A Visual Localization Method Based on Multi-scale Crater Detection and Cluster Matching
Siyuan Li, Jianbin Huang, Tao Li, Shuo Zhang, Jiawei Ren, Jiaxuan Wu, Yuntao He · 2024
Visual localization is a critical component of lunar exploration, it helps avoid hazardous regions on the Moon's complex surface. This paper proposes a novel method based on multi-scale crater detection and cluster matching to address the limitations of existing visual localization methods in terms of real-time performance and accuracy. The method uses the Hough transform to detect large-scale craters in simulated images, calculates the center offsets between the detected large-scale craters and those in the candidate crater list, and applies K-means clustering to group the center offsets, forming large-scale matched pairs. These large-scale matched pairs are then used for pose estimation, achieving rapid coarse localization and obtain the coarse localization error. Subsequently, principal component analysis (PCA) is employed to identify small-scale craters, similarly clustered and matched using K-means. Finally, estimate the pose again and compute the reprojection error for precise localization. Simulation experiments demonstrate that the proposed method achieves a reprojection error of 2.35 pixels (px) and a relative error of only 0.11%, validating the method's effectiveness.