Variational Bayesian unscented Kalman filter target tracking method based on M estimation
By using M estimation and variational Bayesian traceless Kalman filtering methods to deal with noise interference and non-Gaussian measurement problems in underwater target tracking, the stability of the singular value decomposition is used to solve the noise interference and noise interference in the prior art. The impact of non-Gaussian measurement on tracking accuracy achieves higher tracking accuracy and stability.
Patent Information
- Application Number
- CN202411282217.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2024-11-12
AI Technical Summary
The prior art is difficult to effectively deal with noise interference and non-Gaussian measurement problems in nonlinear systems in underwater target tracking, resulting in reduced tracking accuracy and unstable algorithms.
Using the variational Bayesian traceless Kalman filtering method based on M estimation, the noise distribution is measured through variational strategy, the observation data is preprocessed using the M estimation method, and the covariance matrix divergence problem is solved through singular value decomposition, and the algorithm is improved stability.
The accuracy and stability of underwater target tracking are improved, and the impact of noise interference on the algorithm is reduced. Simulation experiments show that there is higher tracking accuracy and better stability under non-Gaussian conditions.
Smart Images

Figure CN118938230A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a variational Bayesian unscented Kalman filtering target tracking method based on M estimation, and belongs to underwater target tracking. Background Art
[0002] The underwater battlefield has become one of the main battlefields of naval warfare today due to its good concealment. Germany's submarine warfare in World War II once caused great damage to the Allied forces' maritime transportation lines, making countries around the world realize the tremendous power of underwater warfare. After the end of World War II, with the development of technologies such as vibration reduction and noise reduction, and intelligent torpedoes, submarines have made great progress in concealment, maneuverability, and attack range. All countries have also attached great importance to the development of underwater combat technology, and are constantly strengthening the construction of underwater platforms and underwater weapons and equipment to improve their underwater combat capabilities.
[0003] Underwater target tracking plays a vital role in national security. For the classic linear Gaussian system problem, the classic Kalman filter can be used to obtain the optimal parameter estimation. However, for the tracking of underwater maneuvering bodies, the underwater active detection platform usually provides the position information and orientation information of the maneuvering body from the observation station, while the passive station can only provide the orientation information of the maneuvering body. The measurement information obtained by both the active detection platform and the passive detection platform has strong nonlinearity. For nonlinear problems, the Kalman filter algorithm will no longer be applicable. The current mainstream nonlinear filtering algorithms mainly include extended Kalman filter, unscented Kalman filter, cubature Kalman filter and particle filter. The core idea of the extended Kalman filter is to transform the nonlinear problem into a linear problem through Taylor expansion. The calculation is simple but not suitable for strong nonlinear problems. The unscented Kalman filter transforms the nonlinear problem into a linear problem through UT transformation based on the Kalman filter. The cubature Kalman filter approximates the integral of the nonlinear function through high-order integration rules, so as to estimate the state and covariance more accurately. The core idea of particle filtering is to use a set of random samples to represent the posterior probability distribution of the state, and to approximate the state distribution of the system by propagating and updating these particles in the state space. Considering the accuracy, computational complexity and stability, the unscented Kalman filter is a good choice as the base filter. Although its computational complexity is slightly higher than that of the extended Kalman filter, it is much lower than that of the volumetric Kalman filter and the particle filter. In addition, the accuracy and stability of the unscented Kalman filter are obvious when dealing with highly nonlinear systems.
[0004] In real underwater acoustic engineering applications, acoustic sensors in detection platforms such as sonar are subject to interference from various environmental noises in the ocean and underwater environments such as seabed topography during actual measurement. Therefore, the measurement values provided by the sonar may not conform to the standard Gaussian white noise, which will lead to mismatch problems in the pre-set measurement noise model.
