CN111582159A - Maneuvering target tracking method facing monitoring system - Google Patents

Maneuvering target tracking method facing monitoring system Download PDF

Info

Publication number
CN111582159A
CN111582159A CN202010377925.3A CN202010377925A CN111582159A CN 111582159 A CN111582159 A CN 111582159A CN 202010377925 A CN202010377925 A CN 202010377925A CN 111582159 A CN111582159 A CN 111582159A
Authority
CN
China
Prior art keywords
target
time
mth
motion
ith particle
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010377925.3A
Other languages
Chinese (zh)
Other versions
CN111582159B (en
Inventor
高磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Aeronautical Radio Electronics Research Institute
Original Assignee
China Aeronautical Radio Electronics Research Institute
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Aeronautical Radio Electronics Research Institute filed Critical China Aeronautical Radio Electronics Research Institute
Priority to CN202010377925.3A priority Critical patent/CN111582159B/en
Publication of CN111582159A publication Critical patent/CN111582159A/en
Application granted granted Critical
Publication of CN111582159B publication Critical patent/CN111582159B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • G06F18/295—Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/277—Analysis of motion involving stochastic approaches, e.g. using Kalman filters
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Data Mining & Analysis (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses a maneuvering target tracking method facing a monitoring system, which comprises the following steps: s1, constructing a maneuvering target motion model, and estimating the motion state of the target; s2, establishing a target association hypothesis set at the moment t and a motion state set corresponding to each target; s3, traversing each target in the target association hypothesis set at the previous moment to calculate various possible motion modes of the motion state of each target at the current moment; s4, constructing a measurement-target distribution matrix according to the possible motion mode of the motion state and the measurement set of the tracking gate at the current moment, and calculating the previous K target association hypotheses with the highest confidence; and S5, updating the motion state set corresponding to each target by adopting a Kalman filter method. The invention can learn unknown motion state change in a monitoring environment, and solves the problems of unknown motion state model set, model selection, transition probability selection between unknown model sets and the like.

Description

Maneuvering target tracking method facing monitoring system
Technical Field
The invention relates to the technical field of target tracking methods, in particular to a maneuvering target tracking method facing a monitoring system, and especially relates to a maneuvering target tracking method based on a layered Dirichlet process-hidden Markov model.
Background
A general target tracking method such as a multi-hypothesis tracking method usually uses a data association technology, and the optimal data association problem is solved in a theoretical level. However, the target tracking system may often use various kinds of feature information of the target to improve the target tracking accuracy, such as the motion features of the target. However, the target maneuvering characteristics are dynamic and complex time sequence data, targets are dynamically switched among various motion models, and the tracking effectiveness can be improved by learning the change rule of the motion models. By constructing the target dynamic system model, the target dynamic maneuvering characteristics and the system parameters can be incorporated into the dynamic system model for estimation.
The existing literature retrieval finds that a maneuvering target tracking system usually uses an interactive multi-model method and various improvement methods thereof, realizes hybrid estimation of motion states by defining a plurality of possible motion model sets, and has the core idea that different model sets are constructed to be converted among different model sets according to the motion states. The interactive multi-model method has the problems of unknown model set and model selection, transition probability selection among unknown model sets and the like. In fact, in an application scenario, the maneuvering motion mode of the target is unknown, and the number of the motion modes is accumulated and increased along with time, so that the ability of self-adaptive learning of model parameters by a nonparametric Bayes method is needed, and the tracking accuracy and stability are improved.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a maneuvering target Tracking method facing a monitoring system, which expands a layered Dirichlet process (HDP) -Hidden Markov Model (HMM) in a nonparametric Bayes method, combines an HDP-HMM filtering algorithm with a Multiple Hypothesis Tracking algorithm (MHT), can learn unknown movement state changes under the condition of clutter interference in a monitoring environment, solves the problems of unknown movement state Model sets, Model selection, transition probability selection among unknown Model sets and the like, and jointly estimates the movement Model and the movement state by adopting methods such as particle filtering and the like in the Tracking process.
The invention aims to be realized by the following technical scheme:
a maneuvering target tracking method facing a monitoring system comprises the following steps:
s1, constructing a maneuvering target motion model, and estimating the motion state of the target;
s2, establishing a target association hypothesis set omega at the time ttTarget association hypothesis set ΩtA set comprising possible association combinations between the motion states and observations of the targets and the motion states of each target;
s3, according to the last time t-1, the target association hypothesis set omegat-1And a motion state set corresponding to each target, traversing the target association hypothesis set
Figure BDA0002480726880000021
Using a Kalman filter to calculate various possible motion modes of the motion state of the target at the current moment tTarget of each target;
s4, constructing a measurement-target distribution matrix according to the obtained possible motion mode of the motion state of each target and the measurement set of each target falling into the tracking gate at the current moment, calculating and reserving the previous K target association hypotheses with the highest confidence degrees, and updating the possible association hypotheses between the targets and the observation in the target association hypothesis set;
and S5, updating the motion state set corresponding to each target by adopting a Kalman filter method based on the current time target association assumption set and the motion state set corresponding to each target.
Compared with the prior art, the invention has the following beneficial effects:
1. the invention provides a maneuvering target tracking method facing a monitoring system, which estimates the number of unknown motion modes and the transition relation between the modes by constructing a target dynamic system model. Compared with a standard IMM method, the method does not need to know the number and the transition probability of the motion modes, improves the tracking precision, and can be applied to the field of high-speed and high-maneuvering target tracking. The algorithm can adopt an implementation mode of measurement iterative processing, so that the calculation complexity meets the actual engineering requirement.
2. The invention combines a track-oriented multi-hypothesis tracking method with a layered Dirichlet process-hidden Markov model to process maneuvering target tracking in a clutter environment, realizes the functions of target track initiation, data association, hypothesis generation, hypothesis management, track maintenance and the like, and simplifies calculation by considering composite estimation in a multi-motion mode, so that the calculation complexity meets the actual engineering requirements.
3. The invention utilizes the particle filtering method, reduces the dimension of the system sampling space, improves the particle sampling efficiency and provides an on-line calculation method. The algorithm framework is clear and easy to realize, so that important technical support is provided for a maneuvering target tracking system in a complex environment.
Drawings
Fig. 1 is a schematic flow chart of a maneuvering target tracking method for a monitoring system.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting.
The maneuvering target tracking method for the monitoring system shown in the embodiment is realized in a pure computer readable program code mode, and comprises the following steps:
and S1, constructing a maneuvering target motion model, and estimating the motion state of the target.
