Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberCN1332220 C
Publication typeGrant
Application numberCN 200410025217
Publication date15 Aug 2007
Filing date17 Jun 2004
Priority date17 Jun 2004
Also published asCN1595195A
Publication number200410025217.4, CN 1332220 C, CN 1332220C, CN 200410025217, CN-C-1332220, CN1332220 C, CN1332220C, CN200410025217, CN200410025217.4
Inventors李建勋, 郑军庭
Applicant上海交通大学
Export CitationBiBTeX, EndNote, RefMan
External Links: SIPO, Espacenet
Super broad band land radar automatic target identification method based on information fusion
CN 1332220 C
Abstract  translated from Chinese
一种基于信息融合的超宽带探地雷达自动目标识别方法,首先利用探地雷达回波信号中的直达波相对目标信号有一个较大的时间差,进行直达波的剔除,利用宽相关处理进行滤波和典型数据提取,提高信号的信噪比,提取纵向和横向典型数据用于目标形状识别,提取典型回波道数据进行Welch功率谱估计,并利用RBF网络进行目标材质分类,最后把目标形状识别和材质识别的结果进行信息融合,达到对不同形状,不同材质的地下目标的全面有效的自动识别。 An ultra-wideband ground penetrating radar automatic target recognition method based on information fusion, the first use of ground penetrating radar echo signals direct wave relative to the target signal has a large time difference, make direct wave removed, use a wide-related processing filter and typical data extraction and improve the signal to noise ratio, extracting longitudinal and transverse typical data for the target shape recognition, data extraction typical echo path Welch power spectrum estimation, using RBF network for target material classification, and finally the target shape recognition and the results of material identification information fusion, to achieve a comprehensive and effective automatic identification of different shapes, different materials in underground targets. 本发明实现了超宽带探地雷达目标的全面自动识别,对于实际的应用系统,特别是手持机具有重要意义和实用价值。 The present invention to achieve a comprehensive automatic identification ultra-wideband ground penetrating radar targets, for practical applications, especially handheld significance and practical value.
Claims(1)  translated from Chinese
1.一种基于信息融合的超宽带探地雷达自动目标识别方法,其特征在于包括如下具体步骤:1)数据处理:包括直达波剔除和信号滤波,将探地雷达的三维回波数据进行横向和纵向方向的平均,获取垂直方向的平均回波数据,从中选择第二和第三个回波的连接点作为截断点进行数据截断,抑制直达波,剔除前面的回波数据部分,将余下的回波数据作为含信号的数据进行后续处理,对截断后的探地雷达回波数据进行宽相关处理,得到三个典型切面和三个回波信号的最大值点处的X、Y、Z坐标值,此坐标值是以探地雷达为原点建立的三维坐标系中的坐标值;2)特征提取:包括用于目标形状识别的纵向和横向典型数据的提取和用于目标材质识别的典型道数据的提取,根据宽相关处理后得到的回波信号最大值点处的X、Y值,得到对应的纵向切面和横向切面数据,再取切面图最大值附近的各道数据对应的最大值,得到两个切面的轮廓点,得到用于形状识别的特征数据,确定不同的X、Y值,得到对应的纵向切面和横向切面交点的典型道数据,然后经Welch功率谱处理后,可以得到用于材质识别的数据;3)分类识别:将得到的形状识别特征数据进行曲线拟合,比较不同曲线对应的平方差,来确定拟合结果,利用不同形状目标回波信号对应不同的拟合曲线,并结合切面图显示,实现目标形状的识别;利用径向基函数RBF神经网络对目标材质进行分类,将与不同材质对应的典型道数据经Welch功率谱估计,得到用于材质识别的样本数据,送入径向基函数RBF神经网络进行训练建立特征量与目标值的函数关系,将上一步特征提取得到的用于材质识别的数据作为特征量输入神经网络,实现目标材质的自动识别;最后把目标形状识别和材质识别的结果进行信息融合,实现对不同材质,不同形状目标的全面自动识别。 A ground penetrating radar-based ultra-wideband automatic target recognition information fusion method, comprising the following specific steps: 1) Data processing: including and excluding the direct wave signal filtering, the three-dimensional ground penetrating radar echo data horizontal and the average vertical direction, vertical direction to obtain an average echo data, choose the connection point of the second and third cut-off point as echo data truncation, direct wave suppression, echo data excluding the previous section, the remaining echo data as the data contained in the subsequent signal processing, ground penetrating radar echo data truncated wide-related treatment to obtain maximum points at three typical section and three echo signals X, Y, Z coordinates value, this value is the coordinate origin GPR build a three-dimensional coordinate system coordinates; 2) feature extraction include: target for extracting longitudinal and lateral shape recognition and typical data for the target material identifying the typical road extracting data, according to the post-correlation processing width of the maximum value of the echo signal obtained at the point X, Y values, to give the corresponding longitudinal section and transverse section data, and then take the maximum value of each cutout in the vicinity of the corresponding data channel, get two section contour points, obtain characteristic data for the identification of shapes, determine the different X, Y value, to give a typical channel data corresponding longitudinal section and transverse section of the intersection, and then after treatment Welch power spectrum can be obtained by data in the material identified; 3) classification: shape recognition feature data obtained by curve fitting, comparing the different curves correspond to the square of the difference, the fitting result is determined, the target echo signal using different shapes corresponding to different fitting curve combined with cutaway view shows the shape of the target to achieve recognition; the use of radial basis function RBF neural network to classify the target material, corresponding to the different materials typical channel data by Welch power spectrum estimation, get sample data for material identification Data used for the identification of the material, into the radial basis function RBF neural network was trained to establish a function of the amount and characteristics of the target, and will get the feature extraction step as a feature of neural network input, automatic target recognition materials; and finally The results of the target shape recognition and material identification information fusion, to achieve full automatic identification of different materials, different shapes goals.
Description  translated from Chinese
基于信息融合的超宽带探地雷达自动目标识别方法 Based on Information Fusion ultra-wideband ground penetrating radar automatic target recognition

