A radar micro-motion target classification method combined with a reliable XGBoost classifier

Abstract

This invention relates to a radar micro-motion target classification method combined with a reliable XGBoost classifier, comprising: acquiring radar echo data of the micro-motion target to be tested; performing clutter suppression preprocessing on the radar echo data; performing feature extraction processing on the preprocessed radar echo data to obtain a feature matrix to be tested; classifying the feature matrix to be tested using a trained reliable XGBoost classifier to obtain a corresponding preliminary classification result; and performing inter-frame decision fusion processing on the preliminary classification result based on the temporal relationship of the radar echo data of the micro-motion target to obtain a classification result for the micro-motion target to be tested. The preliminary classification result output by the reliable XGBoost classifier includes predicted class probabilities and uncertainty results. The predicted class probabilities are used to determine the classification decision result, and the uncertainty results are used to evaluate the reliability of the classification decision result. This invention reduces the uncertainty of the overall classification system and improves the accuracy of the classification results.

Description

Technical Field

[0001] This invention belongs to the field of radar micro-movement target classification, specifically relating to a radar micro-movement target classification method that combines a reliable XGBoost classifier. Background Technology

[0002] In the radar micro-movement target classification process, classifier design is the most critical step, significantly impacting the performance of the entire classification system. A reliable classifier is one that can output decision results from observation data and measure the reliability of those decisions. Further analysis and processing of the reliability of the decision results output by the reliable classifier helps assist subsequent decisions and ensures the reliability of the final decision result of the entire classification system.

[0003] Existing radar micro-motion target classification systems mostly consider only ideal conditions, neglecting the impact of complex environmental factors during radar data acquisition on echo signals in real-world applications. In actual acquisition environments, echo data is affected by atmospheric and ground features, leading to low signal-to-noise ratios or even the inability to observe micro-motion signals. Such echoes are difficult to discriminate against, severely impacting the accuracy and reliability of classification results. Secondly, most systems treat the predicted probability output by the classifier as the reliability of the prediction result, ignoring the different meanings of prediction probability and reliability. Even with a very high prediction probability, there may still be significant uncertainty, resulting in low reliability of the classification result. The prediction probability is insufficient to measure the reliability of the classification result. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a radar micro-motion target classification method combined with a reliable XGBoost classifier. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] This invention provides a radar micro-motion target classification method combining a reliable XGBoost classifier, comprising:

[0006] Step 1: Acquire radar echo data of the micro-moving target to be tested, and perform clutter suppression preprocessing on the radar echo data;

[0007] Step 2: Perform feature extraction processing on the preprocessed radar echo data to obtain the feature matrix to be measured;

[0008] Step 3: Use the trained reliable XGBoost classifier to classify the feature matrix to be tested, and obtain the corresponding preliminary classification result. Based on the temporal relationship of the radar echo data of the micro-moving target to be tested, perform inter-frame decision fusion processing of the preliminary classification result to obtain the classification result of the micro-moving target to be tested.

[0009] The preliminary classification result output by the reliable XGBoost classifier includes predicted class probabilities and uncertainty results. The predicted class probabilities are used to determine the classification decision result, and the uncertainty results are used to evaluate the reliability of the classification decision result.

[0010] In one embodiment of the present invention, step 1 includes:

[0011] The radar echo data of the micro-moving target to be measured is collected, and clutter suppression preprocessing is performed on the radar echo data using the regional CLEAN method or moving target indication technology.

[0012] In one embodiment of the present invention, step 2 includes:

[0013] Step 2.1: Based on the modulation characteristics of the preprocessed radar echo data, extract feature values ​​that reflect the structural and motion information of the micro-moving target;

[0014] Step 2.2: Normalize the eigenvalues ​​to obtain the feature matrix to be tested.

[0015] In one embodiment of the present invention, the training process of the reliable XGBoost classifier includes:

[0016] S1: Obtain radar echo data of multiple categories of micro-moving targets as training samples to form a training sample set, and set corresponding real category labels for the radar echo data in the training sample set.

[0017] S2: Perform clutter suppression preprocessing and feature extraction processing on the radar echo data in the training sample set in sequence to form a preliminary training feature matrix. Normalize the elements in the preliminary training feature matrix to obtain the training feature matrix.

[0018] S3: Initialize the parameters of the reliable XGBoost classifier, including the number of iterations, maximum tree depth, number of classes, and learning rate;

[0019] S4: Input the training feature matrix and the true class labels into the reliable XGBoost classifier and perform multiple iterations of training. The forward addition method is used for multiple iterations. In each iteration, a decision tree corresponding to the number of classes is built. Each decision tree is split into nodes in turn. When splitting the current decision tree, the other decision trees remain unchanged.

[0020] In one embodiment of the present invention, S4 includes:

[0021] S41: Based on the input training feature matrix and the true class label, a greedy algorithm is used to traverse the different feature values ​​of each feature and determine the nodes of the decision tree according to the loss function;

[0022] S42: Repeat S41 to split each node obtained by splitting in turn until the maximum tree depth is reached, and a decision tree is obtained.

[0023] S43: Keep the established decision tree unchanged, repeat S41-S42 until the number of decision trees corresponding to the number of categories is obtained, calculate the loss function value of the current iteration training, update the predicted category label of the current iteration training, and complete one iteration training. The initial value of the predicted category label is in the form of an all-zero vector.

