Airborne radar moving target detection method based on space-time image feature learning

Abstract

The invention discloses an airborne radar moving target detection method based on space-time image feature learning, and the method comprises the following steps: constructing an airborne phased array radar echo signal model, and carrying out the simulation generation of a large amount of space-time echo data containing different target signals, clutter signals and noise signals; performing range pulse compression processing on the space-time echo data to obtain a space-time echo signal vector; performing sparse recovery processing on the space-time echo signal vector by using a sparse Bayesian algorithm to obtain a space-time image corresponding to each piece of space-time echo signal data; according to different feature distribution of targets and clutters on the space-time images, performing target labeling on each space-time image to obtain a labeling data set; training the deep learning network to obtain a deep learning target detection network capable of extracting target space-time image features; according to the method, the target and the clutter are distinguished on the space-time image based on the deep learning technology, so that airborne phased array radar moving target detection is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of airborne radar target detection, and more particularly to an airborne radar moving target detection method based on space-time image feature learning. Background Art

[0002] In actual non-uniform environments, airborne space-time adaptive processing (STAP) radars usually have difficulty obtaining sufficient independent and identically distributed (IID) training range unit data, and cannot accurately estimate the clutter plus noise covariance matrix (CNCM) for clutter suppression and moving target detection. Even if an accurate CNCM can be obtained, the STAP filter will weaken the power of slow-moving targets to a certain extent, affecting the detection performance. Therefore, when the speed of the target is less than the minimum detectable velocity (MDV) of the airborne radar system, even the optimal STAP cannot effectively detect the target.

[0003] In recent years, people have proposed moving target detection methods for airborne STAP radar based on image features. This type of method does not need to estimate CNCM, and directly uses machine learning methods (such as region growing method, bootstrapping method, etc.) to distinguish targets and clutter in space-time (ST) images to effectively detect slow-moving targets. However, there are two problems with such existing methods: ① The quality of the space-time image obtained by Fourier transform is poor (low resolution and high sidelobes). The minimum variance distortion less response (MVDR) spectrum estimation algorithm is used to obtain high-quality space-time images. A certain number of range units are required, but the target may only be located in a specific range unit, not several continuous range units; ② The target is modeled as a point in the space-time image, making more advanced machine learning methods (such as deep learning) inapplicable or inapplicable, and unable to further improve the detection performance. In practice, the target may have more complex space-time image features. For example, different parts of the moving target may have different Doppler frequencies (such as the target has micro-Doppler characteristics), and it is no longer a point in the space-time image. Therefore, how to better realize airborne STAP radar moving target detection based on image features has become an important research topic in this field. Summary of the invention

[0004] The purpose of the present invention is to provide an airborne radar moving target detection method based on space-time image feature learning, which distinguishes targets and clutter on space-time images based on a deep learning target detection network, thereby realizing airborne radar moving target detection.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for detecting moving targets by airborne radar based on space-time image feature learning comprises the following steps:

[0007] Construct an airborne phased array radar echo signal model and simulate and generate a large amount of space-time echo data containing different target signals, clutter signals and noise signals;

[0008] Performing range pulse compression processing on the space-time echo data to obtain a space-time echo signal vector;

[0009] Performing sparse recovery processing on the space-time echo signal vector using a sparse Bayesian algorithm to obtain a space-time image corresponding to each space-time echo signal data;

[0010] According to the feature distribution of different targets on the space-time image, target labeling is performed on each space-time image to obtain a labeling data set;

[0011] Based on the labeled data set, a deep learning network is trained to obtain a deep learning target detection network capable of extracting target space-time image features;

[0012] The trained deep learning target detection network is used to perform target detection on actual airborne radar space-time images.

[0013] Furthermore, the method for constructing the airborne phased array radar echo signal model is specifically as follows:

[0014] The aircraft flies along the x-axis at a constant speed V and an altitude of H. The radar uniform planar array consists of M×N array elements with a spacing of d=λ / 2; where: λ is the signal wavelength;

[0015] The angle between the radar uniform plane array and the flight direction of the carrier aircraft is α, and the azimuth and elevation angles of the clutter scattering block or target are θ and φ respectively; suppose the radar transmits a linear frequency modulated pulse signal with a carrier frequency of f c =c / λ, c is the speed of light, pulse width is T p , the modulation frequency is γ, a coherent processing interval contains K pulses, and the pulse repetition frequency is f r =1 / T r ; Starting from the first pulse emitted by the radar, t is the time, t k =(k-1)T r For slow time, For fast time, k = 1, 2, ..., K;

[0016] Assume that the airborne radar uses two-dimensional beamforming when transmitting signals and one-dimensional beamforming when receiving signals to synthesize the radar uniform planar array into a uniform linear array with N equivalent array elements;

[0017] Divide the clutter into L distance units, and assume that each distance unit contains N c clutter scattering blocks are uniformly distributed in the azimuth range. For the i-th clutter scattering block in the l-th range unit, let the azimuth, elevation and space cone angle be θ respectively. i , and ψ l,i , where: l = 1, 2, ..., L; i = 1, 2, ..., N c .

