A deep neural network-based intelligent tracking method for maneuvering group targets

Abstract

This invention discloses an intelligent tracking method for maneuvering swarm targets based on deep neural networks. This method incorporates deep learning into a Bayesian filtering framework, extracting the motion characteristics and process noise statistical characteristics of the swarm targets from measurement data using multiple deep neural networks. It then estimates the motion state transition matrix and process noise variance matrix of the swarm targets online, and uses Bayesian filtering to accurately estimate the motion state of the swarm targets' centroid. By modeling the swarm target contour as an ellipse centered at the centroid, the high-precision centroid motion state estimation results improve the contour estimation accuracy, ultimately achieving swarm target tracking. Compared to existing swarm target tracking methods, the proposed method does not require prior information to establish a target motion model, enabling high-precision tracking of non-cooperative maneuvering swarm targets in complex environments where prior information is lacking.

Description

Technical Field

[0001] This invention relates to the field of target tracking, and more specifically to a method for intelligent target tracking of a maneuvering swarm based on deep neural networks. Background Technology

[0002] In recent years, with the rapid development of computer, communication, and navigation technologies, cluster weapon systems such as swarm drones and multiple-launch missile salvos have been gradually deployed. On the one hand, these cluster weapon systems contain a large number of combat units, quickly saturating the detection, tracking, and interception capabilities of air defense systems, making effective interception difficult. On the other hand, the combat units within these systems possess a high degree of coordination; if a few units are damaged, the remaining units can quickly fill the gap, without significantly reducing the overall combat capability of the cluster, making effective destruction difficult. Therefore, defending against and intercepting these cluster weapon systems is extremely challenging and has become a major challenge that modern air defense systems must address.

[0003] Accurate tracking of the aforementioned group of targets, obtaining the motion state and the time-varying trajectory and trend of their outline morphology, is a necessary prerequisite for subsequent weapon systems to formulate defense strategies, solve attack elements, and implement effective interception. Currently, for group targets with a large number of densely distributed targets, to improve the real-time performance of the tracking system and avoid tracking capability saturation, the main approach is to use a group-wide tracking method, which involves sequentially estimating the motion state of the group's centroid and the shape of the group's outline. Regarding the estimation of the group centroid's motion state, existing methods and technologies require the pre-establishment of a motion model of the tracked target, belonging to model-driven target tracking methods. However, when tracking non-cooperative targets in complex battlefield environments, factors such as a lack of prior information, variable target motion modes, unknown or even time-varying statistical characteristics of process noise can all lead to mismatch in the pre-established motion model, resulting in a significant decrease in target tracking accuracy and severe degradation of tracking performance. Therefore, existing model-driven group target tracking methods and technologies cannot accurately estimate the state of the aforementioned non-cooperative maneuvering group targets in complex environments. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent tracking of maneuvering swarm targets based on deep neural networks, which solves the problem of high-precision tracking of maneuvering swarm targets in complex environments such as multiple motion modes of swarm targets, lack of prior information about the targets, and time-varying statistical characteristics of process noise.

[0005] The technical solution for achieving the objective of this invention is as follows: Firstly, this invention provides a method for intelligent tracking of maneuvering swarm targets based on deep neural networks, comprising the following steps:

[0006] The first step is to obtain the group target measurement information at the current moment;

[0007] The second step is to estimate the motion state of the group's centroid at the current moment;

[0008] The third step is to estimate the outline shape of the group of targets at the current moment;

[0009] The fourth step is to predict the motion state of the group's center of mass at the next moment;

[0010] The fifth step is to predict the outline shape of the group of targets at the next moment.

[0011] In a second aspect, the present invention provides a maneuvering swarm target intelligent tracking system based on a deep neural network, used to implement the method described in the first aspect, the system comprising:

[0012] The first module is used to obtain the group target measurement information at the current moment;

[0013] The second module is used to estimate the motion state of the group's centroid at the current moment;

[0014] The third module is used to estimate the contour shape of the group of targets at the current moment;

[0015] The fourth module is used to predict the motion state of the group's centroid at the next moment;

[0016] The fifth module is used to predict the outline shape of the group of targets at the next moment.

[0017] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in the first aspect.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0019] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0020] Compared with existing group target tracking methods, the significant advantages of this invention are:

[0021] (1) In the case of multiple motion modes of the centroid of a group of targets and unknown prior information of the targets, this invention uses a deep neural network to extract the motion characteristics of the targets from the measurement data of the group of targets online, and then estimates the motion state transition matrix in real time. This solves the problem that the pre-established motion model is difficult to match the actual motion mode of the maneuvering target with a lack of prior information, and effectively improves the estimation accuracy of the motion state of the centroid of the group of targets.

[0022] (2) In complex environments where the statistical characteristics of process noise are unknown or even time-varying, this invention uses deep neural networks to extract the statistical characteristics of process noise from the measurement data of group targets, and then estimates the variance matrix of process noise in real time. This solves the problem that the statistical characteristics of process noise are difficult to obtain accurately in complex environments, and further improves the estimation accuracy of the motion state of the centroid of group targets.

[0023] (3) The present invention simplifies the modeling of the group target contour into a time-varying ellipse with the center located at the centroid of the group target. On the one hand, it can improve the real-time performance of the group target tracking algorithm, and on the other hand, it can improve the estimation accuracy of the group target contour shape by utilizing the high-precision estimation results of the motion state of the centroid of the group target. Attached Figure Description

[0024] Figure 1 This is a flowchart of a group target tracking method based on deep neural networks.

[0025] Figure 2 A schematic diagram showing the outline shape, centroid, and measurement data of the group of targets.

[0026] Figure 3 This is a diagram showing the overall structure of the motion state transition matrix estimation model.

[0027] Figure 4 This is a diagram showing the overall structure of the process noise variance matrix estimation model.

[0028] Figure 5 This is a schematic diagram of the motion trajectory and measurement data of a group of targets in a typical scenario.

[0029] Figure 6 This is a schematic diagram of the results of group target tracking in a typical scenario. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0031] This invention proposes an intelligent target tracking method for maneuvering groups based on deep neural networks. It can solve the target tracking problem of non-cooperative maneuvering groups under complex environments such as variable target motion modes, unknown target prior information, and unknown process noise variance. The estimation accuracy of the target state of non-cooperative maneuvering groups is significantly better than existing group target tracking methods that require prior information to establish a target motion model in advance.

[0032] The proposed intelligent group target tracking method based on deep neural networks utilizes deep neural networks to extract target motion characteristics and process noise statistical characteristics in real time from group target measurement data. It then estimates the motion state transition matrix and process noise variance matrix online, and subsequently estimates the centroid motion state and contour morphology of the group targets through Bayesian filtering, ultimately completing the group target tracking. The implementation process of the proposed intelligent group target tracking method based on deep neural networks is as follows: Figure 1 As shown, the specific implementation method is as follows:

[0033] Step 1: Obtain the current group target measurement information

[0034] (1.1) As Figure 2 As shown, for a group of targets with densely distributed units within the group, the detection device can acquire multiple measurement data generated by the units within the group at each sampling time, thus forming a measurement dataset. Where N k N represents the number of measurement data acquired at time k. Due to factors such as mutual occlusion among cells within the group, N... k Time-varying, meaning the number of measurement sources varies over time. Therefore, the measurement data z in the measurement dataset... k,l It can be considered as being generated by the l-th measurement source appearing at time k, and the measurement equation is as follows:

[0035] z k,l =h(x k,l )+v k,l (1) In the formula, h(·) represents the measurement function; x k,l v represents the state vector of the l-th measurement source; k,l For measuring noise.

[0036] (1.2) Based on the above measurement dataset, calculate the target centroid measurement information at time k using the following formula.

[0037]

[0038] Step 2: Estimate the motion state of the group's center of mass at the current moment.

[0039] The capacitive Kalman filter method is used to estimate the motion state of the centroid of the group of targets. The specific steps are as follows:

[0040] (2.1) Calculate the volume measurement point

[0041]

[0042] In the formula, j = 1, 2, ..., 2n x n x The motion state vector of the centroid of the group target dimensionality; Φ is the square root factor of the prior error covariance of the centroid motion state of the group of targets; j Let j be the j-th element in set Φ. Set Φ is defined as follows:

[0043]

[0044] (2.2) Calculate the prior measurement of the centroid of the group target

[0045]

[0046] (2.3) Calculate the square root factor of the prior information covariance.

[0047]

[0048] In the formula, Chol(·) denotes Cholesky decomposition; R k Tria(A) is used to measure the covariance of noise. Tria(A) represents the transpose of the upper triangular matrix obtained by orthogonal decomposition of the full-rank column matrix A.

[0049] (2.4) Calculate the cross-covariance

[0050]

[0051] (2.5) Utilizing the centroid measurement information of the group target Estimate the state of motion of the center of mass

[0052]

[0053] G k This is the filter gain matrix.

[0054] Step 3: Estimate the current contour shape of the group of targets.

[0055] To ensure high real-time performance of the group target tracking algorithm, such as Figure 1 As shown, in a two-dimensional scene, the outline of the group of targets is simplified and modeled as a time-varying ellipse as shown in the following formula, with all measurement sources located inside the ellipse.

[0056]

[0057] In the formula, y is the coordinate vector of any point on the elliptical outline of the target group; E k Let be the parameter matrix of the ellipse. This matrix is ​​a symmetric positive definite random matrix, and its transition probability distribution function is shown in the following equation.

[0058]

[0059] In the formula, δ represents the degrees of freedom; A k-1Let C be the transition matrix; W(Y; a, C) represents a symmetric positive definite random matrix C that satisfies the Wishart distribution with degrees of freedom a and parameter matrix Y.

[0060] Based on the above modeling method, using the centroid motion state information of the group targets obtained in the second step, the contour shape of the group targets at the current moment is estimated by Bayesian filtering, as shown in the following formula.

[0061]

[0062] Let D be the covariance matrix of the group centroid measurement distribution. k|k-1 This is an intermediate calculation parameter matrix with no explicit physical meaning. (B) k To expand the target contour morphology modeling error matrix, in represents the degree of freedom parameter in the inverse Wishart distribution;

[0063] If the tracking task for the group target has been completed, exit the group target tracking algorithm; otherwise, continue to execute step four.

[0064] Step 4: Predict the motion state of the group's center of mass at the next moment.

[0065] Using the estimation results of the group target centroid motion state at time k, the motion state of the group target centroid at time k+1 is predicted. The specific steps are as follows:

[0066] (4.1) Estimating the motion state transition matrix based on deep neural networks

[0067] like Figure 3 As shown, the time series input consists of the group target centroid measurement data from time k-w+1 (where w is the sliding window length) to time k, constructed using a dynamic sliding window approach. A deep neural network is then used to extract the motion characteristics of the group target centroids, thus outputting the key parameters in the state transition matrix. First, the centroid measurement data is transformed to the coordinate system of the centroid's motion state. Then, deviation standardization is performed, followed by convolutional filtering to extract features from the data. Next, a bidirectional long short-term memory (Bi-LSTM) network is used to extract the relevant motion characteristic information of the group target centroids from the time series. Finally, a fully connected layer is used to output the key parameters in the motion state transition matrix.

[0068] Without loss of generality, if the motion state vector of the group target's centroid is time (p) x p y Let v be the coordinates of the group's centroid on the x-axis and y-axis of the Cartesian coordinate system. x v y(where ρ represents the velocity of the target's centroid in the x-axis and y-axis directions, respectively). The estimation result of the motion state transition matrix obtained based on the deep neural network is shown in the following equation.

[0069]

[0070] In the formula, f 1,k and f 2,k Two parameters determine the displacement increments of the group's centroid in the x-axis and y-axis directions within the two-dimensional plane; f 3,k and f 4,k Two parameters determine the velocity of the group's centroid in the x-axis and y-axis directions in the two-dimensional plane.

[0071] (4.2) Estimation of process noise variance matrix based on deep neural network

[0072] like Figure 4 As shown, the measurement data of the centroid of the group targets from time k-w+1 (where w is the sliding window length) to time k are input through a dynamic sliding window method and then transformed to the coordinate system where the centroid's motion state is located. After deviation standardization, features in the data are extracted through convolutional filtering. Next, on the one hand, second-order difference is used to determine whether the motion mode of the group targets' centroid has changed, and then a multi-layer Bi-LSTM network is used to extract the motion characteristics of the group targets' centroid from the time series. Then, a multi-layer perceptron and a normalized exponential function are used to output the probability of its motion mode matching with models such as constant velocity motion (CV) and constant turning rate (CT), and the structural form of the process noise variance matrix is ​​selected based on the maximum matching probability. On the other hand, the process noise statistical characteristics are extracted from the time series through a multi-layer Bi-LSTM network, and then the process noise intensity estimation result is output through a fully connected layer. Combining the above results, the estimation result of the process noise variance matrix can be obtained.

[0073] For example, when Figure 4 When a deep neural network determines that the motion mode of the centroid of a group of targets conforms to the CV model, the estimation result of the process noise variance matrix is ​​as follows:

[0074]

[0075] In the formula, T is the sampling period of the detection device; This is the result of the process noise intensity estimation.

[0076] For example, when Figure 4 When the deep neural network in the model determines that the motion mode of the centroid of the group of targets conforms to the CT model, the estimation result of the process noise variance matrix is ​​as follows:

[0077]

[0078] (4.3) Predicting the motion state of the center of mass of the group target

[0079] State transition matrix estimation results obtained using deep neural networks Estimation results of process noise variance matrix Based on the capacitive Kalman filter method, the motion state of the centroid of the group target at time k+1 is predicted. The specific steps are as follows:

[0080] (4.3.1) Calculate the state volume point

[0081]

[0082] (4.3.2) Predicting the motion state of the center of mass of the group target

[0083]

[0084] Step 5: Predict the outline shape of the group of targets at the next moment.

[0085] Using the target group contour morphology estimation result at time k, the target group contour morphology at time k+1 is predicted, as shown in the following formula.

[0086]

[0087] In the formula, d represents the estimation result of the group target's outline parameter matrix. The dimension of λ k E is the elliptical contour parameter matrix k Parameters that affect the degree of influence of measurement noise variance.

[0088] After completing the above steps, return to step one, wait for the next sampling time to begin, and then continue to track the group targets according to the above steps.

[0089] Example

[0090] First, based on the characteristics of the tracked group of targets, training and validation sets are generated. Target motion modes include constant velocity motion, constant speed turning, etc. Figure 3 In the motion state transition matrix estimation model shown, the convolutional layers consist of two-dimensional and one-dimensional convolutions, and the Bi-LSTM network has three layers. During training, the mean squared error loss function is chosen, and the Adam optimizer is used. Figure 4 In the process noise variance matrix estimation model shown, the convolutional layers consist of two-dimensional and one-dimensional convolutions, and the Bi-LSTM networks are all two-layered. The Bi-LSTM network used to train motion pattern determination uses the cross-entropy loss function, and the Bi-LSTM network used to train process noise intensity estimation uses the mean squared error loss function. Both use the Adam optimizer.

[0091] The initial motion state of the group's target centroid is The centroid of the target group moves at a constant speed in a straight line from 0 to 60s, at a constant speed in a turning motion from 61s to 120s, at a constant speed in a straight line from 121s to 180s, at a constant speed in a turning motion from 181s to 240s, and at a constant speed in a straight line from 241s to 300s; the sampling period of the detection equipment is 1s.

[0092] The actual trajectory of the group of targets and the measurement data at each sampling time are as follows: Figure 5 As shown; the tracking results of the intelligent target tracking method for maneuvering swarms based on deep neural networks proposed in this invention on swarm targets are as follows. Figure 6 As shown; based on the above experimental results, the method proposed in this invention can track maneuvering group targets in real time with high precision.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent target tracking in a maneuvering swarm based on deep neural networks, characterized in that, Includes the following steps: The first step is to obtain the current group target measurement information; specifically: 1) For a group of targets with densely distributed units within the group, the detection device can acquire multiple measurement data generated by the units within the group at each sampling time, thus forming a measurement dataset. ,in Let k be the number of measurement data acquired at time k. It is generated by the l-th measurement source that appears at time k, and the measurement equation is as follows: ； In the formula, h(·) represents the measurement function. This represents the state vector of the l-th measurement source. Indicates measurement noise; 2) Based on the above measurement dataset, calculate the target centroid measurement information at time k using the following formula. ； The second step is to estimate the motion state of the group's centroid at the current moment; specifically: The specific steps for estimating the centroid motion state of a group of targets using the capacitive Kalman filter method are as follows: 1) Calculate and measure the volume point ； In the formula, , The motion state vector of the centroid of the group target dimensionality; The square root factor of the prior error covariance of the centroid motion state of the group targets; For set The j-th element in the set Defined as follows ； 2) Calculate the prior measurement of the centroid of the group target ； 3) Calculate the square root factor of the prior information covariance. ； In the formula, Represents Cholesky decomposition; R k To measure the covariance of noise; Represents a column-full rank matrix The transpose of the upper triangular matrix obtained after orthogonal decomposition; 4) Calculate cross-covariance ； 5) Utilizing group target centroid measurement Estimate the state of motion of the center of mass ； The third step is to estimate the outline shape of the group of targets at the current moment; The fourth step is to predict the motion state of the group's center of mass at the next moment; The fifth step is to predict the outline shape of the group of targets at the next moment.

2. The intelligent target tracking method for maneuvering groups based on deep neural networks according to claim 1, characterized in that, The third step is as follows: In a two-dimensional scene, the outline of the group of targets is simplified and modeled as a time-varying ellipse, as shown in the following equation. ； In the formula, y is the coordinate vector of any point on the elliptical outline of the target group; E k The parameter matrix of the ellipse is a symmetric positive definite random matrix, and its transition probability distribution function is shown in the following equation. ； In the formula, For degrees of freedom; Let C be the transition matrix; W(Y; a, C) represents a symmetric positive definite random matrix C that satisfies the Wishart distribution with degrees of freedom a and parameter matrix Y; The contour shape of the target group at the current moment is estimated by Bayesian filtering, as shown in the following formula. ； in, The covariance matrix of the group centroid measurement distribution. For intermediate calculation parameter matrix, To expand the target contour morphology modeling error matrix, ,in d represents the estimation result of the group target's outline parameter matrix. dimensionality represents the degree of freedom parameter in the inverse Wishart distribution; If the tracking task for the group target has been completed, exit the group target tracking algorithm; otherwise, continue to execute step four.

3. The intelligent target tracking method for maneuvering groups based on deep neural networks according to claim 2, characterized in that, The fourth step is as follows: Using the estimation results of the group target centroid motion state at time k, the motion state of the group target centroid at time k+1 is predicted. The specific steps are as follows: 1) Estimating the state transition matrix based on deep neural networks The system uses a dynamic sliding window approach to construct a time series input from the group target centroid measurement data from time k-w+1 to the current time k, where w is the sliding window length. A deep neural network is then used to extract the motion characteristics of the group target centroids, outputting key parameters in the state transition matrix. First, the centroid measurement data is transformed to the coordinate system of the centroid's motion state. Then, deviation standardization is performed, followed by convolutional filtering to extract features from the data. Next, a Bi-LSTM network is used to extract the relevant motion characteristics of the group target centroids from the time series. Finally, a fully connected layer is used to output the key parameters in the state transition matrix. 2) Estimating process noise variance based on deep neural networks The centroid measurement data from time k-w+1 to time k are used as a time series input through a dynamic sliding window approach. This data is then transformed to the coordinate system of the centroid's motion state and standardized. Next, convolutional filtering is used to extract features from the data. Then, second-order difference is used to determine if the group target's motion pattern has changed. A multi-layer Bi-LSTM network is then used to extract the group target's motion characteristics from the time series. A multi-layer perceptron and a normalized exponential function are used to output the probability that the group target's centroid motion pattern matches a constant velocity motion or constant turning rate model. The structure of the process noise variance matrix is ​​selected based on the maximum matching probability. A multi-layer Bi-LSTM network is then used to extract process noise characteristics from the time series, and a fully connected layer outputs the process noise intensity estimation result. Finally, the results from both aspects are combined to obtain the estimated result of the process noise variance matrix. 3) Predict the motion state of the center of mass of the group target State transition matrix estimation results obtained using deep neural networks Estimation results of process noise variance matrix Based on the capacitive Kalman filter method, the motion state of the centroid of the group target at time k+1 is predicted. The specific steps are as follows: (3.1) Calculate the state volume point ； (3.2) Predicting the motion state of the center of mass of the group target 。 4. The intelligent target tracking method for maneuvering swarms based on deep neural networks according to claim 3, characterized in that, The fifth step is as follows: Using the target group contour morphology estimation result at time k, the target group contour morphology at time k+1 is predicted, as shown in the following formula. ； In the formula, Elliptical contour parameter matrix Parameters that affect the degree of influence of measurement noise variance; After completing the above steps, return to step one, wait for the next sampling time to begin, and then continue to complete the tracking of the group target according to the above steps.

5. A maneuvering swarm target intelligent tracking system based on deep neural networks, characterized in that, The system for implementing the method of any one of claims 1-4 comprises: The first module is used to obtain the group target measurement information at the current moment; The second module is used to estimate the motion state of the group's centroid at the current moment; The third module is used to estimate the contour shape of the group of targets at the current moment; The fourth module is used to predict the motion state of the group's centroid at the next moment; The fifth module is used to predict the outline shape of the group of targets at the next moment.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method described in any one of claims 1-4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-4.

Images