Scholarlink

<p>The integration of artificial intelligence technologies with ocean acoustic remote sensing processing substantially improves the automation and accuracy of ocean data analysis. Multibeam echosounder systems are widely used in underwater acoustic remote sensing point clouds, providing precise 3D data for detailed underwater topography. However, existing detection of the outlier in multibeam bathymetric point cloud suffers from low levels of automation and poor performance in complex scenarios. To address these challenges, we proposed a structurally enhanced outlier detection framework for acoustic remote sensing point cloud using quadtree-based preprocessing for terrain slicing, which effectively preserves structural complexity through geometric feature enrichment. Subsequently, we developed an enhanced deep-learning model based on PointCleanNet architectures, incorporating normal vector features for the improved outlier detection, in which global trend features encapsulate the slice-level morphological patterns, localized details are refined for micro-topographic characterization, and oriented normal vector features contribute to structural description. These features are fused through the developed local-global fusion (LGF) module, forming a robust geometric representation that preserves macroscopic structures and critical microscopic details for precise outlier detection. Notably, the model incorporates loss function of a point cloud structural similarity (SSIM) index during training to enhance its capability to detect outliers in complex point cloud structures. To validate the effectiveness of the proposed approach, experiments were conducted on both simulated and in situ datasets, including complex shipwreck and rocky seabed environments. The proposed method effectively eliminated various outliers while preserving terrain features, achieving superior performance with an average Recall of 83.82%, Precision of 90.03%, and Kappa of 99.75%, significantly outperforming traditional methods.</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

A structurally enhanced framework for automated outlier detection of point clouds by underwater acoustic remote sensing

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Fanli Yang,
  • Mingyi Gan,
  • Xianhai Bu,
  • Dianpeng Su,
  • Feng Wang

摘要

The integration of artificial intelligence technologies with ocean acoustic remote sensing processing substantially improves the automation and accuracy of ocean data analysis. Multibeam echosounder systems are widely used in underwater acoustic remote sensing point clouds, providing precise 3D data for detailed underwater topography. However, existing detection of the outlier in multibeam bathymetric point cloud suffers from low levels of automation and poor performance in complex scenarios. To address these challenges, we proposed a structurally enhanced outlier detection framework for acoustic remote sensing point cloud using quadtree-based preprocessing for terrain slicing, which effectively preserves structural complexity through geometric feature enrichment. Subsequently, we developed an enhanced deep-learning model based on PointCleanNet architectures, incorporating normal vector features for the improved outlier detection, in which global trend features encapsulate the slice-level morphological patterns, localized details are refined for micro-topographic characterization, and oriented normal vector features contribute to structural description. These features are fused through the developed local-global fusion (LGF) module, forming a robust geometric representation that preserves macroscopic structures and critical microscopic details for precise outlier detection. Notably, the model incorporates loss function of a point cloud structural similarity (SSIM) index during training to enhance its capability to detect outliers in complex point cloud structures. To validate the effectiveness of the proposed approach, experiments were conducted on both simulated and in situ datasets, including complex shipwreck and rocky seabed environments. The proposed method effectively eliminated various outliers while preserving terrain features, achieving superior performance with an average Recall of 83.82%, Precision of 90.03%, and Kappa of 99.75%, significantly outperforming traditional methods.