Maneuvering target tracking method based on Gaussian process

Through the regression model based on Gaussian process learning target motion state online, the problem of mismatch between the motion model and the target motion law in target tracking is solved, better adaptability and robustness are achieved, and the target tracking effect is improved.

CN119048561AActive Publication Date: 2024-11-29HANGZHOU DIANZI UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411176204.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-29
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

When the existing target tracking technology maneuveres, the motion model does not match the target motion law, resulting in tracking failure and relies on prior information, lacking adaptability and robustness.

Method used

Using a regression model based on Gaussian process, predicts the target motion trajectory and position by learning online target motion state and measurement error, uses Gaussian process motion tracker (GPMT) to capture the complexity and uncertainty of nonlinear dynamic systems, and automatically adjusts the model Parameters to adapt to the target motion characteristics.

Benefits of technology

It improves the adaptability and robustness of target tracking, avoids model mismatch problems, can perform effective model learning on a smaller training data set, reduces dependence on prior information, and improves tracking effect and practicality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048561A_ABST
    Figure CN119048561A_ABST
Patent Text Reader

Abstract

The invention discloses a maneuvering target tracking method based on a Gaussian process. Firstly, a single maneuvering target tracking data set is obtained; then regression prediction based on a Gaussian process is carried out; based on recursive estimation of a Gaussian process, obtaining estimation corresponding to prediction information of the test set; recursively updating training set data and a training set sliding window through a Gaussian process; and finally, carrying out next circulation on the obtained updated data until a set number of cycles is reached, thereby obtaining an optimized model to predict the motion trail and position of the target. According to the method, selection of a fixed model is avoided, the model is established through online learning, and higher robustness is achieved. According to the method, complexity and uncertainty in a nonlinear dynamic system can be captured, effective model learning can be carried out on a small training data set, and compared with the prior art in the field of single maneuvering target tracking, a better tracking effect is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of target tracking, and in particular relates to a mobile target tracking method based on Gaussian process. Background Art

[0002] In common target tracking technology, the main steps can be divided into two parts: the establishment of motion model and the selection of filtering algorithm. The establishment of target motion model is the first step to solve the target tracking problem, and the selection of model is an important factor in determining the target tracking effect. The basic motion models of the target include constant velocity model (CV), constant acceleration model (CA), constant turn rate model (CT) and Singer model. The above models have good tracking performance when the target does not maneuver during motion and the system model matches the motion law. However, when the target maneuvers, if a single model is still used to model the motion trajectory of the maneuvering target, the mismatch between the adopted motion model and the target motion law will lead to the failure of the tracking task.

[0003] To solve the problem of motion model mismatch, the MM method in mechanical modeling and the data-driven method are commonly used. Magill proposed the MM algorithm, the main idea of ​​which is to use a model set, the models in the set can describe the entire motion state of the target. Later, in order to obtain better tracking results, Bar-Shalom believed that the transition between models obeys Markov transition, so he introduced the Markov transition probability matrix (TPM) for model transfer, and proposed the Interacting Multiple Model (IMM) algorithm based on the generalized pseudo-Bayesian algorithm and the MM algorithm [1]. Compared with the MM algorithm, the IMM algorithm is better in tracking results, with better robustness and tracking stability. However, the IMM algorithm still has limitations, is more dependent on prior information, and has the disadvantage of model switching delay.

[0004] To improve the IMM algorithm, the goal is to improve the matching degree between the motion model and the target motion law. There are two main ideas: model set adaptive adjustment and TPM adaptive adjustment. In terms of model set adaptive adjustment, XRLi and Bar-Shalom proposed the variable structure interactive multi-model [2] (VSIMM) algorithm. Later, XRLi and other scholars proposed the possible model set [3] (LMS), which can adaptively increase or decrease models to better adapt to the motion characteristics of the target, thereby improving the tracking accuracy. Later, algorithms such as expected model augmentation [4] (EMA), optimal model augmentation [5] (BMA), and simplified equivalent model interactive multi-model [6] (SEIMM) were proposed. In terms of TPM adaptive adjustment, the current TPM is usually obtained by multiplying the previous TPM by a correction factor, which is still heavily dependent on previous information. In response to this, the literature [7] proposed the ATPM-IMM (adaptive TPM-IMM) [7] algorithm, which uses a Bayesian framework to adaptively update the TPM according to the measurement sequence, thereby reducing the dependence on previous information.

[0005] The above is the development process of the algorithm based on the mechanism model. It can be seen that although the mechanism model is highly interpretable, the model or model set is fixed, and a large amount of correct prior information needs to be mastered in advance. If the prior information is insufficient or wrong, there will be a mismatch between the system model and the law of motion. This topic is proposed to solve the model mismatch problem.

[0006] This paper mainly studies the application of Gaussian process in maneuvering target tracking.

[0007] References:

[0008] [1]Blom H,Bar-Shalom Y.The interacting multiple model algorithm forsystems with Markovian switching coefficients[J].IEEE Transactions onAutomatic Control,1988,33(8):780-783.

[0009] [2]Li XR, Bar-Shalom Y.Multiple-Model Estimation with VariableStructure[J]. IEEE Transactions on Automatic Control, 1996, 41(4): 478-493.

[0010] [3]Xiao-Rong Li and Zhang Y M.Multiple model estimation with variablestructure-Part V:Likely-model set algorithm[J].IEEE Transactions on Aerospaceand ElectronicSystems,2000,36(2):448-466.

[0011] [4]Xiao-Rong Li, Vesselin P Jilkov, Jifeng Ru.Multiple model estimation with variable structure-Part V: Expected-model augmentation algorithm[J]. IEEE Transactions on Aerospace and Electronic Systems, 2005, 41(3):853-867.

[0012] [5]Lan J, Xiao-Rong Li, Mu C. Best-model augmentation for variable-structure multiple-model estimation[J]. IEEE Transactions on Aerospace and Electronic Systems, 2011, 47(3): 2008-202.

[0013] [6]Lan J and Xiao-Rong Li. Equivalent-model augmentation for variable-structure multiple-model estimation[J]. IEEE Transactions on Aerospace and Electronic Systems, 2013, 49(4): 2615-2630.

[0014] [7] Zeng Hao, Mu Wangqiang, Yang Shunping. ATPM-IMM algorithm for high maneuverability target tracking[J]. Journal of Communications, 2022, 43(07): 93-101. Summary of the invention

[0015] The present invention utilizes the online learning capability of Gaussian process regression to learn the target's potential motion model online and obtain a better tracking effect, thereby solving the model mismatch problem of traditional algorithms.

[0016] In view of the deficiencies in the prior art, the present invention provides a maneuvering target tracking method based on Gaussian process, which is called Gaussian process regression tracking algorithm (GPR-MT).

[0017] The present invention utilizes the Gaussian Processes Motion Tracker (GPMT) concept. GPMT is a method for modeling the motion state of a target based on a Gaussian process regression model. The algorithm learns the target motion state and measurement error by training a set of Gaussian process regression models, thereby predicting the motion trajectory and position of the target.

[0018] The advantage of GPMT is that it can capture the complexity and uncertainty in nonlinear dynamic systems and can effectively learn models on smaller training data sets. In addition, the algorithm has good adaptive performance and can automatically adjust model parameters to adapt to different motion characteristics of the target. This method is based on the following assumptions:

[0019] 1) The cross-coordinate coupling is so weak that it can be ignored;

[0020] 2) Coordinate autocorrelations are available in a timely manner;

[0021] 3) weak temporal correlation with input points in the distant past;

[0022] 4) The measurement noise variance is the same for all measurements.

[0023] From the above assumptions, it can be seen that when this method is used to locate and track the target, the dimension of the target will not affect the implementation of the filter, that is, the calculation method of all coordinates is the same, only the data is different. The GPR-MT of the present invention is also based on the above assumptions. The position prediction and update equations of all output dimensions are similar, so the steps designed below take one dimension as an example.

[0024] During the process, the measured values ​​are stored and used for GP training. p,q represents the GP covariance matrix between the input vectors p and q, the n-dimensional vector u and the l-dimensional u ★ The GP covariance matrix between vectors is in the form:

[0025]

[0026] The specific implementation plan is as follows:

[0027] A method for tracking a maneuvering target based on a Gaussian process comprises the following steps:

[0028] Step 1: Obtain a single maneuvering target tracking data set;

[0029] The single maneuvering target tracking dataset is randomly generated based on a single maneuvering target within the limited range of the actual application scenario, and the generated dataset is divided into a training set and a test set.

