航空学报 > 2026, Vol. 47 Issue (10): 533278-533278   doi: 10.7527/S1000-6893.2026.33278

航天遥感图像智能处理与分析专刊

Render3D: 基于三维渲染的月面立体匹配自监督学习方法

孟偲1,2,3(), 李依真1, 李俊博1, 梅既澜1, 白相志1,2, 马江4   

  1. 1.北京航空航天大学 宇航学院,北京 102206
    2.空间飞行器设计优化与动态模拟技术教育部重点实验室,北京 102206
    3.北航-威睛AI+计算光学联合实验室,北京 102206
    4.海军装备部,北京 100071
  • 收稿日期:2025-12-24 修回日期:2025-12-30 接受日期:2026-01-05 出版日期:2026-01-16 发布日期:2026-01-15
  • 通讯作者: 孟偲 E-mail:tsai@buaa.edu.cn
  • 基金资助:
    北京市自然科学基金联合重点项目(L258070)

Render3D: A self-supervised learning method for lunar surface stereo matching based on 3D rendering

Cai MENG1,2,3(), Yizhen LI1, Junbo LI1, Jilan MEI1, Xiangzhi BAI1,2, Jiang MA4   

  1. 1.School of Astronautics,Beihang University,Beijing 102206,China
    2.Key Laboratory of Spacecraft Design Optimization and Dynamic Simulation Technology,Ministry of Education,Beijing 102206,China
    3.Beihang - Vision AI + Computational Optics Joint Laboratory,Beijing 102206,China
    4.Naval Equipment Department,Beijing 100071,China
  • Received:2025-12-24 Revised:2025-12-30 Accepted:2026-01-05 Online:2026-01-16 Published:2026-01-15
  • Contact: Cai MENG E-mail:tsai@buaa.edu.cn
  • Supported by:
    Joint Key Project of Beijing Natural Science Foundation(L258070)

摘要:

立体视觉以其低成本和高可靠性等优势,成为月面巡视探测中感知三维地形的重要技术路径。近年来,基于深度学习的立体匹配方法已成为实现高精度立体视觉的主流方案。然而,受限于数据采集条件,目前缺乏面向月球表面环境的公开立体匹配数据集,严重制约了深度学习模型在月面场景中的训练与微调,进而影响其对月面复杂地形的适应能力。针对该问题,提出了一种基于三维渲染的立体匹配自监督学习方法,记作Render3D。该方法只需单目月面环拍图像输入,通过融合神经辐射场(NeRF)的精确几何重建能力与2D高斯泼溅(2DGS)的高保真表面渲染能力,生成高质量伪标注训练样本,指导立体匹配模型进行微调以适配月面环境。在月面仿真环境与真实物理场景上的实验表明,经Render3D微调后的立体匹配模型在精度上显著优于其他对比方法。在月面场景下,所提方法明显领先于现有自监督学习方法,尤其在纹理稀疏与高阴影区域等复杂环境下,其匹配误差较基线方法降低约50%,达到最优性能。实验结果充分证实,Render3D方法能够有效缓解月面标注数据稀缺的限制,显著提升立体匹配模型在月面环境中的鲁棒性与泛化能力。

关键词: 月面环境, 深度感知, 立体匹配, 自监督学习, 三维渲染

Abstract:

Stereo vision has emerged as a key technical approach for perceiving 3D terrain in lunar exploration, owing to its advantages such as low cost and high reliability. In recent years, the stereo matching methods based on deep learning have become the mainstream solution for achieving high-precision stereo vision. However, limited by data acquisition conditions, there is currently a lack of public stereo matching datasets tailored to the lunar surface environment. This severely restricts the training and fine-tuning of deep learning models in lunar surface scenarios, thereby impairing their adaptability to complex lunar terrain. To address this issue, this paper proposes Render3D, a self-supervised learning method for lunar surface stereo matching based on 3D rendering. The proposed method requires only monocular panoramic lunar surface images as input. By integrating the high-fidelity surface rendering capability of 2D Gaussian Splatting (2DGS) with the accurate geometric reconstruction capability of Neural Radiance Fields (NeRF), it generates high-quality pseudo-annotated training samples. These samples guide the fine-tuning of the stereo matching model to adapt to the lunar surface environment. Experiments conducted in both simulated lunar surface environment and real physical scenario demonstrate that the stereo matching model fine-tuned using Render3D method significantly outperforms other methods in terms of accuracy. In lunar surface scenarios, the proposed method shows clear superiority over existing self-supervised learning approaches, particularly in complex conditions such as textureless regions and areas with high shadows, where the matching error is reduced by approximately 50% compared to baseline methods, achieving state-of-the-art performance. Experimental results fully demonstrate that Render3D can effectively alleviate the constraint of scarce labeled data for lunar surface thus significantly improving the robustness and generalization ability of stereo matching models in the lunar surface environment.

Key words: lunar surface environment, depth perception, stereo matching, self-supervised learning, 3D rendering

中图分类号: