首页|期刊导航|太赫兹科学与电子信息学报|基于PointMamba的3D-InISAR目标识别方法

基于PointMamba的3D-InISAR目标识别方法OA

3D-InISAR target recognition method based on PointMamba

中文摘要英文摘要

干涉逆合成孔径雷达(InISAR)通过多天线干涉测量实现了目标的三维(3D)重构,相较于传统二维成像方法能够提供更为丰富可靠的目标几何特征信息,但InISAR成像结果通常以稀疏无序的三维点云形式呈现,这种数据特性使传统基于手工特征或二维卷积的识别算法难以充分挖掘点云中蕴含的几何结构信息,严重制约了识别性能的提升.对此,本文提出一种基于PointMamba的3D-InISAR目标识别方法,利用该网络针对点云数据的强大序列建模能力,实现对InISAR点云的精准分类.采用计算机辅助设计(CAD)建模与点云降采样结合的方法,高还原度构建了目标散射点模型,并提出一种基于概率密度的特征点提取方法,解决了InISAR稀疏点云稠密化的难题;充分利用PointMamba网络强大的序列建模能力,通过空间填充曲线序列化和选择性状态空间模型,实现对无序点云的高效特征提取.通过多种识别网络的对比实验,所提方法在十类舰船目标数据集上的识别准确率为96.2%,较PointNet++(86.3%)和PCT(93.1%)分别提升了9.9个和3.1个百分点,验证了其在处理稀疏无序InISAR点云数据方面的优越性.

Interferometric Inverse Synthetic Aperture Radar(InISAR)achieves three-dimensional(3D)reconstruction of targets through multi-antenna interferometry,providing richer and more reliable target geometric feature information compared to traditional two-dimensional imaging methods.However,InISAR imaging results are typically presented in the form of sparse and unordered 3D point clouds,making it difficult for conventional recognition algorithms based on handcrafted features or 2D convolutions to fully exploit the geometric structural information contained within the point clouds,thereby severely limiting the improvement of recognition performance.To address this issue,this paper proposes a 3D-InISAR target recognition method based on PointMamba,leveraging the network's powerful sequence modeling capability for point cloud data to achieve accurate classification of InISAR point clouds.A method combining Computer-Aided Design(CAD)modeling with point cloud downsampling is employed to construct target scattering point models with high fidelity,and a feature point extraction method based on probability density is proposed to solve the problem of densifying sparse InISAR point clouds.By fully exploiting the powerful sequence modeling capability of the PointMamba network,efficient feature extraction from unordered point clouds is achieved through space-filling curve serialization and selective state space models.Comparative experiments with various recognition networks demonstrate that the proposed method achieves a recognition accuracy of 96.2%on a 10-class ship target dataset,representing improvements of 9.9 and 3.1 percentage points over PointNet++(86.3%)and PCT(93.1%),respectively,thereby validating its superiority in processing sparse and unordered InISAR point-cloud data.

王勇;夏浩然

哈尔滨工业大学 电子工程技术研究所,黑龙江 哈尔滨 150001哈尔滨工业大学 电子工程技术研究所,黑龙江 哈尔滨 150001

信息技术与安全科学

干涉逆合成孔径雷达雷达成像深度学习雷达点云处理三维目标识别

Interferometric Inverse Synthetic Aperture Radar(InISAR)radar imagingdeep learningradar point cloud processing3D target recognition

《太赫兹科学与电子信息学报》 2026 (9)

1051-1064,14

国家重点研发计划资助项目(2024YFB3909800)国家杰出青年科学基金资助项目(62325104)

10.11805/TKYDA2025276

评论