A deep learning-based two-stage feature matching method for small celestial body 3D shape reconstruction

Abstract

The exploration of small celestial bodies (SCBs) represents a critical frontier in deep space research. Given the weak gravitational fields characteristic of SCBs, the ‘touch-and-go’ approach has been widely adopted in recent missions. This method necessitates the development of high-resolution models of these celestial bodies. However, the significant photometric variations encountered in deep space pose substantial challenges for feature matching between images, thereby limiting the applicability of stereo photogrammetry (SPG) techniques for model construction. To address these challenges, this study proposes a novel stereo photogrammetry (SPG) method incorporating an efficient ‘two-stage’ feature matching algorithm designed to handle images with significant lighting variations. By leveraging deep learning networks and DBSCAN clustering for feature matching, the method achieves robust matching outcomes. Depth information is subsequently reconstructed using structure from motion (SfM) techniques, eliminating the need for external camera parameters. Furthermore, the proposed method enables the construction of a more accurate Bennu model, closely approximating Lidar-based models, while requiring fewer images.