[0024] S44: Repeat S41-S43 for the next iteration of training until the preset number of iterations is reached. Save the split features and corresponding feature values ​​of each node of the decision tree in each iteration process, as well as the leaf node weights of each node, to obtain a reliable XGBoost classifier that has been trained.

[0025] In the iterative training, the leaf node weights of each training sample after the previous iteration are used as the initial values ​​of the leaf node weights in the current iteration. The initial values ​​of the leaf node weights of each training sample in the first iteration are in the form of a vector of all zeros.

[0026] In one embodiment of the present invention, S41 includes:

[0027] S411: Determine the unique feature value corresponding to each feature in the training feature matrix, and establish the feature value set θ corresponding to each feature. m ={θ1,θ2,...,θ q}, m=1,2,…,M, where θ m Let M represent the set of feature values ​​for the m-th feature, M represent the number of features, θ represent unique feature values, and q represent the number of unique feature values.

[0028] S412: Traverse each feature value in the feature value set of each feature, take the feature as a split node, take the feature as a split value, and for the split node, assign the training samples in the training feature matrix whose feature values ​​are less than the split value to the left node, and assign the remaining samples to the right node.

[0029] S413: Determine if the training samples of the left and right nodes are zero. If both the left and right nodes have non-zero training samples, calculate the corresponding loss function value after this split and execute S414; otherwise, return to step S412. The loss function is:

[0030]

[0031]

[0032] In the formula, Let x represent the i-th training sample. i The loss function, where n represents the number of training samples, p ij Indicates training sample x i The probability of belonging to the j-th category, y ij Indicates training sample x i The true category label of the j-th category, α i Indicates training sample x i The Dirichlet distribution parameters, α i =[α i1 ,α i2 ,...α ij ...,α iK ] = e i +1,j=1,...,K, where K represents the number of categories, e i Indicates training sample x i Evidence vector, e i =[e i1 ,e i2 ,..e ij ..,e iK ],j=1,...,K, Indicates training sample x i In this iteration of training, the updated predicted class labels of the current decision tree are... Indicates training sample x i In the previous iteration of training, the predicted class label is used; learning_rate represents the learning rate. Indicates training sample x i In this iteration of training, the weights of the leaf nodes of the current decision tree are given, where l represents the l-th iteration of training; B(α) i ) indicates that the parameter is α i The multivariate Beta distribution, p i Indicates training sample x i The predicted class probability, S i Indicates training sample x i The Dirichlet intensity, where exp represents the exponential operation;

[0033] S414: Based on the calculated loss function value, use gradient descent to solve for the weights of the left and right nodes. Using the calculated weights of the left and right nodes, recalculate the loss function value. Alternately optimize the weights of the left and right nodes until the loss function converges. Save the optimized weights of the left and right nodes, as well as the corresponding loss function value.

[0034] S415: After traversing each feature value in the feature value set of each feature, determine the feature that minimizes the loss function value and its corresponding feature value. Use this feature as the node of the current decision tree, and the corresponding feature value as the split value of the node. Save the weights of the left and right nodes of this node as the weights of the leaf nodes of this node.

[0035] In one embodiment of the present invention, in S43, the predicted class label for each training sample is updated according to the following formula:

[0036]

[0037] In the formula, x represents the training sample in the current iteration of training. i Predicted category labels, x represents the training sample from the previous iteration of training. i The predicted class label, learning_rate represents the learning rate, ω i This represents the training sample x obtained in the current iteration of training. i The leaf node weight vector, ω i =[ω i1 ,ω i2 ,..ω ij ..,ω iK ],j=1,...,K.

[0038] In one embodiment of the present invention, step 3 includes:

[0039] Step 3.1: Input the feature matrix to be tested into the trained reliable XGBoost classifier. Using all the decision trees obtained from each iteration of training, obtain the weight vector of the feature matrix to be tested in the leaf nodes of each decision tree, calculate the corresponding predicted class probability and uncertainty result, and output it as the preliminary classification result.

[0040] Step 3.2: Based on the preliminary classification results corresponding to the radar echo data of the target micro-motion at the current time and the previous time, perform inter-frame decision fusion processing using the inter-frame decision fusion formula to obtain the fused classification result. The inter-frame decision fusion formula is as follows:

[0041]

[0042] In the formula, This indicates the classification result of the fusion. This represents the preliminary classification result at time t. This represents the preliminary classification result at time t-1;

[0043] Step 3.3: Take the category corresponding to the maximum predicted category probability in the fused classification result as the classification result of the micro-movement target to be tested.

[0044] In one embodiment of the present invention, step 3.1, calculating the predicted category probability and uncertainty result, includes:

[0045] Step a: Based on the weight vectors of the feature matrix to be tested in the leaf nodes of each decision tree, calculate the corresponding predicted label by summing them. Based on predicted labels Obtain the corresponding evidence vector e L ,in,

[0046]

[0047] e l j =[e l 1,e l 2,..e l j ..,e l K ],j=1,...,K,l=1,...,L;

[0048]

[0049] In the formula, L represents the number of training iterations, and ω l ω represents the weight vector of the leaf nodes of all decision trees obtained from the l-th iteration of training of the feature matrix to be tested. l =[ω l 1,ω l 2,..ω l j ..,ω l K ],j=1,...,K, Let represent the predicted label obtained using the decision tree trained in the l-th iteration, where This means using the decision tree obtained from the l-th iteration of training to determine the predicted label of the feature matrix to be tested belonging to the j-th class;

[0050] Step b: Based on the predicted labels corresponding to the feature matrix to be tested. and evidence vector e L The predicted category probabilities and uncertainties are calculated according to the following formula:

[0051] α=[α1,α2,...α j ...α K ] = e L+1,j=1,...,K;

[0052]

[0053]

[0054]

[0055] In the formula, p j α represents the probability that the micro-moving target to be tested belongs to each category. j Let represent the j-th element of the Dirichlet distribution parameters of the micro-motion target to be measured, u represent the uncertainty result of the micro-motion target to be measured, S represent the Dirichlet intensity of the micro-motion target to be measured, and e represent the uncertainty result of the micro-motion target to be measured. L j Represents the evidence vector e L The j-th element.

[0056] In one embodiment of the present invention, in step 3.2, the preliminary classification results at time t and time t-1 are respectively expressed as follows:

[0057]

[0058]

[0059] b = [b1, b2, ... b j ...,b K ],j=1,...,K;

[0060]

[0061] In the formula, b represents the confidence vector of the micro-movement target to be measured, b j This represents the j-th element in the trust vector b;

[0062] The classification result of the fusion is expressed as follows:

[0063]

[0064] In the formula, p' j α' represents the probability that the fused micro-movement target belongs to each category. j Let S' represent the j-th element of the Dirichlet distribution parameters of the fused micro-motion target, and let S' represent the Dirichlet intensity of the fused micro-motion target.

[0065] α' j =e' j +1,j=1,...,K;

[0066] e' j =b' j×S',j=1,...,K;

[0067]

[0068]

[0069]

[0070] In the formula, e' j Let b' represent the j-th element in the fused evidence vector. j Let represent the j-th element in the fused confidence vector, u' represent the uncertainty result of the fused micro-motion target, and C represent the degree of conflict between the belief functions of the echo data at two time points. This represents the normalization coefficient.

[0071] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0072] The radar micro-motion target classification method of this invention, combined with a reliable XGBoost classifier, improves the entire process of inter-frame decision fusion by incorporating uncertainty, based on the prediction results and uncertainties obtained from the proposed reliable XGBoost classifier. The reliable XGBoost classifier of this invention can accurately estimate the uncertainty of the decision results and provide the estimation results in numerical form to assist subsequent decision-making. Inter-frame decision fusion is performed using the uncertainty of the obtained decision results, ensuring that prediction results with low uncertainty contribute more to the fused results, reducing the overall uncertainty of the classification system and improving the accuracy of the classification results.

[0073] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of a radar micro-motion target classification method combined with a reliable XGBoost classifier provided by an embodiment of the present invention;

[0075] Figure 2 This is a flowchart of a radar micro-motion target classification method combined with a reliable XGBoost classifier provided by an embodiment of the present invention. Detailed Implementation

[0076] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a radar micro-motion target classification method combined with a reliable XGBoost classifier based on the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0077] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0078] Firstly, this embodiment provides a radar micro-motion target classification method combined with a reliable XGBoost classifier. Please refer to [link to previous document]. Figure 1 and Figure 2 , Figure 1 This is a schematic diagram of a radar micro-motion target classification method combined with a reliable XGBoost classifier provided by an embodiment of the present invention; Figure 2 This is a flowchart of a radar micro-motion target classification method combining a reliable XGBoost classifier, provided by an embodiment of the present invention. As shown in the figure, the method includes the following steps:

[0079] Step 1: Acquire radar echo data of the micro-moving target to be tested, and perform clutter suppression preprocessing on the radar echo data;

[0080] When radar targets or target components move, they are often accompanied by micro-motions such as vibration, rotation, and acceleration, in addition to the translation of the center of mass. These radar targets are collectively referred to as micro-motion targets.

[0081] In an optional implementation, step 1 includes: acquiring radar echo data of the micro-moving target to be measured, and performing clutter suppression preprocessing on the radar echo data using the area CLEAN method or moving target indication technology.

[0082] Ground clutter suppression using the regional CLEAN method or clutter suppression using Moving Target Indication (MTI) are both mature clutter suppression methods, and their specific steps will not be elaborated here. MTI uses MT filters, taking advantage of the difference in Doppler frequencies between clutter and moving targets. This results in a deep stopband in the filter's frequency response at DC and integer multiples of the pulse repetition frequency, while suppression at other frequencies is weaker. This deep "notch" effectively suppresses stationary targets and static clutter.

[0083] Step 2: Perform feature extraction processing on the preprocessed radar echo data to obtain the feature matrix to be measured;

[0084] In an optional implementation, step 2 includes:

[0085] Step 2.1: Based on the modulation characteristics of the preprocessed radar echo data, extract feature values ​​that reflect the structural and motion information of the micro-moving target;

[0086] Optionally, physical mechanism-driven feature extraction methods can be used to extract various features such as peak function sum, peak function variance, peak function entropy, Doppler spectrum l1 norm, and Doppler spectrum difference modulus summation in the time domain, Doppler domain, harmonic domain, and time-spectrum domain.

[0087] Step 2.2: Normalize the eigenvalues ​​to obtain the feature matrix to be tested.

[0088] Step 3: Use the trained reliable XGBoost classifier to classify the feature matrix to be tested and obtain the corresponding preliminary classification results. Based on the temporal relationship of the radar echo data of the micro-moving target to be tested, perform inter-frame decision fusion processing of the preliminary classification results to obtain the classification results of the micro-moving target to be tested.

[0089] The preliminary classification results output by the reliable XGBoost classifier include predicted class probabilities and uncertainty results. The predicted class probabilities are used to determine the classification decision, and the uncertainty results are used to evaluate the reliability of the classification decision.

[0090] In this embodiment, a reliable XGBoost classifier is proposed. Based on the prediction results and uncertainties obtained from the reliable XGBoost classifier, the entire process of inter-frame decision fusion by incorporating uncertainty is improved. The reliable XGBoost classifier can accurately estimate the uncertainty of the decision results and provide the estimation results in numerical form to assist subsequent decision-making. By using the uncertainty of the obtained decision results for inter-frame decision fusion, it is ensured that prediction results with low uncertainty contribute more to the fused results, reducing the uncertainty of the overall classification system and improving the accuracy of the classification results.

[0091] Furthermore, the training process of the reliable XGBoost classifier in this embodiment and the process of combining the preliminary classification results of the reliable XGBoost classifier to perform inter-frame decision fusion to obtain the classification result of the micro-motion target to be tested are described in detail.

[0092] The reliable XGBoost classifier provided in this embodiment uses uncertainty results as an evaluation of decision reliability. This reliable XGBoost classifier combines subjective logic theory with the Dirichlet distribution function to model the uncertainty of the output predicted class probability and the decision result, transforming the probability space into an evidence space. Through uncertainty estimation, it evaluates the reliability of the decision result, enabling the reliable XGBoost classifier to output both the sample prediction probability and the reliability of the classification result.

[0093] In this embodiment, the reliable XGBoost classifier is trained iteratively using a forward addition method, where the result of each iteration is added to the previous result, and each iteration is based on the classification result of the previous iteration. During a single iteration, for a K-class problem, reliable XGBoost builds K decision trees, transforming a K-class problem into K binary classification problems. In a decision tree, all training samples are assigned to different leaf nodes through multiple nodes. The weight ω of the leaf node to which each sample falls is used as the output value of that sample on that decision tree. Samples falling into the same node all use the same leaf node weight ω as their output value. Building K decision trees yields the output value of the sample on the K decision trees, and this output value is a K×1 dimensional vector.

[0094] In an optional implementation, the training process for a reliable XGBoost classifier includes:

[0095] S1: Obtain radar echo data of multiple categories of micro-moving targets as training samples to form a training sample set, and set the corresponding real category label for the radar echo data in the training sample set.

[0096] Optionally, radar echo data of different types of micro-moving targets can be randomly selected from the collected echo data as training samples to form a training sample set.

[0097] S2: Perform clutter suppression preprocessing and feature extraction processing on the radar echo data in the training sample set in sequence to form a preliminary training feature matrix. Normalize the elements in the preliminary training feature matrix to obtain the training feature matrix.

[0098] Optionally, clutter suppression preprocessing can be performed on the radar echo data in the training sample set using the regional CLEAN method or moving target indication technology.

[0099] In this embodiment, the true category labels of the radar echo data in the training sample set are converted to obtain the corresponding one-hot encoding form.

[0100] In this embodiment, the training feature matrix is ​​denoted as X. M×nWhere M is the feature dimension, n is the total number of training samples, the corresponding true class label is denoted as Y, and the predicted label obtained using the reliable XGBoost classifier is denoted as .

[0101] S3: Initialize the parameters of the reliable XGBoost classifier, including the number of iterations, maximum tree depth, number of classes, and learning rate;

[0102] In this embodiment, the number of iterations is denoted as L, the maximum tree depth is denoted as m_depth, the number of categories is denoted as K, and the learning rate is denoted as learning_rate.

[0103] S4: Input the training feature matrix and the true class labels into the reliable XGBoost classifier and perform multiple iterations of training. The forward addition method is used for multiple iterations. In each iteration, a decision tree corresponding to the number of classes is built. Each decision tree is split into nodes in turn. When splitting the node of the current decision tree, the other decision trees remain unchanged.

[0104] In this embodiment, for the K-classification problem, K decision trees are established in a single iteration of training. Each decision tree is split into nodes in sequence. When splitting each decision tree, the remaining decision trees are kept unchanged. When calculating the loss function and gradient, the unchanged decision trees are calculated using the initial values.

[0105] In an optional implementation, S4 includes:

[0106] S41: Based on the input training feature matrix and the true class label, a greedy algorithm is used to traverse the different feature values ​​of each feature and determine the nodes of the decision tree according to the loss function;

[0107] In an optional implementation, S41 includes:

[0108] S411: Determine the unique eigenvalues ​​for each feature in the training feature matrix, and establish the eigenvalue set θ for each feature. m ={θ1,θ2,...,θ q}, m=1,2,…,M, where θ m Let M represent the set of feature values ​​for the m-th feature, M represent the number of features, θ represent unique feature values, and q represent the number of unique feature values.

[0109] S412: Traverse each feature value in the feature value set of each feature, take the feature as a split node, take the feature as the split value, and for the split node, assign the training samples in the training feature matrix whose feature value is less than the split value to the left node, and assign the remaining samples to the right node.

