An interactive multiple model state estimation method based on intention estimation

Abstract

The application discloses an interactive multi-model state estimation method based on intention estimation, which is used to improve the estimation accuracy of a state estimator, reduce estimation errors and over-forecasting quantities when a tracked target makes a maneuver, and further adapt to higher-precision target tracking requirements. Although various filter variants or robust filters appearing after Kalman filters can compensate for the problems of inaccurate and nonlinear state equations to a certain extent, they are also difficult to meet the state estimation requirements in the field of fast tracking of maneuvering targets. The application proposes an interactive multi-model state estimation method based on intention estimation, which breaks through the limitations of traditional filtering methods, can estimate the motion intention of a tracked target and key parameters of a maneuvering equation when the tracked target makes a fast maneuver, significantly improves the estimation accuracy and estimation curve smoothness of an estimator, and optimizes the estimation effect of the estimator.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of target tracking and state estimation, and particularly relates to an interactive multi-model state estimation method based on intention estimation, which is mainly used for improving the estimation accuracy and estimation curve smoothness of a state estimator and optimizing the estimation effect of the state estimator in the case that a tracked target performs a maneuvering motion. BACKGROUND

[0002] The application can be applied to the fields of photoelectric tracking and motion capture. In these application fields, although the motion of a tracked target is often unpredictable, its motion trajectory can usually be summarized as deformation and combination of a limited number of actions by using a statistical induction method (Research on Motion Trajectory Detection and Recognition Technology, Wang Yu, Northeastern University, 2015). The most classic Kalman state estimation method in this field (Principle and Application of Kalman Filter, Huang Xiaoping, 2015) has optimal estimation ability for Gaussian white noise interference, but it is limited by the need for accurate and linear state equations of a tracked target, and the estimation effect cannot meet the tracking accuracy requirements for a maneuvering target (Applied Optimal Estimation, Arthur G. Proceedings of the IEEE, 1974). Although the traditional robust state estimation method can compensate for the uncertainty in the motion model of the estimated object, the complex maneuvering of the tracked target has exceeded the compensable range of the robust state estimation (A Robust State Estimator With Adaptive Factor, Minxing Sun, IEEE Access 2020). The traditional interactive multi-model state estimator can improve the estimation effect for a maneuvering target (Performance Analysis of Interactive Multi-Model Algorithm, Liang Yan, Control Theory and Applications, 2001), but it only estimates the kinematic characteristics of the tracked target rather than the motion intention, which leads to difficulty in further improving the estimation accuracy, especially the prediction accuracy, and there is still much room for improvement (Maneuvering Target State Estimation, Hao Jiankang, System Engineering and Electronics Technology, 1997). With the increasing performance requirements of state estimation methods in various fields, it is particularly important to quickly estimate the current motion intention and perform real-time estimation when the tracked target performs a maneuvering motion if the possible motion methods of the tracked target are known according to statistical methods. We urgently need to improve the current state estimation method so that it can still maintain good estimation performance in the case of noise variation. SUMMARY

[0003] In order to improve the position estimation and prediction ability of the state estimator for a maneuvering target, the application proposes an interactive multi-model state estimator based on intention estimation.

[0004] To achieve the object of the present application, the present application first defines the state equation corresponding to the specific intention state estimator and the meaning of the parameters therein:

[0005] x p (k+1)=f p (x p (k),u(k)),k≥0

[0006] z(k)=h p (x p (k),v(k))

[0007] In the state equation, x p (k), y p (k) are the tracked target state and observation value of the specific intention state estimator in the interactive state estimator for intention p respectively; x p (0), u p (k), v p (k) conform to Gaussian distribution; f p (x i , u i ), h p (x i , v i ) are the state transition equation and observation equation of intention p respectively, and δ ij is the Kronecker symbol equation, i.e. δ ij =1 when i=j; δ ij =0 when i≠j. u(k), v(k) satisfy:

[0008]

[0009] The use of the interactive multi-model state estimator based on intention estimation requires setting the state estimation initial value of the estimator according to the use scenario According to statistical data, the probability transition matrix P of the motion intention of the tracked target is obtained, the model probability μ p of the specific intention state estimator and their estimation initial value Position prediction initial value and posterior estimation covariance

[0010] It should be noted that the element p ij in the probability transition matrix P of the motion intention represents the probability of the tracked target changing from motion intention i to motion intention j.

[0011]

[0012] The model probability μ p of the specific intention state estimator represents the probability of the current motion intention of the tracked target being p.

[0013] The interactive multi-model state estimator based on intention estimation simplifies the input interaction step of traditional interactive multi-model, directly uses specific intention state estimators to estimate respectively, to obtain the estimated value of estimator p Position prediction value And estimation covariance P p (k).

