Air-to-Ground Registration of Point Clouds for Urban Scenes Using Dual Gravity Lines

Wide-area 3D data generation in complex urban environments frequently necessitates the integration of multi-source observations acquired from both aerial and terrestrial platforms, such as aerial surveys and mobile mapping systems. A primary challenge arises from the cross-view nature of these datasets, which are captured from substantially divergent perspectives (often differing by 90° or more). The resulting minimal spatial overlap, due to the heterogeneous sensor lines of sight, makes accurate registration highly challenging and typically requires substantial manual intervention. Furthermore, the registration process is further complicated by scale ambiguities inherent in the street-view data (e.g., GPS-denied or GPS-weak). In this work, we present a robust cross-view registration framework based on the assumption that building façade structures remain consistent across aerial and street-view datasets. Building on this assumption, we propose the dual gravity 3D line feature, characterized by two corners of an individual building façade, as a high-level structural representation of façade geometry, which is effective to handle data with minimal spatial overlap and substantial scale variation. A pair of dual gravity lines across cross-view datasets serves as a minimal solver for estimating the similarity transformation, offering a simpler and more efficient solution compared to standard methods that require three-point correspondence. Furthermore, we incorporate a novel scale-invariant similarity metric into the Random Sample Consensus (RANSAC) framework, enabling more reliable estimation even under significant scale differences. We also released a cross-view registration benchmark, named Cross-View Urban (CV-Urban), adopting two public datasets: ISPRS Multiplatform and ULTRRA Challenge datasets, covering three cities with areas totaling 366,388 m2. It is proposed to comprehensively evaluate the registration methods on cross-view scenarios with various occlusion (average 10%), scale variations, and noisy levels. Extensive experiments on the CV-Urban dataset show that our method effectively aligns cross-view datasets, even in cases of extremely limited overlap and substantial scale differences. The results demonstrate notable improvements in both accuracy and robustness compared to existing approaches under challenging conditions. The code and example dataset can be found at https://github.com/GDAOSU/air-to-ground-registration.