Target state transition matrix detection method under partial information loss

Abstract

The present invention aims to provide a target state transfer matrix detection method under partial information loss, comprising the following steps: constructing a target state transfer matrix detection network, the target state transfer matrix detection network comprising a plurality of sequentially connected feature extraction modules; training the target state transfer matrix detection network to obtain trained target state transfer matrix detection network parameters; obtaining a target state observation sequence of length TS, with a number of states SN, as an input matrix, adding a channel dimension to the matrix to make it 1×TS×SN, inputting the matrix into the trained target state transfer matrix detection network, and sequentially processing it through the plurality of feature extraction modules to obtain a final target transfer matrix. The present invention can accurately detect the target's actual motion state transfer matrix, thereby improving the accuracy of filtered predicted track.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target tracking, and in particular to a target state transfer matrix detection method under partial information loss. Background Art

[0002] The fundamental principle of radar target tracking is based on an iterative computational framework using a Bayesian tracking filter. Given a known target motion model, the target state sequence—in other words, the target track—is calculated from the observation sequence. Mainstream radar target tracking algorithms such as the Kalman filter, extended Kalman filter, unscented Kalman filter, and particle filter all require a known target motion model. If the motion model is inaccurate, the track predicted by the filter will deviate significantly from the actual track.

[0003] However, in practical tracking, especially for maneuvering targets, the target's motion is complex and changeable, making it difficult to accurately predict the target's motion model in advance, resulting in a significant decline in tracking performance. To address this, a large number of model prediction methods have been proposed, such as fixed-structure multi-model fitting algorithms, variable-structure multi-model fitting algorithms, hybrid grid multi-model fitting algorithms, and multi-model fitting algorithms based on unbiased estimation. The basic idea behind these algorithms is to estimate the target's motion by setting a certain number of known models and combining them using a weighted approach. The weights are continuously updated during the iterative calculation of target tracking, ultimately achieving a good model prediction result. However, in practice, the iterative update of the weights is slow and often cannot keep up with the changes in the target's motion. Therefore, Liu et al. proposed a cross-product dot-product network to detect the target's state transition matrix in real time, thereby achieving model prediction. This network uses existing target state information to calculate the target's motion model, demonstrating good performance in motion model detection. However, in practical applications, the motion state is unknown and can only be inferred from observation sequences. However, observation sequences often only contain partial motion state information, significantly limiting the practicality of this algorithm. Summary of the Invention

[0004] The present invention aims to provide a target state transfer matrix detection method under partial information loss. The method provides an accurate target transfer matrix for radar tracking calculation of moving targets, especially fast maneuvering targets, by constructing a target state transfer matrix detection network, obtains target motion model information, and improves the accuracy of the track obtained by filtering prediction.

[0005] The technical solutions of the present invention are as follows:

[0006] The target state transfer matrix detection method under partial information loss includes the following steps:

[0007] Constructing a target state transfer matrix detection network, wherein the target state transfer matrix detection network includes a plurality of sequentially connected feature extraction modules; training the target state transfer matrix detection network to obtain trained target state transfer matrix detection network parameters;

[0008] Get the target state observation sequence of length TS, whose number of states is SN as the input matrix, add a channel dimension to the matrix to make it 1×TS×SN matrix, input the matrix into the trained target state transfer matrix detection network, and process it through multiple feature extraction modules in sequence to obtain the final target transfer matrix; where TS represents the sequence length; SN represents the number of target states;

[0009] In the feature extraction module, the input matrix is ​​processed by a convolutional neural network to obtain a convolution result. After the channel dimension and state dimension are transposed, the convolution result is calculated point by point using the Gaussian activation function and the hyperbolic tangent activation function respectively. The calculation results of these two functions are multiplied to obtain the output result.

[0010] Preferably, the value of TS is 27; the value of SN is 2.

[0011] The number of convolution kernels of the convolutional neural network is a variable K, which is generated by random initialization.

[0012] The processing process of the convolutional neural network is as follows:

[0013] The variable I dimension of the convolutional neural network input is a 1×TS×SN matrix. Z different convolution kernels are generated by random initialization. The dimension of each convolution kernel is 1×SN×SN, the step size is set to 1, the patch value is set to 0, and the z∈[1:Z]th convolution kernel C z Perform dot multiplication with the input variable to obtain the convolution result r. The specific calculation formula is as follows:

[0014]

[0015] Where n = 1, 2..., TS-SN+1, the results of Z convolution kernels are combined into the final output convolution result Its dimension is 1×Z×(TS-SN+1).

[0016] The formula of the Gaussian activation function is as follows:

[0017]

[0018] Among them, r represents the convolution result of the convolutional neural network; x = r 1，i，j , i = 1, 2, ..., TS-SN+1, j = 1, 2, ..., Z; e represents a natural constant; μ and σ represent learnable parameters.

[0019] The formula of the hyperbolic tangent activation function is as follows:

[0020]

[0021] Among them, r represents the convolution result of the convolutional neural network; x = r 1，i，j , i=1,2,…,TS-SN+1,j=1,2,…,Z;e represents a natural constant.

[0022] The 1×TS×SN matrix is ​​processed by four feature extraction modules in sequence, and the number of convolution kernels of the four feature extraction modules is 4, 8, 16, and 8 respectively.

[0023] The training process of the target state transfer matrix detection network is as follows:

[0024] A. Generation of training data:

[0025] The training data is generated by simulation of a noise-free state-space model, and its specific form is as follows:

[0026] x k =f(x k-1 ,M) (4)

[0027] z k =h(x k ) (5)

[0028] Formula (4) is the state transfer function, and formula (5) is the observation function;

[0029] Among them, x k is the target state information at time k, including the coordinate point and speed of the target in the two-dimensional plane, that is, x k =[p x，k , p y，k , v x，k , v y，k ] T , where p k =[p x，k , p y，k ] T is the (x, y) value of the current coordinate point of the target, v k =[v x，k , v y，k ] T is the target's current speed in the x and y directions; z k is the observation information at time k; M is the target motion model;

[0030] B. Preprocessing of training data:

[0031] a) Calculate the first-order difference between the input data and the initial value. The formula is as follows:

[0032] [Δp0, Δp1, …, Δp K-2 ]=[p1-p0, p2-p1,…, p k-1 -p K-2 ] (6)

[0033] b) For each moment k, calculate the sine and cosine values ​​of the difference vector at that moment as follows:

[0034]

[0035]

[0036] The results of formula (7) and formula (8) are combined to form the training input data of the target state transfer matrix detection network: [sc0, sc1, ..., sc K-3 ], where sc k =[sin k , cos k ] T ;

[0037] C. Model training

[0038] [sc0, sc1, ..., sc K-3 ], input into the target state transfer matrix detection network, and calculate the predicted state transfer matrix The loss function is then calculated as follows:

[0039]

[0040] Among them, F represents the actual state transfer matrix;

[0041] The target state transition matrix is ​​then trained through back-propagation to detect the parameters of the network.

[0042] The method of the present invention constructs a target state transfer matrix detection network, which can correctly memorize different maneuvering target state transfer matrices and, when the cross product and dot product information of the observation sequence segment is input, reasonably gives the target state transfer matrix corresponding to the current observation sequence segment based on the existing memory, thereby achieving correct prediction of the target motion model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the structure of the target state transition matrix detection network according to an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of the structure of a feature extraction module according to an embodiment of the present invention;

[0045] Figure 3 This is a simulation diagram of the motion trajectory of a maneuvering target according to Example 2 of the present invention;

[0046] Figure 4 This is a simulation diagram of target azimuth and distance (noise included) measured by the radar according to Example 2 of the present invention;

[0047] Figure 5 This is a simulation diagram of the target azimuth and distance (noise included) measured by the radar in Example 2 of the present invention at 10-13s;

[0048] Figure 6 This is a 10-13s noisy coordinate point diagram of Example 2 of the present invention. DETAILED DESCRIPTION

[0049] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0050] Example 1

[0051] The target state transition matrix detection method under partial information loss includes the following steps:

[0052] (1) constructing a target state transfer matrix detection network, wherein the target state transfer matrix detection network includes four feature extraction modules connected in sequence; the number of convolution kernels of the four feature extraction modules is 4, 8, 16, and 8, respectively; training the target state transfer matrix detection network to obtain a trained target state transfer matrix detection network;

