CN1595195A - Super broad band land radar automatic target identification method based on information fusion - Google Patents

Abstract

It is a ultra-wide band earth-probing radar self object identification method based on information integration, which is the following: first to use the large time lag in the relative object signals to the direct wave in the echo-signal of the earth-probing radar and to filter the direct wave; to use the bandwidth relative process to filter wave and to extract typical data and to improve the signal to noise ratio of the signal and to extract the longitudinal and transverse typical data to object shape identification and to extract typical echo-duct data for Welch power spectrum valuation; to use RBF web to classify the object material and finally to integrate the results of the object shape identification and material identification to achieve the complete and effective automatic identification of the underground object of different materials.

Description

Ultra-wideband ground-penetrating radar (uw-gpr) automatic target recognition method based on information fusion

Technical field

The present invention relates to a kind of ultra-wideband ground-penetrating radar (uw-gpr) automatic target recognition method----based on information fusion and target is carried out comprehensively identification automatically, can be widely used in that underground metal/non-metal pipeline is surveyed, in national security such as location, archaeological site, geologic section exploration, highway quality check and safety inspection and the economic field based on wide relevant treatment, Wei Erqi (Welch) power spectrumanalysis, radial basis function (RBF) neural network and character shape data.

Background technology

Ground penetrating radar just is being widely used in the detection of buried target (as cavity, pipeline, land mine etc.) as the non-destructive detection means, and how to handle to discern the underground target of burying underground to radar echo signal is the difficult problem that the puzzlement ground penetrating radar is used all the time.Main processing means comprise imaging identification and characteristic variable identification at present.

Imaging processing has been obtained and has been buried the object geometric properties, thereby can have been differentiated target according to geometric properties (mainly being profile) by the processing to ground penetrating radar echo signals, mainly is imaged as the master with synthetic aperture radar (SAR).The method that realizes comprises three-dimensional apart from (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, L.P.; 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).Because the decay and the dispersion characteristics of the earth make the ground penetrating radar echo have inconsistency each other, obtain relatively difficulty of distinct image, thereby cause very high false alarm rate.Original further feature information in the signal has been ignored in the identification of imaging simultaneously, especially relatively is difficult to distinguish the similar target of shape.Imaging processing requires high, calculation of complex to experimental facilities simultaneously, is difficult for handling in real time.Result is contained more subjective factor generally by manually explaining.

Identification mainly is to utilize the echoed signal of ground penetrating radar to carry out the extraction of characteristic variable based on characteristic variable, finishes automatic target identification by neural network.Existing relevant ground penetrating radar feature extracting method comprises: continuous wavelet transform (T.Le-Tien, H.Talhami and D.T.Nguyen, " 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) etc.Existent method mainly is to discern according to the bidimensional power spectrum, and the engineering identification that the characteristic quantity complexity is not easy to discern is because the characteristic of echoed signal is mainly emphasized in characteristic variable identification, powerless for the difformity identification of targets simultaneously.

Summary of the invention

The objective of the invention is to deficiency at the prior art existence, a kind of new ultra-wideband ground-penetrating radar (uw-gpr) automatic target recognition method based on information fusion is provided, promptly overcome the equipment requirements height of imaging technique, can not distinguish the shortcoming of the similar target of shape, the complexity that has also overcome existing characteristic variable recognition technology be difficult for to realize and for the helpless deficiency of difformity identification of targets, can carry out effectively identification automatically to the buried target of difformity, unlike material, reach the practical function of through engineering approaches.

For realizing such purpose, in the technical scheme of the present invention, at first the ultra-wideband ground-penetrating radar (uw-gpr) echoed signal is carried out the rejecting of direct wave, utilize wide relevant treatment to carry out signal filtering and typical data extraction.Extract the vertical and horizontal typical data and be used for target shape identification; Extract typical echo track data and carry out Wei Erqi (Welch) power spectrumanalysis, and utilize the RBF neural network that the target material is classified, at last the result of target shape identification and Material Identification is carried out information fusion, thereby realize the identification automatically comprehensively of target.

