A classification method of active target velocity based on time-frequency analysis and Hough transform

Abstract

The present invention discloses a classification method for active target speed based on time-frequency analysis and Hough transform. First, the echo signal received by the active sonar is subjected to time-frequency analysis; then the point-line duality characteristic of the Hough transform is used to map the echo line on the time-frequency plane to a point; then the Hough matrix is ​​mapped, extended and intercepted; finally, the convolutional neural network is used to classify the speed. The method can classify the target speed more accurately and enhance the classification and recognition capability of the active target.

Description

Technical Field

[0001] The invention relates to a classification method of active target speed based on time-frequency analysis and Hough transform, and is particularly suitable for speed classification and recognition of active targets. Background Art

[0002] Signal detection and target recognition are important tasks of detection equipment such as sonar and radar. Active sonar or radar emits a specific form of sound waves or electromagnetic waves, and performs signal detection and feature extraction on the reflected echo of the target, thereby achieving the positioning, classification and tracking of active targets.

[0003] The transmitting signals of active sonar are mainly divided into two types: single-frequency signals and broadband signals. Single-frequency signals have a good effect on measuring target speed, but the time resolution of single-frequency signals is poor, and the measurement of target distance and scale cannot achieve ideal results. Commonly used broadband signals, such as linear frequency modulation signals, can make up for the shortcomings of single-frequency signals in this regard. However, it can be seen from the ambiguity function that due to the poor frequency resolution, broadband signals cannot meet the actual needs for measuring target speed, especially in the presence of noise and reverberation interference. In addition, in more cases, what we need is the category of target speed, such as whether the target is close or far away from the sonar, and whether the movement speed is high or low. Therefore, how to improve the classification and recognition ability of active targets, especially the classification of active target speed when sonar transmits broadband signals, is one of the hot issues currently being studied. Summary of the invention

[0004] The purpose of the present invention is to provide a classification method for active target speed based on time-frequency analysis and Hough transform, which effectively improves the accuracy of target speed classification when sonar transmits broadband signals, thereby improving the classification and recognition ability of active targets.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A classification method for active target speed based on time-frequency analysis and Hough transform comprises the following steps:

[0007] (1) Setting a speed range of the active target, dividing the speed of the active target into several categories within this speed range, and extracting speed values ​​in each category with equal differences;

[0008] (2) Using the known target scale information and the set target velocity value, according to the transmission signal parameters, the bright spot model and the Doppler principle, the echo signal of the simulation training set is obtained;

[0009] (3) performing time-frequency analysis on the echo signal obtained in step (2);

[0010] (4) Perform Hough transform and scale transform on the time-frequency plane;

[0011] (5) Map, extend and intercept the Hough matrix and save it as a grayscale image; thereby constructing a training set, and each grayscale image in the training set is labeled with a speed category;

[0012] (6) Use convolutional neural network for training to obtain the network parameters with the highest accuracy;

[0013] (7) After the echo signal to be classified is denoised by wavelet, the same processing as above is performed and stored as a grayscale image as the input of the neural network to obtain the speed category label, thereby classifying the echo signal by speed.

[0014] In step (1), the velocity range of the active target is v min ~v max , close is a positive number, far is a negative number, within this speed range, the speed is divided into N categories, the upper and lower speed limits of each category are vel1 and vel2 respectively, in each category, M speed values ​​are equally sampled, respectively Therefore, there are N×M speed values ​​in total.

[0015] In the step (2), the transmission signal is selected as a linear frequency modulation signal, and the pulse width of the transmission signal is T p seconds, the period is T seconds, the starting frequency is f0 Hz, the bandwidth is B Hz, and the frequency modulation slope is The sampling frequency of the active sonar system is f s Hertz, the speed of sound waves in water is c meters per second; the size of the target is known to be L meters, according to the bright spot model, using the formula Get the echo signal, where w is a random integer less than or equal to 500, s v =sin(2πf0′t+πk′t 2 ); When the radial velocity of the active target is v m / s, the pulse width of the echo signal is The starting frequency is The bandwidth is The frequency modulation slope is t is in [0,T p ′ ] s The time point of Hertz sampling; therefore, in each speed category, the active target speed v i , i = 0, 1, ..., M-1 can get M echo signal samples s_train(t) i , i=0,1,…,M-1; then there are N×M training set echo signal samples in total for N velocity categories.

[0016] In step (3), according to the pseudo-Wigner-Weir distribution principle, the formula Perform time-frequency analysis on the echo signal s_train(t), where h(t) is the Hamming window function with a window length of N / 4, N = T*f s .

