A cluster multi-hypothesis tracking method that jointly optimizes cluster partitioning and data association

Abstract

This invention discloses a cluster multi-hypothesis tracking method that jointly optimizes cluster partitioning and data association. By incorporating cluster partitioning as part of the global hypothesis of data association within the original multi-hypothesis tracking framework, a unified processing framework for cluster partitioning and data association is constructed. Joint optimization is performed under the maximum a posteriori criterion, and the corresponding optimization problem is transformed into a two-level optimization problem solvable in polynomial time. Furthermore, in cases of complex group structure changes such as cluster splitting and merging, this invention overcomes the problem that existing methods that separately process cluster partitioning and data association struggle to identify cluster structure changes in a timely and accurate manner, leading to a decrease in association accuracy and tracking precision. This invention can also seamlessly track cluster targets and non-cluster targets simultaneously in complex environments, handle complex group structure changes such as cluster splitting and merging, improve cluster target association accuracy and target tracking precision, and reduce track mixing.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to a cluster multi-hypothesis tracking method that jointly optimizes cluster partitioning and data association. Background Technology

[0002] Swarm target tracking, as one of the major challenges in the field of target tracking, has attracted widespread attention in recent years. It has many practical applications in both civilian and military fields, such as drone formations, vehicle formations, robot formations, and animal groups. Compared to the classic multi-target tracking problem, the technical difficulties of swarm target tracking mainly lie in: 1) When the spatial distribution of swarm targets is dense, it leads to poor observation resolution, severe echo crossover, and frequent target disappearance and reappearance, making it difficult to establish stable tracks for targets within the swarm; 2) Treating targets within the swarm as independently distributed individuals is not conducive to eliminating certain impossible situations (collisions, overlaps, etc.), increasing the computational load of the algorithm, and also hindering the full utilization of target motion information within the swarm to improve tracking accuracy and stability; 3) Considering only the motion information of individual targets or the swarm centroid cannot effectively grasp the structural changes of the swarm as a whole (such as splitting / merging), which is not conducive to rapid target identification and classification, and timely and effective judgment of target attempts.

[0003] The key to cluster target tracking methods is the processing of cluster partitioning of targets or observations, and the determination of the association and matching relationship between cluster targets and observation groups. However, due to the close spatial distance and cooperative movement of targets within a group, cluster splitting and merging lead to changes in the group structure, and the large number of targets within a group results in more challenging data association, filtering, and computation problems. AB Poore proposed a practical multi-hypothesis clustering tracking method in the multi-hypothesis tracking framework [Reference 1: Multiple hypothesis clustering and multiple frame assignment tracking, Signal and Data Processing of Small Targets, 2004, 5428:294-307]. This method first obtains possible cluster target partitioning and observation group partitioning using common clustering algorithms under a given clustering threshold, and then achieves group association through multi-hypothesis tracking. The local hypothesis loss coefficient adopts the local hypothesis loss coefficient calculation formula in the multi-hypothesis tracking method. The only difference is that the centroids of the target clusters or observation clusters obtained by the clustering algorithm are replaced by the centroids of the single target states or observations in the original calculation expression. This method can effectively reduce the computational load of association for dense cluster targets. However, in complex situations such as cluster splitting and merging, because it employs a step-by-step optimization strategy to handle cluster partitioning and data association, it struggles to identify cluster structure changes in a timely and accurate manner, leading to a decrease in association accuracy and tracking precision. Therefore, establishing a cluster target tracking method that can improve cluster structure estimation and association accuracy is a pressing issue that needs to be addressed. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a cluster multi-hypothesis tracking method that jointly optimizes cluster partitioning and data association. This method incorporates cluster partitioning as part of the global hypothesis for data association, and jointly optimizes cluster partitioning and data association under the maximum a posteriori criterion, retaining the previous hypothesis at each time step. The cumulative hypothesis with the highest probability can better solve the problem of cluster target tracking, and improve the accuracy of cluster target group structure estimation, data association accuracy and tracking precision.

[0005] The technical solution for achieving the objective of this invention is as follows:

[0006] A cluster multi-hypothesis tracking method that jointly optimizes cluster partitioning and data association includes:

[0007] Step 1: Receive sensor data at Observation at any time; the sensor in The observation at any given moment is: all sensors in the drone formation at... Real-time detection data, or all sensors in the robot formation at any given moment. The detection data at any given time includes radial distance, azimuth angle, and elevation angle.

