High-frequency ground wave radar formation target tracking method and system

Abstract

The invention discloses a high-frequency ground wave radar formation target tracking method and system, and particularly relates to the technical field of target tracking. Comprising the following steps: constructing a position feature vector and an appearance feature vector of a formation target based on high-frequency ground wave radar echoes; constructing nodes by using the position feature vectors and the appearance feature vectors of the formation targets; screening adjacent nodes of the known node Node based on the position feature vector, and calculating a likelihood function; constructing a prototype network model based on the formation target appearance feature vector, and calculating the Euclidean distance between the appearance feature vector of a known node Node and the appearance feature vector of an adjacent node Adjacent; on the basis of a likelihood function Lsiran and an Euclidean distance Euddis, the correlation similarity between a known node Node and an adjacent node AdJacent is calculated, so that a connection edge is constructed; and training based on a graph neural network loss function to realize formation target matching tracking. By adopting the technical scheme of the invention, the problem that the formation target cannot be tracked in the prior art is solved, and accurate tracking of special targets such as formation is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of target tracking, and in particular to a high-frequency ground wave radar formation target tracking method and system. Background Art

[0002] High-frequency ground wave radar is widely used in ship target tracking. It is a challenging problem to achieve long-term stable tracking of targets. The initial target tracking strategy is to detect the target before tracking (DBT), which will lead to the problem of weak targets being missed and track interruption. Due to the serious false alarm of target detection and low positioning accuracy, the performance of detection-based tracking is poor.

[0003] Existing high-frequency ground wave radars mainly track single targets by acquiring information about the target's distance, speed, and azimuth. In order to improve target tracking performance: on the one hand, by improving the filter, the detection performance and trajectory consistency can be improved. On the other hand, by improving the trajectory association algorithm, long-term continuous target tracking can be achieved. At present, there are few technologies that combine the target's range Doppler image appearance features to improve target tracking performance. At the same time, there is no research on tracking formation targets. Summary of the invention

[0004] The present invention aims to provide a high-frequency ground wave radar formation target tracking method and system, which solves the problem that the prior art cannot track formation targets.

[0005] In order to achieve the above object, the present invention provides a technical solution as follows: a high-frequency ground wave radar formation target tracking method, comprising the following steps:

[0006] S1. Constructing the position feature vector and appearance feature vector of the formation target based on the high-frequency ground wave radar echo;

[0007] S2, constructing nodes using the position feature vector and appearance feature vector of the formation target;

[0008] S3, based on the position feature vector, filter the adjacent nodes of the known node Node and calculate the likelihood function;

[0009] S4. Construct a prototype network model based on the appearance feature vector of the formation target, and calculate the Euclidean distance between the appearance feature vectors of the known node Node and the adjacent node Adjacent;

[0010] S5, based on the likelihood function Lsiran obtained in step S3 and the Euclidean distance Eud_dis obtained in step S4, calculate the association similarity between the known node Node and the adjacent node Adjacent, so as to construct a connection edge;

[0011] S6. Perform training based on the graph neural network loss function to achieve formation target matching and tracking.

[0012] Furthermore, the method of step S1 is as follows:

[0013] S101: Perform two-dimensional Fourier transform on the time domain echo received by the high-frequency ground wave radar to obtain a range Doppler image;

[0014] S102: using two-dimensional constant false alarm detection on the range Doppler image to obtain the distance r of the formation target relative to the radar and the radial velocity vr of the formation target;

[0015] S103: Estimating the azimuth of the time domain echo received by the high-frequency ground wave radar to obtain the azimuth θ of the formation target relative to the radar;

[0016] S104: Calculate the position and speed of the formation target in the global coordinate system;

[0017] S105: Using the Sobel operator, detecting the horizontal edge and the vertical edge of the range Doppler image obtained in step S101;

[0018] S106: Construct an appearance feature vector of the formation target.

