A method for long-term target trajectory prediction based on interacting multiple models

By building a uniform speed, uniform acceleration and co-turning model, combining the traceless Kalman filter and likelihood function update, dynamically adjusting the transfer probability matrix, the detection lag and model reliability problems in the prediction of flight target trajectory are solved, and the prediction accuracy is improved and adaptability.

CN118520374BActive Publication Date: 2024-11-08SICHUAN JIUZHOU ELECTRIC GROUP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410993259.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-08
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

The prior art has detection lag problems and reduced reliability caused by rapid model switching in the prediction of flight target trajectory, as well as reduced prediction performance caused by incomplete or excessive model sets.

Method used

Using the target long-term trajectory prediction method based on interactive multi-models, the motion model set is constructed, including uniform speed, uniform acceleration and co-turning models, filtering and updating through the traceless Kalman filter, and dynamically update the transfer probability matrix to adaptively adjust the motion model. , and update the motion state probability with the likelihood function to realize the weighted merge of the model and long-term trajectory prediction.

Benefits of technology

It effectively solves the problems of large prediction errors of a single model and low matching of multiple models, improves the accuracy and reliability of flight target trajectory prediction, and is suitable for multiple trajectory prediction scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118520374B_ABST
    Figure CN118520374B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting a target long-term trajectory based on an interactive multi-model, and relates to the technical field of trajectory prediction. The present invention first constructs a motion model set covering the possible motion states of a flight target, takes the input interaction of the previous moment as the input of the model, and then performs model condition filtering and state prediction through an unscented Kalman filter, and then updates the probability of each motion model, and finally outputs multiple model interaction fusion values, and performs adaptive adjustment of the transfer probability matrix value and accuracy evaluation. The present invention adopts an IMM algorithm, designs a suitable model set and filter, can effectively improve the prediction accuracy of the flight target trajectory when there is observation data, and can perform long-term trajectory prediction when there is no observation data that meets the requirements.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of trajectory prediction, and in particular to a target long-term trajectory prediction method based on interactive multi-models. Background Art

[0002] Flight target trajectory prediction is the process of calculating and estimating the future trajectory state of the target based on the state information of the target at the current moment. It can be equivalently regarded as a prediction problem for time series. This prediction technology is a key technology for tasks such as flight potential conflict warning, anomaly detection, autonomous interval maintenance, and aerial target search. The prediction model of flight target trajectory is generally divided into dynamic model, machine learning model and state estimation model. Since the state estimation model only establishes the motion trajectory equation based on the kinematic law parameters such as the position, velocity and acceleration of the flight target, it is relatively simple and effective to implement, so it is widely used in flight target trajectory prediction.

[0003] In the prediction process, the state estimation model can be divided into two categories: single model state estimation and multi-model state estimation, depending on the different assumptions of whether the flying target has a single flight mode or multiple modes. Although single models have been widely used, a single model cannot estimate a mixed system with different modes well, and the trajectory prediction problem is usually a random linear mixed system estimation problem, which requires a multi-model method to solve. For the multiple flight modes of the flying target, multi-model state estimation can improve the overall performance of the model by using a state estimator that is adapted to the flying target motion model. Among the existing multi-model methods, the interacting multiple model (IMM) algorithm has been successfully applied in target trajectory prediction because of its easy implementation and low computational cost.

[0004] In traditional target trajectory prediction algorithms, different motion models are usually used at different motion stages, and the corresponding model filters detect and switch the target state based on statistical tests. Although this algorithm can also adapt to target maneuvering motion, target maneuvering detection often has a lag, and the rapid switching of models will reduce the reliability of the model.

[0005] Although the IMM algorithm makes full use of prior information and can determine the motion state of the target through historical information and observation data, the algorithm also has certain limitations. Its effect is related to the number and matching degree of the model set. When the model is complete and the matching degree is high, a good prediction effect can be obtained. Some existing methods do not fully consider the diversity of motion models. They often use a single motion model or multiple motion models, but do not fully cover the possible motion states of the flying target. For example, some algorithms only use two models, uniform speed and uniform acceleration; some methods use too many models, resulting in an increase in the amount of calculation, and if the models do not match, it is easy to cause competition between the models, thereby reducing the prediction performance.

