A whole-process adaptive maneuvering target tracking method

Abstract

The application relates to the technical field of target tracking, in particular to a full-process adaptive maneuvering target tracking method. Through preprocessing of multi-source data, preprocessing data is obtained, and maneuvering characteristic parameters of a tracking target are extracted; the maneuvering characteristic parameters of the tracking target are input into a classification model to obtain the maneuvering type of the tracking target; then, an improved particle swarm optimization algorithm is used to introduce inertia weight dynamic attenuation and particle variation mechanism to optimize parameters of a multi-modal submodel, so that an optimized multi-modal submodel is obtained; finally, the maneuvering characteristic parameters of the tracking target are input into the optimized multi-modal submodel to predict the future state of the tracking target. The maneuvering characteristic parameters of the tracking target are extracted from multi-source data, so that the maneuvering type of the tracking target is divided in combination with the maneuvering characteristic parameters, a corresponding multi-modal submodel is selected based on the maneuvering type, and the tracking precision and tracking robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of target tracking technology, and more specifically, to a fully adaptive maneuvering target tracking method. Background Technology

[0002] For a long time, maneuvering target tracking has been a key research area in the field of target tracking, with wide applications in defense, aerospace, traffic control, remote sensing, and many other fields. During maneuvering target tracking, the target's motion state often exhibits nonlinear and time-varying characteristics, such as sudden turns of fighter jets, emergency acceleration and deceleration of vehicles, and orbital adjustments of spacecraft. These "maneuvering behaviors" cause rapid changes in the target's state (position, velocity, acceleration), posing a significant challenge to real-time and accurate state estimation.

[0003] Traditional maneuvering target tracking methods mainly employ CS or Jerk models. However, CS models struggle to handle sudden changes in high maneuverability, while Jerk models suffer from unavoidable errors in steady-state scenarios. The fixed fusion logic of traditional CS-Jerk models cannot be dynamically adjusted based on maneuver intensity, resulting in insufficient tracking accuracy across different maneuver types. Furthermore, model parameters rely on human experience; for example, the attenuation coefficient of the CS model, the segmentation period of the Jerk model, and the fusion weights of the traditional CS-Jerk model all require manual setting based on specific scenarios, lacking an adaptive optimization mechanism. When target types switch or maneuver characteristics change, parameter mismatch leads to a sharp increase in tracking errors. Traditional models also have response times as long as 0.3-0.5 seconds when facing sudden high maneuvers, failing to meet real-time tracking requirements and making them unsuitable for complex real-world application scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a fully adaptive maneuvering target tracking method to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, one objective of this invention is to provide a fully adaptive maneuvering target tracking method, comprising: S1: Collect multi-source data of the tracked target, preprocess the multi-source data to obtain preprocessed data, and extract the maneuver feature parameters of the tracked target based on the preprocessed data; S2: Create a classification model by inputting the maneuver characteristic parameters of the tracked target into the classification model to obtain the maneuver type of the tracked target; the maneuver type of the tracked target includes low maneuver type, medium maneuver type and high maneuver type; S3: Create multiple multimodal sub-models, and introduce dynamic decay of inertial weights and particle mutation mechanism through an improved particle swarm optimization algorithm to optimize the parameters of the multimodal sub-models, resulting in optimized multimodal sub-models; each multimodal sub-model corresponds to a maneuver type. S4: Input the maneuvering characteristic parameters of the tracked target into the optimized multimodal sub-model to predict the future state of the tracked target.

[0006] Preferably, the multi-source data includes location data, velocity data, acceleration data, acceleration rate of change data, and the acquisition time of each data point. The resulting dataset is:

[0007] Where n is the total number of targets being tracked; This is the location data of the i-th tracked target; It is the velocity data of the i-th tracked target; It is the acceleration data of the i-th tracked target; The acceleration rate of change data of the i-th tracked target; Let be the acquisition time corresponding to the i-th tracked target.

[0008] Preferably, the preprocessing step includes: S1.1: Calculate the mean and standard deviation of each multi-source subdata set, and based on... The criteria determine whether outliers exist in multi-source data; S1.2: If outliers exist in the multi-source data, remove the outliers to obtain standard data; S1.3: Based on the acquisition time, the sampling frequency of each multi-source sub-data of the tracking target is adjusted by linear interpolation to achieve spatiotemporal synchronization of each multi-source sub-data and obtain preprocessed data; S1.4: Extract the maneuvering characteristic parameters of the target from the preprocessed data using the sliding window method; the maneuvering characteristic parameters include the magnitude of acceleration change. , frequency of acceleration change rate f The curvature change of the root trajectory of motion .

[0009] Preferably, the classification model uses an improved SVM algorithm as the core algorithm of the adaptive maneuver classifier; the selection rules for the kernel function of the improved SVM algorithm include: when and When the target is determined to be of a low-maneuverability type, a linear kernel function is selected. ,in, It is the input feature vector; when and When the target is determined to be of medium maneuverability, a polynomial kernel function is selected. ,in, , It is the scaling factor. It is a constant term. It is the order of the polynomial; when and When the target is determined to be highly maneuverable, a radial basis function kernel is selected. ,in, .

[0010] Preferably, the multimodal sub-model includes: Low-maneuverability sub-model; applied to tracking targets of low-maneuverability type, the mean update formula of the low-maneuverability sub-model is: ,in, It is a correction factor. It is a tracking target The mean acceleration estimate at time t. It is a tracking target Acceleration observations at time t. It is the attenuation coefficient of the tracking target. It is the magnitude of the acceleration change; Medium maneuvering sub-model; applied to tracking targets of medium maneuvering type, the medium maneuvering sub-model adopts an improved Jerk model, by replacing the "constant Jerk assumption" of the rate of change of acceleration in the traditional Jerk model with the "piecewise linear Jerk assumption"; High-maneuverability sub-model; applied to tracking targets of high maneuverability, the high-maneuverability sub-model dynamically adjusts weights by introducing an attention mechanism. The Where max_val is the value in historical observations. The maximum value, It is the magnitude of acceleration change. f It is the frequency of the rate of change of acceleration. It is the change in curvature of the root locus of motion.

[0011] Preferably, replacing the "constant Jerk assumption" of the rate of change of acceleration in the traditional Jerk model with the "piecewise linear Jerk assumption" includes the following steps: S2.1: Based on the fluctuation frequency of the rate of change of acceleration f Divide the sampling time axis into different lengths; S2.2: Within each sampling time axis, the Jerk linear variation curve is fitted using the least squares method to calculate the Jerk variation parameters; the Jerk variation parameters include the curvature change of the root trajectory. And the initial Jerk value; S2.3: Curvature change based on the root trajectory of motion And calculate the fitting error using the initial Jerk value; S2.4: Set an error threshold and determine whether the fitting error is greater than or equal to the error threshold; S2.5: If the fitting error is greater than or equal to the error threshold, the sampling time axis is divided into multiple segments until the fitting error is less than the error threshold.

[0012] Preferably, step S3 includes: S3.1: Determine the parameters to be optimized for each multimodal sub-model. ,in, These are the correction coefficients for the low-maneuverability sub-model. It is the Jerk segment threshold of the medium-speed sub-model. These are the coefficients of the attention weights in the high-mobility submodel; S3.2: Construct the objective function: ; in, It is the root mean square error of the x-axis. It is the root mean square error of the y-axis. It is the root mean square error of the z-axis. It is the root mean square error of the speed. It is the root mean square error of acceleration. , , , , In order are , , , , The corresponding weighting coefficients; S3.3: By improving the particle swarm optimization algorithm, dynamic decay of inertial weight and particle mutation mechanism are introduced to continuously iterate and optimize the parameters to be optimized for each multimodal sub-model until the termination condition is met, and the iteration stops to obtain the optimized multimodal sub-model.

[0013] Preferably, step S3.3 includes the following steps: S3.31: Set the iteration termination condition; the iteration termination condition includes: When the objective function corresponding satisfy , , , , Stop iteration when the time is right and output the current optimal parameters. ; The iteration stops when the number of iterations reaches the iteration threshold, and the current optimal parameters are output. ; S3.32: Set the initial particle swarm size and initialize the initial position and initial velocity of each particle; each particle corresponds to a set of parameters to be optimized; S3.33: For each particle, iterate the particle's position and velocity based on the individual optimal solution and the global optimal solution, set the stage inertia weight according to the number of iterations, and mutate the tail particles during the iteration process until any iteration termination condition is met, then stop the iteration and output the optimal parameters obtained from the last iteration. The optimized multimodal sub-model was obtained.