[0019] Furthermore, the calculation method of the position and speed in step S104 is as follows:

[0020] x=r·cosθ

[0021] y=r·sinθ

[0022] v x = vr·cosθ

[0023] v y = vr·sinθ

[0024] Among them, x, y represent the position of the formation target in the global coordinate system, v x ,v y Indicates the velocity of the formation target in the global coordinate system;

[0025] Construct the position feature vector of the formation target:

[0026] Pos=(x,y,v x ,v y ,rlen,dlen)

[0027] Among them, Pos represents the position feature vector of the formation target, rlen represents the number of cells occupied by the formation target in the range direction in the range Doppler spectrum, and dlen represents the number of cells occupied by the formation target in the Doppler direction in the range Doppler spectrum.

[0028] Further, the method of step S105 is as follows:

[0029]

[0030] Where I represents the range Doppler image, Gx represents the gradient of image I in the horizontal direction, and Gy represents the gradient of image I in the vertical direction;

[0031] Represents a horizontal template for detecting horizontal edges of an image. Represents a longitudinal template for detecting vertical edges of an image; It means that the horizontal template is convolved with the range Doppler image obtained in step S101. It means that the longitudinal template is plane-convolved with the range Doppler image obtained in step S101;

[0032] Compute the gradient of the range Doppler image:

[0033] GA=Gx+Gy

[0034] Where GA represents the gradient of the range Doppler image;

[0035] The construction method of step S106 is as follows:

[0036] App=(Gx,Gy,GA,Nf,Rgb,Hsv,Gabor)

[0037] Among them, App represents the appearance feature vector of the formation target, Nf represents the normalized first-order moment feature of the range Doppler image, Rgb represents the grayscale feature of the range Doppler image, Hsv represents the hue, saturation, and value features of the range Doppler image, and Gabor represents the texture feature of the range Doppler image.

[0038] Furthermore, the node construction method of step S2 is as follows:

[0039] Node=(Pos,App,k,formation)

[0040] Among them, k represents the time and formation represents the formation type.

[0041] Furthermore, the method of step S3 is as follows:

[0042] S301: Based on the known position feature vector of the node Node, predict the position feature vector of the next moment:

[0043] Pos_1=F k Pos

[0044] Among them, Pos_1 represents the predicted value of the position feature vector at the next moment, and F k represents the state transfer matrix;

[0045] S302: Record the nodes satisfying abs(Pos_adj-Pos_1)<Dis as the adjacent nodes Adjacent of the known node Node, where Pos_adj represents the position feature vector of the adjacent node Adjacent, Dis represents the threshold value, and abs(Pos_adj-Pos_1) represents the distance between Pos_adj and Pos_1;

[0046] S303: Calculate the position vector likelihood function of the known node Node and the adjacent node Adjacent:

[0047]

[0048] Among them, Lsiran represents the likelihood function, H represents the transformation matrix, and H T represents the transposed matrix of matrix H, P represents the covariance matrix, R represents the observation noise covariance, and σ represents the standard deviation of the measurement distribution of the high-frequency ground wave radar at the next moment.

[0049] Furthermore, the method of step S4 is as follows:

[0050] S401: Based on the range Doppler images of the single longitudinal formation, the V-formation and the symmetrical formation, the appearance feature vector is extracted and a training data set is constructed;

[0051] Train the prototype network with Resnet as the backbone network to obtain the prototype network model;

[0052] S402: using a clustering algorithm to classify the appearance feature vectors of the known node Node, calculating the average value of each category, and recording it as a prototype vector;

[0053] Calculate the Euclidean distance Eud_dis between the appearance feature vector of the adjacent node Adjacent and the prototype vector.

[0054] Further, the method of step S5 is as follows:

[0055] S501: Based on the likelihood function Lsiran obtained in step S3 and the Euclidean distance Eud_dis obtained in step S4, the association similarity between the known node Node and the adjacent node Adjacent is calculated. The calculation formula is as follows:

[0056] Simi=Qz1·Lsiran+Qz2·Eud_dis

[0057] Among them, Simi represents the association similarity between the known node Node and the adjacent node Adjacent, Qz1 represents the weight of the likelihood function Lsiran, and Qz2 represents the weight of the Euclidean distance Eud_dis;

[0058] Based on the association similarity, an edge is established between a known node Node and an adjacent node Adjacent.

[0059] Further, the method of step S6 is as follows:

[0060] S601: Construct graph neural network loss function:

[0061] Lsunsi=Lnode+Qz3·Line

[0062] Among them, Lsunsi represents the graph neural network loss function, Lnode represents the node loss function, Line represents the edge loss function, and Qz3 represents the weight of the edge loss function;

[0063] Train the graph neural network and obtain the edges with the highest similarity scores and the corresponding adjacent nodes Adjacent through the known nodes Node;

[0064] S602: Update the position feature vector of the adjacent nodes:

[0065] Pos_2=Pos_1+Kalz(Pos_adj-H·Pos_1)

[0066] Among them, Pos_2 represents the updated value of the position feature vector of the adjacent node, and Kalz represents the Kalman gain.

[0067] Another technical solution provided by the present invention is a high-frequency ground wave radar formation target tracking system, comprising:

[0068] Feature vector construction module: Calculate the position and velocity of the formation target based on the high-frequency ground wave radar echo, and calculate the gradient of the formation target range Doppler image to construct the position feature vector and appearance feature vector;

[0069] Node construction module: construct nodes using position feature vectors and appearance feature vectors;

[0070] Adjacent node calculation module: predicts the position feature vector of the next moment based on the position feature vector, filters the adjacent nodes through the position feature vector, and calculates the likelihood function;

[0071] Adjacent node appearance feature comparison module: construct a training data set through the appearance feature vector, train the prototype network through the training data set, calculate the prototype vector, and finally calculate the Euclidean distance;

[0072] Connection edge construction module: calculates the association similarity between known nodes and adjacent nodes and constructs connection edges;

[0073] Target matching and tracking module: Construct the graph neural network loss function, train the graph neural network, and estimate the position feature vector at the next moment.

[0074] Compared with the prior art, this solution has the following beneficial effects:

[0075] 1. This scheme proposes a high-frequency ground wave radar formation target tracking method and system, which is conducive to utilizing the multi-dimensional characteristics of the formation target to improve the tracking accuracy and continuity. This scheme constructs nodes and edges based on the characteristics and similarity of the formation target, which is conducive to using the training method of the graph neural network to improve the tracking accuracy.

[0076] 2. This scheme constructs a position feature vector based on the position, speed and size of the formation target, which is conducive to utilizing the multi-dimensional motion characteristics of the target and improving the trajectory filtering performance; this scheme constructs an appearance feature vector based on the Sobel operator, which is conducive to fully exploring the characteristics of the range Doppler image and utilizing the appearance characteristics of the formation target.

[0077] 3. This scheme selects adjacent nodes based on the predicted value of the position feature vector at the next moment, reducing the workload of subsequent calculations; the likelihood function provides a measure of the degree of matching between known nodes and adjacent stages; the prototype network model is used to reduce the demand for formation target samples in the training process; the clustering algorithm can improve the accuracy of the prototype vector.

[0078] 4. The present invention sets association similarity based on likelihood function and Euclidean distance, and utilizes multidimensional features to improve association accuracy; constructing a graph neural network loss function is conducive to comprehensive consideration of nodes and edges for training; updating the determined adjacent node position feature vectors is conducive to updating the position of the formation target and improving subsequent tracking accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is a flow chart of the high-frequency ground wave radar formation target tracking method in Example 1;

[0080] Figure 2 is a geometric model diagram of four formation targets obtained by the method of Example 1;

[0081] Figure 3 These are the range Doppler images of four types of formation targets obtained using the method of Example 1. DETAILED DESCRIPTION

[0082] The present invention is further described in detail below through specific embodiments:

[0083] Example 1

[0084] like Figure 1As shown, the high-frequency ground wave radar formation target tracking method includes the following steps:

[0085] S1. Constructing the position feature vector and appearance feature vector of the formation target based on the high-frequency ground wave radar echo, the method is as follows:

[0086] S101: Perform a two-dimensional Fourier transform on the time domain echo received by the high-frequency ground wave radar to obtain a range Doppler image.