In the monitoring system, the motion state of the maneuvering target is generally used as an unknown variable for describing system change, and in the embodiment, an HDP-HMM model is used for constructing a maneuvering target motion model which consists of a system state space equation, an observation equation and a maneuvering state generation process. The system state space equation represents the change condition of the motion state of the maneuvering target; the observation equation characterizes the observable maneuvering target motion state. The system state space equation and the observation equation form an HMM model of the movement of the maneuvering target. The maneuver state generation process characterizes the maneuver state of the maneuver target, which can be assumed to be affected by the motion control variable, which is based on the HDP process, since the maneuver state of the maneuver target is unknown, variable.
Hypothesis monitoringThe target group in the region comprises M ≧ 1 target, and defines a target system variable set { X ≧ 1t,Zt,UtIn which X istAnd UtAs a set of continuous variable vectors, ZtIs a discrete variable vector set. XtInvolving the movement states of M objects, ZtAnd UtThe unknown maneuvering motion mode and the unobservable motion control quantity of the M targets at the time t are respectively contained, and are defined as follows:
Xt＝(xt1,...,xtm,..,xtM),m＝1,...,M
Zt＝(zt1,...,ztm,..,ztM),m＝1,...,M
Ut＝(ut1,...,utm,..,utM),m＝1,...,M
wherein: x is the number oftmRepresenting the motion state estimate at the mth target instant t. The motion state of each target is a 6-dimensional vector representing the position, velocity and acceleration in a two-dimensional space. The two-dimensional space scene is selected to simplify calculation, and the method can be directly expanded to a three-dimensional space.
utmIs the unobservable control input of the mth target at the moment t, obeys the Gaussian distribution utm～N(μk,∑k)，μkIs utmMean value of ∑kIs utmThe covariance matrix of (2). Definition of thetak＝{μk,∑kIs the distribution parameter of the kth motor motion pattern, here assumed to be θkThe conjugate prior distribution of (d) is a Gaussian inverse Westst distribution NIW { κ, θ, v, Δ }, κ being μkTheta expresses the degree of confidence in the mean, α is ∑kV expresses the degree of confidence in the mean.
The system state space equation is:
xtm＝Axt-1,m+Butm(ztm)+wtm
the observation equation is:
ytm＝Cxtm+vtm
the maneuvering state generating process comprises the following steps:
β～GEM(γ)
πkm～DP(α,β)
Figure BDA0002480726880000041
where A is the state transition matrix of the target, B is the control matrix of the target, C is the observation matrix of the target, xt-1,mRepresenting the motion state estimation of the mth target time t-1; y istFor the set of measurements at time t,
Figure BDA0002480726880000051
n is shared at time ttAnd (6) measuring. Each measurement is a two-dimensional vector representing a location in two-dimensional space. Process noise w of the system state space equationtmMeasurement noise v obeying N (0, Q) distribution, observation equationtmFollowing the N (0, R) distribution, Q is the covariance matrix of the process noise and R is the covariance matrix of the observed noise. Definition of ztmThe motion pattern of the mth object, which is an HDP model with concentration parameters of α and gamma and a basic measure of β, defines the system variable πkmFor each movement pattern z of the systemtmβ is generated by a roll-breaking process with a convergence parameter of gamma due to the motion pattern ztmAre infinite and related, therefore
Figure BDA0002480726880000052
I.e. the current moment movement pattern ztmTransition probability density dependent on last time t-1
Figure BDA0002480726880000053
Knowing the current motion pattern ztmThen unknown control input utmFrom z of the sampletmA gaussian distribution. Finally, the HDP-HMM model is combined with the linear system model by utmTo excite the moving state of the maneuvering target. In the motorized object system model, utmIs not observable by y being observabletmAnd realizing the estimation of the motion state of the target.
S2, establishing a target association hypothesis set omega at the time tt。
Target association hypothesis set omegatThe target association hypothesis set at the time t is obtained by associating the target association hypothesis set at the last time with the measurement set at the time t, and the initial target association hypothesis set at the time t can be generated in a random distribution mode.
Target association hypothesis set omegatThe motion state of each target in the set of motion states is independently formed, and the motion state corresponding to each target is estimated by adopting a particle filtering method. Each motion state comprises a global target index, a target life cycle and J particles, the calculation amount and the estimation accuracy of the method are determined by the value of J, and each particle comprises:
{xtm,Ptm,ztm,utm,Ltm,{nljtm},αm,βm,γtm,{mljtm},S1tm,S2tm}
xtmis the motion state estimate for the mth target instant t. PtmIs the covariance matrix of the mth target instant t. z is a radical oftmIs the motion pattern at the mth target instant t. u. oftmIs an unobservable control input for the mth target time t. L istmThe total number of motion patterns traversed for the mth target by time t. n isljtmRepresents the number of transitions from mode l to mode j for the mth target by time t, { n }ljtmα is the matrix of the number of times that the mode has been switched for all the m-th targettmThe concentration parameter representing the target second level of the mth time t, βtmGlobal weight parameter vector, γ, representing the mth target first level at time ttmA concentration parameter representing the mth target first level at time t. m isljtmIs the number of tables at time tth target at restaurant l corresponding to jth dish in HDP-based restaurant model, { mljtmIs the number of tables for all restaurants in the target HDP-based restaurant model at time tth. Defining the mth target control input u at time ttmSufficient statistics of distribution S1tmAnd S2tm。
S3, according to the last time t-1, the target association hypothesis set omegat-1And a motion state set corresponding to each target, traversing the target association hypothesis set
Figure BDA0002480726880000061