技术领域 FIELD

本发明涉及一种基于信息融合的超宽带探地雷达自动目标识别方法----基于宽相关处理、韦尔奇(Welch)功率谱分析、径向基函数(RBF)神经网络以及形状特征数据对目标进行全面自动识别,可广泛应用于地下金属/非金属管道探测、考古遗址定位、地质剖面勘探、高速公路质量检查以及安全检查等国家安全和经济领域中。 The present invention relates to an ultra-wideband ground penetrating radar automatic target recognition method based on information fusion ---- broad-based correlation process, Welch (Welch) power spectrum analysis, radial basis function (RBF) neural network and shape feature data Full automatic target recognition, can be widely used in underground metal / nonmetal pipeline detection, archaeological site location, geological exploration profile, highway quality inspection and safety inspection of national security and economic fields.

背景技术 BACKGROUND

探地雷达作为非破坏性探测手段正被广泛应用于地下目标(如空洞、管道、地雷等)的探测,如何对雷达回波信号进行处理以识别地下埋设的目标始终是困扰探地雷达应用的难题。 GPR as a non-destructive means of detection are being widely used in underground targets (such as voids, pipelines, mines, etc.) detection, how to radar echo signal is processed to identify Buried goal has always been plagued by ground penetrating radar applications problems. 目前主要的处理手段包括成像识别和特征变量识别。 The main treatment methods, including imaging to identify and recognize the characteristic variables.

成像处理通过对探地雷达回波信号的处理,获取了埋藏物体几何特征,从而可以根据几何特征(主要是外形)对目标加以判别,主要以合成孔径雷达(SAR)成像为主。 Imaging processing by ground penetrating radar echo signal is processed to obtain a buried object geometry, which can (mostly appearance) to determine the geometrical characteristics of the target, mainly in synthetic aperture radar (SAR) imaging based. 实现的方法包括三维距(Stanislav Vitebskiy,Lawrence Carin and MarcA.Ressler,Ultra-wideband,short-pulse Ground-penetrating radar:simulation andmeasurement.IEEE Trans.On geoscience and remote sensing.35(3),1997,762-772)和相位处理(Sai,B.;Ligthart,LP;GPR Phase-Based Techniques forProfiling Rough Surfaces and Detecting Small,Low-Contrast Landmines Under FlatGround Geoscience and Remote Sensing,IEEE Transactions on,Volume:42,Issue:2,Feb.2004 Pages:318-326)。 Implemented method includes a three-dimensional distance (Stanislav Vitebskiy, Lawrence Carin and MarcA.Ressler, Ultra-wideband, short-pulse Ground-penetrating radar: simulation andmeasurement.IEEE Trans.On geoscience and remote sensing.35 (3), 1997,762- 772) and phase processing (Sai, B; Ligthart, LP; GPR Phase-Based Techniques forProfiling Rough Surfaces and Detecting Small, Low-Contrast Landmines Under FlatGround Geoscience and Remote Sensing, IEEE Transactions on, Volume:. 42, Issue: 2, Feb.2004 Pages: 318-326). 由于大地的衰减和色散特性,使得探地雷达回波相互间具有不一致性,获取清晰的图像相对比较困难,从而造成很高的虚警率。 Due to attenuation and dispersion characteristics of the earth, making ground-penetrating radar echo each other have inconsistencies, get a clear image is relatively difficult, resulting in a high false alarm rate. 同时成像识别忽略了信号中原有的其它特征信息,尤其是比较难于区分形状相似的目标。 Simultaneous imaging identification ignores any original signal information other features, especially when compared to similar goals difficult to distinguish shape. 同时成像处理对实验设备要求高、计算复杂,不易实时处理。 Meanwhile imaging processing equipment for experiments that require high computational complexity, easy real-time processing. 处理结果一般由人工加以解释,含有较多的主观因素。 The results are generally explained by the artificial, contains more subjective factors.

基于特征变量识别主要是利用探地雷达的回波信号进行特征变量的提取,借助神经网络完成自动目标识别。 Feature-based recognition is the use of variable ground penetrating radar echo signal is extracted characteristic variables, with the neural network to complete the automatic target recognition. 已有的相关探地雷达特征提取方法包括:连续子波变换(T.Le-Tien,H.Talhami and DTNguyen,“Target SignatureExtraction Based on the Continuous Wavelet Transform in Ultra-WidebandRadar,”IEE Electronics Letters,Vol.33,Issue 1,January 1997),和时频分析(Guillermo C.Gaunaurd,Hans C.Strifors,Applications of(Wigner-Type)Time-Frequency Distributions to Sonar and Radar SignalAnalysis,7th.International Wigner Symposium held in College park,MD USA,2001)等。 GPR has been related feature extraction methods include: continuous wavelet transform (T.Le-Tien, H.Talhami and DTNguyen, "Target SignatureExtraction Based on the Continuous Wavelet Transform in Ultra-WidebandRadar," IEE Electronics Letters, Vol. 33, Issue 1, January 1997), and time-frequency analysis (Guillermo C.Gaunaurd, Hans C.Strifors, Applications of (Wigner-Type) Time-Frequency Distributions to Sonar and Radar SignalAnalysis, 7th.International Wigner Symposium held in College park , MD USA, 2001) and the like. 已有的方法主要是依据两维功率谱进行识别,特征量复杂不便于识别的工程识别,同时由于特征变量识别主要强调回波信号的特性,对于不同形状目标的识别却无能为力。 The method has been mainly based on two-dimensional power spectrum recognition, feature quantity is not easy to project identification recognition complex, and because the main emphasis on identifying characteristic variable characteristics of the echo signal, identifying targets for different shapes are powerless.

