CN115099343B - Distributed tag Bernoulli fusion tracking method under limited field of view - Google Patents
Distributed tag Bernoulli fusion tracking method under limited field of view Download PDFInfo
- Publication number
- CN115099343B CN115099343B CN202210738739.7A CN202210738739A CN115099343B CN 115099343 B CN115099343 B CN 115099343B CN 202210738739 A CN202210738739 A CN 202210738739A CN 115099343 B CN115099343 B CN 115099343B
- Authority
- CN
- China
- Prior art keywords
- target
- sensor
- lmb
- targets
- posterior distribution
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 102
- 230000004927 fusion Effects 0.000 title claims abstract description 98
- 238000012423 maintenance Methods 0.000 claims abstract description 32
- 230000000007 visual effect Effects 0.000 claims abstract description 21
- 230000005540 biological transmission Effects 0.000 claims abstract description 17
- 238000001914 filtration Methods 0.000 claims description 43
- 238000000605 extraction Methods 0.000 claims description 22
- 238000004422 calculation algorithm Methods 0.000 claims description 15
- 238000012937 correction Methods 0.000 claims description 15
- 238000005259 measurement Methods 0.000 claims description 15
- 230000033001 locomotion Effects 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 230000009707 neogenesis Effects 0.000 claims description 9
- 238000000638 solvent extraction Methods 0.000 claims description 6
- 235000008694 Humulus lupulus Nutrition 0.000 claims description 5
- 238000009792 diffusion process Methods 0.000 claims description 4
- 238000007499 fusion processing Methods 0.000 claims description 4
- 238000005192 partition Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 abstract description 5
- 238000005516 engineering process Methods 0.000 abstract description 2
- 230000010365 information processing Effects 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 description 10
- 238000007500 overflow downdraw method Methods 0.000 description 8
- 230000006870 function Effects 0.000 description 7
- 238000012544 monitoring process Methods 0.000 description 7
- 238000004891 communication Methods 0.000 description 6
- 239000000284 extract Substances 0.000 description 6
- 238000013507 mapping Methods 0.000 description 6
- 230000004083 survival effect Effects 0.000 description 6
- 101001086191 Borrelia burgdorferi Outer surface protein A Proteins 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- SLXKOJJOQWFEFD-UHFFFAOYSA-N 6-aminohexanoic acid Chemical compound NCCCCCC(O)=O SLXKOJJOQWFEFD-UHFFFAOYSA-N 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a distributed tag multiple Bernoulli fusion tracking method under a limited visual field, and belongs to the fields of intelligent information processing technology and signal processing. Firstly, using flooding transmission to enable a single sensor to have posterior distribution information of all targets in a sensor network, adopting a clustering-based target dividing method to accurately classify the targets, executing GCI fusion on the fusion targets, and directly extracting the target distribution information if the fusion targets are not fusion targets; aiming at a moving target crossing a sensor or a single sensor tracking target, the invention utilizes the target label and the sensor identification information to establish a track maintenance table, and the target local label is recovered by matching with the track maintenance table, so that the target track is continuous. The invention effectively solves the problem of target tracking missing caused by target distribution information loss in a non-overlapping area due to GCI fusion in the limited field sensor network, and can output the state information and complete tracks of all targets in the global field of the network.
Description
Technical Field
The invention relates to a distributed tag multiple Bernoulli fusion tracking method under a limited visual field, and belongs to the fields of intelligent information processing technology and signal processing.
Background
Multi-target tracking refers to estimating multiple target states in a complex environment using sensor measurements, the number of targets and the target states varying with time during tracking. The traditional multi-target tracking method is based on data association, the probability of target and measurement association hypothesis is calculated at each tracking moment, and a typical method is provided with a joint probability data association filter and multi-hypothesis tracking. In theory, the method based on data association can accurately track the targets, but in the actual application process, when the number of tracked targets is large, the problem of 'combined explosion' occurs, namely, the excessive calculation amount is caused by the excessive association assumption.
Mahler in 2003 proposed a probability hypothesis density (Probability Hypothesis Density, PHD) filter based on a random finite set that avoided data correlation during filtering and tracked well. Thereafter, a Multi-Target tracking method based on a random finite set is receiving a great deal of attention, and a Multi-Target tracking method such as a potential probability hypothesis density filter (Cardinality Probability Hypothesis Density, CPHD), a potential equalization Bernoulli filter (Cardinality Balanced Multi-Target Multi-Bernoulli, CBMeMBer), a Labeled Bernoulli (LMB) filter, and the like is used. Wherein, the LMB filter not only can accurately estimate the multi-target state, but also can provide a target track.
And the distributed fusion of the sensors is that each node in the sensor network exchanges filtering posterior with an adjacent sensor through communication, and the computing capability of the sensor is utilized to fuse multi-target posterior distribution so as to improve the target tracking precision. The fusion method commonly used at present is a fusion method based on generalized covariance intersection (Generalized Covariance Intersection, GCI), the posterior distribution of the potential probability hypothesis density, potential equalization Bernoulli, label Bernoulli and other filters is related to GCI fusion, and the multi-target tracking method based on fusion effectively improves the tracking performance of the filters.
Compared with other methods, the LMB-based distributed fusion method provides a target track management function, but the LMB distribution GCI fusion has the problems of high calculation complexity and inconsistent labels. The reason why the targets are inconsistent in the multi-sensor tracking application scene is analyzed in detail by Li Suqi et al, and a robust tag multi-filter distributed fusion method based on no-tag is provided (Li Suqi, hoseinnezhad, reza, et al, robust Distributed Fusion With Labeled Random Finite sets.). Li Suqi et al further propose a label matching fusion algorithm based on optimal matching (Li Suqi, battisteli G, chisci L, et al, computing application Multi-Agent Multi-Object Tracking With Labeled Random Finite Sets [ J ]. IEEE Transactions on Signal Processing, 2018:1-1.), which has lower computational complexity than the robust label Multi-filter distributed fusion method based on label-free, but all of the above methods have the same field of view based on multiple sensors in the sensor network.
When the sensors are placed properly, the overlapping area of each sensor field of View (FoVs) in the sensor network is larger, and the targets only move in the overlapping area, at the moment, the monitoring Fields of each sensor can be considered to be consistent, if the types of the sensors are the same and the performances of the sensors are good, the posterior distribution difference of the filtered multiple targets is not large, and the tracking precision can be effectively improved by adopting a fusion algorithm.
However, in practical application, the single sensor has limited monitoring vision due to the limitation of sensing distance and angle, and cannot cover the whole vision range to be monitored actually or only part of monitoring vision overlaps among the sensors due to station arrangement. When the monitoring visual fields of the sensors are partially overlapped, tracking the movement of the target across the sensors, and directly fusing the sensor filtering posterior by adopting a distributed fusion method based on GCI fusion, the target distribution information outside the non-overlapping area is lost, and the target is missed.
Disclosure of Invention
Aiming at the problems that in an actual distributed sensor network, the visual field of a single sensor is limited, only targets with specific angles and distances can be tracked, only partial detection areas among adjacent communication sensors are overlapped, posterior distribution among the sensors is only partially consistent, and target distribution information in non-overlapping areas is lost and target tracking is missed due to sensor fusion, the invention provides a distributed tag Bernoulli fusion tracking method under the limited visual field, which comprises the following steps:
Step one: deploying a sensor network and initializing sensor parameters;
the sensor network consists of N sensors, and the sensor set is recorded as S= { S i :i∈N}，s i The sensor is identified, and has uniqueness;
step two: sensor s i Operating LMB filtering algorithm to obtain k timeCarved multi-objective posterior distribution LMB parameter set:
wherein,a target tag space at time k, r (l) represents the existence probability of a target with a tag of l, and p (l) represents the probability density distribution of the target with the tag of l;
step three: the sensor s i LMB parameter set for multi-objective posterior distribution of k moment Extracting target to obtain LMB parameter set of posterior distribution of extracted target->And diffusing the extraction target posterior distribution LMB parameter set in the network by a flooding transmission method>Meanwhile, obtaining LMB parameter sets of the posterior distribution of the extraction targets of other sensors;
step four: the sensor s i Obtaining LMB parameter set of posterior distribution of all sensor extraction targetsThen, adopting a clustering-based partitioning method to perform ∈N & lt/M & gt>Dividing targets in the system into a fusible target and an unfused target, wherein the fusible target is a target jointly tracked by a plurality of sensors, and the unfused target is Targets tracked by a single sensor alone;
step five: after the target division is completed, GCI fusion is carried out on the fusible category targets to obtain a fusible category target fusion posterior distribution LMB parameter setAnd label correction is carried out in the fusion process, so that the aim of maintaining the track is fulfilled;
for non-fusible targets, directly extracting a posterior distribution LMB parameter set of the targets and performing label correction according to requirements, wherein the method comprises the following steps:
if there are only sensors s in the class i Directly extracting posterior distribution LMB parameter sets of the targets as part of fusion posterior distribution LMB parameter sets of non-fusible targets to obtain the fusion posterior distribution LMB parameter sets of the targets
If there are only sensors in the classExtracting a posterior distribution LMB parameter set of such a target as part of a fusion posterior distribution LMB parameter set of an unfused class of targets, but performing tag correction on tag information of such a target posterior distribution LMB parameter set so that such a target is located at the sensor s i The whole movement life cycle has the same label under the visual angle to obtain the target fusion posterior distribution LMB parameter set +.>
The fusible targets and the non-fusible targets maintain the track of the targets according to a track maintenance strategy, so that a single sensor outputs a complete track of the moving targets crossing the sensor;
Step six: combining the fused posterior LMB parameter set obtained in the fifth step and the directly extracted target posterior LMB parameter set into a target state extraction posterior LMB parameter set, wherein the sensor s i Extracting a posterior LMB parameter set according to the target state and outputting all target state information in a sensor network tracking area;
step seven: utilizing the posterior parameter set in the fifth stepAnd (3) forming the filtering priori at the next moment, and continuously tracking all targets in the sensor network by repeating the steps two to six. Wherein (1)>LMB parameter set representing fusion posterior distribution of fusion class object,/->Representing sensor s i Fusion posterior distribution LMB parameter sets for tracked non-fusible class targets.
Optionally, the second step includes:
assume that at time k-1, sensor s i Filtering to obtain a posterior LMB set which is as follows:
the target neogenesis model is selected from a self-adaptive neogenesis model, and a neogenesis target LMB parameter set is as follows:
wherein,representing a nascent target tag space;
at time k, an LMB prediction step is firstly performed in each sensor to obtain a predicted LMB parameter set Next, Z is measured at time k using a sensor k For said predicted LMB parameter set +.>Updating to obtain a filtered posterior distribution LMB parameter set +.>And simultaneously, a new target parameter set at the next moment is generated by measurement.
Optionally, the third step includes:
sensor s i Posterior distribution LMB parameter set obtained by LMB filteringComprising a sensor s i Real target and false alarm distribution information in visual field, LMB parameter set of posterior distribution +.>Performing potential estimation and target extraction to obtain potential estimation +.>Extracting target posterior distribution LMB parameter set +.>And false alarm posterior distribution LMB parameter set +.>
Obtaining potential estimatesExtracting a target posterior distribution LMB parameter set +.>After that, sensor s i By passing throughAnd the flooding transmission exchanges and extracts target posterior distribution LMB set information with other sensors in the sensor network:
when T is less than or equal to T, the sensor s i The information diffusion is completed by iteratively exchanging the latest acquired posterior distribution LMB parameter set information with the adjacent interconnected sensors, so that the flooding transmission purpose is achieved; when T > T, indicating completion of flooding transmission, sensor s i Collecting target posterior sets extracted by all sensors in networkT represents the number of iterative exchanges of information.
Optionally, the fourth step includes:
first, a target posterior distribution LMB parameter set is extracted according to all sensorsEstimating the target state of each LMB component to obtain +.>As clustered data sample points, wherein l n Representing target tags, s j Representing sensor identification,/- >Representing sensor s j A target tag space;
second, according to the sensor s i Target extraction potential estimationLet the number of the initial categories of the classification be->With sensors s i State estimation of +.>For initial cluster center->Wherein (1)>Representing sensor s i Estimating a target potential;
finally, calculating the estimation state of each targetTo the cluster center->According to Euclidean distance, completing target clustering division; the cluster partition objective satisfies two constraint conditions: the sample points from the same sensor in each type are at most one, the distance between the sample points in each type and the clustering center is not greater than a distance threshold eta, the points which do not meet the two constraint conditions are added as new clustering centers, and iterative clustering is performed until the constraint is met.
Optionally, the process of maintaining the track of the target according to the track maintenance strategy in the fifth step includes:
when the target moves across the field of view of the sensor, the state of the category where the target is located changes, and when the target moving across the sensor is divided into an unfused category of target, the target only carries out state extraction, so that the target track is divided into two independent parts, and the track maintenance strategy is required to be executed by utilizing track maintenance table information to output a complete track;
the elements of each row in the track maintenance table are Wherein the method comprises the steps ofRepresenting sensor s i Extracting target tag-> A potential estimate is extracted for the target at time k,the representation and label are +>Based on a clustering method, all target labels and sensor identification information in a target set divided into the same class>
If the target is a fusible target, the target is fused for the first timeThe information is added to the track maintenance table, and the follow-up tracking time can be fused with the category target information to be consistent with the track maintenance table information, so that correction is not required;
if from sensor s i The target is a local filtering tracking target, and the target label information does not need to be corrected when the label is unchanged in the local filtering process;
if from a sensorNon-fusible targets based on target labels and sensor identification informationSearching track maintenance table if there is +.>Information, the target tag is corrected to +.>The sensor identity is corrected to s i The method comprises the steps of carrying out a first treatment on the surface of the If there is no->Information representing the sensor s i Extracting the target information for the first time, initializing the target at the sensor s i The middle label is->Add->And the information is sent to a track maintenance table for label correction at the next moment.
Optionally, the extracting a posterior distribution LMB parameter set of the six target states in the step is:
wherein, LMB parameter set representing fusion posterior distribution of fusion class object,/->Representing sensor s i Fusion posterior distribution LMB parameter set of tracked non-fusible class target, ++>Representing sensor s j Fusion posterior distribution LMB parameter sets for tracked non-fusible class targets.
Optionally, the state of the target includes: the position of the target, the speed of the target, and the turn rate.
Optionally, the number N of the sensors is 4.
Optionally, the sensor is a limited field of view sensor.
The invention has the beneficial effects that:
the invention provides a distributed tag multiple Bernoulli fusion tracking method under a limited visual field, which comprises the steps of firstly enabling a single sensor to have posterior distribution information of all targets in a sensor network by using flooding transmission, accurately classifying the targets by adopting a clustering-based target dividing method, executing GCI fusion on the fusible targets in a visual field overlapping region, extracting the distribution information of the targets if the non-fusible targets outside the visual field overlapping region are out, and avoiding losing the non-overlapping monitoring visual field region target distribution information. In addition, aiming at the problem that the continuity of the target label of the sensor-crossing moving target cannot be maintained only by simple state extraction of the target in the non-overlapping area, the invention establishes the track maintenance table by utilizing the target label and the sensor identification information, and the target is tracked by the sensor-crossing moving target or the single sensor, and the target local label is recovered by matching with the track maintenance table, so that the target track is continuous, and the problem of target tracking missing caused by the loss of the target distribution information in the non-overlapping area is effectively solved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flow chart of the method of the present invention.
Fig. 2 is a simulation diagram of a target nonlinear motion track obtained by tracking a three-sensor six-target tracking scene by using a distributed tag bernoulli fusion tracking method under a limited field of view according to an embodiment of the present invention.
Fig. 3 is a graph comparing OSPA experimental results of the method of the present invention with those of other methods in a three-sensor six-target scenario.
Fig. 4 is a graph comparing the experimental results of potential estimation of the method of the invention with other methods in a three-sensor six-target scenario.
FIG. 5 is a graph of single sensor single run output target state estimation for a three sensor six target scenario of the method of the present invention.
Fig. 6 is a simulation diagram of a target nonlinear motion track obtained by tracking in a four-sensor seven-target tracking scene by using the distributed tag bernoulli fusion tracking method under a limited field of view according to an embodiment of the present invention.
Fig. 7 is a graph comparing OSPA experimental results of the method of the present invention with other methods in a four-sensor seven-target scenario.
Fig. 8 is a graph comparing the experimental results of potential estimation of the method of the invention with other methods in a four-sensor seven-target scenario.
FIG. 9 is a graph of single sensor single run output target state estimation for a four sensor seven target scenario of the method of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the embodiments of the present invention will be described in further detail with reference to the accompanying drawings.
First, the related basic theory related to the present application is described as follows:
1. label random finite set
A random finite set (Random Finite Set, RFS) is a set of random set elements and random numbers of elements, and can be used to describe a target state and a target number time-varying tracking scenario. Thus, a random finite set is usedRepresenting multiple target states, and X represents a target state space.
In addition, to manage the target track, unique tag information L ε L= { α may be added for each target state X ε X i I e N, where N represents a set of positive integers, L represents a target tag space, and the tag information can be used for target identification. The above description shows that the tag RFS is defined on the space x×l, and that the tags of different targets have uniqueness.
To describe the tag RFS, a mapping function L is defined, x×l→l, for any target L ((X, L))=l. It can be seen from the mapping function that each target tag represented by the random finite set X has uniqueness if and only if |x|= |l (X) |= |{ L (X): X e X } |. Thus, a function Δ (X) =δ is defined |X| (|L (X) |) is a tag inconsistency indicator.
2. Label Bernoulli
The tags randomly finite set describe the target state using the target existence probability and probability density distribution. Assuming that the target label is L epsilon L, and the existence probability is r l Probability density distribution is p l The label Bernoulli (LMB) probability density distribution is expressed as:
π(X)＝Δ(X)ω(L(X))p(x,l) X
wherein:
and p (x, l) =p l (x) Delta (X) is a tag inconsistency indicator function, and L (X) is the total target tag set obtained based on the mapping L.
To simplify the description, LMB probability densities are typically characterized by a parameter set comprising r and p:
π＝{(r l ,p l )} l∈L
the LMB filtering tracking multi-target can estimate multi-target time-varying states and provide target tracks, and one complete filtering comprises a prediction step and an updating step. And in the predicting step, a new target set is newly added and the motion state of the survival target is predicted. In the updating step, the prediction set is updated by using the sensor measurement, when the new target parameter is updated by measurement, the existence probability of the new target parameter is increased, and when the survival target is not updated by measurement at the last moment, the existence probability of the new target parameter is reduced.
2.1LMB Filter prediction step
Every filtering time there may be a new target, the new target LMB parameter set is expressed as follows:
wherein B represents the nascent target tag space.
At time k-1, the sensor performs LMB filtering, and the multi-target posterior probability distribution after filtering is as follows:
wherein L is k-1 Representing the tag space of the object at time k-1.
At time k, firstly, a prediction step is executed for multi-target probability density, a survival target state set is obtained by prediction through a last time filtering posterior, and the survival target state set and a new target parameter set form multi-target prediction distribution at time k, which is expressed as:
wherein,
η s (l)＝<p s (·,l),p k-1 (·,l)>
in the above, eta s (l) The survival probability of the target labeled i is represented,<f,g>＝∫f(x)g(x)dx。
2.2 LMB filtering update step
After the prediction step, the LMB filter updates the prediction distribution by using the measurement of the k-moment sensor, and the k-moment multi-target posterior probability distribution is obtained after the update, which is expressed as:
in the above, L k Is the tag space of the object at time k.The existence probability and probability density distribution are updated for the label for the l object.
Wherein,
wherein Θ represents the target and metrology mapping space, (I, θ) ∈F (L) k ) XΘ represents a mapping assumption of target and measurement, ω (I,θ) Weights representing mapping hypotheses; p is p θ (x, l) indicates that under the map θ, the target labeled l corresponds to a measure that is used to update the target LMB parameters to get the distribution of the target at time k. P is p D (x, l) represents the probability that an object labeled l is detected, g (z) θ(l) I x, l) is the objective likelihood function.
2.5 generalized covariance intersection fusion method
Multi-sensor network utilization sensorAnd the sensor communicates with the exchange filtering posterior, and fusion is carried out on multiple targets through a proper method to improve the tracking performance targets. The distributed fusion is that the sensor exchanges filtering posterior through communication with the adjacent sensor and completes fusion locally, and a fusion center does not exist. Assuming time k, the multi-target state is denoted as X k All the measurement sets of the sensor from the initial time to the current time are expressed as Z 1:k Filtering by two sensors to track multiple targets and obtain a multi-target tracking posterior Fusion posterior can then be expressed as
When the types of the sensors are the same, similar process noise and measurement noise exist in filtering tracking, so that the multi-target posterior obtained by filtering of the sensors are in non-independent and same distribution and have correlation.There is a repetition of the calculations and the fusion process can be improved with the following
But in the practical application of the present invention,and the method is difficult to calculate and cannot be truly applied to multi-target posterior distribution fusion.
Mahler proposes a generalized covariance intersection (Generalized Covariance Intersection, GCI) method based on covariance intersection (Covariance Intersection, CI), which solves the problem of repeated computation. Assuming that the sensor network is represented as S, S epsilon S is represented as a sensor node. Assuming the k moment, the sensor node S epsilon S is filtered to obtain multi-target posterior distribution pi k,s (X), the posterior distribution for GCI fusion can be represented by the following formula
Wherein,
the multi-target posterior distribution is subjected to GCI fusion to obtain a fusion posterior, the fusion posterior replaces a sensor local filtering posterior to be used for filtering at the next moment, iteration filtering is achieved to continuously carry out multi-target tracking, and the effective fusion can improve the tracking performance of the filter.
Embodiment one:
the embodiment provides a distributed tag bernoulli fusion tracking method under a limited field of view, referring to fig. 1, the method includes:
step one: deploying a sensor network and initializing sensor parameters;
the sensor network consists of N sensors, and the sensor set is recorded as S= { S i :i∈N}，s i The sensor is identified, and has uniqueness;representation and sensor s i A set of sensors having a distance not exceeding t hops and comprising a sensor s i ；/>Representation and sensor s i Sensor set with distance not exceeding t hops and containing no sensor s i ；
Step two: sensor s i And (3) operating an LMB filtering algorithm to obtain a multi-target posterior distribution LMB parameter set at the k moment:
wherein,a target tag space at time k, r (l) represents the existence probability of a target with a tag of l, and p (l) represents the probability density distribution of the target with the tag of l;
step three: the sensor s i LMB parameter set of the multi-objective posterior distribution for the k-time instant Extracting target to obtain LMB parameter set of posterior distribution of extracted target->And diffusing the extraction target posterior distribution LMB parameter set in the network by a flooding transmission method>Meanwhile, obtaining LMB parameter sets of the posterior distribution of the extraction targets of other sensors;
step four: the sensor s i Obtaining LMB parameter set of posterior distribution of extraction targets of all other sensorsThen, adopting a clustering-based partitioning method to perform ∈N & lt/M & gt>Dividing targets in the system into a fusible target and an unfused target, wherein the fusible target is a target jointly tracked by a plurality of sensors, and the unfused target is a target independently tracked by a single sensor;
step five: after the target division is completed, the fusible class is classified Performing GCI fusion on the target to obtain a fusion posterior distribution LMB parameter set of the fusion type targetAnd label correction is carried out in the fusion process, so that the aim of maintaining the track is fulfilled;
for the non-fusible targets, directly extracting the posterior distribution LMB parameter set of the targets, and carrying out label correction according to requirements, wherein the method comprises the following steps:
if there are only sensors s in the class i Directly extracting posterior distribution LMB parameter sets of the targets as part of fusion posterior distribution LMB parameter sets of non-fusion targets to obtain fusion posterior distribution LMB parameter sets of the targets
If there are only sensors in the classExtracting a posterior distribution LMB parameter set of such a target as part of a fused posterior distribution LMB parameter set of an unfused class of targets, but performing tag correction on tag information of such a target posterior distribution LMB parameter set so that such a target is located at the sensor s i The whole movement life cycle has the same label under the visual angle to obtain the target fusion posterior distribution LMB parameter set +.>
The fusible targets and the non-fusible targets maintain the track of the targets according to a track maintenance strategy, so that a single sensor outputs a complete track of the moving targets crossing the sensor;
Step six: combining the fusion posterior distribution LMB parameter sets of the fusible target and the non-fusible target obtained in the step five into a target state extraction posterior distribution LMB parameter set, wherein the sensor s i Extracting posterior distribution LMB parameter sets according to the target state to output all the tracking areas of the sensor networkTarget state information;
step seven: utilizing posterior set in step fiveThe next moment filtering priori is formed, and all targets in the sensor network can be continuously tracked by repeating the second to sixth steps; wherein (1)>LMB parameter set representing fusion posterior distribution of fusion class object,/->Representing sensor s i Fusion posterior distribution LMB parameter sets for tracked non-fusible class targets.
Embodiment two:
the embodiment provides a distributed tag multiple Bernoulli fusion tracking method under a limited field of view, and a tracking method framework is shown in fig. 1, wherein the method comprises the following steps:
step one: deploying a sensor network and initializing sensor parameters;
n visual field limited sensors form a sensor network to track targets in a specific range, and the sensor set is recorded as S= { S i :i∈N}，s i Has uniqueness and uniquely identifies a sensor; in addition, useRepresentation and sensor s i A set of sensors that are no more than t hops apart.
Step two: the sensor performs local LMB filtering;
the sensor runs an LMB filtering algorithm realized based on a sequential Monte Carlo (Sequential Monte Carlo, SMC) method, and multi-target posterior distribution information is exchanged by communicating with adjacent sensors after filtering.
Assume that at time k-1, sensor s i Filtering to obtain a posterior LMB set asThe target neogenesis model is selected from an adaptive neogenesis model, and a neogenesis target LMB parameter set is +.>
At time k, an LMB prediction step is firstly performed in each sensor to obtain a predicted LMB parameter setSecond, measurement Z at time k using a sensor k For prediction of LMB parameter set->Updating to obtain a filtered posterior distribution LMB parameter set +.>And simultaneously, a new target parameter set at the next moment is generated by measurement.
Step three: the sensor processes the local data and performs flooding transmission;
sensor s i Performing LMB filtering to obtain posterior distribution LMB parameter setThe posterior distributed LMB parameter set contains the sensor s i The real targets and the false alarm distribution information in the visual field are only fused and extracted for the target set extracted by each sensor, so that the fusion performance is improved and the calculated amount is reduced. From step two, at time k, sensor s i The posterior distribution LMB parameter set parameter of +. >For->Performing potential estimation and extracting LMB parameter sets of posterior distribution corresponding to the target and false alarm respectively to obtain potential estimation +.>And extracting target posterior distribution LMB parameter set +.>False alarm posterior distribution LMB parameter set +.>
Obtaining potential estimatesExtracting a target posterior distribution LMB parameter set +.>After that, sensor s i And exchanging and extracting posterior distribution information of the target set with other sensors in the sensor network through flooding transmission. When T is less than or equal to T, the sensor s i And information diffusion is completed by iteratively exchanging latest acquired posterior distribution LMB parameter set information with adjacent interconnected sensors, so that the flooding transmission purpose is achieved. When T > T, indicating completion of flooding transmission, sensor s i Collecting LMB parameter set of posterior distribution of target extracted by all sensors in networkT represents the number of iterative exchanges of information.
Step four: dividing targets;
according to the sensor in the third step, the flooding transmission process is carried out, and the sensor s i Obtaining LMB parameter set of posterior distribution of all sensor extraction targetsAdopting a clustering-based partitioning method pair +.>The medium targets are divided, and the targets in the posterior distribution LMB parameter set are divided into fusible targets and non-fusible targets.
First, LMB parameter sets are distributed according to a posteriorEstimating the target state of each LMB component to obtain +. >As clustered data sample points, wherein l n Representing target tags, s j Representing sensor identification,/->Representing sensor s j Target tag space. Second, according to the sensor s i Target extract potential->Let the number of the initial categories of the classification be->With sensors s i State estimation->For initial cluster center->Finally, calculate->And->Estimating the Euclidean distance from the state of each target to the clustering center, and completing the clustering division of the targets according to the distance; the clustering partitioning goal should satisfy two constraints: and if the distance between the sample points from the same sensor in each class and the clustering center is not larger than the distance threshold eta, adding the abnormal points in the class as new clustering centers, and carrying out iterative clustering until the constraint is met.
Clustering is completed through iterative computationThe labels are divided into three categories: the class contains the sensor s at the same time i And (3) withExtracting objects, i.e. sensors s i And sensor s j Jointly tracking a target which is a fusible target; the class contains only the sensor s i Extracting objects, i.e. sensors s i Independently tracking and extracting targets which are non-fusible targets; the class contains only the sensor->Extracting an object, i.e. sensor->The extracted targets are tracked separately, and are also non-fusible targets.
Step five: maintaining the track according to class fusion;
according to the target division result of the step four, performing GCI fusion on targets in the fusible class to obtain a fusible class target fusion posterior distribution LMB parameter setFor non-fusible classes, if there are only sensors s in the class i Directly extracting posterior distribution LMB parameter sets of the targets as part of fusion posterior distribution LMB parameter sets of non-fusion targets to obtain fusion posterior distribution LMB parameter sets of the targets +.>If there are only sensors in the class +.>Extracting posterior distribution LMB parameter sets of the targets as part of fusion posterior distribution LMB parameter sets of non-fusion targets to obtain fusion posterior distribution LMB parameter sets of the targets>All three types of targets execute the track maintenance strategy according to the requirement.
When the object moves across the field of view of the sensor, the state of the category in which the object is located changes. For a sensor-crossing moving target, when the sensor-crossing moving target is divided into an unfused target, the target track is divided into two independent parts by only performing state extraction, and the track maintenance strategy is required to be executed by utilizing track maintenance table information to output a complete track. The elements of each row in the track maintenance table areWherein->Representing local sensor s i Extracting target tag-> The potential is extracted for the target at time k. />The representation and label are +>The targets are based on a clustering method, and all target labels and sensor identification information in a target set which is divided into the same type are +.>
In the case of a fusible class, the fusion will occur for the first timeAdding the target information to a track maintenance table, and fusing target information and track dimensions in the class at the follow-up tracking timeThe table holding information is consistent, and correction is not needed; if from sensor s i And the target is a local filtering tracking target, and the target label information does not need to be corrected when the label is unchanged in the local filtering tracking process. If from sensor s j (s j ∈S i /s i ) Non-fusible targets based on target labels and sensor identification informationSearching track maintenance table if there is +.>Information, the target tag is corrected to +.>The sensor identity is corrected to s i . If there is no->Information representing the sensor s i Extracting the target information for the first time, initializing the target at the sensor s i The middle label is->Add->And the information is sent to a track maintenance table for label correction at the next moment.
Step six: output target state estimation
Fusing the posterior distribution LMB parameter sets according to the fusible class targets and the non-fusible class targets obtained in the step five to form a target state extraction posterior distribution LMB parameter set Sensor s i All target state information in the sensor network tracking area may be output.
Step seven: iterative filtering
Utilizing the posterior parameter set in the fifth stepAnd (3) forming the filtering priori at the next moment, and continuously tracking all targets in the sensor network by repeating the steps two to six.
The effect of the invention can be further illustrated by the following experiments:
1. software and hardware and related parameter setting in experimental process
The method is characterized in that the processor is i5-4210M, 2.6GHz is adopted, the memory is 8GB, and the experiment is written by adopting Matlab R2021b software.
In the experiment, the state of each target includes the position of the target, the speeds in the x and y directions, and the turning rate, i.e., x k ＝[d x ,v x ,d y ,v y ,w k ]Wherein (d) x ,d y ) Representing an object in a rectangular coordinate system, (v) x ,v y ) Representing the velocity in the x and y directions, w, respectively k Indicating the turn rate. The target detection probability is P d =0.98, survival probability P s =0.99. The target state transfer function is
f k|k-1 (x k |x k-1 )＝N(x k ,K(w k-1 )x k-1 ,Q)
Wherein K (w) is a state transition matrix, and Q is the process noise intensity.
Wherein sigma w ＝5，σ u Pi/180 is the process noise standard deviation. Clutter distribution follows poisson distribution, the poisson parameter is lambda=10, and the intensity is alpha=1/2000 pi.
The sensor uses a nonlinear measurement model to generate measurements at each moment
Wherein (d) x k ,d y k ) Represents the coordinate position of the target at time k, (d) i,x ,d i,y ) Representing the coordinate position of the sensor, and the noise obeys the distribution epsilon k ～N(·；0，R k )，σ θ ＝(π/180)，σ r ＝5m。
The application provides a distributed sensor-based label multi-Bernoulli fusion multi-target tracking method, experiments are carried out under two scenes, the method is compared with a single sensor LMB algorithm-based target tracking method and a newer and representative LMB distributed fusion algorithm LM-GCI-LMB-based target tracking method, and effectiveness of the proposed algorithm is verified.
The LMB algorithm is described herein with reference to "S.Reuter, B.T.Vo, B.N.Vo, et al, the Labeled Multi-Bernoulli Filter [ J ]. Signal Processing,2014,62 (12): 3246-3260 ].
LM-GCI-LMB can be referred to as "S.Li, G.Battistelli, L.Chisci, et al, computing office Multi-agent Multi-object tracking with labeled random finite sets [ J ]. IEEE Transactions on Signal Processing,2019,67 (1): 260-275 ].
2. Experiment and result analysis
According to the method provided by the invention, through multi-sensor scene experiments, in the sensor network formed by the sensor with limited vision, a single sensor can effectively track all targets in the global vision of the sensor network and extract the complete track of the targets.
As shown in fig. 2, three sensors in this scenario track six targets, which perform nonlinear motion for a duration of 60 moments. As shown in fig. 6, four sensors in this scenario track seven targets, which likewise move non-linearly, with a target movement duration of 80 moments. In both scenarios, there is an object that moves only within a single sensor field of view, as well as there is a moving object across the sensor field of view.
Experiment one: three sensor scenario
The validity of the method is verified and put forward in the process that three sensors track six target scenes in the experiment, the duration of the whole tracking scene is 60s, and communication among the sensors is Hong Canshu t=2. The three sensors are all limited field sensors, and can only detect and track targets with specific angles and distances. The target motion track and sensor deployment condition in the experimental scene is shown in fig. 2.
The results of fig. 3 show that the average OSPA of the multi-objective tracking method of the present invention is minimal compared to other methods, because:
according to the method, the fusible class targets and the non-fusible class targets are divided through a clustering method, GCI fusion is adopted on the fusible class target distribution information, the non-fusible class targets directly extract the distribution information of the fusible class targets as fusion posterior, and the non-overlapping monitoring field area target distribution information is prevented from being lost. While other methods do not take the necessary partitioning method for the target, the Murty algorithm will perform forced allocation based on the cost matrix of distance or divergence, so that the target may have invalid fusion. For the target in the overlapping area, the target distances are similar and the distribution is similar, so that the target in the area can be accurately matched and fused, and the LM-GCI-LMB can fuse the posterior distribution of the co-tracking targets in the overlapping area; for a target tracked by a single sensor only, the posterior distribution is matched and fused with the non-real target distribution (false alarm) or the target distribution tracked by the single sensor, so that the original posterior distribution of the target is 'polluted', real distribution information is lost, and the target is completely missed. The LMB does not take fusion and is able to track all targets within its monitored field of view.
The result of fig. 4 shows that the potential estimation of the method is the most accurate, and the estimation of the target number in the global field of view of the network is more accurate, but the potential estimation is slightly overestimated due to the fact that the false alarm targets existing in the non-fusible targets cannot be distinguished. Other methods have lower potential estimates than the number of targets in a real sensor network. The LMB potential estimation is low in that the LMB potential estimation cannot have the whole sensor field of view, the LM-GCI-GCI does not distinguish targets, and the target posterior is directly fused by adopting a GCI method, so that the real target has missing heel, and the potential estimation is lowest.
Fig. 5 shows a single filtered tracking state output graph of the method of the present invention. The results in fig. 5 show that the method of the invention can effectively provide the sensor network to monitor the complete track of the target in the visual field, and overcomes the defect that the single-sensor LMB filtering with limited visual field can not provide the complete track of the target for the moving target crossing the sensor.
Experiment II: four sensor scenario
Four sensors track seven targets as shown in fig. 6, the duration of the whole tracking scene is 80s, and inter-sensor communication is Hong Canshu t=3.
Fig. 7 shows an average OSPA graph of the methods in this scenario. The results in fig. 7 show that the average OSPA of the method of the present invention is still the lowest and the fusion effect is the best in the four-sensor tracking experiment scenario. In the target tracking method based on the LM-GCI-LMB, as the number of sensors is increased, the more sensors participating in fusion are caused by posterior distribution differences among different sensors, the worse the effect of a conventional fusion algorithm is, and as the filtering moment is increased, the influence diffusion caused by the communication among the sensors is caused, so that the LM-GCI-LMB fusion effect is worse. The LMB algorithm tracking precision is related to the performance of the LMB algorithm, is not influenced by the increase of the number of the sensor network nodes, can only track the targets in the field of view, and cannot acquire the information of the targets outside the field of view.
As shown in fig. 8, in the present experimental scenario, the method of the present invention is still most accurate for estimating the target potential, and the LMB algorithm performs accurate potential estimation on the motion in the field of view. For the LM-GCI-LMB method, the target leakage heel occurs in the target non-overlapping area due to the adoption of the GCI fusion method, so that the potential estimation of each sensor is low.
Fig. 9 shows a single operation versus target state estimation diagram of the method of the present invention, as shown in fig. 9, and in this experimental scenario, the method of the present invention can still more completely extract the target track and the target state in the entire detection area of the sensor.
In summary, the invention firstly utilizes flooding transmission to enable a single sensor to have all target posterior distribution information in a sensor network, adopts a clustering-based target dividing method to accurately classify targets, can fuse class targets to execute GCI fusion, and can directly extract the target distribution information if the class targets are not fused, thereby avoiding losing the target distribution information of non-overlapping monitoring visual field areas. In addition, aiming at a moving target crossing a sensor or a single sensor tracking target, the invention utilizes the target label and the sensor identification information to establish a track maintenance table, and the target local label is recovered by matching with the track maintenance table, so that the target track is continuous. The method effectively solves the problem that the target tracking is missed due to the loss of target distribution information in a non-overlapping area caused by GCI fusion in the limited visual field sensor network, and can output the state information and complete tracks of all targets in the global visual field of the network.
Some steps in the embodiments of the present invention may be implemented by using software, and the corresponding software program may be stored in a readable storage medium, such as an optical disc or a hard disk.
The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.
Claims (9)
1. A method for distributed tag bernoulli fusion tracking in a limited field of view, the method comprising:
step one: deploying a sensor network and initializing sensor parameters;
the sensor network consists of N sensors, and the sensor set is recorded as S= { S i :i∈N}，s i The sensor is identified, and has uniqueness;representation and sensor s i A set of sensors having a distance not exceeding t hops and comprising a sensor s i ；/>Representation and sensor s i Sensor set with distance not exceeding t hops and containing no sensor s i ；
Step two: sensor s i And (3) operating an LMB filtering algorithm to obtain a multi-target posterior distribution LMB parameter set at the k moment:
wherein,a target tag space at time k, r (l) represents the existence probability of a target with a tag of l, and p (l) represents the probability density distribution of the target with the tag of l;
Step three: the sensor s i LMB parameter set of the multi-objective posterior distribution for the k-time instant Extracting target to obtain LMB parameter set of posterior distribution of extracted target->And diffusing the extraction target posterior distribution LMB parameter set in the network by a flooding transmission method>Meanwhile, obtaining LMB parameter sets of the posterior distribution of the extraction targets of other sensors;
step four: the sensor s i Obtaining LMB parameter set of posterior distribution of extraction targets of all other sensorsThen, adopting a clustering-based partitioning method to perform ∈N & lt/M & gt>Dividing targets in the system into a fusible target and an unfused target, wherein the fusible target is a target jointly tracked by a plurality of sensors, and the unfused target is a target independently tracked by a single sensor;
step five: after the target division is completed, GCI fusion is carried out on the fusible class targets to obtain a fusion posterior distribution LMB parameter set of the fusible class targetsAnd label correction is carried out in the fusion process, so that the aim of maintaining the track is fulfilled;
for the non-fusible targets, directly extracting the posterior distribution LMB parameter set of the targets, and carrying out label correction according to requirements, wherein the method comprises the following steps:
if there are only sensors s in the class i Directly extracting posterior distribution LMB parameter sets of the targets as part of fusion posterior distribution LMB parameter sets of non-fusion targets to obtain fusion posterior distribution LMB parameter sets of the targets
If there are only sensors in the classExtracting a posterior distribution LMB parameter set of such a target as part of a fused posterior distribution LMB parameter set of an unfused class of targets, but performing tag correction on tag information of such a target posterior distribution LMB parameter set so that such a target is located at the sensor s i The whole movement life cycle has the same label under the visual angle to obtain the target fusion posterior distribution LMB parameter set +.>
The fusible targets and the non-fusible targets maintain the track of the targets according to a track maintenance strategy, so that a single sensor outputs a complete track of the moving targets crossing the sensor;
step six: combining the fusion posterior distribution LMB parameter sets of the fusible target and the non-fusible target obtained in the step five into a target state extraction posterior distribution LMB parameter set, wherein the sensor s i Extracting posterior distribution LMB parameter sets according to the target states and outputting all target state information in a sensor network tracking area;
Step seven: utilizing posterior set in step fiveThe next moment filtering priori is formed, and all targets in the sensor network can be continuously tracked by repeating the second to sixth steps; wherein (1)>LMB parameter set representing fusion posterior distribution of fusion class object,/->Representing sensor s i Fusion posterior distribution LMB parameter sets for tracked non-fusible class targets.
2. The method according to claim 1, wherein the second step comprises:
assume that at time k-1, sensor s i Filtering to obtain a posterior LMB set which is as follows:
the target neogenesis model is selected from a self-adaptive neogenesis model, and a neogenesis target LMB parameter set is as follows:
wherein,representing a nascent target tag space;
at time k, an LMB prediction step is firstly performed in each sensor to obtain a predicted LMB parameter set Next, Z is measured at time k using a sensor k For said predicted LMB parameter set +.>Updating to obtain a filtered posterior distribution LMB parameter set +.>And simultaneously, a new target parameter set at the next moment is generated by measurement.
3. The method according to claim 2, wherein the step three comprises:
sensor s i Posterior distribution LMB parameter set obtained by LMB filteringComprising a sensor s i Real target and false alarm distribution information in visual field, LMB parameter set of posterior distribution +. >Performing potential estimation and target extraction to obtain potential estimation +.>Extracting target posterior distribution LMB parameter set +.>And false alarm posterior distribution LMB parameter set +.>
Obtaining potential estimatesExtracting a target posterior distribution LMB parameter set +.>After that, sensor s i Exchanging and extracting target posterior distribution LMB set information with other sensors in a sensor network through flooding transmission:
when T is less than or equal to T, the sensor s i The information diffusion is completed by iteratively exchanging the latest acquired posterior distribution LMB parameter set information with the adjacent interconnected sensors, so that the flooding transmission purpose is achieved; when T > T, indicating completion of flooding transmission, sensor s i Collecting target posterior sets extracted by all sensors in networkT represents the number of iterative exchanges of information.
4. A method according to claim 3, wherein said step four comprises:
first, a target posterior distribution LMB parameter set is extracted according to all sensorsEstimating the target state of each LMB component to obtain +.>As clustered data sample points, wherein l n Representing target tags, s j Representing sensor identification,/->Representing sensor s j A target tag space;
second, according to the sensor s i Target extraction potential estimationLet the number of the initial categories of the classification be->With sensors s i State estimation of +. >For initial cluster center->Wherein (1)>Representing sensor s i Estimating a target potential;
finally, calculating the estimation state of each targetTo the cluster center->According to Euclidean distance, completing target clustering division; the cluster partition objective satisfies two constraint conditions: the maximum number of sample points from the same sensor in each type is one, the distance between the sample points in each type and the center of the cluster is not more than a distance threshold eta, and the points which do not meet the two constraint conditions are added into a new clusterAnd (5) iterating the clustering until the constraint is met.
5. The method of claim 4, wherein maintaining the track of the target according to the track maintenance strategy in step five comprises:
when the target moves across the field of view of the sensor, the state of the category where the target is located changes, and when the target moving across the sensor is divided into an unfused category of target, the target only carries out state extraction, so that the target track is divided into two independent parts, and the track maintenance strategy is required to be executed by utilizing track maintenance table information to output a complete track;
the elements of each row in the track maintenance table areWherein->Representing sensor s i Extracting target tag-> A potential estimate is extracted for the target at time k, The representation is: is +.>Based on a clustering method, all target labels and sensor identification information in a target set divided into the same class>
If the target is a fusible target, the target is fused for the first timeThe information is added to the track maintenance table, and the follow-up tracking time can be fused with the category target information to be consistent with the track maintenance table information, so that correction is not required;
if from sensor s i The target is a local filtering tracking target, and the target label information does not need to be corrected when the label is unchanged in the local filtering process;
if from a sensorNon-fusible targets based on target labels and sensor identification informationSearching track maintenance table if there is +.>Information, the target tag is corrected to +.>The sensor identity is corrected to s i The method comprises the steps of carrying out a first treatment on the surface of the If there is no->Information representing the sensor s i Extracting the target information for the first time, initializing the target at the sensor s i The middle label is->Add->And the information is sent to a track maintenance table for label correction at the next moment.
6. The method of claim 5, wherein the step of extracting a posterior distribution LMB parameter set for six target states is:
wherein,LMB parameter set representing fusion posterior distribution of fusion class object,/- >Representing sensor s i Fusion posterior distribution LMB parameter set of tracked non-fusible class target, ++>Representing sensor s j Fusion posterior distribution LMB parameter sets for tracked non-fusible class targets.
7. The method of claim 1, wherein the state of the target comprises: the position of the target, the speed of the target, and the turn rate.
8. The method of claim 1, wherein the number N of sensors is 4.
9. The method of claim 1, wherein the sensor is a limited field of view sensor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210738739.7A CN115099343B (en) | 2022-06-24 | 2022-06-24 | Distributed tag Bernoulli fusion tracking method under limited field of view |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210738739.7A CN115099343B (en) | 2022-06-24 | 2022-06-24 | Distributed tag Bernoulli fusion tracking method under limited field of view |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115099343A CN115099343A (en) | 2022-09-23 |
| CN115099343B true CN115099343B (en) | 2024-03-29 |
Family
ID=83295036
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210738739.7A Active CN115099343B (en) | 2022-06-24 | 2022-06-24 | Distributed tag Bernoulli fusion tracking method under limited field of view |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115099343B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA2891836A1 (en) * | 2015-05-15 | 2016-11-15 | Sportlogiq Inc. | System and method for tracking moving objects in videos |
| CN108882271A (en) * | 2018-07-05 | 2018-11-23 | 电子科技大学 | It is a kind of to regard multiple sensor integrated method altogether based on more the non-of Bernoulli Jacob's distribution of label |
| CN112113572A (en) * | 2020-09-18 | 2020-12-22 | 桂林电子科技大学 | Multi-target tracking method for solving distributed label fusion |
| CN112305915A (en) * | 2020-10-28 | 2021-02-02 | 深圳大学 | Label multi-Bernoulli multi-target tracking method and system of TSK iterative regression model |
| WO2022116375A1 (en) * | 2020-12-01 | 2022-06-09 | 中国人民解放军海军航空大学 | Method for performing tracking-before-detecting on multiple weak targets by high-resolution sensor |
-
2022
- 2022-06-24 CN CN202210738739.7A patent/CN115099343B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA2891836A1 (en) * | 2015-05-15 | 2016-11-15 | Sportlogiq Inc. | System and method for tracking moving objects in videos |
| CN108882271A (en) * | 2018-07-05 | 2018-11-23 | 电子科技大学 | It is a kind of to regard multiple sensor integrated method altogether based on more the non-of Bernoulli Jacob's distribution of label |
| CN112113572A (en) * | 2020-09-18 | 2020-12-22 | 桂林电子科技大学 | Multi-target tracking method for solving distributed label fusion |
| CN112305915A (en) * | 2020-10-28 | 2021-02-02 | 深圳大学 | Label multi-Bernoulli multi-target tracking method and system of TSK iterative regression model |
| WO2022116375A1 (en) * | 2020-12-01 | 2022-06-09 | 中国人民解放军海军航空大学 | Method for performing tracking-before-detecting on multiple weak targets by high-resolution sensor |
Non-Patent Citations (3)
| Title |
|---|
| 吴莎 ; 杨小军 ; .基于Cauchy-Schwarz散度的多传感器控制.计算机技术与发展.2020,(第06期),全文. * |
| 徐悦 ; 杨金龙 ; 葛洪伟 ; .分布式传感器多目标跟踪改进算法.信号处理.2020,(第08期),全文. * |
| 曹倬 ; 冯新喜 ; 蒲磊 ; 王雪 ; 张琳琳 ; .基于标签随机有限集滤波器的多扩展目标跟踪算法.***工程与电子技术.2018,(第03期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115099343A (en) | 2022-09-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107677997B (en) | Extended target tracking method based on GLMB filtering and Gibbs sampling | |
| Casbeer et al. | Distributed information filtering using consensus filters | |
| CN104778358B (en) | The partly overlapping extension method for tracking target in monitored area be present in multisensor | |
| CN107423686B (en) | Video multi-target fuzzy data association method and device | |
| CN106355604A (en) | Target image tracking method and system | |
| CN104063880B (en) | PSO based multi-cell position outline synchronous accurate tracking system | |
| CN101944234A (en) | Multi-object tracking method and device driven by characteristic trace | |
| CN104133983B (en) | Multiple-target Bernoulli distribution type fusion method based on weighting negative first-order RD (Renyi Divergence) | |
| WO2018227491A1 (en) | Method and device for association of fuzzy data of multiple targets in video | |
| Cao et al. | Correlation-based tracking of multiple targets with hierarchical layered structure | |
| CN106447698B (en) | A kind of more pedestrian tracting methods and system based on range sensor | |
| CN109214432B (en) | Multi-sensor multi-target joint detection, tracking and classification method | |
| CN113296089B (en) | LMB density fusion method and device for multi-early-warning-machine target tracking system | |
| CN115099343B (en) | Distributed tag Bernoulli fusion tracking method under limited field of view | |
| CN112305915B (en) | Label multi-Bernoulli multi-target tracking method and system of TSK iterative regression model | |
| Li et al. | Index similarity assisted particle filter for early failure time prediction with applications to turbofan engines and compressors | |
| Kumar et al. | Crowd behavior recognition using hybrid tracking model and genetic algorithm enabled neural network | |
| CN112347208A (en) | Multi-target detection and tracking method based on distributed sensor network | |
| CN111811515B (en) | Multi-target track extraction method based on Gaussian mixture probability hypothesis density filter | |
| Farrell et al. | Learning higher-order transition models in medium-scale camera networks | |
| Ding et al. | Who is partner: A new perspective on data association of multi-object tracking | |
| Ren et al. | A generalized data association approach for cell tracking in high-density population | |
| Qian et al. | Multiple Object Tracking for Occluded Particles | |
| CN112837351B (en) | Improved label multiple Bernoulli distributed optimization fusion tracking method | |
| KR102589987B1 (en) | Method and Apparatus for Tracking of Online Multi-Object with Visual and Radar Features |
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 | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20240731 Address after: No. 168 Hexiao East Road, Ningbo City, Zhejiang Province, 315000 Patentee after: Ningbo New Quality Intelligent Manufacturing Technology Research Institute Country or region after: China Address before: No. 1800 road 214122 Jiangsu Lihu Binhu District City of Wuxi Province Patentee before: Jiangnan University Country or region before: China |