[0087] S102: using two-dimensional constant false alarm detection on the range Doppler image to obtain the distance r of the formation target relative to the radar and the radial velocity vr of the formation target;

[0088] S103: Estimating the azimuth of the time domain echo received by the high-frequency ground wave radar to obtain the azimuth θ of the formation target relative to the radar;

[0089] S104: Calculate the position and speed of the formation target in the global coordinate system. The calculation method is as follows:

[0090] x=r·cosθ

[0091] y=r·sinθ

[0092] v x = vr·cosθ

[0093] v y = vr·sinθ

[0094] Among them, x, y represent the position of the formation target in the global coordinate system, v x ,v y Indicates the velocity of the formation target in the global coordinate system;

[0095] Construct the position feature vector of the formation target:

[0096] Pos=(x,y,v x ,v y ,rlen,dlen)

[0097] Among them, Pos represents the position feature vector of the formation target, rlen represents the number of cells occupied by the formation target in the range direction in the range Doppler spectrum, and dlen represents the number of cells occupied by the formation target in the Doppler direction in the range Doppler spectrum.

[0098] S105: Using the Sobel operator, the horizontal edge and the vertical edge of the range Doppler image obtained in step S101 are detected. The method is as follows:

[0099]

[0100] Where I represents the range Doppler image, Gx represents the gradient of image I in the horizontal direction, and Gy represents the gradient of image I in the vertical direction;

[0101] Represents a horizontal template for detecting horizontal edges of an image. Represents a longitudinal template for detecting vertical edges of an image; It means that the horizontal template is convolved with the range Doppler image obtained in step S101. It means that the longitudinal template is plane-convolved with the range Doppler image obtained in step S101;

[0102] Compute the gradient of the range Doppler image:

[0103] GA=Gx+Gy

[0104] Wherein, GA represents the gradient of the range Doppler image.

[0105] S106: Constructing the appearance feature vector of the formation target, the construction method is as follows:

[0106] App=(Gx,Gy,GA,Nf,Rgb,Hsv,Gabor)

[0107] Among them, App represents the appearance feature vector of the formation target, Nf represents the normalized first-order moment feature of the range Doppler image, Rgb represents the grayscale feature of the range Doppler image, Hsv represents the hue, saturation, and value features of the range Doppler image, and Gabor represents the texture feature of the range Doppler image.

[0108] S2. Use the position feature vector and appearance feature vector of the formation target to construct a node. The node construction method is as follows:

[0109] Node=(Pos,App,k,formation)

[0110] Among them, k represents the time and formation represents the formation type.

[0111] S3. Filter the adjacent nodes of the known node Node based on the position feature vector and calculate the likelihood function; the method is as follows:

[0112] S301: Based on the known position feature vector of the node Node, predict the position feature vector of the next moment:

[0113] Pos_1=F k Pos

[0114] Among them, Pos_1 represents the predicted value of the position feature vector at the next moment, and F krepresents the state transfer matrix, which is related to the motion state of the formation;

[0115] S302: The nodes satisfying abs(Pos_adj-Pos_1)<Dis are recorded as the adjacent nodes Adjacent of the known node Node. There may be multiple adjacent nodes Adjacent, where Pos_adj represents the position feature vector of the adjacent node Adjacent, Dis represents the threshold value, and abs(Pos_adj-Pos_1) represents the distance between Pos_adj and Pos_1.

[0116] S303: Calculate the position vector likelihood function of the known node Node and the adjacent node Adjacent:

[0117]

[0118] Among them, Lsiran represents the likelihood function, H represents the transformation matrix, and H T represents the transposed matrix of matrix H, P represents the covariance matrix, R represents the observation noise covariance, and σ represents the standard deviation of the measurement distribution of the high-frequency ground wave radar at the next moment.

