CN112098993A - Multi-target tracking data association method and system - Google Patents
Multi-target tracking data association method and system Download PDFInfo
- Publication number
- CN112098993A CN112098993A CN202010971580.4A CN202010971580A CN112098993A CN 112098993 A CN112098993 A CN 112098993A CN 202010971580 A CN202010971580 A CN 202010971580A CN 112098993 A CN112098993 A CN 112098993A
- Authority
- CN
- China
- Prior art keywords
- target
- state
- value
- association
- candidate
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/66—Radar-tracking systems; Analogous systems
- G01S13/72—Radar-tracking systems; Analogous systems for two-dimensional tracking, e.g. combination of angle and range tracking, track-while-scan radar
- G01S13/723—Radar-tracking systems; Analogous systems for two-dimensional tracking, e.g. combination of angle and range tracking, track-while-scan radar by using numerical data
- G01S13/726—Multiple target tracking
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/02—Systems using reflection of radio waves, e.g. primary radar systems; Analogous systems
- G01S13/50—Systems of measurement based on relative movement of target
- G01S13/58—Velocity or trajectory determination systems; Sense-of-movement determination systems
- G01S13/60—Velocity or trajectory determination systems; Sense-of-movement determination systems wherein the transmitter and receiver are mounted on the moving object, e.g. for determining ground speed, drift angle, ground track
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Abstract
The invention discloses a multi-target tracking data association method and a system, wherein the method takes the track of a known initial time period of a target measuring point as a reinforcement learning training process according to the association characteristic of multi-target tracking data, generates random clutter around the known measuring point in one step, and takes the clutter and the known measuring point as radar acquisition measuring points; screening candidate measuring points from the measuring points according to the tracking gate, performing data association on all the candidate measuring points according to the matching degree and the position distribution rule by utilizing motion matching and reinforcement learning according to the target motion characteristics, checking the association result according to the known measuring points in one step, and training an experience matrix of a reinforcement learning data association model; and performing data association on track points of the targets entering the clutter area by combining motion matching, and continuously optimizing the experience matrix according to an association result until track association is completed. The problems of low correct association rate, high calculation complexity and the like are solved, the correct association rate is improved, and the calculation complexity is reduced.
Description
Technical Field
The invention relates to the technical field of multi-target tracking, in particular to a multi-target tracking data association method and system. The method is suitable for multi-target tracking data association in a multi-clutter environment.
Background
The basic concept of multi-target tracking was first proposed by Wax in 1955. In 1964, Sutler made intensive research on multi-target tracking theory and data association problem and made pioneering progress, however, the maneuvering target tracking theory did not really draw attention until the early 70 s. During the period, the multi-target tracking technology which is created by Bar-staff and Singer and organically combines the data correlation technology and the Kalman (kalman) filtering technology into a mark has made a breakthrough. However, target tracking data association in a dense clutter environment is always a difficult problem in the field of multi-target tracking, signals captured by a radar also include spurious measurement caused by clutter besides real measurement, and accurate association of targets is difficult to achieve.
In the research of multi-target tracking data association in a multi-clutter environment, the existing nearest neighbor data association method (NN) is the simplest method for solving data association, but the correct association rate of the nearest neighbor method in the clutter environment is lower; the joint probability data association method (JPDA) calculates association probability by combining all targets and measurements into a joint event according to hypothesis, and can well solve the multi-target measurement association problem in a clutter environment.
Disclosure of Invention
The invention provides a multi-target tracking data association method and a multi-target tracking data association system, which are used for overcoming the defects of low correct association rate, high calculation complexity, combined explosion in the calculation of association probability and the like in the prior art, and realizing the purposes of improving the correct association rate and reducing the calculation complexity.
In order to achieve the above object, the present invention provides a multi-target tracking data association method, which comprises the following steps:
s1, constructing a reinforcement learning data association model for predicting the target position at the current moment by combining the previous moment state and the motion attribute of the target;
s2, simulating random clutter points around the known measurement points at the current target moment, and obtaining intra-gate candidate measurement points and intra-gate candidate measurement point position distribution according to the set wave gate;
s3, selecting a weight in an experience matrix of an association model according to the distribution of the candidate measuring points, and obtaining the association probability of each candidate measuring point according to the fluctuation influence of the weight on the state matching degree of the candidate measuring points and the target and the motion matching degree of the candidate measuring points and the target;
s4, obtaining a one-step estimation value of the actual state of the target at the current moment according to the one-step known measurement point of the target at the current moment and carrying out point track-track association;
s5, obtaining a simulation state one-step estimation value of the target at the current moment according to the association probability and the candidate measuring points, and training an experience matrix by taking the Euclidean distance between the simulation state one-step estimation value and the actual state one-step estimation value as loss;
s6, repeating the steps S2-S5 until all known measurement points in the initial time period are associated and trained to obtain a training model;
s7, taking data points collected by the radar after the target enters the clutter area as measuring points, obtaining candidate measuring points to be measured in the door and distribution according to a set wave gate, combining the training model and motion matching to obtain a state one-step estimation value of the target, and associating the state one-step estimation value with the state one-step estimation value; and calculating a one-step state predicted value of the next moment of the target, optimizing an experience matrix of the training model by taking the Mahalanobis distance between the one-step observation predicted value of the one-step state predicted value of the target and the one-step observation predicted value of the one-step state estimated value of the target as loss, and repeating association and optimization until track association is completed.
In order to achieve the above object, the present invention further provides a multi-target tracking data association system, which includes a processor and a memory, wherein the memory stores a multi-target tracking data association program, and the processor executes the steps of the method when running the multi-target tracking data association program.
According to the multi-target tracking data association method and system provided by the invention, a multi-target tracking data association algorithm based on a reinforcement learning model and motion matching is provided, according to the multi-target tracking data association characteristic, the track association of the initial time period known by target measurement is regarded as a reinforcement learning training process, the track association of the subsequent targets entering a clutter area is regarded as a reinforcement learning association process, and the association event of each target and each measurement is not required to be established, so that the algorithm can reduce the calculation complexity in a clutter dense environment, keep the calculation speed faster and avoid the problem of combined explosion; the method disclosed by the invention utilizes a mode of combining reinforcement learning and motion matching to calculate the association probability of the target and the measurement, and the motion and state characteristics of the target and the distribution rule of the indoor measurement are considered during calculation, so that the association accuracy of the multi-target tracking data is effectively improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the structures shown in the drawings without creative efforts.
Fig. 1 is a flowchart of a radar multi-target tracking data association method based on reinforcement learning and motion matching according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of the true tracks and clutter areas of two targets with fewer clutter according to the present invention;
FIG. 3 is a schematic diagram of the comparison simulation of the true track and the estimated track of two targets when clutter is low according to the present invention;
FIG. 4 is a schematic diagram of the true tracks and clutter areas of two targets when clutter is high in the present invention;
FIG. 5 is a schematic diagram of the comparison simulation of the real track and the estimated track of two targets when clutter is high in the present invention;
FIG. 6 is a schematic diagram of the real tracks and clutter areas of two targets when clutter is dense according to the present invention;
FIG. 7 is a schematic diagram of the comparison simulation of the real track and the estimated track of two targets when the clutter is dense.
The implementation, functional features and advantages of the objects of the present invention will be further explained with reference to the accompanying drawings.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that all the directional indicators (such as up, down, left, right, front, and rear … …) in the embodiment of the present invention are only used to explain the relative position relationship between the components, the movement situation, etc. in a specific posture (as shown in the drawing), and if the specific posture is changed, the directional indicator is changed accordingly.
In addition, the descriptions related to "first", "second", etc. in the present invention are only for descriptive purposes and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
In the present invention, unless otherwise expressly stated or limited, the terms "connected," "secured," and the like are to be construed broadly, and for example, "secured" may be a fixed connection, a removable connection, or an integral part; the connection can be mechanical connection, electrical connection, physical connection or wireless communication connection; they may be directly connected or indirectly connected through intervening media, or they may be connected internally or in any other suitable relationship, unless expressly stated otherwise. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
In addition, the technical solutions in the embodiments of the present invention may be combined with each other, but it must be based on the realization of those skilled in the art, and when the technical solutions are contradictory or cannot be realized, such a combination of technical solutions should not be considered to exist, and is not within the protection scope of the present invention.
Example one
As shown in fig. 1-7, an embodiment of the present invention provides a multi-target tracking data association method, where the scheme is not only applicable to data association in a single-target tracking process, but also applicable to data association in two or more target tracking processes, as shown in fig. 1, specifically including:
step S1, constructing a reinforcement learning data association model for predicting the target position at the current moment by combining the last moment state and the motion attribute of the target;
the method comprises the following steps that real measuring point data (not containing clutter measuring data) of a target in a determined time period by a radar and real track point data of the target in the determined time period are known in a training process at the current moment, measuring point data (containing clutter point data and real data) of the target after the determined time period by the radar are known in a correlation process, and correlation characteristics among reinforced learning real data are simulated through constructing a model to identify clutter zone data, namely clutter points (identify real data) and estimate a target track of the clutter zone;
step S2, simulating random noise points around the known measuring points at the current moment of the target and obtaining intra-gate candidate measuring points and intra-gate candidate measuring point position distribution according to the set wave gate;
determining intra-gate candidate measuring points and distribution according to the Mahalanobis distance between the random clutter point and a one-step predicted point of the target t at the previous moment k-1 obtained by a reinforcement learning data association model and a set wave gate; determining the range of the clutter points to guarantee the calculation speed;
step S3, selecting a weight in an experience matrix of the association model according to the distribution of the candidate measuring points, and obtaining the association probability of each candidate measuring point according to the fluctuation influence of the weight on the state matching degree of the candidate measuring points and the target and the motion matching degree of the candidate measuring points and the target;
the position distribution of the in-door candidate measuring points is related to the motion attributes (position, speed direction, size and the like) of the target at the previous moment, the optimal action is selected according to the parameters matched with the motion in the experience matrix to carry out action weight matching, the association probability is calculated, the association probability represents the probability that the target moves to the track point position corresponding to the clutter point at the next moment, and each in-door candidate measuring point is matched with the weight and the association probability is calculated.
Step S4, obtaining a one-step estimation value of the actual state of the target at the current moment according to the one-step known measurement point of the target at the current moment and carrying out point track-track association;
for example: for the one-step known measurement value Z at the k-th timet(k | k) Kalman filtering to obtain one-step estimation value of actual stateAnd one-step estimation value of target actual state at the kth momentPerforming point track-track association; at the k +1 th moment, the measured value Z is known according to one stept(k +1| k +1) Kalman filtering to obtain one-step estimation value of actual stateAnd one-step estimation value of target actual state at the k +1 th momentPerforming point track-track association; continuously repeating the previous process, and performing point track-track association on the obtained one-step estimation value of the training process state;
step S5, obtaining a simulation state one-step estimation value of the target at the current moment according to the association probability and the candidate measuring points, and training an experience matrix by taking the Euclidean distance between the simulation state one-step estimation value and the actual state one-step estimation value as loss;
calculating Kalman gain K of K target t at current momentt(k) Sum-state covariance one-step estimate Pt(k | k) and from this calculate a one-step estimate X of the simulated state of the target t at time kt(k | k); calculating all intra-door candidate measuring points at the current moment by combining the association probability to obtain a simulation state one-step estimation value of the target; substituting the simulation state one-step estimation value of the target at the current moment into an optimization module of the reinforcement learning data association model to obtain an experience matrix;
step S6, training an experience matrix by taking the Euclidean distance between the actual state one-step estimation value obtained in the step S4 and the simulation state one-step estimation value obtained in the step S5 as loss, and repeating the steps S2-S5 until all known measurement points in the initial time period are associated and trained to obtain a training model;
for example: one-step estimation value of target actual state at the kth momentAnd the one-step estimation value X of the simulation statetTraining an experience matrix by taking Euclidean distance between (k | k) as loss to complete the kth round of training; repeating the steps S2-S6 to obtain the k +1 th timeOne-step estimation value X of simulation state of carved targett(k +1| k +1), and the estimated value is compared with the target actual state at the k +1 th moment by one stepThe Euclidean distance between the training data and the training data is used as loss to train the experience matrix of the previous training cycle again to finish the (k +1) th training cycle; the process is circulated until the track point at the last moment of the known starting time period is used as a known measuring point to be associated and trained, and a training model is obtained;
step S7, using radar collected data points as measuring points after the target enters the clutter area, obtaining candidate measuring points to be measured in the door and distribution according to a set wave gate, combining the training model and motion matching to obtain a state one-step estimation value of the target, and associating the state one-step estimation value with the state one-step estimation value; and calculating the Mahalanobis distance between the target one-step state predicted value and the one-step observation predicted value of the one-step state estimated value, optimizing an experience matrix of the training model by taking the Mahalanobis distance as loss, and repeating association and optimization until track association is completed. The motion matching here refers to the degree of motion matching in the correlation probability calculation process.
Re-executing the step S2 to obtain indoor candidate measuring points and distribution by taking the measuring points of the target entering the clutter area as clutter points, executing the steps S3 and S5 to obtain a state one-step estimation value of the target at the current moment, inputting the state one-step estimation value of the target at the current moment into an experience matrix optimization module of a training model to obtain an experience matrix, and performing point trace-track association on the state one-step estimation value of the target at the current moment to obtain a state one-step prediction value of the target at the next moment; and optimizing the experience matrix of the training model by taking the Mahalanobis distance between the one-step observation predicted value of the one-step state predicted value of the target and the one-step observation predicted value of the one-step state estimated value as loss, and circulating the association and optimization processes until track association is completed.
Firstly, determining a multi-target tracking data association initial condition, regarding track association of an initial time period known by target measurement as a reinforcement learning training process according to the multi-target tracking data association characteristic, and regarding track association of a subsequent target entering a clutter area as a reinforcement learning association process; in the training process, random clutter is generated near a known target measurement point, data association is carried out by combining motion matching with reinforcement learning, and an association result is checked according to a known target measurement value so as to train a reinforcement learning experience matrix; and in the association process, performing data association by combining motion matching according to the trained reinforcement learning experience matrix, and continuously optimizing the experience matrix according to the association result until track association is completed. The invention utilizes the mode of combining motion matching with reinforcement learning to carry out data association, and can obtain accurate association results on the basis of ensuring the calculation speed.
Specifically, the method comprises the following steps: after a target enters a clutter area, acquiring a target t measuring point at the current moment; calculating the Mahalanobis distance between each measuring point and a one-step predicting point calculated by the target t through a training data association model and determining intra-gate candidate measuring points and distribution through a set gate; selecting the optimal action for all candidate measuring points of the target t in an experience matrix Q-table of a reinforcement learning data association model at the moment k according to the distribution of the candidate measuring points in the target t gate to carry out action matching, and calculating association probability; one-step state prediction value according to target tAnd one-step state covariance predictionCalculating the Kalman gain Kt(k) Sum-state covariance one-step estimate Pt(k | k) and calculating a state one-step estimate X of the target t based thereont(k | k); calculating a one-step state prediction value of a target tOne-step observation prediction value ofAnd a one-step state estimate XtOne-step observed prediction of (k | k)Will be provided withAndthe mahalanobis distance between them is considered as the cost ft(k) Calculating reinforcement learning reward factor rt(k) And optimizing the Q-table according to the reinforcement learning reward factor until the track association is completed.
According to the correlation characteristics of the multi-target tracking data, the track correlation of the initial time period with known target measurement is regarded as a reinforcement learning training process, namely a training process before a training model is obtained; and (3) regarding the track association of the subsequent targets entering the clutter area as a reinforcement learning association process, namely an optimization process after the training model is obtained. The target measurement here refers to: the radar sensor obtains actual measurement data, the data are obtained after being removed, and the track of the target in a certain time period can be obtained through calculation of the data.
Preferably, the step S1 of constructing the reinforcement learning data association model includes:
step S11: determining initial conditions of multi-target tracking data association;
determining a known target measurement value Z for a starting time periodt(k|k),k=1,...,KtrainAnd a clutter region measurement value Z (k) for determining a state transition matrix F of the target t at the time kt(k) Observation matrix Ht(k) Process noise covariance matrix Qt(k) And the observed noise covariance matrix Rt(k) Calculating the predicted value of the one-step state of the target t at the moment kOne-step observation prediction valueOne-step state covariance predictionSum innovation covariance matrix St(k) (ii) a Of the target t at time kOne-step state predictionPredicting a one-step predicted value of the state (position, speed, acceleration and the like) at the moment k from the moment k-1 for the target t, and observing the predicted value in one stepObtaining a one-step prediction value of the position of the target t at the time k for the radar, and predicting a state covariance matrix in one stepOne-step prediction of covariance between states at time k of target t, St(k) A covariance matrix of innovation at time k of the target t;
the predicted value of the one-step state of the target t at the moment kOne-step observation prediction valueOne-step state covariance predictionSum innovation covariance matrix St(k) Their respective computational expressions are:
wherein, Ft(k) State transition matrix, H, representing target t at time kt(k) An observation matrix, Q, representing the target t at time kt(k) Process noise covariance matrix, R, representing target t at time kt(k) An observed noise covariance matrix, X, representing the target t at time kt(k-1| k-1) is a one-step estimate of the simulated state of target t at time k-1, Pt(k-1| k-1) is a one-step estimate of the state covariance of target t at time k-1.
Step S12: setting a reinforcement learning discount factor lambda and learning efficiency gamma, establishing an experience matrix Q-table of a reinforcement learning model, wherein a state s is measured distribution, an action a is weight selection of the experience matrix, and the Q-table is initialized to a 0 matrix.
Preferably, the step of simulating a random clutter point in S2 comprises:
the training process requires a step of knowing the measurement value Z of the target t at time kt(k|k),k=1,...,KtrainAmbient generation clutter Zflase,i(k):
Zflase,i(k)=Zt(k|k)+l-2l.rand0,1 (1);
Wherein l is the equivalent square side length of the elliptic wave gate, i is 1,20,1Is a random number between 0 and 1, KtrainFor the upper limit of the training process, T is 1, 2.
Preferably, the step of S2 obtaining the intra-gate candidate measurement points and the position distribution thereof includes:
determining candidate measuring points and distribution in the gate according to the Mahalanobis distance between each measuring point and the target one-step predicting point at the previous moment and the wave gate; substituting the measurement into a wave gate detection module to obtain intra-gate candidate measurement and position distribution thereof;
determination of the measurement values Z (k):
for wave gate detection moduleEach measurement Z (k) and the target t one-step predicted measurement value at the time of k calculationMahalanobis distance g oft(k) If the Mahalanobis distance is smaller than the threshold of the wave gate, the measuring point is positioned in the wave gate and reserved as the candidate measurement of the target t, which is recorded asMahalanobis distance gt(k):
If g ist(k) If the following condition is satisfied, the measurement is retained as a candidate measurement of the target t:
gt(k)≤ζ (4);
where ζ is the gate threshold.
Considering the possible event that all candidate measurements in the target t-wave gate are not the target real measurement, the target t is measured in one stepRandomly generating echoes around and adding candidate measurements
The corresponding correlation probabilities are regarded as the probabilities that the intra-gate measurements are all clutter. Measured value predicted by target t one stepEstablishing a two-dimensional rectangular coordinate system for the origin to divide the wave gate into 4And the area divides the wave gate into a central area and an edge area by taking zeta/2 as a limit, so that the wave gate is divided into 8 areas in total, and the distribution condition of the wave gate is calculated according to the position relation between each target candidate measurement and one-step predicted measurement value.
Preferably, the calculating step calculates the association probability between each candidate measuring point of the target t and the targetStep S3 includes:
step S31, selecting a weight in an experience matrix of the correlation model according to the distribution of the candidate measuring points; the method specifically comprises the following steps:
according to the distribution of candidate measurement in each target gate, selecting the best action best _ action in the Q-table corresponding state:
best_action=max[Q(current s,all actions)] (6);
wherein current s is the current state, and each state corresponds to the position distribution measured in the wave gate; all actions are all actions, and each action represents weight selection:
wherein Δ is a scaling factor;
step S32, calculating Euclidean distance between each candidate measuring point and one-step observation predicted value of the target to obtain the state matching degree of each candidate measuring point and the target; the state reflects the matching degree between the candidate measuring points as radar real measuring data and the positions of the target actual track points, specifically:
calculating all candidate measuring values and one-step observation predicted values of target t at the moment kEuclidean distance of
Step S33, obtaining the fluctuation influence of the weight value on the state matching degree by algebraic operation of the selected weight value and the Euclidean distance; here, the multiplication operation is selected, specifically:
step S34, calculating Euclidean distance between each candidate measuring point and the target three-step observation predicted value to obtain the motion matching degree of each candidate measuring point and the target; the observation prediction from the moment k-3 to the moment k is used, the prediction process is obtained according to the motion characteristics of the target, and the distance difference between the measurement and the three-step observation prediction value can reflect whether the measured position accords with the motion characteristics of the target or not or how much the measured position differs from the motion characteristics, specifically:
performing motion matching on all candidate measurement values of the target t at the moment k and finishing weight selection; performing motion matching on all candidate measuring values of the target t at the moment k, and calculating the association probability of each candidate measuring point of the target t
Calculating a target t point trace X at the k-3 momenttThree-step state prediction value of (k-3| k-3)
Calculating all candidate measuring values and three-step views of target t at the moment kMeasure the predicted valueEuclidean distance of
Step S35, obtaining the association probability of each candidate measuring point according to the motion matching degree and the state matching degree after fluctuation; specifically, the method comprises the following steps:
probability of association of all candidate measurement values of target t at time k with target tComprises the following steps:
preferably, the step of obtaining a one-step state estimation value at S4 includes:
calculating Kalman gain K of target t at moment Kt(k) Sum-state covariance one-step estimate Pt(k|k):
Calculating a state one-step estimation value X of the target t at the moment kt(k|k):
Preferably, the step of training the experience matrix in S5 includes:
state one-step estimation value Xt(k | k) and the state covariance one-step estimate Pt(k | k) is used for point track-track correlation and substituted into the reinforcement learning experience matrix optimization module to train the Q-table: for the training process, its state one-step estimate Xt(k | k) is used to substitute into the reinforcement learning empirical matrix optimization module to train the Q-table, but without the point-track correlation, for the one-step known metrology value Zt(k|k),k=1,...,KtrainPerforming Kalman filtering, and performing one-step estimation on the obtained training process statePerforming point track-track association:
Training the Q-table according to the reinforcement learning reward factor:
wherein Qt(si,aj) Measurement at s representing target t at time kiSelect a in the statejQ value corresponding to the action, lambda is learning factor, gamma is discount factor,measure target t at time k at siMaximum Q value in the state.
Preferably, the step of optimizing the experience matrix of the training model in S7 includes:
for the correlation process, calculating a target t one-step state prediction valueOne-step observation prediction value ofAnd target t one-step state estimation value XtOne-step observed prediction of (k | k)
St(k+1)=Ht(k+1)·Pt(k|k)·Ht(k+1)T (19);
Calculating reinforcement learning reward factor rt(k):
Optimizing the Q-table according to the reinforcement learning reward factor:
the parameters in the model of equation (22) are the same as those in equation (14) above.
Therefore, the radar multi-target tracking data association algorithm based on reinforcement learning and motion matching is finished.
The effect of the present invention is further verified and explained by the following simulation experiment.
And (I) simulation experiment data show.
In order to verify the accuracy of the method, the method is proved by a simulation experiment; the experimental data parameters were as follows:
(II) simulation results and analysis
The simulation results of the invention are respectively shown in fig. 2, fig. 3, fig. 4, fig. 5, fig. 6 and fig. 7, fig. 2 and fig. 4 are schematic diagrams of two target real tracks and clutter areas when clutter is less and more, fig. 3 and fig. 5 are schematic diagrams of comparison simulation of two target real tracks and estimated tracks when clutter is less and more, wherein horizontal coordinates and vertical coordinates are X and Y direction positions, and units are m. As can be seen from fig. 2 and 4, the target track is difficult to be accurately correlated and estimated by means of a conventional data correlation algorithm because the measurements of the two targets are crossed and the clutter area is tightly gathered, and as can be seen from fig. 3 and 5, the target measurements can be accurately separated from the clutter by using the method of the present invention, so that high correlation accuracy is ensured.
As can be seen from fig. 6, as the number of clutter in the clutter region further increases, the clutter distribution around the target metrology point trace is very dense. At this time, if a conventional nearest neighbor algorithm is adopted, the estimation error is large; and the situation of combination explosion can occur by adopting a conventional joint probability data association algorithm, so that association fails. The method can efficiently calculate the association probability by combining experience matching with reinforcement learning, and the effectiveness of the processing method is verified by the simulation experiment result of FIG. 7.
In conclusion, the simulation experiment verifies the correctness, the effectiveness and the reliability of the method.
Example two
Based on the first embodiment, the invention provides a multi-target tracking data association system, which comprises a memory and a processor, wherein the memory stores a multi-target tracking data association program, and the processor executes the steps of any embodiment of the method when running the multi-target tracking data association program.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention, and all modifications and equivalents of the present invention, which are made by the contents of the present specification and the accompanying drawings, or directly/indirectly applied to other related technical fields, are included in the scope of the present invention.
Claims (10)
1. A multi-target tracking data association method is characterized by comprising the following steps:
s1, constructing a reinforcement learning data association model for predicting the target position at the current moment by combining the previous moment state and the motion attribute of the target;
s2, simulating random clutter points around the known measurement points at the current target moment, and obtaining intra-gate candidate measurement points and intra-gate candidate measurement point position distribution according to the set wave gate;
s3, selecting a weight in an experience matrix of an association model according to the distribution of the candidate measuring points, and obtaining the association probability of each candidate measuring point according to the fluctuation influence of the weight on the state matching degree of the candidate measuring points and the target and the motion matching degree of the candidate measuring points and the target;
s4, obtaining a one-step estimation value of the actual state of the target at the current moment according to the one-step known measurement point of the target at the current moment and carrying out point track-track association;
s5, obtaining a simulation state one-step estimation value of the target at the current moment according to the association probability and the candidate measuring points, and training an experience matrix by taking the Euclidean distance between the simulation state one-step estimation value and the actual state one-step estimation value as loss;
s6, repeating the steps S2-S5 until all known measurement points in the initial time period are associated and trained to obtain a training model;
s7, taking data points collected by the radar after the target enters the clutter area as measuring points, obtaining candidate measuring points to be measured in the door and distribution according to a set wave gate, combining the training model and motion matching to obtain a state one-step estimation value of the target, and associating the state one-step estimation value with the state one-step estimation value; and calculating a one-step state predicted value of the next moment of the target, optimizing an experience matrix of the training model by taking the Mahalanobis distance between the one-step observation predicted value of the one-step state predicted value of the target and the one-step observation predicted value of the one-step state estimated value of the target as loss, and repeating association and optimization until track association is completed.
2. The multi-target tracking data association method according to claim 1, wherein the step S1 specifically includes:
step S11: determining initial conditions of multi-target tracking data association;
determining a known target measurement value Z for a starting time periodt(k|k),k=1,...,KtrainAnd a clutter region measurement value Z (k) for determining a state transition matrix F of the target t at the time kt(k) Observation matrix Ht(k) Process noise covariance matrix Qt(k) And the observed noise covariance matrix Rt(k) Calculating the predicted value of the one-step state of the target t at the moment kOne-step observation prediction valueOne-step state covariance predictionSum innovation covariance matrix St(k);
The predicted value of the one-step state of the target t at the moment kOne-step observation prediction valueOne-step state covariance predictionSum innovation covariance matrix St(k) The calculation expression of (a) is:
wherein, Ft(k) State transition matrix, H, representing target t at time kt(k) An observation matrix, Q, representing the target t at time kt(k) Process noise covariance matrix, R, representing target t at time kt(k) An observed noise covariance matrix, X, representing the target t at time kt(k-1| k-1) is a one-step estimate of the simulated state of target t at time k-1, Pt(k-1| k-1) is a one-step estimation value of the state covariance of the target t at the moment k-1;
step S12: setting a reinforcement learning discount factor lambda and learning efficiency gamma, establishing an experience matrix Q-table of a reinforcement learning model, wherein a state s is measured distribution, an action a is weight selection of the experience matrix, and the Q-table is initialized to a 0 matrix.
3. The multi-target tracking data association method as claimed in claim 2, wherein the step of simulating random clutter points in step S2 comprises:
known measured value Z of target t at time kt(k|k),k=1,...,KtrainAmbient generation clutter Zflase,i(k):
Zflase,i(k)=Zt(k|k)+l-2l·rand0,1 (1);
Wherein l is the equivalent square side length of the elliptic wave gate, and i is equal to1, 2.. num _ flase is the number of clutter, rand0,1Is a random number between 0 and 1, KtrainFor the upper limit of the training process, T is 1, 2.
4. The multi-target tracking data association method as claimed in claim 3, wherein the step of obtaining the intra-gate candidate measurement points and the position distribution in step S2 comprises:
determining candidate measuring points and distribution in the gate according to the Mahalanobis distance between each measuring point and the target one-step predicting point at the previous moment and the wave gate;
determination of the measurement values Z (k):
calculating each measurement value Z (k) at the time k and a one-step predicted measurement value of the target tMahalanobis distance g oft(k):
If g ist(k) If the following condition is satisfied, the candidate measurement of the target t is retained and recorded as
gt(k)≤ζ (4);
Where ζ is the gate threshold;
one-step predictive measurement for target tRandomly generating echoes around and adding candidate measurements
The corresponding association probability is regarded as the probability that the intra-gate measurement is clutter; measured value predicted by target t one stepEstablishing a two-dimensional rectangular coordinate system for an origin to divide the wave gate into 4 areas, dividing the wave gate into a central area and an edge area by taking zeta/2 as a limit, dividing the wave gate into 8 areas in total, and calculating the distribution condition of the wave gate according to the position relation of each target candidate measuring point and one-step predicted measuring value.
5. The multi-target tracking data association method according to claim 4, wherein the step S3 specifically includes:
s31, selecting a weight in an experience matrix of an association model according to the distribution of the candidate measuring points;
step S32, calculating Euclidean distance between each candidate measuring point and one-step observation predicted value of the target to obtain the state matching degree of each candidate measuring point and the target;
step S33, obtaining the fluctuation influence of the weight value on the state matching degree by algebraic operation of the selected weight value and the Euclidean distance;
step S34, calculating Euclidean distance between each candidate measuring point and the target three-step observation predicted value to obtain the motion matching degree of each candidate measuring point and the target;
and step S35, obtaining the association probability of each candidate measuring point according to the motion matching degree and the state matching degree after fluctuation.
6. The multi-target tracking data association method according to claim 5, wherein the step S31 is specifically:
selecting the best action best _ action in the Q-table corresponding state according to the distribution of the candidate measuring points in each target gate:
best_action=max[Q(current s,all actions)] (6);
wherein current s is the current state, and each state corresponds to the position distribution measured in the wave gate; all actions are all actions, and each action represents weight selection:
wherein Δ is a scaling factor; finishing weight selection according to the mapping relation;
the step S32 specifically includes: calculating all candidate measuring values and one-step observation predicted values of target t at the moment kEuclidean distance of
In step S33, the multiplication operation is specifically selected, and specifically:
the step S34 specifically includes;
calculating a target t point trace X at the k-3 momenttThree-step state prediction value of (k-3| k-3)
Calculating all candidate measuring values and three-step observation predicted values of target t at the moment kEuclidean distance of
The step S35 specifically includes:
obtaining the association probability of all candidate measurement values of the target t at the moment k and the target t
7. The multi-target tracking data association method according to claim 6, wherein the step S4 specifically includes:
calculating Kalman gain K of target t at moment Kt(k) Sum-state covariance one-step estimate Pt(k|k):
Calculating a state one-step estimation value X of the target t at the moment kt(k|k):
8. The multi-target tracking data association method according to claim 7, wherein the step S5 specifically includes:
state one-step estimation value Xt(k | k) and the state covariance one-step estimate Pt(k | k) is used for point track-track correlation and substituted into the reinforcement learning experience matrix optimization module to train the Q-table:
for one step known measurement value Zt(k|k),k=1,...,KtrainPerforming Kalman filtering, and performing one-step estimation on the obtained training process statePerforming point track-track association:
Training the Q-table according to the reinforcement learning reward factor:
9. The multi-target tracking data association method according to claim 8, wherein the step S7 specifically includes:
for the correlation process, a state prediction value is calculatedOne-step observation prediction value ofAnd the estimated value XtOne-step observed prediction of (k | k)
St(k+1)=Ht(k+1)·Pt(k|k)·Ht(k+1)T (19);
Calculating reinforcement learning reward factor rt(k):
Optimizing the Q-table according to the reinforcement learning reward factor:
10. a multi-target tracking data association system, comprising a processor and a memory, wherein the memory stores a multi-target tracking data association program, and the processor executes the steps of the multi-target tracking data association method according to any one of claims 1 to 9 when running the multi-target tracking data association program.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010971580.4A CN112098993A (en) | 2020-09-16 | 2020-09-16 | Multi-target tracking data association method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010971580.4A CN112098993A (en) | 2020-09-16 | 2020-09-16 | Multi-target tracking data association method and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112098993A true CN112098993A (en) | 2020-12-18 |
Family
ID=73759875
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010971580.4A Pending CN112098993A (en) | 2020-09-16 | 2020-09-16 | Multi-target tracking data association method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112098993A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113095401A (en) * | 2021-04-12 | 2021-07-09 | 吉林大学 | Multi-sensor multi-target association tracking method |
CN113191427A (en) * | 2021-04-29 | 2021-07-30 | 无锡物联网创新中心有限公司 | Multi-target vehicle tracking method and related device |
CN113340308A (en) * | 2021-05-31 | 2021-09-03 | 西安电子科技大学 | Correction logic law flight path starting method based on self-reporting point |
CN113701758A (en) * | 2021-08-23 | 2021-11-26 | 中国北方工业有限公司 | Multi-target data association method and system based on biological search algorithm |
CN113985406A (en) * | 2021-12-24 | 2022-01-28 | 中船(浙江)海洋科技有限公司 | Target track splicing method for marine radar |
CN116628448A (en) * | 2023-05-26 | 2023-08-22 | 兰州理工大学 | Sensor management method based on deep reinforcement learning in extended target |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109901153A (en) * | 2019-03-29 | 2019-06-18 | 西安电子科技大学 | Targetpath optimization method based on information entropy weight and nearest-neighbor data correlation |
CN110824467A (en) * | 2019-11-15 | 2020-02-21 | 中山大学 | Multi-target tracking data association method and system |
CN111007495A (en) * | 2019-12-10 | 2020-04-14 | 西安电子科技大学 | Target track optimization method based on double-fusion maximum entropy fuzzy clustering JPDA |
-
2020
- 2020-09-16 CN CN202010971580.4A patent/CN112098993A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109901153A (en) * | 2019-03-29 | 2019-06-18 | 西安电子科技大学 | Targetpath optimization method based on information entropy weight and nearest-neighbor data correlation |
CN110824467A (en) * | 2019-11-15 | 2020-02-21 | 中山大学 | Multi-target tracking data association method and system |
CN111007495A (en) * | 2019-12-10 | 2020-04-14 | 西安电子科技大学 | Target track optimization method based on double-fusion maximum entropy fuzzy clustering JPDA |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113095401A (en) * | 2021-04-12 | 2021-07-09 | 吉林大学 | Multi-sensor multi-target association tracking method |
CN113191427A (en) * | 2021-04-29 | 2021-07-30 | 无锡物联网创新中心有限公司 | Multi-target vehicle tracking method and related device |
CN113191427B (en) * | 2021-04-29 | 2022-08-23 | 无锡物联网创新中心有限公司 | Multi-target vehicle tracking method and related device |
CN113340308A (en) * | 2021-05-31 | 2021-09-03 | 西安电子科技大学 | Correction logic law flight path starting method based on self-reporting point |
CN113701758A (en) * | 2021-08-23 | 2021-11-26 | 中国北方工业有限公司 | Multi-target data association method and system based on biological search algorithm |
CN113985406A (en) * | 2021-12-24 | 2022-01-28 | 中船(浙江)海洋科技有限公司 | Target track splicing method for marine radar |
CN113985406B (en) * | 2021-12-24 | 2022-05-10 | 中船(浙江)海洋科技有限公司 | Target track splicing method for marine radar |
CN116628448A (en) * | 2023-05-26 | 2023-08-22 | 兰州理工大学 | Sensor management method based on deep reinforcement learning in extended target |
CN116628448B (en) * | 2023-05-26 | 2023-11-28 | 兰州理工大学 | Sensor management method based on deep reinforcement learning in extended target |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110824467B (en) | Multi-target tracking data association method and system | |
CN112098993A (en) | Multi-target tracking data association method and system | |
CN109901153B (en) | Target track optimization method based on information entropy weight and nearest neighbor data association | |
CN105137418B (en) | Multiple target tracking and data interconnection method based on complete adjacent fuzzy clustering | |
CN113091738B (en) | Mobile robot map construction method based on visual inertial navigation fusion and related equipment | |
CN106980114A (en) | Target Track of Passive Radar method | |
CN106872955A (en) | Radar Multi Target tracking optimization method based on Joint Probabilistic Data Association algorithm | |
CN104035083B (en) | A kind of radar target tracking method based on measurement conversion | |
CN105510896B (en) | A kind of weighted nearest neighbor numeric field data correlating method of centralization multi-radar data processing | |
CN104155650A (en) | Object tracking method based on trace point quality evaluation by entropy weight method | |
CN104199022B (en) | Target modal estimation based near-space hypersonic velocity target tracking method | |
CN103729859A (en) | Probability nearest neighbor domain multi-target tracking method based on fuzzy clustering | |
CN107526070A (en) | The multipath fusion multiple target tracking algorithm of sky-wave OTH radar | |
CN106932771A (en) | A kind of radar simulation targetpath tracking and system | |
CN104156984A (en) | PHD (Probability Hypothesis Density) method for multi-target tracking in uneven clutter environment | |
CN107656265A (en) | Particle filter fusion method for tracking short flight path before multi frame detection | |
CN106054151A (en) | Radar multi-target tracking optimization method based on data correlation method | |
CN103985120A (en) | Remote sensing image multi-objective association method | |
CN111830501B (en) | HRRP history feature assisted signal fuzzy data association method and system | |
CN110501671A (en) | A kind of method for tracking target and device based on measurement distribution | |
CN109509207B (en) | Method for seamless tracking of point target and extended target | |
CN105424043A (en) | Motion state estimation method based on maneuver judgment | |
Xia et al. | Extended object tracking with automotive radar using learned structural measurement model | |
CN108712725B (en) | SLAM method based on rodent model and WIFI fingerprint | |
CN104182652B (en) | Typical motor formation target tracking modeling method |
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 |