发明内容 SUMMARY

本发明的目的在于针对现有技术存在的不足,提供一种新的基于信息融合的超宽带探地雷达自动目标识别方法,即克服成像技术的设备要求高,不能区分形状相似目标的缺点,也克服了现有特征变量识别技术的复杂不易实现和对于不同形状目标的识别无能为力的不足,可以对不同形状、不同材质的地下目标进行有效的自动识别,达到工程化的实用效果。 The purpose of the present invention is for a deficiency of the prior art, to provide a new ultra-wideband ground penetrating radar based automatic target recognition information fusion imaging techniques that overcome the equipment requirements, can not distinguish between the target shape similar shortcomings, too characteristic variables to overcome the existing recognition technology for less complex and not easy to achieve the goal of identifying different shapes powerless, can be of different shapes, different materials underground targets for effective automatic identification, to engineering practical effect.

为实现这样的目的,本发明的技术方案中,首先对超宽带探地雷达回波信号进行直达波的剔除,利用宽相关处理进行信号滤波和典型数据提取。 To achieve this purpose, the technical aspect of the present invention, the first ultra-wideband ground penetrating radar echo signals direct wave removed, use a wide-related processing for signal filtering and typical data extraction. 提取纵向和横向典型数据用于目标形状识别;提取典型回波道数据并进行韦尔奇(Welch)功率谱分析,并利用RBF神经网络对目标材质进行分类,最后把目标形状识别和材质识别的结果进行信息融合,从而实现目标的全面自动识别。 Extract data for longitudinal and transverse typical target shape recognition; typical echo channel data extraction and Welch (Welch) power spectrum analysis, and the use of RBF neural network to classify the target material, and finally the target shape recognition and material identification information fusion results in order to achieve full automatic target recognition.

本发明的基于信息融合的超宽带探地雷达自动目标识别方法包括如下具体步骤:1.数据处理数据处理主要包括直达波剔除和信号滤波,用于提取典型纵向和横向切面数据和典型道数据。 Ultra-wideband ground penetrating radar automatic target recognition method based on information fusion of the present invention include the following specific steps: 1. Data processing Data processing includes removing the direct wave and signal filtering to extract typical longitudinal and transverse section data and typical channel data. 将探地雷达的三维回波数据进行横向和纵向方向的平均,获取垂直方向的平均回波数据,从中选择第二和第三个回波的连接点作为截断点进行数据截断,抑制直达波,剔除前面的回波数据部分,将余下的回波数据作为含信号的数据进行后续处理,对截断后的探地雷达回波数据进行宽相关处理,得到三个典型切面和三个回波信号的最大值点处的X、Y、Z坐标值,此坐标值是以探地雷达为原点建立的三维坐标系中的坐标值。 The three-dimensional ground penetrating radar echo data, the average horizontal and vertical directions, get vertical average echo data, choose the connection point of the second and third cut-off point as echo data truncation, inhibit direct wave, remove the front part of the echo data, the remaining data as the data contained echo signal for subsequent processing, ground penetrating radar echo data truncated wide-related treatment, get three typical sections and three echo signals maximum points at the X, Y, Z coordinate values, this is the coordinate values of ground penetrating radar to establish the origin of the coordinate values of the three-dimensional coordinate system.

由于探地雷达回波信号由收发天线间直接耦合波、地面反射波、地下介质不连续产生的后向散射波、随机干扰等构成。 Because ground-penetrating radar echo signal from the direct coupling between the transmitting and receiving antenna wave, ground reflected wave, constitute the scattered wave, random interference after subsurface discontinuities produced. 由直接耦合波和地面反射波组成的直达波直接影响回波目标信号。 By direct coupling wave and the ground reflected wave consisting of direct wave directly affect the target echo signal. 由于直达波相对目标信号有一个较大的时间差,因此本发明通过数据时间轴截断抑制直达波。 Because of the direct wave signal has a relatively large target time difference, thus the present invention is by inhibition of the direct wave data truncation time axis.

信号滤波采用宽相关处理方法实现。 Signal filtering width associated processing methods. 对截断后的探地雷达回波数据进行宽相关处理,可以提高回波信号的信噪比。 GPR echo data truncated wide-related processing can improve the signal to noise ratio of the echo signal. 宽相关处理的主要思想就是通过引入伸缩因子,所得的回波信号与伸缩的母波具有匹配关系。 The main idea of the wide-related processing is through the introduction of stretch factor, the resulting echo signal with retractable mother wavelet with a matching relationship. 经过宽相关处理后,可以得到三个典型切面和三个回波信号的最大值点(X,Y,Z)。 After wide correlation processing, can be obtained maximum point (X, Y, Z) of three typical sections and three echo signals.

