海洋测绘2025，Vol.45Issue(5)：6-10,5.DOI:10.3969/j.issn.1671-3044.2025.05.002
基于神经网络的水下桩体多波束点云滤波
Application and analysis of point cloud neural network in underwater pile filtering
摘要
Abstract
To address the inaccuracy in underwater pile surveying caused by outliers in multibeam bathymetric data,this study proposes a filtering method integrating sample augmentation and a lightweight PCPNet.Four outlier simulation models(isolated points/structured clusters/bad pings/side-lobe effects)were constructed based on sonar physical mechanisms.The improved network adopts random local sampling(N=500 points),removes the STN module,optimizes feature alignment,and employs a weighted loss function.Tests on 35 wind-power pile datasets show:at 90％confidence,the average outlier removal rate reaches 98.7％with pile point retention rate＞99％,significantly outperforming conventional methods.This approach provides reliable technical support for high-accuracy pile monitoring in complex seabed environments.关键词
海洋测绘/点云滤波/多波束测深/点云局部特征学习/水下桩体目标/粗差滤除Key words
hydrographic surveying and charting/point cloud filtering/multibeam bathymetry/local feature learning of point cloud/underwater pile target/outlier removal分类
天文与地球科学引用本文复制引用
夏显文,王炎青,龚权华..基于神经网络的水下桩体多波束点云滤波[J].海洋测绘,2025,45(5):6-10,5.基金项目
国家重点研发专项(2022YFC2808303). （2022YFC2808303）