[0030] Step 2: Regression prediction based on Gaussian process;

[0031] The training set and test set in the single maneuvering target tracking dataset are input into the maneuvering target tracking model for regression prediction based on Gaussian process;

[0032] Step 3: Based on the recursive estimation of the Gaussian process, obtain the estimation corresponding to the prediction information of the test set;

[0033] Step 4: The Gaussian process recursively updates the training set data and the training set sliding window;

[0034] Step 5: The updated training set data and test set data are cycled again until the set number of cycles is reached. The final updated training set and test set are obtained. Thus, an optimized model is obtained to predict the target's motion trajectory and position.

[0035] Further, the method of step 2 is as follows:

[0036] Using K SE The (x,x') kernel function is the square exponential function, also known as the covariance function.

[0037] In the first loop, the measurement vector z is obtained from the dataset k , check whether the sliding window length d is reached. If the length is reached, calculate the autocorrelation matrix K uu , otherwise continue to obtain measurements. Get K uu Then calculate the gain J0 and the prediction covariance B0 to get the predicted mean at the initial moment (Right now ) and the predicted covariance (i.e. B0). Then, the estimated gain M0 and mean estimate are calculated. Variance estimation Then, since this is the first cycle and there is no data updated last time, the training set obtained by updating the training set is still the original training set. At this point, the first cycle is completed and the next cycle begins. The following is a detailed description of the subsequent k-th cycle operation:

[0038] Use the training set data obtained at the k-1 moment (i.e. the updated training set obtained in the previous cycle) and According to Gaussian process regression, the mean prediction and variance prediction of the coordinates on one dimension of the target are obtained:

[0039]

[0040]

[0041] Among them, f represents the training set, Represents the test set. is the measurement vector z at time k k The test set input vector, u is the measurement vector z at time k k The training set input vector. is the input vector and the GP covariance matrix between the input vector u. is the inverse of the GP covariance matrix of the input vector u. is the input vector The GP covariance matrix of J k and B k is the matrix derived from GP regression; J k Similar to the gain in the Kalman filter, it is the cross-covariance between the test set input and the training set input calculated by the kernel function at time k; B k is the variance of the uncertainty of regression at time k, that is, the prediction covariance. and They are k moments respectively exist The updated forecast mean and covariance at . m( ) represents the mean. Same as Gaussian process regression, mean Take 0. is the estimated mean of the training set at time k-1. is the covariance matrix of the training set at k-1 moment.

[0042] Further, the specific method of step 3 is as follows;

[0043] Calculate the corresponding estimate based on the test set prediction information obtained in step 2:

[0044]

[0045] Among them, M k is the Kalman gain matrix, obtained from step 2 Calculated. 2 is a hyperparameter measuring the variance of the noise, and I is the identity matrix. and The corresponding inputs are The estimation of the mean and variance of the test set at the time of , respectively, is obtained from step 2 and The predicted value at time K is calculated.

[0046] Furthermore, the specific method of step 4 is as follows:

[0047] Adopt sliding window update: First, advance the input of the training set by one time, that is, discard the content corresponding to the farthest time in the training set. Assume that the training data earlier than the most recent d measurement samples are irrelevant and deleted, t k =[kd,k-d+1,…,k-1] is the sample position in the sliding window at the previous moment. k =[k-d+1,kd,…,k] is the sample position in the sliding window after update.

[0048] Then, the predicted mean of the test set at the current time k obtained in step 2 is and the measurement vector z at the current time k k The residual of the training set at k-1 time The specific calculation steps are as follows:

[0049]

[0050] Among them, G k is the updated training set gain matrix, and the mean and variance of the training set are estimated by and is the covariance matrix of the training set at the k-1 (i.e., the previous moment) moment. is the mean of the training set at the k-1 (previous moment) moment.

[0051] To update the covariance matrix, you need to first calculate the cross-covariance between the training set f and the test set f

[0052]

[0053] The covariance matrix corresponding to the updated training set output is obtained by using the cross-covariance, thus completing an update operation.

[0054] The beneficial effects of the present invention are as follows:

[0055] ① Different mechanism models have different applicable scenarios. For mechanism modeling, how to select a suitable model according to actual conditions is an important issue in determining excellent tracking effects. However, the GPR-MT method of the present invention avoids this problem and has better adaptability and scalability.

[0056] ② For a fixed single model, when the target maneuvers, a mechanism model mismatch occurs, resulting in poor tracking performance or even tracking failure. However, the GPR-MT method of the present invention avoids the selection of a fixed model and establishes a model through online learning, which has stronger robustness.

[0057] ③ The mechanism model algorithm will introduce more parameters and prior information, such as process noise Q k , measurement noise R k In practical applications, it is difficult to accurately obtain the prior information corresponding to these parameters, resulting in low practicality of the algorithm based on the mechanism model. GP is a parameter-free method, so the GPR-MT method of the present invention avoids this problem and has better practicality.

[0058] ④ The GPR-MT method of the present invention can capture the complexity and uncertainty in nonlinear dynamic systems, and can perform effective model learning on a smaller training data set, achieving better tracking effects than the existing technologies in the field of single maneuvering target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of the algorithm of the embodiment of the present invention;

[0060] Figure 2 It is a schematic diagram of tracking results and error examples based on Gaussian process;

[0061] Figure 3 It is a sample update method, that is, a schematic diagram of a sliding window update method. DETAILED DESCRIPTION

[0062] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0063] like Figure 1 The figure shows a specific block diagram of the GPR-MT of the present invention, a Gaussian process-based maneuvering target tracking method GPR-MT, comprising the following steps:

[0064] Step 1: Regression prediction based on Gaussian process;

[0065] Different types of kernel functions are suitable for different data models and application scenarios, and need to be selected and adjusted according to specific problems. SE (x,x') kernel function, also known as square exponential function (SquareExponential), can also be called covariance function (covariance function). The specific form is:

[0066]

[0067] k( x, x ') defines the similarity between pairs of sample points. Taking a two-dimensional target as an example, the feature length scale l of the two input dimensions x and y x , l y The diagonal matrix is ​​Λ=(l1,l2,…,l d ), σ 2 is the variance of the underlying nonlinear function f, x represents the input vector of the training set, x ' represents the current test input vector. These parameters of the mean and covariance functions together with the hyperparameters of the measurement noise variance are called the hyperparameters of the GP.

[0068] In the first loop, the measurement vector z is obtained from the dataset k , check whether the sliding window length d is reached. If the length is reached, calculate the autocorrelation matrix K uu , otherwise continue to obtain measurements. Get K uu Then calculate the gain J0 and the prediction covariance B0 to get the predicted mean at the initial moment (Right now ) and the prediction variance (i.e. B0). Then, the estimated gain M0 and mean estimate are calculated. Variance estimation Then, since this is the first cycle and there is no data updated last time, the training set obtained by updating the training set is still the original training set. At this point, the first cycle is completed and the next cycle begins. The following is a detailed description of the subsequent k-th cycle operation:

[0069] Use the training set data obtained at the k-1 moment (i.e. the updated training set obtained in the previous cycle) and According to Gaussian process regression, the mean prediction and variance prediction of the coordinates on one dimension of the target are obtained:

[0070]

[0071] Among them, f represents the training set, represents the test set. is the measurement vector z at time k k The test set input vector, u is the measurement vector z at time k k The training set input vector. is the input vector and the GP covariance matrix between the input vector u. is the inverse of the GP covariance matrix of the input vector u. is the input vector The GP covariance matrix of J k and B k is the matrix derived from GP regression; Jk Similar to the gain in the Kalman filter, it is the cross-covariance between the test set input and the training set input calculated by the kernel function at time k; B k is the variance of the uncertainty of regression at time k, that is, the prediction covariance. and They are k moments respectively exist The updated forecast mean and covariance at . m( ) represents the mean. Same as Gaussian process regression, mean Take 0. is the estimated mean of the training set at time k-1. is the covariance matrix of the training set at k-1 moment.

[0072] Step 3: Based on the recursive estimation of the Gaussian process, obtain the estimation corresponding to the prediction information of the test set;

[0073] This step is similar to the update step in Kalman filtering, i.e., the state estimation step, as the actual measurement of the corresponding target. The corresponding estimate is calculated based on the test set prediction information obtained in step 2:

[0074]

[0075] Among them, M k is the Kalman gain matrix, obtained from step 2 Calculated. 2 is a hyperparameter measuring the variance of the noise, and I is the identity matrix. and The corresponding inputs are The estimation of the mean and variance of the test set at the time of , respectively, is obtained from step 2 and The predicted value at time K is calculated.

