A distributed multi-target tracking filtering method, system and device

Abstract

The application discloses a kind of distributed multi-target tracking filtering method, system and equipment, it is related to target tracking field, this method includes: first, establish system model, initialize Gaussian component set;Then, for at each filtering iteration time, improved GM-PHD filter is used to predict, update, prune and merge;Then, for communication between each local fusion center, fusion estimation is carried out using distributed fusion rule on respective node;Then, for each local fusion center extracts target state;Finally, compared with current time and setting time, judge whether to carry out next filtering, until current time is greater than setting time, end filtering.The application can improve target tracking precision and the performance of distributed fusion rule.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of target tracking, and in particular to a distributed multi-target tracking filtering method, system and device. BACKGROUND

[0002] Multi-target tracking refers to jointly estimating the number and state of targets from a set of measurements containing noise, missed detections and clutter. Multi-target tracking algorithms based on the theory of random finite sets (RFS) have received unprecedented attention and development due to their ability to handle the birth / death and missing / false data of random targets. In addition, compared with the classical data association-based multi-target tracking algorithm, the RFS-based method avoids the explicit data association between targets and measurements, which involves a large amount of computational load.

[0003] In the multi-target tracking task in the modern complex environment, the sensor network can improve the target discovery probability and the target state perception accuracy on the premise of low cost, and provide more accurate target position and number related information, etc. Therefore, the collaborative target tracking using the sensor network has become a research hotspot. According to the network structure, the collaborative multi-target tracking can be divided into centralized and distributed: the centralized has a center node that can collect all sensor data for filtering and fusion processing, which is beneficial to obtain the globally optimal target tracking result, but has the disadvantages of strong network structure dependence and system paralysis caused by the attack on the center node; the distributed does not depend on the center node, and has the characteristics of strong robustness and scalability, so it has been widely applied.

[0004] In the past decade, although various collaborative multi-target tracking filtering methods have been well studied, there are still some unsolved problems in this research field. For example, the commonly used RFS filters mainly include the probability hypothesis density (PHD) filter and the cardinalized probability hypothesis density (CPHD) filter, but in large-scale scenarios, the RFS filter not only encounters estimation bias, but also faces serious computational burden due to a large number of observations and computational complexity, and the target tracking accuracy is low. For the distributed fusion rule, the two mainstream fusion rules are the generalized covariance intersection (GCI) and the arithmetic average (AA) fusion rules, but the two have complementarity, and how to better improve their performance is still a technical problem to be broken through. SUMMARY

[0005] The application aims to provide a distributed multi-target tracking filtering method, system and device to solve the problems of low target tracking accuracy and poor performance of distributed fusion rules.

[0006] To achieve the above object, the application provides the following solutions.

[0007] A distributed multi-target tracking filtering method comprises the following steps.

[0008] A single-target state is constructed, and an observation model is constructed according to the single-target state when a target is detected by a sensor to generate an observation value of any sensor;

[0009] A measurement set of a local fusion center at any time is determined according to the observation value of any sensor;

[0010] A Gaussian component set is initialized, a target posterior intensity at a current time is determined, and a target posterior intensity at a previous time is obtained based on the target posterior intensity at the current time;

[0011] A probability hypothesis density of a surviving target is predicted according to the target posterior intensity at the previous time, a surviving component is determined, and an observation subset originating from a new target and clutter is generated by dividing the measurement set according to the surviving component;

[0012] A probability hypothesis density of the new target in the observation subset is predicted, and a predicted target prior intensity is determined according to the probability hypothesis density of the new target;

[0013] A probability hypothesis density of the clutter in the observation subset is estimated, and an updated target posterior intensity is determined by updating the target posterior intensity at the current time according to the probability hypothesis density of the clutter and the predicted target prior intensity;

[0014] The initialized Gaussian component set is pruned and merged to generate a pruned and merged Gaussian component set, and a twice-updated target posterior intensity is determined by updating the updated target posterior intensity according to the pruned and merged Gaussian component set; the twice-updated target posterior intensity is assigned as a probability hypothesis density of the local fusion center;

[0015] The probability hypothesis density of the local fusion center and the probability hypothesis density of a neighbor node are obtained by communicating with the neighbor node;

[0016] The probability hypothesis density of a common target is fused by using a GCI rule based on the probability hypothesis density of the local fusion center and the probability hypothesis density of the neighbor node to generate a fused common target probability hypothesis density, and the probability hypothesis density of an additional target is fused by using an AA rule to generate a fused additional target probability hypothesis density;

[0017] The target posterior strength after the fusion is determined based on the common target probability hypothesis density and the additional target probability hypothesis density after the fusion, and the target posterior strength after the second update is updated based on the target posterior strength after the current fusion, and the target posterior strength after the third update is determined.

[0018] Extract the target state set based on the target posterior strength after the three updates;

[0019] When the current time is less than the set time, return to the step "predict the probability hypothesis density of the surviving target based on the target posterior strength of the previous time, determine the surviving component, divide the measurement set according to the surviving component, and generate an observation subset originating from the new target and clutter", and perform the filtering iteration for the next time until the current time is greater than the set time, and end the filtering iteration.

[0020] Optionally, the target posterior strength at the current moment is:

[0021]

[0022] Among them, v k (x) represents the posterior intensity of the target at the current time k. For having a mean Covariance The Gaussian density, x is the variable represented by the Gaussian distribution, i is the index value of the Gaussian component, and N k The number of Gaussian components. The weight of the component.

[0023] Optionally, predicting the probability hypothesis density of newly formed targets in the observed subset, and determining the prior strength of the predicted targets based on the probability hypothesis density of the newly formed targets, specifically includes:

[0024] Using formula Predict the probability hypothesis density of newly formed targets in the observed subset; where γ k (x) represents the probability hypothesis density of the new target, N b,k For the subset of observations originating from newly formed targets and clutter, For γ k The weight of the i-th Gaussian component in (x), For γ k (x)

[0025] The mean of the i-th Gaussian component. For γ k The covariance of the i-th Gaussian component in (x);

[0026] use Determine the prior strength of the predicted target; where v k|k-1(x) is the predicted target prior intensity, v p,k|k-1 (x) is the predicted target prior intensity, v k|k-1 is the number of prior Gaussian components, is v k|k-1 is the weight of the i-th Gaussian component in (x), is v k|k-1 is the mean of the i-th Gaussian component in (x), is v k|k-1 is the covariance of the i-th Gaussian component in (x).

[0027] Optionally, the updated target posterior intensity is:

[0028] where v k (x)' is the updated target posterior intensity, p D is the detection probability, Z k,q is the observation set from node q, z is any one of the observation vectors in Z k，q , is v k is the weight of the i-th Gaussian component in (x)', is v k is the mean of the i-th Gaussian component in (x)', is v k is the covariance of the i-th Gaussian component in (x)'.

[0029] Optionally, the initialized Gaussian component set is pruned and merged to generate a pruned and merged Gaussian component set, and the updated target posterior intensity is updated according to the pruned and merged Gaussian component set to determine a second updated target posterior intensity, and specifically comprising:

[0030] A pruning threshold, a merging threshold, and a Gaussian component number threshold are set;

[0031] The initialized Gaussian component set is pruned according to the pruning threshold to generate an index set of pruned Gaussian components, and the index of the component corresponding to the current maximum weight is found;

[0032] According to the merging threshold, the index set of pruned Gaussian components, and the index of the component corresponding to the current maximum weight, an index sub-set of components similar to the component corresponding to the maximum weight is found;

[0033] Based on the index sub-set of components similar to the component corresponding to the maximum weight, Gaussian components are merged to generate a pruned and merged Gaussian component set;

[0034] update the updated target posterior intensity according to the pruned and merged Gaussian component set based on the Gaussian component number threshold, to determine a second updated target posterior intensity.

[0035] Optionally, based on the probability hypothesis density of the local fusion center and the probability hypothesis density of the neighbor node, the probability hypothesis density representing the common target is fused by using GCI rule to generate a fused common target probability hypothesis density, and the probability hypothesis density representing the extra target is fused by using AA rule to generate a fused extra target probability hypothesis density, specifically including:

[0036] Based on the probability hypothesis density of the local fusion center, a local fusion weight of the local fusion center is determined, a local Gaussian component set and a local Gaussian component index of the local fusion center are defined;

[0037] Based on the probability hypothesis density of the neighbor node, a neighbor fusion weight of the neighbor node is determined, a neighbor Gaussian component set and a neighbor Gaussian component index of the neighbor node are defined;

[0038] A local decomposition threshold and a neighbor decomposition threshold are set;

[0039] The local fusion weight and the neighbor fusion weight are normalized to generate a normalized local fusion weight and a normalized neighbor fusion weight;

[0040] According to the local decomposition threshold, the local Gaussian component in the local Gaussian component set with the local Gaussian component index is decomposed to generate a component set of a local common target and a component set of a local extra target;

[0041] According to the neighbor decomposition threshold, the neighbor Gaussian component in the neighbor Gaussian component set with the neighbor Gaussian component index is decomposed to generate a component set of a neighbor common target and a component set of a neighbor extra target;

[0042] According to the normalized local fusion weight, the normalized neighbor fusion weight, the component set of the local common target and the component set of the neighbor common target, the probability hypothesis density representing the common target is fused by using GCI rule to generate a fused common target probability hypothesis density;

[0043] According to the component set of the local extra target and the component set of the extra common target, the probability hypothesis density representing the extra target is fused by using AA rule to generate a fused extra target probability hypothesis density.

[0044] Optionally, the target state set is:

[0045]

[0046] wherein, is a target state set.

[0047] A distributed multi-target tracking filtering system comprises:

[0048] an observation value generation module configured to construct a single-target state, and construct an observation model according to the single-target state to generate an observation value of any sensor when a target is detected by the sensor;

[0049] an observation machine determination module configured to determine a measurement set of a local fusion center at any time according to the observation value of any sensor;

[0050] a target posterior intensity determination module of a previous time point configured to initialize a Gaussian component set, determine a target posterior intensity at a current time point, and obtain a target posterior intensity at a previous time point based on the target posterior intensity at the current time point;

[0051] an observation subset generation module of a new target and clutter configured to predict a probability hypothesis density of a survival target according to the target posterior intensity at the previous time point, determine a survival component, and divide the measurement set according to the survival component to generate an observation subset from the new target and the clutter;

[0052] a predicted target prior intensity determination module configured to predict a probability hypothesis density of a new target in the observation subset, and determine a predicted target prior intensity according to the probability hypothesis density of the new target;

[0053] an updated target posterior intensity determination module configured to estimate a probability hypothesis density of a clutter in the observation subset, and update the target posterior intensity at the current time point according to the probability hypothesis density of the clutter and the predicted target prior intensity to determine an updated target posterior intensity;

[0054] a twice-updated target posterior intensity determination module configured to cut and merge the initialized Gaussian component set to generate a cut-and-merged Gaussian component set, and update the updated target posterior intensity according to the cut-and-merged Gaussian component set to determine a twice-updated target posterior intensity; the twice-updated target posterior intensity is assigned as a probability hypothesis density of the local fusion center;

[0055] a probability hypothesis density acquisition module of a neighbor node configured to communicate with a neighbor node to acquire a probability hypothesis density of the neighbor node;

[0056] The fused common target probability hypothesis density generation module is configured to fuse the probability hypothesis densities representing the common targets based on the probability hypothesis densities of the local fusion center and the probability hypothesis densities of the neighbor nodes according to a GCI rule to generate a fused common target probability hypothesis density, and fuse the probability hypothesis densities representing the extra targets according to an AA rule to generate a fused extra target probability hypothesis density;

[0057] The third updated target posterior intensity determination module is configured to determine a current fused target posterior intensity according to the fused common target probability hypothesis density and the extra target probability hypothesis density, and update the second updated target posterior intensity according to the current fused target posterior intensity to determine a third updated target posterior intensity;

[0058] The target state set extraction module is configured to extract a target state set based on the third updated target posterior intensity.

[0059] The filtering iteration module is configured to return to the step of predicting the probability hypothesis density of the surviving target based on the target posterior intensity of the last time, determining the surviving component, and dividing the measurement set according to the surviving component to generate the observation subset from the new target and the clutter when the current time is less than a set time, and perform filtering iteration of the next time until the current time is greater than the set time, and end the filtering iteration.

[0060] An electronic device includes a memory for storing a computer program and a processor for running the computer program to make the electronic device perform the distributed multi-target tracking filtering method.

[0061] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the distributed multi-target tracking filtering method.

[0062] According to the embodiments of the present application, the following technical effects are achieved: the present application provides a distributed multi-target tracking filtering method, system and device, the observation subset from the new target and the clutter is generated based on the divided measurement set, the uncertainty of the new target is better presented, the calculation cost is reduced, and in the filtering update process, the probability hypothesis density of the clutter related to the target is introduced to correct the overestimation of the target base, and the target tracking precision is improved; in addition, each local fusion center only needs to fuse the information of the neighbors, the fusion step is distributed, the fusion mechanism of the AA rule compensating for the GCI rule is adopted, the fusion performance is improved, and the target tracking precision is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed in the embodiments will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and for those skilled in the field, other drawings can be obtained based on these drawings without any creative effort.

[0064] Figure 1 A distributed multi-target tracking filtering method flow chart provided by the first embodiment of the present application;

[0065] Figure 2 A design framework chart of the distributed multi-target tracking filtering method provided by the second embodiment of the present application;

[0066] Figure 3 A multi-target motion trajectory chart provided by the third embodiment of the present application;

[0067] Figure 4 A sensor network deployment structure chart provided by the third embodiment of the present application;

[0068] Figure 5 An OSPA error comparison chart before fusion provided by the third embodiment of the present application;

[0069] Figure 6 A cardinality estimation comparison chart before fusion provided by the third embodiment of the present application;

[0070] Figure 7 A calculation cost comparison chart before fusion provided by the third embodiment of the present application;

[0071] Figure 8 An OSPA error comparison chart under large-scale sensor network provided by the third embodiment of the present application;

[0072] Figure 9 A cardinality estimation comparison chart under large-scale sensor network provided by the third embodiment of the present application. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be described clearly and completely below with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the field without any creative effort belong to the protection scope of the present application.

[0074] The present application aims to provide a distributed multi-target tracking filtering method, system and device, which can improve the target tracking precision and the performance of distributed fusion rules.

[0075] In order to make the above objectives, characteristics and advantages of the present application more obvious and comprehensible, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0076] Embodiment one

[0077] Figure 1 A flow chart of the distributed multi-target tracking filtering method provided by the embodiment one of the present application is shown in Figure 1 The present application provides a distributed multi-target tracking filtering method, which comprises the following steps:

[0078] Step 101: constructing a single-target state, constructing an observation model according to the single-target state when a target is detected by a sensor, and generating an observation value of any sensor.

[0079] Step 102: determining a measurement set of a local fusion center at any time according to the observation value of any sensor.

[0080] Step 103: initializing a Gaussian component set, determining a target posterior intensity at a current time, and obtaining a target posterior intensity at a previous time based on the target posterior intensity at the current time.

[0081] In actual application, the target posterior intensity at the current time is:

[0082]

[0083] Wherein, v k (x) is the target posterior intensity at the current time k, is a Gaussian density with a mean and a covariance x is a variable represented by the Gaussian distribution, i is an index value of the Gaussian component, N k is the number of Gaussian components, is the weight of the component.

[0084] Step 104: predicting a probability hypothesis density of a surviving target according to the target posterior intensity at the previous time, determining a surviving component, dividing the measurement set according to the surviving component, and generating an observation subset derived from a new target and clutter.

[0085] Step 105: predicting a probability hypothesis density of a new target in the observation subset, and determining a predicted target prior intensity according to the probability hypothesis density of the new target.

[0086] In actual application, the step 105 specifically comprises: predicting the probability hypothesis density of the new target in the observation subset by using the formula ; wherein, γ k (x) is the probability hypothesis density of the new target, Nb,k for the observation subset originating from the newborn target and clutter, for the gamma k weight of the i-th Gaussian component in (x), for the gamma k mean of the i-th Gaussian component in (x), for the gamma k covariance of the i-th Gaussian component in (x); the predicted target posterior intensity is determined by where v k|k-1 (x) is the predicted target posterior intensity, v p,k|k-1 (x) is the predicted target prior intensity, N k|k-1 is the number of prior Gaussian components, is the weight of the i-th Gaussian component in v k|k-1 (x), is the mean of the i-th Gaussian component in v k|k-1 (x), is the covariance of the i-th Gaussian component in v k|k-1 (x).

[0087] Step 106: estimate the probability hypothesis density of the clutter in the observation subset, and update the target posterior intensity at the current time according to the probability hypothesis density of the clutter and the predicted target prior intensity, to determine an updated target posterior intensity.

[0088] In practical applications, the updated target posterior intensity is: where v k (x)' is the updated target posterior intensity, p D is the detection probability, Z k,q is the observation set from node q, and z is any observation vector in Z k,q . is the weight of the i-th Gaussian component in v k (x)', is the mean of the i-th Gaussian component in v k (x)', is the covariance of the i-th Gaussian component in v k (x)'.

[0089] Step 107: trim and merge the initialized Gaussian component set to generate a trimmed and merged Gaussian component set, and update the updated target posterior intensity according to the trimmed and merged Gaussian component set to determine a twice-updated target posterior intensity; the twice-updated target posterior intensity is assigned as the probability hypothesis density of the local fusion center.

[0090] In practical application, the step 107 specifically includes: setting a clipping threshold, a merging threshold and a Gaussian component number threshold; clipping the initialized Gaussian component set according to the clipping threshold to generate an index set of the clipped Gaussian components and find an index of a component corresponding to a current maximum weight value; finding an index sub-set of components similar to the component corresponding to the maximum weight value according to the merging threshold, the index set of the clipped Gaussian components and the index of the component corresponding to the maximum weight value; merging Gaussian components based on the index sub-set of the components similar to the component corresponding to the maximum weight value to generate a set of clipped and merged Gaussian components; and updating the updated target posterior intensity according to the set of clipped and merged Gaussian components based on the Gaussian component number threshold to determine a twice-updated target posterior intensity.

[0091] Step 108: communicating with a neighbor node to obtain a probability hypothesis density of the neighbor node.

[0092] Step 109: based on the probability hypothesis density of the local fusion center and the probability hypothesis density of the neighbor node, fusing probability hypothesis densities representing common targets by using a GCI rule to generate a fused common target probability hypothesis density and fusing probability hypothesis densities representing additional targets by using an AA rule to generate a fused additional target probability hypothesis density.

[0093] In practical application, the step 109 specifically includes: determining the local fusion weight of the local fusion center based on the probability hypothesis density of the local fusion center, defining the local Gaussian component set and the local Gaussian component index of the local fusion center; determining the neighbor fusion weight of the neighbor node based on the probability hypothesis density of the neighbor node, defining the neighbor Gaussian component set and the neighbor Gaussian component index of the neighbor node; setting the local decomposition threshold and the neighbor decomposition threshold; normalizing the local fusion weight and the neighbor fusion weight to generate the normalized local fusion weight and the normalized neighbor fusion weight; decomposing the local Gaussian component with the local Gaussian component index in the local Gaussian component set according to the local decomposition threshold to generate the component set of the local common target and the component set of the local additional target; decomposing the neighbor Gaussian component with the neighbor Gaussian component index in the neighbor Gaussian component set according to the neighbor decomposition threshold to generate the component set of the neighbor common target and the component set of the neighbor additional target; adopting the GCI rule to fuse the probability hypothesis density representing the common target based on the normalized local fusion weight, the normalized neighbor fusion weight, the component set of the local common target and the component set of the neighbor common target to generate the fused common target probability hypothesis density; adopting the AA rule to fuse the probability hypothesis density representing the additional target based on the component set of the local additional target and the component set of the additional common target to generate the fused additional target probability hypothesis density.

[0094] Step 110: determining the current fused target posterior intensity according to the fused common target probability hypothesis density and the additional target probability hypothesis density, and updating the twice updated target posterior intensity according to the current fused target posterior intensity to determine the thrice updated target posterior intensity.

[0095] Step 111: extracting the target state set based on the thrice updated target posterior intensity.

[0096] In practical application, the target state set is:

[0097]

[0098] wherein, is the target state set.

[0099] Step 112: when the current time is less than the set time, returning to step 104 for the filtering iteration of the next time until the current time is greater than the set time, and ending the filtering iteration.

[0100] Embodiment two

[0101] The distributed multi-target tracking filtering method provided by the application has eight steps in the second embodiment, wherein step 1 is used for establishing a system model; step 2 is used for initialization; steps 3-5 are used for predicting, updating, pruning and merging by using the improved GM-PHD filter at each filtering iteration time; step 6 is used for communication between each local fusion center, and fusion estimation is performed on each node by using a distributed fusion rule; step 7 is used for extracting the target state by each local fusion center; and step 8 is used for judging whether to return to step 3 for the next filtering.

[0102] Taking a local fusion center q, for example, Figure 2 the design framework diagram of the distributed multi-target tracking filtering method provided by the second embodiment of the application is as shown in Figure 2 , and the specific design steps are as follows:

[0103] Step 1: Establishing a system model.

[0104] 1) Considering the state model of the target. The state of a single target is denoted as wherein and are the position and velocity in the x direction and the y direction at time k, respectively.

[0105] The linear time-invariant state model of a single target is:

[0106] x k =F k|k-1 x k-1 +w k-1 (1)

[0107] wherein F k|k-1 is a state transition matrix, and w k-1 is zero-mean Gaussian white noise with variance Q k-1 .

[0108] The state of multiple targets at time k is described by an RFS:

[0109]

[0110] wherein denotes the state of the i-th target, and takes a value in the state space , wherein n x is the dimension of the target state. |·| represents the cardinality. denotes a finite set. In each time step, the uncertainty of target appearance, disappearance, derivation and survival in the environment leads to the random change of the number and state of the targets.

[0111] 2) Consider the sensor network topology model consisting of isomorphic sensor nodes and local fusion centers. From a mathematical point of view, the sensor network can be described by a topological graph G = {N, C, A}, where N and C are the sets of sensors and local fusion centers, respectively, is the set of edges, representing the connections between local fusion centers.

[0112] For a node q, if (a q ,a η ) e A, where (a q ,a η ) denotes an edge from node q to node η, then node η is a neighbor of node q, and w η > 0, where w η denotes the weight of edge (a q ,a η ). Let C q denote the set of neighbor fusion centers connected to q, and satisfy Let N q denote the set of sensors belonging to q, and each sensor belongs to only one local fusion center, thus,

[0113] 3) Consider the sensor observation model. When a target is detected by a sensor, the measurement is generated according to the following observation model:

[0114] z k = H k x k + v k (3)

[0115] where H k is the observation matrix, is a zero-mean Gaussian white noise with variance R k .

[0116] The measurement set of local fusion center q at time k is described by a RFS

[0117]

[0118] where, denotes an observation from a sensor, taking values in the observation space , where n z is the dimension of the observation. Considering the effects of missed detection and clutter, some targets can not be detected, and the measurements Z k,q may contain false alarms or clutter. Note that each measurement is not labeled to indicate its source, i.e., the correlation between sensors and observations and the correlation between targets / clutter and observations are unknown.

[0119] Step 2: Initialization.

[0120] 1) Set the simulation time T, set the current simulation sampling time k = 0.

[0121] 2) Initialize the Gaussian component set The index set of the Gaussian component is described as J k = {1, 2,..., N k}, and the target posterior intensity is described as:

[0122]

[0123] Wherein, represents the Gaussian density with mean and covariance N k represents the number of Gaussian components, represents the weight of the component.

[0124] Step 3: Prediction process.

[0125] 1) k = k + 1.

[0126] 2) Obtain the target posterior intensity at the last time:

[0127]

[0128] 3) Obtain the measurement set Z q from the sensor set N k,q .

[0129] 4) Predict the PHD of the surviving target:

[0130]

[0131] Wherein, p S represents the survival probability.

[0132]

[0133]

[0134] 5) Pre-divide the measurement set.

[0135] Find the optimal matching binary matrix Y * between the surviving component and the observation:

[0136]

[0137]

[0138] where tr(·) denotes the sum of the traces of matrices, Y is the matching binary matrix between surviving components and observations, v k-1 (x) is the target posterior intensity at the last time, N k-1 is the number of Gaussian components of v k-1 (x), Y n,m is the element of matrix Y at the nth row and mth column, representing the matching between the nth surviving component and the mth observation. The distance matrix D is designed as follows:

[0139]

[0140] where d n,m represents the distance between the nth surviving component and the mth observation, which is defined as follows:

[0141]

[0142]

[0143]

[0144] where, is the innovation of the nth surviving component under the mth observation, S n is the normalization matrix of the nth surviving component, is the mth observation, H k is the observation matrix, R k is the covariance matrix of Gaussian observation noise, is the mean of the predicted prior intensity of the nth surviving component, is the covariance of the predicted prior intensity of the nth surviving component.

[0145] Further, for the associated pair corresponding to and distance d n,m , is the element of matrix Y * at the nth row and mth column, under a given distance threshold T r , if d n,m <T r , it is considered that the association is valid, otherwise the associated pair is discarded, i.e. set

[0146] Then, the observation subset related to the surviving target is expressed by the observations in all remaining associated pairs, which is extracted from the measurement set Z k,q , and the observation subset originating from the newborn target and clutter can be generated, where X p,k is the random finite set corresponding to the surviving target, x is Xp,k any one of the target states.

[0147] 6) Predict PHD of new-born targets k (x):

[0148]

[0149] where,

[0150]

[0151]

[0152]

[0153]

[0154] where, N b,k PHD of new-born targets k (x) is the number of Gaussian components, is the weight of the i-th Gaussian component in k (x), is the covariance of the i-th Gaussian component in k (x), is is the mean of the i-th Gaussian component in k (x), k,q is the set of observations from node q, z is k,q any one of the observation vectors.

[0155] 7) Obtain predicted target intensity v k|k-1 (x):

[0156]

[0157] where, N k|k-1 = N k-1 + N b,k denotes the number of prior Gaussian components, v p,k|k-1 (x) is the prior intensity of the predicted surviving targets, is the weight of the i-th Gaussian component in v k|k-1 (x), is the mean of the i-th Gaussian component in v k|k-1 (x), is the covariance of the i-th Gaussian component in v k|k-1 (x).

[0158] Step 4: Update

[0159] 1) Estimate target-dependent clutter PHD

[0160]

[0161] where p D denotes the detection probability, N q is the set of sensor nodes belonging to node q.

[0162]

[0163] where, is the Gaussian density corresponding to the Bernoulli component of the i-th component of the predicted target intensity and the observation z.

[0164] 2) Update the posterior PHD v k (x)' of the target.

[0165]

[0166] where,

[0167]

[0168]

[0169]

[0170]

[0171] where I denotes the identity matrix, κ k (z) denotes the clutter intensity, p D is the detection probability, Z k,q is the set of observations from node q, and z is any one of the observation vectors in Z k,q is the i-th Gaussian component in v k (x)' with weight, is the i-th Gaussian component in v k (x)' with mean, is the i-th Gaussian component in v k (x)' with covariance, is the i-th Gaussian component in v k (x)' with filtering gain.

[0172] Similarly, the target posterior intensity can also be derived in the Gaussian mixture form

[0173]

[0174] where N k = N k|k-1 × (1 + |Z k,q ​represents the number of posterior Gaussian components.

[0175] Step 5: pruning and merging.

[0176] 1) Given pruning threshold T P , merging threshold T U , Gaussian component number threshold N max , set iteration number l, l = 1.

[0177] 2) Define the index set of pruned Gaussian components as where, is the weight of the i-th Gaussian component in the target posterior intensity v k (x), J k = {1, 2,..., N k} is the Gaussian component index set of the target posterior intensity v k (x), N k is the number of Gaussian components of the target posterior intensity v k (x).

[0178] 3) Find the index id of the component with the current maximum weight:

[0179]

[0180] 4) Find the index subset of components similar to the component with the maximum weight:

[0181]

[0182] 5) Merge Gaussian components:

[0183]

[0184]

[0185]

[0186] 6) Delete the merged Gaussian component index subset I = I \ L. If , go to the next step; otherwise, let l = l + 1 and return to 3);

[0187] 7) If l ≤ N max , go to the next step; otherwise, replace max with the N

[0188] 8) Update the target posterior intensity v k (x) with the pruned and merged Gaussian component set.

[0189] Step 6: Distributed fusion

[0190] 1) For the convenience of representation of the calculation process, the subscript of the representative node is no longer omitted, i.e. let the PHD of node q be denoted as PHDv k (x) and the set of Gaussian components as k,q (x). Each Gaussian component has an index i, The fusion weight is defined as

[0191] 2) Given a decomposition threshold T c ;

[0192] 3) Select a neighbor node η, Get the PHD of the neighbor node PHDv k,η (x), the fusion weight w η , and define the set of Gaussian components as Each Gaussian component has an index j,

[0193] 4) Normalize the fusion weight:

[0194]

[0195]

[0196] 5) Decompose the PHDs.

[0197] For the Gaussian component with index i∈J k,q , decompose it according to the following rules:

[0198]

[0199] where and represent the sets of components representing the common target and the extra target, respectively,

[0200]

[0201] Similarly, for the Gaussian component with index j∈J k,η , decompose it according to the following rules:

[0202]

[0203] where and represent the sets of components representing the common target and the extra target, respectively,

[0204]

[0205] 6) Fuse the PHDs representing the common target using the GCI rule:

[0206]

[0207] where,

[0208]

[0209]

[0210]

[0211] 7) Fuse PHDs representing extra targets using AA rule:

[0212]

[0213] where, 0≤Δ1,Δ2≤1 are adjustable parameters.

[0214] 8) Obtain the current fused target intensity:

[0215]

[0216] 9) Use step 5 to crop and merge Gaussian components.

[0217] 10) Delete the fused neighbor node C q = C q \η, let If then go to next step; otherwise, return to 3).

[0218] 11) Use the fused target intensity to update the target posterior intensity v k,q (x) and update the corresponding Gaussian component set Step 7: Target state extraction.

[0219] For convenience, the subscripts of the representative nodes are still omitted in this step as in steps 1-5, i.e., let the fused target intensity v k,q (x) of a node q be written as v k (x), and the corresponding Gaussian component set be written as Each Gaussian component has an index i,

[0220] 1) Set the extracted state vector set Set the extraction threshold w th .

[0221] 2) Arbitrarily select a Gaussian component with index i, .

[0222] 3) If then go to next step; otherwise, update the state vector set:

[0223]

[0224] where repmat(A,m,n) means to form an m x n block matrix with each element replicated from the matrix A.

[0225] 4) Delete the index J = J \ i from the index set. If then go to next step; otherwise, return to step 3).

[0226] 5) Output the extracted state set

[0227] Step 8: Determine whether the filtering iteration continues.

[0228] If k < T, return to step 3 to enter the next filtering iteration; otherwise, end the filtering iteration.

[0229] Example Three

[0230] The performance of the distributed multi-target tracking filtering method provided by the application under a large-scale network is evaluated from the following two points: first, the I-AB-GM-PHD filter provided by the application is compared with the traditional AB-GM-PHD filter; second, the (GCI-AA-COM fusion rule provided by the application is compared with the GCI fusion, AA fusion, and two improved GCI and AA fusion rules, i.e., the AA(AA-CC) fusion rule based on the cardinality consensus and the CA-GCI fusion rule based on clustering.

[0231] All simulation experiments are run on an Intel(R) Core(TM) i7-10700 CPU@2.90GHz computer using MATLAB 2020a. The performance evaluation results in terms of optimal sub-pattern (OSPA) distance, cardinality estimation, and computational cost are the average values of 200 independent Monte Carlo runs, and the simulation running time step is set to T = 100s, and the sampling interval is T s = 1s. At the same time, the OSPA distance parameters are set to c = 30m, p = 2, and the "tic" and "toc" functions are used in MATLAB for computational cost measurement.

[0232] Consider the target tracking problem of unknown and time-varying target number in a clutter environment, and multiple targets move in a two-dimensional monitoring area [-1000, 1000] x [-1000, 1000] m 2 The state transition matrix and process noise variance are set as:

[0233]

[0234] where I2denotes a 2x2 identity matrix, σ w = 0.1 m / s 2 denotes the standard deviation of process noise.

[0235] The observation matrix and the observation noise variance are set as:

[0236]

[0237] where σ ε = 10 m denotes the standard deviation of observation noise.

[0238] Figure 3 The motion trajectories of 10 targets in the surveillance region are described. The targets are named as T1~T10 in sequence, which enter and exit the region at different times and locations. The target name and the survival time period are marked at the starting point of each target. Table 1 is a target initial setting table, and other initial settings are shown in Table 1.

[0239] Table 1

[0240] Objectives Initial position (m) Initial velocity (m / s) T1 [-900,-900] [14,14] T2 [-700,100] [4,-10] T3 [200,-800] [-3.5,10] T4 [-100,700] [-1,-16] T5 [500,900] [-7,-2] T6 [-300,350] [15,2] T7 [-300,-400] [9,1.5] T8 [50,800] [2,-3] T9 [800,-750] [-6,14] T10 [500,500] [-4,-13]

[0241] Figure 4 The deployment structure of the sensor network is described. |C| = 8 local fusion centers (LFCs) are marked as LFC1~LFC8 in sequence, and |N| = 200 sensor nodes (SNs) are marked with the same color as the LFC to which they belong. The black solid lines between the LFCs represent the undirected topological graph of the LFCs.

[0242] In each local filter, the number of initialized Gaussian components N0= 4 is set, and for the Gaussian component with index i∈{1,2,...,N0}: the weight is set as where randn denotes a random number subject to a standard normal distribution; the covariance is set as where diag([a,b,c,d]) denotes a diagonal matrix with a, b, c, and d as diagonal elements; the mean value is set as where [x pos ,y pos ] = meshgrid([-500,500]), x pos denotes a 2x2 square grid matrix with [-500,500] for each row, and y pos denotes a 2x2 square grid matrix with [-500,500] T for each column. The survival probability p S = 0.99, the merging threshold T U = 4, the clipping threshold T P = 10 -5 , and the Gaussian component number threshold Nmax = 50, state extraction threshold w th = 0.5, pre-partition threshold T r = 10. For fusion rule, decomposition threshold T c = 10, confidence factor Δ1 = Δ2 = 0.7.

[0243] First, in order to verify the effectiveness of the PHD filter algorithm of the present application, the detection probability is set to p D = 1. The performance comparison results before fusion in a specific LFC (LFC1) are shown in Figures 5-7 .

[0244] From Figure 5 and Figure 6 , it can be seen that the filter provided by the present application can effectively solve the problem of cardinality overestimation. In fact, the traditional AB-GM-PHD filter overestimates the cardinality during the entire simulation time, but detects the newly born target in time at the time of 10s, 20s, 30s and 40s. In contrast, the I-AB-GM-PHD filter is almost close to the true cardinality, only with a slight delay in detecting the newly born target. As shown in Figure 7 , the I-AB-GM-PHD filter keeps the average calculation time of each recursion within 0.1s, and the performance is better than that of the AB-GM-PHD filter.

[0245] Then, in order to evaluate the overall algorithm proposed by the present application, the detection probability is set to p D = 0.95, and the performance of the fusion algorithm based on the I-AB-GM-PHD (hereinafter referred to as I-AB) filter is tested. The performance comparison results under large-scale sensor networks are shown in Figure 8 , Figure 9 .

[0246] From Figure 8 , it can be seen that in the interval [0, 40]s, the OSPA error of I-AB-GCI and I-AB-GCI-AA-COM can improve that of the single LFC (I-AB), and the OSPA error of I-AB-GCI-AA-COM is the smallest. At the same time, the OSPA errors of I-AB-AA and I-AB-AA-CC are larger than those before fusion. In the interval [60, 100]s, they all improve the OSPA error of the single LFC. At the same time, it can be seen from Figure 9 that I-AB-GCI-AA-COM is closest to the true cardinality in the interval [0, 40]s, and slightly underestimates in the interval [60, 100]s, but it is not obvious.

[0247] Example Four

[0248] In order to implement the method corresponding to the above-mentioned embodiment one, to realize the corresponding functions and technical effects, the following provides a distributed multi-target tracking filtering system, comprising:

[0249] An observation value generation module is configured to construct a single-target state, and construct an observation model according to the single-target state when a target is detected by a sensor, and generate an observation value of any sensor.

[0250] An observation machine determination module is configured to determine a measurement set of the local fusion center at any time according to the observation value of any sensor.

[0251] A target posterior intensity determination module of a previous time is configured to initialize a Gaussian component set, determine a target posterior intensity at a current time, and obtain a target posterior intensity of a previous time based on the target posterior intensity at the current time.

[0252] A new target and clutter observation subset generation module is configured to predict a probability hypothesis density of a surviving target according to the target posterior intensity of the previous time, determine a surviving component, and divide the measurement set according to the surviving component to generate an observation subset from a new target and clutter.

[0253] A predicted target prior intensity determination module is configured to predict a probability hypothesis density of a new target in the observation subset, and determine a predicted target prior intensity according to the probability hypothesis density of the new target.

[0254] An updated target posterior intensity determination module is configured to estimate a probability hypothesis density of a clutter in the observation subset, and update the target posterior intensity at the current time according to the probability hypothesis density of the clutter and the predicted target prior intensity to determine an updated target posterior intensity.

[0255] A twice-updated target posterior intensity determination module is configured to crop and merge the initialized Gaussian component set to generate a cropped and merged Gaussian component set, and update the updated target posterior intensity according to the cropped and merged Gaussian component set to determine a twice-updated target posterior intensity; the twice-updated target posterior intensity is assigned as a probability hypothesis density of the local fusion center.

[0256] A probability hypothesis density acquisition module of a neighbor node is configured to communicate with a neighbor node to acquire a probability hypothesis density of the neighbor node.

[0257] The fused common target probability hypothesis density generation module is configured to fuse the probability hypothesis densities representing the common targets based on the probability hypothesis densities of the local fusion center and the probability hypothesis densities of the neighbor nodes according to the GCI rule to generate a fused common target probability hypothesis density, and fuse the probability hypothesis densities representing the extra targets according to the AA rule to generate a fused extra target probability hypothesis density.

[0258] The third updated target posterior intensity determination module is configured to determine a current fused target posterior intensity according to the fused common target probability hypothesis density and the extra target probability hypothesis density, and update the second updated target posterior intensity according to the current fused target posterior intensity to determine a third updated target posterior intensity.

[0259] The target state set extraction module is configured to extract a target state set based on the third updated target posterior intensity.

[0260] The filtering iteration module is configured to return to the step of predicting the probability hypothesis density of the surviving targets according to the target posterior intensity of the last time, determining the surviving components, and dividing the measurement set according to the surviving components to generate the observation subset from the new targets and the clutter, to perform filtering iteration of the next time until the current time is greater than the set time, and end the filtering iteration.

[0261] Embodiment five

[0262] The embodiment of the present application also provides an electronic device, including a memory and a processor, the memory is used for storing a computer program, and the processor is used for running the computer program to enable the electronic device to execute the distributed multi-target tracking filtering method in the embodiment one. The electronic device can be a server.

[0263] In addition, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the distributed multi-target tracking filtering method in the embodiment one.

[0264] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, the description is relatively simple because it corresponds to the method disclosed in the embodiments. The relevant parts can be referred to the method part.

[0265] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A distributed multi-target tracking filtering method, characterized in that, The method comprises the following steps: constructing a single-target state, constructing an observation model according to the single-target state when a target is detected by a sensor, and generating an observation value of any sensor; determining a measurement set of a local fusion center at any time according to the observation value of any sensor; initializing a Gaussian component set, determining a target posterior intensity at a current time, and obtaining a target posterior intensity at a previous time based on the target posterior intensity at the current time; predicting a probability hypothesis density of a surviving target according to the target posterior intensity at the previous time, determining a surviving component, and dividing the measurement set according to the surviving component to generate an observation subset derived from a new target and clutter; predicting a probability hypothesis density of the new target in the observation subset, and determining a predicted target prior intensity according to the probability hypothesis density of the new target; estimating a probability hypothesis density of the clutter in the observation subset, and updating the target posterior intensity at the current time according to the probability hypothesis density of the clutter and the predicted target prior intensity to determine an updated target posterior intensity; trimming and merging the initialized Gaussian component set to generate a trimmed and merged Gaussian component set, and updating the updated target posterior intensity according to the trimmed and merged Gaussian component set to determine a twice-updated target posterior intensity; the twice-updated target posterior intensity is assigned as a probability hypothesis density of the local fusion center; communicating with a neighbor node to obtain a probability hypothesis density of the neighbor node; based on the probability hypothesis density of the local fusion center and the probability hypothesis density of the neighbor node, fusing probability hypothesis densities representing common targets by using a GCI rule to generate a fused common target probability hypothesis density, and fusing probability hypothesis densities representing additional targets by using an AA rule to generate a fused additional target probability hypothesis density; determining a current fused target posterior intensity according to the fused common target probability hypothesis density and the additional target probability hypothesis density, and updating the twice-updated target posterior intensity according to the current fused target posterior intensity to determine a thrice-updated target posterior intensity; extracting a target state set based on the thrice-updated target posterior intensity; when the current time is less than a set time, returning to the step of predicting the probability hypothesis density of the surviving target according to the target posterior intensity at the previous time, determining the surviving component, dividing the measurement set according to the surviving component to generate the observation subset derived from the new target and the clutter, and performing a filtering iteration at a next time until the current time is greater than the set time, and ending the filtering iteration.

2. The distributed multi-target tracking filtering method of claim 1, wherein, the target posterior intensity at the current time is: where v k (x) is the target posterior intensity at the current time k, is a Gaussian density with mean and covariance x is the variable represented by the Gaussian distribution, i is the index value of the Gaussian component, N k is the number of Gaussian components, is the weight of the component.

3. The distributed multi-target tracking filtering method of claim 2, wherein, the step of predicting the probability hypothesis density of the new target in the observation subset, and determining the predicted target prior intensity according to the probability hypothesis density of the new target specifically comprises: The probability hypothesis density of the new-born target in the observation subset is predicted by using the formula The probability hypothesis density of the new-born target in the observation subset is predicted by using the formula k (x) is the probability hypothesis density of the new-born target, N b,k is the observation subset derived from the new-born target and clutter, is the weight of the i-th Gaussian component in γ k (x), is the mean of the i-th Gaussian component in γ k (x), is the covariance of the i-th Gaussian component in γ k (x). Utilizing determining a predicted target prior intensity; wherein, v k|k-1 (x) is a predicted target prior intensity, v p,k|k-1 (x) is a predicted target prior intensity, v k|k-1 is a number of prior Gaussian components, is a weight of an i-th Gaussian component in v k|k-1 (x), v is a mean of an i-th Gaussian component in v k|k-1 (x), v is a covariance of an i-th Gaussian component in v k|k-1 (x).

4. The distributed multi-target tracking filtering method of claim 3, wherein, the updated target posterior intensity is: where v k (x)′ is the updated target posterior intensity, p D is the detection probability, Z k,q is the set of observations from node q, z is a Z k,q any one of the observation vectors, is the weight of the i-th Gaussian component in v k (x)′, is the mean of the i-th Gaussian component in v k (x)′, is the covariance of the i-th Gaussian component in v k (x)′.

5. The distributed multi-target tracking filtering method of claim 1, wherein, The initialized Gaussian component set is pruned and merged to generate a pruned and merged Gaussian component set, and the updated target posterior intensity is updated according to the pruned and merged Gaussian component set to determine a twice-updated target posterior intensity, and the method specifically comprises: A pruning threshold, a merging threshold and a Gaussian component number threshold are set; The initialized Gaussian component set is pruned according to the pruning threshold to generate a pruned Gaussian component index set, and an index of a component corresponding to a current maximum weight value is found; An index sub-set of components similar to the component corresponding to the maximum weight value is found according to the merging threshold, the pruned Gaussian component index set and the index of the component corresponding to the current maximum weight value; Based on the index sub-set of the components similar to the component corresponding to the maximum weight value, Gaussian components are merged to generate a pruned and merged Gaussian component set; Based on the Gaussian component number threshold, the updated target posterior intensity is updated according to the pruned and merged Gaussian component set to determine a twice-updated target posterior intensity.

6. The distributed multi-target tracking filtering method of claim 5, wherein, Based on the probability hypothesis density of the local fusion center and the probability hypothesis density of the neighbor node, the probability hypothesis density representing the common target is fused by using the GCI rule to generate a fused common target probability hypothesis density, and the probability hypothesis density representing the additional target is fused by using the AA rule to generate a fused additional target probability hypothesis density, and the method specifically comprises: Based on the probability hypothesis density of the local fusion center, a local fusion weight of the local fusion center is determined, and a local Gaussian component set and a local Gaussian component index of the local fusion center are defined; Based on the probability hypothesis density of the neighbor node, a neighbor fusion weight of the neighbor node is determined, and a neighbor Gaussian component set and a neighbor Gaussian component index of the neighbor node are defined; A local decomposition threshold and a neighbor decomposition threshold are set; The local fusion weight and the neighbor fusion weight are normalized to generate a normalized local fusion weight and a normalized neighbor fusion weight; Based on the local decomposition threshold, a local Gaussian component having the local Gaussian component index in the local Gaussian component set is decomposed to generate a component set of a local common target and a component set of a local additional target; Based on the neighbor decomposition threshold, a neighbor Gaussian component having the neighbor Gaussian component index in the neighbor Gaussian component set is decomposed to generate a component set of a neighbor common target and a component set of a neighbor additional target; Based on the normalized local fusion weight, the normalized neighbor fusion weight, the component set of the local common target and the component set of the neighbor common target, the probability hypothesis density representing the common target is fused by using the GCI rule to generate a fused common target probability hypothesis density; Based on the component set of the local additional target and the component set of the additional common target, the probability hypothesis density representing the additional target is fused by using the AA rule to generate a fused additional target probability hypothesis density.

7. The distributed multi-target tracking filtering method of claim 6, wherein, The target state set comprises: wherein, is a set of target states.

8. A distributed multi-target tracking filtering system, characterized in that, It comprises: The observation value generation module is configured to construct a single-target state, and construct an observation model according to the single-target state when a target is detected by a sensor, and generate an observation value of any sensor; The observation machine determination module is configured to determine a measurement set of the local fusion center at any time according to the observation value of any sensor; The last-time target posterior intensity determination module is configured to initialize a Gaussian component set, determine a target posterior intensity at a current time, and obtain a target posterior intensity at a last time based on the target posterior intensity at the current time; The observation subset generation module of the new-born target and clutter is configured to predict a probability hypothesis density of a surviving target according to the target posterior intensity at the last time, determine a surviving component, divide the measurement set according to the surviving component, and generate an observation subset from a new-born target and clutter; The predicted target prior intensity determination module is configured to predict a probability hypothesis density of a new-born target in the observation subset, and determine a predicted target prior intensity according to the probability hypothesis density of the new-born target; The updated target posterior intensity determination module is configured to estimate a probability hypothesis density of a clutter in the observation subset, and update the target posterior intensity at the current time according to the probability hypothesis density of the clutter and the predicted target prior intensity, to determine an updated target posterior intensity; The twice-updated target posterior intensity determination module is configured to crop and merge the initialized Gaussian component set, generate a cropped and merged Gaussian component set, and update the updated target posterior intensity according to the cropped and merged Gaussian component set, to determine a twice-updated target posterior intensity; the twice-updated target posterior intensity is assigned as a probability hypothesis density of the local fusion center; The neighbor node probability hypothesis density acquisition module is configured to communicate with a neighbor node, and acquire a probability hypothesis density of the neighbor node; The fused common target probability hypothesis density and fused extra target probability hypothesis density generation module is configured to fuse a probability hypothesis density representing a common target by using a GCI rule based on the probability hypothesis density of the local fusion center and the probability hypothesis density of the neighbor node, to generate a fused common target probability hypothesis density, and fuse a probability hypothesis density representing an extra target by using an AA rule, to generate a fused extra target probability hypothesis density; The thrice-updated target posterior intensity determination module is configured to determine a current fused target posterior intensity according to the fused common target probability hypothesis density and extra target probability hypothesis density, and update the twice-updated target posterior intensity according to the current fused target posterior intensity, to determine a thrice-updated target posterior intensity; The target state set extraction module is configured to extract a target state set based on the thrice-updated target posterior intensity. The filtering iteration module is configured to return to the step of "predicting the probability hypothesis density of the survival target according to the target posterior intensity of the last time, determining a survival component, dividing the measurement set according to the survival component, and generating an observation subset from the new target and clutter" when the current time is less than the set time, and to perform filtering iteration of the next time until the current time is greater than the set time, and to end the filtering iteration.

9. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory is configured to store a computer program, and the processor is configured to run the computer program to enable the electronic device to perform the distributed multi-target tracking filtering method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored in the memory and is executed by the processor to implement the distributed multi-target tracking filtering method according to any one of claims 1 to 7.

Images