[0005] To this end, the present invention proposes a variational Bayesian unscented Kalman filtering algorithm based on M estimation for unknown measurement noise and measurement noise that is non-Gaussian. The algorithm uses a variational strategy to estimate the real-time distribution of measurement noise in real time, and uses the M estimation method to pre-process the measurement value during the variational iteration process to reduce the impact of wild values and non-Gaussian measurements on the algorithm. And the introduction of singular value decomposition instead of Cholesky to decompose the covariance matrix can ensure the stability of the iterative process of the filtering algorithm. Simulation experiments show that compared with the existing unknown measurement noise adaptive model, the method provided by the present invention has higher tracking accuracy and certain practical significance. Summary of the invention
[0006] The technical problem solved by the present invention is to propose a variational Bayesian unscented Kalman filter target tracking method based on M estimation.
[0007] The present invention provides a variational Bayesian unscented Kalman filter target tracking method based on M estimation, the method comprising:
[0008] S1. Starting from the actual scenario, considering the impact of unknown measurement noise on tracking accuracy during target tracking, the distribution of model measurement noise is approximated through variational strategy to achieve accurate target estimation;
[0009] S2. For the non-Gaussian data that appear in the actual measurement process, the M estimation method is used to pre-process the observed data, which is beneficial to improve the accuracy of data prediction;
[0010] S3. Singular value decomposition is used to solve the problem of covariance matrix divergence in the iteration process of the traditional unscented Kalman filter algorithm and improve the stability of the algorithm.
[0011] Preferably, in S1, by analyzing the actual ocean environment, it is known that the ocean environment noise conforms to the different degrees of freedom x 2 Distribution, so the inverse Wishart distribution is used to model the distribution of measurement noise to obtain the distribution form of measurement noise, which is expressed as follows:
[0012]
[0013] For the measurement noise covariance matrix R t Then by solving IW(R t ∣η t ,γ t ) to get R t =diag([σ t,1 ,σ t,2 ,…,σ t,n ]), σ t,iis the diagonal element of the measurement matrix, and its calculation formula is as follows:
[0014]
[0015] η t,i =η t-1,i +1
[0016] γ t,i =γ t-1,i +0.5[(z t -Hx t )(z t -Hx t ) T +Hx t H T ]
[0017] Preferably, in S2, the M estimation method is used to construct a residual function with a lower growth rate to obtain the parameters to be estimated, and the Huber loss function is used to deal with non-Gaussian data and outlier problems in the measured values, and its Huber expression is as follows:
[0018]
[0019] Among them, p(·) represents the weight function, τ is the adjustable coefficient, ε t is the residual ε t The i-th component of t =Z t -h(x).
[0020] In order to reduce the impact of outliers in the observed data on other observed data, the Huber loss function is minimized, and the weight function W(ε t ), which is expressed as follows:
[0021]
[0022] Use the weight function to calculate the current weight and calculate the weighted residual It is calculated as follows:
[0023]
[0024] Preferably, in S3, the positive definiteness of the covariance matrix needs to be ensured when UKF generates Sigma points, but the positive definiteness of the covariance matrix will be disturbed by noise in the actual iteration process, so singular value decomposition is used instead of Cholesky decomposition to enhance the stability of the unscented Kalman filter algorithm. Its expression is as follows:
[0025] P=U t-1 Λt-1 V t-1 T
[0026] Where U t-1 and V t-1 denote the left singular value vector matrix and the right singular value vector matrix of P, respectively, t The singular value matrix is called a diagonal matrix.
[0027] Beneficial effects of the present invention. The present invention first gives a two-dimensional rectangular coordinate system in the process of target tracking, with the observation station as the origin, the due north direction as the y-axis, and the due east direction as the x-axis. A variational strategy is introduced to solve the situation where the measurement noise is unknown during the observation process, and the M estimation method is combined to deal with the non-Gaussian problem and wild value problem that may appear in the observation value during the measurement process. Finally, the singular value decomposition algorithm is used to solve the problem of instability of the algorithm during the iteration process. Simulation experiments are used, with the root mean square error as the evaluation index, and the method proposed by the present invention is compared with the algorithm proposed by predecessors, proving the feasibility and effectiveness of the variational Bayesian unscented Kalman filter target tracking method based on M estimation for target tracking. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 Overall diagram of the implementation approach
[0029] Figure 2 Yes Observation Model Diagram
[0030] Figure 3 Tracking comparison chart
[0031] Figure 4 is the position error curve
[0032] Figure 5 This is a comparison of the position errors in the X and Y directions.