[0018] Furthermore, the simulation generates a large number of expressions for space-time echo data containing different targets, clutter and noise, specifically:

[0019] Given a moving target consisting of P scattering points, the three-dimensional position of the target body is [x 0 ,y 0 ,z 0 ], let the azimuth, elevation and space cone angle of the pth target scattering point be θ p , and ψ p , then the k-th space-time echo data received by the n-th equivalent array element is expressed as:

[0020]

[0021] In the formula, and Represent target signal, clutter signal and noise signal respectively;

[0022] The target signal is expressed as:

[0023]

[0024] In the formula, ξ p represents the amplitude of the pth target scattering point, τ 0 =2R 0 / c, is the distance between the target and the radar, P n,p (t k ) is:

[0025]

[0026] In the formula, d n=(n-1)d,v r,p is the radial velocity of the pth target scattering point relative to the airborne radar;

[0027] The expression of the clutter signal is:

[0028]

[0029] In the formula, ξ l,i represents the amplitude of the ith clutter scattering block in the lth range unit, τ l =2R l / c, is the lth clutter distance, P n,l,i (t k ) is:

[0030]

[0031] In the formula,

[0032] Furthermore, the space-time echo data is subjected to range pulse compression processing to obtain a space-time echo signal vector, specifically:

[0033] The space-time echo data is subjected to range pulse compression processing to obtain the NK×1 space-time echo signal vector of the lth range unit, which is expressed as:

[0034]

[0035] In the formula, (·) T represents transpose, represents the noise vector, is the target vector, is the clutter vector;

[0036] The target vector and clutter vector The expressions are:

[0037]

[0038] In the formula, and are the time domain steering vector and spatial domain steering vector of the pth target scattering point, respectively, expressed as:

[0039]

[0040] In the formula, f s p =dcosψ p / λ,f t p=2(Vcosψ' p +v r,p )T r / λ, and is the time domain steering vector and space domain steering vector of the ith clutter scattering block in the lth range unit, expressed as:

[0041]

[0042] In the formula, f s l,i =dcosψ l,i / λ,f t l,i =Vcosψ' l,i T r / λ;

[0043] Furthermore, a sparse representation model is established, specifically:

[0044] Discretize the spatial frequency and Doppler frequency into N s =κ s N and N d =κ d K grid points, where κ s >1 and κ d >1 represents the spatial frequency scale factor and the Doppler frequency scale factor, then the space-time echo signal vector of the lth range unit is approximately expressed as:

[0045]

[0046] In the formula, is the complex amplitude corresponding to the qth grid point, is the complex magnitude vector of all grid points, q=1,2,...,N s N d , and are the time domain steering vector and space domain steering vector corresponding to the qth grid point, respectively, and f t q and f s q are the qth normalized Doppler frequency and spatial frequency, respectively, The dictionary of steering vectors is empty.

[0047] Based on formula (10), the following sparse representation model is established:

[0048]

[0049] In the formula, ||·|| 0 and ||·|| 2L represents the vector 0 Norm and L 2 norm, and ε represents the noise energy.

[0050] Furthermore, the sparse recovery process is performed on the space-time echo signal vector using the sparse Bayesian algorithm to obtain the expression of the space-time image corresponding to each space-time echo data, which is specifically:

[0051] Convex relaxation is performed on formula (11), which can be expressed as:

[0052]

[0053] Where ρ represents the regularization parameter, ||·|| 1 L represents the vector 1 norm;

[0054] Assume that the noise energy follows the power σ 2 The unknown complex Gaussian distribution, then the space-time echo signal vector x l The Gaussian likelihood function of is expressed as:

[0055]

[0056] Let the complex amplitude vector ξ of all grid points l Obeying the complex Gaussian prior distribution, we can get:

[0057]

[0058] In the formula, Γ l =diag(γ l ), Represents the unknown variance parameters corresponding to each grid point on the space-time plane;

[0059] Combined with the space-time echo signal vector x l The Gaussian likelihood function and complex Gaussian prior distribution of all grid points are used to obtain the complex amplitude vector ξ l The posterior probability density is:

[0060]

[0061] Based on equation (15), the sparse Bayesian algorithm is used to transform the complex amplitude vector ξ of all grid points into l Transformed into the unknown variance parameter γ corresponding to each grid point on the space-time plane l The space-time image corresponding to each space-time echo data can be obtained as I l =Mat(γ l ), where Mat(·) represents matrix processing.

[0062] Furthermore, for the unknown variance parameter γ corresponding to each grid point on the space-time plane l The estimation methods include:

[0063] Using the known γ l and σ 2 Calculate ξ l The posterior probability density of l The expectation and variance of

[0064] Utilize l The expected and variance updates of γ l and σ 2 , and repeat until the o+1th iteration satisfies The iteration is stopped when o = 0, 1, ..., O-1, η represents the iteration threshold, so as to obtain the l Estimates; specifically:

[0065] At the o+1th iteration, (x l ,ξ l ) is expressed as:

[0066]

[0067] At the o+1th iteration, the space-time echo signal vector x l The Gaussian likelihood function can be expressed as:

[0068]

[0069] Using the γ obtained in the oth iteration l and σ 2 , calculate the complex amplitude vector ξ of all grid points l The posterior probability density at the o+1th iteration is expressed as:

[0070]

[0071] In formula (18) and for The expectation and variance of are expressed as:

[0072]

[0073] The complex amplitude vector ξ for all grid points l The posterior probability density at the o+1th iteration is maximized, specifically:

[0074] By maximizing Expected estimate of and It is expressed as:

[0075]

[0076] First, using the known Estimate unknown variance parameters

[0077] because and It can be obtained from formula (13), then

[0078]

[0079] In formula (21),

[0080] right The derivative of is expressed as:

[0081]

[0082] get The update method is:

[0083]

[0084] Substitution The expectation and variance of , we get:

[0085]

[0086] Using known calculate The expression is:

[0087]

[0088] By simplifying, we can get:

[0089]

[0090] For σ 2 Taking the derivative, we get:

[0091]

[0092] Let the above formula be 0, and we get the power σ 2 The update method is:

[0093]

[0094] Furthermore, the characteristic distribution of the different targets on the space-time image is as follows:

