CN-121995361-A - Diffusion model radar track information probability prediction method based on condition priori

CN121995361ACN 121995361 ACN121995361 ACN 121995361ACN-121995361-A

Abstract

The invention discloses a diffusion model radar track information probability prediction method based on condition priori, which comprises the steps of obtaining historical track information of a target, inputting the historical track information into a trained condition prediction network to obtain predicted track information of the target, wherein the condition prediction network is used for extracting nonstationary factors and potential variables of the target during movement to obtain nonstationary characteristics and potential variable information, predicting future tracks of the target according to the nonstationary characteristics and the potential variable information, inputting the predicted track information into a trained diffusion model to generate a plurality of possible predicted track sequences through random sampling and gradual denoising processes, and obtaining predicted track distribution of the target, wherein the predicted track distribution further comprises uncertainty information of each predicted track sequence. The prediction information provided by the invention is more and comprises the quantitative information of the uncertainty of each prediction track, and can provide reliable probability reference and firm decision support for subsequent decision tasks.

Inventors

LI YUXIN
GUO JIAN
HU XINYUE
CHEN BO
CHEN WENCHAO
WANG PENGHUI
LIU HONGWEI

Assignees

西安电子科技大学

Dates

Publication Date: 20260508
Application Date: 20251231

Claims (10)

1. The method for predicting the probability of the radar track information of the diffusion model based on the condition prior is characterized by comprising the following steps: acquiring historical track information of a target; inputting the historical track information into a trained condition prediction network to obtain predicted track information of a target; The condition prediction network is used for extracting non-stationary factors and potential variables when the target moves to obtain non-stationary characteristic and potential variable information, and predicting a future track of the target according to the non-stationary characteristic and the potential variable information; And inputting the predicted track information into a trained diffusion model to generate a plurality of possible predicted track sequences through random sampling and gradual denoising processes, so as to obtain the predicted track distribution of the target, wherein the predicted track distribution also comprises uncertainty information of each predicted track sequence.
2. The method of claim 1, wherein the conditional prediction network comprises a non-stationary factor learning module, an embedded layer, an encoder, a latent variable processing module, a decoder; the non-stationary factor learning module is used for extracting local features of the historical track information and learning non-stationary factors to obtain the non-stationary factor features; The embedding layer is used for carrying out embedding processing on the historical track information to obtain a first embedded vector; The encoder is used for extracting high-dimensional latent layer information characterization of the historical track information according to the first embedded vector and the non-stationary factor characteristic; The latent variable processing module is used for extracting the latent variable information according to the high-dimensional latent layer information representation; The embedded layer is further used for carrying out embedded processing on the track sequence vector to obtain a second embedded vector, wherein the track sequence vector is composed of historical track information and a placeholder of the predicted track information; the decoder is configured to determine the predicted track information based on the latent variable information, the non-stationary factor characteristic, and the second embedded vector.
3. The method according to claim 2, wherein the decoder is specifically configured to perform a masked self-attention operation on the second embedded vector, and to perform a cross-attention operation on the second embedded vector as a query vector and the latent variable information as a key vector and a value vector under modulation of the non-stationary factor characteristic to generate the predicted track information.
4. The method according to claim 1, wherein the loss function used by the conditional prediction network in training comprises an expected loss term for improving the prediction accuracy of the conditional prediction network and a KL-divergence loss term for constraining the distribution of potential variables.
5. The method of claim 4, wherein the expected loss term satisfies the following formula: Wherein, the For the term of the expected loss(s), The potential variables are represented by the values of the potential variables, The historical track information is represented and is displayed, For the time step length of each historical track in the historical track information, Representing a mapping function that performs feature extraction and encoding on the historical track information, Representing a posterior distribution of latent variables encoded from the historical track information; Representing the predicted track information in question, For each predicted time step length of the track, Expressed in given latent variable The conditional probability distribution of the track sequence is predicted under the condition, To be used to approximate the desired number of samples, To distribute from the potential variable Sample in (b) to obtain A number of samples of the sample were taken, Indicated at a given first Samples of potential variables Under the condition of (1) predicting track sequence Is a conditional probability distribution of (c).
6. The method according to claim 4, wherein the KL-divergence loss term satisfies the following formula: Wherein, the For the KL divergence loss term, The potential variables are represented by the values of the potential variables, The historical track information is represented and is displayed, For the time step length of each historical track in the historical track information, Representing a mapping function that performs feature extraction and encoding on the historical track information, Represents the posterior distribution of the latent variable encoded from the historical track information, A priori distribution of potential variables; 、 and respectively representing the mean and the variance corresponding to the j potential variable dimension.
7. The method of claim 1 wherein the diffusion model uses a real track sequence as a target sample during training, and wherein the forward diffusion process is constructed by gradually injecting random noise into the real track sequence under the constraint of predicted track information, and wherein the backward denoising process is learned by predicting noise and gradually recovering real track distribution given the noisy track, the predicted track information and the number of diffusion steps.
8. The method of claim 7, wherein the loss function used by the diffusion model in training satisfies the following equation: Wherein, the As a loss function of the diffusion model, In order for the noise to be actually injected, For the noise predicted by the diffusion model, For the diffusion model to true track Track information obtained after the t-th noise addition, Representing the predicted track information in question, For each predicted time step length of the track, In order to diffuse the total number of steps, The L2 norm is represented by the number, Representing the true posterior distribution of the diffuse forward process, i.e. from the first, given the true track and the conditional predicted track Step back to the first The true distribution of the steps is that, The historical track information is represented and is displayed, For the time step length of each historical track in the historical track information, Representing the inverse denoising profile learned from the diffusion model, given the historical track information and the conditional prediction track, from the first Step back to the first And (3) step (c).
9. The diffuse model radar track information probability prediction device based on the condition prior is characterized by comprising an acquisition unit and a processing unit; the acquisition unit is used for acquiring historical track information of the target; The processing unit is used for: inputting the historical track information into a trained condition prediction network to obtain predicted track information of a target; The condition prediction network is used for extracting non-stationary factors and potential variables when the target moves to obtain non-stationary characteristic and potential variable information, and predicting a future track of the target according to the non-stationary characteristic and the potential variable information; And inputting the predicted track information into a trained diffusion model to generate a plurality of possible predicted track sequences through random sampling and gradual denoising processes, so as to obtain the predicted track distribution of the target, wherein the predicted track distribution also comprises uncertainty information of each predicted track sequence.
10. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that the processor implements the method according to any of claims 1-8 when executing the computer program.

Description

Diffusion model radar track information probability prediction method based on condition priori Technical Field The invention belongs to the technical field of radar signal processing, and particularly relates to a diffusion model radar track information probability prediction method based on condition priori. Background The radar track prediction is a core research direction in the fields of radar data processing and target situation awareness, and is essentially based on historical target observation data acquired by radar sensors, and the spatial position, the motion state and the evolution trend of a target in a future period are predicted by modeling the motion rule and the environment interference characteristic of the target, so that the radar track prediction is a decision basis of key scenes such as airspace safety protection, air traffic control, complex airspace situation awareness and the like. In an airspace safety protection scene, accurate radar track prediction can predict the flight track of a target in advance of several seconds to tens of seconds, so that sufficient reaction time is provided for a related protection system, in air traffic control, track prediction needs to avoid track conflicts of multiple aircrafts in real time, the operation safety of an airspace is ensured, in complex airspace situation awareness, the track prediction needs to integrate multiple radar and multiple sensor data, the motion association and operation logic of multiple targets are restored, and related scheduling decisions are supported. However, in practical applications, the target has maneuver characteristics and exhibits complex non-stationary characteristics, which result in limited accuracy and consistency of conventional track distribution predictions. Meanwhile, the traditional radar track prediction method focuses on point estimation of the future state of the target, and the single predicted value cannot fully reflect potential risks and may cause decision bias (such as misjudgment of traffic scheduling opportunity, civil aviation track conflict and the like). And the method lacks accurate quantification of uncertainty of a prediction result, and is difficult to provide reliable reference for subsequent radar-related decisions and regulation. Disclosure of Invention The embodiment of the invention provides a diffusion model radar track information probability prediction method based on condition priori, which can solve the problems that the traditional scheme focuses on point estimation and does not quantify the uncertainty of a prediction result, and the given prediction information has less influence on decision effect. In a first aspect, a method for predicting probability of radar track information of a diffusion model based on condition prior provided by an embodiment of the present invention includes: acquiring historical track information of a target; inputting the historical track information into a trained condition prediction network to obtain predicted track information of a target; The condition prediction network is used for extracting non-stationary factors and potential variables when the target moves to obtain non-stationary characteristic and potential variable information, and predicting a future track of the target according to the non-stationary characteristic and the potential variable information; And inputting the predicted track information into a trained diffusion model to generate a plurality of possible predicted track sequences through random sampling and gradual denoising processes, so as to obtain the predicted track distribution of the target, wherein the predicted track distribution also comprises uncertainty information of each predicted track sequence. In a second aspect, an embodiment of the present invention provides a diffusion model radar track information probability prediction apparatus based on condition prior, including an acquisition unit and a processing unit; the acquisition unit is used for acquiring historical track information of the target; The processing unit is used for: inputting the historical track information into a trained condition prediction network to obtain predicted track information of a target; The condition prediction network is used for extracting non-stationary factors and potential variables when the target moves to obtain non-stationary characteristic and potential variable information, and predicting a future track of the target according to the non-stationary characteristic and the potential variable information; And inputting the predicted track information into a trained diffusion model to generate a plurality of possible predicted track sequences through random sampling and gradual denoising processes, so as to obtain the predicted track distribution of the target, wherein the predicted track distribution also comprises uncertainty information of each predicted track sequence. In a third aspect, an embodiment of the present i