CN-121999011-A - Underwater maneuvering target tracking method and device based on deep learning

CN121999011ACN 121999011 ACN121999011 ACN 121999011ACN-121999011-A

Abstract

The application provides a deep learning-based underwater maneuvering target tracking method and device, wherein the method comprises the steps of obtaining observation data of a target object, determining state estimation values and covariance estimation values of a plurality of motion models on the target object by utilizing the observation data, inputting the state estimation values and covariance estimation values of the motion models on the target object and the observation data into probability estimation models to obtain model probability estimation values of the motion models, and obtaining final state estimation values and final covariance estimation values of the target object by the state estimation values, the covariance estimation values and the model probability estimation values. According to the method, the nonlinear modeling and the deep learning model are combined, the model probability of the motion model is estimated rapidly and accurately, and the accuracy and the instantaneity of tracking the target object are improved.

Inventors

MA XIAOCHUAN
JI TIANHANG
LIU YU
FENG CHAO

Assignees

中国科学院声学研究所

Dates

Publication Date: 20260508
Application Date: 20251219

Claims (10)

1. An underwater maneuvering target tracking method based on deep learning, which is characterized by comprising the following steps: obtaining observation data of a target object at the current moment, wherein the observation data are used for describing the position information of the target object; Determining state estimation values and covariance estimation values of a plurality of motion models on the target object by utilizing the observation data; Inputting the state estimation value and covariance estimation value of each motion model to the target object and the observation data into a probability estimation model to obtain a model probability estimation value of each motion model, wherein the probability estimation model is a deep learning model; and obtaining a final state estimation value and a final covariance estimation value of the target object through the state estimation value, the covariance estimation value and the model probability estimation value, wherein the final state estimation value and the final covariance estimation value are used for determining predicted position information of the target object at the next moment and finally tracking the target object.
2. The method of claim 1, wherein determining state estimates and covariance estimates for the object for a plurality of motion models using the observation data comprises: The method comprises the steps of establishing a nonlinear system state space model, wherein the nonlinear system state space model is used for setting a motion model in the nonlinear system state space model so as to obtain a first state estimated value and a first observation estimated value of the motion model on a target object; And obtaining a state estimated value and a covariance estimated value of each motion model on the target object by using the nonlinear system state space model.
3. The method of claim 2, wherein said using the nonlinear system state space model to obtain a state estimate and a covariance estimate for each of the motion models for the object comprises: Determining a first state estimation value, a first covariance estimation value and model probability of each motion model by using the nonlinear system state space model; determining a mixed state estimated value and a mixed covariance estimated value of each motion model according to the first state estimated value, the first covariance estimated value and the model probability; And obtaining state estimation values and covariance estimation values of the motion models according to the mixed state estimation values and the mixed covariance estimation values.
4. A method according to claim 3, wherein the nonlinear system state space model comprises state estimates And observed estimates ; Wherein, the , And The x-axis and y-axis coordinates at time k, And The speeds in the x-axis and y-axis directions, respectively, k being the time of day; a state transition matrix of the motion model at the moment k-1, For the noise weight matrix to be a weighted matrix of the noise, As a function of the non-linear measurement, In order for the process to be noisy, To measure noise, and And Independent of each other.
5. A method according to claim 3, wherein said determining a hybrid state estimate and a hybrid covariance estimate for each of said motion models based on said first state estimate, said first covariance estimate, and said model probabilities comprises: By fixing the transfer matrix And model probabilities for each of the motion models at time k-1 Obtaining the prediction model probability of the motion model j at the k moment Wherein the predictive model probability The probability that the predicted model j is the correct model based on all the observation information up to the time k-1 at the time k is that the fixed transition matrix Representing the probability of transition from motion model i to motion model j; Model probabilities at time k-1 by the fixed transfer matrix and the motion model Obtaining the prediction model probability of the motion model j at the k moment ; Predictive model probability at time k by motion model j And model probabilities of the motion model at time k-1 Obtaining the mixed probability of the motion model i to the motion model j at the moment k-1 ; Mixing probability of the motion model i to the motion model j through the k-1 moment And a state estimation value of each motion model at the time of k-1 Obtaining the mixed state estimation of the motion model j at the moment k-1 ; Probability of mixing motion model i to motion model j by time k-1 Covariance estimate Hybrid state estimation for motion model j Obtaining a mixed covariance estimated value of the k-1 moment motion model j 。
6. The method of claim 1, wherein the obtaining the final state estimate and the final covariance estimate of the target object from the state estimate, the covariance estimate, and the model probability estimate comprises: -estimating said state estimate of the motion model j The covariance estimate And the model probability estimate Input of , Obtaining the final state estimation value And the final covariance estimate 。
7. The method of claim 1, further comprising training the probabilistic estimation model using state estimate history data, covariance estimate history data, and model probabilistic estimate history data.
8. An underwater maneuver target tracking device based on deep learning, the device comprising: The acquisition module is used for acquiring the observation data of the current moment of the target object, wherein the observation data are used for describing the position information of the target object; the first processing module is used for determining state estimation values and covariance estimation values of a plurality of motion models on the target object by utilizing the observation data; The second processing module is used for inputting the state estimation value and the covariance estimation value of each motion model to the target object and the observation data into a probability estimation model to obtain a model probability estimation value of each motion model, wherein the probability estimation model is a deep learning model; and the third processing module is used for obtaining a final state estimation value and a final covariance estimation value of the target object through the state estimation value, the covariance estimation value and the model probability estimation value, wherein the final state estimation value and the final covariance estimation value are used for determining the predicted position information of the target object at the next moment and finally tracking the target object.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor implements the deep learning based underwater maneuver target tracking method as claimed in any one of claims 1 to 7 when the program is executed by the processor.
10. A computer storage medium having stored thereon a computer program which, when executed by one or more processors, implements the deep learning based underwater maneuver target tracking method as claimed in any of claims 1 to 7.

