Scholarlink

<p>The growing number of space debris and spacecraft has made accurate 3D reconstruction increasingly critical for orbital safety. However, reconstructing non-cooperative space targets (NCTs) remains challenging due to extreme lighting conditions, characterized by the absence of diffuse reflection and dynamic illumination changes during rotation. At present, reconstruction methods based on 2D images are mostly based on good illumination, and there are relatively few researches on extremely poor illumination conditions. We propose a novel 3D Gaussian Splatting (3DGS) framework for handling extreme illumination variations in space target reconstruction. 3DGS effectively addresses varying illumination across regions, but it may produce floaters due to local optimization minima. To address this, we integrate an optimizable decoupled appearance embedding model and implement a lighting mask filtering mechanism that maps 2D lighting masks to 3D point clouds. Experimental verification across multiple datasets shows that compared to state-of-the-art methods, chamfer distance has been reduced by 41.14% and PSNR has been improved by 13.98%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

3D reconstruction of non-cooperative space targets of poor lighting based on 3D gaussian splatting

  • 期刊论文

  • 发表日期：

    2025年5月2日

查看全文
  • Yibin Zhao,
  • Jianjun Yi,
  • Yihan Pan,
  • Liwei Chen

摘要

The growing number of space debris and spacecraft has made accurate 3D reconstruction increasingly critical for orbital safety. However, reconstructing non-cooperative space targets (NCTs) remains challenging due to extreme lighting conditions, characterized by the absence of diffuse reflection and dynamic illumination changes during rotation. At present, reconstruction methods based on 2D images are mostly based on good illumination, and there are relatively few researches on extremely poor illumination conditions. We propose a novel 3D Gaussian Splatting (3DGS) framework for handling extreme illumination variations in space target reconstruction. 3DGS effectively addresses varying illumination across regions, but it may produce floaters due to local optimization minima. To address this, we integrate an optimizable decoupled appearance embedding model and implement a lighting mask filtering mechanism that maps 2D lighting masks to 3D point clouds. Experimental verification across multiple datasets shows that compared to state-of-the-art methods, chamfer distance has been reduced by 41.14% and PSNR has been improved by 13.98%.