Meridian: Metric-Semantic Primitive Matching for Cross-View Geo-Localization Beyond Urban Environments

2026-06-04

AI Summary

This work addresses the challenge of low global localization accuracy and poor generalization in unstructured natural environments devoid of GNSS signals by proposing a cross-view geolocalization method that requires no region-specific training. The approach matches metric-semantic primitives between aerial imagery and ground-level RGB-D data, employing a consistency-based scoring mechanism to generate robust pose hypotheses and reject outlier correspondences. These refined matches are integrated within a pose graph optimization framework to yield high-precision trajectory estimates. Evaluated across a 19-kilometer route encompassing urban areas, parks, campuses, and wilderness, the system achieves an average localization error of 2.4 meters, demonstrating significantly enhanced robustness and generalization in scenes with repetitive geometry or featureless terrain.

Abstract

Successful robot automation requires accurate global localization to support repeatability, task planning, goal specification, and safe operation. However, reliable localization in GNSS-denied environments remains an open problem. Overhead aerial imagery offers a promising solution, but existing approaches primarily target structured urban environments and have been rarely demonstrated in unstructured natural terrain. Limitations of the state-of-the-art include a reliance on models trained for specific environments, as well as difficulty handling repetitive geometries and featureless landscapes commonly found in natural outdoor areas. To overcome these challenges, we present Meridian, a method for matching high-level metric-semantic primitives across aerial images and ground robot RGB-D camera data that achieves accurate global localization and generalizes well across diverse environments, all without any training or algorithmic fine-tuning on area-specific data. We formulate novel consistency metrics to estimate a distribution over robot submap poses and to reject outlier hypotheses in a robust pose graph optimization step for accurate robot trajectory estimation. We demonstrate that our algorithm can localize a ground robot across a wide variety of environments, including an autonomous driving dataset, a park and campus area, and a wilderness camp, with an average optimized trajectory error of 2.4 m over 19 km of ground traversal.

Problem

• cross-view geo-localization
• GNSS-denied environments
• natural terrain
• global localization
• aerial-to-ground matching

Innovation

• metric-semantic primitives
• cross-view geo-localization
• GNSS-denied localization
• pose graph optimization
• environment-agnostic