Radar plot quality evaluation position optimization method and system based on detection-tracking-learning

Abstract

The application relates to a radar plot quality evaluation position optimization method and system based on detection-tracking-learning, which comprises the following steps: after the change of a radar position environment, taking target plot information of a stable track as positive samples of multi-dimensional attributes; taking plots without association in a plot pool as negative samples of multi-dimensional attributes; performing machine learning model training based on the samples, determining attribute values of the samples, determining the quality level of the plots, judging whether there is an effective associated plot in a current antenna cycle input plot sequence, and if yes and unique, taking the plot as a target plot result output; otherwise, using the quality level of the plot and a set low-quality threshold to perform discrimination, taking a non-low-quality plot as a target plot result output, updating the samples based on the discrimination result, and updating the machine learning model. The application effectively enhances the adaptability of a radar plot quality evaluation algorithm and a position environment, and significantly improves combat efficiency such as a detection target detection probability and tracking stability.

Description

Technical Field

[0001] This invention relates to a site optimization method, specifically to a site optimization method for radar spot quality assessment based on detection-tracking-learning. Background Technology

[0002] Site optimization is a crucial means to fully leverage the tactical and technical performance of radar equipment and enhance its combat effectiveness. It represents the optimal match between mission, equipment, and site environment. The effectiveness of site optimization in radar spot quality assessment determines the radar's performance under different combat conditions. As the meteorological / electromagnetic environments faced by radars become increasingly complex and mobility requirements become more stringent, higher demands are placed on the diversity of radar site adaptability measures and the speed of operational adjustments. Therefore, it is necessary to address this by researching and addressing the content, procedures, and methods for general spot quality assessment site optimization adjustments.

[0003] The patent "A Three-Coordinate Radar Track Quality Evaluation Method" (CN108181620A) obtains the final three-coordinate radar track quality according to parameters such as the number of agglomerated range cells and amplitude statistical characteristics, based on the order of range, elevation, and azimuth. However, it is difficult to accurately characterize the boundary between target tracks and clutter tracks by artificially setting thresholds according to predetermined rules. The patent "A Track Filtering Method Based on Support Vector Machine" (CN109613526A) proposes a track filtering method based on support vector machine, which can reliably achieve joint classification of multi-dimensional features of the sample set, effectively remove clutter tracks, and reduce false alarms caused by clutter. However, due to the use of preset sample learning, it is easy to misjudge when the environment changes due to the mismatch between the sample and the new scene. Summary of the Invention

[0004] To overcome the problem of misjudging radar spot attributes due to the mismatch between the detection background and fixed parameters after changes in the radar site environment, this invention proposes a radar spot quality assessment site optimization method and system based on detection-tracking-learning.

[0005] The technical solution to achieve the purpose of this invention is: a radar spot quality assessment site optimization method based on detection-tracking-learning, which realizes online dynamic optimization of spot attribute discrimination, including the following steps:

[0006] After the radar site environment changes, the target point information of the stable track is used as a positive sample of the multidimensional attribute; and the unrelated points in the point pool are used as negative samples of the multidimensional attribute.

[0007] Machine learning models are trained based on positive and negative samples, and the value of each attribute of a sample is determined by the machine learning model.

[0008] The quality level of the dots is determined by the output of the machine learning model;

[0009] The system tracks and acquires associated point tracks, determines whether there are valid associated points in the current antenna cycle input point track sequence, and outputs the target point track result if they exist and are unique; otherwise, it uses the quality level of the point track and the set low quality threshold for discrimination, outputs non-low quality point tracks as target point track results, and updates the samples and machine learning model based on the discrimination results.

[0010] Furthermore, the multidimensional attributes include dot amplitude, dot velocity, environmental information, cloaking flag, clutter flag, Doppler leakage flag, azimuth broadening, pitch broadening, range broadening, azimuth envelope, pitch envelope, range envelope, and the number of echoes crossing the threshold.

[0011] Furthermore, the specific values ​​of each attribute of the sample are determined through a machine learning model, including:

[0012] Initialize the positive sample P and the negative sample N, and then determine the value of each attribute of the sample:

[0013] Calculate the mean vector m of the positive sample P. P The mean vector m of the negative samples N N ;

[0014] Based on the mean vector m P and m N Calculate the within-class scatter matrix s of positive and negative samples. P and s N ;

[0015] Based on the intra-class discreteness matrix s P and s N Calculate the overall class scatter matrix X;

[0016] Determine the optimal projection direction W and classification bias T, and then determine the sample attribute value t = W. T X+T.

[0017] Furthermore, the mean vector m of the positive sample P P The mean vector m of the negative samples N N They are respectively:

[0018]

[0019] Among them, L P L represents the number of positive samples. N w represents the number of negative samples. P w represents the set of positive samples with multidimensional attributes. N This represents the set of negative samples with multidimensional attributes.

[0020] Furthermore, the within-class scatter matrix s of the positive and negative samples P and sN for:

[0021]

[0022] The overall class-specific scatter matrix is: s w =s P +s N .