Ultra-wideband ground-penetrating radar (uw-gpr) automatic target recognition method based on information fusion of the present invention comprises following concrete steps:

1. data processing

Data processing mainly comprises direct wave rejecting and signal filtering, is used to extract typical vertical and horizontal tangent plane data and typical track data.The three-dimensional echo data of ground penetrating radar is carried out the average of horizontal and vertical direction, obtain the average echo data of vertical direction, therefrom select the tie point of second and the 3rd echo to carry out data truncation as truncation points, suppress direct wave, reject the echo data part of front, the echo data of remainder is carried out subsequent treatment as the data that contain signal, ground penetrating radar echo data after blocking is carried out wide relevant treatment, obtain X, Y, the Z value at the maximum of points place of three typical tangent and three echoed signals.

Because ground penetrating radar echo signals is by directly the backward scattered wave of coupled waves, ground-reflected wave, the discontinuous generation of underground medium, random disturbance etc. constitute between dual-mode antenna.Direct wave by direct coupled waves and ground return wave component directly influences the echo echo signal.Because the relative echo signal of direct wave has a bigger mistiming, so the present invention blocks the inhibition direct wave by the data time axle.

Signal filtering adopts wide correlation process method to realize.Ground penetrating radar echo data after blocking is carried out wide relevant treatment, can improve the signal to noise ratio (S/N ratio) of echoed signal.The main thought of wide relevant treatment is exactly by the introducing contraction-expansion factor, and the echoed signal of gained has matching relationship with the female ripple that stretches.After wide relevant treatment, can obtain three typical tangent and three echoed signals maximum of points (X, Y, Z).

2. feature extraction

Feature extraction mainly comprises two parts: be used for target shape identification the vertical and horizontal typical data extraction and be used for the extraction of the typical track data of target Material Identification.X, Y value according to the echoed signal maximum of points place that obtains after the wide relevant treatment, obtain corresponding vertical tangent plane and horizontal tangent plane data, get near the maximal value of each track data correspondence of sectional drawing maximal value again, obtain the point of two tangent planes, obtain being used for the characteristic of shape recognition, determine different X, Y value, obtain the typical track data of corresponding vertical tangent plane and horizontal tangent plane intersection point, after the Welch power spectrum is handled, can obtain being used for the data of Material Identification then.

After wide relevant treatment, can obtain the X at three typical tangent and three echoed signal maximum of points places, Y, Z value.One of them is a horizontal section, the shape information of display-object reflecting surface, and a vertical tangent plane and a horizontal tangent plane, vertically the typical data of the typical data of tangent plane and horizontal tangent plane combines and is used for the identification of target shape; The typical data of the data represented echo of wide relevant treatment of maximal value X, Y correspondence is used for the identification of target material.

X, Y value according to the echoed signal maximum of points place that obtains after the wide relevant treatment, obtain corresponding vertical tangent plane and horizontal tangent plane data, get near the maximal value of each track data correspondence of sectional drawing maximal value again, obtain the point of two tangent planes, obtain being used for the characteristic of shape recognition.Carry out the identification of target shape according to the similarity of two track datas.

Determine different X, Y value, obtain the typical track data of corresponding vertical tangent plane and horizontal tangent plane intersection point, after the Welch power spectrum is handled, can obtain being used for the data of Material Identification then.

Based on the three-dimensional result of the resulting maximal value X of wide relevant treatment, Y and wide relevant treatment, extract corresponding to (X, the wide relevant treatment data of single track Y) form typical road echo data.Because ground penetrating radar echo signals is non-stationary, especially for ultra broadband Transient Electromagnetic scattered signal, traditional spectrum method of estimation based on Fourier transform all can not be used.The average overlapping cyclic spectrum of Welch of considering part scanning can be used for the processing of non-stationary signal and the data volume of one dimension preferably, can be used for the extraction of target signature preferably.The typical track data that extracts can be obtained the power spectrum of one dimension through the processing of Welch power spectrum, and then be used for the identification of material.