[0014] It should be noted that the specific intention state estimator is constructed based on the principle of unscented Kalman filter. The difference between it and the traditional unscented Kalman state estimation method is that the motion intention p and the corresponding motion model z=M p (c1,c2,...,c n ) of the tracked target need to be obtained through trajectory pattern learning algorithm, and then the state transition equation And the observation equation are established according to the motion model, wherein:

[0015]

[0016]

[0017] In most scenarios, the motion model parameters c1,c2,...,c n of the tracked target are constant or constant step change in a maneuver, and at this time That is, it is simplified as:

[0018]

[0019] Where D n is a constant that can be 0. After determining the state transition equation and the observation equation, the specific intention state estimator should be implemented according to the following steps:

[0020] Step (1): Set parameters a, b, K according to the target tracking scene, and set the initial value of state estimation value The initial value of state posterior covariance P p (0), the observation noise covariance R, and the process noise covariance Q;

[0021] Step (2): According to the preset values of a, K and the dimension n of the estimated state , calculate the parameter l:

[0022] l=a 2 (n+K)-n;

[0023] Step (3): According to the estimated state , its dimension n, the preset a, b, the calculated l in the last step, and the state posterior covariance P(k), 2n+1 Sigma points X(l) , ( denotes the i-th column root of the matrix A):

[0024]

[0025] Step (4): Calculate 2n+1 Sigma points X (l) with respective weights w (l) ;

[0026]

[0027] Step (5): 2n+1 Sigma points X (l) are predicted according to the state transition equation

[0028]

[0029] Step (6): Calculate the estimated state with respective weights w (l) , the prior state value and the state prior covariance P(k+1|k) from 2n+1 points and the corresponding weights w

[0030]

[0031]

[0032] Step (7): Perform UT transformation on the prior state value to obtain 2n+1 new Sigma points

[0033]

[0034] Step (8): 2n+1 Sigma points are simulated according to the observation equation to obtain Z (l) :

[0035]

[0036] Step (9): Predict the observation value and the covariance P xz and P zz from 2n+1 Z (l) points and the corresponding weights w (l) :

[0037]

[0038] ​

[0039]

[0040] Step (10): Calculate the Kalman gain matrix K:

[0041]

[0042] Step (11): Calculate the estimation result of the specific intention p state estimator and the covariance P(k+1):

[0043]

[0044] Step (12): Calculate the position estimation value of the tracked target according to the estimation result

[0045]

[0046] After obtaining the position prediction result of each specific intention state estimator , the interactive algorithm needs to evaluate the possibility μ of each motion intention of the tracked target p , and the calculation method is:

[0047]

[0048] In the above formula, v p (k) is the target position estimation error, S p (k) is the corresponding estimation error covariance, Λ p (k) is the likelihood function of intention p, and c is the normalization constant.

[0049] The last step of the interactive multi-model state estimator based on intention estimation is to weight the target position estimation result of the specific intention state estimator according to the possibility μ of each intention p , and obtain the target position prediction value of the entire estimator:

[0050]

[0051]

[0052] Compared with the prior art, the present application has the following advantages:

[0053] (1) Compared with the classical Kalman filtering algorithm and the traditional robust algorithm, the present application effectively compensates for the uncertainty of the motion equation of the tracked target by real-time estimation of the parameters in the motion equation of the tracked target;

[0054] ​(2) Compared with the classic Kalman filter algorithm, the present invention can effectively deal with the tracked target with nonlinear motion;

[0055] (3) Compared with the traditional unscented Kalman filter algorithm, the present invention effectively improves the estimation accuracy and the adaptability of the state estimator by estimating the parameters of the motion equation of the tracked target;

[0056] (4) Compared with the traditional interactive multi-model method, this invention effectively improves the accuracy of state estimation and adaptability to motion trajectory distortion by estimating the motion intention and motion equation parameters of the tracked target. Attached Figure Description

[0057] Figure 1 It is a simulation diagram of the motion trajectory of the tracked target.

[0058] Figure 2 This is a schematic diagram of the position prediction error of a moving target.

[0059] Figure 3 It is a statistical analysis of the cumulative estimation errors of several estimators.

[0060] Figure 4 It is a real-time target tracking result based on an interactive multi-model state estimator with intent inference and other estimators.

[0061] Figure 5 This invention presents a statistical graph of the cumulative estimation errors of different estimators. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0063] To achieve the objectives of this invention, an interactive multi-model state estimation method based on intent estimation is provided, the implementation of which is as follows:

[0064] Before using a state estimator, it is necessary to predict the motion model of the tracked target based on the usage scenario, or obtain the motion model z=M of the tracked target based on measured data and a trajectory pattern learning algorithm. p (c1,c2,...,c n Then, state transition equations are established based on the motion model. and observation equations Next, design the estimator:

[0065] Step (1): Set the initial estimates for the interactive multi-model state estimator based on intent estimation The probability transition matrix P of the tracked target's motion intention, and the probability μ of each motion intention. p Initial values ​​of estimates for each specific intention state estimator Position prediction initial value Posterior estimation covariance initial value P p (0), parameters a, b, k, observation noise covariance R, process noise covariance Q. Then, for each specific intent state estimator, complete the calculation of steps (2)-(12).

