Lightweight Instance-Level Semantic Dense Three-Dimensional Reconstruction for Satellite Components

Vision-based instance-level semantic dense 3D reconstruction (ISDR) for satellite components can significantly enhance on-orbit service robots' perceptual capabilities for autonomous close-range tasks such as component repairs. However, existing ISDR methods, designed for ground scenarios, struggle with satellite targets' weak, repetitive textures and require GPU support, limiting their use on resource-constrained on-orbit platforms. To address these limitations, this paper proposes a lightweight method that builds on prior instance segmentation research. First, a pose estimation algorithm utilizing ORB and improved EDLine features achieves 40–70% higher tracking success rates on satellite flyby datasets compared to benchmarks. Second, the proposed lightweight dense 3D reconstruction method, optimized by accelerating truncated signed distance function fusion and surface extraction, achieves real-time performance at 23 Hz on a CPU with 5 mm voxel resolution. Third, by leveraging adjacent keyframe information, the instance-level semantic fusion improves efficiency by 77% over Voxblox++ at 5 mm resolution. Finally, the proposed ISDR method is validated on synthetic satellite fly-around datasets, achieving interactive-rate ISDR (10 Hz) on non-GPU platforms.