[0095] For the target, the expressions of the spatial frequency and normalized Doppler frequency of the pth scattering point are:

[0096] f s p =cosψ p / 2, f t p =2(Vcosψ' p +v r,p ) / (λf r ) (29)

[0097] Assuming that the airborne radar system cannot distinguish different target scattering points at the same distance or angle, it is assumed that the spatial cone angles of different target scattering points are the same, thus obtaining:

[0098] f s p =f s 0 =cosψ 0 / 2, f t p =2v r,p / (λf r )+F(f s 0 ) (30)

[0099] In the formula, ψ 0 represents the spatial cone angle of the target, F(·) represents the function that converts the target spatial frequency into Doppler frequency, and v r,p It can be expressed as:

[0100] v r,p ≈-[x 0 v x,p +y 0 v y,p +(z 0 -H)v z,p ] / R(0)(31)

[0101] In the formula, [v x,p ,v y,p ,v z,p ] represents the three-dimensional velocity of the pth target scattering point;

[0102] It can be seen from formula (29-31) that: if the target has only one scattering point or all scattering points have the same three-dimensional velocity, the target is a point in the space-time image; if the target has multiple scattering points with large Doppler frequencies, which are sufficient to be detected and distinguished by the airborne radar system, the target in the space-time image is a line segment composed of a set of points with the same spatial frequency but different Doppler frequencies.

[0103] Furthermore, the characteristic distribution of the clutter on the space-time image is as follows:

[0104] The relationship between the normalized Doppler frequency and spatial frequency of the ith clutter scattering block in the lth range unit that generates the clutter signal is expressed as:

[0105] Assume d = λ / 2, β = 4V / (λf r ),but

[0106]

[0107] It can be seen from formula (32) that when α = 0, that is, the radar array is a straight side-view array, the space-time image of the clutter is a straight line with a slope of β; when α = 90° and β ≠ 1, that is, the radar array is a forward-view array, the space-time image of the clutter is a semi-regular ellipse; when α = 90° and β = 1, the space-time image of the clutter is a semicircle; when 0 < α < 90°, that is, the radar array is an oblique-view array, the space-time image of the clutter is a circle whose size and shape are determined by β and the angle α and Determined semi-oblique ellipse.

[0108] Furthermore, based on the labeled data set, a deep learning network is trained to obtain a deep learning target detection network capable of extracting target space-time image features, specifically:

[0109] The labeled dataset is divided into a training set and a test set, 70% of which is used for training and 30% for testing;

[0110] The training set is used to train a deep learning target detection network for extracting target features;

[0111] The test set is used to evaluate the performance of the deep learning target detection network.

[0112] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0113] The present invention proposes a method for distinguishing targets from clutter on space-time images based on a deep learning target detection network, thereby realizing airborne STAP radar moving target detection. The method can obtain space-time images with higher resolution and lower sidelobe level. There is no need to process multiple range unit data, and the range unit to be measured can be directly processed. After obtaining the target space-time image, the deep learning target detection network is used to distinguish the target from clutter for detection, which is expected to break through the MDV limitation of the existing STAP method and realize the rapid detection of slow-moving targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0115] The following is a further description of the airborne radar moving target detection method based on space-time image feature learning of the present invention in conjunction with the accompanying drawings;

[0116] Figure 1 It is a schematic diagram of the overall process of the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0117] Figure 2 It is the airborne STAP radar spatial geometry model in the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0118] Figure 3 It is the space-time image of the clutter signal under different conditions in the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0119] Figure 4 It is the space-time image of the target signal under different conditions in the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0120] Figure 5 It is a schematic diagram of the YOLO-v2 network structure in the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0121] Figure 6 It is a moving target detection result diagram of an airborne STAP radar in the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0122] Figure 7 It is a precision / recall curve diagram of the YOLO-v2 network under different conditions in the airborne radar moving target detection method based on space-time image feature learning provided by the present invention;

[0123] Figure 8 It is a comparison diagram between the airborne radar moving target detection method based on space-time image feature learning provided by the present invention and the optimal STAP. DETAILED DESCRIPTION

[0124] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0125] In order to better understand the purpose, structure and function of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0126] like Figure 1 As shown, the present invention provides an airborne radar moving target detection method based on space-time image feature learning, comprising the following steps:

[0127] Construct an airborne phased array radar echo signal model and simulate and generate a large amount of space-time echo data containing different target signals, clutter signals and noise signals;

[0128] Performing range pulse compression processing on the space-time echo data to obtain a space-time echo signal vector;

[0129] Establishing a sparse representation model, and using a sparse Bayesian algorithm to perform sparse recovery processing on the space-time echo signal vector to obtain a space-time image corresponding to each space-time echo signal data;

[0130] According to the different characteristic distributions of the target and clutter on the space-time image, each space-time image is labeled to obtain a labeled data set;

[0131] Based on the labeled data set, a deep learning network is trained to obtain a deep learning target detection network capable of extracting target space-time image features;

[0132] The trained deep learning target detection network is used to detect moving targets on actual airborne radar space-time images.

[0133] The method for constructing the airborne phased array radar echo signal model is specifically as follows:

[0134] The geometric structure of airborne STAP radar moving target detection is as follows: Figure 2 As shown in the figure, the aircraft flies along the x-axis at a constant speed V and at an altitude of H. The radar uniform planar array (UPA) consists of M×N array elements with a spacing of d=λ / 2 (λ represents the signal wavelength). The angle between the UPA and the aircraft's flight direction is α. The azimuth and elevation angles of the clutter scattering patch (CSP) or target are θ and φ respectively. Assume that the radar transmits a linear frequency modulated pulse signal with a carrier frequency of f c =c / λ, c represents the speed of light, and the pulse width is T p , the modulation frequency is γ, a coherent processing interval (CPI) contains K pulses, and the pulse repetition frequency (PRF) is f r =1 / Tr . From the time the radar transmits the first pulse, t is the time, t k =(k-1)T r For slow time, For fast time, k = 1, 2, ..., K.

[0135] Assume that the airborne radar uses two-dimensional beamforming when transmitting signals and one-dimensional beamforming when receiving signals to synthesize the UPA into a uniform linear array (ULA) with N equivalent array elements. Divide the clutter into L range units, assuming that each range unit contains N c CSPs are uniformly distributed over the azimuth range. For the i-th (i=1,2,...,N)-th range cell in the l-th (l=1,2,...,L) range cell, c )CSP, assuming that its azimuth, elevation and space cone angles are θ i , and ψ l,i .

[0136] Given a moving target consisting of P scattering points, the three-dimensional position of the target body is [x 0 ,y 0 ,z 0 ], let the azimuth, elevation and space cone angle of the pth target scattering point be θ p , and ψ p , then the k-th space-time echo data received by the n-th equivalent array element is expressed as:

[0137]

[0138] In the formula, and Represent target signal, clutter signal and noise signal respectively;

[0139] The target signal is expressed as:

[0140]

[0141] In the formula, ξ p represents the amplitude of the pth target scattering point, τ 0 =2R 0 / c, is the distance between the target and the radar, P n,p (t k ) is:

[0142]

[0143] In the formula, dn =(n-1)d,v r,p is the radial velocity of the pth target scattering point relative to the airborne radar;

[0144] The expression of the clutter signal is:

[0145]

[0146] In the formula, ξ l,i represents the amplitude of the ith clutter scattering block in the lth range unit, τ l =2R l / c, is the lth clutter distance, P n,l,i (t k ) is:

[0147]

[0148] In the formula,

[0149] The space-time echo data is subjected to range pulse compression processing to obtain a space-time echo signal vector, specifically:

[0150] The space-time echo data is subjected to range pulse compression processing to obtain the NK×1 space-time echo signal vector of the lth range unit, which is expressed as:

[0151]

[0152] In the formula, (·) T represents transpose, represents the noise vector, is the target vector, is the clutter vector;

[0153] The target vector and clutter vector The expressions are:

[0154]

[0155] In the formula, and are the time domain steering vector and spatial domain steering vector of the pth target scattering point, respectively, expressed as:

[0156]

[0157] In the formula, f s p =dcosψ p / λ,f t p=2(Vcosψ' p +v r,p )T r / λ, and is the time domain steering vector and spatial domain steering vector of the ith clutter scattering block in the lth range unit, expressed as:

[0158]

[0159] In the formula, f s l,i =dcosψ l,i / λ,f t l,i =Vcosψ' l,i T r / λ.

[0160] Establish a sparse representation model, specifically:

[0161] Discretize the spatial frequency and Doppler frequency into N s =κ s N and N d =κ d K grid points, where κ s >1 and κ d >1 represents the spatial frequency scale factor and the Doppler frequency scale respectively, then the space-time echo signal vector of the lth unit that does not contain the target range is approximately expressed as:

[0162]

[0163] In the formula, is the complex amplitude corresponding to the qth grid point, is the complex magnitude vector of all grid points, q=1,2,...,N s N d , and are the time domain steering vector and space domain steering vector corresponding to the qth grid point, respectively, and f t q and f s q are the qth Doppler frequency and the qth spatial frequency respectively, The dictionary of steering vectors is empty.

[0164] Based on formula (10), the following sparse representation model is established:

[0165]

[0166] In the formula, ||·|| 0 and ||·|| 2L represents the vector 0 Norm and L 2 norm, and ε represents the noise energy.

[0167] The sparse recovery process of the space-time echo signal vector using the sparse Bayesian algorithm to obtain the space-time image corresponding to each space-time echo data is specifically:

[0168] Convex relaxation is performed on formula (11), which can be expressed as:

[0169]

[0170] Where ρ represents the regularization parameter, ||·|| 1 L represents the vector 1 norm;

[0171] Assume that the noise energy follows the power σ 2 The unknown complex Gaussian distribution, then the space-time echo signal vector x l The Gaussian likelihood function of is expressed as:

[0172]

[0173] Let the complex amplitude vector ξ of all grid points l Obeying the complex Gaussian prior distribution, we can get:

[0174]

[0175] In the formula, Γ l =diag(γ l ), Represents the unknown variance parameters corresponding to each grid point on the space-time plane;

[0176] Combined with the space-time echo signal vector x l The Gaussian likelihood function and complex Gaussian prior distribution of all grid points are used to obtain the complex amplitude vector ξ l The posterior probability density is:

[0177]

[0178] Based on formula (15), the complex amplitude vector ξ of all grid points is transformed into l Transformed into the unknown variance parameter γ corresponding to each grid point on the space-time plane l The estimation of each space-time echo data can be expressed as I l =Mat(γ l ), where Mat(·) represents matrix processing.

[0179] It should be noted that: Based on formula (15), the sparse Bayesian algorithm will l The estimate of γ l To determine γ l The number, position and value of non-zero elements in can be processed using the Expectation Maximization (EM) algorithm. l After that, the space-time image of the lth distance unit can be expressed as I l =Mat(γ l ). If the lth range unit does not contain the target, the resulting space-time image only contains clutter.