[0066] Step (2): Calculate parameter l;

[0067] Step (3): Calculate 2n+1 Sigma points X (l) ;

[0068] Step (4): Calculate 2n+1 Sigma points X (l) , respectively, weights w (l) ; of each;

[0069] Step (5): Make a step prediction to obtain

[0070] Step (6): Calculate the prior state value of the estimated state and the state prior covariance P(k+1|k);

[0071] Step (7): Perform UT transformation on the prior state value to obtain 2n+1 new Sigma points

[0072] Step (8): Perform observation simulation on 2n+1 Sigma points , respectively, to obtain Z (l) ;

[0073] Step (9): Predict the observation value and covariance P xz and P zz ;

[0074] Step (10): Calculate the Kalman gain matrix K;

[0075] Step (11): Calculate the estimation result and covariance P(k+1) of the specific intent p state estimator;

[0076] Step (12): Calculate the position estimation value of the tracked target according to the estimation result

[0077] Step (13): Evaluate the possibility p of each motion intent of the tracked target using an interactive algorithm;

[0078] Step (14): Calculate the weighted prediction value of the target position of the tracked target;

[0079] In the following Matlab simulation and state estimation experiment, the estimation effects of the interactive multiple model state estimator based on intention estimation, the classical Kalman state estimator, the robust state estimator, the unscented Kalman estimator for sinusoidal motion (UKF1), and the unscented Kalman estimator for triangular wave (UKF2) will be compared to explain the design process and effects of the application in detail:

[0080] (1.1): In the Matlab state estimation simulation, it is assumed that the tracked target has three kinds of motion intentions, which are uniform straight line motion, variable ellipse motion, and xy-axis compound triangular wave motion. Their motion trajectories are shown in Figure 1 , which are compound triangular wave, uniform straight line, variable ellipse, and uniform straight line motion in time sequence.

[0081] (1.2): The interactive multiple model state estimator based on intention estimation is composed of a classical Kalman filter, an ellipse motion intention estimator, and a compound triangular wave motion intention estimator, which correspond to the straight line motion, ellipse motion, and compound triangular wave motion intentions of the tracked target, respectively. The parameters of the classical Kalman filter are as follows:

[0082] The parameters of the ellipse motion intention estimator are as follows:

[0083]

[0084]

[0085]

[0086]

[0087] α = 0.01, β = 2, κ = 0.

[0088] wherein, is the estimated state, is the position observation value of the tracked target, t x , t y , w x , w y , A x , A y is the time, angular frequency, and amplitude parameters of the sinusoidal motion, p x , p y is the position of the tracked target in the x and y axes, Q is the process noise covariance matrix, R is the observation noise covariance matrix, and α, β, κ are the adjustment parameters.

[0089] The parameters of the compound triangular wave motion intention estimator are as follows:

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] in, It is the estimated state. It is the observed position of the tracked target, t x t y w x w y A x A y T xon T yon These are the parameters of trigonometric motion: time, angular frequency, amplitude, and the proportion of the rising segment. x p y The x and y axes represent the position of the tracked target, Q is the process noise covariance matrix, R is the observation noise covariance matrix, and α, β, and κ are adjustment parameters.

[0096] Other parameters used by the interactive multi-model state estimator are as follows:

[0097]

[0098] (1.3): Perform state estimation simulations on the intention-based interactive multi-model state estimator, the classical Kalman state estimator, the robust state estimator, the sine estimator for sinusoidal motion (UKF1), and the sine estimator for triangular waves (UKF2) according to the methods described in steps (2) to (14).

[0099] (1.4): such as Figure 1 This is a simulation of the motion trajectory of the tracked target. Motion disturbances and observation noise are incorporated into the simulation, which consists of uniform linear motion, composite triangular wave motion, and variable elliptical motion connected together. The position prediction error of the moving target is as follows: Figure 2 As shown, Figure 2 The data includes the true trajectory error on the x-axis, the classical Kalman state estimation error, the robust state estimation error, the unscented Kalman estimation error for sinusoidal motion (UKF1), the unscented Kalman estimation error for triangular waves (UKF2), and the interactive multi-model state estimation error based on intent estimation (IMM). It can be seen that the estimation method proposed in this invention has the smallest estimation error, i.e., the highest estimation accuracy. To more intuitively demonstrate the improvement in estimation accuracy brought about by this invention, Figure 3The cumulative estimation errors of several estimators are counted in the statistics, and it can be seen that the present application (IMM) has the smallest cumulative estimation error, which again proves the superiority of the present application.