[0014] Preferably, the inertia weight of the aforementioned stage changes according to the number of iterations and decreases continuously as the number of iterations increases.

[0015] On the other hand, this application also provides a fully adaptive maneuvering target tracking device, applied to the fully adaptive maneuvering target tracking method described in any one of the foregoing embodiments, comprising: Acquisition terminal; acquires multi-source data of the target being tracked through the acquisition terminal; Tracking end; the tracking end is communicatively connected to the acquisition end, and the tracking end completes the tracking of the target based on multi-source data of the target.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By extracting the maneuver feature parameters of the tracked target from multi-source data, the maneuver type of the tracked target is classified in combination with the maneuver feature parameters, and the corresponding multimodal sub-model is selected based on the maneuver type to improve tracking accuracy and tracking robustness; 2. By improving the PSO algorithm, the core parameters are adaptively optimized, which can be universally applied to multiple types of targets such as vehicles, drones, and missiles. The parameter adaptation accuracy is ≥ 95%, which solves the problem of traditional models relying on human experience for adjustment. In high-mobility and sudden change scenarios, the response time of model switching and parameter adjustment is ≤ 0.1s, which is an improvement over the response time of traditional models and meets the real-time requirements of scenarios such as missile guidance and autonomous driving. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a fully adaptive maneuvering target tracking method. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 like Figure 1 As shown, one of the objectives of this invention is to provide a fully adaptive maneuvering target tracking method, comprising: S1: Collect multi-source data of the tracked target, preprocess the multi-source data to obtain preprocessed data, and extract the maneuver feature parameters of the tracked target based on the preprocessed data; S2: Create a classification model by inputting the maneuver characteristic parameters of the tracked target into the classification model to obtain the maneuver type of the tracked target; the maneuver type of the tracked target includes low maneuver type, medium maneuver type and high maneuver type; S3: Create multiple multimodal sub-models, and introduce dynamic decay of inertial weights and particle mutation mechanism through an improved particle swarm optimization algorithm to optimize the parameters of the multimodal sub-models, resulting in optimized multimodal sub-models; each multimodal sub-model corresponds to a maneuver type. S4: Input the maneuvering characteristic parameters of the tracked target into the optimized multimodal sub-model to predict the future state of the tracked target.

[0020] It should be noted that by collecting multi-source data of the tracked target and preprocessing the multi-source data to obtain preprocessed data, and extracting the maneuver feature parameters of the tracked target based on the preprocessed data, a classification model is created. The maneuver feature parameters of the tracked target are input into the classification model to obtain the maneuver type of the tracked target. Then, multiple multimodal sub-models are created. An improved particle swarm optimization algorithm is used to introduce dynamic decay of inertial weights and particle mutation mechanism to optimize the parameters of the multimodal sub-models, resulting in optimized multimodal sub-models. Finally, the maneuver feature parameters of the tracked target are input into the optimized multimodal sub-models to predict the future state of the tracked target. This application improves tracking accuracy and tracking robustness by extracting the maneuver feature parameters of the tracked target from multi-source data, classifying the maneuver type of the tracked target based on the maneuver feature parameters, and selecting the corresponding multimodal sub-model based on the maneuver type.

[0021] In one embodiment of this application, the multi-source data includes location data, velocity data, acceleration data, acceleration rate of change data, and the acquisition time of each data point. The resulting dataset is:

[0022] Where n is the total number of targets being tracked; This is the location data of the i-th tracked target; It is the velocity data of the i-th tracked target; It is the acceleration data of the i-th tracked target; The acceleration rate of change data of the i-th tracked target; Let be the acquisition time corresponding to the i-th tracked target.

[0023] It should be noted that collecting data from multiple sources ensures the accuracy of location and status prediction for the tracked target.

[0024] In one embodiment of this application, the preprocessing step includes: S1.1: Calculate the mean and standard deviation of each multi-source subdata set, and based on... The criteria determine whether outliers exist in multi-source data; Specifically, The specific criterion is: determine whether there exists an observation whose absolute value differs from its mean by more than 3 times the standard deviation. (Taking location data as an example), if If so, the observation is determined to be an outlier and removed. S1.2: If outliers exist in the multi-source data, remove the outliers to obtain standard data; S1.3: Based on the acquisition time, the sampling frequency of each multi-source sub-data of the tracking target is adjusted by linear interpolation to achieve spatiotemporal synchronization of each multi-source sub-data and obtain preprocessed data; S1.4: Extract the maneuvering characteristic parameters of the target from the preprocessed data using the sliding window method; the maneuvering characteristic parameters include the magnitude of acceleration change. , frequency of acceleration change rate f The curvature change of the root trajectory of motion . It should be noted that because different sub-data acquisition devices may have different acquisition frequencies, it is difficult to match the acquired multi-source data one by one. Therefore, it is necessary to use linear interpolation to synchronize the data from the acquisition device with a low sampling frequency to a high sampling frequency.

[0025] If the sampling time of sampling device 1 is The sampling time of sampling device 2 is For those who fall and Between Through formula Calculate the X-axis position after synchronization (the same applies to velocity and acceleration dimensions) to obtain the preprocessed dataset for spatiotemporal synchronization. .

[0026] Next, for the preprocessed dataset Maneuver characteristic parameters are extracted using the sliding window method, where the length L of the sliding window is determined based on the sampling frequency of the observed data. Adaptive adjustment, specifically: when When the sampling interval is ≤ 0.1s, the length L of the sliding window is set to 5-8 sampling periods, corresponding to a time window of 0.5-0.8s. At this time, there is enough data in the window to accurately reflect short-term maneuver characteristics. when When the sampling interval is >0.1s, the length L of the sliding window is set to 3-5 sampling periods, corresponding to a time window of 0.3-0.5s, to avoid the lag of the maneuver feature parameters due to an excessively long window.

[0027] Amplitude of acceleration change (unit This is used to reflect the fluctuation intensity of the acceleration of the tracked target; The oscillation frequency f of the rate of change of acceleration (in units) The dominant frequency is extracted as the fluctuation frequency f. The higher the dominant frequency, the more frequent the acceleration changes and the more violent the target maneuver. Change in curvature of the target trajectory (unit ), used to reflect the curvature of the target trajectory (such as linear motion). Turning motion ).

[0028] In one embodiment of this application, the classification model uses an improved SVM algorithm as the core algorithm of the adaptive maneuver classifier; the kernel function selection rules of the improved SVM algorithm include: when and When the target is determined to be of a low-maneuverability type, a linear kernel function is selected. ,in, It is the input feature vector; when and When the target is determined to be of medium maneuverability, a polynomial kernel function is selected. ,in, , It is the scaling factor. It is a constant term. It is the order of the polynomial; when and When the target is determined to be highly maneuverable, a radial basis function kernel is selected. ,in, ; Specifically, the classification model also adopts a dynamic update mechanism for the penalty parameter. The penalty parameter C is used to balance the classification error and the model complexity. The traditional fixed value of C cannot cope with the dynamic changes of the feature parameters. In this application, the classification accuracy of the validation set is recalculated every 10 sampling periods (the corresponding time interval is determined according to the sampling frequency, such as 10Hz sampling corresponds to 1 second). Randomly select 30% of the current observation data as the validation set, and calculate the classifier's accuracy Acc on the validation set. If Acc ≥ 90%, keep the current penalty parameter C unchanged; if Acc < 90%, adjust the penalty parameter C according to the error direction: when the misclassified samples are mainly "low maneuver misclassified as medium maneuver", increase the penalty parameter C by 20% from the current value to strengthen the classification constraint on low maneuver samples; when the misclassified samples are mainly "medium maneuver misclassified as high maneuver", decrease the penalty parameter C by 20% from the current value to avoid the model overfitting high maneuver samples; repeat the adjustment until Acc ≥ 90% to ensure that the classifier always maintains a high accuracy.

[0029] It should be noted that by analyzing the maneuver characteristic parameters, the classification of the tracked target can be determined, thus providing a reliable basis for the subsequent selection of the corresponding multimodal sub-model.

