现代防御技术 ›› 2023, Vol. 51 ›› Issue (4): 53-62.DOI: 10.3969/j.issn.1009-086x.2023.04.007

• 目标特性与探测跟踪技术 • 上一篇    下一篇

一种多源集中式空中目标类型综合识别方法

李喆1, 童逸琦2,3, 夏文博1, 应宇欣2,3   

  1. 1.北京电子工程总体研究所, 北京 100854
    2.北京航空航天大学 人工智能研究院, 北京 100191
    3.软件开发环境国家重点实验室, 北京 100191
  • 收稿日期:2022-08-16 修回日期:2022-10-31 出版日期:2023-08-28 发布日期:2023-09-01
  • 作者简介:李喆(1984-),男,浙江杭州人。高工,硕士,研究方向为指挥控制总体技术。
  • 基金资助:
    国家自然科学基金(62176014);科技创新2030-“新一代人工智能”重大项目(2021ZD0113602)

A Multi-source Centralized Air Target Classification Method

Zhe LI1, Yiqi TONG2,3, Wenbo XIA1, Yuxin YING2,3   

  1. 1.Beijing Institute of Electronic System Engineering, Beijing 100854, China
    2.Beihang University, Beijing 100191, China
    3.State Key Laboratory of Software Development Environment, Beijing 100191, China
  • Received:2022-08-16 Revised:2022-10-31 Online:2023-08-28 Published:2023-09-01

摘要:

与单传感器相比,多源传感器收集的多方面、多类型信息对目标类型识别更有意义。针对当前空中类型识别方法特征单一,分类结果无法交叉验证、准确率低的问题,提出一种融合多源传感器特征的目标类型综合识别方法。该方法首先利用深度学习模型提取目标飞行器的光电图像、运动轨迹、RCS以及电子侦收特征。考虑到真实环境下各传感器特征的重要程度会动态变化,采用注意力机制和距离参数对特征权重进行动态分配。在仿真数据集上的实验结果表明,与单传感器模型相比,集中式方法识别准确率平均提升12.89%,识别效果提升明显;与基于层次分析法投票的分布式模型相比,集中式方法不仅能够更有效地融合多源特征,提升识别效果,并且具有更强的鲁棒性,能适应复杂环境的变化。

关键词: 空中目标, 类型识别, 多源特征, 集中式, 注意力机制, 层次分析法

Abstract:

Compared with single sensors, the multi-faceted and multi-type information collected by multi-source sensors is more valuable for air target classification. In view of the current problems of air target recognition methods having single features, inability to cross-validate classification results and low recognition accuracy, an attention-based centralized air target recognition method with dynamic fusion of multi-source sensor features is proposed. This method uses deep learning models to extract the photoelectric image, motion trajectory, RCS and electromagnetic features of the target vehicle. Considering that the importance of each sensor feature will change dynamically in the real environment, the attention mechanism and the distance parameter are used to dynamically allocate the feature weights. The experimental results on the simulation dataset show that compared with the single-sensor model, the centralized method improves the recognition accuracy by 12.89% on average, which is a significant improvement in recognition effect; compared with the distributed model based on hierarchical analysis voting, the centralized method is more robust and better adapted to complex environments while fusing multi-source features more effectively.

Key words: air target, object classification, multi-source features, centralized model, attention mechanism, analytic hierarchy process(AHP)

中图分类号: