A feature extraction method for multi-rotor unmanned aerial vehicle and flying bird target

Abstract

The present application belongs to the technical field of multi-rotor unmanned aerial vehicle target recognition, and particularly relates to a feature extraction method for multi-rotor unmanned aerial vehicle and flying bird targets. The present application firstly performs distance-preserving dimension reduction preprocessing on the target micro-Doppler spectrum, and then extracts locally preserving orthogonal optimization projection features from the preprocessed spectrum data, so as to realize classification and recognition of multi-rotor unmanned aerial vehicle and flying bird targets. Since the distance-preserving dimension reduction method not only reduces the dimension size of the target spectrum and avoids "dimension disaster", but also retains the original distance information between targets, the distance weight coefficient can better reflect the distribution structure of the target data. At the same time, the locally preserving orthogonal optimization projection vector has mutual orthogonality, which can reduce the redundancy in the target features and reduce the overlap between the target feature regions, thereby improving the correct classification rate of the unmanned aerial vehicle and flying bird targets. The experimental results verify the effectiveness of the method of the present application.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-rotor unmanned aerial vehicle target recognition, and particularly relates to a feature extraction method for multi-rotor unmanned aerial vehicle and bird targets. BACKGROUND

[0002] In recent years, multi-rotor unmanned aerial vehicles are widely used in agriculture, industry, military and other fields due to the advantages of low flight height, strong mobility and fast cruising speed. However, many illegal events of using unmanned aerial vehicles have occurred due to the characteristics of low cost, easy purchase, small size and easy concealment. Therefore, it is of great significance to accurately identify the type of unmanned aerial vehicles and effectively counterattack to ensure national security, air route safety and urban security.

[0003] Small multi-rotor unmanned aerial vehicles and birds are typical "low, slow and small" targets, which have the characteristics of low observability, multiple target types and small radar scattering cross section. In addition, birds and unmanned aerial vehicles have similar shapes and motion characteristics, and the detection environment is complex, and the radar echo signal is weak, which causes great difficulty in classification and identification of unmanned aerial vehicle and bird targets.

[0004] At present, the linear subspace method is a relatively effective method for identifying unmanned aerial vehicles and bird targets. The conventional linear subspace method assumes that the sample data distribution structure of the target is globally linear, and the target features are obtained by projecting the sample data into a global Euclidean space. Therefore, the conventional linear subspace method can only extract the global structure features of the target. However, research shows that the local structure information existing in the sample data reflects the detailed difference information between targets, which is more conducive to the classification and identification of targets. For example, the manifold learning method finds the embedded low-dimensional manifold structure in high-dimensional data to obtain the local structure features of the target, thereby improving the recognition rate. However, the projection vectors of these methods are not orthogonal, resulting in large redundancy and overlap of the extracted features, which reduces the recognition performance. Therefore, these methods need to be further improved. SUMMARY

[0005] The purpose of the present application is to provide a feature extraction method for multi-rotor unmanned aerial vehicles and bird targets. First, the distance preserving dimension reduction method is used to preprocess the target micro-Doppler spectrum, and then the locally preserving orthogonal optimization projection features are extracted, so as to realize the classification and identification of multi-rotor unmanned aerial vehicles and bird targets. The preprocessing method in this method not only reduces the dimension size of the target spectrum and avoids the "dimension disaster", but also preserves the distance information between targets required for subsequent construction of the weight matrix. At the same time, the locally preserving orthogonal optimization projection vector is orthogonal, which reduces the redundancy in the features, reduces the overlap of the target feature distribution area, and finally improves the correct recognition rate of the target.

[0006] The technical scheme of the present application is as follows:

[0007] A feature extraction method for multi-rotor unmanned aerial vehicles and bird targets, after the radar transmits continuous wave signals to the target, defines S cf (t) is the fth radar echo of the cth target, where c takes values 1-6, and the 6th target is defined as single-rotor, double-rotor, three-rotor, four-rotor, six-rotor unmanned aerial vehicle target and bird target in turn, f = 1, 2,... n c , n c is the number of echo samples of the cth target, and the total number of echo samples is defined as The method comprises the following steps:

[0008] S1, segmenting the radar echo data S cf (t), and setting an overlap rate between segments and segments:

[0009]

[0010] Where q is the number of segments, respectively represent the 1st segment, the 2nd segment,... the qth segment in the radar echo data S cf (t) ;

[0011] S2, performing short-time Fourier transform on the radar echo segment to obtain the time-frequency spectrum of the target Taking the amplitude of the time-frequency spectrum to form the micro-Doppler spectrum matrix of the target

[0012]

[0013] Where m and n are discrete points of time and frequency, is a two-dimensional real matrix, the row variable is the time information, the column variable is the frequency information, and the matrix element contains the amplitude information corresponding to each time and each frequency;

[0014] If the signal length is P t , then the dimension of the target Doppler spectrum can be represented as M x Q,

[0015]

[0016] Where M is the length of the short-time Fourier transform window function, and L is the number of sliding points of the window function;

[0017] S3, arranging the elements in the target micro-Doppler spectrum in column arrangement to form a column vector with size M x Q, and arranging all target column vectors to form a sample matrix X:

[0018] X = [x1, x2,... xN

[0019] Among them, the number of columns of the sample matrix X is the total number of samples N, and x1, x2,...... x N are the first, second, and Nth column vectors in the sample matrix respectively, and the number of rows is the dimension M×Q of the sample column vector;

[0020] Calculate the transition matrix B, where the element b in the i-th row and j-th column of B ij is:

[0021]

[0022] where d ij is the distance between sample i and sample j:

[0023] d ij = ||x i - x j ||2

[0024] Take the first r (<N) largest non-zero eigenvalues λ1, λ2,...... λ r of the transition matrix B and the corresponding eigenvectors u1, u2,...... u r [[ID=�5]], and calculate the matrix Z:

[0025]

[0026] where Z is the sample matrix after dimensionality reduction preprocessing Z = [z1, z2,...... z N , Λ is a diagonal matrix, and the elements on the diagonal are λ1, λ2,...... λ r , and U is the eigenvector matrix composed of u1, u2,...... u r ;[[ID=4ͅ]]

[0027] S4. Calculate the distance weight matrix W, where the element wij in the i-th row and j-th column of W ij is:

[0028]

[0029] where h is a real constant determined by experiments;

[0030] Calculate the matrix L p [[ID=ȶ]]:

[0031] L p = D - W

[0032] where the matrix D is a diagonal matrix, and the elements on the diagonal are the sums of the elements in the corresponding columns or rows of the distance weight matrix W, that is

[0033] Take (ZDZ​T ) -1 ZL p Z T The eigenvector corresponding to the smallest nonzero eigenvalue of the matrix is called a1, and it constitutes the matrix A (1) = [a1];

[0034] The total scatter matrix S is calculated g :

[0035]

[0036] where I is the identity matrix with diagonal elements of 1, e is a column vector with all elements of 1, and v is the mean column vector of all samples:

[0037]

[0038] The eigenvalue decomposition is performed on the matrix [I-(ZDZ T ) -1 S g A (1) [H (1) ] -1 [A (1) ] T S g (ZDZ T ) -1 ZL p Z T ], and the eigenvector corresponding to the smallest nonzero eigenvalue is taken as a2, where:

[0039] H (1) = [A (1) ] T S g (ZDZ T ) -1 S g A (1)

[0040] The vector a2 is orthogonal to a1 and constitutes the matrix A (2) :

[0041] A (2) = [a1, a2]

[0042] The eigenvalue decomposition is performed on the matrix [I-(ZDZ T ) -1 S g A (k-1) [H (k-1) ] -1 [A (k-1) ] T S g (ZDZ T ) -1 ZLp Z T ] Matrix minimum non-zero eigenvalue corresponding eigenvector, as a k , wherein:

[0043] A (k-1) =[a1,a2,......a k-1 ]

[0044] [H (k-1) ]=[A (k-1) ] T S g (ZDZ T ) -1 S g [A (k-1) ]

[0045] Wherein, a1, a2,.....a k-1 (k-1) mutually orthogonal vectors obtained by solving;

[0046] The k mutually orthogonal vectors a1, a2,.....a k-1 , a k Form a matrix A:

[0047] A=[a1,a2......a k ]

[0048] The obtained matrix A is a locally orthogonal optimization projection matrix, for any dimensionality reduction preprocessed target spectrum column vector z test Project to A:

[0049] y test =A T z test

[0050] Then y test Corresponding to the eigenvector of any target spectrum, defined as a locally orthogonal optimization projection eigenvector.

[0051] The beneficial effects of the present application are that the present application firstly carries out distance preserving dimensionality reduction preprocessing to the target micro-Doppler spectrum, then extracts locally orthogonal optimization projection features from the preprocessed spectrum data, realizes the classification and recognition of multi-rotor unmanned aerial vehicle and bird targets. Since the distance preserving dimensionality reduction method not only reduces the dimension size of the target spectrum, avoids the "dimension disaster", but also can retain the original distance information between targets, so that the distance weight coefficient can better reflect the distribution structure of target data. At the same time, the locally orthogonal optimization projection vector has mutual orthogonality, which can reduce the redundancy of target features and reduce the overlap between target feature regions, thereby improving the correct classification rate of unmanned aerial vehicle and bird targets. The experimental results verify the effectiveness of the method of the present application. DETAILED DESCRIPTION

[0052] The effectiveness and progress achieved by the present application are demonstrated below in connection with simulations:

[0053] The simulation experiment designed single-rotor, double-rotor, triple-rotor, quad-rotor, six-rotor unmanned aerial vehicles and bird targets. The simulation radar parameters include: carrier frequency of 34.6 GHz; pulse repetition frequency of 125000 Hz; the distance between the target and the radar is 100 m; the target scattering coefficient is 1; the radar's pitch angle is 10°, and the azimuth angle is 45°.

[0054] Unmanned aerial vehicle parameters: under the initial condition, the distance between the unmanned aerial vehicle and the radar is 100 m, the motion speed of the unmanned aerial vehicle is 0, the number of blades of each rotor is 2, the distance from the blade tip to the rotor rotation center is 0, the distance from the rotor rotation center to the unmanned aerial vehicle center is 0.8 m, and the observation time is 15 s. The specific parameters of the five types of unmanned aerial vehicles are shown in Table 1:

[0055] Table 1 Simulation parameters of five types of rotor unmanned aerial vehicles

[0056]

[0057]

[0058] Bird parameters: the initial distance between the bird and the radar is 100 m, the flapping frequency of the bird is 14 Hz, the half-wing span is 0.3 m, the maximum wing angle is 75°, and the observation time is 15 s.

[0059] The radar echo of the target is segmented, each segment has a frame length of 0.01 s, and the overlap between segments is 50%. The segmented echo signal is subjected to short-time Fourier transform to obtain the corresponding micro-Doppler spectrum. 200 micro-Doppler spectra are randomly selected from each class, and a total of 1200 micro-Doppler spectra are used as a data set. From each of the 200 micro-Doppler spectra, 60 are selected as training samples for each class, and the remaining are used as test samples. A total of 360 micro-Doppler spectrum samples are obtained for the training set, and 840 micro-Doppler spectrum samples are obtained for the test set. First, the sample data set is preprocessed by dimensionality reduction using the distance preserving method, and then the training set and the test set are projected into the generalized local preserving projection space to obtain their respective generalized local preserving projection features. Finally, classification and recognition are performed according to the nearest neighbor classification criterion.

[0060] Under the condition of signal-to-noise ratio of -5 dB, the principal component feature extraction method, the linear discriminant feature extraction method, the principal component fusion local preserving projection feature method, and the method of the present application are used to identify five types of rotor unmanned aerial vehicles and one type of bird target. To avoid randomness, 20 experimental results are taken as the average recognition rate, and the recognition results are shown in Table 2.

[0061] Table 2 Correct recognition rates (%) of four methods for five types of rotor unmanned aerial vehicles and one type of bird targets

[0062]

[0063] From the results in Table 2, it can be seen that, in the case of low signal-to-noise ratio, the average recognition rate of the method of the present application is increased by 10.32%, 18.47%, and 7.66% respectively compared with the principal component feature extraction method, the linear discriminant feature extraction method, and the principal component fusion local reserve projection feature method, and has a large improvement effect.

Claims

1. A feature extraction method for multi-rotor unmanned aerial vehicles and bird targets, wherein a radar defines S cf (t) is the fth radar echo of the cth target, where c takes values from 1 to 6, and the six targets are defined in turn as single-rotor, double-rotor, triple-rotor, quad-rotor, hexa-rotor unmanned aerial vehicle targets and bird targets, and f = 1, 2,... n c , n c is the number of echo samples of the cth target, and the total number of echo samples is defined as characterized in that comprising the steps of: S1, to the radar echo data S cf (t) segmenting, with an overlap rate between segments: where q is the number of segments, respectively represent radar echo data S cf 1st, 2nd,... qth segments in (t) S2, segmenting the radar return performing a short-time Fourier transform to obtain a time-frequency spectrum of the target taking the magnitude of the time-frequency spectrum to form a micro-Doppler spectrogram matrix of the target where m, n are discrete points in time and frequency, respectively, is a two-dimensional real matrix, with row variable being time information and column variable being frequency information, and the matrix elements contain the amplitude information corresponding to each time and each frequency; If the signal length is P t then the dimension of the target Doppler profile may be represented as M x Q, where M is the length of the short-time Fourier transform window function and L is the number of sliding points of the window function; S3. Obtain the target's micro-Doppler spectrum. The elements in the matrix are arranged column-wise to form a column vector of size M×Q. The sample matrix X is composed of all the target column vectors. X = [x1, x2,... x N ] where the number of columns of the sample matrix X is the total number of samples N, x1, x2,... x N are the 1st, 2nd, and Nth column vectors of the sample matrix, respectively, and the number of rows is the dimension of the sample column vectors M x Q; The transition matrix B is computed, where the element b ij is the i-th row and j-th column of B where d ij is the distance between sample i and sample j: d ij = x i - x j2 Take the first r < N largest non-zero eigenvalues λ1, λ2,... λ r The corresponding eigenvectors u1, u2,... u r , Calculate the matrix Z: Wherein, Z is the sample matrix after dimensionality reduction preprocessing Z = [z1, z2,... z N ], Λ is a diagonal matrix, the elements on the diagonal are λ1, λ2,... λ r , respectively, U is a feature vector matrix composed of u1, u2,... u r ; S4, compute a distance weight matrix W, where the element w ij is: where h is a real constant; Computing the matrix L p : L p = D - W where the matrix D is a diagonal matrix with the diagonal elements being the sum of the elements in the corresponding column or row of the distance weight matrix W, i.e. or Take (ZDZ T ) -1 ZL p Z T The eigenvector corresponding to the smallest non-zero eigenvalue of the matrix is called a1, and the matrix A (1) = [a1]; Computing the total scatter matrix S g : where I is an identity matrix with 1 on the diagonal, e is a column vector with all elements equal to 1, and v is a mean column vector of all samples: to [I - (ZDZ T ) -1 S g A (1) [H (1) ] -1 [A (1) ] T S g (ZDZ T ) -1 ZL p Z T ] matrix, take the minimum non-zero eigenvalue corresponding to the eigenvector, as a2, wherein: H (1) = [A (1) ] T S g (ZDZ T ) -1 S g A (1) Vector a2 is orthogonal to a1 and forms the matrix A (2) : A (2) = [a1, a2] Take [I-(ZDZ T ) -1 S g A (k-1) [H (k-1) ] -1 [A (k-1) ] T S g (ZDZ T ) -1 ZL p Z T ] matrix minimum non-zero eigenvalue corresponding to the eigenvector, as a k , wherein: A (k-1) = [a1, a2,..., a k-1 ] [H (k-1) ] = [A (k-1) ] T S g (ZDZ T ) -1 S g [A (k-1) ] wherein a1, a2,... a k-1 are (k-1) mutually orthogonal vectors that have been solved for; k mutually orthogonal vectors a1, a2,... ak are formed k-1 , a k form a matrix A: A = [a1, a2...an] T k ] The obtained matrix A is a locally preserving orthogonal optimized projection matrix, for any dimension-reduced pre-processed target spectral pattern column vector z test Projection to A: y test = A T z test Then y test is the feature vector corresponding to any target spectrum.

Images