Each of which correlates the hypothesis and the target, and calculates various possible motion patterns of the motion state of each target at the current time t using a kalman filter.
Since the target may be in motion mode discrete variable z during the prediction phasetmTo each possible mode of the control input u, thereby causing the control input u to be inputtmObey different continuous distributions, so the prediction estimation of various possible motion modes of the motion state of each target at the current moment t needs to be calculated based on a Kalman filter and weighted summation is carried out. When suppose ztmK is all possible motion patterns of the target, xt|t-1,mkAnd Pt|t-1,mkMotion state prediction estimates and variances, respectively, for the mth target at time t and assuming its motion pattern is k, are calculated as follows:
xt|t-1,mk＝Axt-1,m+But(ztm)
Pt|t-1,mk＝Q+APt-1,mAT
wherein xt-1,mIs the motion state estimation of the mth target at time t-1, Pt-1,mIs the covariance matrix of the motion state estimate of the mth target at time t-1. z is a radical oftmPredicted probability density p (z)tm|z1:t-1,m,βtm,αtm) The expression is as follows:
Figure BDA0002480726880000071
wherein z is1:t-1,mThe cumulative movement pattern for the mth object from the start time to the time t-1,
Figure BDA0002480726880000072
all moving modes from the starting time to the time t-1Formula ztThe accumulated count of (a) is counted,
Figure BDA0002480726880000073
in motion mode z from the start time to the time t-1tCumulative count of k, βktmIs βtmThe k-th element of (a) is,
Figure BDA0002480726880000074
is βtmL of (1)tm+1 element, is a dirac function,kdenotes ztmThe dirac function value at k,
Figure BDA0002480726880000075
denotes ztm＝LtmDirac function value at + 1. Based on this, the predicted estimate of the motion state of the mth target current time t and the covariance x can be obtainedt|t-1,mAnd Pt|t-1,mThe weighting coefficient is taken as the prior estimated probability p (z) of each modetm|z1:t-1,m,βtm,αtm). The tracking detection threshold size may be set based on a predictive estimate of the target. The measurements falling within the tracking detection threshold are correlated with the targets to generate a new set of targets and used to calculate the negative logarithm of likelihood probability values.
S4, constructing a measurement-target distribution matrix according to the possible motion mode of the motion state of each target obtained in S3 and the measurement set of the tracking gate at the current moment, calculating and keeping the previous K target association hypotheses with the highest confidence, and updating the target association hypothesis set.
Inputting a current time measurement set, wherein each measurement has three possibilities: first, the measurement is a continuation of a target in the current hypothesis; second, the measurement is of a new target; third, the measurement is a false alarm. And generating and calculating a measurement-target distribution matrix based on the three possible motion state sets contained in the measurement set and any current target, wherein the matrix element is a negative logarithm value of the measurement likelihood value estimated based on target prediction. The tracks, new tracks and false alarms in the assigned matrix are represented as columns and the measurements are represented as rows. And aiming at the measurement-target distribution matrix, updating possible association hypotheses between targets and observation in the target association hypothesis set at the current moment by adopting a Murty algorithm to carry out a K-Best hypothesis extraction method.
S5, updating the motion state set corresponding to each target by adopting a Kalman filter method based on the current time target association hypothesis set and the motion state set corresponding to each target, wherein the calculation process is as follows:
a) resampling each particle of the mth target, the weight of the ith particle of the mth target
Figure BDA0002480726880000076
Comprises the following steps:
Figure BDA0002480726880000081
Figure BDA0002480726880000082
Figure BDA0002480726880000083
Figure BDA0002480726880000084
wherein the superscript (i) denotes the ith particle. y istjFor the jth measurement at time t,
Figure BDA0002480726880000085
is the total number of motion patterns traversed by the mth object in the ith particle until time t-1,
Figure BDA0002480726880000086
is the cumulative motion pattern of the mth object in the ith particle from the start time to time t-1,
Figure BDA0002480726880000087
is the motion pattern of the mth target instant t in the ith particle,
Figure BDA0002480726880000088
is the control input for the mth target in the ith particle from the start time to the mth target at time t-1,
Figure BDA0002480726880000089
is the first level global weight parameter vector for the mth target in the ith particle at time t-1,
Figure BDA00024807268800000810
is the second level concentration parameter for the mth target in the ith particle at time t-1.
b) Regenerating J particles of the m-th target estimate, the sampling weight of each new particle
Figure BDA00024807268800000811
Comprises the following steps:
Figure BDA00024807268800000812
c) generating a new motion pattern for the ith particle of the mth target
Figure BDA00024807268800000813
Figure BDA00024807268800000814
Wherein,
Figure BDA00024807268800000815
to represent
Figure BDA00024807268800000816
The dirac function value of time.
d) According to new movement pattern
Figure BDA00024807268800000817
Updating the number of movement patterns by value
Figure BDA00024807268800000818
And number of mode transitions
Figure BDA00024807268800000819
Figure BDA0002480726880000091
Figure BDA0002480726880000092
Wherein
Figure BDA0002480726880000093
Is the total number of motion patterns traversed by the mth object in the ith particle until time t,
Figure BDA0002480726880000094
is the m-th target in the ith particle until the time t is in the slave mode
Figure BDA0002480726880000095
Transition to mode
Figure BDA0002480726880000096
Number of times of
Figure BDA0002480726880000097
Is the m-th target in the ith particle until the time t-1
Figure BDA0002480726880000098
Transition to mode
Figure BDA0002480726880000099
The number of times.
e) Computing
Figure BDA00024807268800000910
Updating
Figure BDA00024807268800000911
And
Figure BDA00024807268800000912
when the motion pattern of the mth target is
Figure BDA00024807268800000913
Then, one can obtain:
Figure BDA00024807268800000914
wherein,
Figure BDA00024807268800000915
is the control input for the mth target instant t in the ith particle,
Figure BDA00024807268800000916
and
Figure BDA00024807268800000917
is the control input for the mth target time t in the ith particle
Figure BDA00024807268800000918
Sufficient statistics of the distribution, superscript-1 represents the inverse of the matrix,
Figure BDA00024807268800000919
and
Figure BDA00024807268800000920
is to calculate the auxiliary variable, the solution formula is:
Figure BDA00024807268800000921
Figure BDA00024807268800000922
Figure BDA00024807268800000923
Figure BDA00024807268800000924
Figure BDA00024807268800000925
Figure BDA00024807268800000926
Figure BDA00024807268800000927
Figure BDA0002480726880000101
Figure BDA0002480726880000102
Figure BDA0002480726880000103
wherein,
Figure BDA0002480726880000104
is xt-1,mIs given by the average value of (a), the superscript T denotes the transpose of the matrix, ∑tAnd KtAre the calculation aid variables. Control input
Figure BDA0002480726880000105
Subject to a gaussian inverse vickers distribution,
Figure BDA0002480726880000106
is a hyper-parameter of the inverse gaussian vickers distribution,
Figure BDA0002480726880000107
is the corresponding auxiliary variable. z is a radical ofsIs the order of eyesThe movement pattern, u, being marked at the time ssIs the control input of the target at time s, { us|zsK, s ≠ t } represents all control input sets corresponding to the target motion mode k before the arrival time t, and | | represents the number of the set elements.
f) Calculating the estimation of the ith particle at the mth target t moment according to a Kalman filtering formula
Figure BDA0002480726880000108
Sum covariance
Figure BDA0002480726880000109
g) Sampling
Figure BDA00024807268800001010
The table number of the jth dish corresponding to the jth dish at the restaurant l in the HDP-based restaurant model at the mth target time t in the ith particle is shown, and a restaurant model-based sampling process is given:
for each k 1, …, klj，
Figure BDA00024807268800001011
Sampling auxiliary variable
Figure BDA00024807268800001012
When η is equal to 1, then
Figure BDA00024807268800001013
Ber is the Bernoulli distribution,
Figure BDA00024807268800001014
is that
Figure BDA00024807268800001015
The first parameter of (1).
h) Sampling auxiliary variable
Figure BDA00024807268800001016
A first level of concentration parameter representing the mth target time t-1 in the ith particleSampling the concentration parameter of the first level at the mth target time t in the ith particle
Figure BDA00024807268800001017
Figure BDA00024807268800001018
Figure BDA0002480726880000111
Wherein,
Figure BDA0002480726880000112
is the number of tables in all restaurants in the HDP-based restaurant model at the mth target time instant t in the ith particle, sigma is an auxiliary variable,
Figure BDA0002480726880000113
compliance parameter is αγAnd bγGamma distribution of (2).
i) Sampling auxiliary variable
Figure BDA0002480726880000114
And
Figure BDA0002480726880000115
the sample is the second level concentration parameter of the mth target in the ith particle at time t
Figure BDA0002480726880000116
Figure BDA0002480726880000117
Wherein,
Figure BDA0002480726880000118
is the number of transitions from mode i to any state for the mth target by time t,
Figure BDA0002480726880000119
compliance parameter is aαAnd bαGamma distribution of (2).
j) Sampling a first level global weight parameter vector of an mth target in an ith particle at a time t
Figure BDA00024807268800001110
Figure BDA00024807268800001111
Where Dir is the dirichlet distribution.
Figure BDA00024807268800001112
Is the number of tables in the ith particle that correspond to the 1 st dish in all restaurants in the HDP-based restaurant model for the mth target,
Figure BDA00024807268800001113
is that the mth target in the ith particle corresponds to the mth in all restaurants in the HDP-based restaurant model
Figure BDA00024807268800001114
The number of tables for serving vegetables.
S6, clipping the target association hypothesis set by adopting an N-Scan hypothesis tree pruning method, outputting the target association hypothesis set at the current time and the previous t-1 time and the motion state sets corresponding to the targets, obtaining the motion state estimation values of the targets, and returning to the step S2 at the next time until the monitoring process is finished.
Since the number of tracking hypotheses grows exponentially as the measurement quantities accumulate over time, the hypothesis tree can be pruned using an N-Scan method to control the hypothesis tree depth. When the depth of the hypothesis tree is larger than N, the N-Scan method searches the leaf nodes of the hypothesis tree with the highest current confidence coefficient, and the confidence coefficient is calculated by a measurement-target distribution matrix contained in each target. And reserving the root node branch where the leaf node with the highest confidence coefficient is positioned, and deleting the rest branches. And finally, outputting the association hypothesis set of the current moment and the previous t-1 moment and the motion state set corresponding to each target so as to obtain the estimation value of each state variable of the target.
The method can learn the unknown time-varying state change of the system, estimate the probability of each motion model in the tracking process, and realize target track initiation, data association and track maintenance through effective multi-hypothesis generation and hypothesis management technology. Due to the non-linear characteristics of the dynamic system and the uncertainty of the model parameters, an analytic solution cannot be obtained, and the maneuvering target tracking and the joint estimation of the motion mode of the dynamic system under the clutter environment are realized by using a particle filtering method. The proposed algorithm is very flexible, so that the computational complexity meets the actual engineering requirements. Therefore, the method can be widely applied to a real scene target tracking and monitoring system, and provides important technical support for good information fusion.
The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention. The embodiments and features of the embodiments of the present application may be combined with each other arbitrarily without conflict.

