Target tracking method and device based on feedback learning

Abstract

The application relates to a target tracking method and device based on feedback learning. Through multiple execution of data association, local estimation and data fusion iteration processes, the target state information at the last moment is fed back to the calculation of the predicted measurement and the predicted measurement error covariance of each sensor about the target, the adjustment of the position and size of the gate during data association and the optimization of the local estimation at the current iteration, until the iteration round reaches the preset iteration number, so that the accuracy of data association and the fusion estimation accuracy during target tracking are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing technology, and in particular to a target tracking method and apparatus based on feedback learning. Background Technology

[0002] In the era of big data, as sensor monitoring scenarios become increasingly complex, single-sensor monitoring systems can no longer meet the demands. To solve the target tracking problem in complex scenarios, extending single-sensor target tracking to multi-sensor target tracking is imperative. As a core technology of modern sensor systems, multi-sensor target tracking and information fusion technology has been widely applied in fields such as autonomous driving, weather monitoring, and biomedicine.

[0003] Target tracking comprises two parts: data association and target state estimation. Data association is a prerequisite for state estimation; erroneous sensor data will degrade estimation performance. In practical applications, ideal target measurement conditions, such as measurements from isolated target sources at every time interval, are rare. On the one hand, sensor observation inevitably involves measurement errors and missed detections. On the other hand, to effectively measure targets employing stealth measures, the sensor detection threshold needs to be lowered, leading to a large influx of clutter and an increase in false alarms, thus increasing the complexity of multi-sensor target data association. Furthermore, outliers generated by land and sea clutter, interference, and deception cause sensor model mismatch, and existing methods, relying on prior knowledge, struggle to cope with model mismatch, easily resulting in degraded fusion performance.

[0004] In target tracking, when there are erroneous data associations, the traditional fusion process is open-loop, difficult to optimize iteratively, and relies on prior models. This leads to a decrease in the accuracy of data association and the accuracy of fusion estimation in the current distributed open-loop fusion processing architecture when the model is mismatched. Summary of the Invention

[0005] Therefore, it is necessary to provide a target tracking method and apparatus based on feedback learning to address the above-mentioned technical problems. By improving the accuracy of data association through feedback learning, the accuracy of fusion estimation can also be directly improved.

[0006] In a first aspect, the present invention provides a target tracking method based on feedback learning, comprising the following steps:

[0007] Establish a nonlinear measurement model for each sensor;

[0008] Establish a discrete-time nonlinear state model of the target at the current moment;

[0009] The measurements of each sensor with respect to the target at the current moment are obtained based on the nonlinear measurement model of each sensor.

[0010] Using the discrete-time nonlinear state model of the target at the current moment and the nonlinear measurement model of each sensor, the predicted measurement and prediction measurement error covariance of each sensor with respect to the target at the current moment are calculated based on the target state information at the previous moment. The target state information at the previous moment includes the state estimate and state estimate error covariance at the previous moment.

[0011] Based on the predicted measurements of each sensor about the target at the current moment and the measurements of each sensor about the target, a data association gate is constructed for each sensor, and a candidate measurement is determined based on each data association gate;

[0012] Local extended Kalman filtering is performed based on each candidate measurement and the target state predicted at the current moment to obtain a local estimate of the target for each sensor;

[0013] The target's state estimate at the current moment is obtained by performing matrix weighted fusion of the local estimates from all sensors.

[0014] Determine if the number of iterations is less than the preset number of iterations. If so, use the target's state estimation information at the current moment as the target's state information at the previous moment in the next iteration process and return to the step of establishing the discrete-time nonlinear state model of the target at the current moment. Otherwise, target tracking ends.

[0015] In one embodiment, a nonlinear measurement model for each sensor is established based on the measurement characteristics of each sensor, and a discrete-time nonlinear state model for the target is established based on the target's current motion.

[0016] In one embodiment, the nonlinear measurement model for each sensor is:

[0017] z i,k =h i (x k )+ν i,k

[0018] In the formula, x represents the measurement of the target by sensor i. k h represents the state vector at time k. i (·) represents the nonlinear measurement function of sensor i, ν i,k The noise of sensor i at time k is represented by i = 1, ..., L, where i = 1, ..., L represents the number of sensors.

[0019] The discrete-time nonlinear state model of the target is as follows

[0020] x k =f(x) k-1 )+ω k-1

[0021] In the formula, Let f(·) represent the state vector at time k-1, and let f(·) represent the nonlinear state transition function used to characterize the target's motion at the current time. ω k-1 This indicates process noise.

[0022] In one embodiment, the predicted measurement of the target at the current moment and the predicted measurement error covariance of each sensor are calculated based on the target state information from the previous moment, using the discrete-time nonlinear state model of the target at the current moment and the nonlinear measurement model of each sensor.

[0023]

[0024]

[0025]

[0026]

[0027] In the formula, Estimate the target state from the previous moment. This represents the predicted target state at the current moment. Q is the predicted measurement of sensor i with respect to the target at the current moment. k-1 Let k be the noise covariance of the Gaussian process at time k-1. This represents the state estimate at the previous time step. The state transition matrix P is obtained by Taylor series expansion. k-1 This represents the error covariance of the state estimation at the previous time step. Indicates in The measurement matrix R is obtained by Taylor series expansion. i,k It is the measurement noise covariance of sensor i.

[0028] In one embodiment, the data association gate is a fan-shaped gate, and the gate center of the data association gate is the predicted measurement of sensor i about the target at the current time;

[0029] The mathematical expression for the data association gate of sensor i is:

[0030]

[0031]

[0032] In the formula, and These are the distance echo and angle echo that constitute the j-th measurement of sensor i falling into the gate at time k. and The gate center is adjusted based on the target state information from the previous moment, and consists of the distance prediction and angle prediction of sensor i with respect to the target at time k. ρ and K θ It is based on the gate probability and χ² 2 The square root of the parameter obtained from the distribution table, and Let be the variances of the measurement errors of the distance ρ and angle θ between sensor i and the target, respectively. yes The prediction variance yes The prediction variance and Together, they determine the gate size.

[0033] In one embodiment, determining a candidate measurement based on a data association gate is as follows: if there is a measurement about the target within the data association gate, then the measurement about the target is a candidate measurement; if there is no measurement about the target within the data association gate, then the predicted measurement about the target is a candidate measurement; if there are multiple measurements about the target within the data association gate, then the measurement about the target that is closest to the predicted measurement about the target is selected as a candidate measurement.

[0034] In one embodiment, the candidate measurement is calculated as follows:

[0035]

[0036] In the formula, For sensor i, Let S be the j-th measurement of sensor i with respect to the target. i,k It is the error covariance matrix of the measurement prediction of sensor i at time k.

[0037] In one embodiment, the state update equation for the locally extended Kalman filter is:

[0038]

[0039]

[0040] In the formula, P represents the local estimate of the target by each sensor at the current moment. i,k express The error covariance.

[0041] In one embodiment, matrix-weighted fusion of the local estimates from all sensors includes:

[0042] The Dixon criterion is used to filter out anomalous local estimates from all sensor local estimates;

[0043] The target's state estimation information at the current moment is calculated based on the residual normal local estimates from the local estimates of all sensors;

[0044] The formula for calculating the target's state estimation information at the current moment is:

[0045]

[0046] P k =(e T Ψe) -1

[0047] In the formula, It is the fused state estimate of the target at the current moment, P. k yes The estimation error covariance, It is an estimation vector composed of normal local estimates, where l is the number of normal local estimates, and Ψ = [P ij [,k],i,j=1,…,l≤L is a matrix consisting of the cross covariances of normal local estimates and I n It is an n-dimensional identity matrix.

[0048] Secondly, the present invention also provides a target tracking device based on feedback learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a target tracking method based on feedback learning.

[0049] The beneficial effects of this invention are as follows: In this invention, the sensor uses the target state information obtained from the previous moment through fusion to adjust the gate center and radius of the data association gate at the current moment, thereby improving the accuracy of data association. Furthermore, based on the target state information obtained from the previous moment through fusion, the sensor data is checked, and normal local estimation is selected for matrix weighted fusion, thereby improving the accuracy of fusion estimation. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the target tracking method based on feedback learning provided in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the data association sector gate provided in an embodiment of the present invention;

[0052] Figure 3 These are schematic diagrams illustrating trajectory tracking using different methods and real trajectories provided in embodiments of the present invention;

[0053] Figure 4This is a schematic diagram comparing the root mean square error curves of the feedback-free correlation fusion algorithm provided in the embodiments of the present invention and the method of the present invention;

[0054] Figure 5 This is a schematic diagram of the association of sensor 1 provided in an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of the association of sensor 2 provided in an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram of the association of sensor 3 provided in an embodiment of the present invention;

[0057] Figure 8 This is a schematic diagram of trajectory tracking using the non-feedback correlation fusion algorithm under maneuvering and interference in this embodiment of the invention and the method of this invention;

[0058] Figure 9 This is a schematic diagram comparing the root mean square error curves of the feedback-free correlation fusion algorithm under maneuvering and interference in the embodiments of the present invention and the method of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] In one embodiment, such as Figure 1 As shown, Figure 1 This is one of the flowcharts illustrating a target tracking method based on feedback learning provided in this embodiment of the invention. When applied to a computer device, this method includes the following steps:

[0061] S101. Establish a nonlinear measurement model for each sensor.

[0062] Specifically, the nonlinear measurement model for each sensor is...

[0063] z i,k =h i (x k )+ν i,k

[0064] In the formula, The measurement of sensor i with respect to the target refers to the observation data received from sensor i, used to describe the state of the target at a certain moment. k The state vector at time k represents the target's true state at a given moment, including attributes such as position, velocity, and acceleration. i (·) represents the nonlinear measurement function of sensor i, ν i,kLet i represent the measurement noise of sensor i at time k, where i = 1, ..., L represents the number of sensors.

[0065] S102. Establish a discrete-time nonlinear state model of the target at the current moment.

[0066] Specifically, the discrete-time nonlinear state model of the target is as follows:

[0067] x k =f(x) k-1 )+ω k-1

[0068] In the formula, Let f(·) represent the state vector at time k-1, and let f(·) represent the nonlinear state transition function used to characterize the target's motion at the current time. ω k-1 This indicates process noise.

[0069] S103. Obtain the measurement of the target by each sensor at the current moment based on the nonlinear measurement model of each sensor.

[0070] The measurements taken by each sensor regarding the target at any given moment contain both real measurements and clutter. Clutter refers to non-target signals or noise from the sensors. This clutter can affect the target's measurement data, potentially causing the tracking algorithm to misjudge the target's position, velocity, or other states. Clutter may be incorrectly identified as a real target, causing the tracking algorithm to focus on false targets and ignore the true target. A large amount of clutter can make data association (matching continuous measurement data with the target to determine its trajectory) difficult, increasing the complexity of data association and reducing the accuracy of the tracking algorithm. Measurement data is inevitably affected by sensor errors, environmental interference, and other factors; therefore, it needs to be processed and filtered to reduce errors and improve tracking accuracy.

[0071] S104. Using the discrete-time nonlinear state model of the target at the current moment and the nonlinear measurement model of each sensor, calculate the predicted measurement and prediction measurement error covariance of each sensor about the target at the current moment based on the target state information of the previous moment.

[0072] The target state information at the previous time step includes the state estimate and the state estimate error covariance at the previous time step.

[0073] Specifically, using the discrete-time nonlinear state model of the target at the current moment and the nonlinear measurement model of each sensor, the predicted measurement of the target at the current moment and the formula for calculating the predicted measurement error covariance of each sensor are calculated based on the target state information from the previous moment.

[0074]

[0075]

[0076]

[0077]

[0078] In the formula, Estimate the target state from the previous moment. This represents the predicted target state at the current moment. Q is the predicted measurement of sensor i with respect to the target at the current moment. k-1 Let k be the noise covariance of the Gaussian process at time k-1. This represents the state estimate at the previous time step. The state transition matrix P is obtained by Taylor series expansion. k-1 This represents the error covariance of the state estimation at the previous time step. Indicates in The measurement matrix R is obtained by Taylor series expansion. i,k It is the measurement noise covariance of sensor i.