[0017] In step (4), the coordinates of each point on the time-frequency plane are (x, y), and the line on the time-frequency plane is mapped to a point in the (θ, ρ) plane using the polar coordinate transformation formula: ρ = xcosθ + ysinθ to perform Hough transformation to obtain the maximum value max and its bit number n max , divide the Hough matrix by To perform a scale transformation, where θ is the angle and ρ is the radius length.

[0018] In step (5), the Hough matrix is ​​mapped to the θ axis and extended, and the target speed is recorded as v min With v max The value of the horizontal coordinate of the peak point at time t is selected, and a square area is cut out from the Hough matrix within this range and stored as a grayscale image; according to the N×M speed values ​​in step (1), N×M samples are constructed as a training set, and each grayscale image in the training set is labeled with a speed category.

[0019] In the step (6), a convolutional neural network is used for training to obtain the required network parameters such as the learning rate, and obtain the network parameters with the highest accuracy.

[0020] In the step (7), the target echo signal to be classified is subjected to wavelet denoising and then PWVD, wherein the mother wavelet is db35, and then Hough transform and scale transform are performed, and then mapping, extension and truncation are performed, and stored as a grayscale image as the input of the neural network to obtain the label of the speed category, thereby performing speed classification on the echo signal.

[0021] Beneficial effect: The present invention provides a classification method for active target speed based on time-frequency analysis and Hough transform, which effectively improves the accuracy of speed classification by processing signals in the time-frequency domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is the echo model of the training set of the present invention;

[0023] Figure 2 It is a time-frequency diagram of the echo signal when v=-5m / s, the signal-to-noise ratio is -5dB, and the signal-to-mixing ratio is 0dB in Example 1 of the present invention;

[0024] Figure 3 for Figure 2The image of the time-frequency graph after Hough transform and scale transformation;

[0025] Figure 4 is the echo model of the test set of the present invention;

[0026] Figure 5 Flowchart of the present invention for classifying active target speed. DETAILED DESCRIPTION

[0027] The present invention will be further described below in conjunction with the accompanying drawings.

[0028] The embodiments of the present invention are further described below in conjunction with the accompanying drawings:

[0029] Embodiment 1:

[0030] (1) The speed range of the active target is set to -20 m / s to 20 m / s, with approaching as a positive number and moving away as a negative number. Within this speed range, the speed is divided into four categories: -20 ≤ v i <-10, -10≤v i <0, 0≤v i ≤10, 10 <v i ≤20, 250 speed values ​​are equally sampled in each category, so there are 4×250 speed values ​​in total.

[0031] (2) Set the active sonar transmission signal to a linear frequency modulation signal with a pulse width of T p is 1 second, the period T is 4 seconds, the starting frequency f0 is 300 Hz, the bandwidth B is 600 Hz, and the sampling frequency f of the active sonar system is s is 4096 Hz, the propagation speed of sound waves in water is c 1500 m / s; the size of the target L is 100 m. Figure 1 The bright spot model shown uses the formula: in w is a random integer less than or equal to 500, s v =sin(2πf0′t+πk′t 2 ); When the radial velocity of the active target is v m / s, the pulse width of the echo signal is The starting frequency is The bandwidth is The frequency modulation slope is t is in [0,T p ′ ] s The time point of Hz sampling. Therefore, in each speed category, the active target speed v i, i=0,1,…,249 can get 250 echo signal samples s_train(t) i , i=0,1,…,249; then there are 4×250 test set echo signal samples in total for the four speed categories.

[0032] (3) Based on the principle of Pseudo Wigner-Ville Distribution (PWVD), using the formula Perform time-frequency analysis on the echo signal s_train(t), where h(t) is the Hamming window function with a window length of N / 4, N = T*f s .

[0033] (4) The coordinates of each point on the time-frequency plane are (x, y). The line on the time-frequency plane is mapped to the point in the (θ, ρ) plane using the polar coordinate transformation formula: ρ = xcosθ + ysinθ to perform Hough transform, and the maximum value max and its number of bits n_max are obtained. The Hough matrix is ​​divided by To perform a scale transformation, where θ is the angle and ρ is the radius length.

[0034] (5) Map the Hough matrix to the θ axis and extend it, record the horizontal coordinate values ​​of the peak points when the target speed is -20 m / s and 20 m / s, and within this range, extract a square area from the Hough matrix and save it as a grayscale image; based on the 4×250 speed values ​​in step (1), construct 4×250 samples as a training set, and each grayscale image in the training set is labeled with a speed category.

[0035] (6) Use convolutional neural network for training to obtain the required network parameters such as learning rate and obtain the network parameters with the highest accuracy.

