Scholarlink

<p>Visual SLAM (simultaneous localization and mapping) has emerged as a critical technology in robotics and autonomous systems, enabling navigation in indoor and GPS-denied environments by simultaneously estimating camera poses and constructing environmental maps using visual inputs. Despite its advantages, traditional Visual SLAM methods suffer from performance degradation under dynamic lighting, motion blur, or textureless regions. To address these limitations, this paper proposes a novel framework that integrates an optical flow-based deep learning network with Probabilistic Volumetric Fusion (PVF) and Neural Radiance Fields (NeRF). The proposed method constructs correlation volumes from sequential images and performs accurate 3D mapping by refining depth estimates and camera poses. Through extensive experiments on TartanAir, EuRoC MAV, and TUM-RGBD datasets, the method demonstrates superior accuracy and robustness compared to existing SLAM approaches, particularly in 3D reconstruction tasks evaluated using <i>L</i><sub>1</sub> Norm and PSNR. This work contributes to improving the reliability and precision of Visual SLAM for real-world deployment.</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

Learning-based visual SLAM with optical flow and probabilistic volumetric fusion

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Min-Hong Park,
  • Jae-Hoon Cho,
  • Yong-Tae Kim

摘要

Visual SLAM (simultaneous localization and mapping) has emerged as a critical technology in robotics and autonomous systems, enabling navigation in indoor and GPS-denied environments by simultaneously estimating camera poses and constructing environmental maps using visual inputs. Despite its advantages, traditional Visual SLAM methods suffer from performance degradation under dynamic lighting, motion blur, or textureless regions. To address these limitations, this paper proposes a novel framework that integrates an optical flow-based deep learning network with Probabilistic Volumetric Fusion (PVF) and Neural Radiance Fields (NeRF). The proposed method constructs correlation volumes from sequential images and performs accurate 3D mapping by refining depth estimates and camera poses. Through extensive experiments on TartanAir, EuRoC MAV, and TUM-RGBD datasets, the method demonstrates superior accuracy and robustness compared to existing SLAM approaches, particularly in 3D reconstruction tasks evaluated using L1 Norm and PSNR. This work contributes to improving the reliability and precision of Visual SLAM for real-world deployment.