WO2021008077A1

WO2021008077A1 - Multi-target tracking method and system under flicker noise

Info

Publication number: WO2021008077A1
Application number: PCT/CN2019/125850
Authority: WO
Inventors: 刘宗香; 黄炳坚; 武宏杰; 李良群
Original assignee: 深圳大学
Priority date: 2019-07-16
Filing date: 2019-12-17
Publication date: 2021-01-21
Also published as: CN110390684A

Abstract

A multi-target tracking method and system under flicker noise, applicable to the technical field of target tracking. The method comprises: by using a distribution function and a labeled multi-Bernoulli filter density of targets at the previous moment, obtaining a predicted distribution function and a predicted labeled multi-Bernoulli filter density of an existing target at the current moment by prediction (S101); setting a preset distribution function and a preset labeled multi-Bernoulli filter density for a newly generated target (S102); respectively merging the preset distribution function and the preset labeled multi-Bernoulli filter density of the newly generated target with the predicted distribution function and the predicted labeled multi-Bernoulli filter density of the existing target at the current moment to obtain a predicted distribution function and a predicted labeled multi-Bernoulli filter density of targets at the current moment (S103); converting the predicted labeled multi-Bernoulli filter density of the targets at the current moment into a predicted δ-generalized labeled multi-Bernoulli filter density, and processing the measurement at the current moment, and the predicted distribution function and the predicted δ-generalized labeled multi-Bernoulli filter density of the targets by means of a variational Bayesian method to obtain an updated distribution function and an updated δ-generalized labeled multi-Bernoulli filter density of the targets (S104); performing joint cropping on the predicted distribution function and the predicted δ-generalized labeled multi-Bernoulli filter density of the targets, and the updated distribution function and the updated δ-generalized labeled multi-Bernoulli filter density of the targets by means of Gibbs sampling, forming a candidate distribution function at the current moment from the remaining predicted distribution function and updated distribution function after cropping, and converting the remaining predicted δ-generalized labeled multi-Bernoulli filter density and updated δ-generalized labeled multi-Bernoulli filter density into a labeled multi-Bernoulli filter density to form a candidate labeled multi-Bernoulli filter density at the current moment (S105); and performing pruning and fusion processing on the candidate distribution function and the candidate labeled multi-Bernoulli filter density at the current moment to obtain the distribution function and the labeled multi-Bernoulli filter density of the targets at the current moment as an input of a filter at the next moment; estimating the number of targets at the current moment according to the labeled multi-Bernoulli filter density of the targets at the current moment, and calculating the existence probability of the targets at the current moment; and sequentially extracting target distribution functions having a large existence probability according to the estimated number of targets, and using the extracted target distribution function as an output of the filter at the current moment (S106). According to the method, the filter can accurately extract the target state of the targets at the current moment in the flicker noise environment, thereby improving the accuracy of multi-target tracking.

Description

Multi-target tracking method and system under flicker noise

cross reference

This application claims the priority of the Chinese application filed on July 16, 2019, the application number is 201910639220.1, the entire content of which is incorporated into this application by reference.

Technical field

This application relates to the technical field of target tracking, and in particular to a method and system for multi-target tracking under flicker noise.

Background technique

Tag Do Bernoulli filter can accurately estimate the number of targets and track target trajectories in clutter and noise environments, so it has been applied to many practical problems, such as radar target tracking, image data tracking, ground target tracking, sensor management, audio And other applications such as video data tracking, visual data tracking and cell tracking, and moving multi-target tracking.

However, the label multi-Bernoulli filter is mostly used when the noise is Gaussian noise, and the multi-target tracking effect in the flicker noise environment is not good. Therefore, how to effectively track multiple targets under flicker noise is a key technical problem that needs to be explored and resolved.

Summary of the invention

The main purpose of this application is to propose a multi-target tracking method and system under flicker noise, so as to solve the problem that the existing filters for multi-target tracking cannot be applied in the flicker noise environment.

To achieve the foregoing objective, the first aspect of the embodiments of the present application provides a multi-target tracking method under flicker noise, including:

Using the distribution function of each target at the previous moment and the label Dobernulli filter density, the predicted distribution function of the target at the current moment and the predicted label Dobernulli filter density are obtained through prediction;

Set the preset distribution function and the preset label Dobernuli filter density for the freshman target;

Combine the preset distribution function and the preset label Do Bernoulli filter density of the new target with the predicted distribution function and the predicted label Do Bernoulli filter density of the existing target at the current moment to obtain the current target Predicted distribution function and predicted label multi-Bernoulli filter density;

Wherein, each target at the current moment includes a target that already exists at the current moment and a new-born target at the current moment;

Convert the predicted label Do Bernoulli filter density of each target at the current moment into the predicted delta-generalized label Do Bernoulli filter density, and use the variational Bayes method to measure the current moment and the predicted distribution function of each target And predict the multi-Bernoulli filter density of δ-generalized label to obtain the updated distribution function of each target and update the multi-Bernoulli filter density of δ-generalized label;

Through Gibbs sampling, joint cropping is performed on the predicted distribution function of each target, predicted δ-generalized label Do Bernoulli filter density and the updated distribution function of each target, and updated δ-generalized label Do Bernoulli filter density ; The remaining predicted distribution function and updated distribution function after cropping form the candidate distribution function at the current moment, and the cropped remaining predicted δ-generalized label multi-Bernoulli filter density and updated δ-generalized label multi-Bernoulli filter density are converted Is the label multi-Bernoulli filter density to form the candidate label multi-Bernoulli filter density at the current moment;

Performing pruning and fusion processing on the candidate distribution function and the candidate label multi-Bernoulli filter density at the current moment to obtain the distribution function and label multi-Bernoulli filter density of each target at the current moment as the input of the next moment filter; According to the label Do Bernoulli filter density of each target at the current moment, the target number at the current moment is estimated, and the existence probability of each target at the current moment is calculated; and according to the estimated target number, the target distribution function with high probability of existence is extracted in turn. The fetched target distribution function is used as the output of the filter at the current moment.

The second aspect of the embodiments of the present application provides a multi-target tracking system under flicker noise, including:

The prediction module is used to use the distribution function of each target at the previous moment and the label Dobernuli filter density to obtain the predicted distribution function of the existing target at the current moment and the predicted label Doberman filter density;

Freshman target acquisition module, used to set a preset distribution function and a preset label Dobernuli filter density for the freshman target;

The merging module is used to merge the preset distribution function and the preset label Dobernulli filter density of the new target with the predicted distribution function and the predicted label Dobernuli filter density of the existing target at the current moment to obtain The predicted distribution function of each target at the current moment and the predicted label Do Bernoulli filter density;

The update module is used to convert the predicted label Do Bernoulli filter density of each target at the current moment into the predicted delta-generalized label Do Bernoulli filter density, and use the variational Bayes method to measure the current moment, each The predicted distribution function of the target and the predicted δ-generalized label multi-Bernoulli filter density are processed to obtain the updated distribution function of each target and the updated δ-generalized label multi-Bernoulli filter density;

The cropping module is used to predict the distribution function of each target, predict the δ-generalized label Do Bernoulli filter density and update the distribution function of each target through Gibbs sampling, update the δ-generalized label Do Bernoulli filter density The filtering density is jointly cropped; the remaining predicted distribution function and updated distribution function after cropping form the candidate distribution function at the current moment, and the remaining predicted δ-generalized label Dobernulli filter density after cropping and updated δ-generalized label Dober The Nouli filter density is converted to the label Dobernuoli filter density to form the candidate label Dobernuoli filter density at the current moment;

The extraction module is used to perform pruning and fusion processing on the candidate distribution function at the current moment and the candidate label Dobernuelli filter density to obtain the distribution function of each target at the current moment and the label Dobernuelli filter density as the next moment The input of the filter; estimate the number of targets at the current time according to the label Do Bernoulli filter density of each target at the current time, and calculate the existence probability of each target at the current time; and according to the estimated number of targets, the target distribution function with high probability is sequentially Extracted, and the extracted target distribution function is used as the output of the filter at the current moment.

The embodiment of the present application proposes a method for tracking multiple targets under flicker noise. The targets are divided into targets at the previous moment, existing targets at the current moment, and new targets according to time. Among them, the distribution functions and labels of the targets at the previous moment are many. The Bernoulli filter density is known, and it is used to predict the distribution function of the existing target at the current moment and the label Dobernulli filter density. At the same time, set the preset distribution function and the preset label Doberman filter density for the new target, and change the current The distribution function of the target at the moment and the label Do Bernoulli filter density are respectively combined with the preset distribution function of the new target and the preset label Do Bernoulli filter density to obtain the predicted distribution function and predicted label Do Bernoulli of each target at the current moment Filtering density, convert the predicted label Dobernullian filter density of each target at the current moment into the form of δ-generalized label Dobernullian filter density, and then update to obtain the updated distribution function of the target at the current moment and update δ-generalized Tag multi-Bernoulli filter density, through Gibbs sampling to predict the target's predicted distribution function, predict δ-generalized label multi-Bernoulli filter density and target update distribution function, update δ-generalized label multi-Bernoulli filter density Joint cropping. After cropping, the remaining predicted distribution function and updated distribution function form the candidate distribution function at the current moment. At the same time, the cropped remaining predicted δ-generalized label multi-Bernoulli filter density and updated δ-generalized label multi-Bernoulli filter The density is converted to the label multi-Bernoulli filter density to form the candidate label multi-Bernoulli filter density at the current moment, and before the candidate distribution function and the candidate label multi-Bernoulli filter density are input to the filter, they are also pruned Fusion processing, so as to extract the effective components again on the basis of the candidate distribution function and the candidate label Dobery filter density, which constitutes the distribution function of the current time target and the label Doberman filter density, the above distribution function and label Doberman The Nuuli filter density, as the input of the filter at the next moment, estimates the number of targets at the current moment based on the label Dobernuoli filter density of each target at the current moment, calculates the existence probability of each target at the current moment, and according to the estimated target number, The target distribution function with high probability of existence is extracted in turn, and the extracted target distribution function is used as the output of the filter at the current moment. Among them, the target with high probability is the tracking target of the filter, and the output of the filter is used to describe the target. The multi-target tracking method under flicker noise provided by the embodiment of the present application can accurately extract the target state of the total target in the flicker noise environment, thereby improving the accuracy of multi-target tracking in the flicker noise environment .

Description of the drawings

FIG. 1 is a schematic diagram of the implementation process of the multi-target tracking method under flicker noise according to Embodiment 1 of the application;

2 is a schematic diagram of the composition structure of a multi-target tracking system under flicker noise provided in the second embodiment of the application;

3 is the measurement data of 50 scanning periods of the sensor provided in the third embodiment of the application;

4 is a filter output result obtained by processing according to the multi-target tracking method under flicker noise in the first embodiment provided by the third embodiment of the application;

FIG. 5 is a filter output result obtained by processing the existing VB-PHD filtering method under flicker noise according to the third embodiment of the present application;

6 is a schematic diagram of the average OSPA distance obtained after 100 experiments according to the multi-target tracking method under flicker noise in the first embodiment and the VB-PHD filtering method;

FIG. 7 is a schematic diagram of the estimation of the number of targets obtained after 100 experiments according to the multi-target tracking method under flicker noise in the first embodiment and the VB-PHD filtering method.

The realization, functional characteristics, and advantages of the purpose of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

Detailed ways

It should be understood that the specific embodiments described here are only used to explain the application, and are not used to limit the application.

It should be noted that in this article, the terms "include", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, It also includes other elements not explicitly listed, or elements inherent to the process, method, article, or device. If there are no more restrictions, the element defined by the sentence "including a..." does not exclude the existence of other identical elements in the process, method, article or device that includes the element.

In this document, suffixes such as “module”, “part” or “unit” used to indicate elements are only used for the description of this application, and have no specific meaning in themselves. Therefore, "module" and "component" can be mixed.

In the following description, the serial number of the application embodiment is only for description, and does not represent the merits of the embodiments.

Example one

As shown in FIG. 1, an embodiment of the present application proposes a multi-target tracking method under flicker noise, including steps S101 to S106.

Among them, the target at the previous moment and the target at the current moment refer to multiple tracking targets at different moments, and the distribution of the targets in the area is represented by the distribution function and the label Dobernulli filter density.

Among them, steps S101 to S103 are prediction steps, step S104 is an update step, step S105 is a cropping step, and step S106 is an output step.

In the embodiment of the present application, the implementation process of step S101 may be:

Let k-1 represent the previous moment, k represent the current moment, t _k-1 represent the time of the previous moment, and t _k represent the time of the current moment;

The observation noise at the current moment obeys the Student's t distribution, expressed as:

Among them, z _k represents the measured value at time k,

Indicates the mean value of the measurement, Λ _k is the accuracy matrix, and λ _k is the degree of freedom of t distribution;

The distribution function of each target at the previous moment is expressed as:

Among them, N represents the Gaussian distribution, IG represents the inverse gamma distribution, x _k-1 represents the state component at the previous moment, m _k-1 represents the state estimate mean, P _k-1 represents the covariance matrix, and R _k-1 represents the noise Variance matrix, d represents the dimensions of the inverse gamma distribution parameters α _k-1 and β _k-1 ;

The label Do Bernoulli filter density of each target at the previous moment is expressed as:

among them,

Represents the label space at time k-1,

Indicates the target tag, t is used to record the corresponding time, i is a non-repeated positive integer to distinguish other targets at the same time,

Is the probability of existence,

Is the probability density,

Is the weight,

According to the distribution function of each target at the previous moment, the predicted distribution function of the target existing at the current moment is obtained, and the formula is:

Among them, m _k,S =F _k-1 m _k-1 ,

α _k,S =ρ _α α _k-1 , β _k,S =ρ _β β _k-1 , x _k is the current state component, F _k-1 is the state transition matrix, Q _k-1 is the process noise variance Matrix, ρ _α and ρ _β are the propagation factors;

According to the label Do Bernoulli filter density of each target at the previous moment, the predicted label Do Bernoulli filter density of the existing target at the current moment is obtained, and the formula is:

among them,

Is the target survival probability, f(x|x') is the single target transition density.

In the embodiment of the present application, the implementation process of step S102 may be:

The preset distribution function of the newborn target is:

Among them, x _k is the state component at time k, m _{k, B} is the estimated mean value of the state of the newborn target, P _{k, B} is the covariance matrix of the newborn target, and α _{k, B} and β _{k, B} are the inverse _G of the newborn target The parameters of the Mar distribution;

The preset label Do Bernoulli filter density of the newborn target is:

among them,

Represents the label space of the freshman goal,

Is the probability of existence of the new goal,

Is the probability density,

Is the weight.

In the embodiment of the present application, the implementation process of step S103 may be:

Combine the preset distribution function of the new-born target and the predicted distribution function of the existing target at the current moment to obtain the predicted distribution function of each target at the current moment. The formula is:

Combining the preset label Do Bernoulli filter density of the newborn target and the predicted label Do Bernoulli filter density of the existing target at the current moment to obtain the predicted label Do Bernoulli filter density of each target at the current moment, The formula is:

among them,

In steps S101 to S103, a heuristic method is used to generate the parameters of the inverse gamma distribution, so as to predict the distribution function of the target at the current moment and the label multiple Bernoulli filter density according to the distribution function of the target at the previous moment and the label multi-Bernoulli filter density. Nouri filter density, and set the preset distribution function and preset label Do Bernoulli filter density for the newborn target; then set the preset distribution function and preset label Do Bernoulli filter density of the newborn target with the current time. The predicted distribution function of the target and the predicted label Dobernulli filter density are combined to obtain the predicted distribution function of each target at the current moment and the predicted label Dobernullian filter density.

In step S104, the predicted label Dobernullian filter density of each target at the current moment is converted into a delta-generalized label Dobernullian filter density, and the variational Bayes method is used to process the predicted distribution function of each target at the current moment. And predict the δ-generalized label multi-Bernoulli filter density, obtain the updated distribution function of each target and update the δ-generalized label multi-Bernoulli filter density, and provide a data basis for the joint cropping in the next step.

In the embodiment of the present application, the implementation process of step S104 may be:

Convert the predicted label Dobernulli filter density of each target at the current moment into the form of δ-generalized label Dobernulli filter density to obtain the predicted δ-generalized label Dobernuli filter density, the formula is:

among them,

for

A limited subset of

Use the variational Bayes method to obtain the updated distribution function of each target at the current moment, the formula is:

Among them, m _k =m _k|k-1 +K _k v _k , P _k =P _k|k-1 -K _k H _k P _k|k-1 ,

among them,

v _k = z _k -H _k m _k|k-1 ,

H _k is the observation matrix;

will

with

Instead of

with

Obtain R _k , and perform iterative update until the difference between the two before and after m _{k in the} iterative process is less than the first threshold or reaches the maximum number of iterations to obtain updated m _k , P _k , α _k and β _k ;

Obtain the updated δ-generalized label Do Bernoulli filter density of each target at the current moment, the formula is:

Among them, θ _k ∈Θ _k represents the 1-1 mapping from the label to the observation set:

among them,

Is the probability of detection,

Is the probability of missed detection, and k(z) is the noise confounding degree that obeys the Poisson distribution.

In step S105, the predicted distribution function of each target, the predicted δ-generalized label Do Bernoulli filter density and the updated distribution function of each target, update the δ-generalized label Do Bernoulli filter density of each target through Gibbs sampling The filtering density is jointly cropped; the remaining predicted distribution function and updated distribution function after cropping form the candidate distribution function at the current moment, and the remaining predicted δ-generalized label Dobernulli filter density after cropping and updated δ-generalized label Dober The Nulli filter density is converted into the label Dobernulli filter density to form the candidate label Dobernulli filter density at the current moment.

In the embodiment of the present application, the implementation process of step S105 may be:

Combining the predicted delta-generalized label Do Bernoulli filter density and the updated delta-generalized label Do Bernoulli filter density to obtain:

among them,

Represents the weight corresponding to the δ-generalized label Do Bernoulli filter density converted from the label Do Bernoulli filter density of the target at the previous moment,

among them,

M is the number of observations at the current moment, and P is the target number at the current moment;

Solve with Gibbs sampling method

get

The larger value of the γ vector set is to select the weight value

Larger components, resulting in larger weights

set;

The predicted distribution function of each target, the predicted delta-generalized label Dobernulli filter density, the updated distribution function of each target, and the updated delta-generalized label Dobernuli filter density are jointly cropped, and the weight value is deleted. The distribution function corresponding to the small target and the δ-generalized label multi-Bernoulli filter density. Take the remaining predicted distribution function and updated distribution function after cropping as the candidate distribution function at the current moment, and convert the cropped remaining predicted δ-generalized label Dobernuoli filter density and updated δ-generalized label Dobernuoli filter density Do Bernoulli filter density for the label, namely

among them,

Is the cropped weight value, which is used to obtain the Do Bernoulli filter density of the candidate label at the current moment.

In step S106, before inputting the candidate distribution function and the candidate label multi-Bernoulli filter density to the filter, the pruning and fusion processing is also performed on them, so that the candidate distribution function and the candidate label multi-Bernoulli filter density are again based on The effective components are extracted to form the distribution function of the target at the current moment and the label Do Bernoulli filter density. The above distribution function and label Do Bernoulli filter density are used as the input of the filter at the next moment. The label multi-Bernoulli filter density estimates the number of targets at the current moment, calculates the existence probability of each target at the current moment, and according to the estimated number of targets, sequentially extracts the target distribution functions with high probability of existence, and the extracted target distribution function is taken as The output of the filter at the current moment.