[0110] S413: Determine if the training samples of the left and right nodes are zero. If both the left and right nodes have non-zero training samples, calculate the corresponding loss function value after this split and execute S414; otherwise, return to step S412. The loss function is:

[0111]

[0112]

[0113] In the formula, Let x represent the i-th training sample. i The loss function, where n represents the number of training samples, p ij Indicates training sample x i The probability of belonging to the j-th category, y ij The training sample x represents the training sample x. i The true category label of the j-th category, α i Indicates training sample x i The Dirichlet distribution parameters, α i =[α i1 ,α i2 ,...α ij ...,α iK ] = e i +1,j=1,...,K, where K represents the number of categories, e i Indicates training sample x i Evidence vector, e i =[e i1 ,e i2 ,..e ij ..,e iK ],j=1,...,K, Indicates training sample x i In this iteration of training, the updated predicted class labels of the current decision tree are... Indicates training sample x i In the previous iteration of training, the predicted class label is used; learning_rate represents the learning rate. Indicates training sample x i In this iteration of training, the weights of the leaf nodes of the current decision tree are given, where l represents the l-th iteration of training; B(α) i ) indicates that the parameter is α i The multivariate Beta distribution, p i Indicates training sample x i The predicted class probability, S i Indicates training sample x i The Dirichlet intensity, where exp represents the exponential operation;

[0114] In this embodiment, training sample x i Trust vector b i =[b i1 ,b i2 ,...,b ij ...,b iK ],j=1,...,K, where, Training sample x i The Dirichlet intensity is expressed as

[0115] S414: Based on the calculated loss function value, use gradient descent to solve for the weights of the left and right nodes. Using the calculated weights of the left and right nodes, recalculate the loss function value. Alternately optimize the weights of the left and right nodes until the loss function converges. Save the optimized weights of the left and right nodes, as well as the corresponding loss function value.

[0116] In this embodiment, the gradient descent method is used to solve for the weight ω of the left node. left The weight ω of the right node right , where ω left and ω right The initial value is 0, the gradient update step size is 0.1, and the current loss function has a weight ω for the left leaf node. left The gradient of is calculated using the following formula:

[0117]

[0118] Update ω using gradient left for: Use the updated ω again left Recalculate the loss function, and then calculate the weight ω of the loss function with respect to the right node based on the recalculated loss function. right The gradient of ω, and with respect to ω right The process involves updating and repeatedly optimizing the weights of the left and right nodes alternately until the loss function converges, yielding the optimized weights of the left and right nodes. The weight ω of the right node... right The gradient calculation method is related to the weight ω of the left node. left The calculation method is similar.

[0119] S415: After traversing each feature value in the feature value set of each feature, determine the feature that minimizes the loss function value and its corresponding feature value. Use this feature as the node of the current decision tree, and the corresponding feature value as the split value of the node. Save the weights of the left and right nodes of this node as the weights of the leaf nodes of this node.

[0120] S42: Repeat S41 to split each node obtained by splitting in turn until the maximum tree depth is reached, and a decision tree is obtained.

[0121] In this embodiment, the nodes obtained from the splitting are split sequentially in the order of left, middle, and right. During the splitting of the current node, the splitting value determined in the previous split is used. Samples with a feature value less than the splitting value are assigned to the left side, and ω is used as the splitting point. left The leaf node weight is used as the weight value for this child node; the remaining samples are assigned to the right node, using ω. right This serves as the weight value of the leaf node of the child node.

[0122] S43: Keep the established decision tree unchanged, repeat S41-S42 until the number of decision trees corresponding to the number of categories is obtained, calculate the loss function value of the current iteration training, update the predicted category label of the current iteration training, and complete one iteration training. The initial value of the predicted category label is in the form of an all-zero vector.

[0123] In this embodiment, the predicted class label for each training sample is updated according to the following formula:

[0124]

[0125] In the formula, x represents the training sample in the current iteration of training. i Predicted category labels, x represents the training sample from the previous iteration of training. i The predicted class label, learning_rate represents the learning rate, ω i This represents the training sample x obtained in the current iteration of training. i The leaf node weight vector, ω i =[ω i1 ,ω i2 ,..ω ij ..,ω iK ],j=1,...,K.

[0126] S44: Repeat S41-S43 for the next iteration of training until the preset number of iterations is reached. Save the split features and corresponding feature values ​​of each node of the decision tree in each iteration process, as well as the leaf node weights of each node, to obtain a reliable XGBoost classifier that has been trained.

[0127] In the iterative training, the leaf node weights of each training sample after the previous iteration are used as the initial values ​​of the leaf node weights in the current iteration. The initial values ​​of the leaf node weights of each training sample in the first iteration are in the form of a vector of all zeros.

[0128] Furthermore, based on the trained reliable XGBoost classifier, a detailed explanation is provided on how to combine the preliminary classification results of the reliable XGBoost classifier to perform inter-frame decision fusion to obtain the classification results of the micro-motion target to be tested.

[0129] In an optional implementation, step 3 includes:

[0130] Step 3.1: Input the feature matrix to be tested into the trained reliable XGBoost classifier. Using all the decision trees obtained from each iteration of training, obtain the weight vector of the leaf node of the feature matrix to be tested in each decision tree, calculate the corresponding predicted class probability and uncertainty result, and output it as the preliminary classification result.