3. Classification and Identification

The shape recognition characteristic that obtains is carried out curve fitting, and the difference of two squares of more different curve correspondences is determined fitting result, utilizes the corresponding different matched curve of difformity target echo signal, and shows in conjunction with sectional drawing, realizes the identification of target shape; Utilize radial basis function RBF neural network that the target material is classified, typical track data that will be corresponding with unlike material is through the Welch power Spectral Estimation, obtain being used for the sample data of Material Identification, send into radial basis function RBF neural network and train the funtcional relationship of setting up characteristic quantity and desired value, the data that are used for Material Identification that the previous step feature extraction is obtained realize the automatic identification of target material as the characteristic quantity input neural network; At last the result of target shape identification and Material Identification is carried out information fusion, realize, the identification automatically comprehensively of difformity target unlike material.

Utilize the data of the point that feature extraction obtains to carry out curve and conic fitting, relatively the difference of two squares of twice fitting curve determines that fitting result is straight line or quafric curve.And in conjunction with the vertical and horizontal typical tangent result in the 3-D display, the difference of the distribution shape of the typical track data of two typical tangent of difformity object.If two tangent plane data fittings all are secondaries, be shown as two peaks, correspond to ball; If one is once, one is secondary, shows that one is that the peak distributes, and one is discontinuous The extreme value distribution, then correspond to pipe.Can realize target shape identification like this.

Utilize radial basis function RBF neural network that the target material is classified, at first typical track data that will be corresponding with unlike material is through the Welch power Spectral Estimation, obtain being used for the sample data of Material Identification, send into radial basis function RBF neural network and train the funtcional relationship of setting up characteristic quantity and desired value, the data that are used for Material Identification that the previous step feature extraction is obtained realize the automatic identification of target material as the characteristic quantity input neural network.

At the typical road characteristic that obtains, utilize radial basis function RBF neural network that the target material is classified.Choose typical soil, iron and PVC data from measurement data at first respectively, respectively by direct wave reject, the Welch power Spectral Estimation obtains the input that characteristic feature is used for neural metwork training, simultaneously with correspondence target information---soil, iron and PVC form the desired output of training with different value representations respectively.The mapping relations of characteristic quantity and target information have promptly been represented when the later network weight of network training convergence.At the power spectrum of the typical track data of feature extraction, can carry out the automatic Classification and Identification of target material by training convergent neural network.

At last the result of target shape identification and Material Identification is carried out information fusion, can realize, the identification automatically comprehensively of difformity target unlike material.

In the method for the present invention, utilize the relative echo signal of the direct wave in the ground penetrating radar echo signals that a bigger mistiming is arranged, carried out the rejecting of direct wave, and utilized wide relevant treatment to carry out signal filtering and typical data extraction, improved the signal to noise ratio (S/N ratio) of signal.Extract the vertical and horizontal typical data in the method and be used for target shape identification, extract typical echo track data and carry out the Welch power spectrumanalysis, and utilize the RBF neural network that the target material is classified, at last the result of target shape identification and Material Identification is carried out information fusion, realization is to unlike material, the difformity Automatic identification of targets.Method of the present invention is easy to realize, promptly overcome the equipment requirements height of existing imaging technique, can not distinguish the shortcoming of the similar target of shape, also overcome the characteristic variable recognition technology for the helpless deficiency of difformity identification of targets, for the through engineering approaches of ground penetrating radar provides an otherwise effective technique implementation method.The present invention is significant and practical value for the application system of reality, particularly hand tool.

Description of drawings

Fig. 1 is the theory diagram that the present invention is based on the ultra-wideband ground-penetrating radar (uw-gpr) automatic target identification of information fusion.

Fig. 2 is the recognition effect contrast figure of difformity object.

Wherein, Fig. 2 (a) (b), is to scheme with showing contrast at the processing of two iron pipes (c), and Fig. 2 (a) is that raw data shows, Fig. 2 (b) is that wide relevant treatment result shows that Fig. 2 (c) is a 3-D display; Fig. 2 (d) (e), is that the cubical processing of aluminium is schemed with showing contrast (f), and Fig. 2 (d) is that raw data shows, Fig. 2 (e) is that wide relevant treatment result shows that Fig. 2 (f) is a 3-D display.

Fig. 3 is the Welch power spectrum contrast figure of the typical track data of unlike material.