Claims (6)

1. A maneuvering target tracking method facing a monitoring system is characterized by comprising the following steps:
s1, constructing a maneuvering target motion model, and estimating the motion state of the target;
s2, establishing a target association hypothesis set omega at the time ttTarget association hypothesis set ΩtA set comprising possible association combinations between the motion states and observations of the targets and the motion states of each target;
s3, according to the last time t-1, the target association hypothesis set omegat-1And a motion state set corresponding to each target, traversing the target association hypothesis set
Figure FDA0002480726870000011
In (1)Each target calculates various possible motion modes of the motion state of each target at the current moment t by using a Kalman filter;
s4, constructing a measurement-target distribution matrix according to the possible motion mode of the motion state of each target obtained in S3 and the measurement set of which the current time falls into the tracking gate, calculating and keeping the previous K target association hypotheses with the highest confidence, and updating the possible association hypotheses between the targets and the observation in the target association hypothesis set;
and S5, updating the motion state set corresponding to each target by adopting a Kalman filter method based on the current time target association assumption set and the motion state set corresponding to each target.
2. The maneuvering target tracking method facing the monitoring system according to claim 1, characterized in that the maneuvering target motion model is composed of three parts of a system state space equation, an observation equation and a maneuvering state generation process;
the system state space equation represents the change situation of the motion state of the maneuvering target:
xtm＝Axt-1,m+Butm(ztm)+wtm；
the observation equation characterizes the observable maneuvering target motion state:
ytm＝Cxtm+vtm；
the maneuver state generation process characterizes a maneuver state of the maneuver object:
β～GEM(γ)
πkm～DP(α,β)
Figure FDA0002480726870000012
wherein: x is the number oftmRepresenting the motion state estimation of the mth target moment t; x is the number oft-1,mRepresenting the motion state estimation of the mth target time t-1; z is a radical oftmIs the motion pattern at the mth target time t; u. oftmIs the unobservable control input of the mth target at the moment t, obeys the Gaussian distribution utm～N(μk,∑k)，μkIs utmMean value of ∑kIs utmThe covariance matrix of (a); thetak＝{μk,∑kIs the distribution parameter of the kth motor motion pattern, here assumed to be θkThe conjugate prior distribution of (d) is a Gaussian inverse Westst distribution NIW { κ, θ, v, Δ }, κ being μkTheta expresses the degree of confidence in the mean, and delta is ∑kV expresses the degree of confidence in the mean; a is the state transition matrix of the target, B is the control matrix of the target, C is the observation matrix of the target, ytIs the measurement set at time t;
process noise w of the system state space equationtmMeasurement noise v obeying N (0, Q) distribution, observation equationtmObeying N (0, R) distribution, Q being the covariance matrix of the process noise, R being the covariance matrix of the observed noise, HDP model with concentration parameters of α and gamma and basic measure of β,. pikmFor each movement pattern z of the systemtmβ is generated by a folding stick process with a concentration parameter of gamma, and the motion mode ztmAre infinite and related, therefore
Figure FDA0002480726870000021
I.e. the current moment movement pattern ztmTransition probability density dependent on last time t-1
Figure FDA0002480726870000022
Knowing the current motion pattern ztmThen unknown control input utmFrom z of the sampletmA Gaussian distribution; finally, pass utmTo excite the motion state of the maneuvering target through observable ytmAnd realizing the estimation of the motion state of the target.
3. The method of claim 1, wherein the motion state of each target comprises a global target index, a target life cycle, and J particles, each particle comprising:
{xtm,Ptm,ztm,utm,Ltm,{nljtm},αm,βm,γtm,{mljtm},S1tm,S2tm}
xtmis the motion state estimation of the mth target instant t, PtmIs the covariance matrix of the mth target time t, ztmIs the motion pattern of the mth target instant t, utmIs an unobservable control input, L, for the mth target time ttmThe total number of motion patterns traversed for the mth object by the time t, nljtmRepresents the number of transitions from mode l to mode j for the mth target by time t, { n }ljtmIs the matrix of the number of mode transitions that occurred for all the mth target, αtmThe concentration parameter representing the target second level of the mth time t, βtmGlobal weight parameter vector, γ, representing the mth target first level at time ttmA concentration parameter, m, representing the mth target first level at time tljtmIs the number of tables at time tth target at restaurant l corresponding to jth dish in HDP-based restaurant model, { mljtm"is the number of tables for all restaurants in the HDP-based restaurant model for the mth target at time t, defining a control input u for the mth target at time ttmSufficient statistics of distribution S1tmAnd S2tm。
4. The method for tracking the maneuvering target facing the monitoring system according to claim 3, characterized by the step S5 comprising the steps of:
a) resampling each particle of the mth target, the weight of the ith particle of the mth target
Figure FDA0002480726870000031
Comprises the following steps:
Figure FDA0002480726870000032
Figure FDA0002480726870000033
Figure FDA0002480726870000034
wherein the superscript (i) denotes the ith particle, ytjFor the jth measurement at time t,
Figure FDA0002480726870000035
is the total number of motion patterns traversed by the mth object in the ith particle until time t-1,
Figure FDA0002480726870000036
is the cumulative motion pattern of the mth object in the ith particle from the start time to time t-1,
Figure FDA0002480726870000037
is the motion pattern of the mth target instant t in the ith particle,
Figure FDA0002480726870000038
is the control input for the mth target in the ith particle from the start time to the mth target at time t-1,
Figure FDA0002480726870000039
is the first level global weight parameter vector for the mth target in the ith particle at time t-1,
Figure FDA0002480726870000041
is the second level of the concentration parameter of the mth target in the ith particle at time t-1;
b) regenerating J particles of the m-th target estimate, the sampling weight of each new particle
Figure FDA0002480726870000042
Comprises the following steps:
Figure FDA0002480726870000043
c) generating a new motion pattern for the ith particle of the mth target
Figure FDA0002480726870000044
Figure FDA0002480726870000045
Wherein,
Figure FDA0002480726870000046
to represent
Figure FDA0002480726870000047
A function value of the time dirac;
d) according to new movement pattern
Figure FDA0002480726870000048
Updating the number of movement patterns by value
Figure FDA0002480726870000049
And number of mode transitions
Figure FDA00024807268700000410
Figure FDA00024807268700000411
Figure FDA00024807268700000412
Wherein
Figure FDA00024807268700000413
Traversed by the mth target in the ith particle by the time tThe total number of motion patterns to be used,
Figure FDA00024807268700000414
is the m-th target in the ith particle until the time t is in the slave mode
Figure FDA00024807268700000415
Transition to mode
Figure FDA00024807268700000416
The number of times of the operation of the motor,
Figure FDA00024807268700000417
is the m-th target in the ith particle until the time t-1
Figure FDA00024807268700000418
Transition to mode
Figure FDA00024807268700000419
The number of times of (c);
e) computing
Figure FDA00024807268700000420
Updating
Figure FDA00024807268700000421
And
Figure FDA00024807268700000422
when the motion pattern of the mth target is
Figure FDA00024807268700000423
Then, one can obtain:
Figure FDA00024807268700000424
wherein,
Figure FDA00024807268700000425
is the control input for the mth target instant t in the ith particle,
Figure FDA00024807268700000426
and
Figure FDA00024807268700000427
is the control input for the mth target time t in the ith particle
Figure FDA00024807268700000428
Sufficient statistics of the distribution, superscript-1 represents the inverse of the matrix,
Figure FDA0002480726870000051
and
Figure FDA0002480726870000052
is to calculate the auxiliary variable, the solution formula is:
Figure FDA0002480726870000053
Figure FDA0002480726870000054
Figure FDA0002480726870000055
Figure FDA0002480726870000056
Figure FDA0002480726870000057
Figure FDA0002480726870000058
Figure FDA0002480726870000059
Figure FDA00024807268700000510
Figure FDA00024807268700000511
Figure FDA00024807268700000512
wherein,
Figure FDA00024807268700000513
is xt-1,mIs given by the average value of (a), the superscript T denotes the transpose of the matrix, ∑tAnd KtIs to calculate auxiliary variables, control inputs
Figure FDA00024807268700000514
Subject to the inverse Gaussian Weight distribution, { kappa, theta, nu, delta } is a hyperparameter of the inverse Gaussian Weight distribution,
Figure FDA00024807268700000515
is the corresponding auxiliary variable, zsIs the motion pattern of the object at time s, usIs the control input of the target at time s, { us|zsK, s ≠ t } represents all control input sets corresponding to the target motion mode k before the arrival time t, and | | represents the number of the elements of the set;
f) calculating the estimation of the ith particle at the mth target t moment according to a Kalman filtering formula
Figure FDA00024807268700000516
Sum covariance
Figure FDA00024807268700000517
g) Sampling
Figure FDA00024807268700000518
Figure FDA00024807268700000519
The table number of the jth dish corresponding to the jth dish at the restaurant l in the HDP-based restaurant model at the mth target time t in the ith particle is shown, and a restaurant model-based sampling process is given:
for each k 1, …, klj，
Figure FDA0002480726870000061
Sampling auxiliary variable
Figure FDA0002480726870000062
When η is equal to 1, then
Figure FDA0002480726870000063
Ber is the Bernoulli distribution,
Figure FDA0002480726870000064
is that
Figure FDA0002480726870000065
The first parameter of (1);
h) sampling auxiliary variable
Figure FDA0002480726870000066
Figure FDA0002480726870000067
Representing the first level of the concentration parameter at the mth target time t-1 in the ith particle, and sampling the first level of the concentration parameter at the mth target time t in the ith particle
Figure FDA0002480726870000068
Figure FDA0002480726870000069
Figure FDA00024807268700000610
Wherein,
Figure FDA00024807268700000611
is the number of tables in all restaurants in the HDP-based restaurant model at the mth target time instant t in the ith particle, sigma is an auxiliary variable,
Figure FDA00024807268700000612
compliance parameter is αγAnd bγGamma distribution of (2);
i) sampling auxiliary variable
Figure FDA00024807268700000613
And
Figure FDA00024807268700000614
the sample is the second level concentration parameter of the mth target in the ith particle at time t
Figure FDA00024807268700000615
Figure FDA00024807268700000616
Wherein,
Figure FDA00024807268700000617
is the number of transitions from mode i to any state for the mth target by time t,
Figure FDA00024807268700000618
compliance parameter is aαAnd bαGamma distribution of (2);
j) sampling a first level global weight parameter vector of an mth target in an ith particle at a time t
Figure FDA00024807268700000619
Figure FDA00024807268700000620
Where Dir is the dirichlet distribution,
Figure FDA00024807268700000621
is the number of tables in the ith particle that correspond to the 1 st dish in all restaurants in the HDP-based restaurant model for the mth target,
Figure FDA00024807268700000622
is that the mth target in the ith particle corresponds to the mth in all restaurants in the HDP-based restaurant model
Figure FDA0002480726870000071
The number of tables for serving vegetables.
5. The method for tracking the maneuvering target facing the monitoring system according to claim 1, characterized in that in step S4, a Murty algorithm is adopted to perform a K-Best hypothesis extraction method to update possible association hypotheses between the target and the observation in the target association hypothesis set at the current time.
6. The method for tracking the maneuvering target facing the monitoring system according to claim 1, characterized by further comprising a step S6 of clipping the target association hypothesis set by using an N-Scan hypothesis tree pruning method, outputting the target association hypothesis set and the corresponding motion state set by the current time and the previous t-1 time, and obtaining the motion state estimation value.
CN202010377925.3A 2020-05-07 2020-05-07 Maneuvering target tracking method facing monitoring system Active CN111582159B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010377925.3A CN111582159B (en) 2020-05-07 2020-05-07 Maneuvering target tracking method facing monitoring system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010377925.3A CN111582159B (en) 2020-05-07 2020-05-07 Maneuvering target tracking method facing monitoring system

Publications (2)

Publication Number Publication Date
CN111582159A true CN111582159A (en) 2020-08-25
CN111582159B CN111582159B (en) 2022-11-04

Family

ID=72112028

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010377925.3A Active CN111582159B (en) 2020-05-07 2020-05-07 Maneuvering target tracking method facing monitoring system

Country Status (1)

Country Link
CN (1) CN111582159B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113793533A (en) * 2021-08-30 2021-12-14 武汉理工大学 Collision early warning method and device based on vehicle front obstacle recognition

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105699964A (en) * 2016-02-29 2016-06-22 无锡南理工科技发展有限公司 Road multi-target tracking method based on automobile anti-collision radar
CN109633589A (en) * 2019-01-08 2019-04-16 沈阳理工大学 The Multi-target Data Associations assumed are optimized based on multi-model more in target following
CN110120066A (en) * 2019-04-11 2019-08-13 上海交通大学 Robust multiple targets tracking and tracking system towards monitoring system

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105699964A (en) * 2016-02-29 2016-06-22 无锡南理工科技发展有限公司 Road multi-target tracking method based on automobile anti-collision radar
CN109633589A (en) * 2019-01-08 2019-04-16 沈阳理工大学 The Multi-target Data Associations assumed are optimized based on multi-model more in target following
CN110120066A (en) * 2019-04-11 2019-08-13 上海交通大学 Robust multiple targets tracking and tracking system towards monitoring system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
胡子军等: "基于高斯混合带势概率假设密度滤波器的未知杂波下多机动目标跟踪算法", 《电子与信息学报》 *
韩玉兰等: "非线性系统的多扩展目标跟踪算法", 《计算机应用》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113793533A (en) * 2021-08-30 2021-12-14 武汉理工大学 Collision early warning method and device based on vehicle front obstacle recognition

Also Published As

Publication number Publication date
CN111582159B (en) 2022-11-04

Similar Documents

Publication Publication Date Title
CN109508445B (en) Target tracking method with color measurement noise and variational Bayesian self-adaptive Kalman filtering
Cox A review of statistical data association techniques for motion correspondence
Bar-Shalom et al. Multisensor track-to-track association for tracks with dependent errors
Hue et al. Tracking multiple objects with particle filtering
CN110849369B (en) Robot tracking method, device, equipment and computer readable storage medium
CN107677997B (en) Extended target tracking method based on GLMB filtering and Gibbs sampling
CN114638855A (en) Multi-target tracking method, equipment and medium
CN111693984A (en) Improved EKF-UKF moving target tracking method
JP2009501334A (en) Adaptation of model sets by stochastic mass diffusion
CN111582159B (en) Maneuvering target tracking method facing monitoring system
CN116628448A (en) Sensor management method based on deep reinforcement learning in extended target
Malleswaran et al. IMM-UKF-TFS model-based approach for intelligent navigation
CN116734860A (en) Multi-AUV self-adaptive cooperative positioning method and system based on factor graph
CN116520281A (en) DDPG-based extended target tracking optimization method and device
CN113709662A (en) Ultra-wideband-based autonomous three-dimensional inversion positioning method
CN114339595B (en) Ultra-wide band dynamic inversion positioning method based on multi-model prediction
CN113537302B (en) Multi-sensor randomized data association fusion method
CN103888100A (en) Method for filtering non-Gaussian linear stochastic system based on negentropy
CN115578425A (en) Dynamic tracking method applied to fry counter and based on unscented Kalman filtering
CN104467742A (en) Sensor network distribution type consistency particle filter based on Gaussian mixture model
Hadi et al. Behavior formula extraction for object trajectory using curve fitting method
CN114061592A (en) Adaptive robust AUV navigation method based on multiple models
Turner et al. Receding Horizon Tracking of an Unknown Number of Mobile Targets using a Bearings-Only Sensor
Mora et al. Multirate obstacle tracking and path planning for intelligent vehicles
Weng Tracking multiple objects using visual-based sensors

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant