Target trajectory prediction method based on maneuvering intention inference
By combining state estimation with inequality constraints and sequential likelihood ratio testing with a maneuver model, the problems of weak noise resistance and low efficiency in traditional maneuver intent recognition and trajectory prediction methods are solved, achieving high-precision and fast target trajectory prediction.
Patent Information
- Application Number
- CN202511582905.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-02-06
AI Technical Summary
Existing methods for identifying maneuvering intent have weak noise resistance, and trajectory prediction methods are inefficient and have limited accuracy. Traditional methods are less efficient and less accurate under complex maneuvering intents, making it difficult to meet the application requirements of modern military and civilian fields.
An intention inference method combining state estimation with inequality constraints and sequential likelihood ratio test is adopted. The tracking model set and prediction model set are separated. The identified maneuver model is used for trajectory prediction, and the target trajectory is predicted by combining the intention inference results, thereby improving noise resistance and prediction accuracy.
It achieves higher accuracy and faster convergence trajectory prediction under complex maneuvering intentions, improves the noise resistance of intention inference, and improves the problem of decreased prediction accuracy caused by too many mismatched models in interactive multi-model.
Smart Images
Figure CN121480943A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of target tracking, maneuver intention inference and trajectory prediction, and in particular to a target trajectory prediction method based on maneuver intention inference. BACKGROUND
[0002] For a target monitoring system, target tracking is a very important supporting technology, and accurate target tracking is the premise of accurately identifying the maneuver intention of the enemy target and conducting trajectory prediction, which involves key research fields including modeling of maneuvering targets, selection of filters and multi-model estimation methods.
[0003] In the military field, the maneuver intention of air combat refers to the flight maneuvers performed by fighter aircraft to achieve their tactical objectives, such as snake maneuver, spiral maneuver, half-roll inversion, etc. Traditional air combat decision-making relies on the combat experience and intuitive judgment of pilots, which is not accurate and rapid enough under modern high-tech warfare conditions. By analyzing the enemy flight data and real-time sensor information for target tracking, the maneuvering actions of the target flight can be identified to determine the tactical intention of the enemy, and based on this, the trajectory prediction of the enemy aircraft can provide important reference information for air battlefield situation awareness and command planning, which is of great significance for the improvement of the air combat victory rate of the own side.
[0004] In the civil field, trajectory prediction based on intention inference also plays an important role: in air traffic control, it helps to monitor the flight trajectory of the target, helps to ensure the safety distance and thus reduces the risk of air conflict; in the automatic driving system, clearly identifying the flight intention of other targets in the current airspace and predicting the heading can help the system better understand the environmental changes and thus make better decisions; in aircraft safety monitoring, by identifying the intention and behavior of the target, abnormal situations can be detected early, and preventive measures can be taken in time to reduce the accident rate.
[0005] Traditional maneuver intention recognition methods have the disadvantages of weak noise resistance, etc.: time series feature recognition cannot avoid the problems of trajectory segmentation and threshold selection, and due to the dependence on real trajectory information, the noise resistance is poor. However, in actual application, the real trajectory information is limited, and the accuracy of feature calculation is also affected by the precision limitation of the sensor, which affects the accuracy and efficiency of this method. Machine learning and neural network-based methods require a large amount of training materials, and the requirements for data are more stringent, and the lack of training data limits the effectiveness of this method.
[0006] Traditional trajectory prediction has the disadvantages of low prediction efficiency and limited prediction accuracy: it is usually based on tracking models or combined with statistical features of target-related states, and in order to improve the prediction accuracy, a multi-model method is often used. However, in the face of complex maneuver intention, due to the direct use of tracking information, the model set is often too complex, resulting in low calculation efficiency and decreased accuracy, which is difficult to meet the application requirements.
[0007] Therefore, it is necessary to propose a new trajectory prediction method based on intention inference on the basis of traditional maneuver intention recognition and trajectory prediction method, give enough information for decision-making and game, which has high application value in modern military field and civil field. SUMMARY
[0008] In order to solve the problems of weak anti-noise ability of the existing intention inference method, low efficiency and limited precision of the trajectory prediction method, the intention inference method with inequality constraint combined with sequential likelihood ratio test is provided, the anti-noise ability of the maneuver intention inference method is improved, the target trajectory prediction is carried out based on the maneuver intention inference, the tracking model set and the prediction model set are separated, the identified maneuver model is used instead of the interactive multiple model, the trajectory prediction is carried out combined with the intention inference result, the prediction precision is higher than that of the trajectory prediction method based on the multiple model, and higher precision and faster convergence of the trajectory prediction for the targets with different maneuver intentions are realized.
[0009] The present application is implemented by the following technical solutions.
