A Radar Target Track Initiation Method Based on Quantum Particle Swarm and LGBM

Abstract

The present invention proposes a radar target track initiation method based on quantum particle swarm and LGBM, including: acquisition and marking of original data; feature extraction of data, establishing an initial feature vector and further extracting target features from the spatio-temporal dimensions respectively; establishment and training of the LightGbm model, using grid search and cross-validation for parameter selection of the model; optimization of the probability threshold, establishing a function of the objective function with respect to the probability threshold parameter, using the QPSO algorithm for parameter optimization, and obtaining the probability threshold when the objective function converges to the maximum value; deployment and application of the LightGbm model. This method is based on extracting multi-dimensional features of the target's initial track, establishing a LightGbm model for false discrimination of the target track, and using the QPSO algorithm to optimize the probability threshold parameter, so as to maximize the track initiation efficiency.

Description

Technical Field

[0001] The present invention relates to a method for starting a radar target track, in particular to a method for starting a radar target track based on quantum particle swarm and LGBM. Background Art

[0002] Radar target tracking technology has been widely applied in military, civilian and other fields. Track initiation is one of the important links in target tracking technology. Good track initiation can lay a foundation for the subsequent stable tracking of targets. Track initiation is a process of automatically establishing a track for a target entering the radar coverage area. As Figure 1 shown, it mainly includes three aspects: candidate track generation, track initiation judgment and formal track establishment. The quality of track initiation is mainly evaluated from the following two aspects: one is to establish a true track as accurately as possible, and the other is to avoid establishing false tracks as much as possible. This means that high-quality track initiation corresponds to a low miss rate and a low false alarm rate, which is very important for realizing fast and stable tracking of multiple targets. Due to the influence of sea clutter interference, enemy electronic interference, echo loss and target maneuvering, it is still very difficult to achieve fully automatic track initiation, and it is necessary to further explore track information and establish a more suitable initiation discrimination model.

[0003] Classic track initiation methods are mainly divided into two categories: sequential processing technology and batch processing technology. Among them, representative methods of sequential processing technology include heuristic rule method, logic method, etc. Its basic idea is to process the echo data obtained from each scan one by one, and judge whether to establish a track according to the point track association result within a certain time window. Sequential processing technology has a low computational load and is suitable for the case of weak background clutter. Representative methods of batch processing technology include Hough transform method and its improved algorithms. Its main idea is to jointly process the echo data obtained from multiple scans, and use Hough transform to achieve non-coherent accumulation of echo signals, so as to improve the track initiation performance. Batch processing technology is suitable for the case of strong clutter background, but due to its large computational load and weak real-time performance, it is often used for offline processing. In a complex background, it is very difficult for these two types of algorithms to achieve real-time and fully automatic track initiation in a complex background.

[0004] Machine learning algorithms have been preliminarily applied to the problem of track initiation. This type of method regards the problem of track initiation as a problem of judging whether a target is real based on radar observation combined data, which belongs to the category of classification. Compared with traditional decision-making methods, machine learning algorithms have better adaptability and decision-making accuracy. However, at present, neither the initial track feature information has been deeply explored, nor the different effects brought by the miss rate and false alarm rate have been considered. Summary of the Invention

[0005] Objective of the Invention: The technical problem to be solved by the present invention is to provide a radar target track initiation method based on quantum particle swarm and LGBM in view of the deficiencies of the prior art.

[0006] To solve the above technical problem, the present invention discloses a radar target track initiation method based on quantum particle swarm and LGBM, including the following steps:

[0007] Step 1: Record the initial track information that meets the basic starting batch conditions and label it, and form an initial feature matrix and its corresponding label value after preliminary processing;

[0008] Step 2: Obtain the features in the time and space dimensions of the initial track through statistical calculation, and combine them with the initial feature matrix to form the final input feature data;

[0009] Step 3: Establish a LightGbm model according to the final input feature data and the corresponding labels, perform model training, use grid search and cross-validation to select the parameters of the model, and save the optimal LightGbm model during training;

[0010] Step 4: Calculate the correct rate P1 of the true track and the correct rate P2 of the false track according to the probability output of the optimal LightGbm model and the probability threshold, design an objective function, and use the QPSO algorithm to optimize the objective function to obtain the optimal probability threshold;

[0011] Step 5: Apply the LightGbm model for online real-time prediction, and perform statistical calculation on the time predicted by the LightGbm model to verify the online real-time prediction ability of the LightGbm model;

[0012] Step 6: Regularly update the LightGbm model, regularly check whether the time accumulation reaches the LightGbm model update cycle or check whether the prediction error of the LightGbm model reaches the LightGbm model update condition. If the update cycle or update condition is reached, update it.