[0180] For the unknown variance parameter γ corresponding to each grid point on the space-time plane l The estimation methods include:

[0181] Step E, using the known γ l and σ 2 Calculate ξ l The posterior probability density of l The expectation and variance of

[0182] Step M, using ξ l The expected and variance updates of γ l and σ 2 ;

[0183] Repeat steps E and M until the o+1th iteration satisfies The iteration is stopped when o = 0, 1, ..., O-1, η represents the iteration threshold, so as to obtain the l Estimates; specifically:

[0184] At the o+1th iteration, (x l ,ξ l ) is expressed as:

[0185]

[0186] At the o+1th iteration, the space-time echo signal vector x l The Gaussian likelihood function can be expressed as:

[0187]

[0188] Using the γ obtained in the oth iteration l and σ 2 , calculate the complex amplitude vector ξ of all grid points l The posterior probability density at the o+1th iteration is expressed as:

[0189]

[0190] In formula (18) and for The expectation and variance of are expressed as:

[0191]

[0192] The complex amplitude vector ξ for all grid points l The posterior probability density at the o+1th iteration is maximized, specifically:

[0193] By maximizing Expected estimate of and It is expressed as:

[0194]

[0195] First, using the known Estimate unknown variance parameters

[0196] because and It can be obtained from formula (13), then

[0197]

[0198] In formula (21),

[0199] right The derivative of is expressed as:

[0200]

[0201] get The update method is:

[0202]

[0203] Substitution The expectation and variance of , we get:

[0204]

[0205] Using known calculate The expression is:

[0206]

[0207] By simplifying, we can get:

[0208]

[0209] For σ 2 Taking the derivative, we get:

[0210]

[0211] Let the above formula be 0, and we get the power σ 2 The update method is:

[0212]

[0213] The characteristic distribution of the target on the space-time image is as follows:

[0214] For the target, the expressions of the spatial frequency and normalized Doppler frequency of the pth scattering point are:

[0215] f s p =cosψ p / 2, f t p =2(Vcosψ' p +v r,p ) / (λf r ) (29)

[0216] Assuming that the airborne radar system cannot distinguish different target scattering points at the same distance or angle, it is assumed that the spatial cone angles of different target scattering points are the same, thus obtaining:

[0217] f s p =f s 0 =cosψ 0 / 2, f t p =2v r,p / (λf r )+F(f s 0 ) (30)

[0218] In the formula, ψ 0 represents the spatial cone angle of the target, F(·) represents the function that converts the target spatial frequency into Doppler frequency, and v r,p It can be expressed as:

[0219] v r,p ≈-[x 0 v x,p +y 0 v y,p +(z 0 -H)v z,p ] / R(0)(31)

[0220] In the formula, [v x,p ,v y,p ,v z,p ] represents the three-dimensional velocity of the pth target scattering point;

[0221] It can be seen from formula (29-31) that: if the target has only one scattering point or all scattering points have the same three-dimensional velocity, the target is a point in the space-time image; if the target has multiple scattering points with large Doppler frequencies, which are sufficient to be detected and distinguished by the airborne radar system, the target in the space-time image is a line segment composed of a set of points with the same spatial frequency but different Doppler frequencies.

[0222] The characteristic distribution of the clutter on the space-time image is as follows:

[0223] The relationship between the normalized Doppler frequency and spatial frequency of the ith clutter scattering block in the lth range unit that generates the clutter signal is expressed as:

[0224] Assume d = λ / 2, β = 4V / (λf r ),but

[0225]

[0226] It can be seen from formula (32) that when α = 0, that is, the radar array is a straight side-view array, the space-time image of the clutter is a straight line with a slope of β; when α = 90° and β ≠ 1, that is, the radar array is a forward-view array, the space-time image of the clutter is a semi-regular ellipse; when α = 90° and β = 1, the space-time image of the clutter is a semicircle; when 0 < α < 90°, that is, the radar array is an oblique-view array, the space-time image of the clutter is a circle whose size and shape are determined by β and the angle α and Determined semi-oblique ellipse.

[0227] Assume that the aircraft altitude of the airborne radar is H = 3 km, the number of array elements is N = 20, the number of pulses is K = 20, the array element spacing is d = 0.1 m, the carrier wavelength is λ = 0.2 m, and the pulse repetition frequency is f r =2kHz, clutter distance is 30km, clutter-to-noise ratio CNR=60dB, the range of spatial frequency and Doppler frequency is [-0.5,0.5], and the number of grids is N d =N s =100, N c = 361 CSPs are evenly distributed in the azimuth [0,π] range, the scattering coefficient obeys the complex Gaussian distribution, and the power is determined by the CNR. The clutter space-time images obtained based on the constructed sparse representation model and sparse Bayesian algorithm under four different conditions are as follows Figure 3 It can be found that the space-time image characteristics (line, semi-oblique ellipse, semicircle) presented by clutter under different conditions are consistent with the theoretical analysis.

[0228] Assume that the airborne radar carrier speed V = 100m / s, the non-positive side view angle α = 0, the target distance is 30km, the target azimuth is 90°, and the other parameters are the same as Figure 3 The target space-time images obtained based on the constructed sparse representation model and sparse Bayesian algorithm under two different conditions are as follows: Figure 4 As shown in the figure, the target in the left figure contains only one scattering point, with a signal-to-noise ratio of 30dB and a radial velocity of 20m / s; the target in the right figure contains 16 scattering points, one of which has a signal-to-noise ratio of 30dB and a radial velocity of 20m / s, representing the main body of the target, and the signal-to-noise ratio of the other 15 scattering points varies randomly in the range of [15,30]dB, and the radial velocity varies randomly in the range of [0,40]m / s. It can be found that the space-time image features (i.e., points and line segments) presented by the target under two different conditions are consistent with the theoretical analysis.

