Group target tracking method based on virtual leader model
By introducing virtual leader model and graph theory in multi-objective tracking, establishing collaborative relationships and structural information between goals, the problem of difficult to accurately estimate the number and status of goals in multi-objective tracking is solved, and more efficient group goal tracking is achieved .
Patent Information
- Application Number
- CN202411409634.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2025-02-18
AI Technical Summary
In multi-objective tracking scenarios, it is difficult for the prior art to accurately estimate the number and state of multiple targets with certain motion patterns, especially in group targets, where the cooperative relationship and structural information between targets are not fully utilized.
A group goal tracking method based on the virtual leader model is adopted, and the collaborative relationship and structural information between the targets are established through random finite sets and graph theory, and state prediction and quantity prediction are used using tag RFS and GLMB algorithms.
It significantly improves the tracking accuracy and robustness of the distinguishable group targets, and can effectively utilize the interaction of within-group targets to achieve more accurate state and quantity predictions.
Smart Images
Figure CN119471667A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of radar signal processing, and in particular to a group target tracking method based on a virtual leader model. Background Art
[0002] In the scenario where radar tracks and monitors aircraft targets, the tracking object is generally composed of multiple targets with certain movement patterns. In radar information processing technology, multiple targets with similar movement patterns are often grouped as group targets, and multiple group targets are tracked to achieve multi-target tracking scenarios. In the multi-target tracking implementation technology, a random finite set (RFS) is defined as a random element with a finite value. Unlike a random variable, the number of elements in a random finite set is random, and the values of the elements themselves are also random, and the elements are arranged in disorder. Due to this feature of random finite sets, it is very suitable for multi-target tracking scenarios.
[0003] Different from single target tracking, multi-target tracking requires not only accurate estimation of the target state, but also accurate estimation of the number of targets in the space. Therefore, when modeling multi-targets with random finite sets, it is also necessary to use the probability distribution of the RFS potential to predict the number of elements in the set, and use the joint probability density of RFS to achieve multi-target state prediction. Summary of the invention
[0004] To achieve the above purpose, the present application provides a group target tracking method based on a virtual leader model, comprising the following steps:
[0005] Obtain the measurement values of N(k-1) targets at time k-1 in the state space X; the measurement values include the number and state of the targets, expressed in the form of a finite set: X k-1 ={x k-1,1 ,…,x k-1,j}∈X, where x k-1,i , i=1,2,...,N(k-1) is the state of the i-th target in the state space X at time k-1, j≤N(k-1), N(k-1) is the number of targets in the set at time k-1;
[0006] Load a random label set and create a measurement space-based The observed RFS is expressed as: The measurement space The state of the i-th target at time k is expressed as: z k,i ,i=1,2,...,M(k); establish the multi-objective state set X in the state space X at time k k ; is in the state space The state x of each target in k,i The tag of the binding identification is expressed as l = (k, i), where k is the time when the target is generated, and l means that the target is the i-th target generated at time k;
[0007] Obtaining the relationship between multiple targets in the state space X, and determining the subgroup target according to the relationship;
[0008] Calculate the status and quantity of subgroup targets at time k.
[0009] The process of determining subgroup targets based on relationships includes:
[0010] Define a group ο consisting of n targets in the state space X. The undirected graph of the group is represented by g ο =(V ο ,E ο ), the node set is represented by V ο = {v1,K,v n}, where node v i Contains the motion state information of target i;
[0011] Determine whether any two nodes in the node set belong to the same subgroup;
[0012] The adjacency matrix is used to describe the collaborative relationship between targets in the group. The adjacency matrix is expressed as:
[0013] Among them, when the i-th target is the j-th target parent node, a(i,j)=1, otherwise a(i,j)=0;
[0014] The relationship between the multiple targets is determined, including whether it is a parent-child relationship, and a group target motion model is defined to distinguish the group target motion model.
[0015] Among them, judging whether any two nodes in the node set belong to the same subgroup includes:
[0016] Extract any two nodes v i and v j , determine whether there is an edge between the nodes, if so, the two nodes belong to the same subgroup g;
[0017] The edge set corresponding to the subgroup g satisfies E g ={(i,j)|i≠j; v i ,v j ∈V g ;d ij <ε}.