[0013] Advantageous Effects:

[0014] (1) In view of the scenarios of the radar in complex environments, such as strong sea clutter, artificial electronic interference, etc., reduce problems such as many false batches, track interruption, and many jump points existing in the track initiation process.

[0015] (2) Transform the track initiation problem into a common classification problem in machine learning, deeply mine the feature information in the initial track, and use the intelligent model LightGbm for prediction and classification. Compared with the traditional method, the LightGbm model fully combines various features for non-linear judgment, and can achieve better results.

[0016] (3) The traditional method uses prior knowledge to manually select the classification threshold, while the machine learning method is used to replace it, significantly reducing the dependence on prior information and improving the self - adaptability of the model.

[0017] (4) The objective function distinguishes the different costs brought by the correct starting true track and the wrong starting false track. Under the given cost - weighted combination condition, the QPSO algorithm is used to obtain the optimal probability threshold to achieve the optimal effect. Brief Description of the Drawings

[0018] The following further specifically describes the present invention in conjunction with the drawings and specific embodiments, and the above - mentioned and / or other advantages of the present invention will become clearer.

[0019] Figure 1 It is a schematic block diagram of the target track starting process.

[0020] Figure 2 It is a flowchart of the intelligent full - automatic track starting method based on feature extraction in the complex background environment of the present invention.

[0021] Figure 3 It is a screenshot of the full - automatic starting and tracking of an implementation example of the present invention applied in a certain radar tracking system. Specific Embodiments

[0022] The present invention proposes an intelligent full - automatic radar target track starting method in a complex environment, that is, a radar target track starting method based on quantum particle swarm and LGBM (Light Gradient Boosting Machine), which is used to improve the track starting efficiency of radar targets in a complex background environment. The specific steps are as follows:

[0023] Step 1: Record the initial track information that meets the basic starting batch conditions through the display control software and perform manual tagging, and form an initial feature matrix and its corresponding label values after preliminary processing.

[0024] Step 2: Obtain the features of the initial track in the time and space dimensions through statistical calculation, and combine them with the initial features to form the final input feature data.

[0025] Step 3: Establish a LightGbm integrated learning model (Reference: Qi M. LightGBM: A Highly Efficient Gradient Boosting Decision Tree [C] / / Neural Information Processing Systems. Curran Associates Inc. 2017.), train the model, select the model parameters using grid search and cross-validation, and save the optimal model;

[0026] Step 4: Calculate the correct rate P1 of the true track and the correct rate P2 of the false track based on the model probability output and the probability threshold, design an objective function based on P1 and P2, and use the QPSO (Quantum-behaved Particle Swarm Optimization) algorithm to optimize the objective function (Reference: Sun J, Xu W, Feng B. A global search strategy of quantum-behaved particle swarm optimization [C] / / Proc IEEE Conference on Cybernetics & Intelligent Systems. 2004.) to obtain the optimal probability threshold;

[0027] Step 5: Apply the model for online real-time prediction. To verify the online real-time prediction ability of the model, we calculate the model prediction time statistics on the corresponding platform;

[0028] Step 6: Regularly update the model, and regularly check whether the model prediction error meets the conditions for model update.

[0029] In Step 1; meeting the basic batch starting condition means that the number of associated points included in the initial track reaches the batch starting point requirement, and the batch starting point requirement can be specified according to actual needs. In the application of the present invention in engineering, it is set to 5;

[0030] In Step 1; the method of manual labeling is mainly through manual observation. Since we set that the starting target needs to basically conform to uniform linear motion in the batch starting stage, the original video information and the continuity and jitter of each cycle track can be referred to determine the true track and the false track, so as to label each track;

[0031] In Step 1; the initial processing to obtain the initial feature matrix information mainly includes the following two aspects of information:

[0032] On the one hand, it contains the latest N trace information, where N is set as the batch starting point requirement, and each trace information includes: distance r, azimuth a, amplitude amp, range width re, azimuth width ae, and number of time frames t.