[0030] After extracting the maneuver feature parameters, construct the maneuver feature parameter vector of the tracked target. and the maneuver feature parameter vector The data is input into a classification model, which then outputs the maneuver type of the tracked target through an improved SVM algorithm.

[0031] Since traditional SVM algorithms use fixed kernel functions, they cannot adapt to feature distributions with different maneuver intensities. Therefore, this application improves the adaptability to maneuver features of different intensities by setting corresponding kernel functions for different maneuver features. Low-maneuver type kernel functions can reduce the computational complexity of SVM algorithms while ensuring classification accuracy; medium-maneuver type kernel functions can balance linear fitting and nonlinear adaptation capabilities; high-maneuver type kernel functions fit complex nonlinear features through Gaussian distribution characteristics.

[0032] In one embodiment of this application, the multimodal sub-model includes: Low-maneuverability sub-model; applied to tracking targets of low-maneuverability type, the mean update formula of the low-maneuverability sub-model is: ,in, It is a correction factor. It is a tracking target The mean acceleration estimate at time t. It is a tracking target Acceleration observations at time t. It is the attenuation coefficient of the tracking target. It is the magnitude of the acceleration change; Medium maneuvering sub-model; applied to tracking targets of medium maneuvering type, the medium maneuvering sub-model adopts an improved Jerk model, by replacing the "constant Jerk assumption" of the rate of change of acceleration in the traditional Jerk model with the "piecewise linear Jerk assumption"; High-maneuverability sub-model; applied to tracking targets of high maneuverability, the high-maneuverability sub-model dynamically adjusts weights by introducing an attention mechanism. The ,in, This refers to the attention weights for maneuver features, where max_val is the value from historical observations. The maximum value, It is the magnitude of acceleration change. f It is the frequency of the rate of change of acceleration. It is the change in curvature of the root locus of motion.

[0033] It should be noted that by constructing sub-models adapted to different maneuver types based on the maneuver classification results of the aforementioned steps, "one type of maneuver corresponds to one optimization model" is achieved.

[0034] The low-maneuverability sub-model is designed for low-maneuverability scenarios. Based on the traditional CS model (which assumes non-zero mean exponential correlation acceleration), a dynamic correction term for the mean acceleration is introduced. The improved mean update formula is obtained. The value ranges from 0.1 to 0.3, and is dynamically adjusted according to the low maneuverability intensity. The closer , Take 0.3; The closer to 0, Taking a value of 0.1, the introduction of the correction term can compensate for the lag deviation of the mean acceleration in real time, for example, when the target moves from a constant velocity ( Suddenly enters a slight acceleration ( When ), the correction item ,make Rapidly approaching actual acceleration This reduces mean lag error.

[0035] The medium maneuver sub-model is designed for medium maneuver scenarios. and .

[0036] The high-mobility sub-model is designed for high-mobility scenarios (when...) or (At the time), this invention integrates the non-zero mean acceleration characteristics of the traditional CS model with the high dynamic response characteristics of the Jerk model to construct a novel motion model, and introduces an attention mechanism to dynamically adjust the weights; Attention weights for maneuver features The value range is [0,1]. The larger the value, the higher the current maneuver intensity and the stronger the nonlinearity of the target being tracked. The output weights of the CS and Jerk modules are dynamically adjusted: when the maneuver feature attention weights... At that time, the Jerk module has a better ability to adapt to highly dynamic changes, so setting the weight of the Jerk module is important. CS module weights When the attention weight of the maneuver feature At that time, the CS module has better stability, so set the weight of the CS module. Jer module weights The final acceleration estimate By dynamically allocating weights, optimal estimation can be achieved under different maneuver intensities.

[0037] In one embodiment of this application, replacing the "constant Jerk assumption" of the rate of change of acceleration in the traditional Jerk model with the "piecewise linear Jerk assumption" includes the following steps: S2.1: Based on the fluctuation frequency of the rate of change of acceleration f Divide the sampling time axis into different lengths; S2.2: Within each sampling time axis, the Jerk linear variation curve is fitted using the least squares method to calculate the Jerk variation parameters; the Jerk variation parameters include the curvature change of the root trajectory. And the initial Jerk value; S2.3: Curvature change based on the root trajectory of motion And calculate the fitting error using the initial Jerk value; S2.4: Set an error threshold and determine whether the fitting error is greater than or equal to the error threshold; S2.5: If the fitting error is greater than or equal to the error threshold, the sampling time axis is divided into multiple segments until the fitting error is less than the error threshold.

