| 摘要: |
| 本文针对复杂杂波环境下的雷达目标检测问题，提出了一种基于图神经网络的多尺度特征融合雷达目标检测方法，该方法利用多个脉冲回波之间的特征关联性检测目标。具体而言，其首先利用多个级联的特征提取模块从回波中提取多尺度特征。然后，该方法利用多尺度特征构造多个有向完全图，图中每个节点对应一个脉冲。之后，每个节点利用图神经网络的消息传播机制聚合其邻居节点的信息，以此学习脉间回波高阶相关性。进一步地，该方法融合多尺度特征以丰富对目标和杂波的特征表示。最后，对融合后特征进行非线性映射，以二分类的形式得到检测结果。基于公开雷达数据集的试验验证了所提方法的有效性。 |
| 关键词: 雷达目标检测 杂波环境 图神经网络 多尺度特征融合 |
| DOI： |
| 分类号:TN959.1+1 |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Graph Neural Network Based RadarTarget Detection Method with Multi-scale Feature Fusion |
|
王彦斌, 汪育苗, 崔国龙
|
|
|
| Abstract: |
| This paper deals with the problem of detection in complex clutter environment. We propose a graph neural network-based radar target detection method with multi-scale feature fusion, which uses the feature correlations among multiple pulse echoes for the detection. Specifically, the proposed method first utilizes multiple cascaded feature extraction modules to extract muti-scale features. Subsequently, it constructs multiple directed complete graphs using muti-scale features, where a node in a graph corresponds to a pulse, and the node features of each graph are the features of the corresponding scale. Then, each node aggregates the information of its neighbors using the graph neural network, and the proposed method can therefore learn the high-order correlation among the pulse echoes. Further, the proposed method fuses multi-scale features to enrich the featur representation of the target and the clutter. Finally, the fused features are mapped nonlinearly, and the detection results are obtained in the form of binary classification. The effectiveness of the proposed method is verified using public radar database. |
| Key words: radar target detection clutter environment graph neural network multi-scale feature fusion |