[0033] On the other hand, it contains other information such as the radar working mode workmode and weather environment envi information. When in use, one-hot encoding needs to be performed on the working mode and weather environment. For example, when the radar working mode includes three types: short range, medium range, and long range, the three working modes after one-hot encoding are [1, 0, 0], [0, 1, 0], [0, 0, 1] respectively; the weather environment information can include three types: sunny, rainy / snowy, and cloudy. Similarly, the three weather information after one-hot encoding are [1, 0, 0], [0, 1, 0], [0, 0, 1] respectively.

[0034] Then the initial feature matrix is:

[0035] M1 = [r1, r2, r3, r4, r5, a1, a2, a3, a4, a5, amp1, amp2, amp3, amp4, amp5, re1, re2, re3, re4, re5, ae1, ae2, ae3, ae4, ae5, t1, t2, t3, t4, t5, workmode, envi];

[0036] In step 1; the label values are 0 and 1, where 0 represents a false track and 1 represents a true track;

[0037] In step 2; statistical analysis and calculation are performed on the original data from two dimensions of time and space, and the further obtained feature information includes the following:

[0038] Time dimension: amplitude variance varAmp, amplitude mean avgAmp, range width variance reVar, range width mean avgRe, azimuth width variance varAe, azimuth width mean avgAe, average associated distance avgConnDis, association probability cibnnProba;

[0039] Space dimension: radial velocity vRad, tangential velocity vTan, radial velocity change ΔvRad, tangential velocity change ΔvTan, point-track associated distance connDis;

[0040] The calculation formulas for the features are as follows:

[0041] Amplitude mean:

[0042] Amplitude variance:

[0043] Range width mean:

[0044] Range width variance:

[0045] Azimuth width mean:

[0046] Azimuth width variance:

[0047] The i-th associated range:

[0048] Associated range mean:

[0049] Association probability:

[0050] Radial velocity:

[0051] Tangential velocity:

[0052] Adjacent radial velocity difference: ΔvRad j = |vRad j+1 - vRad j | j ≤ N - 2

[0053] Adjacent tangential velocity difference ΔvTan j = |vTan j+1 - vTan j | j ≤ N - 2

[0054] Where, ΔR i represents the difference in radial distance during the association of the i-th track, ΔA i represents the difference in azimuth during the association of the i-th track, r represents the range accuracy of the radar, a represents the azimuth accuracy of the radar, dotCnt represents the number of associated points, period represents the number of processing cycles, ΔvRad j represents the difference between the (j + 1)-th radial velocity and the j-th radial velocity, ΔvTan j represents the difference between the (j + 1)-th tangential velocity and the j-th tangential velocity.

[0055] In step 3, the parameter setting method of the LightGbm model is as follows: Fix the parameters determined by the model according to empirical knowledge, and use the grid search and cross-validation methods to adjust the parameters that need to be optimized; the reference settings of the model parameters that can be determined are shown in Table 1:

[0056] Table 1 LightGbm model parameter table determined

[0057] Parameter Value Frame type (boosting_type) Gradient Boosting Decision Tree (gbdt) Task (objective) Binary classification (binary) Evaluation function (metrics) binary_logloss Feature fraction used (feature_fraction) 0.8

[0058] The model parameters to be adjusted and optimized and their optimization ranges are shown in Table 2:

[0059] Table 2 LightGbm model parameters to be optimized

[0060] Parameter Optimization range Learning rate (learning_rate) [0.01,0.02,0.05,0.1] Number of base learners (n_estimators) [100,200,500,800] Maximum depth of the tree (max_depth) [3,4,5,6]

[0061] Explanation of optimization parameters:

[0062] learning_rate: Learning rate. The smaller the learning rate, the slower the model training. Generally, choosing a smaller learning rate can obtain better model performance with better stability;

[0063] n_estimators: The number of base learners, which is also the number of boosting iterations. Generally, choosing a larger number of iterations will result in better performance, but if it is too large, it is easy to cause model overfitting;

[0064] max_depth: The maximum depth of the tree model, generally set between 3 and 6. This parameter is an important parameter to prevent model overfitting and has a decisive impact on model performance and generalization ability;

[0065] Use five-fold cross-validation (cross_validation), and traverse the above given parameter combinations through the method of grid search (reference: SVM mill load prediction based on grid search and cross-validation [J]. Chinese Journal of Testing, 2017(1).) to select the optimal parameters and save the corresponding optimal model;

