A target motion prediction method, system, device and medium

Abstract

The present invention discloses a method, system, device and medium for predicting the motion of a target, which relates to the field of motion prediction technology and includes the following steps: obtaining a plurality of one-dimensional motion data of the target's basic motion mode in different directions; combining the plurality of one-dimensional motion data in different directions in pairs to generate a two-dimensional position increment, and using a sliding window to select the true values ​​of the two-dimensional position increments of multiple unit time periods in real time, and inputting the true values ​​of the two-dimensional position increments of the first few unit time periods into a Gaussian mixture model, and updating the Gaussian mixture model with the last two-dimensional position increment of the unit time period as the output result; inputting the data to be predicted into the updated Gaussian mixture model to obtain a Gaussian mixture regression prediction result of the motion. The present invention combines a rolling time domain model optimization method to achieve continuous rolling of historical information and model prediction values ​​to optimize the Gaussian mixture model, thereby predicting the motion of the target to be detected, greatly improving the accuracy of the prediction.

Description

Technical Field

[0001] The present invention relates to the field of motion prediction technology, and in particular to a method, system, device and medium for predicting the motion of a target. Background Art

[0002] As self-driving cars, social robots, and intelligent transportation systems become an integral part of daily life, predicting the future trajectory of a target becomes an urgent need.

[0003] Trajectory prediction involves predicting a target's subsequent position based on its motion characteristics. Common trajectory prediction methods include Gaussian mixture models. These models can model the distribution of target motion, adapt to different motion patterns, provide probability density estimates, quantify the uncertainty of predictions, and offer flexibility in handling general nonlinear relationships.

[0004] However, when performing motion prediction, the Gaussian mixture model is often an invariant model, and the prediction results are directly obtained through the input motion data. Due to factors such as time variation, model mismatch, and interference in the actual system, when using an invariant model for motion prediction, it will lead to situations that are inconsistent with the actual situation and the prediction accuracy is low. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide a method, system, device and medium for motion prediction of a target, so as to solve the problem that the prior art often uses an unchanging model to directly obtain the prediction result through the input motion data. Due to factors such as time variation, model mismatch, interference, etc. in the actual system, when motion prediction is performed through an unchanging model, it will lead to situations that are inconsistent with the actual situation and low prediction accuracy.

[0006] The present invention specifically provides the following technical solutions:

[0007] A method for predicting target motion includes the following steps:

[0008] Obtaining several one-dimensional motion data of the target basic motion mode in different directions;

[0009] Combine several one-dimensional motion data in different directions in pairs to generate two-dimensional position increments, and use a sliding window to select the true values ​​of the two-dimensional position increments in multiple unit time periods in real time. The true values ​​of the two-dimensional position increments in the first few unit time periods are input into the Gaussian mixture model, and the last two-dimensional position increment in the unit time period is used as the output result to update the Gaussian mixture model;

[0010] The target data to be predicted is input into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion.

[0011] Preferably, the basic motion modes include uniform motion, uniformly accelerated motion, circular motion and random accelerated motion, and the specific expressions are:

[0012] x1=p 11 t+p 12 ;

[0013] x2=p 21 t 2 +p 22 t+p 23 ;

[0014] x3=p 31 sin(p 32 t)+p 33 ;

[0015] x4=p 41 +∫v x dt,v x =p 42 +∫p 43 *randndt;

[0016] z1=q 11 t+q 12 ;

[0017] z2=q 21 t 2 +q 22 t+q 23 ;

[0018] z3=q 31 sin(q 32 t)+q 33 ;

[0019] z4=q 41 +∫v z dt,v z =q 42 +∫q 43 *randndt;

[0020] Where t is time, x i and z i are the x-direction position and z-direction position of uniform motion, uniform acceleration motion, circular motion and random acceleration motion, respectively, v x and v z are the x-direction velocity and z-direction velocity of random acceleration motion, p ij and q ij are motion parameters, i = 1, 2, 3, 4, j = 1, 2, 3, and randn is a function that generates random numbers or matrices with standard normal distribution.

[0021] Preferably, generating a plurality of one-dimensional motion data in different directions of the motion modality comprises the following steps:

[0022] Get the input data set X;

[0023] The mean function m(x) and covariance function k(x,x') are used to determine the random variable of the input data set X; the specific expression is:

[0024] f(x)~GP(m(x)k(x,x'));

[0025] Where m(x) = E[f(x)], k(x,x') = E[(f(x) - m(x))(f(x') - m(x'))], m(x) is the motion mode, X = [x1, x2, ... x i ,…,x n ] is a d×n dimensional input matrix, x i ∈R d is a d-dimensional input vector, GP is a Gaussian process, E is the mean, f(x) is a random variable expression, and x and x′ are two different parameters in X;

[0026] Subtract the position coordinates of the two moments before and after to obtain the time series of position increments, and use this time series as the motion data of the target under this mode; the specific expression is:

[0027]

[0028] Among them, x i (t k ) and z i (t k ) is the discretized position information in the x and z directions, Δx i (t k ) and Δz i (t k ) are the time series of position increments in the x direction and the z direction at time t, respectively, Δx i (t k+1 ) and Δz i (t k+1 ) are the time series of position increments in the x direction at time t+1 and in the z direction at time t+1 respectively;

[0029] Random weights are introduced to weight the target motion data under the four motion modes, and linear transition processing is performed when the random weights are switched to obtain two sets of one-dimensional motion data sets of the target;

[0030]

[0031] Among them, Δx(t k) and Δz(t k ) is the target motion data set expressed in position increments; w xi and w zi Represents the weights of different modes, i = 1, 2, 3, 4, and the specific expression is:

[0032]

[0033]

[0034] Among them, w xij and w zij Indicates that with t k Three different weight combinations were selected, j = 1, 2, 3, at t i The weights are switched when t0 < t k ≤t1, t2<t k ≤t3 and t4<t k ≤t f During the period, t1<t k ≤t2 and t3<t k ≤t4 is the weight transition period; the specific expression is:

[0035]

[0036] Preferably, the step of combining a plurality of one-dimensional motion data in different directions in pairs to generate a two-dimensional position increment includes:

[0037] The multiple one-dimensional motion data of the target in the x direction and the multiple one-dimensional motion data in the z direction are combined in pairs (Δx(t k ),Δz(t k )) to generate multiple ground target motion data.

[0038] Preferably, inputting the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion includes:

[0039] Input the target data to be predicted into the updated Gaussian mixture model to obtain Gaussian mixture regression prediction, and then obtain the predicted value y * , specifically:

[0040]

[0041] Among them, M represents the number of clusters, o is the input, that is, the two-dimensional position increment of the past two time periods, y is the output, that is, the two-dimensional position increment of the last unit time period, o * is the predicted input, f^(y) is the predicted function value, f i ^(y) is the mean of each prediction function, φi (y) represents the mixed weight, and the mixed weight calculation formula is:

[0042]

[0043] Among them, ω i is the mixing coefficient, μ iy is the mean of the output, ∑ iy is the output covariance matrix, N is the probability density of the multidimensional Gaussian model, and the specific expression is:

[0044]

[0045] Among them, μ is a d×1 vector, representing the mean of each dimension of data; ∑ is a d×d matrix, representing the covariance matrix of the model.

[0046] The present invention provides a target motion prediction system, comprising:

[0047] An acquisition module is used to obtain a plurality of one-dimensional motion data of the target basic motion mode in different directions;

[0048] The model training module is used to combine several one-dimensional motion data in different directions in pairs to generate two-dimensional position increments, and use a sliding window to select the true values ​​of the two-dimensional position increments for multiple unit time periods in real time. The true values ​​of the two-dimensional position increments for the first few unit time periods are input into the Gaussian mixture model, and the last two-dimensional position increment for the unit time period is used as the output result to update the Gaussian mixture model;

[0049] The prediction module is used to input the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion.

[0050] The present invention provides a computer device, comprising a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the above-mentioned target motion prediction method.

[0051] The present invention provides a storage medium storing a computer program, wherein the computer program is characterized in that when executed by a processor, the computer program implements the steps of the motion prediction method for the above-mentioned goal.

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

[0053] The present invention obtains a plurality of one-dimensional motion data according to the motion mode of the target, obtains a two-dimensional position increment by combining the one-dimensional data motion, selects the true values ​​of the two-dimensional position increments of multiple unit time periods in real time through a sliding window, inputs the true values ​​of the two-dimensional position increments of the first few unit time periods into a Gaussian mixture model, and updates the Gaussian mixture model with the last two-dimensional position increment of the unit time period as the output result. This realizes the continuous rolling of historical information and model prediction values ​​in the time domain, optimizes the Gaussian mixture model, and thus predicts the motion of the target to be detected, thereby greatly improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a ground target motion data set diagram of the present invention;

[0055] Figure 2 A Gaussian mixture model of the ground moving target position increment of the present invention;

[0056] Figure 3 This is the GMM-GMR trajectory prediction result diagram of the present invention;

[0057] Figure 4 is the GMM-GMR trajectory prediction error diagram of the present invention;

[0058] Figure 5 This is an overall flow chart of a motion prediction method for one object of the present invention. DETAILED DESCRIPTION

[0059] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0060] The present invention establishes a ground target motion data set and designs a target trajectory prediction algorithm based on Gaussian mixture model and Gaussian mixture regression.

[0061] like Figure 5 As shown, the present invention provides a method for predicting target motion, which specifically includes the following steps:

[0062] Step S1: Acquire a plurality of one-dimensional motion data of the target basic motion mode in different directions.

[0063] The basic motion modes include uniform motion, uniformly accelerated motion, circular motion and random acceleration motion. The specific expressions are:

[0064] x1=p 11 t+p 12 ;

[0065] x2=p 21 t 2 +p 22 t+p 23 ;

[0066] x3=p 31 sin(p 32 t)+p 33 ;

[0067] x4=p 41 +∫v x dt,v x =p 42 +∫p 43 *randndt;

[0068] z1=q 11 t+q 12 ;

[0069] z2=q 21 t 2 +q 22 t+q 23 ;

[0070] z3=q 31 sin(q 32 t)+q 33 ;

[0071] z4=q 41 +∫v z dt,v z =q 42 +∫q 43 *randndt;

[0072] Where t is time, x i and z i are the x-direction position and z-direction position of uniform motion, uniform acceleration motion, circular motion and random acceleration motion, respectively, v x and v z are the x-direction velocity and z-direction velocity of random acceleration motion, p ij and q ij are motion parameters, i = 1, 2, 3, 4, j = 1, 2, 3, and randn is a function that generates random numbers or matrices with standard normal distribution.

[0073] Using the basic motion mode as the mean, the standard exponential covariance function is applied to generate the covariance matrix of the target position at different times, and Gaussian Process Regression (GPR) is used to generate ten one-dimensional motion data of this motion mode. The basic principle of Gaussian Process Regression is as follows:

[0074] For the input training data set Among them, x i ∈R d is a d-dimensional input vector, X=[x1,x2,…,x n ] is the d×n dimensional input matrix, y i ∈R d is the corresponding output scalar; generating a number of one-dimensional motion data in different directions of the motion mode, including the following steps:

[0075] Get the input data set X; it is known that the input data set X can constitute a random variable set [f(x1),f(x2),…,f(x n )].

[0076] The mean function m(x) and covariance function k(x,x') are used to determine the random variable of the input data set X; the specific expression is:

[0077] f(x)~GP(m(x)k(x,x'));

[0078] Where m(x) = E[f(x)], k(x,x') = E[(f(x) - m(x))(f(x') - m(x'))], m(x) is the motion mode, X = [x1, x2, ... x i ,…,x n ] is a d×n dimensional input matrix, x i ∈R d is a d-dimensional input vector, GP is a Gaussian process, E is the mean, f(x) is a random variable expression, x and x′ are two different variables in X, and the kernel function used is the standard exponential covariance function, that is:

[0079]

[0080] Among them, θ0 is the exponential weight, θ1 represents the length scale, δ ij is the Dirac function. When i=j, the function δ ij =1, otherwise δ ij =0.

[0081] Incremental representation of the data set: Subtract the position coordinates of the two moments before and after to obtain the time series of position increments, and use this time series as the motion data of the target under this mode; the specific expression is:

[0082]

[0083] Among them, x i (t k ) and z i (tk ) is the discretized position information in the x and z directions, Δx i (t k ) and Δz i (t k ) are the time series of position increments in the x direction and the z direction at time t, respectively, Δx i (t k+1 ) and Δz i (t k+1 ) are the time series of position increments in the x direction at time t+1 and in the z direction at time t+1, respectively.

[0084] Two sets of one-dimensional motion data sets are obtained by weighting: random weights are introduced to weight the target motion data under the four motion modes, and linear transition processing is performed when the random weights are switched to obtain two sets of one-dimensional motion data sets of the target.

[0085]

[0086] Among them, Δx(t k ) and Δz(t k ) is the target motion data set expressed in position increments; w xi and w zi Represents the weights of different modes, i = 1, 2, 3, 4, and the specific expression is:

[0087]

[0088] Among them, w xij and w zij Indicates that with t k Three different weight combinations were selected, j = 1, 2, 3, at t i The weights are switched when t0 < t k ≤t1, t2<t k ≤t3 and t4<t k ≤t f During the period, t1<t k ≤t2 and t3<t k ≤t4 is the weight transition period; the specific expression is:

[0089]

[0090] Step S2: Combine several one-dimensional motion data in different directions in pairs to generate two-dimensional position increments, and use a sliding window to select the true values ​​of the two-dimensional position increments of multiple unit time periods in real time. The true values ​​of the two-dimensional position increments of the first few unit time periods are input into the Gaussian mixture model, and the Gaussian mixture model is updated with the last two-dimensional position increment of the unit time period as the output result.

[0091] The target will get multiple one-dimensional motion data in the x direction and multiple one-dimensional motion data in the z direction, and combine them two by two (Δx(t k ),Δz(t k )) to generate multiple ground target motion data; in the embodiment, Figure 1 The ten ground target motion data are shown.

[0092] A sliding window is used to select the true value of the two-dimensional position increment obtained by measurement, and the model is updated using the true value.

[0093] In the real-time prediction process, in order to improve the adaptability of the algorithm, it is necessary to update the model according to the current data. The present invention selects the data volume used by the real-time updated model in the prediction process as N=100. The specific method is as follows:

[0094] A sliding window (framing the time series data using a specified unit length, equivalent to sliding a slider of a specified length across a scale, recording a set of data within the slider for each unit of sliding) is used to select the true values ​​of the measured 2D position increments. The sliding window length is three time units. Corresponding to step 2), each data set consists of three time periods of 2D position increments, with the first two time periods being the input and the last time period being the output.

[0095] Based on the data obtained from the sliding window, only the N groups (if there are less than N groups, the training set data is used to fill the gap) of the latest data are retained to update the model. The new model is updated in the same way as 1) using the latest data.

[0096] In this way, the sliding window method is used to realize the rolling time domain, and a GMM-GMR trajectory prediction algorithm with real-time update of the model based on small data artificial intelligence is established.

[0097] Step S3: Input the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion.

[0098] Input the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction results of the motion, including:

[0099] Design a Gaussian Mixture Regression (GMR) algorithm with four inputs and two outputs. The input is the two-dimensional position increment over the past two time periods, denoted by o; the output is the two-dimensional position increment over the next time period, denoted by y.

[0100] Input the target data to be predicted into the updated Gaussian mixture model to obtain Gaussian mixture regression prediction, and then obtain the predicted value y *, specifically:

[0101]

[0102] Among them, M represents the number of clusters, o is the input, that is, the two-dimensional position increment of the past two time periods, y is the output, that is, the two-dimensional position increment of the last unit time period, o * is the predicted input, f^(y) is the predicted function value, f i ^(y) is the mean of each prediction function, φ i (y) represents the mixed weight, and the mixed weight calculation formula is:

[0103]

[0104] Among them, ω i is the mixing coefficient, μ iy is the mean of the output, ∑ iy is the output covariance matrix, N is the probability density of the multidimensional Gaussian model, and the specific expression is:

[0105]

[0106] Among them, μ is a d×1 vector, representing the mean of each dimension of data; ∑ is a d×d matrix, representing the covariance matrix of the model.

[0107] A six-dimensional Gaussian mixture model (GMM) is established for the two-dimensional position increments of the ground target motion in three adjacent time periods, that is, d = 6, as shown in Figure 2 shown.

[0108] Prediction algorithm simulation verification and result analysis:

[0109] The data set is divided into the first 9 motion trajectory data as the training set and the 10th motion trajectory data as the test set. The designed GMM-GMR trajectory prediction algorithm is simulated and verified. The results are as follows Figure 3 and Figure 4 shown.

[0110] The mean X-direction position increment for a ground moving target is 0.2096m, and the corresponding root mean square error of the predicted X-direction position increment is 0.0035m. The mean Z-direction position increment is 0.1028m, and the corresponding root mean square error of the predicted Z-direction position increment is 0.0040m. This shows that the GMM-GMR trajectory prediction performance based on small data is sufficient and can effectively achieve target trajectory prediction.

[0111] Based on the above method, the present invention provides a target motion prediction system, which includes: an acquisition module, a model training module and a prediction module.

[0112] Among them, the acquisition module is used to obtain several one-dimensional motion data of the target basic motion mode in different directions; the model training module is used to combine several one-dimensional motion data in different directions in pairs to generate two-dimensional position increments, and use a sliding window to select the true values ​​of the two-dimensional position increments of multiple unit time periods in real time, and input the true values ​​of the two-dimensional position increments of the first few unit time periods into the Gaussian mixture model, and update the Gaussian mixture model with the last two-dimensional position increment of the unit time period as the output result; the prediction module is used to input the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion.

[0113] The present invention also provides a computer device, comprising a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of a target motion prediction method.

[0114] According to the disclosed embodiments, a computing device may communicate with one or more external devices (e.g., a keyboard, a pointing device, Bluetooth communications, etc.), or with any device that enables a computing device to communicate with one or more other computing devices (e.g., a router, a modem, etc.).

[0115] The present invention also provides a storage medium storing a computer program, which implements the steps of a target motion prediction method when the computer program is executed by a processor.

[0116] According to the disclosed embodiments, the storage medium may be a non-volatile computer-readable storage medium, such as, but not limited to, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0117] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. For those skilled in the art to which the present invention belongs, several simple deductions or replacements can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for predicting target motion, characterized in that: The steps include: Obtaining several one-dimensional motion data of the target basic motion mode in different directions; Combine several one-dimensional motion data in different directions in pairs to generate two-dimensional position increments, and use a sliding window to select the true values ​​of the two-dimensional position increments in multiple unit time periods in real time. The true values ​​of the two-dimensional position increments in the first few unit time periods are input into the Gaussian mixture model, and the last two-dimensional position increment in the unit time period is used as the output result to update the Gaussian mixture model; Input the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion; The step of obtaining a plurality of one-dimensional motion data of the target basic motion mode in different directions comprises the following steps: Get the input data set X ; Using the mean function and covariance function Determine the input data set X The specific expression is: ； in, , , is the motion mode, is a d×n dimensional input matrix, For d-dimensional uniform motion, uniform acceleration motion, circular motion and random acceleration motion x Direction position, GP is a Gaussian process, E is the mean, is the random variable expression, and They are X Two different variables in Subtract the position coordinates of the two moments before and after to obtain the time series of position increments, and use this time series as the motion data of the target under this mode; the specific expression is: ； in, and yes x Discretized position information in the direction and z direction, and They are t time x Direction and t time z Time series of position increments in a direction, and They are t +1 moment x Direction and t +1 moment z Time series of position increments in a direction; Random weights are introduced to weight the target motion data under the four motion modes, and linear transition processing is performed when the random weights are switched to obtain two sets of one-dimensional motion data sets of the target; ； in, and is the target motion dataset expressed in position increments; and represents the weights of different modes, , the specific expression is: ； ； in, and Indicates that with Three different weight combinations were selected for the change. ,exist When switching weights, the three sets of weight combinations are applied to 、 and During the period, and is the weight transition period; the specific expression is: ； 。 2. The method for predicting target motion according to claim 1, wherein: The basic motion modes include uniform motion, uniformly accelerated motion, circular motion and random acceleration motion, and the specific expressions are: ； ； ； ； ； ； ； ； in, For time, and They are uniform motion, uniformly accelerated motion, circular motion and random acceleration motion. x Direction position and z Direction position, and They are random acceleration motion x Direction speed and z Directional speed, and is the motion parameter, , , randn A function that generates random numbers or matrices from a standard normal distribution.

3. The method for predicting target motion according to claim 1, wherein: The step of combining a plurality of one-dimensional motion data in different directions in pairs to generate a two-dimensional position increment includes: The target will be x Multiple one-dimensional motion data in the direction and z Multiple one-dimensional motion data in the direction are combined in pairs , generating multiple ground target motion data.

4. A method for predicting target motion according to claim 3, characterized in that: The step of inputting the target data to be predicted into the updated Gaussian mixture model to obtain a Gaussian mixture regression prediction result of the motion includes: Input the target data to be predicted into the updated Gaussian mixture model to obtain Gaussian mixture regression prediction, and then obtain the predicted value , specifically: ； in, M represents the number of clusters, o is the input, i.e. the two-dimensional position increment in the past two time periods, y is the output, i.e. the two-dimensional position increment of the last unit time period, is the predicted input, is the predicted function value, For each prediction function mean, Represents the mixed weight, and the mixed weight calculation formula is: ； in, is the mixing coefficient, is the mean of the output, is the covariance matrix of the output, N is the probability density of the multidimensional Gaussian model, and the specific expression is: ； in, Is a d×1 vector representing the mean of each dimension of data; It is a d×d matrix representing the covariance matrix of the model.

5. A motion prediction system according to any one of claims 1 to 4, characterized in that: include: An acquisition module is used to obtain a plurality of one-dimensional motion data of the target basic motion mode in different directions; The model training module is used to combine several one-dimensional motion data in different directions in pairs to generate two-dimensional position increments, and use a sliding window to select the true values ​​of the two-dimensional position increments for multiple unit time periods in real time. The true values ​​of the two-dimensional position increments for the first few unit time periods are input into the Gaussian mixture model, and the last two-dimensional position increment for the unit time period is used as the output result to update the Gaussian mixture model; The prediction module is used to input the target data to be predicted into the updated Gaussian mixture model to obtain the Gaussian mixture regression prediction result of the motion.

6. A computer device, characterized in that: The method comprises a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the target motion prediction method according to any one of claims 1 to 4.

7. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the motion prediction method according to any one of claims 1 to 4 are implemented.

Images