【作者】 袁莉；

【导师】 保铮；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2007， 博士

【摘要】 自20世纪60年代以来，战场感知对信息获取提出越来越高的要求，雷达自动目标识别(ATR)这一新的研究领域应运而生。高分辨率雷达的发展为ATR的研究提供了强有力的技术支持。由于雷达高分辨距离像(HRRP)能够提供目标沿距离方向的几何结构信息，且具有易于获取和处理的独特优势，使得基于HRRP的ATR技术受到越来越多的关注。在前人工作的基础上，本文着重研究HRRP的特征提取和目标分类两个方面的相关理论与技术，主要工作概况如下：从散射点模型出发，讨论了高分辨距离像的特性及用于识别时的需要考虑的几个关键问题。比较了利用逆Fourier变换(IFFT)得到的通常意义上的高分辨距离像、利用RELAX超分辨参数估计得到的距离像与利用MUSIC算法从目标回波中得到的超分辨距离像。结合仿真实验，详细分析了MUSIC超分辨距离像用于目标识别时对目标姿态、散射点个数、信噪比的敏感性等问题，说明MUSIC不能作为提高识别性能的超分辨算法用于距离像识别中。研究了高分辨距离像识别中的功率谱特征的特征压缩问题。仿真实验说明功率谱低频特征分量之间存在较强的相关性。详细分析了几种现有的特征压缩方法，结合功率谱的统计特性，提出一种基于标准化数据Fisher判决分析(FDA)的特征压缩方法，并分析了其性能提高的机理。基于外场实测数据的识别结果表明，本文给出的方法在降低维数的同时，可以明显提高识别性能，且具有较好的鲁棒性。研究了高分辨距离像的平移不变特征—中心矩特征向量的识别方法。考虑到中心矩特征的目标姿态敏感性，由一定方位角内的平均距离像松弛HRRP方位敏感性，然后提取平均像的中心矩特征向量。采用具有良好推广能力的多类支撑向量机进行分类。与基于中心矩特征的MCC方法和基于原始距离像的MCC方法相比较，本文方法减少计算量的同时具有较高的识别率，基于实测数据的仿真实验结果表明本文方法是有效的。研究了多次距离像的识别问题。由于距离像的目标方位敏感性，单次距离像的识别性能常常受到限制。在许多雷达系统中，可以得到多个独立距离像样本。本文在生成模板特征和测试特征时均利用了平均像和方差像作为特征，给出一种基于最小Kullback-Leibler(K-L)距离准则的HRRP序列识别方法，使用与识别准则相同的最小K-L距离准则进行平移匹配，并给出了快速算法。与最大似然(ML)准则及AGC方法进行比较，本文方法性能相当但大大降低了平移匹配的计算量。给出了基于实测数据和仿真数据的识别实验，结果表明本文方法是有效的。研究了距离像识别中的一个关键问题—目标角域数目的确定及角域划分。雷达高分辨率距离像具有目标姿态敏感性的特点，在识别时的一种解决方法是对目标不同角域建立不同的统计模型。在给定系统参数条件下，选择目标划分角域个数及每个角域覆盖范围是影响识别器运算量及识别性能的关键。本文给出了一种基于数据的自适应学习上述分类器参数的算法，采用联合高斯分布来描述HRRP在目标一个角域内的统计特性，在贝叶斯框架下通过迭代算法来确定数据划分边界，并自动确定目标角域个数，在目标识别中的应用中具有现实意义。基于联合高斯分布的数据模型通过迭代算法来确定数据划分边界，并自动确定目标角域个数。与等间隔数据划分方法相比，本文方法在降低识别运算量的同时，可以提高识别性能。更多还原

【Abstract】 With the new requirement for radar to get more information from the battlefield, radar automatic target recognition (ATR) as a new research area appeared in the 1960s. The development of high range resolution radar has given strong support to RATR. High range resolution profile (HRRP) contains the target structure signatures and is easy to be acquired, which makes the HRRP ATR to be received intensive attention. The work of this dissertation is focused on feature extraction and target recognition of high range resolution profile. The main content is summarized as follows.Starting from the scatters model of high-resolution radar target, the property of HRRP and several key problems of HRRP recognition is discussed. The range profiles obtained by conventional inverse fast Fourier transform (IFFT) is compared with that obtained by the RELAX and multiple signal classification (MUSIC) super- resolution technique. Through simulations, we detail the problems such as the target orientation, the number of scatters and the signal-to-noise ratio sensitivity which should be paid attention to when using MUSIC super-resolution range profile for radar automatic target recognition.The dimensionality reduction of power spectrum feature is concerned in HRRP recognition. The stronger correlation between the low frequency bins is shown in simulation results. A concise overview of the linear feature compress methods is given. With the statistical property of power spectrum discussed, a new fisher discriminant analysis feature compress method, which is based on the standard data, is presented for dimensionality reduction of power spectrum feature. And the performance improvement mechanism of the proposed method is discussed. The experimental results based on the measured data show that the proposed technique achieves robust good recognition performance with low feature dimension.HRRP recognition based on translation invariant feature—the central moments feature vector is concerned. To handle the target aspect sensitivity of moments feature, the average HRRPs associated with different target aspect sectors are used to extract the central moments feature vector for further recognition. A multi-class support vector machine (SVM) classifier with better generalization is designed to classify airplane objects. Comparing with the moments based MCC method, and the HRRP based MCC, the proposed approach can achieve better recognition performance and reduce the computation complexity and storage requirement. Experiment results based on the measured data are given to show the efficiency of the proposed method. The multiples HRRPs recognition problem is concerned. Generally, using single profile can not achieve good recognition performance, because of HRRP’s target aspect sensitivity. Actually, a sequence of independent HRRPs can be obtained in many radar systems. The average range profile and the variance profile are extracted together as the feature vectors for both training data and test data representation. A decision rule is established for HRRP sequence recognition based on the minimum Kullback-Leibler (K-L) distance criterion. And the same criterion is used for time-shift compensation of HRRP, with fast algorithm proposed. Comparing with the maximum likelihood criterion and adaptive Gaussian classifier (AGC), the propose method is much more computational efficient but with comparable recognition performance. Experimental results based on both the measured and the simulation data show that the minimum K-L distance classifier is effective.Radar High Range Resolution Profile (HRRP) is very sensitive to target aspect variation. To deal with this problem, usually, multiple statistical models are built for different target aspect sector when using HRRP for target recognition. Therefore, how to determine target aspect sector number and how to divide target aspect sector play an important role in classifier training. A data driven adaptive learning algorithm is proposed in this paper, which determines the target aspect sector boundary based on a multivariate Gaussian statistical data model and an iteration algorithm, and the target aspect sector number can be determined simultaneously. Comparing with the traditional equal interval target aspect partition approach, the proposed approach can achieve better recognition performance with lower computation complexity. Experimental results based on the measured data show the efficiency of the proposed method.更多还原