[0033] Figure 6 This is the speed error comparison diagram between the X direction and the Y direction. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0036] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.
[0037] This embodiment provides a variational Bayesian unscented Kalman filter target tracking method based on M estimation. This method estimates the observation noise, then uses the M method to pre-process the measurement noise, and solves the problem of filter divergence in the iterative process through singular value decomposition. The root mean square error is selected as the evaluation index, and a simulation experiment is carried out. The overall block diagram of this embodiment is shown in the figure. Figure 1 As shown, the method of this embodiment includes:
[0038] S1. Use the variational strategy to estimate the measurement noise, and obtain the posterior distribution of the measurement noise through the variational strategy to solve the problem of performance degradation in traditional target tracking algorithms when the noise does not meet the assumptions. The variational Bayesian method can effectively reduce the reliance on the noise model assumptions by optimizing the approximation function of the posterior distribution, thereby improving the robustness and stability of the algorithm. The measurement noise solution is as follows:
[0039] R t =diag([σ t,1 ,σ t,2 ,…,σ t,n ])
[0040]
[0041] η t,i =η t-1,i +1
[0042] γ t,i =γ t-1,i +0.5[(z t -Hx t )(z t -Hx t ) T +Hx t H T ]
[0043] Because the measurement noise matrix itself has a certain time-varying characteristic, a time-varying factor ρ∈(0,1] is introduced, which represents the fluctuation degree of the measurement noise between two moments. The closer the value is to zero, the greater the fluctuation degree is, and the closer it is to 1, the smaller the fluctuation degree is.
[0044] S2. Homogenize the observation station and the maneuvering target, and establish a two-dimensional rectangular coordinate system with the observation station position as the origin. The observation station position and the target position are as follows: Figure 2 As shown. For the observation values obtained by sonar and radar, the M estimation theory is used to pre-process the observation data:
[0045] In order to get closer to the actual measurement value, the measurement noise that changes over time is used to simulate the actual situation. The expression is as follows:
[0046]
[0047] Among them, r=10.
[0048] The time-varying measurement noise is used to simulate the observation environment. Two methods are used to simulate the environmental noise. The expressions are as follows:
[0049]
[0050] Among them, p1=0.01, p2=0.05.
[0051] Using the above observations to simulate the real environment will be closer to the real observation distribution. Therefore, in order to reduce the impact of outliers in the observation data on other observations, the Huber function is derived to obtain the weight function W(ε t ), which is expressed as follows:
[0052]
[0053] Among them, p(·) represents the weight function, τ is the adjustable coefficient, ε t is the residual ε t The i-th component of t =Z t -h(x).
[0054] Use the weight function to calculate the current weight and calculate the weighted residual It is calculated as follows:
[0055]
[0056] The weighted residual method can dynamically adjust the size of the weight, and can handle deviations from Gaussian distribution, spike noise or other forms of non-Gaussian noise. Small residuals correspond to larger weights, and large residuals correspond to smaller weights. In this way, estimates can be made based on most normal observations, while ignoring or reducing the impact of outliers, thereby improving the accuracy and stability of the estimate.
[0057] S3. When using singular value decomposition instead of unscented Kalman filtering to ensure the positive definiteness of the covariance matrix, the expression of the singular value decomposition is as follows:
[0058] P=U t-1 Λ t-1 V t-1 T
[0059] Where Ut-1 and V t-1 denote the left singular value vector matrix and the right singular value vector matrix of P, respectively, t The singular value matrix is called a diagonal matrix.
[0060] The new Sigma point is obtained by using the singular value information of the matrix as follows:
[0061] χ t-1 =[χ 0 ,χ1 i ,χ2 i ]
[0062]
[0063] Table 1M-VBUKF algorithm flow
[0064]
[0065] Experimental data processing and analysis
[0066] 1. Experimental environment
[0067] The variational Bayesian unscented Kalman filter target tracking method based on M estimation proposed in this embodiment is implemented by Matlab programming. The experimental equipment uses a processor Inter(R)Core(TM)i7-9700 CPU@3.00GHz, 32G installed memory, 64-bit Windows 10 operating system, and GPU model GEFORCE RTX 2080Ti.