[0053] The target state transfer matrix detection network training process is as follows:

[0054] A. Generation of training data:

[0055] The training data is generated by simulation of a noise-free state-space model, and its specific form is as follows:

[0056] x k =f(x k-1 ,M) (4)

[0057] z k =h(x k ) (5)

[0058] Formula (4) is the state transfer function, and formula (5) is the observation function;

[0059] Among them, x k is the target state information at time k, including the coordinate point and speed of the target in the two-dimensional plane, that is, x k =[p x，k , p y，k , v x，k , v y，k ]T , where p k =[p x，k , p y，k ] T is the (x, y) value of the current coordinate point of the target, v k =[v x，k , v y，k ] T is the target's current speed in the x and y directions; z k is the observation information at time k; M is the target motion model;

[0060] B. Preprocessing of training data:

[0061] a) Calculate the first-order difference between the input data and the initial value. The formula is as follows:

[0062] [Δp0, Δp1, ..., Δp K-2 ]=[p1-p0, p2-p1, ..., p k-1 -p K-2 ] (6)

[0063] b) For each moment k, calculate the sine and cosine values ​​of the difference vector at that moment as follows:

[0064]

[0065]

[0066] The results of formula (7) and formula (8) are combined to form the training input data of the target state transfer matrix detection network: [sc0, sc1, ..., sc K-3 ], where sc k =[sin k , cos k ] T ;

[0067] C. Model training

[0068] [sc0, sc1, ..., sc K-3 ], input into the target state transfer matrix detection network, and calculate the predicted state transfer matrix The loss function is then calculated as follows:

[0069]

[0070] The target state transition matrix is ​​then trained through back-propagation to detect the parameters of the network.

[0071] (2) Obtain a target state observation sequence of length TS, with the number of states SN as the input matrix, add a channel dimension to the matrix to make it a 1×TS×SN matrix, with the value of TS being 27 and the value of SN being 2; input the matrix into the trained target state transition matrix detection network, and process it through the four feature extraction modules in sequence to obtain the final target transfer matrix;

[0072] In the feature extraction module, the input matrix is ​​processed by a convolutional neural network to obtain a convolution result. After the channel dimension and state dimension are transposed, the convolution result is calculated point by point using the Gaussian activation function and the hyperbolic tangent activation function respectively. The calculation results of these two functions are multiplied to obtain the output result.

[0073] The processing process of the convolutional neural network is as follows:

[0074] The variable I dimension of the convolutional neural network input is a 1×TS×SN matrix. Z different convolution kernels are generated by random initialization. The dimension of each convolution kernel is 1×SN×SN, the step size is set to 1, the patch value is set to 0, and the z∈[1:Z]th convolution kernel C z Perform dot multiplication with the input variable to obtain the convolution result r. The specific calculation formula is as follows:

[0075]

[0076] Where n = 1, 2..., TS-SN+1, the results of Z convolution kernels are combined into the final output convolution result Its dimension is 1×Z×(TS-SN+1);

[0077] The formula of the Gaussian activation function is as follows:

[0078]

[0079] Among them, r represents the convolution result of the convolutional neural network; x = r 1，i，j ,i=1,2,…,TS-SN+1,j=1,2,…,Z;μ and σ Represents learnable parameters.

[0080] The formula of the hyperbolic tangent activation function is as follows:

[0081]

[0082] Among them, r represents the convolution result of the convolutional neural network; x = r 1，i，j ,i=1,2,…,TS-SN+1,j=1,2,…,Z.

[0083] Example 2

[0084] This embodiment detects the uniform turning maneuver motion model. The specific implementation process is as follows:

[0085] (1) constructing a target state transfer matrix detection network, wherein the target state transfer matrix detection network includes four feature extraction modules connected in sequence; the number of convolution kernels of the four feature extraction modules is 4, 8, 16, and 8, respectively; training the target state transfer matrix detection network to obtain a trained target state transfer matrix detection network;

[0086] A. Generation of training data:

[0087] The training data is generated by simulation of a noise-free state-space model, and its specific form is as follows:

[0088] x k =f(x k-1 ,M) (4)

[0089] z k =h(x k ) (5)

[0090] Formula (4) is the state transfer function, and formula (5) is the observation function;

[0091] Among them, x k is the target state information at time k, including the coordinate point and speed of the target in the two-dimensional plane, that is, x k =[p x，k , p y，k , v x，k , v y，k ] T , where p k =[p x，k , p y，k ] T is the (x, y) value of the current coordinate point of the target, v k =[v x，k , v y，k ] T is the target's current speed in the x and y directions; z k is the observation information at time k; M is the target motion model;

[0092] Whatever form of motion model you plan to detect, generate training data according to its corresponding target state transfer matrix. Here, the uniform speed turning state transfer matrix is ​​used to generate training data.