[0079] S105. Construct a data association gate for each sensor based on the predicted measurement of the target by each sensor at the current time and the measurement of the target by each sensor, and determine a candidate measurement based on each data association gate.

[0080] In this embodiment, as Figure 2 As shown, the data association gate is a fan-shaped gate, and the center of the data association gate is the predicted measurement of sensor i about the target at the current time.

[0081] Specifically, the mathematical expression for the data association gate of sensor i is:

[0082]

[0083]

[0084] In the formula, and These are the distance echo and angle echo that constitute the j-th measurement of sensor i falling into the gate at time k. and The gate center is adjusted based on the target state information from the previous moment, and consists of the distance prediction and angle prediction of sensor i with respect to the target at time k. ρ and K θ It is based on the gate probability and χ² 2 The square root of the parameter obtained from the distribution table, and Let be the variances of the measurement errors of the distance ρ and angle θ between sensor i and the target, respectively. yes The prediction variance yes The prediction variance and Together, they determine the gate size.

[0085] Determining a candidate measurement based on a data association gate is as follows: if there is a measurement about the target within the data association gate, then the measurement about the target is a candidate measurement; if there is no measurement about the target within the data association gate, then the predicted measurement about the target is a candidate measurement; if there are multiple measurements about the target within the data association gate, then the measurement about the target that is closest to the predicted measurement about the target is selected as the candidate measurement.

[0086] The formula for calculating the candidate measurement is:

[0087]

[0088] In the formula, For sensor i, Let S be the j-th measurement of sensor i with respect to the target. i,k It is the error covariance matrix of the measurement prediction of sensor i at time k.

[0089] S106. Perform local extended Kalman filtering based on each candidate measurement and the target state predicted at the current moment to obtain the local estimate of the target by each sensor.

[0090] Specifically, the state update equation for the locally extended Kalman filter is:

[0091]

[0092]

[0093] In the formula, P represents the local estimate of the target by each sensor at the current moment. i,k express The error covariance.

[0094] S107. Perform matrix weighted fusion of the local estimates from all sensors to obtain the target's state estimate information at the current moment.

[0095] Specifically, the matrix-weighted fusion of local estimates from all sensors includes:

[0096] (1) Use the Dixon criterion to filter out abnormal local estimates in the local estimates of all sensors.

[0097] Dixon's criterion is

[0098]

[0099]

[0100] x remain ={x (i) ,i=1,…,L|max(D n ,D′ n )≤D}

[0101] Among them, D L Let D′ be the statistic for the upper-side test. L Let {x} be the statistic for the lower-side test. (i) Let {x, i = 1, ..., L} be the set of local estimates arranged in ascending order. remain It is a local estimate set after removing outliers, and D is the critical value of the test statistic obtained by looking up the table based on the detection level α.

[0102] (2) Calculate the target's state estimation information at the current moment based on the remaining normal local estimates in the local estimates of all sensors.

[0103] The formula for calculating the target's state estimation information at the current moment is:

[0104]

[0105] P k =(e T Ψe) -1

[0106] In the formula, It is the fused state estimate of the target at the current moment, P. k yes The estimation error covariance, It is an estimation vector composed of normal local estimates, where l is the number of normal local estimates, and Ψ = [P ij [,k],i,j=1,…,l≤L is a matrix consisting of the cross covariances of normal local estimates and I n It is an n-dimensional identity matrix.

[0107] S108. Determine if the number of iterations is less than the preset number of iterations. If so, use the target's state estimation information at the current moment as the target's state information at the previous moment in the next iteration process and return to the step of establishing the discrete-time nonlinear state model of the target at the current moment. Otherwise, target tracking ends.

[0108] Specifically, in this embodiment, the preset number of iterations is the ratio of the total detection time of the actual sensor to the detection interval time of the sensor.