[0068] 2. Evaluation indicators
[0069] The root mean square error is used to evaluate the performance of the algorithm. It is a statistical indicator used to measure the prediction accuracy of the model. It is used to evaluate the deviation between the predicted value of the model and the actual value. Due to the existence of the square term, the abnormal value of RMSE is very sensitive and can intuitively explain and understand the prediction accuracy of the model. Its expression is as follows:
[0070]
[0071] Among them, Y i represents the true value, f(x i ) represents the predicted value, N represents the number of Monte Carlo experiments and the experimental parameter
[0072] 3. Experimental parameters
[0073] The experiment adopts a uniform motion model. The initial state of the maneuvering target is X = [2000m, -10m / s, 1500m, -6m / s]. A two-dimensional rectangular coordinate system is established with the observation station as the origin. The observation station provides the distance and angle information of the maneuvering target from the observation station. Assume that the process noise of the maneuvering target is Q = diag ([5, 0.1, 5, 0.1]), the initial covariance matrix P = diag ([1000, 10, 1000, 10]), and the initial variation parameter η t-1 =[0,0],γ t-1 =diag([150,0.1]). Since the process noise covariance fluctuates greatly, ρ is set to 0.85. The non-Gaussian measurement noise is set to r = 10; the total simulation time is 200s, the sampling time T = 1s. In order to better test the performance of the algorithm, the noise changes on both sides during the tracking process are as follows: Among them, p1=0.01, p2=0.05.
[0074] 4. Simulation and result analysis
[0075] (1) Monte Carlo simulations were performed on the target tracking system 100 times, and a target tracking position comparison diagram was randomly selected. The M-VBUKF was compared with the existing filters, including the tracking results of the UKF algorithm with real measurement noise input and the A-SRTCKF algorithm in the entire tracking interval. Figure 3 shown.
[0076] Depend on Figure 3 It can be seen that in the entire maneuvering target tracking interval, the three filtering algorithms can stably track the target, and the prediction effect is better than the observation value. Combined with the local enlarged image near point A (550, 650) and point B (1250, 1000) for local analysis, it can be found that the predicted value of the M-VBUKF algorithm proposed in the present invention is closer to the true value at point A, and the observations of the UKF algorithm and A-SRTCKF with real measurement noise have certain fluctuations. For point B, a local analysis shows that the measurement noise at point B is higher than that at point A, but it can be seen from the figure that the algorithm proposed in the present invention can still maintain a prediction result similar to that of the UKF with real measurement noise for strong noise interference. Therefore, it can be seen from the position comparison diagram of maneuvering target tracking that the algorithm proposed in the present invention has a better tracking effect for non-Gaussian uncertain measurements and higher tracking accuracy.
[0077] (2) In order to analyze the performance of the M-VBUKF algorithm more intuitively, the root mean square error is used as the evaluation index of the algorithm performance, and the error curves of the three algorithms, M-VBUKF, UKF with real measurement noise input, and A-SRTCKF, are plotted. Figure 4 Position error curves of the three algorithms. Figure 5 The position error curves of the three algorithms in the X and Y directions are shown in Figure 2. Figure 6 The speed error curves of the three algorithms.
[0078] Depend on Figure 4 and Figure 5 It can be seen that the prediction errors of the three algorithms, M-VBUKF, A-SRTCKF and UKF with real measurement input, are all less than about 50% of the observed value, and in the entire tracking interval, it can be found that the error value of the algorithm proposed by the present invention is smaller than the prediction result of UKF with real measurement input in most of the time. Figure 6 From the velocity errors in the X and Y directions, it can be seen that the maximum velocity errors of M-VBUKF in both directions are smaller than those of the other two algorithms, and it has a faster convergence accuracy, while the overall error value is smaller. Therefore, the analysis of position error and velocity error shows that the M-VBUKF algorithm has higher tracking accuracy and convergence speed.
[0079] (3) Based on (2), the tracking error of the system is quantified, and the mean values of position error and velocity error of the three algorithms, M-VBUKF, A-SRTCKF and UKF with real measurement input, are obtained during the entire tracking period, as shown in Table 1.
