CraterSense: Efficient and Robust Crater-Based Lunar Localization System With Structure-Aware Sparse Transformer

Crater-based localization, which estimates spacecraft position by matching observed lunar craters with global catalogs, is critical for autonomous lunar surface navigation. However, traditional geometric matching approaches are highly sensitive to detection noise and computationally intensive, limiting their suitability for real-time onboard applications, particularly in lost-in-space (LIS) scenarios where no prior pose estimate is available. We propose CraterSense, a lightweight and robust localization system that integrates a multiscale crater detection module with a novel identification framework. By abstracting crater layouts into sparse structural images, visual representations that encode only the centers and diameters of the craters, CraterSense reformulates crater identification as a classification task. This abstraction fundamentally improves robustness to detection errors and allows efficient joint optimization of detection and identification without reliance on prior pose information. To effectively process these texture-free and geometrically sparse inputs, we introduce CATSformer, a compact Vision Transformer that combines convolutional stems, token selection, central attention, and structural position embedding. Experiments on two representative lunar regions demonstrate that CraterSense achieves up to 98.2% identification accuracy, strong robustness to perturbations, and low-latency inference, supporting real-time deployment in planetary missions. This work provides a structure-aware solution for crater-based localization, achieving high precision and onboard autonomy without prior pose information.