Clutter Background Extraction Method Based on Gaussian Mixture Model

Abstract

This invention discloses a clutter background extraction method based on a Gaussian mixture model. The steps are as follows: obtaining a training sample set and a test sample set; constructing a Gaussian mixture model; iteratively training the Gaussian mixture model; predicting the category of each test sample; and obtaining the clutter background extraction result. The Gaussian mixture model used in this invention can effectively determine information with target features. By using the target information to exclude false target information, the true target can be determined. By removing the true target signal from the samples, clutter background information, including high-intensity clutter background, can be extracted, ultimately obtaining more comprehensive clutter background information. This avoids the shortcomings of existing technologies in extracting high-intensity clutter background in Doppler-distance images and effectively improves the ability to suppress clutter.

Description

Technical Field

[0001] This invention belongs to the field of signal processing technology and relates to a clutter background extraction method, specifically a clutter background extraction method based on a Gaussian mixture model. The extracted background clutter can be used for clutter suppression in radar Doppler-range images. Background Technology

[0002] During target detection, in addition to receiving the target's echo signal, radar receivers can also receive many different types of clutter and noise signals. Some clutter signals have strong amplitudes, which can affect detection performance. Therefore, clutter background extraction is one of the key problems that urgently needs to be solved in the field of radar signal processing technology.

[0003] Currently, the commonly used clutter background extraction method is thresholding. Thresholding is an image segmentation technique that segments the image based on the significant difference in gray levels between the target and the background, using a threshold determined by a clear peak in the image's histogram. Thresholding typically employs single-threshold and multi-threshold methods, and the choice of threshold determines the quality of the segmentation. The principle of clutter background extraction methods is to classify image pixels into different categories based on grayscale values ​​or color features by setting different thresholds. However, this method cannot effectively separate the target from clutter in Doppler-distance images; clutter located close to the target often fails to be extracted due to its higher signal intensity.

[0004] For example, patent application CN111060880A, entitled "A Meteorological Clutter Suppression Method Based on the Constant False Alarm Rate Detection Principle," discloses a meteorological clutter suppression method based on the constant false alarm rate detection principle. The steps are as follows: 1) Obtain the spectrum of all echo signals output after the first constant false alarm rate detection process; 2) Select the frequency points to be measured sequentially from the effective spectrum of all echo signals and record their power; 3) Set up a first detection window, a second detection window, a third detection window, and a fourth detection window on both sides of the frequency point to be measured in the velocity and range dimensions, respectively, and sample the echo amplitude power of each frequency point within the window; 4) If the power of the frequency point to be measured is greater than the sum of the average power of all windows and a preset threshold, then retain the frequency point to be measured; otherwise, filter out the frequency point to be measured; 5) Through a second constant false alarm rate detection and target screening, perform amplitude detection, comparison, and identification processing on each frequency point to be measured in the velocity and range dimensions, filtering out the extracted clutter in advance so that it cannot enter the subsequent target tracking processing stage. This method can adaptively adjust the threshold size according to the strength of clutter echoes, which improves the performance of clutter extraction caused by global thresholding. However, it has poor ability to extract clutter backgrounds with high intensity in Doppler-distance images, resulting in low clutter suppression capability. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing a clutter background extraction method based on a Gaussian mixture model, which solves the technical problem of low clutter suppression capability in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Obtain the training sample set and the test sample set:

[0008] (1a) Obtain N×L pulse echo data of the airborne Ka-band radar continuously tracking the UAV target with a pulse repetition frequency of N within a tracking time of L seconds in pulse-Doppler mode, and sample each pulse echo data M times to obtain M×N×L discrete digital signals. Then, preprocess each discrete digital signal, and use all the preprocessed time-domain pulse signals to construct a time-domain pulse signal matrix with dimension M×(N×L), where N≥32000Hz, L≥2, and M≥319;

[0009] (1b) Perform coherent pulse accumulation on the time-domain pulse signal matrix at intervals of N / 40, with a range element number of R, to obtain an image set D consisting of J = L × 40 Doppler-range images of dimension M × R. Use T images from image set D as the training sample set D. train The remaining Q = JT images are used as the test sample set D. test Where R ≥ 128, J / 2 < T;

[0010] (2) Constructing a Gaussian mixture model:

[0011] Construct a Gaussian mixture model consisting of M×R parallel sub-models, and sample the sub-model P corresponding to the r-th distance unit in the m-th sampling. m,r It contains K Gaussian distribution functions, where 3≤K≤7, m=1,2,…,M, r=1,2,…,R;

[0012] (3) Iteratively train the Gaussian mixture model:

[0013] (3a) Initialize the number of iterations to i, the maximum number of iterations to I, I≥J / 2, and randomly select the training sample set D. train any training sample Initialize the m×r-th sub-model P m,r The k-th Gaussian distribution function G k The weights, mean, and variances are w, respectively. k,0 μ k,0 , And let i = 1;

