【作者】 李明；

【导师】 刘梅；

【作者基本信息】 哈尔滨工业大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 随着空天战场环境的日益复杂和多模态数据融合技术的广泛应用,目标特征提取与识别技术在防空预警、战略防御等国防领域发挥着不可替代的作用。本文针对飞机类、弹道类与卫星目标,深入研究目标运动、微动及图像等多模态特征提取与智能识别技术,主要工作可以概括为以下四个方面:(1)目标运动特征提取与识别:在对飞机类、弹道类与卫星目标运动轨迹建模以及采用GLMB跟踪得到目标运动轨迹的基础上,研究了目标运动特征的提取与识别技术,提取到飞行高度、速度、轨迹曲率、平滑度以及轨迹斜率五种运动特征。针对弹道类与卫星目标在某些飞行区域运动特征模糊造成识别率下降的问题,提出了基于最小径矢的弹道目标与卫星识别方法,从二者运动轨迹的本质出发,实现了对弹道类目标与卫星目标的准确识别。(2)目标微动特征提取与参数估计:为更加细致地区分飞机类目标(直升机、螺旋桨飞机与喷气式飞机)与弹道类目标(锥体弹头与诱饵),需要进一步提取精细化特征。在目标微动建模以及电磁波仿真获得雷达回波的基础上,研究了目标微多普勒曲线提取与微动频率估计方法。针对时频变换由于分辨率限制导致时频谱图能量泄漏,造成微多普勒曲线提取困难的问题,本文提出差异化解决方案:一是针对飞机类目标,提出聚类-跟踪的方法,目标旋翼旋转会产生非交叉闪烁条带,在提取曲线前采用聚类算法对时频图进行能量集中处理;二是针对弹道导弹目标,提出STVMD-GLMB多信号分量跟踪方法,目标微动会产生多条交叉的微多普勒曲线,通过改进的短时变分模态分解方法分离信号分量,结合广义参数化时频变换与聚类算法实现能量集中。最终将聚类结果作为量测输入,采用GLMB多目标跟踪算法完成微多普勒曲线提取与微动频率估计。(3)可见光、红外图像及一维距离像融合识别:随着雷达反探测技术的不断发展以及电磁空间环境的日益复杂化,多源数据融合技术逐渐成为了研究热点。可见光、红外图像以及雷达一维距离像各具独特的成像特性,能够从不同角度为目标识别提供关键信息。在对三种传感器成像原理以及特点分析的基础上,研究了多源图像数据库的生成方法,利用Vega Prime与FEKO仿真软件生成目标仿真图像,并结合内在质量与环境质量标准,构建了多种环境下贴近实际的训练图像库。针对传统图像融合识别算法在特征提取与融合阶段,过度依赖人工干预,主观性过强的问题,提出了基于Res Net-Vi T的多源图像动态自适应融合识别方法,综合Res Net与Vi T的网络优势以及多源图像信息互补优势,相比于单源图像识别准确率更高。(4)基于深度学习的多模态数据端到端识别:为整合雷达量测、回波信号及多源图像等多模态数据,简化特征提取的复杂过程,研究了基于深度学习的端到端识别方法,根据不同数据的特点分别设计编码器,并利用LSTM网络处理变长序列的特性,解决了多模态数据中同一目标的样本数量不一致,导致融合特征长度不固定的问题,实现端到端识别。更多还原

【Abstract】 With the increasing complexity of the aerospace battlefield environment and the widespread application of multimodal data fusion technologies,target feature extraction and recognition technology plays an irreplaceable role in national defense fields such as air defense early warning and strategic defense.This paper focuses on aircraft,ballistic,and satellite targets,conducting in-depth research on multimodal feature extraction and intelligent recognition technologies for target motion,micro-motion,and imaging characteristics.The main contributions can be summarized into four aspects:(1)Target Motion Feature Extraction and Recognition:Based on motion trajectory modeling of aircraft,ballistic,and satellite targets and GLMB tracking results,this study investigates motion feature extraction and recognition technologies,extracting five motion features:flight altitude,velocity,trajectory curvature,smoothness,and trajectory slope.To address the recognition rate decline caused by motion feature ambiguity between ballistic and satellite targets in specific flight regions,a minimum path vector-based recognition method is proposed,achieving accurate discrimination between ballistic and satellite targets through essential analysis of their motion trajectories.(2)Target Micro-Motion Feature Extraction and Parameter Estimation:To more meticulously distinguish between aircraft-type targets(helicopters,propeller planes,and jet planes)and ballistic-type targets(conical warheads and decoys),it is necessary to further extract refined features.Based on target micro-motion modeling and electromagnetic wave simulation to obtain radar echoes,this paper studies methods for micro-Doppler curve extraction and micro-motion frequency estimation.Aiming at the problem that time-frequency transformation causes energy leakage in time-frequency spectrograms due to resolution limitations,making micro-Doppler curve extraction difficult,this paper proposes differentiated solutions:First,for aircraft-type targets,a clustering-tracking method is proposed.The rotational motion of the rotor produces non-crossing flickering strips,and a clustering algorithm is adopted to perform energy concentration processing on the time-frequency diagram before curve extraction;Second,for ballistic missile targets,an STVMD-GLMB multi-signal component tracking method is proposed.The micro-motion of the target generates multiple intersecting micro-Doppler curves.The improved Short-Time Variational Mode Decomposition(STVMD)method is used to separate signal components,and energy concentration is achieved by combining generalized parameterized time-frequency transformation with a clustering algorithm.Finally,the clustering results are used as measurement inputs,and the GLMB(Generalized Labeled Multi-Bernoulli)multi-target tracking algorithm is employed to complete micro-Doppler curve extraction and micro-motion frequency estimation.(3)Fusion Recognition of Visible Light/Infrared Images and HRRP:With the continuous development of radar counter-detection technologies and the growing complexity of electromagnetic environments,multi-source data fusion technology has become a research hotspot.Visible light/infrared images and HRRP possess unique imaging characteristics that provide complementary information for target recognition.Based on analysis of three sensor imaging principles and characteristics,this paper studies methods for generating multi-source image databases.Using Vega Prime and FEKO simulation software to generate target simulation images,combined with intrinsic quality and environmental quality standards,a practical training image library under various conditions is constructed.To address the excessive reliance on manual intervention in traditional image fusion recognition algorithms,a dynamic adaptive fusion recognition method based on Res Net-Vi T is proposed,integrating the advantages of Res Net and Vision Transformer networks,achieving higher accuracy compared with single-source image recognition.(4)End-to-End Recognition of Multimodal Data Using Deep Learning:To integrate multimodal data including radar measurements,echo signals,and multi-source images while simplifying complex feature extraction processes,an end-to-end recognition method based on deep learning is investigated.Dedicated encoders are designed for different data types,and LSTM networks are employed to handle variable-length sequences,resolving the issue of inconsistent sample quantities for the same target in multimodal data that causes variable fusion feature lengths,ultimately realizing end-to-end recognition.更多还原