A coding feature extraction method for multi-rotor UAV identification

Abstract

The present invention belongs to the technical field of multi-rotor unmanned aerial vehicle target recognition, and specifically relates to a coding feature extraction method for multi-rotor unmanned aerial vehicle recognition. The present invention first obtains a time-frequency spectrum by performing a short-time Fourier transform on the radar echo data sequence of the unmanned aerial vehicle target, performs multiple differential calculations on the time-frequency spectrum in some directions, obtains a high-order difference domain image, and then divides the high-order difference domain image into multiple non-overlapping sub-blocks. The element value of each sub-block is differentiated from the sub-block mean, and a coding sequence is formed according to the positive and negative coding and counterclockwise order of the differential value. The coding sequences of all sub-blocks of the entire image in all directions form a vector, which serves as the high-order difference domain sub-block multi-directional coding feature to complete the recognition of the multi-rotor unmanned aerial vehicle. Since the directional detailed information in the sub-block structure is fully utilized, the local fine features of the relevant target can be extracted from the radar echo data, thereby improving the recognition rate of the target.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-rotor unmanned aerial vehicle (UAV) target recognition, and in particular relates to a coding feature extraction method for multi-rotor UAV recognition. Background Art

[0002] Currently, the projective transformation method is a classic approach for identifying drones. This method primarily uses a training dataset of drone targets to establish a transformation matrix and extract target classification features. However, conventional projective transformation methods can only extract global structural features from a macroscopic perspective, while ignoring local structural features and reducing differences in target details. Therefore, it is expected that the target recognition rate of conventional multi-rotor drone projective transformation methods will be further improved. Summary of the Invention

[0003] The purpose of the present invention is to propose a coding feature extraction method for multi-rotor UAV identification. By performing high-order difference processing on the time-frequency spectrum, dividing it into non-overlapping sub-blocks, and encoding the sub-blocks in multiple directions as identification features, the local structural features reflecting the target directional details can be better extracted, the degree of difference in the target directional details is increased, and thus the recognition rate of the target is improved.

[0004] The technical solution of the present invention is:

[0005] A coding feature extraction method for multi-rotor UAV identification includes the following steps:

[0006] S1. Define the acquired radar echo training data sequence of the multi-rotor drone as an n-dimensional column vector x ij , i=1,2,…g,j=1,2…N i , where i represents the category of the drone, g represents the number of categories, and N i represents the number of training samples of the i-th type of UAV target, then the total number of training samples is

[0007] S2, training sample data x for the i-th type of drone target ij Perform short-time Fourier transform to obtain the time-frequency spectrum S ij :

[0008] S ij =[s ij,km ] K×M

[0009] Among them, s ij,km Represents the time spectrum diagram S ij The elements in, k=1,2,…K, m=1,2,…M, k is the row subscript, indicating the frequency change direction, m is the column subscript, indicating the time change direction, K is S ij The number of rows, M is Sij The number of columns;

[0010] S3. Perform high-order difference processing on the time-frequency spectrum:

[0011] Taking each element as the center, perform differences with the elements adjacent to the right, upper right corner, directly above, and upper left corner to obtain 4 amplitude difference maps:

[0012]

[0013]

[0014]

[0015]

[0016] in, and Respectively represent the image of the first-order difference between the element and the adjacent elements on the right, upper right corner, directly above, and upper left corner, and Respectively represent the difference between the element in the kth row and the mth column and its adjacent elements to the right, upper right, directly above, and upper left;

[0017] Based on this, we get the q>1 order difference domain images in the four directions:

[0018]

[0019]

[0020]

[0021]

[0022] in, and They represent the difference between the elements in the (q-1) order difference domain image and the adjacent elements on the right, upper right corner, directly above, and upper left corner, respectively. and Respectively represent the difference between the element in the kth row and the mth column in the (q-1)-order difference domain image and the adjacent elements on the right, upper right, directly above, and upper left;

[0023] S4, the q-order difference domain images in four directions and Divide into sub-blocks and obtain the sub-block multi-directional coding sequence:

[0024] Use a 3x3 window to filter the image Non-overlapping sliding is performed in , d = 1, 2, 3, 4, and multiple 3x3 sub-blocks are obtained to form a sub-block set

[0025]

[0026]

[0027] l=1,2,…L

[0028] in, Represents a sub-block set The l-th sub-block matrix in, Represents the sub-block matrix In the element, L represents the total number of sub-blocks;

[0029] Calculate the sub-block matrix The mean of all elements in

[0030]

[0031] Encode as follows:

[0032]

[0033] in, yes Corresponding to the encoding, the encoding of each element in the sub-block is composed of a differential encoding sequence of 0 and 1 in the order of rows

[0034]

[0035] S5. Extract sub-block multi-directional coding features:

[0036] The differential multi-directional coding sequence in the same direction Make up a vector

[0037]

[0038] Then all the vectors in the direction The following vectors are formed:

[0039]

[0040] Obtained is the q-order difference domain sub-block multi-directional coding feature vector.