Wherein, Fig. 3 (a) is the Welch power spectrum of typical track data, and Fig. 3 (b) is the Welch power spectrum of the typical track data of PVC, and Fig. 3 (c) is the Welch power spectrum of the typical track data of soil.

Embodiment

In order to understand technical scheme of the present invention better, embodiments of the present invention are further described below in conjunction with accompanying drawing.

The present invention is based on information fusion ultra-wideband ground-penetrating radar (uw-gpr) automatic target identification theory diagram as shown in Figure 1, comprise three major parts altogether, i.e. data processing, feature extraction and Classification and Identification.Wherein data processing section mainly comprises the direct wave rejecting and adopts wide correlation process method to realize signal filtering, is used to extract typical transverse and vertical tangent plane data and typical track data.Feature extraction partly comprise the extraction of the horizontal and vertical typical data that is used for target shape identification and be used for the target Material Identification typical track data extraction and extract after power Spectral Estimation.Classification and Identification partly utilizes horizontal and vertical two typical datas to finish the identification and the classification of target shape, utilizes the RBF neural network target material is discerned and to be classified to the target Material Identification characteristic that obtains.Thereby at last the result of target shape identification and Material Identification is carried out information fusion and obtain the Target Recognition result.

The concrete implementation detail of each several part is as follows:

1. data processing

At each road test data, it is as follows to set up the ultra-wideband ground-penetrating radar (uw-gpr) echo model:

The direct impulse of ground penetrating radar ultra-wideband antenna emission is r ₁(t)=and x (t), then echoed signal is:

$S\left(t\right)=S_0\left(t\right)+\underset{j=1}{\overset{m+1}{\mathrm{\Σ}}}\underset{\mathrm{\φ}}{\mathrm{\Σ}}{K}_{i,j}x\left(s_{i,j}(t-{\mathrm{\τ}}_{i,j})\right)+\underset{j=1}{\overset{m+1}{\mathrm{\Σ}}}\underset{\stackrel{\‾}{\mathrm{\φ}}}{\mathrm{\Σ}}{K}_{i,j}s\left(s_{i,j}(t-{\mathrm{\τ}}_{i,j})\right)$

$+n\left(t\right)$

Wherein: S ₀(t) be direct wave, i represents i secondary reflection ripple, and j represents j layer reflection wave.M represents that the ground identity distance buries the number of plies that target can be divided.φ={ i| τ _{I, j}In the ∈ target echo signal width }, φ is the φ supplementary set.N (t) is a Gaussian noise.K _{I, j}Be attenuation constant (corresponding reflection coefficient), s _{L, m+1}And τ _{L, m+1}Be unknown parameter to be estimated, represent time delay, the video stretching of target.

Can be described as through the echoed signal after the direct wave rejecting:

$S^{\′}\left(t\right)+\underset{j=1}{\overset{m+1}{\mathrm{\Σ}}}\underset{\mathrm{\φ}}{\mathrm{\Σ}}{K}_{i,j}x\left(s_{i,j}(t-{\mathrm{\τ}}_{i,j})\right)+\underset{j=1}{\overset{m+1}{\mathrm{\Σ}}}\underset{\stackrel{\‾}{\mathrm{\φ}}}{\mathrm{\Σ}}{K}_{i,j}x\left(s_{i,j}(t-{\mathrm{\τ}}_{i,j})\right)+n\left(t\right)$

Under the uniform dielectric condition, ignore medium and the repeatedly influence of reflection wave, but the effective echoed signal approximate description that then is used for target detection and parameter estimation is:

$r_2\left(t\right)=\underset{i}{\mathrm{\Σ}}{K}_{i,T}x\left(s_{i,T}(t-{\mathrm{\τ}}_{i,T})\right)+n\left(t\right)$

The broadband associative processor is output as:

$\mathrm{WC}(s,\mathrm{\τ})=\sqrt{s}{{\∫r}^{*}}_1\left(s(t-\mathrm{\τ})\right)r_2\left(t\right)\mathrm{dt}$

Under the nonhomogeneous media situation, vertical or horizontal average by the multiple tracks data, exchange correct coupling and parameter for and get steadily and surely and estimate with the cost that is reduced to of vertical or horizontal resolution.