[0109] The target tracking method based on feedback learning in this embodiment introduces the Dixon criterion and multi-sensor fusion feedback learning on the basis of data association and extended Kalman filtering. It uses the target state information obtained from the previous time step obtained by fusion to adjust the gate center and radius of the data association gate at the current time step, thereby improving the accuracy of data association. Furthermore, it uses the target state information obtained from the previous time step obtained by fusion to verify the sensor data and selects normal local estimation for matrix weighted fusion, thereby improving the accuracy of target state estimation.

[0110] In a specific embodiment, taking a wheeled mobile robot as an example, simulation experiments are conducted on the method of the present invention, the association weighted fusion algorithm without feedback in dense clutter environment, and the association weighted fusion algorithm with feedback in case of erroneous association, when the target does not move. Secondly, simulation experiments are conducted in target movement and interference environment, and the association accuracy is compared under target trajectory, association status, root mean square error and different clutter densities to verify that the present invention has high association accuracy and good state estimation accuracy.

[0111] In this embodiment, the discrete-time nonlinear state model is as follows:

[0112]

[0113] In the formula, Represents the state vector. and These represent the positions of the wheeled mobile robot in the X and Y directions, respectively. Indicates the direction of the vehicle relative to the overall frame, Δ T Let f(·) represent the sampling period, v be the linear velocity of the wheeled mobile robot, and w be the angular velocity of the wheeled mobile robot. f(·) represents the nonlinear state transition function, and ω... k-1 Let represent the Gaussian process noise at time k-1, and satisfy ω k-1 ~N(0,Q) k-1 ), 0 and Q k-1 These are the mean and variance of the noise, respectively.

[0114] The nonlinear measurement model for each sensor is as follows:

[0115]

[0116] In the formula, The position of sensor i is indicated, and i = 1, ..., L represents the number of sensors.

[0117] The relevant parameters for tracking a wheeled mobile robot using a feedback learning-based target tracking method are: initial state x0 = [0, 0, 0] T Sampling period Δ T =1s, a total of 200 samples are taken (preset iteration number), i.e., k=200. Linear velocity v=20cm / s, angular velocity w=0.01πrad / s, system process noise covariance matrix. Initial covariance matrix Sensor position The measurement noise variances are (0cm, 5cm), (10cm, 1cm), and (10cm, 15cm), respectively. Clutter density λ = 2, parameter K ρ =3 and K θ =3, α=0.95, D=0.97.

[0118] like Figure 3 As shown, Figure 3 This is a schematic diagram of trajectory tracking using different methods and real trajectories provided in the embodiments of the present invention. The solid square lines represent trajectory tracking with feedback correlation fusion, the solid circular lines represent trajectory tracking with feedback correlation fusion when sensors track clutter at different time periods, and the solid asterisk lines represent trajectory tracking without feedback correlation fusion. Figure 3 This demonstrates that fusion with feedback outperforms open-loop fusion without feedback, and that feedback learning can effectively prevent performance degradation even with a large amount of measurement data association errors. The trajectory fused with feedback has the smallest error compared to the true trajectory, therefore, fusion with feedback outperforms open-loop fusion without feedback.

[0119] To further demonstrate the good estimation accuracy of the proposed fusion algorithm, we introduce root mean square error (RMSE) and a feedback-free fusion algorithm for comparison. Figure 4 The comparison of the position RMSE between the feedback-free correlation fusion algorithm and the proposed fusion algorithm is presented.

[0120]

[0121] Where k is the sampling time, and It is the position and state estimation information of a wheeled mobile robot. and This represents the actual positional state information of the wheeled mobile robot. The RMSE of the position obtained by the feedback-based correlation fusion algorithm is lower than that obtained by the non-feedback-based correlation fusion algorithm, demonstrating better estimation accuracy.

[0122] To verify that feedback can improve the accuracy of sensor correlation with target measurements, the correlation results of correlation fusion algorithms with and without feedback learning for sensors 1, 2, and 3 were compared under the same parameter settings. Figure 5 , Figure 6 and Figure 7 Sensors 1, 2, and 3 are compared with actual measurements, and a comparison is given between the non-feedback association fusion algorithm and the method of this invention. It can be seen that compared with the non-feedback learning case, the method of this invention improves the association accuracy, that is, increases the selection probability of the actual measurement.