Description

Underwater maneuvering target tracking method and device based on deep learning Technical Field One or more embodiments of the present disclosure relate to the field of target tracking technology, and in particular, to a method and apparatus for tracking an underwater maneuvering target based on deep learning. Background In underwater target detection and tracking, targets often show nonlinear, non-Gaussian and uncertain complex motions, and particularly, the realization of accurate state estimation and continuous tracking under the maneuvering states of high speed, large turning and the like is a key technical problem. Currently, the main stream is a state estimation method based on Bayesian theory, such as Kalman filtering, extended Kalman filtering, unscented Kalman filtering and particle filtering. However, these methods rely heavily on a priori assumptions about the motion model of the target, which can significantly degrade when the model is mismatched or the target maneuvers severely. Content of the application The application describes an underwater maneuvering target tracking method and device based on deep learning, which can solve the technical problems. According to a first aspect, a method for tracking underwater maneuvering targets based on deep learning is provided. The method comprises the following steps: obtaining observation data of a target object at the current moment, wherein the observation data are used for describing the position information of the target object; Determining state estimation values and covariance estimation values of a plurality of motion models on the target object by utilizing the observation data; Inputting the state estimation value and covariance estimation value of each motion model to the target object and the observation data into a probability estimation model to obtain a model probability estimation value of each motion model, wherein the probability estimation model is a deep learning model; and obtaining a final state estimation value and a final covariance estimation value of the target object through the state estimation value, the covariance estimation value and the model probability estimation value, wherein the final state estimation value and the final covariance estimation value are used for determining predicted position information of the target object at the next moment and finally tracking the target object. In some embodiments, the determining, using the observation data, a state estimate and a covariance estimate for the object for a plurality of motion models comprises: The method comprises the steps of establishing a nonlinear system state space model, wherein the nonlinear system state space model is used for setting a motion model in the nonlinear system state space model so as to obtain a first state estimated value and a first observation estimated value of the motion model on a target object; And obtaining a state estimated value and a covariance estimated value of each motion model on the target object by using the nonlinear system state space model. In some embodiments, the obtaining, by using the nonlinear system state space model, a state estimation value and a covariance estimation value of each motion model for the target object includes: Determining a first state estimation value, a first covariance estimation value and model probability of each motion model by using the nonlinear system state space model; determining a mixed state estimated value and a mixed covariance estimated value of each motion model according to the first state estimated value, the first covariance estimated value and the model probability; And obtaining state estimation values and covariance estimation values of the motion models according to the mixed state estimation values and the mixed covariance estimation values. In some embodiments, the nonlinear system state space model includes state estimatesAnd observed estimates: Wherein, the ,AndThe x-axis and y-axis coordinates at time k,AndThe speeds in the x-axis and y-axis directions, respectively, k being the time of day; a state transition matrix of the motion model at the moment k-1, For the noise weight matrix to be a weighted matrix of the noise,As a function of the non-linear measurement,In order for the process to be noisy,To measure noise, andAndIndependent of each other. In some embodiments, the determining the hybrid state estimate and the hybrid covariance estimate for each of the motion models based on the first state estimate, the first covariance estimate, and the model probabilities comprises: By fixing the transfer matrix And model probabilities for each of the motion models at time k-1Obtaining the prediction model probability of the motion model j at the k momentWherein the predictive model probabilityThe probability that the predicted model j is the correct model based on all the observation information up to the time k-1 at the time k is that the fixed transition matrixRepresenting t