[0066] In step 4, the steps to calculate the correct rate P1 of the true track and the correct rate P2 of the false track based on the model probability output and the probability threshold are as follows:

[0067] In step 4-1, according to the trained optimal model, we can obtain the probability output proba of the kth training sample k , in order to obtain the final classification result labelpred k , it is necessary to further determine the probability threshold α, and judge according to the following formula:

[0068]

[0069] In step 4-2, according to the true labels and predicted labels of all samples, calculate P1 and P2 as follows:

[0070]

[0071] Among them, TP represents the number of true tracks that are predicted as true tracks by the model among all training samples, TN represents the number of false tracks that are predicted as false tracks by the model among all training samples, FP represents the number of false tracks that are predicted as true tracks by the model among all training samples, and FN represents the number of true tracks that are predicted as false tracks by the model among all training samples;

[0072] In step 4, the design of the objective function is mainly based on the objectives of track initiation: one is to establish true tracks as accurately as possible, and the other is to establish false tracks as few as possible. Therefore, it is necessary to take these two parts of the cost into consideration and form the final objective function based on P1 and P2 obtained in the previous step, as shown below:

[0073] Fitness function =(λ*P1+P2) / (λ+1), λ>0

[0074] Among them, λ is the weight coefficient. The larger λ is, the more we value the accuracy rate of true tracks, that is, we hope that as many true tracks as possible can be normally initiated, representing that the omission rate is as low as possible; the smaller λ is, the more we value the accuracy rate of false tracks, that is, we hope that as few false tracks as possible are initiated, representing that the false alarm rate is as low as possible. Combining the needs of users, what users hope more is not to miss any true tracks under the premise of a certain false alarm rate, that is, a small number of false tracks are allowed to be initiated, because a small number of false tracks can also be judged as true or false and deleted through the track evaluation module after further data accumulation, so the impact is small. Therefore, λ can be set to a larger value. In the engineering application of the present invention, it is set to 5. Then the final optimized objective function can be:

[0075] Fitness function =(5*P1+P2) / (5+1)

[0076] Fitness function =(5*TP / (TP+FN)+TN / (TN+FP) / 6

[0077] In step 4, to obtain the optimized threshold α, the QPSO algorithm is used to solve the maximum value of the objective function. The settings of the QPSO model parameters are shown in Table 3:

[0078] Table 3 QPSO model parameter settings

[0079] Parameter Value Number of particles 50 Particle dimension 1 Control coefficient 0.7 Maximum number of iterations 100 Maximum value of probability threshold α 1 Minimum value of probability threshold α 0

[0080] Obtain the probability threshold α0 when the objective function converges to the maximum value;

[0081] In step 5, the steps for online real-time calling the model to implement classification prediction include:

[0082] In step 5-1, load the trained optimal model;

[0083] In step 5-2, for each newly arrived track sample, process its original data to obtain the initial feature matrix, the time-dimensional feature matrix, and the space-dimensional feature matrix, and merge them to obtain the final model input feature matrix;

[0084] In step 5-3, call the model to obtain the predicted probability output. By comparing it with the optimized threshold α0, if the predicted probability output is greater than the threshold α0, it is judged as a real track; if the predicted probability output is less than the threshold α0, it is judged as a false track;

[0085] In step 5, to verify the real-time performance of the model, on the domestic platform, the total time required to predict 5000 samples is statistically 250 ms, so the time required to predict each sample is 0.05 ms, meeting the real-time requirement.

[0086] In step 6, the regular update of the model mainly judges two conditions: whether the time accumulation reaches the model update period or statistically the prediction accuracy of the model in the recent period.

[0087] The present invention is an intelligent fully automatic track initiation method based on feature extraction. By combining multi-dimensional features of manually extracted initial tracks, a track discrimination LightGbm model is established. The objective function considers the different costs brought by the correct discrimination rate of real tracks and the correct discrimination rate of false tracks, and uses the QPSO algorithm for optimization to obtain the optimal probability threshold, finally realizing the optimal fully automatic track initiation, greatly reducing the dependence on prior knowledge, improving the self-adaptability of the algorithm, and ultimately improving the track initiation performance in complex environments.