[0014] (3b) The training sample set D train As input to the Gaussian mixture model, each sub-model P m,r For each training sample The value of each pixel in Predict the category and obtain Prediction results

[0015]

[0016]

[0017] in, This represents the predicted class result of the signal at the r-th distance cell position of the m-th sampling point in the t-th training sample. This indicates that the predicted category result is the target signal. This indicates that the predicted category result is clutter signal;

[0018] (3c) Apply a global learning rate α to each sub-model P m,r Each Gaussian distribution function G in k weight w k,t Mean μ k,t ,variance The update is performed to obtain the parameters of each Gaussian distribution function. The updated sub-model;

[0019] (3d) Select the sub-model P after parameter update. m,r middle The largest The sub-model consists of Gaussian distribution functions after the current iteration number i. Where B represents the condition that the first b w are satisfied. k The sum of the values ​​of b at the time when it exceeds the threshold (Threshold), where 1 ≤ b ≤ K;

[0020] (3e) Determine whether i≥J / 2 is true. If yes, obtain the trained Gaussian mixture model. Otherwise, let i = i+1 and execute step (3b).

[0021] (4) Predict the category for each test sample:

[0022] Test sample set D test Each sub-model P is used as input to the trained Gaussian mixture model. m,r For each test sample The value of each pixel in Predict the category and obtain D. test The predicted result Y test ;

[0023] (5) Obtain the clutter background extraction results:

[0024] Statistical test sample set D test The number of times b corresponding to the Q prediction results has a category prediction value of 1 at the r-th distance cell position of the m-th sampling point. mr And will satisfy b mr The qth prediction result of >Q / 2 Corresponding test samples The pixel value at the r-th distance unit position of the m-th sampling point is set to zero to obtain the clutter background in the Q test samples.

[0025] This invention has the following advantages over existing technologies:

[0026] The Gaussian mixture model constructed in this invention comprises multiple sub-models arranged in parallel. During the iterative training of the Gaussian mixture model and the acquisition of prediction results for test samples, the Gaussian mixture model can effectively determine moving target information, eliminate false target information to determine the real target using the moving target information, and extract clutter background information including high-intensity clutter background by removing the real target signal from the sample. This results in more comprehensive clutter background information, avoiding the shortcomings of existing technologies in extracting high-intensity clutter background in Doppler-distance images and effectively improving the ability to suppress clutter. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0028] Figure 2 This is a flowchart illustrating the implementation of iterative training of the Gaussian mixture model of the present invention.

[0029] Figure 3 This is a flowchart illustrating the implementation of obtaining clutter background according to the present invention;

[0030] Figure 4 The images shown are Doppler-distance images used in the simulation experiments of this invention, as well as simulation results of this invention and prior art. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0032] Reference Figure 1 The present invention includes the following steps:

[0033] Step 1) Obtain the training sample set and the test sample set:

[0034] Step 1a) First, acquire N×L pulse echo data of the airborne Ka-band radar continuously tracking the UAV target within a tracking time of L seconds using pulse-Doppler mode and with a pulse repetition frequency of N. Sample each pulse echo data point M times to obtain M×N×L discrete digital signals. Perform digital down-conversion processing on each of the obtained discrete digital signals to obtain a baseband signal with a lower frequency than the original signal after mixing. Downsample each digitally downconverted baseband signal, and then perform pulse compression processing on each downsampled signal. Construct a time-domain pulse signal matrix of dimension M×(N×L) using all the time-domain pulse signals after pulse compression processing. In this embodiment, N=32000Hz, L=2, M=319.

[0035] Step 1b) Perform coherent pulse accumulation on the time-domain pulse signal matrix at intervals of N / 40, with a range element number of R, to obtain an image set D consisting of J = L × 40 Doppler-range images of dimension M × R. Use T images from image set D as the training sample set D. train The remaining Q = JT images are used as the test sample set D. test In this embodiment, R = 128 and T = 74;

[0036] Step 2) Construct a Gaussian mixture model:

[0037] Construct a Gaussian mixture model consisting of M×R parallel sub-models, where P is the sub-model corresponding to the r-th distance unit on the m-th sampling. m,r It consists of K Gaussian distribution functions. In this embodiment, K = 5, m = 1, 2, ..., 319, r = 1, 2, ..., 128.

[0038] Step 3) Iteratively train the Gaussian mixture model, the implementation steps of which are as follows: Figure 2 As shown.

[0039] Step 3a) Initialize the number of iterations to i, the maximum number of iterations to I, I≥J / 2, and randomly select the training sample set D. train any training sample Initialize the m×r-th sub-model P m,r The k-th Gaussian distribution function G k The weights, mean, and variances are w, respectively. k,0 μ k,0 , Let i = 1; in this embodiment, I = 74.