[0100] (2.1): The estimation effects of the interactive multiple model state estimator based on intention estimation, the classical Kalman state estimator, the robust state estimator, the unscented Kalman estimator (UKF1) for sinusoidal motion, and the unscented Kalman estimator (UKF2) for triangular waves are also compared in the experiment. The estimator parameters used are the same as in (1.2) and are not repeated. Through experiments and data statistics, the present application has Figure 4 the real-time target tracking results of the interactive multiple model state estimator based on intention estimation and other estimators in the statistics, and it can be seen that the practical application effect of the estimator of the present application is excellent, and the real-time position of the tracked target is well estimated. The present application further reflects the superiority of the present application in Figure 5 the cumulative estimation errors of different estimators are counted in the statistics, and the estimation error curve of the interactive multiple model state estimator based on intention estimation is always at the bottom of the several curves, that is, the estimation error is the smallest, and the estimation accuracy is the highest, which further reflects the superiority of the present application.

Claims

1. An interactive multiple model state estimation method based on intent estimation, characterized by: The specific implementation steps are as follows: Step (1): Setting the initial value of the estimated value of the position of the tracked target of the interactive multiple model state estimator , the probability transition matrix P of the different intentions of the tracked target, the probabilities of the current intentions of the tracked target , the initial value of the state estimated value of each specific intention state estimator , the initial value of the target position prediction value , and the initial value of the state posterior covariance ; Step (2): State estimation is performed using each specific intent state estimator based on the state estimate value , the posteriori estimate covariance , and the observation value , to obtain a new state estimate value , a target position prediction value , and a posteriori estimate covariance ; Specifically include: Step (A1): Set parameters , state estimation initial value , state posterior covariance initial value , observation noise covariance , process noise covariance , state transition equation of unscented Kalman filter , and observation equation ; Step (A2): Calculate the parameter according to the value of the preset and the dimension n of the state estimate value ; Step (A3): Calculate the state estimate value and its dimension n, the preset , the state posterior covariance , the state posterior covariance , and 2n+1 Sigma points ; Step (A4): Compute 2n+1 Sigma points Respective weights ; Step (A5): 2n+1 Sigma points According to the state transition equation , one-step prediction is made respectively to obtain ; Step (A6): Based on 2n+1 Points and corresponding weights Calculate the posterior state estimate Prior state values State prior covariance ; Step (A7): Update the prior state value UT transform to get 2n+1 new Sigma points ; Step (A8): 2n+1 Sigma points According to the observation equation , respectively, the observation simulation ; Step (A9): Predict the observation and covariance from 2n+1 points and corresponding weights and ;​​​ Step (A10): Calculate Kalman gain matrix ; Step (A11): Compute posterior state estimate of the system and covariance ; Step (A12): Compute the position prediction of the tracked object under the intent p ; Step (3): Calculate the probability of each motion intention of the tracked target ; Step (4): Compute the overall target position prediction of the interacting multiple model state estimator and the state estimation covariance ; Step (5): return to step (2) for a new round of state estimation.

2. The intent-augmented interactive multiple model state estimation method of claim 1, wherein: The state estimation using the specific intention state estimators in step (2) needs to acquire the motion data of the tracked target in a specific tracking scene in advance, and then use a trajectory pattern learning algorithm to cluster and induce several motion intentions of the tracked target and corresponding motion functions Here, the output of the motion function is the position observation value of the tracked target, is the parameter obtained by the trajectory pattern learning.

3. The intent-based interactive multiple model state estimation method of claim 1, wherein: the state estimate in step (A1) The state vector consists of the parameters of the corresponding motion function consists of: 。 4. The intent-based interactive multiple model state estimation method of claim 1, wherein: The state transition equation in step (A1) The change rule of the parameter is composed of the change rule of the parameter The change rule is obtained through trajectory mode learning.

5. The intent-based interactive multiple model state estimation method of claim 1, wherein: the observation equation in step (A1) , according to the estimated motion function parameters obtaining the tracked target position values: 。 6. The intent-augmented interactive multiple model state estimation method according to any one of claims 1-5, characterized in that: Through the simultaneous estimation of the state estimator for different motion intentions of the tracked target, the estimation results under different motion intention assumptions are obtained. The interactive multiple model further calculates the possibility of different motion intentions of the tracked target according to the estimation results and the observation results, and obtains the optimal target position estimation or prediction result through weighting, so as to improve the accuracy and smoothness of state estimation and enhance the estimation effect of the estimator.

Images