[0088] Embodiment:

[0089] Combined with Figure 2 The intelligent fully automatic track initiation method based on feature extraction in a complex background environment of the present invention includes the following steps:

[0090] In the first step, establish a fully automatic area in the sea clutter area through the display control software, and record all the initial track information that meets the basic initiation conditions of the radar in different working modes and different working environments. Meeting the basic initiation batch conditions means that the number of associated points included in the initial track reaches the requirement of the initiation batch points, and the requirement of the initiation batch points can be specified according to actual needs. In this project, it is set to 5.

[0091] Through manual observation, combined with the original video information and the continuity and jitter of each cycle track, judge whether the target basically conforms to the uniform linear motion state, so as to judge the authenticity of the track. If the track is a real track, label it with 1; if it is a false track, label it with 0;

[0092] The original track information is initially processed to form an initial feature matrix. The information in the initial feature matrix mainly includes the following two aspects:

[0093] On the one hand, it includes the information of the latest N point traces. N is set as the batch starting point requirement, and each point trace information contains: distance r, azimuth a, amplitude amp, range width re, azimuth width ae, and time frame number t.

[0094] On the other hand, it includes other information such as the radar working mode workmode and weather environment envi information. When using, one-hot encoding needs to be performed on the working mode and weather environment. For example, when the radar working mode includes three types: short range, medium range, and long range, the three working modes after one-hot encoding are [1, 0, 0], [0, 1, 0], [0, 0, 1] respectively; the weather environment information can include three types: sunny, rainy / snowy, and cloudy. Similarly, the three weather information after one-hot encoding are [1, 0, 0], [0, 1, 0], [0, 0, 1] respectively.

[0095] Then the initial feature matrix is organized as:

[0096] M1 = [r1, r2, r3, r4, r5, a1, a2, a3, a4, a5, amp1, amp2, amp3, amp4, amp5, re1, re2, re3, re4, re5, ae1, ae2, ae3, ae4, ae5, t1, t2, t3, t4, t5, workmode, envi].

[0097] In the second step, further statistical analysis is performed on the obtained original data from both the time and space dimensions to further extract feature information, mainly including the following types of features:

[0098] Time dimension: amplitude variance varAmp, amplitude mean avgAmp, range width variance reVar, range width mean avgRe, azimuth width variance varAe, azimuth width mean avgAe, average associated distance avgConnDis, association probability connProba;

[0099] Space dimension: radial velocity vRad, tangential velocity vTan, radial velocity change ΔvRad, tangential velocity change ΔvTan, point-track association distance connDis;

[0100] The calculation formulas for the features are as follows:

[0101] Amplitude mean:

[0102] Amplitude variance:

[0103] Range width mean:

[0104] Range width variance:

[0105] Azimuth width mean:

[0106] Azimuth width variance:

[0107] The i-th associated range:

[0108] Associated range mean:

[0109] Association probability:

[0110] Radial velocity:

[0111] Tangential velocity:

[0112] Adjacent radial velocity difference: ΔvRad j = |vRad j+1 - vRad j | j ≤ N - 2

[0113] Adjacent tangential velocity difference ΔvTan j = |vTan j+1 - vTan j | j ≤ N - 2

[0114] Where, ΔR i represents the difference in radial distance during the association of the i-th track point, ΔA i represents the difference in azimuth during the association of the i-th track point, r represents the range accuracy of the radar, a represents the azimuth accuracy of the radar, dotCnt represents the number of associated points, period represents the number of processing cycles, ΔvRad j represents the difference between the (j + 1)-th radial velocity and the j-th radial velocity, ΔvTan j represents the difference between the (j + 1)-th tangential velocity and the j-th tangential velocity.

[0115] Combine the above features and the initial features obtained in the second step to obtain the final input feature matrix.

[0116] In the third step, establish a LightGbm model, fix the parameters determined by the model according to empirical knowledge, and use grid search and cross-validation methods to adjust the parameters that need to be optimized; the reference settings for the determinable parameters are shown in Table 4 as follows:

[0117] Table 4 Determined LightGbm model parameter table

[0118] Parameter Value Frame type (boosting_type) Gradient Boosting Decision Tree (gbdt) Task (objective) Binary classification (binary) Evaluation function (metrics) binary_logloss Feature fraction used (feature_fraction) 0.8

[0119] The parameters to be adjusted and the optimization range are shown in Table 5 as follows:

[0120] Table 5 LightGbm model parameters to be optimized

[0121] Parameter Optimization range Learning rate (learning_rate) [0.01,0.02,0.05,0.1] Number of base learners (n_estimators) [100,200,500,800] Maximum depth of the tree (max_depth) [3,4,5,6]