[0010] In one aspect of the present application, a target trajectory prediction method based on maneuver intention inference is provided, comprising: Receiving the measurement information of the target tracked by the radar at the initial time, giving the target state information at the initial time, selecting the tracking model set and the prediction model set of the interactive multiple model method, obtaining the threshold matrix of the sequential likelihood ratio test algorithm corresponding to the inequality constraint of different maneuver intention assumptions; Tracking the target by using the interactive multiple model method, obtaining the target state estimation at the current time according to the target state information at the previous time and the measurement information at the current time; According to the target state information at the previous time and the measurement information at the current time, the measurement likelihood of the target under different maneuver intention assumptions is calculated by using the projection method; According to the measurement likelihood value, the matrix likelihood ratio statistic is calculated, and the target maneuver intention is identified by using the matrix sequential likelihood ratio test algorithm; According to the target maneuver intention at the current time, the trajectory prediction is carried out by using the maneuver model corresponding to the target maneuver intention, and the target trajectory prediction result is obtained.
[0011] As preferred, the measurement information of the target tracked by the radar at the initial time is received, the target state information at the initial time is given, the model set and the prediction model set of the interactive multiple model method are selected, the threshold matrix of the sequential likelihood ratio test algorithm corresponding to the inequality constraint of different maneuver intention assumptions is obtained, comprising: Receiving the radar measurement information of the target tracked at the initial time, giving the initial process noise covariance and the initial measurement noise covariance; Given the state information of the target at the initial time, the initial state vector of the target and the state estimation error covariance matrix are obtained; According to the prior information, a tracking model set is selected as the tracking model set of the interactive multiple model method, and the initial model probability and the transition probability matrix are given; According to the prior information, a maneuver intention hypothesis is set, and the threshold matrix of the sequential likelihood ratio test algorithm corresponding to different maneuver intention hypotheses is given.
[0012] Preferably, the radar measurement information includes initial distance, initial azimuth angle and initial elevation angle; and the state information includes position, velocity, angular velocity and acceleration.
[0013] Preferably, the target state estimation at the current time is obtained, including: According to the transition probability matrix, the state estimation and the state estimation error covariance matrix of the target in the input interactive model are calculated; The volume Kalman filter is performed on each tracking model in the tracking model set to obtain the state estimation and the state estimation error covariance matrix of each model at the current time, and the measurement prediction error vector and the innovation covariance matrix; The likelihood function and the model probability of each model are calculated according to the measurement prediction error vector and the innovation covariance matrix; The fused state estimation and the state estimation error covariance matrix are calculated according to the updated model probability.
[0014] Preferably, the volume Kalman filter is performed on each tracking model, including: According to the third-order volume rule, the volume points and the weights corresponding to the state vector estimation at the previous time are calculated; According to the state equation, the volume points are nonlinearly transformed to obtain the state prediction and the state prediction error covariance matrix; According to the volume rule, the volume points and the weights corresponding to the state prediction vector are calculated; According to the measurement equation, the volume points are nonlinearly transformed to obtain the volume point measurement prediction, the innovation covariance matrix and the mutual covariance matrix of the state and the measurement; The state estimation and the state estimation error covariance matrix are calculated according to the innovation covariance matrix and the mutual covariance matrix of the state and the measurement.
[0015] Preferably, the projection method is used to calculate the measurement likelihood of the target under different maneuver intention hypotheses, including: According to the third-order volume rule, the volume points and the weights corresponding to the target state estimation at the previous time are calculated; For each maneuver intention hypothesis, the projection of the volume points not in the constraint region under different hypotheses to the constraint region is calculated; According to the volume point projection, the measurement prediction and innovation covariance matrix under different assumptions are calculated; According to the measurement prediction and innovation covariance matrix, the likelihood function of the maneuver assumption of the target is calculated.
[0016] As preferred, the matrix likelihood ratio statistic is calculated, and the target maneuver intention is identified by using the matrix sequential likelihood ratio test algorithm, including: The likelihood function value of each maneuver assumption of the target at the current time is received, and the log-likelihood ratio statistic is calculated; The likelihood function values at the current time are multiplied and normalized to obtain the probability of each maneuver assumption of the target being identified at the current time; The target maneuver intention is obtained by comparing the log-likelihood ratio statistic with a threshold matrix: If the likelihood ratio statistic of a certain assumption is greater than the set threshold value, it is considered that the target performs the maneuver corresponding to the assumption at the current time; if the above condition is not met, the target maneuver at the current time cannot be determined.
[0017] As preferred, according to the target maneuver intention at the current time, the maneuver model corresponding to the target maneuver intention is used to perform trajectory prediction to obtain the target trajectory prediction result, including: The intention inference result at the current time is received, and the model for trajectory prediction is selected: if the target maneuver at the current time cannot be determined, the interacting multiple model algorithm is selected to perform trajectory prediction; if the target maneuver at the current time can be determined, a certain assumption is accepted, and the maneuver model corresponding to the assumption is selected to perform trajectory prediction; The selected prediction model is received, and fixed-step trajectory prediction is performed under the model to calculate the state prediction and state prediction error covariance matrix of the target.