[0131] In this embodiment, the calculation of the predicted category probability and uncertainty results includes the following steps:

[0132] Step a: Based on the weight vectors of the feature matrix to be tested in the leaf nodes of each decision tree, calculate the corresponding predicted label by summing them. Based on predicted labels Obtain the corresponding evidence vector e L ,in,

[0133]

[0134] e l j =[e l 1,e l 2,..e l j ..,e l K ],j=1,...,K,l=1,...,L(6);

[0135]

[0136] In the formula, L represents the number of training iterations, and ω l ω represents the weight vector of the leaf nodes of all decision trees obtained from the l-th iteration of training of the feature matrix to be tested. l =[ω l 1,ω l 2,..ω l j ..,ω l K ],j=1,...,K, Let represent the predicted label obtained using the decision tree trained in the l-th iteration, where This means using the decision tree obtained from the l-th iteration of training to determine the predicted label of the feature matrix to be tested belonging to the j-th class;

[0137] Step b: Based on the predicted labels corresponding to the feature matrix to be tested and evidence vector e L The predicted category probabilities and uncertainties are calculated according to the following formula:

[0138] α=[α1,α2,...α j ...α K ] = e L +1,j=1,...,K(8);

[0139]

[0140]

[0141]

[0142] In the formula, p j α represents the probability that the micro-moving target to be tested belongs to each category. j Let represent the j-th element of the Dirichlet distribution parameters of the micro-motion target to be measured, u represent the uncertainty result of the micro-motion target to be measured, S represent the Dirichlet intensity of the micro-motion target to be measured, and e represent the uncertainty result of the micro-motion target to be measured. L j Represents the evidence vector e L The j-th element.

[0143] Step 3.2: Based on the preliminary classification results corresponding to the radar echo data of the target at the current moment and the previous moment, perform inter-frame decision fusion processing using the inter-frame decision fusion formula to obtain the fused classification result. The inter-frame decision fusion formula is as follows:

[0144]

[0145] In the formula, This indicates the classification result of the fusion. This represents the preliminary classification result at time t. This represents the preliminary classification result at time t-1;

[0146] In this embodiment, the preliminary classification results at time t and time t-1 are respectively expressed as follows:

[0147]

[0148]

[0149] b = [b1, b2, ... b j ...,b k],j=1,...,K(15);

[0150]

[0151] In the formula, b represents the confidence vector of the micro-movement target to be measured, b j This represents the j-th element in the trust vector b;

[0152] The fusion classification result is represented as follows:

[0153]

[0154] In the formula, p' j α' represents the probability that the fused micro-movement target belongs to each category. j Let S' represent the j-th element of the Dirichlet distribution parameters of the fused micro-motion target, and let S' represent the Dirichlet intensity of the fused micro-motion target.

[0155] α' j =e' j +1,j=1,...,K(18);

[0156] e' j =b' j ×S',j=1,...,K(19);

[0157]

[0158]

[0159]

[0160] In the formula, e' j Let b' represent the j-th element in the fused evidence vector. j Let represent the j-th element in the fused confidence vector, u' represent the uncertainty result of the fused micro-motion target, and C represent the degree of conflict between the belief functions of the echo data at two time points. This represents the normalization coefficient.

[0161] Step 3.3: Take the category corresponding to the maximum predicted category probability in the fused classification results as the classification result of the micro-movement target to be tested.

[0162] Secondly, this embodiment demonstrates the effectiveness of the radar micro-movement target classification method through simulation experiments using measured data.

[0163] 1. Experimental conditions and experimental content

[0164] The software platform used in this experiment is: Windows 11 operating system, Matlab R2019b, and Python 3.6.

[0165] The hardware platform used in this experiment is: Lenovo IUGBR9H8 workstation, CPU: Intel Core™ i7-10510U

[0166] The dataset used in this experiment is a radar micro-motion target measurement dataset, which consists of four types of targets: four-wheeled vehicles, motorcycles, pedestrians, and quadcopter drones. To ensure data diversity, four-wheeled vehicles include various types, such as large buses and small cars; motorcycle and quadcopter drone data are obtained from at least two different models each; pedestrians are measured from people of different body types, and their movements are mainly normal walking. Vehicle echoes were measured on ordinary roads, with the maximum speed not exceeding the radar's maximum unambiguous speed. This measurement dataset includes various data acquisition scenarios, and the movement patterns of each target are diverse, with different moving targets ranging from 400m to 3km from the radar. The radar used is a narrowband pulse radar, and its specific parameters are shown in the table below.

[0167] Table 1.1 Dataset Description

[0168] Dataset Name Radar target dataset Radar operating system Pulse radar carrier frequency Ku (16GHz) Repetition frequency 5KHz CPI duration 51.2ms bandwidth 75M Distance resolution 2m Radial velocity resolution 0.2m / s Target types Four-wheeled vehicles, motorcycles, pedestrians, quadcopter drones Target distance from radar 400m~3km

[0169] The dataset is processed as follows:

[0170] (1) Based on the trajectory number in the dataset, all data for a given trajectory is divided into training data and test data. The training samples consist of 17 trajectories totaling 1695 frames, and the test samples consist of 4 trajectories totaling 413 frames. All data includes raw radar echo data from four categories: four-wheeled vehicles, motorcycles, pedestrians, and drones. The sample size statistics for each category are shown in Table 1.2.

[0171] (2) Both training and test data use the CLEAN algorithm for ground clutter suppression.

[0172] (3) After removing clutter, feature extraction is performed. When evaluating the model performance, the classification metric used is classification accuracy.

[0173] Table 1.2 Statistics on the number of samples in the four categories of radar target data

[0174] Target Category Number of training samples Number of test samples Four-wheeled vehicle 332 80 motorcycle 357 87 pedestrian 647 160 drones 359 86

[0175] 2. Experimental performance analysis

[0176] The method of this invention aims to obtain an accurate estimate of the uncertainty of the decision result while obtaining the target classification result. Therefore, the experimental evaluation is carried out from two aspects: uncertainty estimation result and classification accuracy.

