Toward Robust Cross-View Vehicle Localization in Complex Urban Environments

Abstract

In urban environments, global navigation satellite system (GNSS)-based localization can be unreliable due to signal occlusion and multipath interference. As a result, cross-view image-based geo-localization has attracted increasing interest for its ability to operate without GNSS. However, dense buildings, occlusions, and dynamic objects make it difficult to reliably match ground-view and aerial images in urban areas. To address this, we propose a cross-view vehicle localization approach that uses road structures as stable and informative anchors. First, we transform ground-level images into the bird’s-eye view (BEV) domain to reduce viewpoint disparities and achieve consistent spatial alignment with satellite imagery. Next, we introduce a multidimensional feature matching module that captures deep structural information of road elements, including their continuity and geometric characteristics, by jointly integrating semantic, topological, and morphological cues. Moreover, a center-focused attention mechanism is employed to prioritize the central region of the image, improving alignment accuracy and suppressing background noise. The experiments on the nuScenes and Argoverse datasets demonstrate that our method consistently outperforms existing approaches across diverse urban scenes and spatial sampling conditions, highlighting its effectiveness and robustness in real-world geo-localization scenarios.