[0018] As preferred, if the target maneuver at the current time cannot be determined, the interacting multiple model algorithm is selected to perform trajectory prediction, including: According to the prior information, a prediction model set is selected as the model set of the interacting multiple model algorithm, the initial model probability and transition probability matrix are given, and the model probability of the target after input interaction is calculated; The model state estimation and state estimation error covariance matrix of the target after input interaction are calculated; According to the third-order volume rule, the volume point and weight corresponding to the state vector estimation value at the current time are calculated; The volume point is nonlinearly transformed according to the state equation of the model to obtain the state prediction and state prediction error covariance matrix; The state prediction and state prediction covariance of each model are weighted and fused to calculate the state prediction vector and state prediction error covariance matrix based on the interacting multiple model algorithm, and the target trajectory prediction result is obtained.
[0019] As preferred, if the target maneuver at the current time can be judged, a certain assumption is accepted, a maneuver model corresponding to the assumption is selected to perform trajectory prediction, including: The volume point and weight value corresponding to the state estimation at the current time are calculated according to the third-order volume rule; The volume point is projected and transformed according to the selected prediction model; The volume point projection is nonlinearly transformed according to the state equation; The state prediction vector and the state prediction error covariance matrix of the set step length are calculated to obtain the target trajectory prediction result.
[0020] The present application has the following beneficial effects due to the above technical scheme: The present application provides a high-efficiency and accurate target trajectory prediction method based on maneuver intention inference, effectively solves the problems of weak anti-noise ability of the existing intention inference method, low efficiency and limited precision of the trajectory prediction method, etc. On the one hand, in the case that different maneuver intentions of the target have the same kinematic model, the sequential likelihood ratio test recognition technology based on inequality constraints is used, the state estimation method with inequality constraints based on the volume rule is combined with the matrix sequential likelihood ratio test algorithm, it is ensured that the volume point projection is located in the constraint region and retains its sampling characteristics, the anti-noise ability of the intention inference is improved, and efficient intention inference is realized for the above-mentioned case. On the other hand, a new trajectory prediction method based on intention inference is proposed, the target trajectory is predicted combined with the intention inference result, the trajectory is predicted using the maneuver model corresponding to the recognized maneuver intention, the problem of prediction precision decline caused by too many unmatched models in the interactive multiple model is effectively improved, and faster model convergence is realized. BRIEF DESCRIPTION OF DRAWINGS
[0021] The drawings described herein are used to provide further understanding of the present application, constitute a part of the present application, and do not constitute an improper limitation on the present application, and in the drawings: Figure 1 It is a flow chart of the trajectory prediction algorithm based on maneuver intention inference; Figure 2 It is a schematic diagram of simulation scenario 1; Figures 3(a)-3(e) It is a result diagram of target tracking, intention inference and trajectory prediction under simulation scenario 1; Among them Figures 3(a)-3(b) It is the root mean square error of position and velocity of target tracking using the interactive multiple model algorithm, Fig. 3(c) is the inference probability of each maneuver intention, Figures 3(d)-3(e) It is a comparison diagram of the root mean square error of target position and velocity of trajectory prediction based on maneuver intention inference and trajectory prediction using only the interactive multiple model algorithm.
[0022] Figure 4Fig. 2 is a schematic diagram of simulation scenario 2; Figures 5(a)-5(e) Fig. 3 is a result diagram of target tracking, intention inference and trajectory prediction under simulation scenario 2; wherein Figures 5(a)-5(b) Fig. 4 is a root mean square error diagram of position and velocity of target tracking using an interacting multiple model algorithm, Fig. 5(c) is an inference probability of each maneuvering intention, Figures 5(d)-5(e) Fig. 6 is a root mean square error diagram of target position and velocity of trajectory prediction based on maneuvering intention inference and trajectory prediction using only an interacting multiple model algorithm. DETAILED DESCRIPTION
[0023] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments, which are used to explain the present application but not to limit the present application.
[0024] As shown in Fig. 1, the embodiment of the present application provides a target trajectory prediction method based on maneuvering intention inference, which comprises the following steps: Figure 1 S101: receiving measurement information of a target tracked by a radar at an initial time, giving target state information at the initial time, selecting a model set and a prediction model set of an interactive multiple model method, and obtaining inequality constraints corresponding to different maneuvering intention hypotheses and a threshold matrix of a sequential likelihood ratio test algorithm.
[0025] Specifically, the steps include: 1) receiving radar measurement information of a target tracked at an initial time (denoted as 0 time) , wherein, is an initial distance, is an initial azimuth angle, is an initial pitch angle, an initial process noise covariance and an initial measurement noise covariance are given; 12) giving state information of a target at an initial time: position , velocity , angular velocity and acceleration , obtaining a target initial state vector :
[0026] and a state estimation error covariance matrix ; 13) selecting tracking models as a model set of an interactive multiple model algorithm according to prior information, giving an initial model probability , and a transition probability matrix , wherein, probability that the target is in the i-th model at the initial time, probability that the target transfers from the i-th model to the j-th model, probability that the target transfers from the i-th model to the j-th model; 14) setting up the number of maneuver intention hypotheses according to prior information , given the inequality constraints corresponding to different maneuver intention hypotheses , and the threshold matrix of the sequential likelihood ratio test algorithm .
[0027] S102: track the target by using the interactive multiple model method, and obtain the target state estimation at the current time according to the target state information at the previous time and the measurement information at the current time.