[0076] Step 4: The Gaussian process recursively updates the training set data and the training set sliding window;

[0077] In the application of Gaussian processes in the field of maneuvering target tracking, in order to more accurately describe the target motion, the training set needs to be updated recursively online. This paper adopts a sliding window update: first, the input of the training set is advanced by one moment, that is, the content corresponding to the farthest moment in the training set is discarded. It is assumed that the training data older than the most recent d measurement samples are irrelevant and deleted. This idea determines the depth or sliding window length of the tracker. Figure 3 As shown, where t k =[kd,k-d+1,…,k-1] is the sample position in the sliding window at the previous moment. k =[k-d+1,kd,…,k] is the sample position in the sliding window after update.

[0078] Then, the predicted mean of the test set at the current time k obtained in step 2 is and the measurement vector z at the current time k k The residual of the training set at k-1 time The specific calculation steps are as follows:

[0079]

[0080] Among them, G k is the updated training set gain matrix, and the mean and variance of the training set are estimated by and is the covariance matrix of the training set at the k-1 (i.e., the previous moment) moment. is the mean of the training set at the k-1 (previous moment) moment.

[0081] To update the covariance matrix, you need to first calculate the training set f and the test set The cross-covariance

[0082]

[0083] The covariance matrix corresponding to the updated training set output is obtained by using the cross-covariance, thus completing an update operation.

[0084] Step 5: loop the updated training set data and test set data k+1 times until the set number of loops is reached. Finally, the updated training set and test set are obtained. Thus, the optimized model is obtained to predict the target's motion trajectory and position.

[0085] Figure 2 The following are the tracking results and error examples based on Gaussian process. The starting point of the target is (-1000m, 2000m). The target moves in a uniform straight line from 0 to 30s, with a speed of 200m / s in the x direction and -80m / s in the y direction. The acceleration changes from 30 to 100s, and a sharp turn is made. The acceleration is ax = -2m / s, ay = 4m / s. The radar scanning period is T = 1s, and the standard deviation of the observation noise is 100m. The software simulation is completed in Matlab_R2020b software.

[0086] Tracking performance is evaluated by the root mean square error, which is a metric used to measure the difference between the model's predicted values ​​and the true values.

[0087] Mean Squared Error (MSE), the average of the squares of the differences between the predicted value and the true value, is given by the following formula:

[0088]

[0089] Root Mean Squared Error (RMSE) is based on MSE and squared. The formula is as follows:

[0090]

[0091] from Figure 2 It can be seen that in this experimental scenario, the GPR-MT algorithm has better tracking effect than the IMM algorithm, and the RMSE fluctuation of the GPR-MT algorithm is smaller and the peak error is lower, that is, the tracking effect is more stable and more robust.

[0092] Table 1 lists the RMSE ranges of each algorithm in different time periods to better compare the tracking accuracy of other algorithms. Since filtering actually starts from the third second, in order to be more convincing, the actual RMSE range of the first section starts from the fifth second; in order to make it easier to compare and highlight the tracking accuracy gap between different algorithms, the time period with large fluctuations is not considered here, and the second section takes the range when the RMSE is relatively stable, starting from about sixty seconds.

[0093]

[0094] Table 1

[0095] It can be clearly seen from Table 1 that before the target maneuvers, all algorithms can track successfully, but the RMSE of the CT-KF algorithm is significantly larger, and the CV-KF and Singer-KF algorithms perform best in terms of RMSE. After the target maneuvers, the CT-KF and GP-based algorithms perform best in terms of RMSE. Overall, the GPR-MT and IMM algorithms have the most stable and excellent performance in terms of RMSE in this experimental scenario.

[0096] The above contents are further detailed descriptions of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can also make several substitutions or modifications to these described embodiments without departing from the concept of the present invention, and these substitutions or modifications should be regarded as belonging to the protection scope of the present invention.

[0097] Parts of the present invention that are not described in detail belong to the well-known technologies of those skilled in the art.