[0023] Furthermore, the optimal projection direction is:

[0024] Furthermore, the classification bias is:

[0025] Furthermore, the quality level of the dots is:

[0026]

[0027] Where w m Let x represent the weight of the attribute of the m-th sample. m This represents the evaluation result of the m-th sample attribute, where M represents the total number of trace sample attributes, and t m This represents the calculated value of the attribute of the m-th sample in the input point sequence, i.e., the output of the machine learning model. This represents the threshold for the m-th attribute of a sample.

[0028] Furthermore, updating samples based on the discrimination results specifically includes: if a point is judged to be of low quality, but is actually a high-quality point through tracking, then the erroneous sample needs to be corrected; otherwise, no sample correction is required.

[0029] A radar spot quality assessment site optimization system based on detection-tracking-learning includes a tracker, a learner, a detector, and a synthesizer, wherein:

[0030] The tracker is used to track associated point tracks and determine the sample set. If a target exists within the associated gate and the Euclidean distance R between the points in the current input point track sequence and the historical point tracks is less than a set threshold, the target information is sent to the synthesizer. The sample set includes positive and negative samples with multidimensional attributes. Positive samples are target point track information of stable tracks, and negative samples are points without association in the point track pool. The machine learning model is trained through the sample set to determine the value of each attribute of the sample.

[0031] The learner includes a machine learning model, which is trained using a set of samples to determine the value of each attribute of the sample.

[0032] The detector is used to determine the quality level of the dots;

[0033] The synthesizer makes a judgment: if the target information output by the tracker exists and is unique, it is output as the target point result; otherwise, it uses the quality level of the point output by the detector and the set low quality threshold to make a judgment, and outputs the non-low quality point as the target point result. At the same time, it updates the samples and the machine learning model based on the judgment result.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows: by adopting the method described in the present invention, and by using a radar spot quality assessment site optimization method based on detection-tracking-learning, online learning and dynamic optimization of spot quality can be achieved, which can effectively enhance the adaptability of the radar spot quality assessment algorithm to the site environment and significantly improve the operational effectiveness such as the detection probability and tracking stability of the detected target. Attached Figure Description

[0035] Figure 1 This is a flowchart of a radar spot quality assessment site optimization method based on detection-tracking-learning.

[0036] Figure 2 The detector outputs a map showing the distribution of dots before and after filtering. Detailed Implementation

[0037] The present invention will be further explained below with reference to the accompanying drawings and embodiments.

[0038] The implementation process of this invention is as follows: Figure 1 As shown, the preferred implementation process is specifically described as follows:

[0039] Because radar site environments are complex and diverse, mobile early warning and detection radars need to have good site adaptability. To adapt well to these environments, fine-grained adjustments to site parameters are necessary based on the specific conditions. When the site environment changes, the radar first operates in a conventional detection mode for a period, using target point information that has already formed stable tracks to generate a small number of positive samples P (high-quality points); and uses unrelated points in the point pool to generate a large number of negative samples N (low-quality points).

[0040] S2 The learner then uses a multi-dimensional attribute w based on "point amplitude, point velocity, environmental information, cloaking marker, clutter marker, Doppler leakage marker, azimuth broadening, pitch broadening, range broadening, azimuth envelope, pitch envelope, range envelope, and number of echoes crossing the threshold". After the initial sample labeling is completed, the detector is trained in the first frame using the initialized positive samples P and negative samples N. First, the mean vector m of the positive samples P is calculated. P The mean vector m of the negative samples N N :

[0041]

[0042] Among them, L P L represents the number of positive samples. N w represents the number of negative samples. P w represents the set of multidimensional attributes of positive samples. N This represents the set of multidimensional attributes of negative samples. Then, the within-class scatter matrix s of positive and negative samples is calculated. P and s N :

[0043]

[0044] Calculate the overall within-class scatter matrix:

[0045] s w =s P +s N

[0046] Calculate the optimal projection direction:

[0047]

[0048] Calculate the classification bias of positive and negative samples of the dot pattern:

[0049]

[0050] Obtain the sample attribute values ​​of the dot sequence, t = W T X+T, and then use the comprehensive constraint rules of point association and quality level to judge the detector output results. If the detector result is a low quality point, but the tracker can stably track and correctly associate, then the erroneous samples should be corrected and the learner should be updated with a new sample library.

[0051] Based on the premise that the target track is visible and the inter-frame motion is limited, the S3 tracker determines whether there are valid associated points in the current antenna period input track sequence for associated point tracks. If a target exists within the associated gate, and the Euclidean distance R between the point in the current input track sequence and the historical point track is less than a set threshold T, the tracker will be activated. dis =800 meters, then the target information will be sent to the synthesizer.

[0052] The S4 detector calculates the quality level of the input point sequence based on multiple attributes, including point amplitude, point velocity, environmental information, cloaking marker, clutter marker, Doppler leakage marker, azimuth broadening, elevation broadening, range broadening, azimuth envelope, elevation envelope, range envelope, and number of echoes crossing the threshold.