[0123] To further verify that the proposed correlation fusion algorithm can improve the correlation accuracy, under the same parameter settings, the correlation accuracy of the correlation fusion algorithm with feedback learning (the method of this invention) and the correlation fusion algorithm without feedback learning under different clutter densities were compared. The comparison results are shown in Table 1. Table 1 shows the accuracy of the two correlation fusion algorithms under different clutter densities. It can be seen that the clutter density affects the correlation accuracy, and the correlation fusion algorithm with feedback learning can improve the correlation accuracy.

[0124] Table 1. Association accuracy of different algorithms

[0125]

[0126] like Figure 8 and Figure 9 As shown, Figure 8 and Figure 9 The tracking performance and position RMSE of the feedback-free correlation fusion algorithm and the method of this invention are compared respectively. To further demonstrate the superior performance of the method of this invention compared to the open-loop feedback-free fusion method, in the case of simultaneous target maneuvering and interference, the feedback learning endows the method with the ability of data discrimination, model learning, and fusion optimization, thus achieving higher estimation accuracy than the feedback-free fusion method. Assume that the target maneuvers when k=100 (in... Figure 8 This manifests as a left turn during a circular motion of a wheeled mobile robot, causing its trajectory to deviate from a circular pattern. The target's maneuverability affects its motion mode, velocity, acceleration, and other state parameters. In other words, the discrete-time nonlinear state model of the target cannot accurately describe its current motion, specifically resulting in increased modeling error (Q...). 100 =10Q 100 Q 100 (Expressed as the Gaussian process noise covariance at 100 seconds), similarly, when measurement interference exists, it leads to an increase in sensor observation error, i.e., R. 3,100 =2R 3,100 (R 3,100(This represents the observation error of the measurement equation of sensor 3 at 100 seconds). At this time, the correlation accuracy of the non-feedback correlation fusion algorithm is 82%, 82%, and 81.5%, respectively, while the correlation accuracy of the feedback learning correlation fusion algorithm is 83.5%, 82%, and 85%, respectively. This fully demonstrates that the target tracking method based on feedback learning in this invention improves the data correlation accuracy and fusion estimation precision.

[0127] Based on the same inventive concept, embodiments of the present invention also provide a feedback-based target tracking device for implementing the above-described feedback-based target tracking method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more feedback-based target tracking device embodiments provided below can be found in the limitations of the feedback-based target tracking method described above, and will not be repeated here.

[0128] In one embodiment, a target tracking device based on feedback learning includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a target tracking method based on feedback learning.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A target tracking method based on feedback learning, characterized in that, Includes the following steps: Establish a nonlinear measurement model for each sensor; Establish a discrete-time nonlinear state model of the target at the current moment; The measurements of each sensor with respect to the target at the current moment are obtained based on the nonlinear measurement model of each sensor. Using the discrete-time nonlinear state model of the target at the current moment and the nonlinear measurement model of each sensor, the predicted measurement and prediction measurement error covariance of each sensor with respect to the target at the current moment are calculated based on the target state information at the previous moment. The target state information at the previous moment includes the state estimate and state estimate error covariance at the previous moment. Based on the predicted measurements of each sensor about the target at the current moment and the measurements of each sensor about the target, a data association gate is constructed for each sensor, and a candidate measurement is determined based on each data association gate; Local extended Kalman filtering is performed based on each candidate measurement and the target state predicted at the current moment to obtain a local estimate of the target for each sensor; The target's state estimate at the current moment is obtained by performing matrix weighted fusion of the local estimates from all sensors. Determine if the number of iterations is less than the preset number of iterations. If so, use the target's state estimation information at the current moment as the target's state information at the previous moment in the next iteration process and return to the step of establishing the discrete-time nonlinear state model of the target at the current moment. Otherwise, target tracking ends.

2. The target tracking method based on feedback learning according to claim 1, characterized in that, A nonlinear measurement model for each sensor is established based on its measurement characteristics, and a discrete-time nonlinear state model for the target is established based on the target's current motion.