2.特征提取特征提取主要包括两部分:用于目标形状识别的纵向和横向典型数据的提取和用于目标材质识别的典型道数据的提取。 2. The feature extraction feature extraction consists of two parts: the target shape recognition for extracting vertical and horizontal typical data and extract the target material identifying typical channel data. 根据宽相关处理后得到的回波信号最大值点处的X、Y值,得到对应的纵向切面和横向切面数据,再取切面图最大值附近的各道数据对应的最大值,得到两个切面的轮廓点,得到用于形状识别的特征数据,确定不同的X、Y值,得到对应的纵向切面和横向切面交点的典型道数据,然后经Welch功率谱处理后,可以得到用于材质识别的数据。 After the correlation processing based on a wide echo signal obtained at the maximum point X, Y values, to give the corresponding longitudinal section and transverse section data, and then take the cutout near the maximum value of each track data corresponding to the maximum value, to obtain two tangent plane contour points, to obtain characteristic data for the identification of shapes, determine the different X, Y value, to give a typical channel data corresponding longitudinal section and transverse section of the intersection, and then Welch power spectrum after treatment, can be used for material identification data.

经过宽相关处理后,可以得到三个典型切面和三个回波信号最大值点处的X,Y,Z值。 After wide correlation processing, resulting in three typical sections and three echo signal maximum point X, Y, Z values. 其中一个是水平切面,显示目标反射面的形状信息,一个纵向切面和一个横向切面,纵向切面的典型数据和横向切面的典型数据相结合用于目标形状的识别;最大值X、Y对应的宽相关处理数据代表回波的典型数据,用于目标材质的识别。 One is a horizontal section, the shape information of the target display reflecting surface, a longitudinal section and a transverse section, longitudinal section of a typical transverse section of a typical data and combined data for identifying the shape of the target; the maximum X, Y corresponding wide correlation processing echo data representative of typical data for identifying the target material.

根据宽相关处理后得到的回波信号最大值点处的X、Y值,得到对应的纵向切面和横向切面数据,再取切面图最大值附近的各道数据对应的最大值,得到两个切面的轮廓点,得到用于形状识别的特征数据。 After the correlation processing based on a wide echo signal obtained at the maximum point X, Y values, to give the corresponding longitudinal section and transverse section data, and then take the cutout near the maximum value of each track data corresponding to the maximum value, to obtain two tangent plane contour points, to obtain characteristic data for shape recognition. 根据两道数据的相似性进行目标形状的识别。 Identified by the similarity of the two target shape data.

确定不同的X、Y值,得到对应的纵向切面和横向切面交点的典型道数据,然后经Welch功率谱处理后,可以得到用于材质识别的数据。 Determine the different X, Y value, to obtain the corresponding longitudinal section and transverse section of typical data channel intersection, and then after Welch power spectrum processing, the data can be obtained for material identification.

基于宽相关处理所得到的最大值X、Y以及宽相关处理的三维结果,提取对应于(X,Y)的单道宽相关处理数据形成典型道回波数据。 Based on the obtained correlation processing width maximum X, Y, and wide dimensional result of the correlation process, the extraction corresponds to (X, Y) of the single-channel width-related data forming a typical channel processing echo data. 由于探地雷达回波信号的非平稳性,尤其是对于超宽带瞬态电磁散射信号,传统的基于傅立叶变换的谱估计方法都将不能使用。 Due to non-stationary ground penetrating radar echo signal, especially for ultra-wideband signal transient electromagnetic scattering, the traditional spectral estimation method based on Fourier transform will not be used. 考虑部分扫描的Welch平均重叠周期谱可以较好的用于非平稳信号的处理和一维的数据量,可以较好的用于目标特征的提取。 Welch considered part of the scan cycle average spectral overlap can be better used for processing and the amount of data one-dimensional non-stationary signals can be used to better target feature extraction.

将提取的典型道数据经Welch功率谱处理即可得到一维的功率谱,进而用于材质的识别。 Typical channel data extracted by Welch power spectrum can be obtained by processing the one-dimensional power spectrum, and then used to identify the material.

3.分类识别将得到的形状识别特征数据进行曲线拟合,比较不同曲线对应的平方差,来确定拟合结果,利用不同形状目标回波信号对应不同的拟合曲线,并结合切面图显示,实现目标形状的识别;利用径向基函数RBF神经网络对目标材质进行分类,将与不同材质对应的典型道数据经Welch功率谱估计,得到用于材质识别的样本数据,送入径向基函数RBF神经网络进行训练建立特征量与目标值的函数关系,将上一步特征提取得到的用于材质识别的数据作为特征量输入神经网络,实现目标材质的自动识别;最后把目标形状识别和材质识别的结果进行信息融合,实现对不同材质,不同形状目标的全面自动识别。 3. The shape recognition feature data classification was subjected to curve fitting, comparing the different curves correspond to the square of the difference, to determine fitting results, the use of different shapes corresponding to different target echo signal curve fitting, in conjunction with cutout display, achieve recognition target shape; the use of radial basis function RBF neural network to classify the target material, corresponding to the different materials typical channel data by Welch power spectrum estimation, get sample data for material identification, into the radial basis function RBF neural network is trained to establish a function of feature quantity and the target value, the data for the identification of the material obtained in the previous step, feature extraction as feature quantity input of the neural network, automatic identification of the target material; and finally the target shape and the material identifying Recognition The results of data fusion, to achieve full automatic identification of different materials, different shapes goals.