[0229] contrast Figure 3 and Figure 4 It can be found that clutter and target have different space-time image characteristics. Clutter obeys straight line or curve distribution on the space-time image, while the target obeys point or line segment distribution. Therefore, the difference in space-time image characteristics between target and clutter can be used to distinguish them, realizing the detection of moving targets by airborne STAP radar.

[0230] According to the different characteristic distributions of the target and clutter on the space-time image, each space-time image is labeled to obtain a labeled data set, which is specifically:

[0231] The labeled dataset is divided into a training set and a test set, 70% of which is used for training and 30% for testing;

[0232] The training set is used to train a YOLO-v2 target detection network for extracting target features;

[0233] The test set is used to evaluate the performance of the YOLO-v2 target detection network.

[0234] A YOLO-v2 target detection network is mainly composed of two sub-networks, namely the feature extraction sub-network and the target detection sub-network. The former is a deep convolutional neural network, and the latter is a relatively small convolutional neural network, which consists of a series of convolutional layers, batch normalization layers and ReLU layers, followed by transformation layers and output layers. The present invention selects a typical deep convolutional neural network ResNet-18 for feature extraction, and the network layer after the last ReLU layer is replaced by the target detection sub-network of YOLO-v2, and finally the image target detection network structure of the present invention is as follows: Figure 5 As shown in the figure, the network layers in the dotted box belong to the feature extraction subnetwork, and the other network layers belong to the target detection subnetwork.

[0235] The present invention verifies the application performance of the proposed intelligent target detection method in airborne STAP radar through simulation experiments, and compares and analyzes it with the optimal STAP method. The simulation parameters are: aircraft height H = 3km, number of array elements N = 20, array element spacing d = 0.1m, carrier wavelength λ = 0.2m, number of pulses K = 20, pulse repetition frequency f r = 2000 Hz, the range of spatial frequency and Doppler frequency is [-0.5, 0.5], the number of grids is N s =N d =100, CNR = 60dB. Based on different carrier speeds V and non-positive side view angles α, four different situations are considered, as shown in Table 1. Figure 3 Corresponding.

[0236] Table 1 Simulation parameters of airborne STAP radar

[0237]

[0238] To generate clutter signal, let N c = 361 CSPs are uniformly distributed in the azimuth range [0,π], their scattering coefficients obey complex Gaussian distribution, and the power is determined by CNR. To generate the target signal, the number of scattering points is randomly selected from the range [11,21], one of which is defined as the target body, its signal-to-noise ratio is fixed at 30dB, and the normalized Doppler frequency and spatial frequency are randomly selected from the range [-0.2, 0.2], and the signal-to-noise ratio of other scattering points is randomly selected in the range [15, 30] dB, and the spatial frequency is equal to The normalized Doppler frequency is For each simulation case shown in Table 1, 200 radar space-time signal vectors are generated according to formula (2), and then a total of 800 clutter and target space-time images are generated using the SBL algorithm. After the targets are labeled, these space-time images are randomly divided into two data sets, 70% for training and 30% for testing.

[0239] The present invention provides a technical solution for training the YOLO-v2 network model:

[0240] Assume that the network input size is 672×672, the number of classes is 1, the number of anchor boxes is 5, the batch size is 20, the number of iterations is 100, and the initial learning rate is 10 -4 , use the Adam optimizer to train YOLO-v2. Use the trained YOLO-v2 to detect moving targets on the test dataset. Some results are shown below: Figure 6As shown. It can be found that under different radar parameters (aircraft speed and non-positive side view angle) and target parameters (spatial and Doppler frequency), the proposed method can always effectively detect the target. For the first two sub-images on the right, the target and clutter overlap in the space-time image due to the very small speed. The existing STAP method is usually unable to detect this type of target, while the proposed method utilizes the difference in space-time image features between the target and clutter, and makes it possible by using a target detector based on deep learning technology.

[0241] In order to quantitatively evaluate the target detection performance of the proposed method, the average precision (AP) index of the YOLO-v2 network under four different simulation conditions is calculated to represent the correct classification ability (precision) and the ability to detect all targets (recall) of the target detector. Figure 7 The precision / recall (PR) curves of the YOLO-v2 network under four different test conditions are given, with average accuracies of 96.18%, 94.07%, 91.17% and 95.81%, respectively, indicating that the proposed method has high target detection accuracy under different conditions.

[0242] It should be noted that the proposed method is compared with the optimal STAP method assuming that CNCM is known. Assuming the carrier speed V = 100m / s, the non-frontal angle α = 0, keeping the spatial frequency of the target body at 0, and making its normalized Doppler frequency change linearly in the range of [-0.4, 0.4], the comparison results are as follows: Figure 8 As shown. For STAP and the proposed method, SCNR loss and detection confidence are used as performance metrics, respectively. For the latter, 20 target and clutter space-time images are generated at each normalized Doppler frequency, and the average value of YOLO detection confidence is calculated. It can be seen that: when the normalized Doppler frequency of the target is close to 0, the SCNR loss of the STAP method is large, indicating that it has poor slow target detection performance; in contrast, regardless of the target speed, the proposed method can always detect moving targets with a high confidence (≥0.85), indicating that it is not limited by the minimum detectable speed of the target.

[0243] The specific steps of the moving target detection method proposed in the present invention are as follows: ① Based on equation (6), a large amount of airborne STAP radar echo data is obtained by simulation, including targets, clutter and noise with different parameters; ② For each echo data, the sparse Bayesian algorithm is used to solve equation (11), and matrix processing is performed to obtain a two-dimensional space-time image corresponding to each echo data; ③ According to the characteristic distribution of the target on the space-time image (such as Figure 3④ Based on the labeled space-time image set, deep neural networks (such as YOLO, FasterRCNN, etc.) are trained to extract target features and realize intelligent target detection.

