A UAV target feature extraction method based on fuzzy locality preserving projection
The micro Doppler spectrum features of the drone radar echo are extracted by the fuzzy local retention projection method, and the weight matrix is constructed using the fuzzy membership degree, which solves the problem of insufficient recognition performance in the traditional method and achieves a higher drone target recognition rate.
Patent Information
- Application Number
- CN202310439746.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-23
- Publication Date
- 2025-08-22
- Estimated Expiration
- 2043-04-23
AI Technical Summary
Traditional drone recognition methods are based on the fact that optical images are susceptible to interference from light and occlusions, and conventional locally retained projection feature extraction methods cannot fully describe the data distribution, resulting in insufficient recognition performance, especially when heterogeneous targets are approaching.
The method based on fuzzy local reserve projection is adopted to obtain the micro Doppler spectrum through the time-frequency transformation of radar echo, and the weight matrix is constructed using fuzzy membership, and the fuzzy local reserve projection features are extracted to realize the classification recognition of multi-rotor drones and flying birds.
The drone target recognition rate has been improved, especially under low signal-to-noise ratio conditions, which has increased by 8.41% compared to conventional methods, effectively solving the problem of insufficient recognition performance of traditional methods.
Smart Images
Figure BDA0004193483090000022 
Figure BDA0004193483090000027 
Figure BDA0004193483090000031
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) identification, and in particular relates to a UAV target feature extraction method based on fuzzy local preservation projection. Background Art
[0002] Drones are typically "low, slow, and small" (flying at low altitude, slow speed, and small size). Due to their ease of operation and high mobility, they are widely used in a variety of fields, including disaster relief, pesticide spraying, and crop monitoring. However, as drone use becomes increasingly widespread, "illegal" drone incidents, such as disrupting air traffic and intruding into private spaces for eavesdropping, have become a frequent occurrence. These unauthorized drones pose a serious threat to social order and public safety. Therefore, identifying drone types and implementing appropriate countermeasures has become a pressing issue.
[0003] Traditional drone identification methods are primarily based on optical image recognition, but optical image recognition of drones is susceptible to interference from light and obstructions. Radar offers 24-hour observation and immunity to adverse weather conditions, leading to the increasing application of target recognition methods based on radar signals. Extracting micro-motion features from target radar echoes enables the classification and identification of multi-rotor drones. Conventional local-preserving projection feature extraction is a commonly used manifold learning method that aims to maintain the relative distance relationship between target samples in the high-dimensional original space and the low-dimensional feature space. However, this method only characterizes local information based on Euclidean distance and cannot fully describe the distribution of the original samples. Furthermore, in actual data distributions, samples of heterogeneous targets may be very close together, reducing the method's performance. Therefore, the recognition performance of conventional local-preserving projection feature extraction methods has room for further improvement. Summary of the Invention
[0004] To address these issues, the present invention proposes a method for extracting target features from multi-rotor drones and flying birds based on fuzzy local-preserving projections. First, a micro-Doppler spectrum of the target is obtained by performing a time-frequency transformation on the target's radar echo. From this, fuzzy local-preserving projection features are extracted to enable classification and recognition of multi-rotor drones and flying birds. By constructing a weight matrix using the fuzzy membership of samples containing category information, this method more effectively describes the spatial distribution of the data, thereby improving target recognition.
[0005] The technical solution of the present invention is:
[0006] A UAV target feature extraction method based on fuzzy local preservation projection is proposed. The radar transmits a continuous wave signal to the target and defines the fth radar echo received from the cth target as S cf(t), in practical applications, there are usually 6 types of targets, namely c = 1, 2, 3, 4, 5, 6 represent single-rotor, dual-rotor, three-rotor, four-rotor, six-rotor UAV targets and flying bird targets respectively. For the c-th type target, n c radar echoes constitute the training samples, that is, f=1,2,......n c , the total number of training samples is N, The feature extraction method comprises:
[0007] S1. Segment the target's radar echo data and set the overlap rate between segments:
[0008]
[0009] in, Represent the first, second and qth segments of data respectively, where q is the number of segments;
[0010] S2. Segmented radar echo data Perform short-time Fourier transform to obtain the target's time-frequency spectrum p=1~q, so the target's micro-Doppler spectrum matrix for:
[0011]
[0012] in, It is a two-dimensional real number matrix, the row variable m is the time information, the column variable n is the frequency information, and the matrix elements contain the amplitude information corresponding to each time and each frequency;
[0013] Define the signal length as P t , the target Doppler spectrum The dimension is expressed as M×Q,
[0014]
[0015] 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;
[0016] S3. Extract the fuzzy local preservation projection features of the target micro-Doppler spectrum:
[0017] S31, the target's micro-Doppler spectrum The elements in are arranged in columns to form a column vector of size Z, Z = M × Q, and all target column vectors form the sample matrix X:
[0018] X=[x1,x2,...x i ...x j ...x N ]
[0019] Among them, the number of columns of the sample matrix X is the total number of training samples N, and the number of rows is the dimension Z of a single sample, x i and x j are the i-th and j-th Z-dimensional column vectors, also called sample vectors, i, j = 1 to N;
[0020] S32. Construct an adjacency graph:
[0021] ① Calculate the Euclidean distance between any two samples and obtain an N×N symmetric sample distance matrix D. The elements in D are:
[0022]
[0023] Among them, d ij Represents the sample vector x i and the sample vector x j The Euclidean distance between them, Z is the dimension of the sample vector, and Represents the sample vector x i and the sample vector x j The e-th dimension value of , e = 1 to Z;
[0024] ②Because the elements on the diagonal of the sample distance matrix D represent the Euclidean distance between the sample itself and have no practical research significance, they are all set to infinite NAN;
[0025] ③ Sort the sample distance matrix obtained in ② in ascending order according to the distance value of each row or column;
[0026] ④ Select the first k1 samples in each row or column of the arranged distance matrix, obtain the k1 nearest neighbor samples of each sample, and obtain their category information;
[0027] S33. Calculate the elements in the distance weight matrix W:
[0028]
[0029] Among them, w ij is the i-th row and j-th column element of the distance weight matrix W, and σ is the kernel parameter;
[0030] S34. Calculate the fuzzy membership matrix of sample categories
[0031] Calculate the membership m of the i-th sample to the k-th class ik :
[0032]
[0033] Among them, n ikrepresents the number of samples belonging to the kth category among the k1 nearest neighbor samples of the i-th sample, l(i) represents the true category of the i-th sample; m ik The fuzzy membership matrix H = [m ik ] N×c ;
[0034] S35. Calculate the category association between the i-th sample and the j-th sample:
[0035] R ij =m il(j) +m jl(i)
[0036] Among them, l(i) and l(j) represent the true categories of sample i and sample j respectively, and m il(j) represents the fuzzy membership of sample i to class l(j), m jl(i) Represents the fuzzy membership of sample j to class l(i); if samples i and j belong to the same target sample, their sample category association R ij The value is large. If they belong to different categories, then R ij The value is small; the sample category association matrix R is obtained from the category association matrix R = [R ij ] N×N ;
[0037] S36, take (XVX T ) -1 X(VR⊙W)X T The eigenvectors corresponding to the first g smallest non-zero eigenvalues of the matrix form the g-dimensional projection space A; where ⊙ represents the dot product of two matrices, and the matrix V is a diagonal matrix with elements v on its diagonal. ii is the sum of the elements in the corresponding columns or rows of the dot product matrix R⊙W, that is or
[0038] For any sample vector x test Projection to A:
[0039] y test =A T x test
[0040] Among them, y test The fuzzy local projection feature vector of any sample vector is retained and is also the extracted feature.
[0041] The beneficial effect of the present invention is that the method of the present invention uses the micro-Doppler spectrum of the target as the research object, and realizes the classification and recognition of multi-rotor drones and flying bird targets by extracting the fuzzy local preservation projection features of the spectrum. Compared with the local preservation projection method that only constructs the weight matrix based on Euclidean distance information, the method of the present invention uses the sample fuzzy membership to construct the weight matrix. On the one hand, it can more accurately describe the distribution of the data space. At the same time, because the fuzzy membership includes the target category information, it is more conducive to the classification of the target, thereby improving the target recognition rate. Simulation experiments were conducted on five types of targets. The experimental results show that the target recognition rate of the method of the present invention is significantly improved compared with the conventional local preservation projection method, verifying the effectiveness of the method of the present invention. DETAILED DESCRIPTION
[0042] The practicality of the present invention is illustrated below in conjunction with simulation:
[0043] The simulation experiment designed single-rotor, dual-rotor, tri-rotor, quad-rotor, and hexacopter drones, as well as flying birds. The simulated radar parameters included: a carrier frequency of 34.6 GHz; a pulse repetition frequency of 125,000 Hz; a target-to-radar distance of 100 meters; a target scattering coefficient of 1; and radar pitch angles of 10° and 45°.
[0044] UAV parameters: Under initial conditions, the distance between the UAV and the radar is 100m, the UAV's speed is 0, the number of blades on each rotor is 2, the distance from the blade tip to the rotor's rotation center is 0, the distance from each rotor's rotation center to the center of the UAV is 0.8m, and the observation time is 15s. The specific parameters of the five types of UAVs are shown in Table 1:
[0045] Table 1 Simulation parameters of five types of rotary-wing UAVs
[0046]
[0047] Bird parameters: The initial distance between the bird and the radar is 100m, the bird's flapping frequency is 14Hz, the half wingspan is 0.3m, the maximum wing angle is 75°, and the observation time is 15s.
[0048] The target's radar echo is segmented, with each segment having a frame length of 0.01s and a 50% overlap between segments. The segmented echo signals are subjected to a short-time Fourier transform (SFT) to obtain the corresponding micro-Doppler spectra. 200 micro-Doppler spectra are randomly sampled from each class, for a total of 1200 micro-Doppler spectra as the data set. From each of the 200 micro-Doppler spectra, 60 are selected as training samples, and the remaining are used as testing samples, resulting in a training set of 360 micro-Doppler spectra and a test set of 840 micro-Doppler spectra. After preprocessing the sample dataset using a distance-preserving method for dimensionality reduction, the training and test sets are projected onto a generalized locality-preserving projection space to obtain their respective generalized locality-preserving projection features. Finally, classification and recognition are performed using the nearest neighbor classification criterion.
[0049] Under a signal-to-noise ratio of 0 dB, the conventional local preservation projection feature method and the proposed method were used to identify five types of rotary-wing drones and one type of flying bird. The average recognition rate was calculated from 20 experimental results, and the overall average recognition results for the six types of targets are shown in Table 2.
[0050] Table 2 Correct recognition rates of the two methods for five types of rotorcraft drones and one type of flying bird targets (%)
[0051]
[0052]
[0053] From the results in Table 2, it can be seen that under low signal-to-noise ratio conditions, the average recognition rate of the method of the present invention is improved by 8.41% compared with the conventional local preservation projection method, which has a significant improvement effect, indicating that the method of the present invention is practical.
Claims
1. A UAV target feature extraction method based on fuzzy local preservation projection is used. A radar transmits a continuous wave signal to the target and defines the fth radar echo received from the cth target as S cf (t), where c = 1, 2, 3, 4, 5, 6, respectively defining 6 types of targets as single-rotor, dual-rotor, tri-rotor, quad-rotor, hexagonal drone targets and flying birds. For the cth type of target, n c radar echoes constitute the training samples, that is, f=1,2,......n c , the total number of training samples is N, It is characterized by: The feature extraction method comprises: S1. Segment the target's radar echo data and set the overlap rate between segments: in, Represent the first, second and qth segments of data respectively, where q is the number of segments; S2. Segmented radar echo data Perform short-time Fourier transform to obtain the target's time-frequency spectrum p=1~q, get the target micro-Doppler spectrum matrix for: in, It is a two-dimensional real number matrix, the row variable m is the time information, the column variable n is the frequency information, and the matrix elements contain the amplitude information corresponding to each time and each frequency; Define the signal length as P t , the target Doppler spectrum The dimension is expressed as M×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. Extract the fuzzy local preservation projection features of the target micro-Doppler spectrum: S31, the target's micro-Doppler spectrum The elements in are arranged in columns to form a column vector of size Z, Z = M × Q, and all target column vectors form the sample matrix X: X=[x1,x2,...x i ...x j ...x N ] Among them, the number of columns of the sample matrix X is the total number of training samples N, and the number of rows is the dimension Z of a single sample, x i and x j are the i-th and j-th Z-dimensional column vectors, also called sample vectors, i, j = 1 to N; S32. Construct an adjacency graph: ① Calculate the Euclidean distance between any two samples and obtain an N×N symmetric sample distance matrix D. The elements in D are: Among them, d ij Represents the sample vector x i and the sample vector x j The Euclidean distance between them, Z is the dimension of the sample vector, and Represents the sample vector x i and the sample vector x j The e-th dimension value of , e = 1 to Z; ②Because the elements on the diagonal of the sample distance matrix D represent the Euclidean distance between the sample itself and have no practical research significance, they are all set to infinite NAN; ③ Sort the sample distance matrix obtained in ② in ascending order according to the distance value of each row or column; ④ Select the first k1 samples in each row or column of the arranged distance matrix, obtain the k1 nearest neighbor samples of each sample, and obtain their category information; S33. Calculate the elements in the distance weight matrix W: Among them, w ij is the i-th row and j-th column element of the distance weight matrix W, and σ is the kernel parameter; S34. Calculate the membership m of the i-th sample to the k-th class ik : Among them, n ik represents the number of samples belonging to the kth category among the k1 nearest neighbor samples of the i-th sample, l(i) represents the true category of the i-th sample; m ik The fuzzy membership matrix H = [m ik ] N×c ; S35. Calculate the category association between the i-th sample and the j-th sample: R ij =m il(j) +m jl(i) Among them, l(i) and l(j) represent the true categories of sample i and sample j respectively, and m il(j) represents the fuzzy membership of sample i to class l(j), m jl(i) Represents the fuzzy membership of sample j to class l(i); if samples i and j belong to the same target sample, their sample category association R ij The value is large. If they belong to different categories, then R ij The value is small; the sample category association matrix R is obtained from the category association matrix R = [R ij ] N×N ; S36, take (XVX T ) -1 X(VR⊙W)X T The eigenvectors corresponding to the first g smallest non-zero eigenvalues of the matrix form the g-dimensional projection space A; where ⊙ represents the dot product of two matrices, and the matrix V is a diagonal matrix with elements v on its diagonal. ii is the sum of the elements in the corresponding columns or rows of the dot product matrix R⊙W, that is or For any sample vector x test Projection to A: and test =A T x test Among them, y test The projected feature vector is retained locally for the fuzzy portion of any sample vector, i.e., the extracted feature.