[0008] Step 2: Enter to save Before the moment The most probable cumulative hypothesis And the target state estimate and state estimate error covariance under each assumption; if each cumulative assumption If it is not an empty set, then prediction is performed based on the state transition model, and the values ​​of all targets in the set are calculated. The forecast status and forecast covariance at each time point; the targets include confirmed targets and newly emerging targets;

[0009] Step 3: Use gate technology to confirm the target's position. The predicted status at each time point is preprocessed, and all confirmed targets are divided into multiple clusters. Confirmed targets located in the same cluster are further divided into group targets, forming a group target partitioning of all confirmed targets. And denote the set of all possible group objective partitions as ;according to The sensor observations received at each time point are preprocessed using a gate technique to divide them into multiple clusters. Then, observations within the same cluster are further subdivided into grouped observations, forming observation group partitions. And denote the set of all possible observation groups as ;

[0010] Step 4: Obtain the predicted state and predicted covariance of the confirmed targets through Step 2, as well as the possible group target partitioning set obtained in Step 3. and observation group division set Under the maximum a posteriori criterion, a joint optimization problem of cluster partitioning and data association is established, and the solution is obtained from the prior art. The hypothesis with the highest probability is as follows:

[0011] 4.1 Optimization Problem Modeling: For The first moment -best cumulative assumption ,make Indicates that it is in The first time derived from time The best assumption is obtained by solving the following problem:

[0012]

[0013] in, It is by exist The set of all possible hypotheses derived from time. From the initial moment to All sensor observations accumulated over time, yes In the given and The conditional probability density function under the given conditions, -best indicates that the probability is the highest. The largest one;

[0014] 4.2 Problem Transformation and Solution: The above problem is solved through a variant of a two-layer optimization problem, the inner layer of which involves solving a series of two-dimensional allocation problems. -best solution, the specific form of the two-dimensional allocation problem is as follows:

[0015]

[0016] Among them, optimization variables The value can be 0 or 1. Represents the group objective division The first in Group targets and observation group division The first in The observation groups are related; It is a group target division The first in Group targets and observation group division The first in The loss coefficient associated with each observation group Is it a selection of target partitioning? and observation group division The loss coefficient is calculated using the following expression:

[0017]

[0018] in, and These are the likelihood coefficients for cluster partitioning and the likelihood ratios for group association, respectively. and These are the first two parts of cluster target partitioning and observation group partitioning. Individual group goals and the first One observation group, and These are the expected number of observations generated within the corresponding clutter group and target area, respectively. It is the expected number of clutter groups generated in each scan. and These are the cluster targets. The center and detection probability, It is the first The sensor received the first time at the nth moment. One observation, and Observation Likelihood density and clutter density functions originating from the cluster center;

[0019] The two-dimensional allocation problem is solved using an allocation algorithm;

[0020] 4.3 Searching -best solution, i.e., solving the outer layer optimization problem: sorting the solutions obtained from the inner layer and finding the one with the smallest total loss. One, obtained exist Momentary -best assumption ;

[0021] Step 5: If If not empty, then for the obtained Sort the total losses corresponding to the assumptions at each time point and find the minimum. One, that is, to obtain Before the moment The most probable cumulative hypothesis And delete the remaining assumptions; otherwise, Before the moment The most probable cumulative hypothesis Set to an empty set;

[0022] against Each cumulative hypothesis retained at every moment Perform steps 6-8:

[0023] Step 6: If If a confirmed target exists, then the cumulative hypothesis is used. Given Time and The cluster structure information at any given time is used to determine changes in the cluster structure using the following cluster splitting and merging identification methods; otherwise, proceed to step 8.

[0024] The cluster splitting and merging identification method is as follows:

[0025] 6.1 Construct an undirected graph of cluster targets at adjacent time points: Let express Moment Individual group goals, each group goal Includes some individual goals; express Moment Individual group goals; if group goals and An edge exists connecting two groups if and only if both groups have the same group member in their target group. First, define two sets of nodes as follows:

[0026]

[0027] in, and Representing group goals and Construct an undirected graph to indicate Time and The connection relationships of the group target set at time t, where and These are undirected graphs The set of nodes and the set of edges are defined as follows:

[0028]

[0029] 6.2 Finding group targets with interactive relationships at adjacent time points: Finding undirected graphs using graph search algorithms. In all maximally connected subgraphs, the group objectives represented by the nodes in each maximally connected subgraph are the group objectives that have interactive relationships at adjacent time steps.

[0030] 6.3 Cluster Structure Determination: Assume that step 6.2 yielded an undirected graph. of A maximal connected subgraph is defined as ,in and These are maximal connected subgraphs. The set of nodes and the set of edges; let and Represent the vertex set of a maximal connected subgraph, respectively. Chinese representative Time and The set of group target vertices at time t, i.e.:

[0031]

[0032] Then, through each maximal connected subgraph corresponding to and The cardinality of the set is used to determine changes in the cluster structure, specifically in the following four cases:

[0033] 1) If and satisfy Then from Time to At any moment, the group target Keep the group structure unchanged;

[0034] 2) If and satisfy Then from Time to At any moment, a group goal Split into multiple subgroups of targets ;

[0035] 3) If and satisfy Then from Time to At any given moment, multiple group targets Merge into a single group target ;

[0036] 4) If and satisfy Then from Time to At any given moment, multiple group targets There exists a group of objectives that splits into multiple subgroups of objectives. At the same time, these subgroups merge with other groups of objectives, forming multiple groups of objectives. ;

[0037] Step 7: If There is a confirmed target, based on the cumulative hypothesis. Given the cluster partitioning and data associations, for each maximal connected subgraph found in step 6... For cluster targets, the state estimate and state estimate error covariance of the confirmed target are calculated and confirmed using the cluster center-intra-cluster target filtering method; otherwise, proceed to step 8.

[0038] The specific steps of the cluster center-intra-cluster target filtering method are as follows:

[0039] 7.1 Splitting a cluster into sub-clusters while preserving the cluster structure: If the cluster target is in Time to If the group structure remains unchanged, proceed to step 7.2; otherwise, split the cluster objective into sub-group objectives that maintain the same group structure, as follows:

[0040] 7.1.1 If a cluster target is in Time to If it splits into multiple subgroups at a given time, then according to... The subgroup formed at each moment will The cluster target at time t is divided into corresponding subgroups, at which point... The corresponding subgroups obtained by time splitting The subgroup at each moment maintains the group structure unchanged;

[0041] 7.1.2 If multiple cluster targets are in Time to If they are merged into a single cluster target at any given time, then according to... Multiple cluster targets at any given time will The large group formed at time is divided into corresponding subgroups, at which point... Cluster target at time and The subgroups obtained by splitting at each time step maintain the group structure unchanged;

[0042] 7.1.3 If cluster splitting and merging occur simultaneously, the split cluster target is processed according to step 7.1.1, and the merged cluster target is processed according to step 7.1.2, to obtain the results. Time to A subgroup whose group structure remains unchanged at all times;

[0043] 7.2 Group Center Filter Update: For each group center filter update... Time to For cluster targets or subgroups that maintain an unchanged group structure at all times, first calculate... The observation cluster center, which is the average of the observations within the observation cluster, is then used to update the data through a filtering method. Estimation of the cluster center state at time t;

[0044] 7.3 Intra-group target status update: Calculate the intra-group target status update for each cluster target or sub-group. The difference between the state estimate at time t and the corresponding cluster center state estimate, assuming that the displacements of the intra-group targets and the cluster center remain constant, is calculated by... The cluster center state estimate at a given time is used to update the target within the cluster. State estimation at time;

[0045] Step 8: Use cumulative hypotheses middle If there are no observations associated with the confirmed target at any given time, proceed... The connection between the new goals that started before the time and The process begins with the creation of a new target, followed by track confirmation and termination checks. Finally, the cumulative hypothesis is updated based on the track start, confirmation, and termination results. ;

[0046] Step 9: Retain Moment -best cumulative assumption and targeting The cumulative assumption with the highest probability at time is used to calculate and output all confirmed targets based on the estimation criteria. State estimation at time step, state estimation error covariance, and group structure information; setting Return to step 1.

[0047] The beneficial effects of this invention are:

[0048] 1) This invention constructs a unified processing framework for cluster partitioning and data association by incorporating cluster partitioning as part of the global hypothesis of data association in the original multi-hypothesis tracking framework. It performs joint optimization under the maximum a posteriori criterion and transforms the corresponding optimization problem into a two-level optimization problem that can be solved in polynomial time.

[0049] 2) In cases of complex group structure changes such as cluster splitting and merging, this invention overcomes the problem that existing methods for separating and processing cluster partitioning and data association are unable to identify cluster structure changes in a timely and accurate manner, leading to a decrease in association accuracy and tracking precision.

[0050] 3) This invention can seamlessly track cluster targets and non-cluster targets in complex environments simultaneously, handle complex group structure changes such as cluster splitting and merging, improve the accuracy of cluster target association and target tracking accuracy, and reduce track mixing. Attached Figure Description

[0051] Figure 1 This is a flowchart of the cluster multi-hypothesis tracking algorithm of the present invention.

[0052] Figure 2 The data shown in the example represents the actual flight path.

[0053] Figure 3 The track obtained in this example is processed using the classic multi-hypothesis tracking method.

[0054] Figure 4 The track obtained by applying the clustered multi-hypothesis tracking method of the present invention in the embodiment is shown.

[0055] Figure 5 Error curves are shown for the multi-hypothesis tracking algorithm in the embodiments, the multi-hypothesis clustering tracking algorithm combined with the DBSCAN clustering algorithm, and the cluster multi-hypothesis tracking algorithm of the present invention. Detailed Implementation

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] The flowchart of the cluster multi-hypothesis tracking algorithm of this invention is as follows: Figure 1 As shown, the specific steps are as follows:

[0058] Step 1: Enter to save to Before the moment The cumulative hypothesis with the highest probability (i.e., the corresponding cluster partitioning and data association). And the target state estimate and state estimate error covariance information under each assumption; receive sensor data at... The detection data at each moment (i.e., all observations generated by the sensors, which can be sensor observation data of typical cluster targets such as drone formations, robot formations, and animal groups, usually including information such as radial distance, azimuth angle, and pitch angle). At the initial moment, it is assumed that there is no target, that is, the initial cumulative hypothesis is an empty set. When the observations generated by the sensors at the next moment after the initial moment are received, the track is started by executing step 8 to complete the initialization of the cumulative hypothesis.

[0059] Terminology Explanation:

[0060] 1) Target group segmentation: dividing multiple targets The targets are assigned to some groups, and in any possible partition, each target can only be assigned to one group. The partitioning of unknown target groups into groups can be determined by index vectors. It indicates that the first one Dimensional components It is the first Group index of targets:

[0061]

[0062] in, Indicate the objective It is assigned to a new group outside of the existing group objectives.

[0063] 2) Observation Group Partitioning: Multiple observations obtained by the sensors (typically including radial distance, azimuth, elevation, etc.) are assigned to groups. Similar to target group partitioning, observation group partitioning can be achieved using index vectors. express.

[0064] 3) Group target association: and The pairing relationship between the divided target groups and the observed groups. This pairing relationship can be represented by an association vector. It indicates that the first one is... Dimensional components Is with the first Observation cluster index associated with each cluster target:

[0065]

[0066] 4) Global assumption: at any time global assumptions It can be divided by target group Observation group division Association with the corresponding group To define, that is .

[0067] 5) Cumulative assumption: Accumulating from the initial time (first time) to... The global assumption at time t is denoted as . .

[0068] 6) Confirmed targets and new targets: The status of a target track can be simply divided into new and confirmed targets. Confirmed targets will be output and displayed. Target track confirmation can adopt the classic M / N logic method: if a target is associated with observations at M times within N consecutive times, then the new target will be confirmed, that is, called a confirmed target.

[0069] For each hypothesis in the input If it is not an empty set, then perform the following steps 2-4:

[0070] Step 2: For Given The state estimates and corresponding state estimate error covariances of all targets (i.e., confirmed targets and newly formed targets) at any given time are used for prediction based on a given state transition model (e.g., common uniform linear motion (CV) models, uniform turning motion (CT) models, etc.). The target's position in time is calculated. The forecast status and forecast covariance at each time point.

[0071] Step 3: Use gate techniques (such as commonly used rectangular gates, ellipsoidal gates, etc.) to confirm the target's position. The predicted status at each time point is preprocessed, dividing all confirmed targets into multiple clusters (targets within the same cluster may belong to the same group of targets, while targets in different clusters are considered not to belong to the same group of targets). Next, only targets within the same cluster are further divided into groups, thus forming the group target classification for all confirmed targets. And denote the set of all possible group objective partitions as In addition, regarding The sensor observations received at each time point are processed in the same way to form an observation group division. And denote the set of all possible observation group partitions as .

[0072] Step 4: Obtain the predicted state and predicted covariance of the confirmed targets through Step 2, as well as the potential group target partitioning set obtained in Step 3. and observation group division set Under the maximum a posteriori criterion, a joint optimization problem of cluster partitioning and data association is established, and the solution is obtained from the prior art. The hypothesis with the highest probability;

[0073] The specific solution steps are described below:

[0074] Step 4.1 (Optimization Problem Modeling): For The first moment -best cumulative assumption ,make Indicates that it is in The first time derived from time The best assumption can be obtained by solving the following problem:

[0075]

[0076] in, It is by exist The set of all possible hypotheses derived from time. From the initial moment to All sensor observations accumulated over time, yes In a given and The conditional probability density function under the given conditions, -best indicates that the probability is the highest. The largest one.

[0077] Step 4.2 (Optimization Problem Transformation and Solution): Further, the above problem is solved through a variant of a two-layer optimization problem, the inner layer of which involves solving a series of two-dimensional allocation problems. The best solution can be obtained efficiently using an allocation algorithm.

[0078] The specific form of the two-dimensional allocation problem is as follows:

[0079]

[0080] Among them, optimization variables The value can be 0 or 1. Represents the group objective division The first in Group targets and observation group division The first in The observation groups are related; It is a group target division The first in Group targets and observation group division The first in The loss coefficient associated with each observation group Is it a selection of target partitioning? and observation group division The loss coefficient is calculated using the following expression:

[0081]

[0082] in, and These are the likelihood coefficients for cluster partitioning and the likelihood ratios for group association, respectively. and These are the first two parts of cluster target partitioning and observation group partitioning. Individual group goals and the first One observation group, and These are the expected number of observations generated within the corresponding clutter group and target area, respectively. It is the expected number of clutter groups generated in each scan. and These are the cluster targets. The center and detection probability, It is the first The sensor received the first time at the nth moment. One observation, and Observation Likelihood density and clutter density functions originating from the cluster center.

[0083] The above two-dimensional assignment problem can be solved efficiently by assignment algorithms, such as the classic Murty algorithm [Reference 2: An algorithm for ranking all the assignment in order of increasing cost, Operations Research, 1968, 16(3):682-687].

[0084] Step 4.3 (Searching) -best solution): Finally, the outer layer optimization problem is to sort the total loss (i.e., the objective function value of the optimization problem) corresponding to the solutions obtained from the inner layer and find the minimum. One, and you can get exist Momentary -best assumption ( and in The moment it derives It consists of (The cumulative assumption of time).

[0085] Step 5: If Not empty, by retaining to Each cumulative hypothesis at time Execute steps 2-4 to obtain We make assumptions about the time points, and then sort the total losses for each assumption to find the minimum. One, that is, to obtain Before the moment The most probable cumulative hypothesis And delete the remaining assumptions; otherwise, Before the moment The most probable cumulative hypothesis Set to an empty set.

[0086] against Each cumulative hypothesis retained at every moment Perform the following steps 6-8.

[0087] Step 6: If If a confirmed target exists, then the cumulative hypothesis is used. Given Time and The cluster structure information at any given time is used to determine changes in the cluster structure using the following cluster splitting and merging identification methods; otherwise, proceed to step 8.

[0088] The specific steps of the cluster splitting and merging identification method are as follows:

[0089] Step 6.1 (Constructing an undirected graph of cluster targets at adjacent time points): Let express Moment Individual group goals, each group goal It includes some individual goals. Let express Moment Individual group goals. If group goals and An edge exists connecting two groups if and only if both groups have the same group member in their target group. First, define two sets of nodes as follows:

[0090]

[0091] in, and Representing group goals and Construct an undirected graph to indicate Time and The connection relationships of the group target set at time t, where and These are undirected graphs The set of nodes and the set of edges are defined as follows:

[0092] .

[0093] Step 6.2 (Finding group targets with interaction relationships in adjacent time steps): Search for an undirected graph using a graph search algorithm (e.g., Tarjan's algorithm based on depth-first search). All maximal connected subgraphs in the graph (the vertex set of a maximal connected subgraph is actually the group before and after the group structure change, and the edge set is actually the group connection relationship before and after the structure change).

[0094] Step 6.3 (Cluster Structure Determination): Assume that step 6.2 yielded an undirected graph. of A maximal connected subgraph is defined as ,in and These are maximal connected subgraphs. The set of nodes and the set of edges; let and Represent the vertex set of a maximal connected subgraph, respectively. Chinese representative Time and The set of group target vertices at time t, i.e.:

[0095]

[0096] Then, through each maximal connected subgraph corresponding to and The cardinality of the set is used to determine changes in the cluster structure, specifically in the following four cases:

[0097] 1) If and satisfy Then from Time to At any moment, the group target Keep the group structure unchanged;

[0098] 2) If and satisfy Then from Time to At any moment, a group goal Split into multiple subgroups of targets ;

[0099] 3) If and satisfy Then from Time to At any given moment, multiple group targets Merge into a single group target ;

[0100] 4) If and satisfy Then from Time to At any given moment, multiple group targets There exists a group of objectives that splits into multiple subgroups of objectives. At the same time, these subgroups merge with other groups of objectives, forming multiple groups of objectives. .

[0101] Step 7: If There is a confirmed target, based on the cumulative hypothesis. Given the cluster partitioning and data associations, for each maximal connected subgraph found in step 6... For cluster targets, the state estimate and state estimate error covariance of the confirmed target are calculated using the cluster center-intra-cluster target filtering method; otherwise, proceed to step 8.

[0102] The specific steps of the cluster center-intra-cluster target filtering method are as follows:

[0103] 7.1 Splitting a cluster into sub-clusters while preserving the cluster structure: If the cluster target is in Time to If the group structure remains unchanged at all times, proceed to step 7.2; otherwise, the cluster objective needs to be split into sub-group objectives that maintain the group structure unchanged. Specifically: if a cluster objective is in Time to Split into multiple subgroups at any time, first according to The subgroup formed at each moment will The cluster target at time t is divided into corresponding subgroups (at this time, The corresponding subgroups obtained by time splitting The subgroup at time can be assumed to be at Time to (Keep the group structure unchanged at all times); if multiple cluster targets are in Time to If they are merged into a single cluster target at any given time, then according to... Multiple cluster targets at any given time will The large group formed at time is divided into corresponding subgroups (at this time, Cluster target at time and The corresponding subgroups obtained by time-shaping can be assumed to be in Time to (Maintaining the group structure unchanged at all times); if cluster splitting and merging occur simultaneously, a similar splitting can be performed to obtain... Time to A subgroup whose group structure remains unchanged at all times;

[0104] 7.2 Group Center Filter Update: For each group center filter update... Time to For cluster targets or subgroups that maintain an unchanged group structure at all times, first calculate... The observation cluster centers (i.e., the average of observations within the observation cluster) are correlated at each time step, and then the observation cluster centers are updated using common filtering methods (such as extended Kalman filtering, tasteless filtering, etc.) to obtain the results. Estimation of the cluster center state at time t;

[0105] 7.3 Intra-group target status update: Finally, calculate the intra-group target status update for each cluster target or sub-group. The difference between the state estimate at time 1 and the corresponding cluster center state estimate is calculated by maintaining the displacement of the target state within the cluster and the cluster center state unchanged. The cluster center state estimate at a given time is used to update the target within the cluster. State estimation at time step.

[0106] Step 8: Use cumulative hypotheses middle If there are no observations associated with the confirmed target at any given time, then... The connection between the new goals that started before the time and The process begins with the creation of a new target, followed by track confirmation and termination checks (track start, confirmation, and termination can employ common methods such as logical start, M / N confirmation, and termination criteria). Finally, the cumulative hypothesis is updated based on the track start, confirmation, and termination results. .

[0107] Step 9: Retain Moment -best cumulative assumption and targeting The cumulative hypothesis with the highest probability at time is used to calculate and output all confirmed targets based on the corresponding criteria (e.g., minimum mean square error estimation). The state estimate at each time step, the state estimate error covariance, and the group structure information are then determined; subsequently, the following settings are configured. Return to algorithm step 1.

[0108] This embodiment specifically describes the cluster multi-hypothesis tracking algorithm proposed in this invention for target association tracking in scenarios where cluster targets and non-cluster targets coexist and are accompanied by cluster splitting and merging, and simulation experiments were conducted.

[0109] Scenario Description: The following CV motion model and CT motion model are used to simulate and generate clustered targets and non-clustered targets, with the clustered targets undergoing multiple cluster splits and mergers; the sensor is set at the origin of the coordinate system, and target observations are generated under Gaussian noise perturbation, as well as clutter of a random number following a given Poisson distribution and uniformly distributed in the monitoring area; then the tracking algorithm is used to process the clustered targets and non-clustered targets to obtain the estimated track.

[0110] Data source: This example simulation demonstrates a scenario where a single sensor tracks clustered targets and non-clustered targets in a Cartesian two-dimensional coordinate system. System state They represent The target's position and velocity components at each moment are shown. The target is generated by CV and CT models. The specific state transition equations, observation equations, and motion models used are as follows:

[0111]

[0112]

[0113] in, It is the sensor sampling time interval. It refers to the turning rate. Specific scenarios include... Figure 2 As shown, there are 9 targets within the two-dimensional monitoring area. Circles represent the starting positions of the targets, squares represent the ending positions, and triangles represent the sensor positions. The sensors performed a total of 100 scans. The initial positions of the targets are shown below. Figure 2 As shown in Table 1, the survival time and initial velocity magnitude are as follows.

[0114] Table 1 Target survival time and initial velocity magnitude

[0115]

[0116] During the tracking process, the standard deviations of the sensor's radial distance and azimuth angle observation errors were respectively... All algorithms use CV model tracking, and the process noise is...

[0117]

[0118] The target movement in this scenario can be roughly divided into the following stages: 1) Targets 1-6 are newly created at the initial moment, forming two groups of targets (group target 1 contains targets 1-4, and group target 2 contains targets 5-6); 2) Targets 7-8 are newly created and then merge with group target 2; 3) Group target 1 and group target 2 merge to form a large group that moves in coordination, and then the group splits again to form two groups of targets (the new group target 1 contains targets 1-3, and group target 2 contains targets 4-8. Note that target 4 moved from one group to another before merging and after splitting); 4) The track of group target 2 terminates, group target 1 splits again, the track of the group containing targets 2-3 terminates, and then target 1 moves as a single target (i.e., a non-group target), accompanied by the creation of target 9; 5) The entire scenario ends with the termination of the track of target 9.

[0119] Evaluation metrics: This example uses... Distance metrics are used as evaluation indicators.

[0120] Comparison algorithm: using 100 Monte Carlo simulations The average error is used to compare the tracking accuracy of the classic multi-hypothesis tracking algorithm, the multi-hypothesis clustering tracking algorithm combined with the DBSCAN clustering algorithm, and the cluster multi-hypothesis tracking algorithm proposed in this invention.

[0121] Effect Analysis: Figure 3-4 The images show target trajectory estimation using a multi-hypothesis tracking algorithm and a cluster multi-hypothesis tracking algorithm that retains the 3-best hypothesis in a Monte Carlo experiment. Figure 3-4 As shown, the track obtained by the multi-hypothesis tracking algorithm has multiple batch mixings during the entire tracking process, while the estimated track of the clustered multi-hypothesis tracking algorithm is continuous, stable and has no batch mixing. Figure 5 These are estimated tracks obtained through different algorithms. Error curves plotted for distance over time. For example... Figure 5 As shown, in the complex group target tracking stage accompanied by cluster splitting and merging, the cluster multi-hypothesis tracking algorithm proposed in this invention (retaining 1-best and 3-best hypotheses) has a smaller tracking error than both the multi-hypothesis tracking algorithm and the multi-hypothesis clustering tracking algorithm combined with DBSCAN clustering. Furthermore, it retains more hypotheses at each time step, and the cluster multi-hypothesis tracking algorithm retaining 3-best hypotheses achieves better performance than the cluster multi-hypothesis tracking algorithm retaining 1-best hypotheses. In addition, after the 60th time step, since there are no cluster targets in the scene and all algorithms use the same track starting settings, the tracking accuracy of different algorithms gradually becomes consistent. It can be seen that the cluster multi-hypothesis tracking method provided in this invention can more accurately identify changes in group structure such as cluster splitting and merging when tracking and processing cluster targets, improving the association accuracy and tracking precision, and reducing track batch mixing.