[0036] (7) Input the echo signal to be classified. The signal used in this example is a simulated signal. Use the formula: in n1=0, w is a random integer in the range [0,500], s v =sin(2πf0′t+πk′t 2 ); Set the radial velocity of the active target to v = -5 m / s, and the pulse width of the echo signal to The starting frequency is The bandwidth is The frequency modulation slope is t is in [0,T p ′ ] sThe time point of Hertz sampling; p1=p2=1, p3=p4=0.5 is the normalized amplitude of the bright spot; n(t) is Gaussian white noise, and the signal-to-noise ratio is set to -5dB; r(t) is the reverberation signal, and the signal-to-mixing ratio is set to 0dB. The target echo signal to be classified is subjected to wavelet denoising and PWVD, where the mother wavelet: db35, and then Hough transform and scale transform are performed, and then mapped, extended and intercepted, and stored as a grayscale image as the input of the neural network to obtain the speed category label, thereby classifying the echo signal. The measured speed range is -10≤v i <0, which meets the target speed v = -5m / s. The network trained by the training set generated by this method has certain ability to resist noise, reverberation and small bright spot interference.

[0037] Example 2

[0038] (1) The speed range of the active target is set to -20 m / s to 20 m / s, with approaching as a positive number and moving away as a negative number. Within this speed range, the speed is divided into four categories: -20 ≤ v i <-10, -10≤v i <0, 0≤v i ≤10, 10 <v i ≤ 20. 250 speed values ​​were randomly selected from each category, so there were 4×250 speed values ​​to make the training set; 50 speed values ​​were randomly selected from each category, so there were 4×50 speed values ​​to make the test set.

[0039] (2) Set the active sonar transmission signal to a linear frequency modulation signal with a pulse width of T p is 1 second, the period T is 4 seconds, the starting frequency f0 is 300 Hz, the bandwidth B is 600 Hz, and the sampling frequency f of the active sonar system is s is 4096 Hz, the propagation speed of sound waves in water is c 1500 m / s; the size of the target L is 100 m. Figure 1 The bright spot model shown uses the formula: in w is a random integer less than or equal to 500, s v =sin(2πf0′t+πk′t 2 ); When the radial velocity of the active target is v m / s, the pulse width of the echo signal is The starting frequency is The bandwidth is The frequency modulation slope is t is in [0,T p ′ ] s The time point of Hz sampling. Therefore, in each speed category, the active target speed vi , i=0,1,…,249 can get 250 echo signal samples s_train(t) i , i=0,1,…,249; then there are 4×250 training set echo signal samples in total for the four velocity categories.

[0040] (3) Based on the principle of Pseudo Wigner-Ville Distribution (PWVD), using the formula Perform time-frequency analysis on the echo signal s_train(t), where h(t) is the Hamming window function with a window length of N / 4, N = T*f s .

[0041] (4) The coordinates of each point on the time-frequency plane are (x, y). The line on the time-frequency plane is mapped to the point in the (θ, ρ) plane using the polar coordinate transformation formula: ρ = xcosθ + ysinθ to perform Hough transform, and the maximum value max and its number of bits n_max are obtained. The Hough matrix is ​​divided by To perform a scale transformation, where θ is the angle and ρ is the radius length.

[0042] (5) Map the Hough matrix to the θ axis and extend it, record the horizontal coordinate values ​​of the peak points when the target speed is -20 m / s and 20 m / s, and within this range, extract a square area from the Hough matrix and save it as a grayscale image; based on the 4×250 speed values ​​in step (1), construct 4×250 samples as a training set, and each grayscale image in the training set is labeled with a speed category.

[0043] (6) Use convolutional neural network for training to obtain the required network parameters such as learning rate and obtain the network parameters with the highest accuracy.

[0044] (7) Basis Figure 4 The bright spot model shown uses the formula: The echo signal of the simulation test set. n1=0, w is a random integer in the range [0,500], s v = sin(2πf0′t ′ +πk′t ′2 ); When the radial velocity of the active target is v, the pulse width of the echo signal is The starting frequency is The bandwidth is The frequency modulation slope is t ′ For [0,T p ′] s The time point of Hertz sampling; p1=p2=1, p3=p4=0.5 is the normalized amplitude of the bright spot; n(t) is Gaussian white noise, and the signal-to-noise ratio is set to 0dB, -5dB and -10dB respectively; r(t) is the reverberation signal, and the signal-to-mixing ratio is set to 0dB. Therefore, in each speed category, the active target speed v i , i=0,1,…,49 can get 50 echo signal samples s_test(t) i , i=0,1,…,49; then there are 4×50 test set echo signal samples in total for the four speed categories. The target echo signal to be classified is subjected to wavelet denoising and then PWVD, wherein the mother wavelet: db35 is subjected to Hough transform and scale transform, and then mapped, extended and intercepted, and stored as a grayscale image. The test set is used as the input of the neural network, the label of the speed category is predicted, and compared with the label of the actual classification, the accuracy of the network is calculated, ten classification experiments are carried out and the average value of the accuracy is taken. The time-frequency diagram obtained in step (3) is used as the training set to train the network, and the time-frequency diagram obtained in this step is used as the test set and input into the network, ten classification experiments are carried out and the average value of the accuracy is taken. The results are shown in Table 1. Compared with directly introducing the time-frequency diagram into the network, the present invention can achieve more accurate active target classification effect.