[0119] S4. Based on the appearance feature vector of the formation target, a prototype network model is constructed to calculate the Euclidean distance between the appearance feature vectors of the known node Node and the adjacent node Adjacent. The method is as follows:

[0120] S401: Based on the range Doppler images of the single longitudinal formation, the V-formation and the symmetrical formation, the appearance feature vector is extracted to construct a training data set;

[0121] Train the prototype network with Resnet as the backbone network to obtain the prototype network model;

[0122] The prototype network training steps are as follows:

[0123] Based on the range Doppler images of single vertical formation, V-formation and symmetrical formation, the appearance feature vectors are extracted and the training data set is constructed.

[0124] Divide the training dataset into a support set and a query set;

[0125] In the support set, the Resnet network is used to learn features, and in the query set, the recognition performance of the Resnet network is verified;

[0126] By continuously adjusting the learning parameters, a trained prototype network model is obtained.

[0127] S402: using a clustering algorithm to classify the appearance feature vectors of the known node Node, calculating the average value of each category, and recording it as a prototype vector;

[0128] Calculate the Euclidean distance Eud_dis between the appearance feature vector of the adjacent node Adjacent and the prototype vector.

[0129] S5. Based on the likelihood function Lsiran obtained in step S3 and the Euclidean distance Eud_dis obtained in step S4, the association similarity between the known node Node and the adjacent node Adjacent is calculated to construct a connection edge; the method is as follows:

[0130] S501: Based on the likelihood function Lsiran obtained in step S3 and the Euclidean distance Eud_dis obtained in step S4, the association similarity between the known node Node and the adjacent node Adjacent is calculated. The calculation formula is as follows:

[0131] Simi=Qz1·Lsiran+Qz2·Eud_dis

[0132] Among them, Simi represents the association similarity between the known node Node and the adjacent node Adjacent, Qz1 represents the weight of the likelihood function Lsiran, and Qz2 represents the weight of the Euclidean distance Eud_dis;

[0133] Based on the association similarity, an edge is established between a known node Node and an adjacent node Adjacent.

[0134] S6. Training based on the graph neural network loss function to achieve formation target matching and tracking is as follows:

[0135] S601: Construct graph neural network loss function:

[0136] Lsunsi=Lnode+Qz3·Line

[0137] Among them, Lsunsi represents the graph neural network loss function, Lnode represents the node loss function, Line represents the edge loss function, and Qz3 represents the weight of the edge loss function;

[0138] The graph neural network is trained to obtain the edge with the highest similarity score and the corresponding adjacent node Adjacent through the known node Node; the graph neural network training steps are as follows:

[0139] A batch training algorithm is used, with a known node Node as the root. Under a fixed sample capacity, the neighborhood of the root node is recursively expanded to sample the tree.

[0140] S602: Update the position feature vector of the adjacent nodes:

[0141] Pos_2=Pos_1+Kalz(Pos_adj-H·Pos_1)

[0142] Among them, Pos_2 represents the updated value of the position feature vector of the adjacent node, and Kalz represents the Kalman gain.

[0143] Figure 2 The geometric models of four formation targets are single column, oblique column, V-shaped column and symmetrical column (the dimensions marked in the figure are for reference only and the actual dimensions may vary). Figure 3 for Figure 2 The range Doppler image corresponding to the target in the formation, taking a single column, an oblique column, a V-shaped formation, and a symmetrical formation as examples, shows the range Doppler image obtained in step S101. This scheme includes but is not limited to the four formation modes of a single column, an oblique column, a V-shaped formation, and a symmetrical formation.

[0144] Example 2

[0145] High frequency ground wave radar formation target tracking system, including:

[0146] Feature vector construction module: Calculate the position and velocity of the formation target based on the high-frequency ground wave radar echo, and calculate the gradient of the formation target range Doppler image to construct the position feature vector and appearance feature vector;

[0147] Node construction module: construct nodes using position feature vectors and appearance feature vectors;

[0148] Adjacent node calculation module: predicts the position feature vector of the next moment based on the position feature vector, filters the adjacent nodes through the position feature vector, and calculates the likelihood function;

[0149] Adjacent node appearance feature comparison module: construct a training data set through the appearance feature vector, train the prototype network through the training data set, calculate the prototype vector, and finally calculate the Euclidean distance;

[0150] Connection edge construction module: calculates the association similarity between known nodes and adjacent nodes and constructs connection edges;

[0151] Target matching and tracking module: Construct the graph neural network loss function, train the graph neural network, and estimate the position feature vector at the next moment.

[0152] The above are only embodiments of the present invention, and the common knowledge such as the known specific structures and / or characteristics in the scheme are not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A high frequency ground wave radar formation target tracking method, characterized in that: The steps include: S1. Constructing the position feature vector and appearance feature vector of the formation target based on the high-frequency ground wave radar echo; S2, constructing nodes using the position feature vector and appearance feature vector of the formation target; S3, based on the position feature vector, filter the adjacent nodes of the known node Node and calculate the likelihood function; S4. Construct a prototype network model based on the appearance feature vector of the formation target, and calculate the Euclidean distance between the appearance feature vectors of the known node Node and the adjacent node Adjacent; S5, based on the likelihood function Lsiran obtained in step S3 and the Euclidean distance Eud_dis obtained in step S4, calculate the association similarity between the known node Node and the adjacent node Adjacent, so as to construct a connection edge; S6. Perform training based on the graph neural network loss function to achieve formation target matching and tracking.

2. The high-frequency ground wave radar formation target tracking method according to claim 1, characterized in that: The method of step S1 is as follows: S101: Perform two-dimensional Fourier transform on the time domain echo received by the high-frequency ground wave radar to obtain a range Doppler image; S102: using two-dimensional constant false alarm detection on the range Doppler image to obtain the distance r of the formation target relative to the radar and the radial velocity vr of the formation target; S103: Estimating the azimuth of the time domain echo received by the high-frequency ground wave radar to obtain the azimuth θ of the formation target relative to the radar; S104: Calculate the position and speed of the formation target in the global coordinate system; S105: Using the Sobel operator, detecting the horizontal edge and the vertical edge of the range Doppler image obtained in step S101; S106: Construct an appearance feature vector of the formation target.

3. The high frequency ground wave radar formation target tracking method and system according to claim 2, characterized in that: The calculation method of position and speed in step S104 is as follows: x=r·cosθ y=r·sinθ v x =vr·cosθ v y =vr sinθ Among them, x, y represent the position of the formation target in the global coordinate system, v x ,v y Indicates the velocity of the formation target in the global coordinate system; Construct the position feature vector of the formation target: Pos=(x,y,v x ,v y ,rlen,dlen) Among them, Pos represents the position feature vector of the formation target, rlen represents the number of cells occupied by the formation target in the range direction in the range Doppler spectrum, and dlen represents the number of cells occupied by the formation target in the Doppler direction in the range Doppler spectrum.

4. The high frequency ground wave radar formation target tracking method according to claim 2, characterized in that: The method of step S105 is as follows: Where I represents the range Doppler image, Gx represents the gradient of image I in the horizontal direction, and Gy represents the gradient of image I in the vertical direction; Represents a horizontal template for detecting horizontal edges of an image. Represents a longitudinal template for detecting vertical edges of an image; It means that the horizontal template is convolved with the range Doppler image obtained in step S101. It means that the longitudinal template is plane-convolved with the range Doppler image obtained in step S101; Compute the gradient of the range Doppler image: GA=Gx+Gy Where GA represents the gradient of the range Doppler image; The construction method of step S106 is as follows: App=(Gx,Gy,GA,Nf,Rgb,Hsv,Gabor) Among them, App represents the appearance feature vector of the formation target, Nf represents the normalized first-order moment feature of the range Doppler image, Rgb represents the grayscale feature of the range Doppler image, Hsv represents the hue, saturation, and value features of the range Doppler image, and Gabor represents the texture feature of the range Doppler image.

5. The high frequency ground wave radar formation target tracking method according to claim 1, characterized in that: The node construction method of step S2 is as follows: Node=(Pos,App,k,formation) Among them, k represents the time and formation represents the formation type.

6. The high frequency ground wave radar formation target tracking method according to claim 1, characterized in that: The method of step S3 is as follows: S301: Based on the known position feature vector of the node Node, predict the position feature vector of the next moment: Pos_1=F k Pos Among them, Pos_1 represents the predicted value of the position feature vector at the next moment, and F k represents the state transfer matrix; S302: Record the nodes satisfying abs(Pos_adj-Pos_1)<Dis as the adjacent nodes Adjacent of the known node Node, where Pos_adj represents the position feature vector of the adjacent node Adjacent, Dis represents the threshold value, and abs(Pos_adj-Pos_1) represents the distance between Pos_adj and Pos_1; S303: Calculate the position vector likelihood function of the known node Node and the adjacent node Adjacent: Among them, Lsiran represents the likelihood function, H represents the transformation matrix, and H T represents the transposed matrix of matrix H, P represents the covariance matrix, R represents the observation noise covariance, and σ represents the standard deviation of the measurement distribution of the high-frequency ground wave radar at the next moment.

7. The high frequency ground wave radar formation target tracking method according to claim 1, characterized in that: The method of step S4 is as follows: S401: Based on the range Doppler images of the single longitudinal formation, the V-formation and the symmetrical formation, the appearance feature vector is extracted to construct a training data set; Train the prototype network with Resnet as the backbone network to obtain the prototype network model; S402: using a clustering algorithm to classify the appearance feature vectors of the known node Node, calculating the average value of each category, and recording it as a prototype vector; Calculate the Euclidean distance Eud_dis between the appearance feature vector of the adjacent node Adjacent and the prototype vector.

8. The high frequency ground wave radar formation target tracking method according to claim 1, characterized in that: The method of step S5 is as follows: S501: Based on the likelihood function Lsiran obtained in step S3 and the Euclidean distance Eud_dis obtained in step S4, the association similarity between the known node Node and the adjacent node Adjacent is calculated. The calculation formula is as follows: Simi=Qz1·Lsiran+Qz2·Eud_dis Among them, Simi represents the association similarity between the known node Node and the adjacent node Adjacent, Qz1 represents the weight of the likelihood function Lsiran, and Qz2 represents the weight of the Euclidean distance Eud_dis; Based on the association similarity, an edge is established between a known node Node and an adjacent node Adjacent.

9. The high frequency ground wave radar formation target tracking method according to claim 1, characterized in that: The method of step S6 is as follows: S601: Construct graph neural network loss function: Lsunsi=Lnode+Qz3·Line Among them, Lsunsi represents the graph neural network loss function, Lnode represents the node loss function, Line represents the edge loss function, and Qz3 represents the weight of the edge loss function; Train the graph neural network and obtain the edges with the highest similarity scores and the corresponding adjacent nodes Adjacent through the known nodes Node; S602: Update the position feature vector of the adjacent nodes: Pos_2=Pos_1+Kalz(Pos_adj-H·Pos_1) Among them, Pos_2 represents the updated value of the position feature vector of the adjacent node, and Kalz represents the Kalman gain.

10. High frequency ground wave radar formation target tracking system, characterized by: include: Feature vector construction module: Calculate the position and velocity of the formation target based on the high-frequency ground wave radar echo, and calculate the gradient of the formation target range Doppler image to construct the position feature vector and appearance feature vector; Node construction module: construct nodes using position feature vectors and appearance feature vectors; Adjacent node calculation module: predicts the position feature vector of the next moment based on the position feature vector, filters the adjacent nodes through the position feature vector, and calculates the likelihood function; Adjacent node appearance feature comparison module: construct a training data set through the appearance feature vector, train the prototype network through the training data set, calculate the prototype vector, and finally calculate the Euclidean distance; Connection edge construction module: calculates the association similarity between known nodes and adjacent nodes and constructs connection edges; Target matching and tracking module: Construct the graph neural network loss function, train the graph neural network, and estimate the position feature vector at the next moment.

Images