After wide relevant treatment, can obtain three typical tangent and three echoed signals maximum of points (X, Y, Z).One is horizontal section, the shape information of display-object reflecting surface, and a vertical tangent plane and a horizontal tangent plane, vertically the typical data of the typical data of tangent plane and horizontal tangent plane combines and is used for the identification of target shape.The track data of two tangent plane intersection points is represented the typical data of echo, is used for the identification of target material.

2. feature extraction

Feature extraction mainly comprises two parts: be used for target shape identification the vertical and horizontal typical data extraction and be used for the extraction of the typical track data of target Material Identification.

After wide relevant treatment, can obtain X, Y, the Z value at echoed signal maximum of points place, get X, Y value respectively, can obtain corresponding vertical tangent plane and horizontal tangent plane data, get near the maximal value of each track data correspondence of sectional drawing maximal value again, obtain the point of two tangent planes, so just obtained being used for the typical track data of shape recognition.

Part scans the Welch power spectrum and is proved to be effective identification that can be used for the target material, and Welch method spectrum is estimated to take the data sementation windowing process to ask average way again, obtains earlier every section spectrum respectively and estimates, carries out overall average then.Prove that according to Probability Statistics Theory be divided into the K section as if the data that with raw footage are N, every segment length is got M=N/K, independent each other as each segment data, then the variance of Gu Jiing will have only the 1/K of original not segmentation, reach the purpose of consistent Estimation.But if K increases, M reduces, then resolution descends.On the contrary, if K reduces, M increases, though deviation reduces, estimation variance increases.So the value of K and M is suitably chosen in the requirement that must take into account resolution and variance in practice.

The computation process of Welch power Spectral Estimation is as follows: the length of establishing signal s (n) is 512, is divided into the K=7 section, and every segment length is N=128, overlapping 50%.And each subclass added a hanmin window w (n) (n=128).

The Welch power Spectral Estimation is calculated as follows:

$P_w=\frac{1}{\mathrm{UK}}\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{S}_i\left(w\right)S_i^{*}\left(w\right)$

$S_i\left(w\right)=S_i\left(n\right)w\left(n\right)e^{-\frac{2\mathrm{\π}}{m}\mathrm{wn}}$

$U=\frac{1}{m}\underset{n=0}{\overset{m-1}{\mathrm{\Σ}}}{w}^2\left(n\right)$

Fig. 3 is the Welch power spectrum contrast figure of the typical track data of unlike material, and contrast can be seen and exists bigger difference between the three, therefore can be used as the detection of the Material Identification and the target of target.Determine different X, the Y value obtains the vertical tangent plane of correspondence and the typical track data of horizontal tangent plane intersection point, after the Welch power spectrum is handled, can obtain being used for the data of Material Identification then.

3. Classification and Identification

The data that the test of target shape identification of the present invention is adopted are respectively the measurement data at ball and pipe.The method of testing is at first to carry out wide coherent signal at the data of measuring to handle, and obtains horizon-slice map, horizontal sectional drawing and vertical sectional drawing.Carry out target shape identification in conjunction with the typical data in the vertical and horizontal tangent plane.

Through wide relevant treatment, the signal to noise ratio (S/N ratio) of echoed signal has obtained enhancing.Utilize the data of the point that feature extraction obtains to carry out curve and conic fitting, relatively the difference of two squares of twice fitting curve determines that fitting result is straight line or quafric curve.And in conjunction with the vertical and horizontal typical tangent result in the 3-D display, the difference of the distribution shape of the typical track data of two typical tangent of difformity object.As shown in Figure 2,, be shown as two peaks, correspond to ball if two tangent plane data fittings all are secondaries; If one is once, one is secondary, shows that one is that the peak distributes, and one is discontinuous The extreme value distribution, then correspond to pipe.Can realize target shape identification like this.

