Intelligent tracking and trajectory forecasting method for high-speed maneuvering target based on neural process

Abstract

The invention particularly relates to an intelligent tracking and trajectory forecasting method for a high-speed maneuvering target based on a neural process, which comprises the following steps of: obtaining a maneuvering target motion data set, automatically learning distribution information from data by adopting an encoder-decoder architecture, and automatically predicting the trajectory of the maneuvering target through a unique hidden variable design. The target motion time sequence characteristics and the uncertainty of the track random process are captured, the problem that a traditional neural network tracking method lacks the random process modeling capability is solved, accurate estimation of the target state and the track is achieved, and meanwhile the network is endowed with good generalization capability and interpretability. The method is suitable for maneuvering target tracking and trajectory forecasting tasks in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of target tracking, specifically to a method for intelligent tracking and trajectory prediction of high-speed maneuvering targets based on neural processes. Background Technology

[0002] Target tracking and trajectory prediction is a process in which engineers use sensors to estimate and predict the state of a target in real time, and it has a wide range of applications. Traditional model-driven target tracking algorithms are often limited by prior information and dynamic models that rely on "single-step prediction," making them difficult to apply to scenarios where prior information is scarce and targets are highly maneuverable and move in complex ways.

[0003] Recent research shows that Recurrent Neural Networks (RNNs) and attention-based Transformer networks can effectively capture long-term evolution patterns of target states from time-series data, overcoming the limitations of single-step prediction in classical dynamic models. These technologies have become cutting-edge techniques for target tracking and trajectory prediction. For example, the patent "A High-Speed ​​Maneuvering Target Tracking Error Compensation Method Based on Intelligent Smoothing of Time Series" uses a Long Short-Term Memory (LSTM) network to learn the tracking error patterns of a filtering method, and improves tracking accuracy by compensating for fitting errors.

[0004] However, such neural network-based target tracking methods essentially approximate a fixed end-to-end mapping function through the network. Since the network parameters remain unchanged after training, the output is often deterministic for the same set of input measurements, making it difficult to effectively capture steady-state uncertainties in the data and the similarity of trajectories under different maneuvering modes. When the target undergoes strong maneuvers, and the maneuvering mode and target trajectory type exceed the distribution range of the training dataset (e.g., a new trajectory generated by simply changing a single aerodynamic parameter), the network performance degrades significantly, sometimes requiring retraining to restore generalization ability. To address this problem, stochastic processes offer a new solution. They characterize the function's variation pattern at different input points through the joint distribution of random variables at different times, enabling the capture of long-term temporal correlations of states in the input space and the similarity of spatiotemporal correlations of states under different maneuvering modes.

[0005] Furthermore, constrained by the objective physical laws of aerodynamics, the state of high-speed maneuvering targets exhibits strong spatiotemporal correlation and regularity in the spatiotemporal domain. This is because factors such as environmental changes, fuel consumption, and the external forces controlling the target's motion all exhibit strong spatiotemporal correlation characteristics and steady-state properties, with the spatiotemporal changes of some factors approaching steady-state stochastic processes. Therefore, the evolution of the target's state in the spatiotemporal domain is not random and irregular; the generated trajectory is smooth and regular, suitable for describing stochastic processes. In response, the patent "A High-Speed ​​Maneuvering Target Tracking Method Based on a Stochastic Process of Continuous Spatiotemporal Track Function" uses stochastic processes to model the evolution law of the target's state, effectively capturing the long-term, refined motion characteristics of the target trajectory. It can effectively handle some simple, smooth motion trajectories, but it does not fully utilize the powerful nonlinear complex function learning ability of the network, limiting the types of moving targets it can handle.

[0006] In summary, the core starting point of this invention is to combine the advantages of neural networks in nonlinear fitting with the ability of stochastic processes in uncertainty modeling, thereby balancing the interpretability and generalization performance of the model, overcoming the shortcomings of existing technologies, and realizing intelligent tracking and accurate trajectory prediction of complex high-speed maneuvering targets.

[0007] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] This invention provides a method for intelligent tracking and trajectory prediction of high-speed maneuvering targets based on neural processes, which enables intelligent tracking and accurate trajectory prediction of complex high-speed maneuvering targets, thereby overcoming the shortcomings of existing technologies to a certain extent.