[0122] Introduction to optimized parameters:

[0123] learning_rate: Learning rate. The smaller the learning rate, the slower the model training. Generally, choosing a smaller learning rate can obtain better model performance with better stability;

[0124] n_estimators: The number of base learners, which is also the number of boosting iterations. Generally, choosing a larger number of iterations will result in better performance, but if it is too large, it is easy to cause overfitting of the model;

[0125] max_depth: The maximum depth of the tree model, generally set between 3 and 6. This parameter is an important parameter to prevent model overfitting and has a decisive impact on model performance and generalization ability;

[0126] Use five-fold cross-validation (cross_validation), and traverse the above given parameter combinations through the method of grid search (Gride search) to select the optimal parameters and save the corresponding optimal model;

[0127] In the fourth step, according to the trained optimal model, we can obtain the probability output proba of the k-th training sample k , in order to obtain the final classification result labelPred k , it is necessary to further determine the probability threshold α and judge according to the following formula:

[0128]

[0129] According to the true labels and predicted labels of all samples, the correct rate P1 of the true track and the correct rate P2 of the false track can be obtained through the following calculations:

[0130]

[0131] Among them, TP represents the number of true tracks predicted as true tracks by the model among all true tracks, TN represents the number of false tracks predicted as false tracks by the model among all false tracks, FP represents the number of false tracks predicted as true tracks by the model among all false tracks, and FN represents the number of true tracks predicted as false tracks by the model among all true tracks;

[0132] When designing the objective function, two main target tasks of track initiation need to be considered: one is to establish the true track as accurately as possible, and the other is to establish as few false tracks as possible. Therefore, the costs of these two parts need to be taken into account to form the final objective function, as shown below:

[0133] Fitness function =(λ * P1 + P2) / (λ + 1), λ > 0

[0134] Among them, λ is the weight coefficient. The larger λ is, the more we value the correct rate of the true track, that is, we hope that as many true tracks as possible can start normally, which means the omission rate is as low as possible; the smaller λ is, the more we value the correct rate of the false track, that is, we hope that as few false tracks as possible start, which means the false alarm rate is as low as possible. Combining the needs of users, users hope more that no true tracks are missed on the premise of a certain false alarm rate, that is, a small number of false tracks are allowed to start, because a small number of false tracks can be judged and deleted as true or false through the track evaluation module after further data accumulation, so the impact is small. Therefore, λ can be set to a larger value. In the engineering application of the present invention, it is set to 5. Then the final optimized objective function can be:

[0135] Fitness function =(5 * P1 + P2) / (5 + 1)

[0136] Fitness function =(5 * TP / (TP + FN) + TN / (TN + FP)) / 6

[0137] Among them, TP, TN, FN, and FP are all functions of the probability threshold α;

[0138] After obtaining the objective function, the QPSO algorithm is used to optimize the objective function to obtain the optimal probability threshold. The common parameter settings of the QPSO algorithm are shown in Table 6:

[0139] Table 6 QPSO model parameter settings

[0140] Parameter Value Number of particles 50 Particle dimension 1 Control coefficient 0.7 Maximum number of iterations 100 Maximum value of probability threshold α 1 Minimum value of probability threshold α 0

[0141] Obtain the probability threshold α0 when the objective function converges to the maximum value;

[0142] Step 5, apply the model for online real-time prediction, which mainly includes the following steps:

[0143] 1. Load the trained optimal model;

[0144] 2. For each newly arrived track sample, process its original data to obtain the initial feature matrix, time-dimensional feature matrix, and space-dimensional feature matrix, and merge them to obtain the final model input feature matrix;

[0145] 3. Call the model for prediction to obtain the probability output. By comparing it with the probability threshold α0, if the output is greater than the threshold, it is judged as a true track; if the output is less than the threshold, it is judged as a false track.

[0146] In addition, in order to verify the real-time performance of the model to meet the requirements of engineering applications, the total time required to predict 5000 samples was counted on the domestic platform as 250 ms. Then, the time required to predict each sample is 0.05 ms, which meets the real-time requirements of engineering applications.

[0147] Step 6: Regular update of the model. Judge whether the time accumulation reaches the model update period or calculate the prediction efficacy of the model during this period based on the observed real data and the model prediction results. If the time accumulation reaches the model update or the prediction error is greater than the set prediction error threshold, then use all the accumulated data for retraining the model.