[0045] Table 1

[0046]

[0047] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A classification method for active target speed based on time-frequency analysis and Hough transform, Features: The following steps are involved: (1) Setting a speed range of the active target, dividing the speed of the active target into several categories within this speed range, and extracting speed values ​​in each category with equal differences; (2) Using the known target scale information and the set target velocity value, according to the transmission signal parameters, the bright spot model and the Doppler principle, the echo signal of the simulation training set is obtained; In step (2), the transmission signal is selected as a linear frequency modulation signal, and the pulse width of the transmission signal is T p seconds, the period is T seconds, and the starting frequency is f 0 Hz, the bandwidth is B Hz, and the FM slope is The sampling frequency of the active sonar system is f s Hertz, the speed of sound waves in water is c meters per second; the size of the target is known to be L meters, according to the bright spot model, using the formula Get the echo signal, where w is a random integer less than or equal to 500, s v =sin(2πf 0 ′t+πk′t 2 ); When the radial velocity of the active target is v m / s, the pulse width of the echo signal is The starting frequency is The bandwidth is The frequency modulation slope is t is in [0,T′ p ] s The time point of Hertz sampling; therefore, in each speed category, the active target speed v i , i = 0, 1, ..., M-1 can get M echo signal samples s_train(t) i , i=0,1,…,M-1; then there are N×M training set echo signal samples in N velocity categories; (3) performing time-frequency analysis on the echo signal obtained in step (2); (4) Perform Hough transform and scale transform on the time-frequency plane; (5) Map, extend and intercept the Hough matrix and save it as a grayscale image; thereby constructing a training set, and each grayscale image in the training set is labeled with a speed category; (6) Use convolutional neural network for training to obtain the network parameters with the highest accuracy; (7) The target echo signal to be classified is subjected to wavelet denoising and then PWVD, where the mother wavelet is db35, and then Hough transform and scale transform are performed, and then mapping, extension and truncation are performed, and stored as a grayscale image as the input of the neural network to obtain the label of the speed category, thereby classifying the echo signal by speed.

2. The classification method of active target speed based on time-frequency analysis and Hough transform according to claim 1, Features: In step (1), the velocity range of the active target is v min ~v max , approaching is a positive number, and moving away is a negative number. Within this speed range, the speed is divided into N categories, and the upper and lower speed limits of each category are vel 1 Andvel 2 , in each category, M speed values ​​are equally sampled, which are Therefore, there are N×M speed values ​​in total.

3. The classification method of active target speed based on time-frequency analysis and Hough transform according to claim 1, Features: In step (3), according to the pseudo-Wigner-Weir distribution principle, the formula Perform time-frequency analysis on the echo signal s_train(t), where h(t) is the Hamming window function with a window length of N / 4, N = T*f s .

4. The classification method of active target speed based on time-frequency analysis and Hough transform according to claim 1, Features: In step (4), the coordinates of each point on the time-frequency plane are (x, y). The line on the time-frequency plane is mapped to a point in the (θ, ρ) plane using the polar coordinate transformation formula: ρ = x cosθ + y sinθ to perform Hough transform, and the maximum value max and its bit number n are obtained. max , divide the Hough matrix by To perform a scale transformation, where θ is the angle and ρ is the radius length.

5. The classification method of active target speed based on time-frequency analysis and Hough transform according to claim 1, Features: In step (5), the Hough matrix is ​​mapped to the θ axis and extended, and the target speed is recorded as v min With v max The value of the horizontal coordinate of the peak point at time t is obtained, and a square area is cut out from the Hough matrix within this range and stored as a grayscale image; according to the N×M speed values ​​in step (1), N×M samples are constructed as a training set, and each grayscale image in the training set is labeled with a speed category.

6. The classification method of active target speed based on time-frequency analysis and Hough transform according to claim 1, Features: In step (6), a convolutional neural network is used for training to obtain the parameters required by the network and obtain the network parameters with the highest accuracy.

Images