[0018] Defining the motion model for distinguishing group targets includes:
[0019] Determine the resolution group target motion model at time k-1, the resolution group target motion model is expressed as: x k-1,i ∈X k-1 ,in, is the position on the x-axis, is the velocity in the x-axis direction, is the position on the y-axis, is the velocity in the y-axis direction;
[0020] The distinguishing group target motion model can be used to predict the state and quantity of each subgroup target at time k.
[0021] Furthermore, the state and quantity of each subgroup target at time k are predicted, including:
[0022] If the target has a single parent node, the state predicted by the group target motion model at time k is expressed as:
[0023] x k,i =F k-1,l x k-1,l +b k-1 (l,i)+B k-1,i w k-1,i ,
[0024] z k,i =C k x k,i +v k,i ,
[0025] Among them, F k-1,l , B k-1,i They represent the state transfer matrix and state noise coefficient matrix at time k-1, respectively, and C k is the observation matrix at time k, w and v are the system noise and observation noise at the corresponding time respectively, and both obey the normal distribution.
[0026] Furthermore, if the specified target has no parent node, the target is defined as the head node, and the state of the head node at time k-1 is x k-1,l , where the compensation vector b k-1 (l,i)=0;
[0027] The state of the child nodes of the head node at time k is expressed as:
[0028]
[0029] in, represents all nodes of target i.
[0030] Among them, the multi-objective state set X in the state space X at time k is k Given the k-1 time multi-objective Xk-1 The survival part, the predicted derivative part and the new part are composed of:
[0031] Among them, x k-1 is k-1 time point X k-1 The individual in S k∣k-1 (x k-1 ) is the RFS of the target at time k-1 and the target surviving at time k, B k∣k-1 (x k-1 ) is X k-1 The RFS of the derived target at time Γ k is the RFS of the individual born at time k.
[0032] Specifically, before calculating the state of the subgroup target at time k, the state of the target at time k is estimated by the GLMB algorithm, which is expressed as:
[0033] Defining the Deviation Matrix The difference in the target state is expressed as:
[0034]
[0035] Definition k (i,j), expressed as:
[0036] Get the predicted adjacency matrix at time k, expressed as:
[0037] Among them, η λ Represents η k The value of (i,j) is k When (i,j) is less than this value, is equal to 1, otherwise it is 0; Q k,n is the covariance matrix of the process noise.
[0038] When calculating the number of subgroup targets, the Laplacian matrix is defined as: Among them, E k is a symmetric matrix and where e k,i Represents node v k,i The degrees of freedom of is the number of edges connected to the node; is the estimated adjacency matrix;
[0039] Get the eigenvalues of the Laplacian matrix, where the number of times 0 appears is the number of subgroups.
[0040] According to the present invention, the structural information of the group can be combined with the movement of the targets within the group, thereby significantly improving the tracking accuracy of the distinguishable group targets; on the other hand, the present invention introduces the concept of label RFS, assigns tracks to different targets in the state space, and uses the adjacency matrix to describe the collaborative relationship between the targets in the group, making full use of the interaction between the targets in the group, thereby significantly improving the accuracy and robustness of distinguishable group target tracking. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 is a step diagram of a group target tracking method provided according to an embodiment of the present invention;
[0042] Figure 2 is an undirected graph of a group goal provided by an embodiment of the present invention;
[0043] Figure 3 is a schematic diagram of the dependency relationship between targets provided according to an embodiment of the present invention;
[0044] Figure 4 is a schematic diagram of a multi-group target structure provided according to an embodiment of the present invention. DETAILED DESCRIPTION
[0045] In the state space where the observed targets in the radar monitoring range are located, multiple targets move in a certain cooperative mode, and the cooperative mode can be reflected by a graph structure mode. In the present invention, a single target in the observation space is identified through a random finite set to distinguish the cooperative movement of the graph structure mode of multiple targets in the group target, distinguish the structure of the group target, and describe the relationship between the targets in the group through graph theory; with the help of the similarity of the graph structure and the group structure, the adjacency matrix is used to describe the distinguishable group target structure, and the number prediction and state prediction of the group target are realized.
[0046] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings.
[0047] Figure 1 A step diagram of a group target tracking method based on a virtual leader model is provided, as shown in the figure, including the following steps:
[0048] Step S100: Obtain the measurement values of N(k-1) targets at time k-1 in the state space X; the measurement values of these targets, that is, the quantity and state, can be expressed in the form of a finite set as follows:
[0049] X k-1 ={x k-1,1 ,…,x k-1,j}∈X (1),
[0050] Among them, x k-1,i,i=1,2,...,N(k-1) is the state of the i-th target in the state space X at time k-1, j≤N(k-1), N(k-1) is the number of targets in the set at time k-1.
[0051] In the present invention, the target is the tracking target of the radar in the tracking scene.
[0052] Step S110: Load a random label set and create a measurement space-based Observation RFS;
[0053] The observed RFS is expressed as:
[0054] Among them, z k,i ,i=1,2,...,M(k) represents the measurement space The state of the i-th target at time k.
[0055] Multiple targets in the state space X at time k can establish a multi-target state set X k , the multi-objective state set X at time k k From the k-1 time multi-objective X k-1 Collective prediction generation. Specifically, the multi-objective state set X at time k k Given the k-1 time multi-objective X k-1 The surviving part, the predicted derivative part and the new part are composed of:
[0056] (1) Survival part, k-1 time X k-1 Individual x in k-1 , with p s,k (x k-1 ) is the survival probability, and the RFS of the surviving target is expressed as S k∣k-1 (x k-1 ), the corresponding x k-1 With p s,k (x k-1 ) survives with a probability of 1-p s,k (x k-1 ) disappears;
[0057] (2) Derivative part, X k Some of the targets in may be represented by X k-1 The target split in is generated, that is, the derived target, whose RFS is represented by B k∣k-1 (x k-1 );
[0058] (3) New part: New targets may appear in the observation area at any time. The RFS of new targets is expressed as Γ k .
[0059] On this basis, the multi-objective state set X in the state space X at time k k It is expressed as:
[0060]
[0061] Among them, x k-1 is k-1 time point X k-1 The individual in S k∣k-1 (x k-1 ) is the RFS of the target at time k-1 and the target surviving at time k, B k∣k-1 (x k-1 ) is X k-1 The RFS of the derived target at time Γ k is the RFS of the individual born at time k.
[0062] In this step, a label space is created. With state space Phase distinction, and corresponding to the state space The state x of each target in k,i The tag of the binding identification is expressed as l = (k, i), where k is the time when the target is generated, and l means that the target is the i-th target generated at time k;
[0063] Define the target RFSX as defined in the state space and label space The RFS on the There is a one-to-one correspondence between labels and targets, defining the joint space to the label space The mapping function For the tag RFSX, is its label set. At the same time, the number of labels is equal to the number of targets, expressed as: where |·| represents the cardinality of the set.
[0064] On this basis, the uniqueness indicative function of the defined label is expressed as:
[0065] and:
[0066] On the other hand, we calculate the multi-objective X at a given k-1 time k-1 When the survival part is s,k (x k-1 ): Since the label Multi-Bernoulli RFS is composed of the Multi-Bernoulli parameter set {r (ξ) ,p (ξ) (·)} ξ∈ΨDefinition, where Ψ is the index set, r represents the probability of the target, and p represents the probability distribution of the target state; for each individual Bernoulli RFS, if it exists, it can get a label;
[0067] Therefore, the definition in The label multi-Bernoulli RFS probability density of the joint space is expressed as:
[0068]
[0069] in,
[0070] The probability density formula for multi-Bernoulli RFS is abbreviated as:
[0071]
[0072] Step S120: acquiring the relationship between multiple targets in the state space X, and determining the subgroup target according to the relationship;
[0073] In the present invention, graph theory is introduced to correspond the state space to a group, wherein each target in the state space corresponds to a node, and the state of each node can reflect the relationship between the targets in the state space; when predicting the state and quantity of the target at each moment, multiple targets with related states are formed into subgroups, and the state characteristics such as the movement direction and quantity at subsequent moments are predicted as a whole target, i.e., the subgroup target. That is: the group is split into multiple subgroups according to the states of some targets in the state space, and one group includes multiple subgroups; one group target is composed of one or more subgroup targets.
[0074] The specific implementation process includes:
[0075] Define a group ο consisting of n targets in the state space X, where the undirected graph of the group is represented by g ο =(V ο ,E ο ), the node set is represented by V ο = {v1,K,v n}, where node v i Contains the motion state information of target i;
[0076] Determine whether any two nodes in the node set belong to the same subgroup. The determination process includes:
[0077] Extract any two nodes v i and v j , if the distance between them is d ij If it is less than the threshold ε, then it is determined whether there is an edge between the nodes. If so, it can be considered that node v i and v j , that is, target i and target j belong to the same subgroup g.
[0078] At this time, the edge set corresponding to the subgroup g satisfies E g ={(i,j)|i≠j; v i ,v j ∈V g ;d ij <ε}.
[0079] Define the mapping function Γ g (i) = {j|j≠i; v i ,v j ∈V} points to other targets in the same group as target i, and defines G = {g1,K,g m} is a set of multiple undirected graphs.
[0080] Through the similarity between graph structure and group structure, the adjacency matrix is introduced to distinguish the group target structure. Since the adjacency matrix can describe the cooperative relationship between the targets in the group, an asymmetric adjacency matrix is used, that is, the adjacency matrix describes the parent-child relationship of the interconnected targets.
[0081] The adjacency matrix is represented as:
[0082]
[0083] Among them, when the i-th target is the j-th target parent node, a(i,j)=1, otherwise a(i,j)=0.
[0084] On this basis, the relationship between the target and the parent node is obtained, and the motion model for distinguishing group targets is defined.
[0085] At K-1 moment, x k-1,i ∈X k-1 ,in, is the position on the x-axis, is the velocity in the x-axis direction, is the position on the y-axis, is the velocity in the y-axis direction;
[0086] If the target has a single parent node, define the distinguishing group target motion model to predict the state at time k, expressed as:
[0087] x k,i =F k-1,l x k-1,l +b k-1 (l,i)+B k-1,i w k-1,i (11)
[0088] z k,i =C k x k,i +vk,i (12)
[0089] Among them, F k-1,l , B k-1,i They represent the state transfer matrix and state noise coefficient matrix at time k-1, respectively, and C k is the observation matrix at time k, w and v are the system noise and observation noise at the corresponding time, respectively, and both obey the normal distribution.
[0090] If a target has no parent node, then the target is called the head node as a virtual leader. The movement of the head node will affect its child targets, while the head node's own movement is not affected by other targets. Therefore, the compensation vector in the head node motion model is defined, and x k-1,l is its own state at time k-1; if the target has a parent node and the target's motion is affected by its parent node, the compensation vector b in the target motion model is k-1 (l,i) contains the direction and distance information between the node and its parent node. When the target has multiple parent nodes, x k,i It is expressed as follows:
[0091]
[0092] in, represents all nodes of target i.
[0093] In this step, the target i as the head node and the node with the target i as the head node are the same subgroup.
[0094] like Figure 2 As shown in the undirected graph example, the 7 targets belong to 2 subgroups, and the arrows indicate the movement direction of the group. Although there is no edge connection between v6 and v7, they have a common connection node v5, so it is judged that v6 and v7 belong to the same subgroup.
[0095] Based on the motion model of distinguishing group targets, the state and quantity of each subgroup target at the next moment, i.e., moment k, can be calculated based on moment k-1.
[0096] Step S130: Calculate the state and quantity of the subgroup targets at time k.
[0097] Before calculating the state and number of subgroup targets, the GLMB filtering algorithm is used to estimate the next state of the target in the state space based on the given value at the previous moment. The standard GLMB algorithm is defined as:
[0098]
[0099] In order to facilitate calculation, the above formula is transformed into:
[0100]
[0101] This formula is abbreviated as δ-GLMB.
[0102] The prediction step calculation formula of δ-GLMB is as follows:
[0103]
[0104] in:
[0105]
[0106]
[0107] Specifically, Is a new label The weight of It is a survival tag The weight of is the probability density of the new target, The prior density The density of the surviving targets obtained, Represents the probability density of surviving targets.
[0108] On this basis, the update step calculation formula is:
[0109]
[0110] Among them, in a fixed (I,ξ), Θ (M) ={ξ (1) ,…,ξ (M)}The set represents the maximum weight When Θ is M, is the normalized weight after truncation.
[0111] Next, the deviation matrix and deviation vector can be calculated to obtain the adjacency matrix at time k.
[0112] In this step, the next state of the target is estimated by the GLMB filtering algorithm. Represents the state estimation set of the group target at time k. In order to obtain the structure of the group at time k, we first estimate the adjacency matrix at time k. Since the adjacency matrix cannot be obtained directly, we introduce the deviation matrix Represents the difference between the states of each target, expressed as:
[0113]
[0114] Among them, d k (i,j) is defined as:
[0115] Furthermore, through the deviation vector bk (l,i) and the deviation matrix can predict the adjacency matrix, expressed as:
[0116]
[0117] Among them, η λ Represents η k The value of (i,j) is k When (i,j) is less than this value, is equal to 1, otherwise it is 0; Q k,n is the covariance matrix of the process noise.
[0118] Through the above method, the adjacency matrix of the distinguishable group targets at time k is finally predicted.
[0119] By k (i,j) is defined as d k (i,j)=d k (j,i), so the deviation matrix is a symmetric matrix. The predicted adjacency matrix is obtained from the predicted deviation matrix, so the predicted adjacency matrix is also a symmetric matrix.
[0120] The undirected graph of the target can be obtained from the symmetric adjacency matrix. The undirected graph can obtain the mutual connection information of each node in the graph but cannot obtain the parent-child relationship between nodes. In implementation, the parent node often moves in front of the child node. Therefore, the present invention defines the target in front of two cooperatively moving targets as the parent node, c to determine the mutual relationship between the targets.
[0121] For example, in Figure 3 In the two moving groups of targets, group 1 moves upward, target 3 is in front of target B, so target A is the parent node of target B; group 2 moves downward, and target C is in front of target D, so target C is the parent node of target D. From the properties of the set, it can be seen that the elements in the set have no order. Therefore, the adjacency matrix estimate obtained by the state estimation set of the group at the same time is not fixed.
[0122] On the other hand, the present invention predicts the number of subgroups, ie, subgroup targets, based on the connectivity graph through the Laplacian matrix.
[0123] In practical applications, at every moment, due to the change in the status of each target in the subgroup target, the status of the target individuals that constitute the subgroup target changes, and the number of subgroups will also change.
[0124] This state change phenomenon is reflected in the target state set. Since the order of elements is irrelevant, the corresponding adjacency matrix will also change over time. Figure 4It can be seen that each subgroup is regarded as a connected graph, that is, any two points in the group are connected, but the entire group is not a connected graph, because there is no connection between subgroups 1, 2 and 3. The non-connected graph composed of these multiple connected graphs is called a forest.
[0125] The present invention adopts the concept of connected graph to predict the number of subgroups.
[0126] When specifically calculating the number of group targets, the Laplacian matrix is defined as:
[0127]
[0128] Among them, E k is a symmetric matrix and equal to e k,i Represents node v k,i The degrees of freedom of is the number of edges connected to the node; is the estimated adjacency matrix;
[0129] In this step, the number of subgroups is defined as the number of times 0 appears in the eigenvalues of the Laplacian matrix.
[0130] The group target tracking method based on the diversified formation constraint technology of the virtual leader model proposed in the present invention can combine the structural information of the group with the movement of the targets in the group, thereby significantly improving the tracking accuracy of the distinguishable group targets; on the other hand, compared with the traditional RFS multi-target tracking algorithm, the present invention introduces the concept of label RFS, adds identity-identifying labels to the states of different targets in the state space, solves the problem that the traditional RFS multi-target tracking algorithm cannot assign tracks to the targets, and outputs the tracks of distinguishable group targets under the random set framework; the present invention introduces graph theory to describe the relationship between the nodes in the group, introduces the adjacency matrix to describe the collaborative relationship between the targets in the group, and makes full use of the interaction between the targets in the group, thereby significantly improving the accuracy and robustness of distinguishable group target tracking.
[0131] The above disclosures are only several specific embodiments of the present invention; however, the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A group target tracking method based on a virtual leader model, characterized in that: The following steps are involved: Obtain the measurement values of N(k-1) targets at time k-1 in the state space X; the measurement values include the number and state of the targets, expressed in the form of a finite set: X k-1 ={x k-1,1 ,…,x k-1,j }∈X, where x k-1,i , i=1,2,...,N(k-1) is the state of the i-th target in the state space X at time k-1, j≤N(k-1), N(k-1) is the number of targets in the set at time k-1; Load a random label set and create a measurement space-based The observed RFS is expressed as: The measurement space The state of the i-th target at time k is expressed as: z k,i ,i=1,2,...,M(k); establish the multi-objective state set X in the state space X at time k k ; is in the state space The state x of each target in k,i The tag of the binding identification is expressed as l = (k, i), where k is the time when the target is generated, and l means that the target is the i-th target generated at time k; Obtaining the relationship between multiple targets in the state space X, and determining the subgroup target according to the relationship; Calculate the status and quantity of subgroup targets at time k.
2. The group target tracking method according to claim 1, characterized in that: Determining the subgroup target according to the relationship includes: Define a group o consisting of n targets in the state space X, and the undirected graph of the group is represented by g ο =(V ο ,E ο ), the node set is represented by V ο ={v1,K,v n }, where node v i Contains the motion state information of target i; Determine whether any two nodes in the node set belong to the same subgroup; The adjacency matrix is used to describe the collaborative relationship between targets in the group. The adjacency matrix is expressed as: Among them, when the i-th target is the j-th target parent node, a(i,j)=1, otherwise a(i,j)=0; The relationship between the multiple targets is determined, including whether it is a parent-child relationship, and a group target motion model is defined to distinguish the group target motion model.
3. The group target tracking method according to claim 2, characterized in that: The determining whether any two nodes in the node set belong to the same subgroup includes: Extract any two nodes v i and v j , determine whether there is an edge between the nodes, if so, the two nodes belong to the same subgroup g; The edge set corresponding to the subgroup g satisfies E g ={(i,j)|i≠j; v i ,v j ∈V g ;d ij <ε}.
4. The group target tracking method according to claim 2, characterized in that: The definition of the group target motion model includes: Determine the resolution group target motion model at time k-1, the resolution group target motion model is expressed as: in, is the position on the x-axis, is the velocity in the x-axis direction, is the position on the y-axis, is the velocity in the y-axis direction; The distinguishing group target motion model can be used to predict the state and quantity of each subgroup target at time k.
5. The group target tracking method according to claim 4, characterized in that: The state and quantity of each subgroup target at the predicted k time include: If the target has a single parent node, the state predicted by the group target motion model at time k is expressed as: x k,i =F k-1,l x k-1,l +b k-1 (l,i)+B k-1,i w k-1,i , z k,i =C k x k,i +v k,i , Among them, F k-1,l , B k-1,i They represent the state transfer matrix and state noise coefficient matrix at time k-1, respectively, and C k is the observation matrix at time k, w and v are the system noise and observation noise at the corresponding time respectively, and both obey the normal distribution.
6. The group target tracking method according to claim 4, characterized in that: If the target has no parent node, define the target as the head node, and the state of the head node at time k-1 is x k-1,l , where the compensation vector b k-1 (l,i)=0; The state of the child nodes of the head node at time k is expressed as: in, represents all nodes of target i.
7. The group target tracking method according to claim 1, characterized in that: The multi-objective state set X in the state space X at time k k Given the k-1 time multi-objective X k-1 The survival part, the predicted derivative part and the new part are composed of: Among them, x k-1 is k-1 time point X k-1 The individual in S k∣k-1 (x k-1 ) is the RFS of the target at time k-1 and the target surviving at time k, B k∣k-1 (x k-1 ) is X k-1 The RFS of the derived target at time Γ k is the RFS of the individual born at time k.
8. The group target tracking method according to claim 1, before calculating the state of the subgroup target at time k, the state of the target at time k is estimated by the GLMB algorithm, which is expressed as: Defining the Deviation Matrix The difference in the target state is expressed as: Definition k (i,j), expressed as: Get the predicted adjacency matrix at time k, expressed as: in, η λ Represents η k The value of (i,j) is k When (i,j) is less than this value, Equal to 1, otherwise 0; Q k,n is the covariance matrix of the process noise.
9. The group target tracking method according to claim 8, characterized in that: When calculating the number of subgroup targets, the Laplacian matrix is defined as: Among them, E k is a symmetric matrix and where e k,i Represents node v k,i The degrees of freedom of is the number of edges connected to the node; is the estimated adjacency matrix; Get the eigenvalues of the Laplacian matrix, where the number of times 0 appears is the number of subgroups.