The present invention adopts the RBF radial basis function neural network to carry out Target Recognition.RBF chooses three layers of feedforward network with single hidden layer, comprises input layer, middle layer and output layer.The input layer number choose sampling number according to the proper vector chosen.Consider the length of useful information in the echoed signal, this sampling number is taken as 128.The selection principle of middle layer number is that 2 times input layer number deducts the output layer number.The output layer number is 1, uses kind---soil, iron and the PVC of 0,1,2 representatives object to be identified respectively according to different application.

Wide relevant treatment result at real data, different X on earth and the target of fetching earth respectively, the Y value, with the different typical track data of correspondence through the Welch power Spectral Estimation, obtain being used for the sample data of Material Identification, contrast can be seen and exists bigger difference between the three, therefore can be used as the detection of the Material Identification and the target of target.The power spectrum characteristic amount is sent into radial basis function RBF neural network trains.The echoed signal maximal value that while obtains by wide relevant treatment at measurement data to be identified.It is corresponding that (X, typical road signal Y) passes through the Welch power Spectral Estimation, and then carries out Classification and Identification by neural network.Scope according to the output valve of network is carried out the automatic identification of target material.When output valve ∈ (0.5,0.5), be judged to be soil; When output valve ∈ (0.5,1.5), be judged to be iron; When output valve ∈ (1.5,2.5), be judged to be PVC; Other output valve is judged other.

As shown in Figure 3.For the training and the identification of the neural network of pseudo-iron pipe and pvc pipe, the output result is a table 1, and reflection Welch power spectrum can be effectively finished identification to underground target material by neural network.

Table 1

Material	The learning sample number	The test specimens given figure	Discrimination
				Iron	????200	????20	????90％
????PVC	????200	????20	????80％
				Soil	????200	????20	????75％

Existing imaging recognition technology of contrast and characteristic variable identification, the present invention can be effectively to difformity, and the buried target of unlike material carries out effectively identification automatically, can reach the practical function of through engineering approaches.Simultaneously from whole performing step as can be known, method of the present invention is easy to realize, thereby provides a technology implementation method for the through engineering approaches of ground penetrating radar.

Claims (1)

1, a kind of ultra-wideband ground-penetrating radar (uw-gpr) automatic target recognition method based on information fusion is characterized in that comprising following concrete steps:

1) data processing: comprise direct wave rejecting and signal filtering, the three-dimensional echo data of ground penetrating radar is carried out the average of horizontal and vertical direction, obtain the average echo data of vertical direction, therefrom select the tie point of second and the 3rd echo to carry out data truncation as truncation points, suppress direct wave, reject the echo data part of front, the echo data of remainder is carried out subsequent treatment as the data that contain signal, ground penetrating radar echo data after blocking is carried out wide relevant treatment, obtain the X at the maximum of points place of three typical tangent and three echoed signals, Y, the Z value;

2) feature extraction: comprise the extraction of the vertical and horizontal typical data that is used for target shape identification and be used for the extraction of the typical track data of target Material Identification, X according to the echoed signal maximum of points place that obtains after the wide relevant treatment, the Y value, obtain corresponding vertical tangent plane and horizontal tangent plane data, get near the maximal value of each track data correspondence of sectional drawing maximal value again, obtain the point of two tangent planes, obtain being used for the characteristic of shape recognition, determine different X, the Y value, obtain the typical track data of corresponding vertical tangent plane and horizontal tangent plane intersection point, after the Welch power spectrum is handled, can obtain being used for the data of Material Identification then;

3) Classification and Identification: the shape recognition characteristic that obtains is carried out curve fitting, the difference of two squares of more different curve correspondences is determined fitting result, utilizes the corresponding different matched curve of difformity target echo signal, and, realize the identification of target shape in conjunction with the sectional drawing demonstration; Utilize radial basis function RBF neural network that the target material is classified, typical track data that will be corresponding with unlike material is through the Welch power Spectral Estimation, obtain being used for the sample data of Material Identification, send into radial basis function RBF neural network and train the funtcional relationship of setting up characteristic quantity and desired value, the data that are used for Material Identification that the previous step feature extraction is obtained realize the automatic identification of target material as the characteristic quantity input neural network; At last the result of target shape identification and Material Identification is carried out information fusion, realize, the identification automatically comprehensively of difformity target unlike material.

Images