Claims

1. A maneuvering target tracking method based on Gaussian process, characterized in that: The steps include: Step 1: Obtain a single maneuvering target tracking data set; The single maneuvering target tracking dataset is randomly generated based on the single maneuvering target within the limited range of the actual application scenario, and the generated dataset is divided into a training set and a test set; Step 2: Regression prediction based on Gaussian process; The training set and test set in the single maneuvering target tracking dataset are input into the maneuvering target tracking model for regression prediction based on Gaussian process; Step 3: Based on the recursive estimation of the Gaussian process, obtain the estimation corresponding to the prediction information of the test set; Step 4: The Gaussian process recursively updates the training set data and the training set sliding window; Step 5: The updated training set data and test set data are cycled again until the set number of cycles is reached; Get the final updated training set and test set; Thereby obtaining an optimized model to predict the motion trajectory and position of the target.

2. A Gaussian process-based maneuvering target tracking method according to claim 1, characterized in that: Step 2: Using K SE (x,x') kernel function, in the first loop, obtains the measurement vector z from the data set k , check whether the sliding window length d is reached; if the length is reached, calculate the autocorrelation matrix K uu , otherwise continue to obtain the measurement; get K uu Then calculate the gain J0 and the prediction covariance B0 to get the predicted mean at the initial moment and the prediction variance After that, the estimated gain M0 and mean estimate are calculated Variance estimation Then, since this is the first cycle and there is no data updated last time, the training set obtained by updating the training set is still the original training set; at this point, the first cycle is completed and enters the next cycle; the following describes the subsequent k-th cycle operation in detail: Using the training set data obtained at k-1 time and According to Gaussian process regression, the mean prediction and variance prediction of the coordinates on one dimension of the target are obtained: Among them, f represents the training set, represents the test set; is the measurement vector z at time k k The test set input vector, u is the measurement vector z at time k k The training set input vector of is the input vector The GP covariance matrix between and the input vector u; is the inverse of the GP covariance matrix of the input vector u; is the input vector The GP covariance matrix of J k and B k is the matrix derived from GP regression, which is the cross-covariance between the test set input and the training set input calculated by the kernel function at time k; B k is the uncertainty variance of regression at time k, i.e., the prediction covariance; and They are k moments respectively exist The updated forecast mean and covariance at the time; m() represents the mean; Same as Gaussian process regression, mean Take 0; is the estimated mean of the training set at time k-1; is the covariance matrix of the training set at k-1 moment.

3. A Gaussian process-based maneuvering target tracking method according to claim 2, characterized in that: Step 3: The specific method is as follows; Calculate the corresponding estimate based on the test set prediction information obtained in step 2: Among them, M k is the Kalman gain matrix, obtained from step 2 Calculated; σ 2 is a hyperparameter measuring the noise variance, I is the identity matrix; and The corresponding inputs are The estimation of the mean and variance of the test set at the time of , respectively, is obtained from step 2 and The predicted value at time K is calculated.

4. The Gaussian process-based maneuvering target tracking method according to claim 3, characterized in that: Step 4: Adopt sliding window update: First, advance the input of the training set by one time, that is, discard the content corresponding to the farthest time in the training set; assume that the training data earlier than the most recent d measurement samples are irrelevant and delete them; Then, the predicted mean of the test set at the current time k obtained in step 2 is and the measurement vector z at the current time k k The residual of the training set at k-1 time The specific calculation steps are as follows: Among them, G k is the updated training set gain matrix, and the mean and variance of the training set are estimated by and is the covariance matrix of the training set at the k-1 moment; is the mean of the training set at time k-1; To update the covariance matrix, you need to first calculate the training set f and the test set The cross-covariance The covariance matrix corresponding to the updated training set output is obtained by using the cross-covariance, thus completing an update operation.

Citation Information

Patent Citations

  • Method for tracking hypersonic velocity reentry vehicle based on Gaussian mixture approximate

    CN107544067A

  • Anti-occlusion moving target tracking device and method based on ROI prediction and multi-module learning

    CN114972735A

  • Space-time causal mapping driven mixed importance Gaussian filtering method and device

    CN116611232A

  • Variational filtering maneuvering target tracking method based on Gaussian process regression

    CN116662733A

  • Method for tracking an aerial target, and radar implementing such a method

    EP3514571A1