CN-122020052-A - Ultra-ellipse extended target robust tracking method and system for abnormal noise environment

CN122020052ACN 122020052 ACN122020052 ACN 122020052ACN-122020052-A

Abstract

The invention discloses a super-ellipse extended target robust tracking method and system for an abnormal noise environment. The invention can effectively solve the problem that the performance of the traditional tracking technology is obviously reduced in a non-Gaussian and non-steady abnormal noise environment. By introducing an advanced noise modeling and inference mechanism, the robustness of the filter to the burst abnormal measurement value is remarkably improved, and tracking misalignment or divergence caused by noise interference is avoided. Meanwhile, the method enhances the adaptability to the geometric shape of the expansion target, can accurately track and estimate various regular and irregular shape targets such as ellipses, rectangles and the like, and realizes the stable and reliable joint estimation of the motion state and the shape parameters. Finally, under a complex noise scene, the method can obtain a more accurate and more stable target tracking track and shape outline, and the practicability and reliability of the whole tracking system are improved.

Inventors

CHEN HUI
QU HAIPING
CAI ZIYUE
SUN JIAWEN
MA ENZE
LI PENGFEI
CAI WENPENG
WANG LI
WANG XUXIN
ZHAO YONGHONG

Assignees

兰州理工大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (8)

1. The ultra-ellipse extended target robust tracking method facing to the abnormal noise environment is characterized by comprising the following steps of: S1, acquiring an initialized target state space based on a preset sampling period and a total time step; s2, based on the initialized target state space, a state space model and a measurement model are established, and measurement noise is generated; S3, modeling the process noise into Gaussian noise based on the established state space model and the measurement model, and obtaining a prediction result of the target state; S4, modeling abnormal measurement noise as GSTM distribution based on a prediction result of a target state to construct a layered Gaussian model, obtaining a closed updating formula through variable dB leaf inference, recursively updating a motion state by a Kalman filter according to the closed updating formula, and updating a shape state by particle filtering.
2. The method for robust tracking of a super-elliptical expansion target for an abnormal noise environment according to claim 1, wherein S1 comprises: Setting a sampling period and a total time step, initializing a motion state parameter and a shape state parameter of a target, wherein the motion state parameter comprises a target center position and a target speed, the shape state parameter is described by adopting a super-ellipse model, modeling process noise as zero-mean Gaussian noise, setting a motion state covariance and a shape state covariance, and modeling measurement noise as non-stationary non-Gaussian noise.
3. The method for robust tracking of ultra-elliptical expansion targets for an anomaly-oriented noise environment of claim 2, wherein S2 comprises: constructing a state vector of the expansion target based on the motion state parameter and the shape state parameter; based on the state vector, a state transition equation of the motion state parameter and a state transition equation of the shape state parameter are established; Based on the state transfer equation and the super elliptic geometry, a nonlinear measurement model for generating observation points is established; based on the nonlinear measurement model, a measurement model for filtering and equivalent measurement noise are derived.
4. The method for robust tracking of a super-elliptical expansion target for an abnormal noise environment according to claim 1, wherein said S3 comprises: modeling the measurement noise by adopting GSTM distribution, and modeling a measurement likelihood probability density function into a layered Gaussian model; based on the established layered Gaussian model, performing time update prediction on the motion state parameters by using a standard linear Gaussian system assumption; Based on the prediction result of the motion state parameter, the shape state parameter is predicted by using a random walk model in a time updating way.
5. The method of claim 1, wherein in S4, the step of modeling the anomaly measurement noise as a GSTM distribution to construct a layered gaussian model comprises: introducing Bernoulli random variables, gamma scaling variables and mixed probability variables, and representing GSTM distribution as a layered Gaussian model; based on the layered Gaussian model, a joint probability distribution including motion states, shape states, and hidden variables is constructed.
6. The method for robust tracking of a super-elliptical expansion target for an abnormal noise environment according to claim 1, wherein in S4, the step of obtaining a closed update formula by variational bayesian inference comprises: Factorization is carried out on posterior distribution of the joint probability distribution, so that a decomposition result is obtained; based on the decomposition result, deducing a closed updating formula of the posterior distribution of the variation of each hidden variable through iteration of minimizing KL divergence; based on the closed updating formula, a fixed point iteration is performed to update all variation parameters.
7. The method for robust tracking of ultra-elliptical expansion targets for an anomaly-based environment of claim 6, wherein in S4, the motion state is recursively updated by a kalman filter, and the step of updating the shape state by particle filtering comprises: In each iteration, based on the updated equivalent measurement noise covariance, a Kalman filtering formula is adopted to update the motion state parameters and the covariance thereof; And taking the updated motion state estimation as input, embedding the noise estimation provided by the variation inference into a shape likelihood function, and updating the shape state parameters by adopting a particle filtering algorithm.
8. An ultra-elliptical expansion target robust tracking system oriented to an abnormal noise environment, the system being configured to implement the method of any one of claims 1-7, comprising an initialization module, a construction module, a prediction module, and an update module; the initialization module is used for acquiring an initialized target state space based on a preset sampling period and a total time step; The construction module is used for establishing a state space model and a measurement model based on the initialized target state space and generating measurement noise; The prediction module is used for modeling the process noise into Gaussian noise based on the established state space model and the measurement model, and obtaining a prediction result of the target state; The updating module is used for modeling abnormal measurement noise as GSTM distribution based on a prediction result of a target state to construct a layered Gaussian model, obtaining a closed updating formula through variable dB leaf inference, recursively updating a motion state by a Kalman filter according to the closed updating formula, and updating a shape state by particle filtering.

Description

Ultra-ellipse extended target robust tracking method and system for abnormal noise environment Technical Field The invention relates to the field of signal processing and automatic control, in particular to a super-ellipse extended target robust tracking method and system for an abnormal noise environment. Background In the context of rapid development of multi-sensor fusion and intelligent perception technologies, extended target tracking (Extended Object Tracking, EOT) has become a critical task in modern radar, vision, lidar, etc. systems. Unlike point target tracking, the extended target not only has dynamic position and velocity, but also has a non-point spatial region. In order to improve modeling accuracy and physical interpretation capability, target shape modeling is incorporated into the state estimation process, forming an EOT framework for joint estimation of states and shapes. The current extended target modeling method mainly comprises a random matrix model (Random matrix model, RMM), a random hypersurface model (Random hypersurface model, RHM), a Gaussian process regression model (Gaussian process regression, GPR) and the like, wherein the hypersurface model expresses continuous shape change from ellipse to rectangle with few parameters, so that the problems of more parameters and high calculation cost of the RHM/GPR model are solved, and good compromise of modeling accuracy and calculation efficiency is achieved. The super-ellipse extended target tracking method has wide potential in the scenes of intelligent driving, target identification, abnormal state detection and the like due to the unified representation capability of the super-ellipse extended target tracking method on rectangular, square and elliptical shapes. In particular, the squareness parameter epsilon is a key for describing flexible and complex shapes of objects by controlling the transition of the shape from smooth to angular. However, under an abnormal noise or non-gaussian measurement environment, the parameter is very prone to estimation deviation or numerical oscillation, and becomes a bottleneck affecting overall tracking accuracy. Although most current extended target tracking methods commonly construct a filtering model based on gaussian white noise assumptions, this premise is often difficult to hold in a real-world environment. The data collected by the actual sensor is often affected by various factors, such as electromagnetic interference, occlusion, signal reflection anomalies, etc., which can deviate the measurement from the true value, exhibiting anomalies offset and thick tail distribution far above what is expected by the gaussian model. Under such conditions, the process noise and the measurement noise no longer follow the ideal gaussian distribution, but rather more closely follow the statistical properties of the thick-tailed or mixed distribution. In this case, the recursive assumption of the conventional Kalman filter is no longer valid, and direct use will lead to systematic deviations or fluctuations in the state estimate. To cope with this problem, researchers have proposed various Robust kalman filters (Robust KALMAN FILTERS, RKFS), such as improved filtering methods based on M-estimation, maximum correlation entropy, and minimum error entropy. These algorithms enhance the tolerance of the system to outliers to some extent, but most methods do not explicitly exploit the thick tail structure of the noise and still face the problem of loss of accuracy. In recent years, researchers have gradually turned to modeling optimization of the noise distribution itself. Among them, filters based on student t distribution have been widely used for stationary thick tail noise scenarios, such as t-filters and their robust variants. Although these methods have good performance in thick-tail noise environments, conventional t-filtering methods are difficult to flexibly adapt when the noise distribution is not smoothly varying (i.e., gaussian in time and variable in parameter thick-tail distribution). To address this problem, gaussian-Student's t Mixture (GSTM) was proposed to more flexibly characterize the uncertainty and stationarity of noise. The model adaptively models different types of noise by introducing mixed probability variables, and can automatically adjust the sensitivity of the filter to abnormal values according to observation data in the reasoning process, and flexibly switch between stable noise and unsteady noise. However, the existing GSTM model is mainly focused on position and velocity state estimation, and has not been expanded to the field of extended target tracking for complex shape modeling. Particularly, when a superellipse model is adopted to model an irregular target, a key parameter, namely square parameter epsilon, is particularly sensitive to abnormal measurement, and obvious estimation deviation is easily generated, so that the whole shape estimation is invalid. Therefore, under abnorm