[0009] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0010] According to a first aspect of the present invention, a method for intelligent tracking and trajectory prediction of high-speed maneuvering targets based on neural processes is provided, the method comprising: Step 1: Obtain the motion dataset of the maneuvering target, which includes measurement sequences; Step 2: Use an RNN network to extract the temporal information features of the measurement sequence and map them into a high-dimensional representation space to obtain a latent representation containing temporal information. ,Right now ,in Represents an RNN network; Step 3: Use a multilayer perceptron (MLP) based encoder to process the input time. and high-dimensional measurement information Encode the data, learn the mapping from time to the target state, and generate the code. This is used to further generate latent variables, i.e. ,in Indicates encoder; Step 4: Using a mapping network based on MLP, obtain the distribution of latent variables, i.e., the mean. Covariance ,satisfy Latent variables are obtained through reparameter sampling. ; Step 5: Utilize the latent variables learned by the encoder and the current moment Joint modeling is performed, and the information is subjected to multi-level nonlinear transformation through an MLP-based decoder to obtain the target state estimate. ,Right now ,in Indicates decoder; When performing a tracking task, the decoder input is the current time, and the decoder outputs a state estimate of the target; when performing a trajectory prediction task, the decoder input is a future time, and the decoder outputs a state prediction of the target. Step 6: Perform inverse normalization on the state estimate and then concatenate the state estimates at each time step to form a continuous track.

[0011] In some exemplary embodiments, obtaining the motion dataset of the maneuvering target specifically involves: Define a state-space model, and the state transition equations for the objective are as follows:

[0012] The state of the target is represented as: Including location information Speed ​​information and acceleration information ; Represents the state transition matrix. Represents the noise matrix. Indicates process noise; The measurement of the target is , including distance and azimuth The measurement equation is:

[0013] in, This refers to measurement noise, including distance noise and azimuth noise.

[0014] In some exemplary embodiments, the RNN network has 2 layers, with 128 hidden units in each layer.

[0015] In some exemplary embodiments, the encoder has 3 layers, with 128 units in the first and second layers and 64 units in the last layer.

[0016] In some exemplary embodiments, the mapping network has two layers, with 64 hidden layer units in each layer.

[0017] In some exemplary embodiments, the decoder has 5 layers, with 128 units in the first and second layers, 128 units in the third layer, 64 units in the fourth layer, and 2 units in the last layer.

[0018] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the high-speed maneuvering target intelligent tracking and trajectory prediction method based on neural processes described in the first aspect.

[0019] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the high-speed maneuvering target intelligent tracking and trajectory prediction method based on neural processes described in the first aspect above.

[0020] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the neural process-based intelligent tracking and trajectory prediction method for high-speed maneuvering targets as described in the first aspect by executing the executable instructions.

[0021] The high-speed maneuvering target intelligent tracking and trajectory prediction method based on neural processes provided in this invention uses stochastic processes to model the target tracking scenario. It employs an encoder-decoder architecture to automatically learn distribution information from the data. Through a unique latent variable design, it captures the temporal characteristics of target motion and the uncertainty of the stochastic trajectory process, solving the problem of traditional neural network tracking methods lacking stochastic process modeling capabilities. This achieves accurate estimation of the target state and trajectory, while endowing the network with good generalization ability and interpretability. This method is suitable for maneuvering target tracking and trajectory prediction tasks in complex scenarios.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0024] Figure 1 This is a flowchart of the method of the present invention; Figure 2 It contains 2000 trajectory images from the dataset; Figure 3 This is a framework diagram of the method described in this paper; Figure 4 It is a simulation scene diagram; Figure 5 This is a comparison chart of tracking errors. Detailed Implementation

[0025] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0026] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0027] For target tracking tasks, the target's motion state It typically exhibits significant long-term time-series characteristics and can be represented by time as the independent variable. function And by a random process The sampling yielded: (1) in, It is defined in the time domain (input space). Above, in state space For the output random process, , By modeling stochastic processes, the state estimation problem can be transformed into an inference problem of stochastic functions. Specifically, that is, at a given time... and measurement sequence Under the condition of, realize the function The inference is then used to obtain the posterior estimate of the state based on the inference function. : (2) NP is an algorithmic model for estimating stochastic processes using neural networks. It not only possesses the ability to model the uncertainty of stochastic processes but also leverages the powerful function fitting advantages of neural networks. Its core idea is to use neural networks to approximate the function distribution and extract high-dimensional vectors from measurement information. And thus parameterize the random process Therefore, for the function Sampling is equivalent to sampling the latent variables. Sampling: (3) To address the shortcomings and deficiencies of existing technologies, this example embodiment provides a high-speed maneuvering target intelligent tracking and trajectory prediction method based on neural processes (NP). This invention aims to utilize neural networks to learn the temporal characteristics of target motion, overcome the limitations of current neural network target tracking algorithms in lacking uncertainty and stochastic process modeling capabilities, achieve high-precision target tracking, and provide a novel model and data-driven solution for target tracking and other fields.

[0028] refer to Figure 1 As shown, the specific steps may include: Step 1: Dataset Creation First, a motion dataset of the maneuvering target needs to be established for network training, and a state-space model needs to be defined. The state transition equation of the target is shown below: (6) The state of the target is represented as: Including location information Speed ​​information and acceleration information . Represents the state transition matrix. Represents the noise matrix. This indicates process noise.