[0006] In addition, the existing IMM-based prediction methods lack an evaluation method for the accuracy of predicted positions in some application tasks, while the long-term motion trajectory prediction of flying targets requires the evaluation of the accuracy of the prediction results. Therefore, it is necessary to design an evaluation method for quantitative evaluation. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for long-term target trajectory prediction based on interactive multi-models, which mainly solves the following two problems: (1) solving the detection lag problem caused by using a single model for prediction at different motion stages, and the problem of reduced prediction reliability caused by rapid switching of different motion models; (2) solving the problem of low model matching due to incomplete number of model sets in the IMM algorithm, or reducing prediction performance due to too many models.

[0008] The objective of the present invention is achieved through the following technical solutions:

[0009] A method for predicting target long-term trajectory based on interacting multiple models, comprising:

[0010] Step 1: Construct a motion model set, which includes a uniform speed model, a uniform acceleration model, and a coordinated turning model;

[0011] Step 2: According to k The target state estimate at time -1 and the probability of each model filter are used to calculate the mixed state estimate and mixed covariance, which are used as the initial state values ​​of each motion model;

[0012] Step 3: The mixed state estimation, mixed covariance and observation value are used as filter inputs, and each sub-model is filtered and updated using UKF to obtain the updated state and covariance;

[0013] Step 4: Use the likelihood function to update the target k The state of motion at the moment belongs to j Motion Model M j probability;

[0014] Step 5: Adaptively adjust the transfer probability matrix value of each motion model according to different characteristics of the target motion state to dynamically update the motion model;

[0015] Step 6: Determine whether there is any observation data input into the model. If so, weight the estimation results of each filter and combine them to calculate the total state estimate and total covariance estimate. The calculated output is used as the input for the next recursion, and return to step 2 for the next recursion. If not, perform long-term trajectory prediction based on the historical model probability value and the transition probability matrix, and obtain the future long-term trajectory state of the flight target through multiple recursions.

[0016] Furthermore, the step 1 specifically includes:

[0017] According to the target motion characteristics, a motion model set covering the motion state of the flying target is established. M 1 , M 2 ,…, M n ], including a uniform speed model, a uniform acceleration model, and a cooperative turning model, wherein the cooperative turning model includes a left cooperative turning model and a right cooperative turning model. n Indicates the number of models interacting in the motion model set;

[0018] Set up j Motion Model M j The target state equation is expressed as , the observation equation is , Where: represents the state transfer matrix; X j ( k ) indicates that the system is k The state vector at the moment; G j ( k ) is the process noise matrix; W j ( k ) is zero-mean and has a covariance matrix of Q j White noise sequence; H ( k ) is the observation matrix of the system; V ( k ) is zero-mean and has a covariance matrix of R j Gaussian observation noise sequence;

[0019] Determine the state transfer matrix corresponding to the motion model set , measurement matrix and the initial value of the model probability .

[0020] Furthermore, the step 2 specifically includes:

[0021] according to k Target state estimation at time -1 and each model filter probability , calculate the mixed state estimate and mixed covariance As the initial state value of the current cycle, the calculation process of each parameter is as follows:

[0022] Calculated in k -1 moment transferred from other models M j The total predicted probability of i Motion Model M i Transfer to M j The mixing probability is calculated as:

[0023]

[0024] in, is the normalization constant, p ij is the Markov transition probability matrix P , indicating that the target is i The model is transferred to j The probability of a model;

[0025] Calculate the j Motion Model M j The mixed state estimation is calculated as:

[0026]

[0027] Calculate the j Motion Model M j The mixed covariance estimate is calculated as:

[0028]

[0029] in For model M i Covariance at time k-1.

[0030] Furthermore, the step 3 specifically includes:

[0031] The mixed state estimate , mixed covariance And the observed value Z ( k ) as the initial state input of each motion model;

[0032] The unscented Kalman filter is used to estimate the state of each motion model. The steps are as follows:

[0033] Obtain a set of sampling points through unscented transformation , called Sigma point set, and the corresponding mean weight and covariance weights , calculate the one-step prediction value of the system state quantity:

[0034]

[0035] In the formula represents the state transfer matrix;

[0036] The predicted values ​​of the Sigma point set are weighted and averaged to obtain a one-step prediction of the system state quantity. and covariance , the calculation formulas are:

[0037] ,

[0038]

[0039] in , r The state dimension of the model input;

[0040] According to the one-step prediction value, the unscented transformation is used again to construct a new Sigma point set, which is substituted into the system observation equation to obtain the further prediction value of the observation of the Sigma point set:

[0041]

[0042] In the formula is the observation matrix of the system;

[0043] The mean of the system prediction is obtained by weighted summation And covariance:

[0044] ,

[0045] ,

[0046] ,

[0047] In the formula For the j The observed covariance of the model, For the j The covariance of the observed and predicted values ​​of each model;

[0048] The calculated gain matrix is:

[0049]

[0050] The updated state and covariance are:

[0051] ,

[0052] .

[0053] Furthermore, the step 4 specifically includes:

[0054] Use the likelihood function to update the target k The state of motion at the moment belongs to j Motion Model M j Probability :

[0055]

[0056] in, is the likelihood function, exp(.) means calculating the power series with the natural logarithm e as the base; det(.) means calculating the value of the determinant; .

[0057] Furthermore, the step 5 specifically includes:

[0058] For the motion submodel M j , defined in k-1 Moment and k The probabilities at each moment are and , set the adjustment function for: ,in is the probability difference, which indicates the change in the degree of matching with the true motion model at consecutive time points;

[0059] For the transition probability matrix value After adjustment and normalization, we get:

[0060]

[0061] in, and They are k- 1 moment andk The transition probability matrix value at the moment;

[0062] The transition probability matrix P The diagonal elements of p ii Setting the Threshold , and there is ,when When , keep the adjustment result unchanged; when When , the matrix elements are adjusted twice, and the adjustment method is

[0063]

[0064] The updated and adjusted transition probability matrix value is used in the next state estimation calculation.

[0065] Furthermore, the weighted combination of the estimation results of each filter to calculate the total state estimation and the total covariance estimation includes:

[0066] After the parallel calculation of each sub-model filter, its posterior probability is obtained, and then the estimation results of each filter are weighted and combined to calculate the total state estimate:

[0067]

[0068] The total covariance is estimated as:

[0069] .

[0070] Furthermore, the long-term trajectory prediction is performed based on the historical model probability value and the transition probability matrix, and the future long-term trajectory state of the flight target is obtained through multiple recursions, including:

[0071] According to the state transition probability matrix value p ij , the probability of each model filter at k-1 time , calculate the sub-motion model M at time k j The predicted probability for ;

[0072] According to the predicted probability of each sub-model , predicted state value and the predicted covariance , the corresponding state prediction value and prediction covariance are:

[0073] ,

[0074] ;

[0075] The prediction results of the flight state of the flight target at the next moment and the probability values ​​of each sub-model after multiple sampling periods obtained by multiple recursions are used as inputs for subsequent predictions to obtain the future long-term trajectory state of the flight target.

[0076] Furthermore, it also includes:

[0077] Step 7: Evaluate the prediction accuracy of the state estimation of the flight target, specifically: Calculate k Prediction accuracy at the moment , and its calculation formula is

[0078]

[0079] in, Long-term prediction start time The maximum sub-model probability value, is the model number, , For the sub-model k The probability of time.

[0080] The beneficial effects of the present invention are:

[0081] 1) Based on the IMM algorithm, the present invention establishes a set of motion models such as a uniform speed model, a uniform acceleration model, and a coordinated turning model, which can cover typical flight modes of flying targets, such as turning, diving, and serpentine flight modes. The proposed long-term trajectory prediction method of flying targets solves the problems of large errors in the prediction formula of a single motion model and low completeness and matching of multiple motion models.

[0082] 2) The present invention dynamically updates the value of the transfer probability matrix according to the different characteristics of the target motion state, which solves the problem that the transfer probability matrix in the interactive multi-model is usually set to a fixed value according to prior information, resulting in unreasonable parameter settings, reduces the errors introduced by humans, and effectively optimizes the trajectory prediction process.

[0083] 3) The present invention adopts multiple unscented Kalman filters for parallel filtering. This filter technology can effectively solve the problem of nonlinear system tracking. However, the conventional linear Kalman filter cannot effectively handle nonlinear system filtering problems such as coordinated turning.