利用特征提取得到的轮廓点的数据进行一次曲线和二次曲线拟合,比较两次拟合曲线的平方差,来确定拟合结果是直线还是二次曲线。 The data obtained by feature extraction contour points once and quadratic curve fitting, comparing the difference between the two squares curve fitting to determine the fitting result is a straight line or a quadratic curve. 并结合三维显示中的纵向和横向典型切面结果,不同形状物体的两个典型切面的典型道数据的分布形状的不同。 Different distribution profiles and combined three-dimensional display of vertical and horizontal section a typical result, two objects of different shapes of the typical section of a typical channel data. 如果两个切面数据拟合都是二次的,显示为两个高峰,对应为球;如果一个为一次的,一个为二次的,显示一个为高峰分布,一个为不连续极值分布,则对应为管。 If the two facets are quadratic fit the data displayed as two peaks, corresponding to the ball; if one is one and one for the secondary, showing a peak distribution for an extreme value distribution is not continuous, then corresponding to the tube. 这样可以实现目标形状识别。 This allows the target shape recognition.

利用径向基函数RBF神经网络对目标材质进行分类,首先将与不同材质对应的典型道数据经Welch功率谱估计,得到用于材质识别的样本数据,送入径向基函数RBF神经网络进行训练建立特征量与目标值的函数关系,将上一步特征提取得到的用于材质识别的数据作为特征量输入神经网络,实现目标材质的自动识别。 Using Radial Basis Function RBF neural network to classify the target material, first with the different materials corresponding to a typical channel data by Welch power spectrum estimation, get sample data for material identification, into the radial basis function RBF neural network training establishing a function of feature quantity and the target value, the material identifying data for the feature extraction step as a feature quantity obtained by the neural network input, automatic identification of the target material.

针对得到的典型道特征数据,利用径向基函数RBF神经网络对目标材质进行分类。 Road characteristics typical data obtained using radial basis function RBF neural network to classify the target material. 首先分别从测量数据选取典型的土壤、铁和PVC数据,分别通过直达波剔除、Welch功率谱估计得到典型特征用于神经网络训练的输入,同时将对应的目标信息——土壤、铁和PVC分别用不同的值表示形成训练的期望输出。 First, the typical soils were selected, iron and PVC data from measured data were excluded by the direct wave, Welch power spectrum estimation to get input for the typical characteristics of neural network training, while the corresponding target information - soil, iron and PVC respectively formed with a different value indicates a desired output training. 当网络训练收敛以后的网络权值即代表了特征量与目标信息的映射关系。 When the weights of the network convergence after network training, which represents the mapping between the characteristics of the target amount of information. 针对特征提取的典型道数据的功率谱,通过训练收敛的神经网络即可进行目标材质的自动分类识别。 Power spectrum for a typical feature extraction channel data through the convergence of the neural network can be trained for automatic classification and identification of the target material.

最后把目标形状识别和材质识别的结果进行信息融合,可以实现对不同材质,不同形状目标的全面自动识别。 Finally, the results of the target shape and the material identifying the identification information fusion, can achieve full automatic identification of different materials, different shape of the target.

本发明的方法中,利用了探地雷达回波信号中的直达波相对目标信号有一个较大的时间差,进行直达波的剔除,并利用宽相关处理进行信号滤波和典型数据提取,提高了信号的信噪比。 The method of the present invention, the use of ground-penetrating radar echo signals direct wave relative to the target signal has a large time difference, make direct wave removed, and the use of wide-correlation signal filtering and processing of typical data extraction and improve the signal signal to noise ratio. 方法中提取纵向和横向典型数据用于目标形状识别,提取典型回波道数据并进行Welch功率谱分析,并利用RBF神经网络对目标材质进行分类,最后把目标形状识别和材质识别的结果进行信息融合,实现对不同材质,不同形状目标的自动识别。 Method to extract longitudinal and transverse typical data for the target shape recognition, data extraction typical echo path and Welch power spectrum analysis, and the use of RBF neural network to classify the target material, and finally the results of target shape recognition and material identification information fusion, automatic identification of different materials, different shape of the target. 本发明的方法易于实现,即克服现有成像技术的设备要求高,不能区分形状相似目标的缺点,也克服特征变量识别技术的对于不同形状目标的识别无能为力的不足,为探地雷达的工程化提供了一个有效的技术实现方法。 The method of the present invention is easy to realize that to overcome the existing imaging technology equipment requirements, can not distinguish the shape of the shortcomings similar objectives, but also to overcome the inability to identify the targets of different shapes characteristic variables insufficient recognition technology, for ground penetrating radar engineering provides an effective technology implementation. 本发明对于实际的应用系统,特别是手持机具有重要意义和实用价值。 The present invention is useful for practical applications, in particular handset significance and practical value.

附图说明 Brief Description

图1为本发明基于信息融合的超宽带探地雷达自动目标识别的原理框图。 Figure 1 is a block diagram of information fusion of ultra-wideband ground penetrating radar automatic target recognition based on the principle.

图2为不同形状物体的识别效果对照图。 Figure 2 is to identify the effect of the different shape of the object with reference to Figs.

其中,图2(a),(b),(c)为针对两根铁管的处理与显示对照图,图2(a)为原始数据显示,图2(b)为宽相关处理结果显示,图2(c)为三维显示;图2(d),(e),(f)为铝立方体的处理与显示对照图,图2(d)为原始数据显示,图2(e)为宽相关处理结果显示,图2(f)为三维显示。 Wherein, Fig. 2 (a), (b), (c) are for the two iron pipe processing and display control, and Fig. 2 (a) is the original data, Fig. 2 (b) is related to the processing result display width, Figure 2 (c) three-dimensional display; Fig. 2 (d), (e), (f) aluminum cube processing and display control, and Fig. 2 (d) show the original data, Fig. 2 (e) for the wide correlator processing result display, FIG. 2 (f) three-dimensional display.

图3为不同材质的典型道数据的Welch功率谱对照图。 Figure 3 is a control diagram Welch power spectrum of different materials typical channel data.