[0148] The following is combined with Figure 3 description to further illustrate the present invention through the implementation examples and their effect evaluations applied in the project. Figure 3 It is a screenshot of the radar display during the full-automatic start and tracking process of a certain sea radar in the sea clutter area. Among them, TVXXX is the track number. For example, the track numbered TV0005 indicates that the track with batch number 0005 is in the normal tracking state. Through multi-cycle manual observation and confirmation, all the targets that have been started and tracked in the picture are real targets, and there are basically no false targets throughout the process, indicating that this method can achieve the full-automatic start of target tracks in a complex background environment and can greatly reduce the operation load of radar operators.

[0149] In specific implementation, the present application provides a computer storage medium and a corresponding data processing unit. Among them, the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, it can run the content of the invention of a method for starting radar target tracks based on quantum particle swarm and LGBM provided by the present invention and some or all of the steps in each embodiment. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), etc.

[0150] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention, in essence, or the parts that contribute to the prior art can be embodied in the form of a computer program, that is, a software product. This computer program software product can be stored in a storage medium, including several instructions for causing a device (which can be a personal computer, server, single-chip microcomputer, MUU or network device, etc.) containing a data processing unit to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0151] The present invention provides an idea and method for a radar target track initiation method based on quantum particle swarm and LGBM. There are many methods and ways to specifically implement this technical solution. The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented by the prior art.

Claims

1. A radar target track initiation method based on quantum particle swarm and LGBM, characterized in that, The process of track initiation includes the following steps: Step 1: Record the initial track information that meets the basic batch starting conditions and label it. After preliminary processing, form an initial feature matrix and its corresponding label values; Step 2: Obtain the features of the initial track in the time and space dimensions through statistical calculations, and combine them with the initial feature matrix to form the final input feature data; Step 3: Based on the final input feature data and the corresponding labels, establish a LightGbm model, perform model training, use grid search and cross-validation to select the model parameters, and save the optimal LightGbm model during training; Step 4: Calculate the correct rate P1 of the true track and the correct rate P2 of the false track according to the probability output of the optimal LightGbm model and the probability threshold. Design an objective function, and use the QPSO algorithm to optimize the objective function to obtain the optimal probability threshold; Step 5: Apply the LightGbm model for online real-time prediction, and perform statistical calculations on the time predicted by the LightGbm model to verify the online real-time prediction ability of the LightGbm model; Step 6: Regularly update the LightGbm model. Regularly check whether the time accumulation reaches the LightGbm model update period or whether the prediction error of the LightGbm model reaches the LightGbm model update condition. If the update period or update condition is reached, perform an update.

2. The radar target track initiation method based on quantum particle swarm and LGBM according to claim 1, characterized in that In Step 1, it is described that: Meeting the basic batch starting conditions means that the number of associated points included in the initial track reaches the batch starting point requirement. The batch starting point requirement is specified according to actual needs and is set to 5 in engineering applications; The labeling in Step 1, that is, manual labeling, the method includes: through manual observation, referring to the original video information and the continuity and jitter of each cycle track to determine the true track and the false track, so as to label each track; The preliminary processing in Step 1 is the process of forming the initial feature matrix. The initial feature matrix M1 is as follows: M1 = [r1, r2, …, r i , …, r N , a1, a2, …, a i , …, a N , amp1, amp2, …, amp i , …, amp N , re1, re2, …, re i , …, re N , ae1, ae2, …, ae i , …, ae N , t1, t2, …, t i , …, t N , workmode,envi] Among them, N is the required minimum batch point number, that is, the initial feature matrix M1 contains the latest N point track information in the initial track, r i represents the distance of the i-th point track, a i represents the azimuth of the i-th point track, amp i represents the amplitude of the i-th point track, re i represents the range width of the i-th point track, ae i represents the azimuth width of the i-th point track, t i represents the number of time frames of the i-th point track, workmode represents the radar working mode, and envi represents the weather environment; Perform one-hot encoding on the radar working mode workmode and the weather environment envi; The label values in Step 1 are used to represent false tracks and true tracks.

3. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 2, characterized in that The features of the initial track in the time and space dimensions described in Step 2 include: The features of the initial track in the time dimension include: amplitude variance varAmp, amplitude mean avgAmp, range width variance varRe, range width mean avgRe, azimuth width variance varAe, azimuth width mean avgAe, average associated distance avgConnDis, association probability connProba; The features of the initial track in the space dimension include: radial velocity vRad, tangential velocity vTan, radial velocity change ΔvRad, tangential velocity change ΔvTan, point-track association distance connDis.

4. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 3, characterized in that, The method of obtaining the features of the initial track in the time and space dimensions through statistical calculations described in Step 2 specifically includes: Amplitude mean value: Amplitude variance: Distance width mean: Distance width variance: Mean azimuth width: Azimuth width variance: The i-th associated distance: Mean associated distance: Association probability: Radial velocity: Tangential velocity: Adjacent radial velocity difference: ΔvRad j = |vRad j+1 - vRad j | j ≤ N - 2 Adjacent tangential velocity difference: ΔvTan j = |vTan j+1 - vTan j | j ≤ N - 2 where, ΔR i represents the difference in radial distance when the i-th trace is associated, ΔA i represents the difference in azimuth when the i-th trace is associated, r represents the range accuracy of the radar, a represents the azimuth accuracy of the radar, dotCnt represents the number of associated points, period represents the number of processing cycles, ΔvRad j represents the difference between the (j + 1)-th radial velocity vRad j+1 and the j-th radial velocity vRad j , ΔvTan j represents the difference between the (j + 1)-th tangential velocity vTan j+1 and the j-th tangential velocity vTan j .

5. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 4, characterized in that The method of using grid search and cross-validation for the model parameters described in Step 3 specifically includes: Select the model parameters to be adjusted and optimized in the LightGbm model and their optimization ranges. The model parameters to be adjusted and optimized include: learning rate, number of base learners, and maximum depth of the tree model. Use five-fold cross-validation, and traverse the above parameter combinations through grid search to select the optimal parameters, and save the corresponding model as the optimal model.

6. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 5, characterized in that, In step 4, calculate the correct rate P1 of the true track and the correct rate P2 of the false track based on the probability output and probability threshold of the optimal LightGbm model. The steps are as follows: Step 4-1: Obtain the probability output proba of the k-th training sample according to the trained optimal LightGbm model k , and then the final classification result label Pred k Make a judgment according to the following formula: Where α is the probability threshold. Step 4-2, according to the true labels and predicted labels of all training samples, perform the following calculations to obtain the correct rate P1 of the true track and the correct rate P2 of the false track: P1 = TP / (TP + FN) P2 = TN / (TN + FP) Where TP represents the number of true tracks predicted as true by the optimal LightGbm model among all training samples, TN represents the number of false tracks predicted as false by the optimal LightGbm model among all training samples, FP represents the number of false tracks predicted as true by the optimal LightGbm model among all training samples, and FN represents the number of true tracks predicted as false by the optimal LightGbm model among all training samples.

7. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 6, characterized in that, The objective function Fitness described in step 4 function is expressed as follows: Fitness function =(λ * P1 + P2) / (λ + 1), λ > 0 Where λ is the weight coefficient, and the final optimization objective function is: Fitness function =(λ * TP / (TP + FN) + TN / (TN + FP)) / (λ + 1) Where TP, TN, FN, and FP are all functions of the probability threshold α.

8. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 7, characterized in that The method of using the QPSO algorithm to optimize the objective function in step 4 includes: using the QPSO algorithm to solve the maximum value of the objective function, and the probability threshold α0 when the objective function converges to the maximum value is the optimal probability threshold.

9. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 8, characterized in that, The specific steps of applying the LightGbm model for online real-time prediction in step 5 include: Step 5-1, load the trained optimal LightGbm model. Step 5-2, for each new track sample, process its original data to obtain the initial feature matrix, time-dimensional feature matrix, and space-dimensional feature matrix, and merge them to obtain the final model input feature matrix. Step 5-3, call the loaded model to obtain the predicted probability output, and compare it with the optimized threshold α0. If the predicted probability output is greater than the threshold α0, it is judged as a true track, otherwise it is judged as a false track. Step 5-4, count the total time for predicting a preset number of samples using the loaded model, calculate the time consumed for each sample prediction, and complete the verification of the online real-time prediction ability of the LightGbm model.

10. A radar target track initiation method based on quantum particle swarm and LGBM according to claim 9, characterized in that, The required starting batch point number N in step 1 is set to 5; the weight coefficient λ in step 4 is set to 5.

Images