[0028] Specifically, the following steps are included: 21) calculating the state estimation and the state estimation error covariance matrix of the target in the i-th model after input interaction according to the transition probability matrix; According to the probability that the target is in the i-th model at the initial time and the probability transition matrix of the models , the probability that the target is in the i-th model after input interaction can be calculated as:
[0029] According to the conditional probability calculation formula, the probability that the target is in the i-th model at the initial time under the condition that the target is in the j-th model at the initial time can be calculated as:
[0030] Therefore, the state estimation of the target in the i-th model after input interaction can be calculated as:
[0031] In the formula, is the state estimation vector of the target in the i-th model at the initial time; The state estimation error covariance matrix is:
[0032] wherein, the target is in the state estimation error covariance matrix of the th model at time 22) performing cubature Kalman filtering on the th model respectively to obtain the state estimation and the state estimation error covariance matrix of the th model at time , the measurement prediction error vector and the innovation covariance matrix ; wherein, measurement information of the target at time
[0033] wherein, the step of performing cubature Kalman filtering on the th model is: 221) calculating the state vector estimation of the th model at time -1 according to the third-order cubature rule, the corresponding cubature points and weights , wherein is the number of cubature points, denotes the dimension of ;
[0034]
[0035] wherein, is the state estimation error covariance matrix of the target at time -1, denotes the th column of the standard cubature point generation matrix .
[0036]
[0037] 222) performing nonlinear transformation on the cubature points according to the state equation of the th model to obtain the state prediction vector corresponding to the cubature points :
[0038] Accordingly, the state prediction :
[0039] and state prediction error covariance matrix :
[0040] where is the process noise covariance matrix at time k; 223) Calculate the state prediction vector according to the cubature rule the corresponding cubature points and weights :
[0041]
[0042] where is the state prediction error covariance matrix of the target at time k, , denotes the j-th column of the standard cubature point generation matrix .
[0043]
[0044] 224) Perform a nonlinear transformation on the cubature points according to the measurement equation to obtain the cubature point measurement prediction :
[0045] Accordingly, the measurement prediction :
[0046] innovation covariance matrix :
[0047] and cross-covariance matrix between state and measurement :
[0048] where is the measurement noise covariance matrix at time k; 225) Calculate the innovation covariance matrix and the cross-covariance matrix between state and measurement Calculate state estimation :
[0049] and state estimation error covariance matrix :
[0050] In the formula, express Measurement information at time.
[0051] 23) Predict the error vector based on the measurement. and the new covariance matrix Calculate the first The likelihood function of each model :
[0052] Normalizing the likelihood yields the target value. At the moment in the first The probability of each model :
[0053] 24) Based on the updated model probabilities Perform weighted summation to calculate the fused state estimate :
[0054] and state estimation error covariance matrix :
[0055] In the formula, for Time of the first State estimation of the model for Time of the first The state estimation error covariance matrix of the model for The target is at the first moment The probability of each model.
[0056] S103: Based on the target state information of the previous moment and the measurement information of the current moment, the projection method is used to solve the state estimation problem with inequality constraints, and the measurement likelihood of the target under different maneuver intention assumptions is calculated.
[0057] Specifically, the steps include the following: 31) Calculate according to the third-order volume rule State estimation at time 1 corresponding volume point and weight value , here represents dimension of
[0058]
[0059] wherein, is a standard volume point, represents the column of the standard volume point generation matrix ;
[0060] 32) For each of the maneuver intention hypotheses, the projection of the volume points not in the intention constraint region to the constraint region under the th hypothesis is calculated respectively; The intention under the th hypothesis can be expressed using the following inequality:
[0061] wherein, and are the lower and upper bounds of the constraint respectively, which correspond to the values according to the intention hypothesis; is a constraint function about , which can be linear or nonlinear, and in the linear case it can be simplified as represents, wherein is a linear constraint matrix. Then, the calculation of the projection of the volume points not in the intention constraint to the constraint region is equivalent to solving the following inequality constraint problem:
[0062] wherein, is an optimization variable, is a volume point not in the constraint region, is a weighted matrix, is the solution of the inequality constraint problem, i.e., the projection of the volume point in the constraint region, and are the upper and lower bounds of the constraint respectively wherein, the solution of the constraint problem is as follows: First, consider the linear equality constraint problem:
[0063] wherein, For optimization variables A linear constraint, For weighted matrices, For the constraint matrix, For constraint values; The solution to this constrained problem can be obtained using the Lagrange multiplier method.
[0064] In the formula, It is a projection matrix.
[0065] Secondly, consider the projection problem under linear inequality constraints:
[0066] In the formula, For optimization variables A linear constraint, For weighted matrices, For the constraint matrix, and These are the upper and lower bounds of the constraint, respectively.
[0067] Same record ,by To optimize the variables, this problem is equivalent to
[0068] In the formula, yes The function, This is a solution to the equivalent constraint problem. Therefore, the solution to the generalized linear constraint problem is...
[0069] Furthermore, if If it is one-dimensional, then its analytical solution is:
[0070]
[0071] Finally, consider the nonlinear constraint problem, where the nonlinear constraint function can be... At point A first-order Taylor expansion is then performed to transform the nonlinear problem into a linear one.
[0072] In the formula, yes exist gradient at; The volume point set can be obtained based on this solution. Projection points within the constrained area ; 33) Based on the volume point projection Calculate the first Measurement predictions and information covariance matrices under several assumptions; According to the measurement equation Projection of volume points Measurement prediction of volume point projection obtained by performing nonlinear transformation :
[0073] Based on this, the measurement prediction can be calculated. :
[0074] New covariance matrix :
[0075] 34) The likelihood function of the computer's intention hypothesis based on measurement prediction and the information covariance matrix; use d Describe the dimension of the measurement vector, and calculate the dimensionality of the measurement vector. Likelihood function of a maneuver intention hypothesis :
[0076] In the formula, This represents the error vector of measurement prediction.
[0077] S104: Calculate the matrix likelihood ratio statistic based on the measured likelihood, and use the matrix sequential likelihood ratio test algorithm to identify the target's maneuvering intention.
[0078] Specifically, the steps include the following: 41) Receiving time The likelihood function values of each maneuver intention hypothesis are calculated, and the log-likelihood ratio statistic is also calculated. according to Likelihood ratio statistic at time and time The likelihood function value of each maneuver intention hypothesis is calculated. Likelihood ratio statistic at time :
[0079] In the formula, for No. Line 1 The element values of the column, for The row, the element value of the column, is the likelihood function of the model at the moment, is the likelihood function of the model at the moment, is ; The logarithmic calculation of the likelihood ratio statistics at the moment is the logarithmic likelihood ratio statistics at the moment :
[0080] In the formula, is the element value of the row, the column, is the logarithmic likelihood ratio statistics at the moment the element value of the row, the column; 42) The likelihood function value at the moment is multiplied and normalized to obtain the probability that the th maneuver intention is recognized at the current moment : In the formula, is the likelihood function of the model at the moment, is the multiplication of the likelihood function of the
[0081] th model from the initial moment to the moment ; is the multiplication of the likelihood function of the th model from the initial moment to the moment ; is the multiplication of the likelihood function of the th model from the initial moment to the moment ; 43) The target maneuver intention is obtained by comparing the logarithmic likelihood ratio statistics and the threshold matrix ; If the likelihood ratio statistics of the th assumption is greater than all the thresholds relative to other assumptions, that is,
[0082] the assumption is accepted, that is, it is considered that the target at the current moment performs the maneuver corresponding to the th assumption, is No. Line number The element values of the column, for No. Line number The column's element values are updated; and the log-likelihood ratio statistic is updated to avoid historical cumulative bias, i.e.
[0083] If there is no log-likelihood ratio statistic that satisfies the above conditions, then all hypotheses are rejected, meaning that the target maneuver cannot be determined at the current moment.
[0084] S105: Based on the target's maneuvering intent at the current moment, use the maneuvering model corresponding to the target's maneuvering intent to predict the trajectory.
[0085] Specifically, the steps include the following: 51) Receive the current moment's intention inference results and select the model for trajectory prediction; like If the target's maneuver cannot be determined at any given time, an interactive multi-model algorithm is selected for trajectory prediction; if It is possible to judge the target's movement at any time and receive the first When a certain assumption is made, the corresponding model maneuvering model is selected for trajectory prediction. 52) Receive the selected prediction model and perform predictions under that model. Trajectory prediction for each step length, calculating the target State prediction vector at time step and state prediction error covariance matrix The target trajectory prediction result is obtained.
[0086] 521) If the target's maneuver at the current moment cannot be determined, the Interactive Multi-Model Algorithm (IMAMA) is used for trajectory prediction to obtain the target trajectory prediction result. The specific steps for trajectory prediction using the IMAMA are as follows: a) Select based on prior information A prediction model As a model set for an interactive multi-model algorithm, given the initial model probabilities as follows: The transition probability matrix is ,in, Indicates that the target is at the initial time. The probability of each model, Indicates the target starts from the first The model is transferred to the first... The probability of each model; according to The target is at the first moment The probability of each model and the probability transition matrix of the model The probability of the target being in the i-th model after the input interaction can be calculated as
[0087] In the formula, is the probability of the target being in the i-th model at time t; According to the conditional probability calculation formula, the probability of the target being in the i-th model at time t, given that the target is in the i-th model at time t, can be calculated as In the formula,
[0088] In the formula, is the probability of the target being in the i-th model after the input interaction; b) The state estimation of the target being in the i-th model after the input interaction can be calculated as
[0089] The state estimation error covariance matrix of the target being in the i-th model at time t is In the formula,
[0090] is the state estimation error covariance matrix of the target being in the i-th model at time t; c) According to the third-order cubature rule, the state vector estimation value at time t is calculated as The corresponding cubature points and weights are where is the number of cubature points, represents the dimension of ;
[0091]
[0092] In the formula, is the state estimation error covariance matrix of the target at time t. is the standard cubature point, , represents the standard cubature point generation matrix of the first column;
[0093] d) a nonlinear transformation of the cubature points according to the state equation of the th model to obtain the state prediction corresponding cubature points :
[0094] Accordingly, the state prediction :
[0095] and the state prediction error covariance matrix :
[0096] e) the state prediction and the state prediction covariance of the models are weighted and fused, and the state prediction based on the interacting multiple model algorithm :
[0097] and the state prediction error covariance matrix :
[0098] In the formula, is the state prediction error covariance matrix of the th model at time , and is the state prediction of the th model at time ; 522) If the target maneuver at the current time can be determined, the trajectory prediction steps are as follows: 5221) according to the third-order cubature rule, the state vector estimation value at time corresponding cubature points and weights , , where is the number of cubature points, represents the dimension of ;
[0099]
[0100] wherein, , denotes the standard volume point generation matrix of the first column;
[0101] 5222) Projecting the volume point set according to the selected prediction model; If the target at time performs the first assumption corresponding maneuver, the intention under the first assumption can be expressed using the following inequality:
[0102] wherein, and are the lower and upper bounds of the constraint respectively, and the values are determined according to the intention assumption; is the constraint function about , which can be linear or nonlinear, and in the linear case it can be simplified as denotes. Then under the first l assumption, the volume point is projected as follows:
[0103] The projection point of the volume point set in the constraint region can be obtained; 5223) Nonlinear transformation of volume point projection according to the state equation; If the target at time performs the first assumption corresponding maneuver, the nonlinear transformation of the volume point projection is as follows:
[0104] wherein, is the state prediction of the volume point projection ; 5224) Calculate the state prediction vector and the state prediction error covariance matrix of step length, and obtain the target trajectory prediction result; If the target at time performs the first assumption corresponding maneuver, the state prediction vector can be calculated according to the nonlinear transformation result of the volume point:
[0105] And the state prediction error covariance matrix:
[0106] S106: Results processing and output.
[0107] Output target tracking results Intent inference results and trajectory prediction results .
[0108] The present invention will be further illustrated below through specific embodiments.
[0109] (1) Receive radar measurement data at the initial moment. Given the target's initial state information, including the target's position, velocity, acceleration, and angular velocity, this embodiment initializes the tracking model and prediction model of the interactive multi-model algorithm based on prior information. Both include a near-uniform motion model, a near-uniform acceleration motion model, and a near-uniform circular motion model. Set the maneuver intention assumptions as left turn, right turn, up somersault, and down somersault. Given the inequality constraints and prediction models corresponding to different maneuver intention assumptions, set the threshold matrix of the sequential likelihood ratio test algorithm.
[0110] (2) Receive the target status information from the previous moment. and measurement information at the current moment State estimation is performed using an interactive multiple model method based on capacitive Kalman filtering to obtain the target's state. State estimation at time 1 and state estimation error covariance matrix .
[0111] (3) Based on the target state information of the previous moment and measurement information at the current moment Calculation based on third-order volume rule State estimation at time 1 Corresponding volume point and weight Calculate the first one using the projection method. Under the assumption of a maneuvering intention, the volume point set Projection points within the constrained area And calculate the first based on the volume point projection. The likelihood function value of the hypothetical maneuver intention.
[0112] (4) Calculate the log-likelihood ratio statistic based on the likelihood function values of all maneuver intentions hypothesized at the current time. and the current moment The probability of a maneuvering intent being recognized The matrix sequential likelihood ratio test algorithm is used to identify the target's maneuvering intent: if the first... If the likelihood ratio statistic of the first hypothesis is greater than the set threshold, then the hypothesis is accepted, meaning that the target is considered to be performing the first hypothesis at the current time. The maneuver corresponding to each hypothesis is determined, and the log-likelihood ratio statistic is updated. If there is no log-likelihood ratio statistic that satisfies the above conditions, all hypotheses are rejected, meaning that the target maneuver cannot be determined at the current time.
[0113] (5) Receive the intent inference result at the current moment and perform... Predicting the target trajectory for a step length, if If the target's maneuver cannot be determined at any given time, an interactive multi-model algorithm is selected for trajectory prediction; if Always accept the first We make several assumptions and select the model corresponding to each assumption for trajectory prediction. This yields... State prediction vector at time step and state prediction error covariance matrix Then proceed to step (2) and repeat the subsequent process until the termination condition is met, such as when the current trajectory has been tracked.
[0114] (6) Proceed to step (1) and repeat the subsequent process until the number of experiments reaches the set number of Monte Carlo experiments. Then, organize and output the results, and calculate the root mean square error of target position estimation and velocity estimation, the recognition probability of the intent model and the root mean square error of position prediction and velocity prediction.
[0115] Figure 2 In the simulation scenario 1 shown, the target makes a left turn with an initial position of [0m, 0m, 2500m], an initial velocity of [50m / s, 100m / s, 0m / s], and an initial angular velocity of [0rad / s, 0rad / s, 0.2rad / s]. The simulation duration is 18s, the sampling interval is 0.5s, and the number of Monte Carlo simulations is 100.
[0116] Figures 3(a)-3(b) The root mean square error of the position and velocity of the target tracking using the interactive multi-model algorithm is shown in simulation scenario 1. The position estimation stabilizes after 4 seconds and the velocity estimation stabilizes after 7 seconds.
[0117] Figure 3(c) shows the inference probability of each maneuver intention. The correct intention probability is highest after the 3rd second, and the intention is accurately inferred after the 4th second. This corresponds to the stability of the target position estimation, with fast convergence speed and stable intention inference results.
[0118] Figures 3(d)-3(e)Fig. 6 is a comparison diagram of root mean square errors of target position and velocity for trajectory prediction based on maneuver intention inference and trajectory prediction using only an interacting multiple model algorithm (a comparative algorithm). After the 8th second, the target position prediction based on maneuver intention inference is slightly better than the prediction result of the comparative algorithm; after the 6th second, the target velocity prediction based on maneuver intention inference is slightly better than the prediction result of the comparative algorithm. Considering that the prediction length is 3 seconds, the above results show that after the target maneuver intention is correctly inferred, the trajectory prediction based on intention inference converges faster and the result is more stable than the comparative algorithm.
[0119] Figure 4 In the illustrated simulation scenario 2, the target performs an upper somersault, the initial position is [0 m, 0 m, 2500 m], the initial velocity is [0 m / s, 100 m / s, 0 m / s], the initial angular velocity is [0.2 rad / s, 0 rad / s, 0 rad / s], the simulation duration is 15 seconds, the sampling interval is 0.5 seconds, and the number of Monte Carlo is 100 times.
[0120] Figures 5(a)-5(b) Fig. 5 (c) shows the root mean square errors of position and velocity for target tracking using an interacting multiple model algorithm in a simulation scenario 3, wherein the position estimation is stable after the 5th second, and the velocity estimation is stable after the 7th second.
[0121] Fig. 5 (c) is the inference probability of each maneuver intention, the probability of correct intention is the highest after the 3rd second, the intention is correctly inferred after the 5th second, which corresponds to the stable target position estimation, the convergence speed is fast, and the intention inference result is stable.
[0122] Figures 5(d)-5(e) Fig. 6 is a comparison diagram of root mean square errors of target position and velocity for trajectory prediction based on maneuver intention inference and trajectory prediction using only an interacting multiple model algorithm (a comparative algorithm). After the 8th second, the target position prediction based on maneuver intention inference is slightly better than the prediction result of the comparative algorithm; after the 6th second, the target velocity prediction based on maneuver intention inference is slightly better than the prediction result of the comparative algorithm. Considering that the prediction length is 3 seconds, the above results show that after the target maneuver intention is correctly inferred, the trajectory prediction based on intention inference converges faster and the result is more stable than the comparative algorithm.
[0123] The above experimental results verify that the intention inference result in the present application is stable, and the accuracy and convergence efficiency of target trajectory prediction are improved. The accurate intention inference and trajectory prediction are achieved in multiple scenarios with the same target motion model but different maneuver intentions, which has great theoretical and practical value.
[0124] The present application is not limited to the above-mentioned embodiments, and based on the technical solutions disclosed in the present application, those skilled in the art can make some substitutions and modifications to some technical features without creative labor, and these substitutions and modifications are all within the protection scope of the present application.
Claims
1. A target trajectory prediction method based on maneuver intent inference, characterized in that, include: Receive measurement information at the initial moment of radar tracking target, given target state information at the initial moment, select model set and prediction model set of interactive multi-model method, and obtain inequality constraints and threshold matrix of sequential likelihood ratio test algorithm corresponding to different maneuver intention assumptions. The interactive multi-model approach is used to track the target. Based on the target state information of the previous moment and the measurement information of the current moment, the target state estimate at the current moment is obtained. Based on the target state information of the previous moment and the measurement information of the current moment, the measurement likelihood of the target under different maneuvering intentions is calculated using the projection method. Based on the measured likelihood, the matrix likelihood ratio statistic is calculated, and the matrix sequential likelihood ratio test algorithm is used to identify the target's maneuvering intentions. Based on the target's maneuvering intent at the current moment, the maneuvering model corresponding to the target's maneuvering intent is used to predict the trajectory, and the target trajectory prediction result is obtained.
2. The target trajectory prediction method based on maneuver intent inference according to claim 1, characterized in that, Receive measurement information from radar tracking the target at the initial moment, given the target state information at the initial moment, select the model set and prediction model set of the interactive multi-model method, and obtain the inequality constraints corresponding to different maneuver intent assumptions and the threshold matrix of the sequential likelihood ratio test algorithm, including: Receive radar measurement information at the initial moment of target tracking, and provide the initial process noise covariance and the initial measurement noise covariance; Given the target's initial state information, we obtain the target's initial state vector and the state estimation error covariance matrix; The tracking model is selected as the tracking model set for the interactive multi-model method based on prior information, given the initial model probability and transition probability matrix; Based on prior information, we set maneuver intent assumptions, and provide inequality constraints and threshold matrices for sequential likelihood ratio testing algorithms corresponding to different maneuver intent assumptions.
3. The target trajectory prediction method based on maneuver intent inference according to claim 2, characterized in that, The radar measurement information includes initial range, initial azimuth angle, and initial elevation angle; the status information includes position, velocity, angular velocity, and acceleration.
4. The target trajectory prediction method based on maneuver intent inference according to claim 1, characterized in that, Obtain the target state estimate at the current moment, including: Calculate the target's state estimate and state estimate error covariance matrix after input interaction based on the transition probability matrix; Perform capacitive Kalman filtering on each tracking model in the tracking model set to obtain the state estimate and state estimate error covariance matrix of each model at the current time, as well as the measurement prediction error vector and innovation covariance matrix; The likelihood function and model probability of each model are calculated based on the measurement prediction error vector and the information covariance matrix. The fused state estimate and state estimate error covariance matrix are calculated based on the updated model probabilities.
5. The target trajectory prediction method based on maneuver intent inference according to claim 4, characterized in that, Perform capacitive Kalman filtering on each tracking model, including: The state vector at the previous moment is estimated based on the third-order volume rule, along with the corresponding volume points and weights. The state prediction and the state prediction covariance matrix are obtained by performing a nonlinear transformation on the volume point based on the state equation. Calculate the volume points and weights corresponding to the state prediction vector according to the volume rule; Based on the measurement equation, a nonlinear transformation is performed on the volume point to obtain the volume point measurement prediction, the information covariance matrix, and the cross-covariance matrix between the state and the measurement. Calculate the state estimate and the state estimate error covariance matrix based on the information covariance matrix and the cross-covariance matrix of the state and measurement.
6. The target trajectory prediction method based on maneuver intent inference according to claim 1, characterized in that, The measurement likelihood of a target under different maneuvering intent assumptions is calculated using the projection method, including: Calculate the volume points and weights corresponding to the target state estimate of the previous time step according to the third-order volume rule; For each maneuver intent assumption, calculate the projection of the volume point outside the intent constraint area to the constraint area under different assumptions; Calculate the measurement prediction and information covariance matrix under different assumptions based on volume point projection; The likelihood function is based on the measurement prediction and the new information covariance matrix, which is the computer's intention assumption.
7. The target trajectory prediction method based on maneuver intent inference according to claim 1, characterized in that, Calculate the matrix likelihood ratio statistic and use the matrix sequential likelihood ratio test algorithm to identify the target's maneuvering intent, including: Receive the likelihood function values of each maneuver intention hypothesis at the current time, and calculate the log-likelihood ratio statistic; The probability of each maneuver intention being recognized is obtained by multiplying the likelihood function values at the current time and then normalizing them. By comparing the log-likelihood ratio statistic and the threshold matrix, the target's maneuvering intention can be obtained: If the likelihood ratio statistic of a certain hypothesis is greater than the set threshold, then the target is considered to be performing the maneuver corresponding to that hypothesis at the current moment; if the above condition is not met, then the target maneuver at the current moment cannot be determined.
8. The target trajectory prediction method based on maneuver intent inference according to claim 1, characterized in that, Based on the target's maneuvering intent at the current moment, a trajectory prediction is performed using the maneuvering model corresponding to the target's maneuvering intent, resulting in a target trajectory prediction result, including: Receive the current moment's intent inference result and select the model for trajectory prediction: If the target maneuver at the current moment cannot be determined, select the interactive multi-model algorithm for trajectory prediction; if the target maneuver at the current moment can be determined, accept a certain hypothesis and select the maneuver model corresponding to that hypothesis for trajectory prediction. The system receives the selected prediction model, performs trajectory prediction with a fixed step size under the model, calculates the target's state prediction vector and state prediction error covariance matrix, and obtains the target trajectory prediction result.
9. The target trajectory prediction method based on maneuver intent inference according to claim 8, characterized in that, If the target's current maneuver cannot be determined, the interactive multi-model algorithm is selected for trajectory prediction. The specific steps are as follows: Based on prior information, a set of prediction models is selected as the model set for the interactive multi-model algorithm. Given the initial model probability and transition probability matrix, the model probability of the target after input interaction is calculated. The target's model state estimate and state estimate error covariance matrix after input interaction are calculated. Calculate the volume point and weight corresponding to the estimated state vector at the current moment according to the third-order volume rule; Based on the state equation of the model, a nonlinear transformation is performed on the volume points to obtain the state prediction and the state prediction error covariance matrix; The state predictions and state prediction covariances of each model are weighted and fused to calculate the state prediction vector and state prediction error covariance matrix based on the interactive multi-model algorithm, thus obtaining the target trajectory prediction result.
10. The target trajectory prediction method based on maneuver intent inference according to claim 8, characterized in that, If the target's maneuver at the current moment can be determined, and a certain assumption is accepted, the maneuver model corresponding to that assumption is selected for trajectory prediction. The specific steps are as follows: Calculate the volume point and weight corresponding to the current state estimate based on the third-order volume rule; The volume points are projected and transformed according to the selected prediction model; A nonlinear transformation is performed on the volume point projection based on the state equation; Calculate the state prediction vector and state prediction error covariance matrix for the set step size to obtain the target trajectory prediction result.