[0093] When the target turns at a constant speed, the target's motion model is the following state transfer matrix:

[0094]

[0095] In the above formula, α is the turning rate, s τ is the sampling time interval, and Equation (4) degenerates into the following linear form

[0096] x k =F·x k-1 (12)

[0097] For actual observation functions, radar generally observes the target's range d and azimuth θ. These are only related to the coordinate point information of the target state. The specific relationship is as follows:

[0098]

[0099] As long as we give an initial target state x0, a given turning rate α, and the number of target movement time points K, we can calculate the target's true trajectory: [x0, x1, ..., x K-1 Correspondingly, the target observation sequence [z0, z1, ..., z K-1 ]. Because according to formula (13), the sequence lacks speed information, but corresponds one-to-one with the actual coordinate points of the target, then the actual coordinate point sequence [p0, p1, ..., p K-1 ] as the final observation result. The above results and the original target state transfer matrix (11) together constitute the input and output data pairs for network model parameter training. In fact, considering the actual structure of the target state transfer matrix detection network, the number of target motion time points K should be set to 29.

[0100] In order to generate a large amount of data and cover the calculation range of target tracking to a certain extent, we randomly obtain p from the interval [1000, 10000] for each input and output data pair. k , randomly obtain v from the interval [50, 350] k , α is randomly obtained from the interval [-90, 90] × π / 180. This way, each training session is conducted on different input and output data pairs, improving the robustness of the model.

[0101] B. Preprocessing of training data:

[0102] a) Calculate the first-order difference between the input data and the initial value. The formula is as follows:

[0103] [Δp0, Δp1, ..., Δp K-2 ]=[p1-p0, p2-p1, ..., p k-1 -p K-2 ] (6)

[0104] b) For each moment k, calculate the sine and cosine values ​​of the difference vector at that moment as follows:

[0105]

[0106]

[0107] The results of formula (7) and formula (8) are combined to form the real training input data of the target state transfer matrix detection network: [sc0, sc1, ..., sc K-3 ], where sc k =[sin k , cos k ] T ;

[0108] C. Model training

[0109] [sc0, sc1, ..., sc K-3 ], input into the target state transfer matrix detection network, and calculate the predicted state transfer matrix The loss function is then calculated as follows:

[0110]

[0111] The target state transition matrix is ​​then trained through back-propagation to detect the parameters of the network.

[0112] D. Final results of model training

[0113] After 1 million iterations of training, the error of the output model converged to 10 -10 , the model error is very small and can be used to predict the target state transfer matrix.

[0114] (2) Use of the model

[0115] In the actual observation sequence, we intercept a sequence of length 29 and use the following formulas (13) and (14) to restore the coordinate trajectory of the target movement. Then, we use the preprocessing method of 1.2 training data to calculate the cross product dot product information, and input it into the target state transfer matrix detection network to detect the target state transfer matrix.

[0116] p x，k =d k cos(θ k ) (13)

[0117] p y，k =d k sin(θ k ) (14)

[0118] (3) Specific testing cases

[0119] Figure 3 It is a simulated target motion track, which consists of three parts. The turning rate of the first part is -3° / s, the turning rate of the second part is 8° / s, and the turning rate of the third part is 0° / s.