[0122] Test conclusion: The cluster multi-hypothesis tracking method proposed in this invention can seamlessly handle cluster targets and non-cluster targets in complex environments.

Claims

1. A cluster multi-hypothesis tracking method that jointly optimizes cluster partitioning and data association, characterized in that, include: Step 1: Receive sensor data at Observation at any time; the sensor in The observation at any given moment is: all sensors in the drone formation at... Real-time detection data, or all sensors in the robot formation at any given moment. The detection data at any given time includes radial distance, azimuth angle, and elevation angle. Step 2: Enter to save Before the moment The most probable cumulative hypothesis , and the target state estimate and state estimate error covariance under each assumption; If each cumulative hypothesis If it is not an empty set, then prediction is performed based on the state transition model, and the values ​​of all targets in the set are calculated. The forecast status and forecast covariance at each time point; the targets include confirmed targets and newly emerging targets; Step 3: Use gate technology to confirm the target's position. The predicted status at each time point is preprocessed, and all confirmed targets are divided into multiple clusters. Confirmed targets located in the same cluster are further divided into group targets, forming a group target partitioning of all confirmed targets. And denote the set of all possible group objective partitions as ;according to The sensor observations received at any time are preprocessed using the gate technique and divided into multiple clusters. Then, observations within the same cluster are further divided into groups to form observation group partitions. And denote the set of all possible observation groups as ; Step 4: Obtain the predicted state and predicted covariance of the confirmed targets through Step 2, as well as the possible group target partitioning set obtained in Step 3. and observation group division set Under the maximum a posteriori criterion, a joint optimization problem of cluster partitioning and data association is established, and the solution is obtained from the prior art. The hypothesis with the highest probability; as follows: 4.1 Optimization Problem Modeling: For The first moment -best cumulative assumption ,make Indicates that it is in The first moment derived from time The best assumption is obtained by solving the following problem: ； in, It is by exist The set of all possible hypotheses derived from time. From the initial moment to All sensor observations accumulated over time, yes In the given and The conditional probability density function under the given conditions, -best indicates that the probability is the highest. The largest one; Indicates in , … hour, exist The first moment derived from time The set consisting of the best assumptions; 4.2 Problem Transformation and Solution: The above problem is solved through a variant of a two-layer optimization problem, the inner layer of which involves solving a series of two-dimensional allocation problems. -best solution, the specific form of the two-dimensional allocation problem is as follows: ； Among them, optimization variables The value can be 0 or 1. Represents the group objective division The first in Group target and observation group division The first in The observation groups are related; It is a group target division The first in Group targets and observation group division The first in The loss coefficient associated with each observation group Is it a selection of target partitioning? and observation group division The loss coefficient, , They are respectively by , The number of target groups and the number of observation groups; The specific calculation expression is as follows: ； in, and These are the likelihood coefficients for cluster partitioning and the likelihood ratios for group association, respectively. and These are the first two parts of cluster target partitioning and observation group partitioning. Individual group goals and the first One observation group, and These are the expected number of observations generated within the corresponding clutter group and target area, respectively. It is the expected number of clutter groups generated in each scan. and These are the cluster targets. The center and detection probability, It is the first The sensor received the first time at the nth moment. One observation, and Observation Likelihood density and clutter density functions originating from the cluster center; The two-dimensional allocation problem is solved using an allocation algorithm; 4.3 Searching -best solution, i.e., solving the outer layer optimization problem: sorting the solutions obtained from the inner layer and finding the one with the smallest total loss. One, obtained exist Momentary -best assumption ; Step 5: If If not empty, then for the obtained Sort the total losses corresponding to the assumptions at each time point and find the minimum. One, that is, to obtain Before the moment The most probable cumulative hypothesis And delete the remaining assumptions; otherwise, Before the moment The most probable cumulative hypothesis Set to an empty set; against Each cumulative hypothesis retained at every moment Perform steps 6-8: Step 6: If If a confirmed target exists, then the cumulative hypothesis is used. Given Time and The cluster structure information at any given time is used to determine changes in the cluster structure using the following cluster splitting and merging identification methods; otherwise, proceed to step 8. The cluster splitting and merging identification method is as follows: 6.1 Construct an undirected graph of cluster targets at adjacent time points: Let express Moment Individual group goals, each group goal Includes some individual goals; express Moment Individual group goals; if group goals and An edge exists connecting two groups if and only if both groups have the same group member in their target group. First, define two sets of nodes as follows: ； in, and Representing group goals and Construct an undirected graph to indicate Time and The connection relationships of the group target set at time t, where and These are undirected graphs The set of nodes and the set of edges are defined as follows: ； 6.2 Finding group targets with interactive relationships at adjacent time points: Finding undirected graphs using graph search algorithms. In all maximally connected subgraphs, the group objectives represented by the nodes in each maximally connected subgraph are the group objectives that have interactive relationships at adjacent time steps. 6.3 Cluster Structure Determination: Assume that step 6.2 yielded an undirected graph. of A maximal connected subgraph is defined as ,in and These are maximal connected subgraphs. The set of nodes and the set of edges; let and Represent the vertex set of a maximal connected subgraph, respectively. Chinese representative Time and The set of group target vertices at time t, i.e.: ； Then, through each maximal connected subgraph corresponding to and The cardinality of the set is used to determine changes in the cluster structure, specifically in the following four cases: 1) If and satisfy Then from Time to At any moment, the group target Keep the group structure unchanged; 2) If and satisfy Then from Time to At any moment, a group goal Split into multiple subgroups of targets ; 3) If and satisfy Then from Time to At any given moment, multiple group targets Merge into a single group target ; 4) If and satisfy Then from Time to At any given moment, multiple group targets There exists a group of objectives that splits into multiple subgroups of objectives. At the same time, these subgroups merge with other groups of objectives, forming multiple groups of objectives. ; Step 7: If There is a confirmed target, based on the cumulative hypothesis. Given the cluster partitioning and data associations, for each maximal connected subgraph found in step 6... For cluster targets, the state estimate and state estimate error covariance of the confirmed target are calculated and confirmed using the cluster center-intra-cluster target filtering method; otherwise, proceed to step 8. The specific steps of the cluster center-intra-cluster target filtering method are as follows: 7.1 Splitting a cluster into sub-clusters while preserving the cluster structure: If the cluster target is in Time to If the group structure remains unchanged, proceed to step 7.2; otherwise, split the cluster objective into sub-group objectives that maintain the same group structure, as follows: 7.1.1 If a cluster target is in Time to If it splits into multiple subgroups at a given time, then according to... The subgroup formed at each moment will The cluster target at time t is divided into corresponding subgroups, at which point... The corresponding subgroups obtained by time splitting The subgroup at each moment maintains the group structure unchanged; 7.1.2 If multiple cluster targets are in Time to If they are merged into a single cluster target at any given time, then according to... Multiple cluster targets at any given time will The large group formed at time is divided into corresponding subgroups, at which point... Cluster target at time and The subgroups obtained by splitting at each time step maintain the group structure unchanged; 7.1.3 If cluster splitting and merging occur simultaneously, the split cluster target is processed according to step 7.1.1, and the merged cluster target is processed according to step 7.1.2, to obtain the results. Time to A subgroup whose group structure remains unchanged at all times; 7.2 Group Center Filter Update: For each group center filter update... Time to For cluster targets or subgroups that maintain an unchanged group structure at all times, first calculate... The observation cluster center, which is the average of the observations within the observation cluster, is then used to update the data through a filtering method. Estimation of the cluster center state at time t; 7.3 Intra-group target status update: Calculate the intra-group target status update for each cluster target or sub-group. The difference between the state estimate at time t and the corresponding cluster center state estimate, assuming that the displacements of the intra-group targets and the cluster center remain constant, is calculated by... The cluster center state estimate at a given time is used to update the target within the cluster. State estimation at time; Step 8: Use cumulative hypotheses middle If there are no observations associated with the confirmed target at any given time, proceed... The connection between the new goals that started before the time and The process begins with the creation of a new target, followed by track confirmation and termination checks. Finally, the cumulative hypothesis is updated based on the track start, confirmation, and termination results. ; Step 9: Retain Moment -best cumulative assumption and targeting The cumulative assumption with the highest probability at time is used to calculate and output all confirmed targets based on the estimation criteria. State estimation at time step, state estimation error covariance, and group structure information; setting Return to step 1.

2. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 2, the state transition model is either a uniform linear motion model or a uniform turning motion model.

3. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 3, the gate technique is a rectangular gate or an ellipsoidal gate.

4. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 4.2, the two-dimensional allocation problem is solved using an allocation algorithm, specifically the Murty algorithm or the improved Jonker-Volgenant algorithm.

5. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 6.2, the graph search algorithm is the Tarjan algorithm based on depth-first search.

6. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 7.2, the filtering method is extended Kalman filtering or odorless filtering.

7. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 8, the execution of track confirmation and track termination judgment adopts the logical start method and M / N confirmation or termination criteria.

8. The cluster multi-hypothesis tracking method for jointly optimizing cluster partitioning and data association as described in claim 1, characterized in that, In step 9, the estimation criterion is either the minimum mean square error criterion or the maximum a posteriori criterion.

Images