其中,图3(a)为典型道数据的Welch功率谱,图3(b)为PVC的典型道数据的Welch功率谱,图3(c)为土壤的典型道数据的Welch功率谱。 Wherein, Welch power Figure 3 (a) is a typical data spectrum channel, Welch power Figure 3 (b) is a typical PVC channel data spectrum, Figure 3 (c) for a typical soil Welch power spectrum of the data channel.

具体实施方式 DETAILED DESCRIPTION

为了更好地理解本发明的技术方案,以下结合附图对本发明的实施方式作进一步描述。 To better understand the technical solution of the present invention, the following with reference to the embodiment of the present invention will be further described.

本发明基于信息融合的超宽带探地雷达自动目标识别的原理框图如图1所示,总共包括三个主要部分,即数据处理、特征提取和分类识别。 The invention is based on information fusion of ultra-wideband ground penetrating radar automatic target recognition block diagram shown in Figure 1, a total of three main components, namely, data processing, feature extraction and classification. 其中数据处理部分主要包括直达波剔除和采用宽相关处理方法实现信号滤波,用于提取典型横向和纵向切面数据和典型道数据。 Wherein the data processing section includes direct wave with wide rejection and associated processing methods to achieve signal filtering to extract the typical horizontal and vertical slice data and typical channel data. 特征提取部分包括用于目标形状识别的横向和纵向典型数据的提取和用于目标材质识别的典型道数据的提取及提取后的功率谱估计。 Feature extraction section for extracting target shape recognition, including horizontal and vertical typical data and is used to extract the target material identifying typical channel data and extracted power spectrum estimation. 分类识别部分利用横向和纵向两个典型数据完成目标形状的识别和分类,对得到的目标材质识别特征数据利用RBF神经网络对目标材质进行识别和分类。 Classification utilizing both horizontal and vertical section of typical target shape data to complete the identification and classification of target material identifying characteristic data obtained by the use of RBF neural network to identify and classify the target material. 最后把目标形状识别和材质识别的结果进行信息融合从而获得目标识别结果。 Finally, the target shape identification and recognition of the results of the material in order to obtain information fusion target recognition results.

各部分具体实施细节如下:1.数据处理针对每一道测试数据,可建立超宽带探地雷达回波模型如下:探地雷达超宽带天线发射的探测脉冲为r1(t)=x(t),则回波信号为:S(t)=S0(t)+Σj=1m+1ΣφKi,jx(si,j(t-τi,j))+Σj=1m+1Σφ-Ki,jx(si,j(t-τi,j))+n(t)]]>其中:S0(t)为直达波,i表示第i次反射波,j表示第j层反射波。 Specific implementation details of each part as follows: 1. Data processing test data for each channel can be established ultra-wideband ground penetrating radar echo model is as follows: ground penetrating radar pulse ultra-wideband antennas as r1 (t) = x (t), the echo signal is: S (t) = S0 (t) + & Sigma; j = 1m + 1 & Sigma; & phi; Ki, jx (si, j (t- & tau; i, j)) + & Sigma; j = 1m + 1 & Sigma; & phi; -Ki, jx (si, j (t- & tau; i, j)) + n (t)]]> where: S0 (t) is a direct wave, i represents the i-th reflected wave, j represents j-layer reflection waves. m表示地面距埋藏目标可分的层数。 m represents the burial ground from the target can be divided into layers. φ={i|τi,j∈目标回波信号宽度内}, {| inside τi, j∈ target echo signal width i}, φ = 为φ补集。 To complement φ. n(t)为高斯噪声。 n (t) is Gaussian noise. Ki,j为衰减常数(对应反射系数),s1,m+1和τ1,m+1是待估计的未知参数,代表目标的时延、频谱展宽。 Ki, j is the attenuation constant (corresponding to the reflection coefficient), s1, m + 1 and τ1, m + 1 are unknown parameters to be estimated, on behalf of the target delay, spectral broadening.

经过直达波剔除后的回波信号可描述为:S′(t)=Σj=1m+1ΣφKi,jx(si,j(t-τi,j))+Σj=1m+1Σφ-Ki,jx(si,j(t-τi,j))+n(t)]]>在均匀介质条件下,忽略介质和多次反射波的影响,则用于目标检测和参数估计的有效回波信号可近似描述为:r2(t)=ΣiKi,Tx(si,T(t-τi,T))+n(t)]]>宽带相关处理器的输出为:WC(s,τ)=s∫r*1(s(t-τ))r2(t)dt]]>在非均匀介质情况下,通过多道数据纵向或横向平均,以纵向或横向分辨率的降低为代价换取正确的匹配和参数得稳健估计。 After the direct wave echo signals excluding can be described as: S & prime; (t) = & Sigma; j = 1m + 1 & Sigma; & phi; Ki, jx (si, j (t- & tau; i, j)) + & Sigma; j -Ki, jx (si, j (t- & tau;; i, j)) + n (t)]]> in a homogeneous medium conditions, ignoring the medium and multiple reflected waves impact, then = 1m + 1 & Sigma; & phi effective echo signal for target detection and parameter estimation can be approximately described as: r2 (t) = & Sigma; iKi, Tx (si, T (t- & tau; i, T)) + n (t)]]> Broadband output of the correlation processor for: WC (s, & tau;) = s & Integral; r * 1 (s (t- & tau;)) r2 (t) dt]]> in inhomogeneous media case, the multi-channel data longitudinal or The average transverse to longitudinal or transverse resolution reduces the cost in exchange for the right match and get robust estimation parameters.