[0038] It should be noted that the state vector is expanded to ,in Jerk (unit: m / s³), based on the oscillation frequency of the rate of change of acceleration. Determine the segmentation threshold: when At that time, the acceleration changes slowly, so the sampling time axis is divided into segments of 5 sampling periods; when When the acceleration changes moderately, the sampling time axis is divided into segments of 3 sampling periods; when (When acceleration changes rapidly), the sampling time axis is divided into segments of two sampling periods; Let the observed acceleration values ​​over a certain time period be... ,…, (corresponding time) ,…, Jerk time linear change is The acceleration can be obtained by integration. Minimize the error function Solving for the given information yields the following results. and ; The error threshold in this application is set to 0.5 m / s. 3 Therefore, the fitting error ,like Lock the Jerk variation parameters for the current segment; if The current sampling time axis is split into two shorter segments, and the model is refitted until the error meets the requirements, ensuring that the Jerk model can accurately depict the acceleration variation in the medium maneuvering scenario.

[0039] In one embodiment of this application, step S3 includes: S3.1: Determine the parameters to be optimized for each multimodal sub-model. ,in, These are the correction coefficients for the low-maneuverability sub-model. It is the Jerk segment threshold of the medium-speed sub-model. These are the coefficients of the attention weights in the high-mobility submodel; S3.2: Construct the objective function: ; in, It is the root mean square error of the x-axis. It is the root mean square error of the y-axis. It is the root mean square error of the z-axis. It is the root mean square error of the speed. It is the root mean square error of acceleration. , , , , In order are , , , , The corresponding weighting coefficients; S3.3: By improving the particle swarm optimization algorithm, dynamic decay of inertial weight and particle mutation mechanism are introduced to continuously iterate and optimize the parameters to be optimized for each multimodal sub-model until the termination condition is met, and the iteration stops to obtain the optimized multimodal sub-model.

[0040] It should be noted that the correction coefficients for the low-mobility sub-model... The value range is 0.1-0.3, which is the Jerk segment threshold for the medium-speed sub-model. according to Determine the critical frequency for the segmented period, such as 0.5Hz or 1Hz, with a value range of 0.3-1.2Hz. They are The values ​​0.4, 0.3, and 0.3 in the calculation formula satisfy the following conditions: ,and , , .

[0041] The objective function aims to minimize the tracking error, specifically the root mean square error. ,in, For the observed values, These are the model's predicted values. , , , , Adjustments should be made based on the actual scenario. For example, in missile guidance scenarios, where positional accuracy is the highest requirement, the settings should be adjusted accordingly. , , In traffic tracking scenarios, speed accuracy is even more important, so setting... , , , .

[0042] In one embodiment of this application, step S3.3 includes the following steps: S3.31: Set the iteration termination condition; the iteration termination condition includes: When the objective function corresponding satisfy , , , , Stop iteration when the time is right and output the current optimal parameters. ; The iteration stops when the number of iterations reaches the iteration threshold, and the current optimal parameters are output. ; Specifically, the iteration threshold set in this application is 50 times; S3.32: Set the initial particle swarm size and initialize the initial position and initial velocity of each particle; each particle corresponds to a set of parameters to be optimized; Specifically, the initial particle swarm size is set to 50 in this application; S3.33: For each particle, iterate the particle's position and velocity based on the individual optimal solution and the global optimal solution, set the stage inertia weight according to the number of iterations, and mutate the tail particles during the iteration process until any iteration termination condition is met, then stop the iteration and output the optimal parameters obtained from the last iteration. The optimized multimodal sub-model was obtained.

[0043] It should be noted that, since the traditional PSO algorithm is prone to getting trapped in local optima, this invention introduces dynamic decay of inertial weights and particle mutation mechanisms to optimize the PSO algorithm.

[0044] In this application, the 10% of particles with the worst fitness value, i.e., the objective function value, are denoted as tail particles. Mutation refers to randomly selecting a parameter from the parameter set and reassigning it randomly within its value range. The reason for setting up the mutation mechanism is to avoid the particle swarm getting stuck in a local optimum.