[0244] The present invention proposes an intelligent method for distinguishing targets and clutter on space-time images based on deep learning technology to achieve airborne STAP radar moving target detection. Considering that targets and clutter only account for a small part in space-time images (space-time images), the proposed method uses a sparse Bayesian Learning (SBL) algorithm to generate space-time images. Compared with the Fourier transform method, the SBL algorithm can obtain space-time images with higher resolution and lower sidelobe levels; compared with the MVDR algorithm, the SBL algorithm directly processes the distance unit to be measured without processing multiple distance unit data. After obtaining the target space-time image, the present invention uses a typical deep neural network YOLO (You Only LookOnce) to learn the image features of the target, distinguish it from clutter for detection. Through these processes, the method proposed by the present invention is expected to break through the MDV limitation of the existing STAP method and achieve rapid detection of slow-moving targets.

[0245] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for detecting moving targets by airborne radar based on space-time image feature learning, characterized in that: The following steps are involved: Construct an airborne phased array radar echo signal model and simulate and generate a large amount of space-time echo data containing different target signals, clutter signals and noise signals; Performing range pulse compression processing on the space-time echo data to obtain a space-time echo signal vector; Establishing a sparse representation model, and using a sparse Bayesian algorithm to perform sparse recovery processing on the space-time echo signal vector to obtain a space-time image corresponding to each space-time echo signal data; According to the different characteristic distributions of the target and clutter on the space-time image, each space-time image is labeled to obtain a labeled data set; Based on the labeled data set, a deep learning network is trained to obtain a deep learning target detection network capable of extracting target space-time image features; The trained deep learning target detection network is used to detect moving targets on actual airborne radar space-time images.

2. The airborne radar moving target detection method based on space-time image feature learning according to claim 1 is characterized in that: The method for constructing the airborne phased array radar echo signal model is specifically as follows: The aircraft flies along the x-axis at a constant speed V and an altitude of H. The radar uniform planar array consists of M×N array elements with a spacing of d=λ / 2; where: λ is the signal wavelength; The angle between the radar uniform plane array and the aircraft's flight direction is α, and the azimuth and elevation angles of the clutter scatterer or target are θ and Assume that the radar transmits a linear frequency modulated pulse signal with a carrier frequency of f c =c / λ, c is the speed of light, pulse width is T p , the modulation frequency is γ, a coherent processing interval contains K pulses, and the pulse repetition frequency is f r =1 / T r ; Starting from the first pulse emitted by the radar, t is the time, t k =(k-1)T r For slow time, For fast time, k = 1, 2, ..., K; Assume that the airborne radar uses two-dimensional beamforming when transmitting signals and one-dimensional beamforming when receiving signals to synthesize the radar uniform planar array into a uniform linear array with N equivalent array elements; Divide the clutter into L distance units, and assume that each distance unit contains N c clutter scattering blocks are uniformly distributed in the azimuth range. For the i-th clutter scattering block in the l-th range unit, let the azimuth, elevation and space cone angle be θ respectively. i , and ψ l,i , where: l = 1, 2, ..., L; i = 1, 2, ..., N c .

3. The airborne radar moving target detection method based on space-time image feature learning according to claim 1 is characterized in that: The simulation generates a large amount of space-time echo data containing different targets, clutter and noise, specifically: Given a moving target consisting of P scattering points, the three-dimensional position of the target body is [x0, y0, z0], and the azimuth, pitch angle and spatial cone angle of the pth target scattering point are θ p , and ψ p , then the k-th space-time echo data received by the n-th equivalent array element is expressed as: In the formula, and Represent target signal, clutter signal and noise signal respectively; The target signal is expressed as: In formula (2), ξ p represents the amplitude of the pth target scattering point, τ0=2R0 / c, is the distance between the target and the radar, P n,p (t k ) is: In the formula, d n =(n-1)d,v r,p is the radial velocity of the pth target scattering point relative to the airborne radar; The expression of the clutter signal is: In the formula, ξ l,i represents the amplitude of the ith clutter scattering block in the lth range unit, τ l =2R l / c, is the lth clutter distance, P n,l,i (t k ) is: In the formula, 4. The airborne radar moving target detection method based on space-time image feature learning according to claim 1 is characterized in that: The space-time echo data is subjected to range pulse compression processing to obtain a space-time echo signal vector, specifically: The space-time echo data is subjected to range pulse compression processing to obtain the NK×1 space-time echo signal vector of the lth range unit, which is expressed as: In the formula, (·) T represents transpose, represents the noise vector, is the target vector, is the clutter vector; The target vector and clutter vector The expressions are: In the formula, and are the time domain steering vector and spatial domain steering vector of the pth target scattering point, respectively, expressed as: In the formula, f s p =dcosψ p / λ,f t p =2(Vcosψ' p +v r,p )T r / λ, and is the time domain steering vector and spatial domain steering vector of the ith clutter scattering block in the lth range unit, expressed as: where f s l,i = d cos ψ l,i / λ, f t l,i = V cos ψ' l,i T r / λ.

5. The airborne radar moving target detection method based on space-time image feature learning according to claim 1 is characterized in that: The method for establishing the sparse representation model is specifically as follows: Discretize the spatial frequency and Doppler frequency into N s =κ s N and N d =κ d K grid points, where κ s >1 and κ d >1 represents the spatial frequency scale factor and the Doppler frequency scale respectively, then the space-time echo signal vector of the lth unit that does not contain the target range is approximately expressed as: In the formula, is the complex amplitude corresponding to the qth grid point, is the complex magnitude vector of all grid points, q=1,2,...,N s N d , and are the time domain steering vector and space domain steering vector corresponding to the qth grid point, respectively, and f t q and f s q are the qth Doppler frequency and the qth spatial frequency respectively, is a dictionary of space-time directional vectors; Based on formula (10), the following sparse representation model is established: Where ||·||0 and ||·||2 represent the L0 norm and L2 norm of the vector respectively, and ε represents the noise energy.

