation of an innovative dataset, elucidating its methodology and content, which encompasses data collected from the vehicle (ground-view images, Birds-Eye-View masks, localization) and aerial imagery data paired with semantic annotation. Subsequently, our ego-localization methodology is outlined, encompassing ground view image processing via a Birds-Eye-View Mapping module, the utilization of rough GPS coordinates for accessing aerial imagery, and the extraction of semantic masks using our Aerial Imagery Segmentation Module. The outputs are then employed in a Localization Error Estimation module. In testing the various parts of our architecture, we observe commendable results in the first two components, while the third component exhibits average performance that has a potential of improvement in future works.