[0040] Step 3b) The training sample set D train As input to the Gaussian mixture model, each sub-model P is used. m,r For each training sample The value of each pixel in To predict the category, the steps are as follows:

[0041] Calculate each training sample Value of each pixel Each sub-model P in the Gaussian mixture model m,r Each Gaussian distribution function G k mean μ k,t-1 distance And judge If true, then let the value of the r-th distance cell position of the m-th sampling point be... This indicates that the predicted category result is clutter signal; otherwise, let... This indicates that the predicted category result is the target signal. (The result is...) Prediction results

[0042]

[0043]

[0044] in, This represents the category prediction result of the r-th distance cell position of the m-th sampling point in the t-th training sample.

[0045] For each training sample, a Gaussian mixture model is used. The value of each pixel in Category prediction can effectively identify information with target features, i.e. Location information m, r at that time;

[0046] Step 3c) Using the global learning rate α and For each sub-model P m,r Each Gaussian distribution function G in k weight w k,t Mean μ k,t ,variance The update is performed to obtain the parameters of each Gaussian distribution function. The updated sub-model, where:

[0047] when At that time, the update formula is:

[0048] pass μ k,t-1 σ k,t-1 Calculation submodule P m,r Central Order The index l of the smallest Gaussian distribution function:

[0049]

[0050] Where, argmin k [·] indicates that the variable value k is returned when the objective function func(·) is minimized;

[0051] Through α, l pairs P m,r Gaussian distribution function G in k weight w k,t-1 Update to get w k,t And the weight w k,t Perform normalization to obtain the normalized weights. in:

[0052]

[0053]

[0054] Through α, for each Gaussian distribution function G k μ k,t-1 , Update to obtain

[0055]

[0056]

[0057]

[0058] Where ρ represents the change factor.

[0059] when At that time, the update formula is:

[0060]

[0061]

[0062]

[0063] By updating the parameters in the Gaussian mixture model, the updated Gaussian mixture model is more resistant to interference from information with moving target features in the sample.

[0064] Step 3d) Select the sub-model P with updated parameters m,r middle The largest The sub-model consists of Gaussian distribution functions after the current iteration number i. Where B represents the condition that the first b w are satisfied. k The sum of the values ​​of b at the time when it exceeds the threshold (Threshold), where 1 ≤ b ≤ K;

[0065] Step 3e) Determine whether i≥J / 2 is true. If yes, obtain the trained Gaussian mixture model. Otherwise, let i = i+1 and execute step 3b).

[0066] Step 4) Predict the category for each test sample:

[0067] Test sample set D test Each sub-model P is used as input to the trained Gaussian mixture model. m,r For each test sample The value of each pixel in Predict the category and obtain D. test The predicted result Y test ;

[0068] Step 5) Obtain the clutter background extraction results, the implementation steps of which are as follows: Figure 3 As shown.

[0069] Step 5a) First, statistically analyze the test sample set D. test The number of times b corresponding to the Q prediction results has a category prediction value of 1 at the r-th distance cell position of the m-th sampling point. mr ;

[0070] Step 5b) Record the location of the actual target, i.e., satisfy b. mr The position of >Q / 2;

[0071] Step 5c) The q-th prediction result Corresponding test samples The pixel value at the r-th distance cell position of the m-th sampling point is set to zero to obtain the test sample. The clutter background signals in the test samples constitute Q clutter backgrounds.

[0072] By using a trained Gaussian mixture model and Q prediction results, false target information can be eliminated to determine the real target. By removing the real target signal from the sample, clutter background information, including high-intensity clutter background, can be extracted, and finally a more comprehensive clutter background information can be obtained.

[0073] The technical effects of the present invention will be further illustrated by the following simulation experiments.

[0074] 1. Simulation experimental conditions:

[0075] The simulation experiment used PyCharm with Python version 3.8 as the development tool, and Python as the development language. The Doppler-distance images used in the simulation were as follows: Figure 4 As shown in (a).

[0076] 2. Simulation content and result analysis:

[0077] Simulations were performed on the clutter suppression capabilities of this invention and existing clutter suppression algorithms based on the constant false alarm rate (CFAR) principle. The protection and reference cells in the CFAR-based clutter suppression algorithm were set to 48 and 72, respectively. The results are as follows: Figure 4 (b) and Figure 4 As shown in (c), the clutter suppression rates of the two methods were calculated and compared based on simulation results, as shown in Table 1:

[0078] Table 1

[0079] method This invention Existing CFAR algorithm Clutter suppression rate 88.73 81.62

[0080] Figure 4 In the three sub-graphs, the horizontal axis represents velocity, and the vertical axis represents distance. For example... Figure 4 (b) The clutter background image extracted using the present invention is shown in the figure. Figure 4 Compared to the clutter background image extracted by the existing CFAR method shown in (c), more clutter signals with higher intensity are preserved.