[0120] The noisy observation data of this target can be found in Figure 4 Simulation diagram of target azimuth and distance (noise included) measured by radar;

[0121] Assume that the target motion state needs to be detected for 10-13 seconds, and the target azimuth angle sequence and distance sequence with a length of 29 in this time period are intercepted as follows: Figure 5 Simulation diagram of target azimuth and distance (noise included) measured by radar at 10-13s;

[0122] The intercepted observation sequence is used to reversely calculate the noisy target coordinate point sequence using formula (13) (14) as follows: Figure 6 As shown;

[0123] The noisy coordinate point sequence is preprocessed according to step B above, and then input into the target state transfer matrix detection network to calculate the target transfer matrix as shown below:

[0124]

Claims

1. A radar target state prediction method based on the state transfer matrix under partial information loss, characterized in that The following steps are involved: Constructing a target state transfer matrix detection network, wherein the target state transfer matrix detection network includes a plurality of sequentially connected feature extraction modules; Training the target state transfer matrix detection network to obtain trained target state transfer matrix detection network parameters; Get the target state observation sequence of length TS, whose number of states is SN as the input matrix, add a channel dimension to the matrix to make it 1×TS×SN matrix, input the matrix into the trained target state transfer matrix detection network, and process it through multiple feature extraction modules in sequence to obtain the final target transfer matrix; where TS represents the sequence length; SN represents the number of target states; In the feature extraction module, the input matrix is ​​processed by a convolutional neural network to obtain a convolution result. After the channel dimension and state dimension are transposed, the convolution result is calculated point by point using the Gaussian activation function and the hyperbolic tangent activation function respectively. The calculation results of these two functions are multiplied to obtain the output result. The cross product dot product information of the observation sequence fragment is input into the target state transfer matrix detection network, and the target state transfer matrix corresponding to the current observation sequence fragment is given, thereby achieving correct prediction of the target motion model.

2. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 1, characterized in that: The TS value is 27; the SN value is 2.

3. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 1, characterized in that: The number of convolution kernels of the convolutional neural network is a variable K, which is generated by random initialization.

4. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 3, characterized in that: The processing process of the convolutional neural network is as follows: Convolutional Neural Network Input Variables The dimension is 1×TS×SN matrix, and Z different convolution kernels are generated by random initialization. The dimension of each convolution kernel is 1×SN×SN, the step size is set to 1, the patch value is set to 0, and the zth [1:Z] convolution kernels Perform dot multiplication with the input variable to obtain the convolution result r. The specific calculation formula is as follows: (1) Where n=1,2…,TS-SN+1, the results of Z convolution kernels are combined into the final output convolution result , whose dimension is 1×Z×(TS-SN+1).

5. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 1, characterized in that: The formula of the Gaussian activation function is as follows: (2) Among them, r represents the convolution result of the convolutional neural network; , =1,2,…,TS-SN+1, =1,2,…,Z; e represents a natural constant; and Represents learnable parameters.

6. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 1, characterized in that: The formula of the hyperbolic tangent activation function is as follows: (3) Among them, r represents the convolution result of the convolutional neural network; , =1,2,…,TS-SN+1, =1,2,…,Z; e represents a natural constant.

7. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 1, characterized in that: The 1×TS×SN matrix is ​​processed by four feature extraction modules in sequence, and the number of convolution kernels of the four feature extraction modules is 4, 8, 16, and 8 respectively.

8. The radar target state prediction method based on the state transfer matrix under partial information loss as claimed in claim 1, characterized in that: The training process of the target state transfer matrix detection network is as follows: A. Generation of training data: The training data is generated by simulation of a noise-free state-space model, and its specific form is as follows: (4) (5) Formula (4) is the state transfer function, and formula (5) is the observation function; in, yes The target state information at the moment includes the target's coordinate point and speed in the two-dimensional plane, that is, ,in is the (x,y) value of the current coordinate point of the target, It is the value of the target's current movement speed in the x and y directions; yes Observation information at the moment; is the target motion model; B. Preprocessing of training data: a) Calculate the first-order difference between the input data and the initial value. The formula is as follows: (6) b) For every moment , calculate the sine and cosine values ​​of the difference vector at this moment as follows: (7) (8) The results of formula (7) and formula (8) are combined to form the training input data of the target state transfer matrix detection network: ,in ; C. Model training Will , input into the target state transfer matrix detection network, and calculate the predicted state transfer matrix , and then calculate the loss function as follows: (9) Among them, F represents the actual state transfer matrix; The target state transition matrix is ​​then trained through back-propagation to detect the parameters of the network.

Images