In the embodiment of the present application, the implementation process of step S106 may be:

Obtaining the candidate distribution function of each target at the current moment and the Dobernuelli filter density of the candidate label through a filter to obtain the target trajectory of each target at the current moment;

Selecting the target trajectory whose existence probability is greater than a second threshold;

Perform pruning and fusion on the selected components in the target trajectory, and delete components with a weight value less than a third threshold;

Perform a weighted average of the remaining components to obtain the fused components, so as to obtain the distribution function of each target at the current moment and the label Dobernuli filter density;

Using the distribution function of each target at the current moment and the label Dobernuli filter density as the input of the filter at the next moment;

Estimate the number of targets at the current time based on the label Do Bernoulli filter density of each target at the current time, calculate the target existence probability of each target at the current time, and according to the estimated target number, sequentially calculate the target distribution function with high probability Extract out

The extracted target distribution function is used as the output of the filter.

In practical applications, the mean value of the distribution function output at the kth moment is the state estimate of each target at the current moment, and the covariance of the distribution function output at the kth moment is the error estimate of each target at the current moment.

Example two

As shown in FIG. 2, the second embodiment of the present application also provides a multi-target tracking system 20 under flicker noise, including but not limited to the following components:

The prediction module 21 is used to use the distribution function of each target at the previous moment and the label Dobernuli filter density to obtain the predicted distribution function of the existing target at the current moment and the predicted label Dobermanu filter density;

The freshman target acquisition module 22 is used to set a preset distribution function and a preset label Dobernuli filter density for the freshman target;

The merging module 23 is used for merging the preset distribution function and the preset label Do Bernoulli filter density of the new target with the predicted distribution function and the predicted label Do Bernoulli filter density of the existing target at the current moment, respectively, to obtain The predicted distribution function of the target and the multi-Bernoulli filter density of the predicted label;

Among them, the goals at the current moment include the existing goals at the current moment and the new-born goals at the current moment;

The update module 24 is used to convert the predicted label Dobernulli filter density of each target at the current moment into a delta-generalized label Dobernuli filter density, and use the variational Bayes method to measure the current moment and predict each target Processing the distribution function and predicting the δ-generalized label Dobernuelli filter density to obtain the updated distribution function of each target and the updated δ-generalized label Dobernuelli filter density;

The cropping module 25 is used to predict the distribution function of each target through Gibbs sampling, predict the δ-generalized label Do Bernoulli filter density and update the distribution function of each target, update the δ-generalized label Do Bernoulli filter The density is jointly cropped; the remaining predicted distribution function and the updated distribution function after cropping form the candidate distribution function at the current moment, and the remaining predicted δ-generalized label Do Bernoulli filter density after cropping and the updated δ-generalized label Dobernu Convert the benefit filter density to the label Dobernulli filter density to form the candidate label Dobernuli filter density at the current moment;

The extraction module 26 is used to perform pruning and fusion processing on the joint distribution function at the current moment and the joint label Dobernuelli filter density to obtain the distribution function of each target at the current moment and the label Dobernuelli filter density as the next time filter The input of the device; estimate the number of targets at the current time according to the label Dobernulli filter density of each target at the current time, calculate the existence probability of each target at the current time; and according to the estimated number of targets, sequentially extract the target distribution function with high probability of existence Then, the extracted target distribution function is used as the output of the filter at the current moment.

Example three

The embodiment of the present application also uses practical applications to illustrate the target tracking effect of the multi-target tracking method under flicker noise and the multi-target tracking system under flicker noise in the first and second embodiments above.

In the embodiment of this application, 6 targets moving in the two-dimensional observation space [-1000(m), 1000(m)]×[-1000(m), 1000(m)] are selected as tracking targets.

The target motion observation time is 50s.

The target state is composed of position and speed, expressed as

Where η _i,x and η _i,y represent position components

with

Represents the velocity component;

The state transition matrix of the target is

The process noise covariance matrix is

Δt=t _k -t _k-1 is the time difference between the current moment and the previous moment, and σ _v is the process noise standard deviation.

The observation matrix is

The observed noise covariance matrix is

σ _w is the standard deviation of the observation noise, and the observation noise is the t-distribution noise that obeys λ _k =2.

In order to generate simulation data, relevant parameters are set in the simulation experiment: p _S =1.0, p _D =0.90, k(z)=5.0×10 ^-7 m ^-2 , σ _v =1ms ^-2 , and σ _w =1m.

As shown in FIG. 3, in the embodiment of the present application, 50 scanning periods of the sensor are used as an experimental simulation, and FIG. 3 is the simulation observation data collected by the sensor.