[0029] Based on the common motion states of the target, the motion models set include constant velocity (CV), constant acceleration (CA), and constant turn (CT).

[0030] The measurement of the target is , including distance and azimuth The measurement equation is: (7) in, This refers to measurement noise, including distance noise and azimuth noise.

[0031] To cover as many possible motion scenarios as possible, this invention categorizes the motion of targets in the dataset into two cases: 1) Single motion model state: The target does not perform any maneuvers during its motion, and the probability of each of the three motion models is 1 / 3. 2) Maneuver state: The target randomly performs one maneuver during its motion, and the motion model switches.

[0032] Step 2: Temporal Feature Extraction To further enhance the processing capability of time-series information, this invention introduces an RNN network. First, the time-series information features of the measurement are extracted and mapped into a high-dimensional representation space to obtain a latent representation containing the time-series information. ,Right now ,in This represents an RNN network.

[0033] Step 3: Encoding Using a multilayer perceptron (MLP) based encoder, the input time... and high-dimensional measurement information Encode the data, learn the mapping from time to the target state, and generate the code. This is used to further generate latent variables, i.e. ,in This indicates the encoder.

[0034] Step 4: Generation of latent variables After all the encodings are aggregated, the distribution (i.e., the mean) of the latent variables is obtained through an MLP-based mapping network. Covariance ),satisfy Then, latent variables are obtained through reparameter sampling, explicitly giving them the ability to express uncertainty.

[0035] Step 5: Decoding Latent variables learned by the encoder and the current moment Joint modeling is performed, and the information is subjected to multi-level nonlinear transformation through an MLP-based decoder to obtain the state estimate. ,Right now ,in This indicates the decoder.

[0036] Step 6: Loss Function Calculation To train the various parts of the network structure, the loss function needs to be derived. In a probabilistic optimization problem, this requires maximizing the likelihood. However, since the true posterior distribution is often difficult to solve analytically, directly calculating the likelihood is quite challenging. Therefore, based on variational Bayesian inference, the lower bound of evidence in equation (3) is calculated, and the likelihood is approximately maximized using the true values ​​of the states generated from the dataset. The lower bound of evidence in equation (3) is shown below: (4) However, the prior distribution of latent variables and It is usually unknown, but the posterior distribution can be obtained using a mapping network. and To replace the prior distribution. After rearranging, we can obtain the loss function. : (5) in, This represents the distribution of the decoder's output states.

[0037] The steps in this exemplary embodiment will now be described in more detail with reference to the accompanying drawings and embodiments.

[0038] I. Training Phase Step 1: Dataset Generation Based on the state-space model defined above, the parameters of the dataset are set as shown in the table below: Table 1 Dataset Parameter Settings

[0039] To fully train the network, this invention generated 1 million tracks, with the training set and validation set ratios being 80% and 20%, respectively.

[0040] Step 2: Data Preprocessing First, convert the distance and azimuth information into coordinate information: (8) in, and They represent and Measurement data in direction, record the measurement coordinates. .

[0041] The track is then segmented using a sliding window to obtain fixed-length subsequences that meet the network input requirements. Let the length of the sliding window be... If the value is 20, then after track segmentation, a measurement coordinate subsequence can be obtained. and state truth subsequence .

[0042] Even after processing using equation (8), the ranges of different tracks still vary significantly. To prevent the dynamic characteristics and noise features of the tracks from being masked by the coordinate values, normalization is also required: (9) In the above formula, . and These represent the normalized measurement coordinates and the true value of the target state, respectively. and This represents the maximum and minimum values ​​of the measured coordinates within the sliding window.

[0043] Step 3: Network parameter settings Model parameter settings: latent variable dimension set to 64; RNN network. The encoder has two layers, with 128 hidden layer units in each layer; The network has 3 layers, with 128 units in the first and second layers and 64 units in the last layer; it is a mapping network. The decoder has two layers, with 64 hidden units in each layer; The building has 5 layers. The first and second layers have 128 units each, the third layer has 128 units each, the fourth layer has 64 units each, and the last layer has 2 units each.

[0044] Training parameter settings: 100 training epochs, initial learning rate of 0.001, multiplied by 0.5 every 20 steps, batch size of 512, and Adam optimization algorithm.

[0045] The network structure and parameter settings described in this invention are not limited to the above-described form, and other suitable network structures and parameter settings can be adopted according to specific application scenarios.

[0046] II. Testing Phase Step 1: Simulated trajectory generation Simulation Scenario 1: The initial state of the target is First, a CV motion lasting 50 seconds is performed, followed by a motion with an angular velocity of 6. The CT motion lasts for 50 seconds, followed by a CV motion lasting for 50 seconds, and then an angular velocity of... The CT motion lasts for 50 seconds, followed by a CV motion lasting for 50 seconds. The total motion time is 250 seconds.