[0081] The clutter suppression rate in Table 1 represents the ratio of the extracted clutter background to the overall clutter background. The clutter suppression rate of this invention is 88.73%, which is significantly higher than the existing clutter suppression rate of 81.62% based on the constant false alarm rate (CFAR) algorithm. This shows that the clutter background extraction method based on the Gaussian mixture model proposed in this invention has a better clutter suppression effect.

Claims

1. A clutter background extraction method based on Gaussian mixture model, characterized in that, Comprising the following steps: (1) Obtain a training sample set and a test sample set: (1a) acquiring airborne The frequency band radar tracks the unmanned aerial vehicle target continuously through the pulse-Doppler working mode in seconds with pulses per second as the pulse repetition frequency, obtains pulse echo data of the unmanned aerial vehicle target, samples each pulse echo data times to obtain discrete digital signals, pre-processes each discrete digital signal, and constructs a time domain pulse signal matrix with a dimension of by using all the pre-processed time domain pulse signals, wherein , , ; (1b) Based on the number of pulses The number of distance elements for the time-domain pulse signal matrix is ​​[number to be filled in]. The coherent pulse accumulation yields the result from... Aspect dimension is Image set composed of Doppler-distance images and the image set In Images as training sample set , the rest Images as test sample set ,in, , ; (2) Construct a Gaussian mixture model: Construct a Gaussian mixture model composed of a plurality of sub-models arranged in parallel The first sub-model corresponds to the first distance unit The second sub-model corresponds to the second distance unit The third sub-model corresponds to the third distance unit The fourth sub-model corresponds to the fourth distance unit Each of the sub-models contains a plurality of Gaussian distribution functions , , ; (3) Iteratively train the Gaussian mixture model: (3a) Initialize the number of iterations to be The maximum number of iterations is , Randomly select training sample set any training sample Initialize the first Sub-model The Middle Gaussian distribution function The weights, mean, and variance are respectively , , and order ; (3b) training the set of training samples As input to the Gaussian mixture model, each sub-model For each training sample predicts the class of each pixel value in the training sample, resulting in a prediction of the training sample : ； ； wherein, represents the class prediction result of the signal at the jth distance unit position of the ith sampling point in the ith training sample, represents that the predicted class result is a target signal, represents that the predicted class result is a clutter signal;​​​ (3c) by a global learning rate for each sub-model for each Gaussian distribution function in the sub-model , mean , variance is updated to obtain an updated sub-model with parameters , , ​ (3d) selecting a sub-model updated with parameters In The largest Gaussian distribution function consisting of the current iteration number is After the sub-model , wherein Indicates that the previous The sum of When the value is greater than the threshold The numerical value of At the moment, ; (3e) determining whether it is true, if yes, obtaining the trained Gaussian mixture model, otherwise, letting and performing step (3b); (4) Predict the class of each test sample: The test sample set As an input of the trained Gaussian mixture model, each sub-model For each test sample Each pixel value The class is predicted, and the prediction result of ;​ (5) Obtain the extraction result of the clutter background: statistical test sample set corresponding the number of times that the category prediction value is 1 at the first distance unit position of the first sampling point of the corresponding test sample the first prediction result that satisfies the number of times that the category prediction value is 1 at the first distance unit position of the first sampling point of the corresponding test sample the first prediction result that satisfies the first prediction result that satisfies the first prediction result that satisfies the corresponding test sample the first prediction result that satisfies the first prediction result that satisfies the first prediction result that satisfies clutter background in the first test sample.

2. The method according to claim 1, wherein, The pre-processing of each discrete digital signal in step (1a) is implemented by: performing digital down-conversion on each discrete digital signal, and performing down-sampling on each digital down-converted baseband signal, and then performing pulse compression processing on each down-sampled signal to obtain a time-domain pulse signal. 3.The clutter background extraction method based on Gaussian mixture model according to claim 1, characterized in that, each sub-model in the step (3b) for each pixel value in each training sample for each pixel value in each training sample for each pixel value in each training sample for each pixel value in each training sample for each pixel value in each training sample for each pixel value in each training sample for each pixel value in each training sample for each pixel value in each training sample​​​​​​ 4. The Gaussian mixture model based clutter background extraction method according to claim 1, characterized in that, the global learning rate in step (3c) for each sub-model the weight , mean , variance of each Gaussian distribution function is updated, wherein: When the update formula is: ； ； ； ； ； ； wherein, denotes returning the variable value at which the objective function takes its minimum value , denotes the change factor; When the update formula is: ； ； 。

Images