Neural Implicit Monocular Visual SLAM for 3D Reconstruction in Planetary Environments

Abstract. The application of SLAM technology in planetary environments has become a research frontier for autonomous rovers. Existing visual SLAM methods often exhibit low accuracy in pose estimation and reconstruction due to poor feature extraction and mismatched correspondences. This paper introduces a novel strategy that integrates neural implicit networks within a visual SLAM framework. By jointly optimizing camera poses and implicit scene representations using neural radiance fields, we achieve high-precision visual localization in the Mars scene without requiring loop closure. We validate our method using data from NASA’s Perseverance rover and compare its performance with OV²SLAM. The results demonstrate that our method significantly outperforms OV²SLAM in localization accuracy, achieving an 85.16% reduction in absolute trajectory errors and maintaining translation errors within 1 m across the entire trajectory. Moreover, our framework delivers compelling novel view synthesis despite sparse inputs and a fixed forward-facing viewpoint. The 3D point cloud models, synthesized from estimated depth maps and poses, further highlight the feasibility and effectiveness of our method for reconstruction in planetary environments.