In order to process the above simulation data, the relevant parameters of the probability hypothesis density filter under flicker noise are set as: transfer factor ρ _α =ρ _β =0.98, the first threshold is 0.01, the second threshold is 10 ^-3 , and the third threshold is 10 ^-5 , the initial value of the inverse gamma distribution parameter α ₀ =[160,160] ^T ,β ₀ =[2300,2300] ^T , set the covariance of the newborn target as P _B =(diag(50,25,50,25 )) ² , the existence probability is r _B =0.03, and the weight is w _B =1.

Based on the above simulation data, FIG. 4 shows the filter output results obtained by processing the existing VB-PHD filtering method under flicker noise, and Figure 5 shows the processing obtained by the multi-target tracking method under flicker noise proposed in the embodiment of the present application. The output result of the filter, the circle represents the tracked target, showing the target tracking effect based on Figure 3. In Figures 4 and 5, the abscissa represents the horizontal distance between the target and the origin, in m, and the ordinate represents the longitudinal distance between the target and the origin, in m.

The embodiment of the application also processes the simulation data of FIG. 3 separately according to the multi-target tracking method under flicker noise provided in Embodiment 1, and the existing VB-PHD filtering method under flicker noise, and performs 100 Monte Carlo experiments. Two statistical results were obtained. Among them, Fig. 6 is the statistical result of the average OSPA (Optimal Subpattern Assignment) distance of the two, and Fig. 7 is the statistical result of the target quantity estimation of the two.

In Figures 6 and 7, the curve with "+" represents the filtering effect of using the VB-PHD filter, and the curve with "*" represents the filtering effect of the multi-target tracking method under flicker noise provided in the first embodiment. For VB-LMB filter, the curve with "|" indicates the number of real targets. In Figure 6, the ordinate represents the average OSPA distance, the unit is m, the abscissa represents the time, the unit is s; in Figure 7, the ordinate represents the target number estimation, the unit is 1, the abscissa represents the time, the unit is s.

It can be seen that, comparing the existing VB-PHD filtering method based on flicker noise with the multi-target tracking method under flicker noise provided in the first embodiment of this application, the multi-target tracking method under flicker noise of this application can be more accurate Estimate the number of targets, the OSPA distance is smaller than the OSPA distance obtained by the existing method. Therefore, the multi-target tracking method under flicker noise and the multi-target tracking system under flicker noise in the first and second embodiments of the present application can be The filter can accurately estimate the number of targets and extract the target distribution function in the flicker noise environment, thereby improving the accuracy of multi-target tracking in the flicker noise environment.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit it; although the foregoing embodiments describe the present application in detail, those of ordinary skill in the art should understand that they can still compare the foregoing embodiments. The recorded technical solutions are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the application, and should be included in this Within the scope of protection applied for.

Claims

A multi-target tracking method under flicker noise is characterized in that it comprises:

Using the distribution function of each target at the previous moment and the label Dobernulli filter density, the predicted distribution function of the target at the current moment and the predicted label Dobernulli filter density are obtained through prediction;

Set the preset distribution function and the preset label Dobernuli filter density for the freshman target;

Combine the preset distribution function and the preset label Do Bernoulli filter density of the new target with the predicted distribution function and the predicted label Do Bernoulli filter density of the existing target at the current moment to obtain the current target Predicted distribution function and predicted label multi-Bernoulli filter density;

Wherein, each target at the current moment includes an existing target at the current moment and a newborn target at the current moment;

Convert the predicted label Do Bernoulli filter density of each target at the current moment into the predicted δ-generalized label Do Bernoulli filter density, and use the variational Bayes method to measure the current moment, the predicted distribution function of each target, and Predict the δ-generalized label multi-Bernoulli filter density for processing, obtain the updated distribution function of each target and update the δ-generalized label multi-Bernoulli filter density;

Through Gibbs sampling, joint cropping is performed on the predicted distribution function of each target, predicted δ-generalized label Do Bernoulli filter density and the updated distribution function of each target, and updated δ-generalized label Do Bernoulli filter density ; The remaining predicted distribution function and updated distribution function after cropping form the candidate distribution function at the current moment, and the cropped remaining predicted δ-generalized label multi-Bernoulli filter density and updated δ-generalized label multi-Bernoulli filter density are converted Is the label multi-Bernoulli filter density to form the candidate label multi-Bernoulli filter density at the current moment;

Performing pruning and fusion processing on the candidate distribution function and the candidate label multi-Bernoulli filter density at the current moment to obtain the distribution function and label multi-Bernoulli filter density of each target at the current moment as the input of the next moment filter; According to the label Do Bernoulli filter density of each target at the current moment, the target number at the current moment is estimated, and the existence probability of each target at the current moment is calculated; and according to the estimated target number, the target distribution function with high probability of existence is extracted in turn. The fetched target distribution function is used as the output of the filter at the current moment.
The method for multi-target tracking under flicker noise according to claim 1, wherein the distribution function of each target at the previous time and the label Dobernulli filter density are used to obtain the predicted distribution function of the existing target at the current time through prediction. And predict the label multi-Bernoulli filter density, including:

Let k-1 represent the previous moment, k represent the current moment, t k-1 represent the time of the previous moment, and t k represent the time of the current moment;

The observation noise at the current moment obeys the Student's t distribution, expressed as:

Among them, z k represents the measured value at time k,
Indicates the mean value of the measurement, Λ k is the accuracy matrix, and λ k is the degree of freedom of t distribution;

The distribution function of each target at the previous moment is expressed as:

Among them, N represents the Gaussian distribution, IG represents the inverse gamma distribution, x k-1 represents the state component at the previous moment, m k-1 represents the state estimate mean, P k-1 represents the covariance matrix, and R k-1 represents the noise Variance matrix, d represents the dimensions of the inverse gamma distribution parameters α k-1 and β k-1 ;

The label Do Bernoulli filter density of each target at the previous moment is expressed as:

among them,
Represents the label space at time k-1,
Indicates the target tag, t is used to record the corresponding time, i is a non-repeated positive integer to distinguish other targets at the same time,
Is the probability of existence,
Is the probability density,
Is the weight,

According to the distribution function of each target at the previous moment, the predicted distribution function of the target existing at the current moment is obtained, and the formula is:

Among them, m k,S =F k-1 m k-1 ,
α k,S =ρ α α k-1 , β k,S =ρ β β k-1 , x k is the current state component, F k-1 is the state transition matrix, Q k-1 is the process noise variance Matrix, ρ α and ρ β are the propagation factors;

According to the label Do Bernoulli filter density of each target at the previous moment, the predicted label Do Bernoulli filter density of the existing target at the current moment is obtained, and the formula is:

among them,

among them,
Is the target survival probability, f(x|x') is the single target transition density.
The method for multi-target tracking under flicker noise according to claim 1, wherein the setting of a preset distribution function and a preset label multi-Bernoulli filter density for new-born targets comprises:

The preset distribution function of the newborn target is:

Among them, x k is the state component at time k, m k, B is the estimated mean value of the state of the newborn target, P k, B is the covariance matrix of the newborn target, and α k, B and β k, B are the inverse G of the newborn target The parameters of the Mar distribution;

The preset label Do Bernoulli filter density of the newborn target is:

among them,
Represents the label space of the freshman goal,
Is the probability of existence of the new goal,
Is the probability density,
Is the weight.
The method for multi-target tracking under flicker noise according to claim 1, wherein the preset distribution function and the preset label Dobernullian filter density of the new target are respectively compared with those of the existing target at the current moment. The prediction distribution function and the prediction label Do Bernoulli filter density are combined to obtain the prediction distribution function and the prediction label Do Bernoulli filter density of each target at the current moment, including:

Combine the preset distribution function of the new-born target and the predicted distribution function of the existing target at the current moment to obtain the predicted distribution function of each target at the current moment. The formula is:

Combining the preset label Do Bernoulli filter density of the newborn target and the predicted label Do Bernoulli filter density of the existing target at the current moment to obtain the predicted label Do Bernoulli filter density of each target at the current moment, The formula is:

among them,
The method for multi-target tracking under flicker noise according to claim 1, wherein the predicted label Dobernullian filter density of each target at the current moment is converted into a predicted delta-generalized label Dobernullian filter density , Through the variational Bayesian method to process the current time measurement, the predicted distribution function of each target and the predicted δ-generalized label Dobery filter density to obtain the updated distribution function of each target and the updated δ-generalized label Dober Nuuli filter density, including:

Convert the predicted label Dobernulli filter density of each target at the current moment into the form of δ-generalized label Dobernulli filter density to obtain the predicted δ-generalized label Dobernuli filter density, the formula is:

among them,
for
A limited subset of

Use the variational Bayes method to obtain the updated distribution function of each target at the current moment, the formula is:

Among them, m k =m k|k-1 +K k v k , P k =P k|k-1 -K k H k P k|k-1 ,

among them,
v k = z k -H k m k|k-1 ,
H k is the observation matrix;

will
with
Instead of
with
Obtain R k , and perform iterative update until the difference between the two before and after m k in the iterative process is less than the first threshold or reaches the maximum number of iterations to obtain updated m k , P k , α k and β k ;

Obtain the updated δ-generalized label Do Bernoulli filter density of each target at the current moment, the formula is:

Among them, θ k ∈Θ k represents the 1-1 mapping from the label to the observation set:

among them,

Is the probability of detection,
Is the probability of missed detection, and k(z) is the noise confounding degree that obeys the Poisson distribution.
The method for multi-target tracking under flicker noise according to claim 1, wherein the predicted distribution function of each target, the predicted delta-generalized label Dobernulli filter density and the each target are predicted by Gibbs sampling. The updated distribution function of the target and the updated δ-generalized label multi-Bernoulli filter density are jointly cropped; the remaining predicted distribution function after cropping and the updated distribution function form the candidate distribution function at the current moment, and the remaining predicted δ-generalized after cropping The label multi-Bernoulli filter density and update δ-generalized label multi-Bernoulli filter density are converted into the form of label multi-Bernoulli filter density to form the candidate label multi-Bernoulli filter density at the current moment, including:

Combining the predicted δ-generalized label Do Bernoulli filter density and the updated δ-generalized label Do Bernoulli filter density to obtain:

among them,
Represents the weight corresponding to the δ-generalized label Do Bernoulli filter density converted from the label Do Bernoulli filter density of the target at the previous moment,

Among them, S i,j =1 {1:M} (j)δ γi [j]+δ M+j [j]δ γi [0]+δ M+P+j [j]δ γi [-1] ,

M is the number of observations at the current moment, and P is the target number at the current moment;

Solve with Gibbs sampling method
get
The set of γ vectors with larger values, that is, select the weight
Larger components, resulting in larger weights
set,

The predicted distribution function of each target, the predicted delta-generalized label Dobernulli filter density, the updated distribution function of each target, and the updated delta-generalized label Dobernuli filter density are jointly cropped, and the weight value is deleted. The distribution function corresponding to the small target and the δ-generalized label multi-Bernoulli filter density.

Take the remaining predicted distribution function and updated distribution function after cropping as the candidate distribution function at the current moment, and convert the cropped remaining predicted δ-generalized label Do Bernoulli filter density and updated δ-generalized label Do Bernoulli filter density into Label multi-Bernoulli filter density, namely

among them,
Is the cropped weight value, which is used to obtain the Do Bernoulli filter density of the candidate label at the current moment.
The multi-target tracking method under flicker noise according to claim 1, wherein the current time candidate distribution function and candidate label multi-Bernoulli filter density are pruned and fused to obtain the current time target The distribution function and the label Dobernullian filter density are used as the input of the filter at the next moment; according to the label Dobernery filter density of each target at the current moment, estimate the number of targets at the current moment, and calculate the existence probability of each target at the current moment; According to the estimated target number, the target distribution functions with high probability of existence are sequentially extracted, and the extracted target distribution function is used as the output of the filter at the current moment, including:

Obtaining the candidate distribution function of each target at the current moment and the Dobernuelli filter density of the candidate label through a filter to obtain the target trajectory of each target at the current moment;

Selecting the target trajectory whose existence probability is greater than a second threshold;

Perform pruning and fusion on the selected components in the target trajectory, and delete components with a weight value less than a third threshold;

Perform a weighted average of the remaining components to obtain the fused components, so as to obtain the distribution function of each target at the current moment and the label Do Bernoulli filter density;

Taking the distribution function of each target at the current moment and the label Dobernulli filter density as the input of the filter at the next moment;

Estimating the number of targets at the current moment according to the label Do Bernoulli filter density of each target at the current moment, calculating the existence probability of each target, and sequentially extracting target distribution functions with high probability of existence according to the estimated target number;

The extracted target distribution function is used as the output of the current time filter.
A multi-target tracking system under flicker noise is characterized in that it comprises:

The prediction module is used to use the distribution function of each target at the previous moment and the label Dobernuli filter density to obtain the predicted distribution function of the existing target at the current moment and the predicted label Dobernuli filter density;

Freshman target acquisition module, used to set a preset distribution function and a preset label Dobernuli filter density for the freshman target;

The merging module is used to merge the preset distribution function and the preset label Dobernulli filter density of the new target with the predicted distribution function and the predicted label Dobernuli filter density of the existing target at the current moment to obtain The predicted distribution function of each target at the current moment and the predicted label Do Bernoulli filter density;

Wherein, each target at the current moment includes a target that already exists at the current moment and a new-born target at the current moment;

The update module is used to convert the predicted label Do Bernoulli filter density of each target at the current moment into the predicted delta-generalized label Do Bernoulli filter density, and use the variational Bayes method to measure the current moment and each target The predicted distribution function and predicted δ-generalized label multi-Bernoulli filter density are processed to obtain the updated distribution function of each target and the updated δ-generalized label multi-Bernoulli filter density;

The cropping module is used to predict the distribution function of each target, predict the δ-generalized label Do Bernoulli filter density and update the distribution function of each target through Gibbs sampling, update the δ-generalized label Do Bernoulli filter density The filtering density is jointly cropped; the remaining predicted distribution function and the updated distribution function after cropping form the candidate distribution function at the current moment, and the remaining predicted δ-generalized label multi-Bernoulli filter density after cropping and the updated δ-generalized label Dober The Nouli filter density is converted to the label Dobernuli filter density to form the candidate label Dober Noli filter density at the current moment;

The extraction module is used to perform pruning and fusion processing on the candidate distribution function at the current moment and the candidate label Dobernuelli filter density to obtain the distribution function of each target at the current moment and the label Dobernuelli filter density as the next moment The input of the filter; estimate the number of targets at the current time according to the label Do Bernoulli filter density of each target at the current time, and calculate the existence probability of each target at the current time; and according to the estimated number of targets, the target distribution function with high probability is sequentially Extracted, and the extracted target distribution function is used as the output of the filter at the current moment.