经过宽相关处理后,可以得到三个典型切面和三个回波信号的最大值点(X,Y,Z)。 After wide correlation processing, can be obtained maximum point (X, Y, Z) of three typical sections and three echo signals. 一个是水平切面,显示目标反射面的形状信息,一个纵向切面和一个横向切面,纵向切面的典型数据和横向切面的典型数据相结合用于目标形状的识别。 A is a horizontal section, the shape information of the target display reflecting surface, a longitudinal section and a transverse section, longitudinal section of a typical data and typical transverse section of the combined data for identifying the target shape. 两个切面交点的道数据代表回波的典型数据,用于目标材质的识别。 Two facets Road intersection echo data represent typical data for identifying the target material.

2. 2. 特征提取特征提取主要包括两部分:用于目标形状识别的纵向和横向典型数据的提取和用于目标材质识别的典型道数据的提取。 Feature Extraction Feature extraction consists of two parts: the target shape recognition for extracting vertical and horizontal typical data and extract the target material identifying typical channel data.

经过宽相关处理后,可以得到回波信号最大值点处的X、Y、Z值,分别取X、Y值,可以得到对应的纵向切面和横向切面数据,再取切面图最大值附近的各道数据对应的最大值,得到两个切面的轮廓点,这样就得到了用于形状识别的典型道数据。 After wide correlation processing, can be X, Y, Z values of the echo signal at a maximum point, were taken X, Y values can be obtained the corresponding longitudinal section and transverse section data, and then take the cutout near the maximum value of each corresponding maximum channel data, get two facets of contour points, so get a typical channel data for shape recognition.

部分扫描Welch功率谱被证明可以用于目标材质的有效识别,Welch法谱估计采取数据分段加窗处理再求平均的办法,先分别求出每段的谱估计,然后进行总平均。 Partial scan proved Welch power spectrum can be used to effectively identify the target material, Welch spectral estimation method adopted windowed data segment and then the averaging approach to treatment, were obtained before each spectral estimation, then the overall average. 根据概率统计理论证明,若将原长度为N的数据分成K段,每段长度取M=N/K,如各段数据互为独立,则估计的方差将只有原来不分段的1/K,达到一致估计的目的。 According to the theory of probability and statistics prove that, if the original data length N into K segments, the length of the take M = N / K, independent of each other as each segment of data, it is estimated that the variance will not only original segment 1 / K reach consensus estimate purposes. 但若K增加、M减小,则分辨率下降。 However, if the increase in K, M decreases, the decrease in resolution. 相反,若K减小、M增加,虽偏差减小,但估计方差增大。 Conversely, if K is reduced, M increases, although the deviation is reduced, it is estimated that the variance increases. 所以在实际中必须兼顾分辨率与方差的要求适当选取K与M的值。 So in practice, you must take into account the requirements of the resolution and variance of K and M values appropriate selection.

Welch功率谱估计的计算过程如下:设信号s(n)的长度为512,将其分成K=7段,每段长度为N=128,重叠50%。 Welch power spectrum estimation calculation is as follows: Let the signal s (n) of length 512, K = 7 will be divided into segments, each of length N = 128, 50% overlap. 并对每个子集加上一个hanmin窗w(n)(n=128)。 And each subset plus a hanmin window w (n) (n = 128).

Welch功率谱估计按下式计算:Pw=1UKΣi=1kSi(w)Si*(w)]]>Si(w)=Si(n)w(n)e-2πmwn]]>U=1mΣn=0m-1w2(n)]]>图3为不同材质的典型道数据的Welch功率谱对照图,对比可以看到三者之间存在着较大的差别,因此可以用来作为目标的材质识别和目标的检测。 Welch power spectrum estimation is calculated as follows: Pw = 1UK & Sigma; i = 1kSi (w) Si * (w)]]> Si (w) = Si (n) w (n) e-2 & pi; mwn]]> U = 1m & Sigma; n = 0m-1w2 (n)]]> Welch power spectrum of FIG. 3 is a control diagram of a typical channel data of different materials, the contrast can be seen there are large differences between the three, and therefore can be used as the target identification and detection of the target material. 确定不同的X,Y值,得到对应的纵向切面和横向切面交点的典型道数据,然后经Welch功率谱处理后,可以得到用于材质识别的数据. Determine the different X, Y value, to obtain the corresponding longitudinal section and transverse section of typical data channel intersection, and then after Welch power spectrum processing, the data can be obtained for material identification.

3.分类识别本发明目标形状识别的试验采用的数据分别为针对球和管的测量数据。 3. The test data used in target classification and recognition of this invention shape recognition were measured data for the ball and tube. 实验的方法是首先针对测量的数据进行宽相关信号处理,获得水平切面图、横向切面图和纵向切面图。 The method of experiment is the first measurement data for the wide-correlation signal processing, obtains the horizontal cutaway view, a longitudinal section view and a transverse section in FIG. 结合纵向和横向切面中的典型数据进行目标形状识别。 Combined longitudinal and transverse section in the shape of a typical data for target recognition.