[0177] (1) Analysis of uncertainty estimation results of reliable XGBoost classifier

[0178] This study compares the accuracy of the reliable XGBoost classifier with other uncertainty measurement methods. The comparison methods are ensemble learning-based uncertainty estimation methods and consistency prediction-based uncertainty estimation methods. The Area Under the Receiver Operating Characteristic curve (AUROC) is used as the evaluation metric to measure the accuracy of the uncertainty measurement; a larger AUROC indicates more accurate uncertainty estimation.

[0179] Table 1.3 Performance Comparison of Different Uncertainty Estimation Methods

[0180] Estimation methods Ensemble learning Consistent prediction method Reliable XGBoost AUROC 0.9035 0.8846 0.9182

[0181] As can be seen, the reliable XGBoost classifier proposed in this invention can better estimate the uncertainty of the prediction results and use it as the basis for judging the reliability of the prediction results.

[0182] (2) Evaluation of classification performance of reliable XGBoost classifier after inter-frame decision fusion

[0183] Table 1.4 Classification accuracy of reliable XGBoost classifier at different inter-frame decision fusion frame numbers

[0184]

[0185] It can be seen that by using a reliable XGBoost classifier to obtain the predicted probability and uncertainty, and then combining it with an inter-frame decision fusion method, the accuracy and reliability of the final classification can be effectively improved. This invention's radar micro-motion target classification method, for the first time, combines an uncertainty estimation method with the XGBoost classifier to achieve a reliable XGBoost classifier. This allows the classifier to obtain both the predicted and uncertainty estimation results for the samples, and to fuse the uncertainty estimation results to assist subsequent decision-making, thereby improving the overall accuracy and reliability of the classification system.

[0186] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or apparatus comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or apparatus that includes said element.

[0187] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A radar micro-motion target classification method combining a reliable XGBoost classifier, characterized in that, include: Step 1: Acquire radar echo data of the micro-moving target to be tested, and perform clutter suppression preprocessing on the radar echo data; Step 2: Perform feature extraction processing on the preprocessed radar echo data to obtain the feature matrix to be measured; Step 3: Use the trained reliable XGBoost classifier to classify the feature matrix to be tested, and obtain the corresponding preliminary classification result. Based on the temporal relationship of the radar echo data of the micro-moving target to be tested, perform inter-frame decision fusion processing of the preliminary classification result to obtain the classification result of the micro-moving target to be tested. The preliminary classification result output by the reliable XGBoost classifier includes predicted class probabilities and uncertainty results. The predicted class probabilities are used to determine the classification decision result, and the uncertainty results are used to evaluate the reliability of the classification decision result. The training process of the reliable XGBoost classifier includes: S41: Based on the input training feature matrix and the true class label, a greedy algorithm is used to traverse the different feature values ​​of each feature and determine the nodes of the decision tree according to the loss function; S42: Repeat S41 to split each node obtained by splitting in turn until the maximum tree depth is reached, and a decision tree is obtained. S43: Keep the established decision tree unchanged, repeat S41-S42 until the number of decision trees corresponding to the number of categories is obtained, calculate the loss function value of the current iteration training, update the predicted category label of the current iteration training, and complete one iteration training. The initial value of the predicted category label is in the form of an all-zero vector. S44: Repeat S41-S43 for the next iteration of training until the preset number of iterations is reached. Save the split features and corresponding feature values ​​of each node of the decision tree in each iteration process, as well as the leaf node weights of each node, to obtain a reliable XGBoost classifier that has been trained. In the iterative training, the leaf node weights of each training sample after the previous iteration are used as the initial values ​​of the leaf node weights in the current iteration. The initial values ​​of the leaf node weights of each training sample in the first iteration are in the form of a vector of all zeros.

2. The radar micro-motion target classification method combined with a reliable XGBoost classifier according to claim 1, characterized in that, Step 1 includes: The radar echo data of the micro-moving target to be measured is collected, and clutter suppression preprocessing is performed on the radar echo data using the regional CLEAN method or moving target indication technology.

3. The radar micro-motion target classification method combining a reliable XGBoost classifier according to claim 1, characterized in that, Step 2 includes: Step 2.1: Based on the modulation characteristics of the preprocessed radar echo data, extract feature values ​​that reflect the structural and motion information of the micro-moving target; Step 2.2: Normalize the eigenvalues ​​to obtain the feature matrix to be tested.

4. The radar micro-motion target classification method combining a reliable XGBoost classifier according to claim 1, characterized in that, The training process of the reliable XGBoost classifier includes: S1: Obtain radar echo data of multiple categories of micro-moving targets as training samples to form a training sample set, and set corresponding real category labels for the radar echo data in the training sample set. S2: Perform clutter suppression preprocessing and feature extraction processing on the radar echo data in the training sample set in sequence to form a preliminary training feature matrix. Normalize the elements in the preliminary training feature matrix to obtain the training feature matrix. S3: Initialize the parameters of the reliable XGBoost classifier, including the number of iterations, maximum tree depth, number of classes, and learning rate; S4: Input the training feature matrix and the true class labels into the reliable XGBoost classifier and perform multiple iterations of training. The forward addition method is used for multiple iterations. In each iteration, a decision tree corresponding to the number of classes is built. Each decision tree is split into nodes in turn. When splitting the current decision tree, the other decision trees remain unchanged.