[0045] In one embodiment of this application, the stage inertia weight changes according to the number of iterations and decreases continuously as the number of iterations increases.

[0046] It should be noted that the iteration extremes are defined based on the number of iterations. In the early stages of iterative optimization, the number of iterations is between 1 and 10, and the inertia weight is used. Set the value to 0.8-1.0 to enhance global search capabilities and avoid getting trapped in local optima; during the mid-stage of iterative optimization, the number of iterations is between 11 and 20. The value decreases by 0.1 every 5 iterations, gradually decreasing from 0.8 to 0.3; in the later stages of iterative optimization, i.e., when the number of iterations exceeds 21, Maintaining a value between 0.2 and 0.3 enhances local convergence and accurately approximates the optimal solution. In one embodiment of this application, a fully adaptive maneuvering target tracking device is also provided, applied to the fully adaptive maneuvering target tracking method described in any one of the foregoing. The device includes a data acquisition end and a tracking end. The data acquisition end acquires multi-source data of the target being tracked. The tracking end is communicatively connected to the data acquisition end. The tracking end completes the tracking of the target based on the multi-source data of the target being tracked.

[0047] It should be noted that this application collects multi-source data of the tracked target through the acquisition end and inputs the multi-source data into the tracking end. The tracking end determines the maneuver type of the tracked target based on the multi-source data and selects the corresponding multimodal sub-model based on the maneuver type, thereby effectively improving the tracking accuracy.

[0048] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fully adaptive maneuvering target tracking method, characterized in that, include: S1: Collect multi-source data of the tracked target, preprocess the multi-source data to obtain preprocessed data, and extract the maneuver feature parameters of the tracked target based on the preprocessed data; S2: Create a classification model by inputting the maneuver characteristic parameters of the tracked target into the classification model to obtain the maneuver type of the tracked target; the maneuver type of the tracked target includes low maneuver type, medium maneuver type and high maneuver type; S3: Create multiple multimodal sub-models, and introduce dynamic decay of inertial weights and particle mutation mechanism through an improved particle swarm optimization algorithm to optimize the parameters of the multimodal sub-models, resulting in optimized multimodal sub-models; each multimodal sub-model corresponds to a maneuver type. S4: Input the maneuvering characteristic parameters of the tracked target into the optimized multimodal sub-model to predict the future state of the tracked target.

2. The fully adaptive maneuvering target tracking method according to claim 1, characterized in that: The multi-source data includes location data, velocity data, acceleration data, acceleration rate of change data, and the acquisition time for each data point. The resulting dataset is: Where n is the total number of targets being tracked; This is the location data of the i-th tracked target; It is the velocity data of the i-th tracked target; It is the acceleration data of the i-th tracked target; The acceleration rate of change data of the i-th tracked target; Let be the acquisition time corresponding to the i-th tracked target.

3. The fully adaptive maneuvering target tracking method according to claim 2, characterized in that: The preprocessing includes the following steps: S1.1: Calculate the mean and standard deviation of each multi-source subdata set, and based on... The criteria determine whether outliers exist in multi-source data; S1.2: If outliers exist in the multi-source data, remove the outliers to obtain standard data; S1.3: Based on the acquisition time, the sampling frequency of each multi-source sub-data of the tracking target is adjusted by linear interpolation to achieve spatiotemporal synchronization of each multi-source sub-data and obtain preprocessed data; S1.4: Extract the maneuvering characteristic parameters of the target from the preprocessed data using the sliding window method; the maneuvering characteristic parameters include the magnitude of acceleration change. , frequency of acceleration change rate f The curvature change of the root trajectory of motion .

4. The fully adaptive maneuvering target tracking method according to claim 3, characterized in that: The classification model uses the improved SVM algorithm as the core algorithm of the adaptive maneuver classifier. The kernel function selection rules for the improved SVM algorithm include: when and When the target is determined to be of a low-maneuverability type, a linear kernel function is selected. ,in, It is the input feature vector; when and When the target is determined to be of medium maneuverability, a polynomial kernel function is selected. ,in, , It is the scaling factor. It is a constant term. It is the order of the polynomial; when and When the target is determined to be highly maneuverable, a radial basis function kernel is selected. ,in, .