[0080] Table 1 Mean error of speed and position
[0081]
[0082] As shown in the data in Table 2, the M-VBUKF algorithm proposed by the present invention has a smaller error mean for position error, and has significantly improved prediction effects in the X and Y directions compared with the A-SRTCK algorithm, with the error values both increased by about 7m, which is still about 1m higher than the UKF algorithm with real measurement input. For the speed direction, the filtering accuracy of the M-VBUKF algorithm is improved by 0.368m / s and 0.167m / s in the X direction, and by 0.454m / s and 0.152m / s in the Y direction, respectively, compared with the other two algorithms. Figure 3 The error curve and the mean speed and position error in Table 2 verify that the M-VBUKF algorithm still has high tracking accuracy under the condition of unknown measurement noise, and the tracking effect is improved to a certain extent compared with the two comparative experiments.
Claims
1. A variational Bayesian unscented Kalman filter target tracking method based on M estimation, characterized in that: include: S1. Starting from the actual scenario, considering the impact of unknown measurement noise on tracking accuracy during target tracking, the distribution of model measurement noise is approximated through variational strategy to achieve accurate target estimation; S2. For the non-Gaussian data that appear in the actual measurement process, the M estimation method is used to pre-process the observed data, which is beneficial to improve the accuracy of data prediction; S3. Singular value decomposition is used to solve the problem of covariance matrix divergence in the iteration process of the traditional unscented Kalman filter algorithm and improve the stability of the algorithm.
2. The variational Bayesian unscented Kalman filter target tracking method based on M estimation according to claim 1, characterized in that: In S1: Through the analysis of the actual ocean environment, it is known that the ocean environmental noise conforms to the different degrees of freedom χ 2 Distribution, so the inverse Wishart distribution is used to model the distribution of measurement noise to obtain the distribution form of measurement noise, which is expressed as follows: For the measurement noise covariance matrix R t Then by solving IW(R t ∣η t ,γ t ) to get R t =diag([σ t,1 ,σ t,2 ,…,σ t,n ]), σ t,i is the diagonal element of the measurement matrix, and its calculation formula is as follows: or t,i =the t-1,i +1 γ t,i =γ t-1,i +0.5[(z t -Hx t )(z t -Hx t ) T +Hx t H T ] 3. The variational Bayesian unscented Kalman filter target tracking method based on M estimation according to claim 1, characterized in that: In S2: The M estimation method is used to construct a residual function with a lower growth rate to obtain the parameters to be estimated, and the Huber loss function is used to deal with non-Gaussian data and outlier problems in the measured values. The Huber expression is as follows: Among them, p(·) represents the weight function, τ is the adjustable coefficient, ε t is the residual ε t The i-th component of t =Z t -h(x). In order to reduce the impact of outliers in the observed data on other observed data, the Huber loss function is minimized, and the weight function W(ε t ), which is expressed as follows: Use the weight function to calculate the current weight and calculate the weighted residual It is calculated as follows: The weighted residual method can dynamically adjust the size of the weight, and can handle deviations from Gaussian distribution, spike noise or other forms of non-Gaussian noise. Small residuals correspond to larger weights, and large residuals correspond to smaller weights. In this way, estimates can be made based on most normal observations, while ignoring or reducing the impact of outliers, thereby improving the accuracy and stability of the estimate.
4. The variational Bayesian unscented Kalman filter target tracking method based on M estimation according to claim 1, characterized in that: In the S3: For the traditional unscented Kalman filter algorithm, the positive definiteness of the covariance matrix needs to be guaranteed when generating Sigma points. However, in the actual iteration process, the positive definiteness of the covariance matrix will be disturbed by noise. Therefore, singular value decomposition is used instead of Cholesky decomposition to enhance the stability of the unscented Kalman filter algorithm. Its expression is as follows: P t-1 =U t-1 Λ t-1 V t-1 T Where U t-1 and V t-1 Respectively represent P t-1 The left singular value vector matrix and the right singular value vector matrix of , Λ t The singular value matrix is called a diagonal matrix.
Citation Information
Cited By
GNSS / IMU (Global Navigation Satellite System / Inertial Measurement Unit) loose combination positioning method based on improved RSUKF smoothing algorithm
CN121165141A