3. The target tracking method based on feedback learning according to claim 2, characterized in that, The nonlinear measurement model for each sensor is as follows: z i,k =h i (x k )+ν i,k In the formula, Let xk represent the measurement of sensor i with respect to the target, and hk represent the state vector at time k. i (·) represents the nonlinear measurement function of sensor i, ν i,k The noise of sensor i at time k is represented by i = 1, ..., L, where i = 1, ..., L represents the number of sensors. The discrete-time nonlinear state model of the target is as follows: x k =f(x k-1 )+ω k-1 In the formula, Let f(·) represent the state vector at time k-1, and let f(·) represent the nonlinear state transition function used to characterize the target's motion at the current time. ω k-1 This indicates process noise.

4. The target tracking method based on feedback learning according to claim 3, characterized in that, Using the discrete-time nonlinear state model of the target at the current moment and the nonlinear measurement model of each sensor, the predicted measurement of the target at the current moment and the formula for calculating the predicted measurement error covariance of each sensor are calculated based on the target state information from the previous moment. In the formula, Estimate the target state from the previous moment. This represents the predicted target state at the current moment. Q is the predicted measurement of sensor i with respect to the target at the current moment. k-1 Let k be the noise covariance of the Gaussian process at time k-1. This represents the state estimate at the previous time step. The state transition matrix P is obtained by Taylor series expansion. k-1 This represents the error covariance of the state estimation at the previous time step. Indicates in The measurement matrix R is obtained by Taylor series expansion. i,k It is the measurement noise covariance of sensor i.

5. The target tracking method based on feedback learning according to claim 1, characterized in that, The data association gate is a fan-shaped gate, and the gate center of the data association gate is the predicted measurement of sensor i about the target at the current time; The mathematical expression for the data association gate of sensor i is: In the formula, and These are the distance echo and angle echo that constitute the j-th measurement of sensor i falling into the gate at time k. and The gate center is adjusted based on the target state information from the previous moment, and consists of the distance prediction and angle prediction of sensor i with respect to the target at time k. ρ and K θ It is based on the gate probability and χ² 2 The square root of the parameter obtained from the distribution table, and Let be the variances of the measurement errors of the distance ρ and angle θ between sensor i and the target, respectively. yes The prediction variance yes The prediction variance and Together, they determine the gate size.

6. The target tracking method based on feedback learning according to claim 1, characterized in that, Determining a candidate measurement based on a data association gate is as follows: if there is a measurement about the target within the data association gate, then the measurement about the target is a candidate measurement; if there is no measurement about the target within the data association gate, then the predicted measurement about the target is a candidate measurement; if there are multiple measurements about the target within the data association gate, then the measurement about the target that is closest to the predicted measurement about the target is selected as the candidate measurement.

7. The target tracking method based on feedback learning according to claim 1, characterized in that, The calculation formula for the candidate measurement is as follows: In the formula, For sensor i, Let S be the j-th measurement of sensor i with respect to the target. i,k It is the error covariance matrix of the measurement prediction of sensor i at time k.

8. The target tracking method based on feedback learning according to claim 1, characterized in that, The state update equation for the locally extended Kalman filter is: In the formula, P represents the local estimate of the target by each sensor at the current moment. i,k express The error covariance.

9. The target tracking method based on feedback learning according to claim 1, characterized in that, The matrix-weighted fusion of local estimates from all sensors includes: The Dixon criterion is used to filter out anomalous local estimates from all sensor local estimates; The target's state estimation information at the current moment is calculated based on the residual normal local estimates from the local estimates of all sensors; The formula for calculating the target's state estimation information at the current moment is: P k =(e T Ψe) -1 In the formula, It is the fused state estimate of the target at the current moment, P. k yes The estimation error covariance, It is an estimation vector composed of normal local estimates, where l is the number of normal local estimates, and Ψ = [P ij,k The matrix ], i, j = 1, ..., l ≤ L is composed of the cross covariances of normal local estimates and e = [I n ,…,I n ] T I n It is an n-dimensional identity matrix.

10. A target tracking device based on feedback learning, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the target tracking method based on feedback learning as described in any one of claims 1 to 9.

Images