[0084] 4) The present invention adopts the IMM algorithm and designs a suitable motion model set and filter. When there is observation data, it can effectively improve the prediction accuracy of the flight target trajectory. When there is no observation data that meets the requirements, it can perform long-term trajectory prediction and is applicable to a variety of trajectory prediction scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1A target trajectory prediction process flow chart provided by an embodiment of the present invention;

[0086] Figure 2 A principle block diagram of a target long-term trajectory prediction method based on interactive multiple models provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0088] See also Figure 1-Figure 2 , the present invention provides a technical solution:

[0089] A method for long-term trajectory prediction of targets based on interactive multiple models, the flow chart of which is as follows Figure 1 As shown, including:

[0090] Step 1: Construct a motion model set, which includes a uniform speed model, a uniform acceleration model, and a coordinated turning model;

[0091] Step 2: According to k The target state estimate at time -1 and the probability of each model filter are used to calculate the mixed state estimate and mixed covariance, which are used as the initial state values ​​of each motion model;

[0092] Step 3: The mixed state estimation, mixed covariance and observation value are used as filter inputs, and each sub-model is filtered and updated using UKF to obtain the updated state and covariance;

[0093] Step 4: Use the likelihood function to update the target k The state of motion at the moment belongs to j Motion Model M j probability;

[0094] Step 5: Adaptively adjust the transfer probability matrix value of each motion model according to different characteristics of the target motion state to dynamically update the motion model;

[0095] Step 6: Determine whether there is any observation data input into the model. If so, weight the estimation results of each filter and combine them to calculate the total state estimate and total covariance estimate. The calculated output is used as the input for the next recursion, and return to step 2 for the next recursion. If not, perform long-term trajectory prediction based on the historical model probability value and the transition probability matrix, and obtain the future long-term trajectory state of the flight target through multiple recursions.

[0096] The principle diagram of the target long-term trajectory prediction method provided in a specific embodiment is as follows Figure 2 As shown, specifically, the step 1 specifically includes:

[0097] According to the target motion characteristics, a motion model set [M1, M2,…,M n ], including a uniform speed model, a uniform acceleration model, and a cooperative turning model, wherein the cooperative turning model includes a left cooperative turning model and a right cooperative turning model, and n represents the number of models interacting in the motion model set;

[0098] Usually, the plane motion trajectory of a flying target is composed of straight segments and turning arcs, and there are three motion processes: uniform speed straight flight, uniform acceleration straight flight, and uniform turning rate constant speed turning. Therefore, its motion model can be divided into a constant velocity (CV) model, a constant acceleration (CA) model, and a coordinated turn (CT) model. According to the various complex motion characteristics of the flying target, the present invention establishes a motion model set with a limited number to cover the possible motion states of the flying target, and further subdivides the coordinated turn model into a left CT motion model and a right CT motion model to describe various complex nonlinear motions such as target speed change and turning, and realize switching between multiple motion modes.

[0099] The transfer between multiple motion models is described as a Markov process, and the transfer between each motion model is determined by the Markov transition probability matrix P, which is expressed as follows:

[0100]

[0101] The elements p ij Indicates that the target is i The model is transferred to j The final target state is obtained by weighting the results of multiple motion model filtering in proportion.

[0102] Set up j Motion Model M j The target state equation is expressed as , the observation equation is , where: represents the state transfer matrix; X j ( k ) indicates that the system is k The state vector at the moment; G j (k ) is the process noise matrix; W j ( k ) is zero-mean and has a covariance matrix of Q j White noise sequence; H ( k ) is the observation matrix of the system; V ( k ) is zero-mean and has a covariance matrix of R j Gaussian observation noise sequence;

[0103] Determine the state transfer matrix corresponding to the motion model set , measurement matrix and the initial value of the model probability .

[0104] In some specific embodiments, the step 2 specifically includes:

[0105] according to k Target state estimation at time -1 and each model filter probability , calculate the mixed state estimate and mixed covariance As the initial state value of the current cycle, the calculation process of each parameter is as follows:

[0106] Calculated in k -1 moment transferred from other models M j The total predicted probability of i Motion Model M i Transfer to M j The mixing probability is calculated as:

[0107]

[0108] in, is the normalization constant, p ij is the Markov transition probability matrix P , indicating that the target is i The model is transferred to j The probability of a model;

[0109] Calculate the j Motion Model M j The mixed state estimation is calculated as:

[0110]

[0111] Calculate the j Motion Model M j The mixed covariance estimate is calculated as:

[0112]

[0113] in For model M i exist k Covariance at moment -1.

[0114] In some specific embodiments, the step 3 specifically includes:

[0115] The mixed state estimate , mixed covariance And the observed value Z ( k ) as the initial state input of each motion model;

[0116] The unscented Kalman filter is used to estimate the state of each motion model. The steps are as follows:

[0117] Obtain a set of sampling points through unscented transformation , called Sigma point set, and the corresponding mean weight and covariance weights , calculate the one-step prediction value of the system state quantity:

[0118]

[0119] In the formula represents the state transfer matrix;

[0120] The predicted values ​​of the Sigma point set are weighted and averaged to obtain a one-step prediction of the system state quantity. and covariance , the calculation formulas are:

[0121] ,

[0122] ,

[0123] in , r The state dimension of the model input;

[0124] According to the one-step prediction value, the unscented transformation is used again to construct a new Sigma point set, which is substituted into the system observation equation to obtain the further prediction value of the observation of the Sigma point set:

[0125]

[0126] In the formula is the observation matrix of the system;

[0127] The mean of the system prediction is obtained by weighted summation And covariance:

[0128] ,

[0129] ,

[0130] ,

[0131] In the formula For the j The observed covariance of the model, For the j The covariance of the observed and predicted values ​​of each model;

[0132] The calculated gain matrix is:

[0133]

[0134] The updated state and covariance are:

[0135] ,

[0136] .

[0137] The Unscented Kalman Filter (UKF) algorithm uses unscented transformation to deal with the nonlinear transmission problem of mean and covariance. It approximates the probability density distribution of nonlinear functions and uses a series of determined samples to approximate the posterior probability density of the state, rather than approximating the nonlinear function. Therefore, it has high calculation accuracy for the statistics of nonlinear distribution, can effectively overcome the linearization error of traditional Kalman filters, better handle the nonlinear motion of targets, and is simple to implement and has high accuracy. The unscented Kalman filter used in the present invention has outstanding advantages in terms of calculation amount, tracking accuracy and real-time performance.

[0138] In some specific embodiments, step 4 specifically includes:

[0139] Use the likelihood function to update the target k The state of motion at the moment belongs to j Motion Model M j Probability :

[0140]

[0141] in, is the likelihood function, exp(.) means calculating the power series with the natural logarithm e as the base; det(.) means calculating the value of the determinant; .

[0142] In some specific embodiments, step 5 specifically includes:

[0143] For the motion submodel M j , defined in k-1 Moment and k The probabilities at each moment are and , set the adjustment function for: ,in is the probability difference, which indicates the change in the degree of matching with the true motion model at consecutive time points;

[0144] Adjust and normalize the transition probability matrix values ​​to obtain:

[0145]

[0146] in, and They are k- 1 moment and k The transition probability matrix value at the moment;

[0147] The transition probability matrix P The diagonal elements of p ii Setting the Threshold , and there is ,when When , keep the adjustment result unchanged; when When , the matrix elements are adjusted twice, and the adjustment method is

[0148]

[0149] The diagonal elements p ii The purpose of setting the threshold is to ensure that the main diagonal of the matrix is ​​dominant and to prevent the probability of the model's own transfer from becoming smaller and smaller;

[0150] The updated and adjusted transition probability matrix value is used in the next state estimation calculation.

[0151] The transition probability matrix in the interactive multi-model is usually set to a fixed value based on prior information, but there is a problem of unreasonable parameter setting. Therefore, in order to reduce human-introduced errors and optimize the processing, the present invention dynamically updates the matrix value according to the different characteristics of the target motion state.

[0152] In some specific embodiments, the weighted combination of the estimation results of each filter to calculate the total state estimation and the total covariance estimation includes:

[0153] After the parallel calculation of each sub-model filter, its posterior probability is obtained, and then the estimation results of each filter are weighted and combined to calculate the total state estimate:

[0154]

[0155] The total covariance is estimated as:

[0156] .

[0157] When there is observation data of the flying target, the results of steps 1-5 above can be used to calculate the total state estimate and the total covariance estimate for adaptive target tracking. However, considering the situation when there is no observation data that meets the requirements, the present invention then designs a long-term prediction algorithm for memory tracking.

[0158] From step 2 above, we can see that with a specific model M j Matched initial mixture state estimate and mixed covariance The calculation of is not affected by the observed data. According to the state transfer equation and covariance formula in step 3, the predicted state value can be obtained and the predicted covariance .

[0159] The long-term trajectory prediction is performed based on the historical model probability value and the transition probability matrix, and the future long-term trajectory state of the flight target is obtained through multiple recursions, including:

[0160] According to the state transition probability matrix value p ij , k -1 The probability of each model filter at the moment , calculated k Moment sub-motion model M j The predicted probability for ;

[0161] According to the predicted probability of each sub-model , predicted state value and the predicted covariance , the corresponding state prediction value and prediction covariance are:

[0162] ,

[0163] ;

[0164] The prediction results of the flight state of the flight target at the next moment and the probability values ​​of each sub-model after multiple sampling periods obtained by multiple recursions are used as inputs for subsequent predictions to obtain the future long-term trajectory state of the flight target.

[0165] In addition to using the root mean square error to measure the difference between the model prediction value and the actual observation value, in order to evaluate the accuracy of long-term trajectory prediction, the present invention also provides a new accuracy evaluation method in combination with application scenarios.

[0166] In some specific embodiments, the target long-term trajectory prediction method further includes:

[0167] Step 7: Evaluate the prediction accuracy of the state estimation of the flight target, specifically: Calculate k Prediction accuracy at the moment , the calculation formula is

[0168]

[0169] in, Long-term prediction start time The maximum sub-model probability value, is the model number, , For the sub-model k The probability of time.

[0170] In order to solve the problems of large error in the prediction formula of a single motion model, low completeness and matching of multiple motion models, the present invention proposes a long-term trajectory prediction method for a flight target based on IMM. First, the input interaction of the previous moment is used as the input of the model, and then the model condition filtering and state prediction are performed through UKF, and then the probability of each motion model is updated, and finally multiple model interaction fusion values ​​are output, and the transfer probability matrix value is adaptively adjusted and the accuracy is evaluated. The present invention adopts the IMM algorithm, designs a suitable model set and filter, can effectively improve the prediction accuracy of the flight target trajectory when there is observation data, and can perform long-term trajectory prediction when there is no observation data that meets the requirements.

[0171] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for long-term target trajectory prediction based on interactive multiple models, characterized in that: include: Step 1: Construct a motion model set, which includes a uniform speed model, a uniform acceleration model, and a coordinated turning model; Step 2: According to k The target state estimate at time -1 and the probability of each model filter are used to calculate the mixed state estimate and mixed covariance, which are used as the initial state values ​​of each motion model; Step 3: The mixed state estimation, mixed covariance and observation value are used as filter inputs, and each sub-model is filtered and updated using UKF to obtain the updated state and covariance; Step 4: Use the likelihood function to update the target k The state of motion at the moment belongs to j Motion Model M j probability; Step 5: Adaptively adjust the transfer probability matrix value of each motion model according to different characteristics of the target motion state to dynamically update the motion model; Step 6: Determine whether there is observation data input into the model. If so, weight the estimation results of each filter, calculate the total state estimation and total covariance estimation, and use the calculated output as the input of the next recursion, and return to step 2 for the next recursion; if not, perform long-term trajectory prediction based on the historical model probability value and the transition probability matrix, and obtain the future long-term trajectory state of the flight target through multiple recursions; The step 1 specifically includes: According to the target motion characteristics, a motion model set covering the motion state of the flying target is established. M 1 , M 2 ,…, M n ], including a uniform speed model, a uniform acceleration model, and a cooperative turning model, wherein the cooperative turning model includes a left cooperative turning model and a right cooperative turning model. n Indicates the number of models interacting in the motion model set; Set up j Motion Model M j The target state equation is expressed as , the observation equation is , Where: represents the state transfer matrix; X j ( k ) indicates that the system is k The state vector at the moment; G j ( k ) is the process noise matrix; W j ( k ) is zero-mean and has a covariance matrix of Q j White noise sequence; H ( k ) is the observation matrix of the system; V ( k ) is zero-mean and has a covariance matrix of R j Gaussian observation noise sequence; Determine the state transfer matrix corresponding to the motion model set , measurement matrix and the initial value of the model probability ; The step 2 specifically includes: according to k Target state estimation at time -1 and each model filter probability , calculate the mixed state estimate and mixed covariance As the initial state value of the current cycle, the calculation process of each parameter is as follows: Calculated in k -1 moment transferred from other models M j The total predicted probability of i Motion Model M i Transfer to M j The mixing probability is calculated as: in, is the normalization constant, p ij is an element of the Markov transition probability matrix P, indicating that the target is i The model is transferred to j The probability of a model; Calculate the j A motion model M j The mixed state estimation is calculated as: Calculate the j Motion Model M j The mixed covariance estimate is calculated as: in For Model M i exist k -1 moment covariance; The step 3 specifically includes: The mixed state estimate , mixed covariance And the observed value Z ( k ) as the initial state input of each motion model; The unscented Kalman filter is used to estimate the state of each motion model. The steps are as follows: Obtain a set of sampling points through unscented transformation , called Sigma point set, and the corresponding mean weight and covariance weights , calculate the one-step prediction value of the system state quantity: In the formula represents the state transfer matrix; The predicted values ​​of the Sigma point set are weighted and averaged to obtain a one-step prediction of the system state quantity. and covariance , the calculation formulas are: , in , r The state dimension of the model input; According to the one-step prediction value, the unscented transformation is used again to construct a new Sigma point set, which is substituted into the system observation equation to obtain the further observation prediction value of the Sigma point set: In the formula is the observation matrix of the system; The mean of the system prediction is obtained by weighted summation And covariance: , , , In the formula For the j The observed covariance of the model, For the j The covariance of the observed and predicted values ​​of each model; The calculated gain matrix is: The updated state and covariance are: , ; The step 4 specifically includes: Use the likelihood function to update the target k The state of motion at the moment belongs to j Motion Model M j Probability : in, is the likelihood function, exp(.) means calculating the power series with the natural logarithm e as the base; det(.) means calculating the value of the determinant; ; The step 5 specifically includes: For the motion submodel M j , defined in k-1 Moment and k The probabilities at each moment are and , set the adjustment function for: ,in is the probability difference, which indicates the change in the degree of matching with the true motion model at consecutive time points; Adjust and normalize the transition probability matrix values ​​to obtain: in, and They are k- 1 moment and k The transition probability matrix value at the moment; The transition probability matrix P The diagonal elements of p ii Setting the Threshold , and there is ,when When , keep the adjustment result unchanged; when When , the matrix elements are adjusted twice, and the adjustment method is The updated and adjusted transition probability matrix value is used in the next state estimation calculation; The long-term trajectory prediction is performed based on the historical model probability value and the transition probability matrix, and the future long-term trajectory state of the flight target is obtained through multiple recursions, including: According to the state transition probability matrix value p ij , k -1 The probability of each model filter at the moment , calculated k Moment Motion Model M j The predicted probability for ; According to the predicted probability of each sub-model , predicted state value and the predicted covariance , the corresponding state prediction value and prediction covariance are: , ; The prediction results of the flight state of the flight target at the next moment and the probability values ​​of each sub-model after multiple sampling periods obtained by multiple recursions are used as inputs for subsequent predictions to obtain the future long-term trajectory state of the flight target.

2. The method for predicting target long-term trajectory based on interacting multiple models according to claim 1, characterized in that: The weighted combination of the estimation results of each filter to calculate the total state estimation and the total covariance estimation includes: After the parallel calculation of each sub-model filter, its posterior probability is obtained, and then the estimation results of each filter are weighted and combined to calculate the total state estimate: The total covariance is estimated as: 。 3. The method for long-term target trajectory prediction based on interacting multiple models according to claim 1, characterized in that: Also includes: Step 7: Evaluate the prediction accuracy of the state estimation of the flight target, specifically: Calculate k Prediction accuracy at the moment , and its calculation formula is in, Long-term prediction start time The maximum sub-model probability value, is the model number, , For the sub-model k The probability of time.