[0041] After obtaining the feature vector, the minimum distance profile classifier can be used to classify and identify the multi-rotor UAV target.

[0042] The beneficial effects of the present invention are as follows: first, the present invention performs a short-time Fourier transform on the radar echo data sequence of the UAV target to obtain a time-frequency spectrum diagram, then performs multiple differential calculations on the time-frequency spectrum diagram in some directions to obtain a high-order differential domain image, and then divides the high-order differential domain image into multiple non-overlapping sub-blocks, and the element value of each sub-block is differentiated from the sub-block mean, and a coding sequence is formed according to the positive and negative coding and counterclockwise order of the differential value. The coding sequences of all sub-blocks of the entire image in all directions constitute a vector, which serves as the multi-directional coding feature of the high-order differential domain sub-block to complete the identification of the multi-rotor UAV. Since the directional detailed information in the sub-block structure is fully utilized, the local fine features of the relevant target can be extracted from the radar echo data, thereby improving the recognition rate of the target. The simulation experimental results of four types of multi-rotor UAVs verify the effectiveness of the method. DETAILED DESCRIPTION

[0043] The following simulation is used to demonstrate the effectiveness and progress of the present invention:

[0044] Four types of UAVs were designed for the simulation experiment: a tri-rotor, a quad-rotor, a hexacopter, and an octo-rotor. Their simulation parameters are shown in Table 1. Simulated radar parameters include: a radar carrier frequency of 24 GHz; a pulse repetition frequency of 100 kHz; a target-to-radar distance of 200 m; and a pitch angle of 10° and an azimuth angle of 30° relative to the radar.

[0045] Table 1 Simulation parameters of four types of UAVs

[0046]

[0047] For each target type, radar echo signals were recorded for 10 seconds and divided into segments of fixed length 0.05 seconds (containing at least one rotation period), with a 50% overlap between segments. Each segment contained 0.05 × 100,000 = 5,000 radar echo sampling data points, for a total of 400 segments per type. Of these 400 segments, 200 were randomly selected as the training dataset, and the remaining 200 as the test dataset. The training dataset for the four target types consisted of 800 segments, and the test dataset consisted of 800 segments. The proposed method was used to extract high-order difference domain sub-block multi-directional encoding features from the selected training datasets and establish a template library. Similarly, high-order difference domain sub-block multi-directional encoding features were extracted from the test samples. A minimum distance classifier was used for classification, achieving an average correct recognition rate of 96% for the four types of multi-rotor drones. The signal-to-noise ratio (SNR) was 10 dB.

Claims

1. A coding feature extraction method for multi-rotor drone identification, characterized in that: The following steps are involved: S1. Define the acquired radar echo training data sequence of the multi-rotor drone as an n-dimensional column vector x ij , i=1,2,…g,j=1,2…N i , where i represents the category of the drone, g represents the number of categories, and N i represents the number of training samples of the i-th type of UAV target, then the total number of training samples is S2, training sample data x for the i-th type of drone target ij Perform short-time Fourier transform to obtain the time-frequency spectrum S ij : S ij =[s ij,km ] K×M Among them, s ij,km Represents the time spectrum diagram S ij The elements in, k=1,2,…K, m=1,2,…M, k is the row subscript, indicating the frequency change direction, m is the column subscript, indicating the time change direction, K is S ij The number of rows, M is S ij The number of columns; S3. Perform high-order difference processing on the time-frequency spectrum: Taking each element as the center, perform differences with the elements adjacent to the right, upper right corner, directly above, and upper left corner to obtain 4 amplitude difference maps: in, and Respectively represent the image of the first-order difference between the element and the adjacent elements on the right, upper right corner, directly above, and upper left corner, and Respectively represent the difference between the element in the kth row and the mth column and its adjacent elements to the right, upper right, directly above, and upper left; Based on this, we get the q>1 order difference domain images in the four directions: in, and They represent the difference between the elements in the (q-1) order difference domain image and the adjacent elements on the right, upper right corner, directly above, and upper left corner, respectively. and Respectively represent the difference between the element in the kth row and the mth column in the (q-1)-order difference domain image and the adjacent elements on the right, upper right, directly above, and upper left; S4, the q-order difference domain images in four directions and Divide into sub-blocks and obtain the sub-block multi-directional coding sequence: Use a 3x3 window to filter the image Non-overlapping sliding is performed in , d = 1, 2, 3, 4, and multiple 3x3 sub-blocks are obtained to form a sub-block set in, Represents a sub-block set The l-th sub-block matrix in, Represents the sub-block matrix In the element, L represents the total number of sub-blocks; Calculate the sub-block matrix The mean of all elements in Encode as follows: in, yes Corresponding to the encoding, the encoding of each element in the sub-block is composed of a differential encoding sequence of 0 and 1 in the order of rows S5. Extract sub-block multi-directional coding features: The differential multi-directional coding sequence in the same direction Make up a vector Then all the vectors in the direction The following vectors are formed: Obtained is the q-order difference domain sub-block multi-directional coding feature vector.

Images