6. The airborne radar moving target detection method based on space-time image feature learning according to claim 1 is characterized in that: The sparse recovery process of the space-time echo signal vector using the sparse Bayesian algorithm to obtain the space-time image corresponding to each space-time echo data is specifically: Convex relaxation is performed on formula (11), which can be expressed as: In the formula, ρ represents the regularization parameter, ||·||1 represents the L1 norm of the vector; Assume that the noise energy follows the power σ 2 The unknown complex Gaussian distribution, then the space-time echo signal vector x l The Gaussian likelihood function of is expressed as: Let the complex amplitude vector ξ of all grid points l Obeying the complex Gaussian prior distribution, we can get: In the formula, Γ l =diag(γ l ), Represents the unknown variance parameters corresponding to each grid point on the space-time plane; Combined with the space-time echo signal vector x l The Gaussian likelihood function and complex Gaussian prior distribution of all grid points are used to obtain the complex amplitude vector ξ l The posterior probability density is: Based on formula (15), the complex amplitude vector ξ of all grid points is transformed into l Transformed into the unknown variance parameter γ corresponding to each grid point on the space-time plane l The estimation of each space-time echo data can be expressed as I l =Mat(γ l ), where Mat(·) represents matrix processing.

7. The airborne radar moving target detection method based on space-time image feature learning according to claim 5 is characterized in that: For the unknown variance parameter γ corresponding to each grid point on the space-time plane l The estimation methods include: Using the known γ l and σ 2 Calculate ξ l The posterior probability density of l The expectation and variance of Utilize l The expected and variance updates of γ l and σ 2 , and repeat until the o+1th iteration satisfies The iteration is stopped when o = 0, 1, ..., O-1, η represents the iteration threshold, so as to obtain the l Estimates; specifically: At the o+1th iteration, (x l ,ξ l ) is expressed as: At the o+1th iteration, the space-time echo signal vector x l The Gaussian likelihood function can be expressed as: Using the γ obtained in the oth iteration l and σ 2 , calculate the complex amplitude vector ξ of all grid points l The posterior probability density at the o+1th iteration is expressed as: In formula (18) and for The expectation and variance of are expressed as: The complex amplitude vector ξ for all grid points l The posterior probability density at the o+1th iteration is maximized, specifically: By maximizing Expected estimate of and It is expressed as: First, using the known Estimate unknown variance parameters because and It can be obtained from formula (13), then Formula (21) right The derivative of is expressed as: get The update method is: Substitution The expectation and variance of , we get: Using known calculate The expression is: By simplifying, we can get: For σ 2 Taking the derivative, we get: Let the above formula be 0, and we get the power σ 2 The update method is:

8. The method for detecting moving targets by airborne radar based on space-time image feature learning according to claim 6, characterized in that: The characteristic distribution of the target on the space-time image is as follows: For the target, the expressions of the spatial frequency and normalized Doppler frequency of the pth scattering point are: f s p =cosψ p / 2,f t p =2(Vcosψ' p +v r,p ) / (λf r ) (29) Assuming that the airborne radar system cannot distinguish different target scattering points at the same distance or angle, it is assumed that the spatial cone angles of different target scattering points are the same, thus obtaining: f s p =f s 0 =cosψ0 / 2,f t p =2v r,p / (λf r )+F(f s 0 ) (30) Where ψ0 represents the spatial cone angle of the target, F(·) represents the function that converts the target spatial frequency into Doppler frequency, and v r,p It can be expressed as: v r,p ≈-[x0v x,p +y0v y,p +(z0-H)v z,p ] / R(0) (31) In the formula, [v x,p ,v y,p ,v z,p ] represents the three-dimensional velocity of the pth target scattering point; It can be seen from formula (29-31) that: if the target has only one scattering point or all scattering points have the same three-dimensional velocity, the target is a point in the space-time image; if the target has multiple scattering points with large Doppler frequencies, which are sufficient to be detected and distinguished by the airborne radar system, the target in the space-time image is a line segment composed of a set of points with the same spatial frequency but different Doppler frequencies.

9. The airborne radar moving target detection method based on space-time image feature learning according to claim 6 is characterized in that: The characteristic distribution of the clutter on the space-time image is as follows: The relationship between the normalized Doppler frequency and spatial frequency of the ith clutter scattering block in the lth range unit that generates the clutter signal is expressed as: Let d = λ / 2, β = 4V / (λf r ), then It can be seen from formula (32) that when α = 0, that is, the radar array is a straight side-view array, the space-time image of the clutter is a straight line with a slope of β; when α = 90° and β ≠ 1, that is, the radar array is a forward-view array, the space-time image of the clutter is a semi-regular ellipse; when α = 90° and β = 1, the space-time image of the clutter is a semicircle; when 0 < α < 90°, that is, the radar array is an oblique-view array, the space-time image of the clutter is a circle whose size and shape are determined by β and the angle α and Determined semi-oblique ellipse.

10. The airborne radar moving target detection method based on space-time image feature learning according to claim 1, characterized in that: Based on the labeled data set, a deep learning network is trained to obtain a target detection network capable of extracting target space-time image features, specifically: The labeled dataset is divided into a training set and a test set, 70% of which is used for training and 30% for testing; The training set is used to train a deep learning target detection network for extracting target features; The test set is used to evaluate the performance of the deep learning target detection network.

Images