经过宽相关处理,回波信号的信噪比得到了增强。 After wide correlation processing, echo signal to noise ratio is enhanced. 利用特征提取得到的轮廓点的数据进行一次曲线和二次曲线拟合,比较两次拟合曲线的平方差,来确定拟合结果是直线还是二次曲线。 The data obtained by feature extraction contour points once and quadratic curve fitting, comparing the difference between the two squares curve fitting to determine the fitting result is a straight line or a quadratic curve. 并结合三维显示中的纵向和横向典型切面结果,不同形状物体的两个典型切面的典型道数据的分布形状的不同。 Different distribution profiles and combined three-dimensional display of vertical and horizontal section a typical result, two objects of different shapes of the typical section of a typical channel data. 如图2所示,如果两个切面数据拟合都是二次的,显示为两个高峰,对应为球;如果一个为一次的,一个为二次的,显示一个为高峰分布,一个为不连续极值分布,则对应为管。 2, if two slice data fit are quadratic, appear as two peaks, corresponding to the ball; if one is one, and a secondary, and a display for the peak distribution, a is not continuous extreme value distribution for the corresponding pipe. 这样可以实现目标形状识别。 This allows the target shape recognition.

本发明采用RBF径向基函数神经网络进行目标识别。 The invention uses radial basis function RBF neural network target recognition. RBF选取具有单隐层的三层前馈网络,包括输入层、中间层和输出层。 RBF select three with a single hidden layer feedforward networks, including an input layer, an intermediate layer and output layer. 输入层个数的选取依据选取的特征向量的采样点数。 The number of input layer according to the selected feature vectors selected sampling points. 考虑回波信号中有用信息的长度,本采样点数取为128。 Echo signals in consideration of the length of the useful information, the present number of samples taken as 128. 中间层个数的选取原则为2倍的输入层个数减去输出层个数。 The principle of selecting the number of input layer of the intermediate layer is 2 times the number minus the number of output layer. 输出层个数为1,根据不同的应用分别用0,1,2代表待识别物体的种类---土壤、铁和PVC。 The number of output layer 1, respectively, depending on the application object to be identified using the representative 0,1,2 --- of the kind of soil, iron and PVC.

针对实际数据的宽相关处理结果,分别取土壤和目标上不同的X,Y值,将对应的不同的典型道数据经Welch功率谱估计,得到用于材质识别的样本数据,对比可以看到三者之间存在着较大的差别,因此可以用来作为目标的材质识别和目标的检测。 For wide actual correlation process result data, were taken a different X, Y value of the soil and the target, the corresponding channel data via different Welch typical power spectrum estimation, obtained for the sample material identifying data, the contrast can be seen three there is a large difference between the person and therefore can be used to detect a target material and target identification. 将功率谱特征量送入径向基函数RBF神经网络进行训练。 The power spectrum characteristic quantities into radial basis function RBF neural network training. 同时针对待识别的测量数据通过宽相关处理得到的回波信号最大值。 While for the data to be identified by measuring the maximum width of the echo signal obtained by the correlation process. 对应(X,Y)的典型道信号经过Welch功率谱估计,进而通过神经网络进行分类识别。 Corresponding to (X, Y) of a typical channel signal through Welch power spectrum estimation, and then identified by a neural network to classify. 根据网络的输出值的范围进行目标材质的自动识别。 Automatic identification of the target material in accordance with the output value of the range of the network. 当输出值∈(-0.5,0.5),判定为土壤;当输出值∈(0.5,1.5),判定为铁;当输出值∈(1.5,2.5),判定为PVC;其它输出值,判定其它。 When the output value ∈ (-0.5,0.5), it is determined that the soil; when the output value ∈ (0.5,1.5), it is determined that iron; when the output value ∈ (1.5,2.5), it is determined that PVC; other output value, determining other.

如图3所示。 As shown in Figure 3. 对于伪铁管和PVC管的神经网络的训练与识别,输出结果为表1,反映Welch功率谱可以有效的借助神经网络完成对地下目标材质的识别。 For training and recognition neural network pseudo iron pipe and PVC pipe, the output results in Table 1, reflecting Welch power spectrum by means of neural networks can effectively complete the identification of underground target material.

表1 Table 1

对比现有成像识别技术和特征变量识别,本发明可以有效地对不同形状,不同材质的地下目标进行有效的自动识别,能够达到工程化的实用效果。 Contrast imaging to identify existing technologies and identify characteristic variables, the present invention can be effectively different shapes, different materials underground targets for effective automatic identification, engineered to achieve practical results. 同时从整个实现步骤可知,本发明的方法易于实现,从而为探地雷达的工程化提供了一个技术实现方法。 At the same time we can see from the entire implementation steps, the method of the present invention is easy to implement, so as ground penetrating radar engineering provides a technical implementation.

Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US5384715 *27 Aug 199324 Jan 1995The Texas A&M Univeristy SystemSystem identification and analysis of subsurface radar signals
US5644314 *29 Mar 19961 Jul 1997The United States Of America As Represented By The Secretary Of The ArmyPortable geophysical system using an inverse collocation-type metehodology
US5912639 *23 May 199715 Jun 1999Power Spectra, Inc.Ground penetrating radar with synthesized end-fire array
US6222481 *4 Jul 199724 Apr 2001Forsvarets ForskningsanstaltMethod of detecting and classifying objects by means of radar
Non-Patent Citations
Reference
1 *基于模糊神经网络和D-S推理的智能特征信息融合研究 敬忠良,杨永胜,李建勋,戴冠中,信息与控制,第26卷第2期 1997;基于小波能量特征的探雷方法研究 王群,倪宏伟,徐毅刚,数据采集与处理,第18卷第2期 2003;利用复合特征进行模式识别的探雷研究 王群,倪宏伟,徐毅刚,应用科学学报,第21卷第1期 2003
Classifications
International ClassificationG01S7/41, G01V3/17, G01S13/02
Legal Events
DateCodeEventDescription
16 Mar 2005C06Publication
18 May 2005C10Request of examination as to substance
15 Aug 2007C14Granted
24 Aug 2011C17Cessation of patent right