NeRF-based simultaneous pose estimation and 3D reconstruction for non-cooperative space target

Abstract

Accurate pose estimation and 3D reconstruction of Non-Cooperative Space Targets (NCSTs) are critical for proximity operations in active debris removal and on-orbit servicing. In this paper, we propose a novel NeRF-based Simultaneous Pose Estimation and 3D Reconstruction (SPAR) framework to address the challenges of efficiency and reliability in traditional point-based methods. Our framework contains three key components: a multi-resolution hash encoder to reduce computational cost, a 2D keyframe feature enhancement to guide view generalization, and a direct photometric constraint to stabilize pose estimation. Furthermore, the proposed framework is evaluated on a newly constructed Spacecraft Pose Estimation and 3D Reconstruction Dataset (SPARD), comprising both synthetic and real RGB-D images and the experimental results demonstrate its effectiveness. Our framework achieves real-time processing at 51 Hz with a pose estimation accuracy of 1.26 cm translation error and 0.97° rotation error. In 3D reconstruction, the framework updates at a frequency of 32 Hz, and attains a peak signal-to-noise ratio of at least 40 dB for RGB-D images. The results show improvements over traditional and NeRF-based baselines, validating its applicability to space missions with NCST. The source code and dataset are available at https://dazhuang-yang.github.io/.