[0047] Simulation Scenario 2: The initial state of the target is First, a CV motion lasting 30 seconds is performed, followed by an acceleration of... The first motion consists of a CA motion lasting 80 seconds, followed by a CV motion lasting 50 seconds. The total motion time is 160 seconds.

[0048] In both simulation scenarios, the process noise and measurement noise settings are consistent with the dataset.

[0049] Step 2: Data Preprocessing This step is consistent with the data preprocessing steps in the network training phase.

[0050] Step 3: Network Testing The simulated flight path is input into the network and processed by the trained model. During the tracking task, the decoder input is the current time. The decoder outputs the target's state estimate; when performing trajectory prediction tasks, the decoder input is the future time. The decoder outputs the state prediction of the target.

[0051] Step 4: Data Post-processing To obtain a complete estimated track of the target, the output needs to be denormalized, and the state estimates at each time step need to be concatenated to form a continuous track.

[0052] Step 5: Performance Analysis and Comparison To compare the tracking performance of different methods, two evaluation metrics were set: Root Mean Square Error (RMSE) and Mean RMSE (MRMSE). These metrics were based on 100 Monte Carlo simulations, and the calculation formulas are as follows: (10) (11) in, Indicates the number of Monte Carlo simulations. Indicates the length of the trajectory. Indicates the first Second simulation. Indicates the first The state estimate of the network output during the next simulation. Indicates the first The true state of the target during the next simulation.

[0053] Compared to the Interactive Multiple Model (IMM) algorithm, in simulation scenario 1, the MRMSE of this invention is 9.89, while the MRMSE of the IMM algorithm is 10.28; in simulation scenario 2, the MRMSE of this invention is 13.79, while the MRMSE of the IMM algorithm is 16.48. Furthermore, [the following text is incomplete and requires further context]. Figure 5 It can be observed that the RMSE of this method is smaller than that of the IMM algorithm in most cases. In summary, this invention has higher accuracy compared to the IMM algorithm.

[0054] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments.

[0055] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0056] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0057] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0058] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A method for intelligent tracking and trajectory prediction of high-speed maneuvering targets based on neural processes, characterized in that, The method includes: Step 1: Obtain the motion dataset of the maneuvering target, which includes measurement sequences; Step 2: Use an RNN network to extract the temporal information features of the measurement sequence and map them into a high-dimensional representation space to obtain a latent representation containing temporal information. ,Right now ,in Represents an RNN network; Step 3: Use a multilayer perceptron (MLP) based encoder to process the input time. and high-dimensional measurement information Encode the data, learn the mapping from time to the target state, and generate the code. This is used to further generate latent variables, i.e. ,in Indicates encoder; Step 4: Using a mapping network based on MLP, obtain the distribution of latent variables, i.e., the mean. Covariance ,satisfy Latent variables are obtained through reparameter sampling. ; Step 5: Utilize the latent variables learned by the encoder and the current moment Joint modeling is performed, and the information is subjected to multi-level nonlinear transformation through an MLP-based decoder to obtain the target state estimate. ,Right now ,in Indicates decoder; When performing a tracking task, the decoder input is the current time, and the decoder outputs a state estimate of the target; when performing a trajectory prediction task, the decoder input is a future time, and the decoder outputs a state prediction of the target. Step 6: Perform inverse normalization on the state estimate and then concatenate the state estimates at each time step to form a continuous track.

2. The method according to claim 1, characterized in that, The acquisition of the motion dataset of the maneuvering target specifically involves: Define a state-space model, and the state transition equations for the objective are as follows: The state of the target is represented as: Including location information Speed ​​information and acceleration information ; Represents the state transition matrix. Represents the noise matrix. Indicates process noise; The measurement of the target is , including distance and azimuth The measurement equation is: in, This refers to measurement noise, including distance noise and azimuth noise.

3. The method according to claim 1, characterized in that, The RNN network has two layers, with 128 hidden units in each layer.

4. The method according to claim 4, characterized in that, The encoder has 3 layers, with 128 units in the first and second layers and 64 units in the last layer.

5. The method according to claim 4, characterized in that, The mapping network has two layers, with 64 hidden units in each layer.

6. The method according to claim 5, characterized in that, The decoder has 5 layers, with 128 units in the first and second layers, 128 units in the third layer, 64 units in the fourth layer, and 2 units in the last layer.

7. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the high-speed maneuvering target intelligent tracking and trajectory prediction method based on neural processes as described in any one of claims 1 to 6.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the high-speed maneuvering target intelligent tracking and trajectory prediction method based on neural processes as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the neural process-based intelligent tracking and trajectory prediction method for high-speed maneuvering targets according to any one of claims 1 to 6 by executing the executable instructions.

Images