[0053]

[0054] Where w m Let x represent the weight of the attribute of the m-th sample.m This represents the evaluation result of the m-th sample attribute, where M represents the total number of trace sample attributes, and t m X represents the calculated value of the attribute of the m-th sample in the input point sequence (determined by the output of the learned model). m This represents the attribute of the m-th sample in the input point sequence X in a subsequent frame. This represents the threshold for the m-th attribute, which is trained using initial labeled samples provided by the tracker. A typical set of parameters is set: w1 = 0.1, w2 = 0.3, w3 = 0.1, w4 = 0.2, w5 = 0.1, w6 = 0.2, w7 = 0.3, w8 = 0.2, w9 = 0.3, w... 10 =0.3, w 11 =0.2, w 12 =0.3, w 13 =0.3; m∈[1,13].

[0055] The S5 synthesizer integrates the tracking and detection results. First, it determines whether there is a valid correlated point in the historical point track within the current antenna cycle. If it exists and is unique, it is output as the target point track result; otherwise, it uses the detector result and a low-quality threshold Q. Thr =1.6 is used for discrimination, and non-low-quality dots are output as target dots. The detector outputs the dot distribution information before and after filtering, as follows: Figure 2 As shown. The synthesized result is then used to update the learner.

Claims

1. A radar spot quality assessment site optimization method based on detection-tracking-learning, characterized in that, Includes the following steps: After the radar site environment changes, the target point information of the stable track is used as a positive sample of the multidimensional attribute; and the unrelated points in the point pool are used as negative samples of the multidimensional attribute. Machine learning models are trained based on positive and negative samples, and the value of each attribute of a sample is determined by the machine learning model. The quality level of the dots is determined by the output of the machine learning model; Track and acquire associated point tracks, determine whether there are valid associated points in the current antenna cycle input point track sequence, and output the target point track result if there are valid and unique associated points. Otherwise, the quality level of the dots is used to make a judgment based on the set low quality threshold. Non-low quality dots are output as the target dots. At the same time, the samples are updated and the machine learning model is updated based on the judgment results. Determining the value of each attribute of a sample using a machine learning model specifically includes: Positive samples and negative samples Initialization, followed by determining the value of each attribute of the sample: Calculate positive samples mean vector and negative samples mean vector ; Based on mean vector and Calculate the within-class scatter matrix of positive and negative samples. and ; Based on the intra-class discreteness matrix and Calculate the overall class scatter matrix X; Determine the optimal projection direction W and classification bias T, and then determine the sample attribute values. .

2. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, The multidimensional attributes include dot amplitude, dot velocity, environmental information, cloaking marker, clutter marker, Doppler leakage marker, azimuth broadening, pitch broadening, range broadening, azimuth envelope, pitch envelope, range envelope, and the number of echoes crossing the threshold.

3. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, The positive sample mean vector and negative samples mean vector They are respectively: in, Indicates the number of positive samples. Indicates the number of negative samples. Represents a set of positive samples with multidimensional attributes. This represents the set of negative samples with multidimensional attributes.

4. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, The intra-class scatter matrix of the positive and negative samples and for: The overall class-specific scatter matrix is: .

5. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, The optimal projection direction is: .

6. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, The classification bias is: .

7. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, The quality level of the dots is: in Indicates the first The weights of each sample attribute, Indicates the first Evaluation results of individual sample attributes This represents the total number of attributes for the dot sample. Indicates the input point sequence number 1 The calculated value of each sample attribute, i.e., the output of the machine learning model. Represents the sample's first Thresholds for each attribute.

8. The radar spot quality assessment site optimization method based on detection-tracking-learning according to claim 1, characterized in that, Updating samples based on the discrimination results specifically includes: if a point is judged to be of low quality, but is actually a high-quality point through tracking, then the incorrect sample needs to be corrected; otherwise, no sample correction is required.

9. A radar spot quality assessment site optimization system based on detection-tracking-learning according to the radar spot quality assessment site optimization method described in claim 1, characterized in that, It includes a tracker, a learner, a detector, and a synthesizer, among which: The tracker is used to track associated point tracks and determine a sample set if a target exists within the associated gate and the Euclidean distance between points in the current input point track sequence and historical point tracks is equal. If the value is less than a set threshold, the target information is sent to the synthesizer. The sample set includes positive and negative samples with multidimensional attributes. Positive samples are target point information of stable tracks, and negative samples are unrelated points in the point pool. The machine learning model is trained through the sample set to determine the value of each attribute of the sample. The learner includes a machine learning model, which is trained using a set of samples to determine the value of each attribute of the sample. The detector is used to determine the quality level of the dots; The synthesizer makes a judgment: if the target information output by the tracker exists and is unique, it is output as the target point result; otherwise, it uses the quality level of the point output by the detector and the set low quality threshold to make a judgment, and outputs the non-low quality point as the target point result. At the same time, it updates the samples and the machine learning model based on the judgment result.

Images