5. The fully adaptive maneuvering target tracking method according to claim 4, characterized in that: The multimodal sub-model includes: Low-maneuverability sub-model; applied to tracking targets of low-maneuverability type, the mean update formula of the low-maneuverability sub-model is: ,in, It is a correction factor. It is a tracking target The mean acceleration estimate at time t. It is a tracking target Acceleration observations at time t. It is the attenuation coefficient of the tracking target. It is the magnitude of the acceleration change; Medium maneuvering sub-model; applied to tracking targets of medium maneuvering type, the medium maneuvering sub-model adopts an improved Jerk model, by replacing the "constant Jerk assumption" of the rate of change of acceleration in the traditional Jerk model with the "piecewise linear Jerk assumption"; High-maneuverability sub-model; applied to tracking targets of high maneuverability, the high-maneuverability sub-model dynamically adjusts weights by introducing an attention mechanism. The Where max_val is the value in historical observations. The maximum value, It is the magnitude of acceleration change. f It is the frequency of the rate of change of acceleration. It is the change in curvature of the root locus of motion.

6. The fully adaptive maneuvering target tracking method according to claim 5, characterized in that: The replacement of the "constant Jerk assumption" of the rate of change of acceleration in the traditional Jerk model with the "piecewise linear Jerk assumption" includes the following steps: S2.1: Based on the fluctuation frequency of the rate of change of acceleration f Divide the sampling time axis into different lengths; S2.2: Within each sampling time axis, the Jerk linear variation curve is fitted using the least squares method to calculate the Jerk variation parameters; the Jerk variation parameters include the curvature change of the root trajectory. And the initial Jerk value; S2.3: Curvature change based on the root trajectory of motion And calculate the fitting error using the initial Jerk value; S2.4: Set an error threshold and determine whether the fitting error is greater than or equal to the error threshold; S2.5: If the fitting error is greater than or equal to the error threshold, the sampling time axis is divided into multiple segments until the fitting error is less than the error threshold.

7. The fully adaptive maneuvering target tracking method according to claim 6, characterized in that: Step S3 includes the following steps: S3.1: Determine the parameters to be optimized for each multimodal sub-model. ,in, These are the correction coefficients for the low-maneuverability sub-model. It is the Jerk segment threshold of the medium-speed sub-model. These are the coefficients of the attention weights in the high-mobility submodel; S3.2: Construct the objective function: ； in, It is the root mean square error of the x-axis. It is the root mean square error of the y-axis. It is the root mean square error of the z-axis. It is the root mean square error of the speed. It is the root mean square error of acceleration. , , , , In order are , , , , The corresponding weighting coefficients; S3.3: By improving the particle swarm optimization algorithm, dynamic decay of inertial weight and particle mutation mechanism are introduced to continuously iterate and optimize the parameters to be optimized for each multimodal sub-model until the termination condition is met, and the iteration stops to obtain the optimized multimodal sub-model.

8. The fully adaptive maneuvering target tracking method according to claim 7, characterized in that: Step S3.3 includes the following steps: S3.31: Set the iteration termination condition; the iteration termination condition includes: When the objective function corresponding satisfy , , , , Stop iteration when the time is right and output the current optimal parameters. ; The iteration stops when the number of iterations reaches the iteration threshold, and the current optimal parameters are output. ; S3.32: Set the initial particle swarm size and initialize the initial position and initial velocity of each particle; each particle corresponds to a set of parameters to be optimized; S3.33: For each particle, iterate the particle's position and velocity based on the individual optimal solution and the global optimal solution, set the stage inertia weight according to the number of iterations, and mutate the tail particles during the iteration process until any iteration termination condition is met, then stop the iteration and output the optimal parameters obtained from the last iteration. The optimized multimodal sub-model was obtained.

9. The fully adaptive maneuvering target tracking method according to claim 8, characterized in that: The stage inertia weight changes depending on the number of iterations and decreases continuously as the number of iterations increases.

10. A fully adaptive maneuvering target tracking system, applied to the method of any one of claims 1 to 9, characterized in that, include: Data acquisition end; The acquisition terminal collects multi-source data of the target being tracked. Tracking end; The tracking end is communicatively connected to the acquisition end, and the tracking end completes the tracking of the target based on multi-source data of the target.

Images