5. The radar micro-motion target classification method combined with a reliable XGBoost classifier according to claim 4, characterized in that, S41 includes: S411: Determine the unique feature value corresponding to each feature in the training feature matrix, and establish a feature value set corresponding to each feature. ,in, Indicates the first The set of feature values ​​for each feature Indicates the number of features, This represents unique eigenvalues. This represents the number of unique eigenvalues. S412: Traverse each feature value in the feature value set of each feature, take the feature as a split node, take the feature as a split value, and for the split node, assign the training samples in the training feature matrix whose feature values ​​are less than the split value to the left node, and assign the remaining samples to the right node. S413: Determine if the training samples of the left and right nodes are zero. If both the left and right nodes have non-zero training samples, calculate the corresponding loss function value after this split and execute S414; otherwise, return to step S412. The loss function is: ； ； In the formula, Indicates the first training samples loss function, Indicates the number of training samples. Indicates training samples Belongs to the The class probability of each category. Indicates training samples In the The actual category labels for each category, Indicates training samples Dirichlet distribution parameters, , Indicates the number of categories. Indicates training samples The vector of evidence, , , , Indicates training samples In this iteration of training, the updated predicted class labels of the current decision tree are... Indicates training samples The predicted class label from the previous training iteration, Indicates the learning rate. Indicates training samples The weights of the leaf nodes in the current decision tree during this iteration of training. Indicates the first Training iterations; The parameter is The multivariate Beta distribution Indicates training samples Predicted class probability, Indicates training samples Dirichlet strength, Indicates exponentiation; S414: Based on the calculated loss function value, use gradient descent to solve for the weights of the left and right nodes. Using the calculated weights of the left and right nodes, recalculate the loss function value. Alternately optimize the weights of the left and right nodes until the loss function converges. Save the optimized weights of the left and right nodes, as well as the corresponding loss function value. S415: After traversing each feature value in the feature value set of each feature, determine the feature that minimizes the loss function value and its corresponding feature value. Use this feature as the node of the current decision tree, and the corresponding feature value as the split value of the node. Save the weights of the left and right nodes of this node as the weights of the leaf nodes of this node.

6. The radar micro-motion target classification method combined with a reliable XGBoost classifier according to claim 5, characterized in that, In step S43, the predicted class label for each training sample is updated according to the following formula: ； In the formula, This represents the training samples in the current iteration of training. Predicted category labels, This represents the training samples from the previous training iteration. Predicted category labels, Indicates the learning rate. This represents the training samples obtained in the current iteration of training. The leaf node weight vector, .

7. The radar micro-motion target classification method combining a reliable XGBoost classifier according to claim 6, characterized in that, Step 3 includes: Step 3.1: Input the feature matrix to be tested into the trained reliable XGBoost classifier. Using all the decision trees obtained from each iteration of training, obtain the weight vector of the feature matrix to be tested in the leaf nodes of each decision tree, calculate the corresponding predicted class probability and uncertainty result, and output it as the preliminary classification result. Step 3.2: Based on the preliminary classification results corresponding to the radar echo data of the target micro-motion at the current time and the previous time, perform inter-frame decision fusion processing using the inter-frame decision fusion formula to obtain the fused classification result. The inter-frame decision fusion formula is as follows: ； In the formula, This indicates the classification result of the fusion. This represents the preliminary classification result at time t. This represents the preliminary classification result at time t-1; Step 3.3: Take the category corresponding to the maximum predicted category probability in the fused classification result as the classification result of the micro-movement target to be tested.

8. The radar micro-motion target classification method combining a reliable XGBoost classifier according to claim 7, characterized in that, In step 3.1, the predicted category probability and uncertainty results are calculated, including: Step a: Based on the weight vectors of the feature matrix to be tested in the leaf nodes of each decision tree, calculate the corresponding predicted label by summing them. According to the predicted label Obtain the corresponding evidence vector ,in, ； ； ； In the formula, Indicates the number of training iterations. Indicates the feature matrix to be measured at the th... The weight vectors of the leaf nodes of all decision trees obtained from the next iteration of training. , Indicates the use of the first The predicted labels obtained from the decision tree trained in the next iteration are as follows: , Indicates the use of the first The decision tree obtained from the second iteration of training determines whether the feature matrix to be tested belongs to the first... Predicted labels for the class; Step b: Based on the predicted labels corresponding to the feature matrix to be tested and evidence vector The predicted category probabilities and uncertainties are calculated according to the following formula: ； ； ； ； In the formula, This indicates the probability that the micro-moving target to be measured belongs to each category. The first parameter represents the Dirichlet distribution parameter of the micro-moving target to be measured. One element, This represents the uncertainty result of the micro-motion target being measured. The Dirichlet intensity represents the intensity of the micro-movement target being measured. Representing the evidence vector The Each element.

9. The radar micro-motion target classification method combining a reliable XGBoost classifier according to claim 8, characterized in that, In step 3.2, the preliminary classification results at time t and time t-1 are expressed as follows: ； ； ； ； In the formula, This represents the confidence vector of the micro-movement target to be measured. Represents the trust vector The first in One element; The classification result of the fusion is expressed as follows: ； In the formula, This represents the probability of the fused micro-movement target belonging to each category. The first parameter represents the Dirichlet distribution parameter of the fused micro-motion target. One element, The fused Dirichlet intensity represents the measured micro-motion target, where, ； ； ； ； ； In the formula, The first element in the fused evidence vector represents the... One element, The first element in the merged trust vector represents the... One element, This represents the uncertainty result of the measured micro-motion target after fusion. This represents the degree of conflict between the belief functions of echo data from two different times. , This represents the normalization coefficient.

Images