Improved trajectory extraction multi-maneuvering target tracking method

Abstract

The invention discloses an improved trajectory extraction multi-maneuvering target tracking method, which comprises the following steps of: firstly, finishing long-time and continuous trajectory extraction of various targets including derivative targets, new targets and the like by adopting a labeled nearest neighbor trajectory management algorithm; then, introducing a multi-model thought, and proposing a multi-model Gaussian mixture probability hypothesis density filtering algorithm based on a jump Markov process; based on a tagged'nearest neighbor 'trajectory management algorithm and a multi-model Gaussian mixture probability hypothesis density filtering algorithm, a tagged jump Markov Gaussian mixture probability hypothesis density filtering algorithm is designed, and effective tracking and trajectory extraction of multiple maneuvering targets in a complex environment are realized. According to the method, the application scene of the Gaussian mixture probability hypothesis density filtering algorithm is expanded, the target trajectory can be accurately and effectively extracted in the face of a multi-target tracking scene including derivative target hypothesis, and the tracking precision under the complex condition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of space exploration and positioning, and particularly to an improved trajectory extraction multi-maneuvering target tracking method. Background Art

[0002] Target positioning and tracking technology is essentially a method for estimating the uncertain motion of a target. It uses one or more detection devices to obtain a series of observations of the target in the environment, and then estimates and predicts the motion states such as the position and velocity of the maneuvering target. Target positioning and tracking technology has extensive application requirements in the military field, such as precision guidance, ballistic missile defense, air defense warning, etc. Due to the increasing complexity of the existing target maneuverability and the increasing requirement for target tracking accuracy, the existing target positioning and tracking technology has encountered great challenges. How to achieve stable, reliable, and timely tracking of the target in the future complex environment and solve the problem of tracking maneuvering targets with continuously changing parameters and discrete jumps in system modes has become a new battlefield requirement.

[0003] The Gaussian Mixing Probability Hypothesis Density (GMPHD) filtering algorithm describes the multi-target state and measurement values using a random finite set, and adopts the multi-target Bayesian formula for recursion to achieve the purpose of dynamically estimating the motion states of multiple targets, solving the target tracking problem in a complex tracking environment, and having a series of advantages such as low computational complexity and high real-time performance. Its appearance has opened up a new direction for multi-target tracking. With the continuous enhancement of the target maneuverability and the aggravation of the system nonlinearity, the GMPHD filtering algorithm based on a single model is not ideal in terms of stability and tracking accuracy, and cannot explicitly distinguish different target trajectories. Although the labeled GMPHD filtering algorithm based on a single model can effectively distinguish different target trajectories, it usually assumes that the derived targets do not exist or ignores the generation of derived targets, which limits its application scenarios.

[0004] To properly solve the above problems, it is necessary to find a multi-maneuvering target tracking method that is applicable to various targets and has a wide range of application scenarios. Summary of the Invention

[0005] The purpose of the present invention is to provide an improved trajectory extraction multi-maneuvering target tracking method to achieve high-precision tracking and effective trajectory extraction of multiple maneuvering targets by multiple aircraft in a complex scenario.

[0006] The technical solution for achieving the purpose of the present invention is as follows: An improved trajectory extraction multi-maneuvering target tracking method includes the following steps:

[0007] S1. Adopt a labeled nearest neighbor trajectory management algorithm to complete the extraction of continuous-time and continuous trajectories of various targets including derived targets and newly generated targets;

[0008] S2. Introduce the idea of jump Markov multi-model, and propose a multi-model Gaussian mixture probability hypothesis density filtering algorithm based on the jump Markov process; design a labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm based on the labeled nearest neighbor trajectory management algorithm and the multi-model Gaussian mixture probability hypothesis density filtering algorithm to achieve effective tracking and trajectory extraction of multiple maneuvering targets in complex environments.

[0009] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above method are implemented.

[0010] A computer-readable storage medium stores a computer program, and when the program is executed by a processor, the steps of the above method are implemented.

[0011] A computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0013] (1) A labeled "nearest neighbor" trajectory management algorithm is proposed. By combining the labeling process with the "nearest neighbor", long-term and continuous trajectory extraction of various targets including derivative targets and newly born targets is completed, and the problem of trajectory discretization caused by filtering missed detections is effectively solved.

[0014] (2) The idea of jump Markov multi-model is introduced, and a multi-model Gaussian mixture probability hypothesis density filtering algorithm based on the jump Markov process is proposed to achieve stable tracking of maneuvering targets with variable motion patterns.

[0015] (3) Based on the labeled "nearest neighbor" trajectory management algorithm and the multi-model Gaussian mixture probability hypothesis density filtering algorithm, a labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm is designed to achieve effective tracking and trajectory extraction of multiple maneuvering targets in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is the flow chart of the labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm in the specific embodiment of the present invention.

[0017] Figure 2 is a schematic diagram of the coordinated turn model in the embodiment of the present invention.

[0018] Figure 3 is a three-dimensional schematic diagram of the target motion trajectory in the embodiment of the present invention.

[0019] Figure 4It is a schematic diagram of the XY plane of the target motion trajectory in the embodiment of the present invention.

[0020] Figure 5 It is a schematic diagram of the XZ plane of the target motion trajectory in the embodiment of the present invention.

[0021] Figure 6 It is the positioning and tracking scenario in the embodiment of the present invention.

[0022] Figure 7 It is a schematic diagram of the XY plane of the tracking trajectory in the embodiment of the present invention.

[0023] Figure 8 It is a schematic diagram of the XZ plane of the tracking trajectory in the embodiment of the present invention.

[0024] Figure 9 It is a three-dimensional image of the tracking trajectory in the embodiment of the present invention.

[0025] Figure 10 It is the XY plane filtering result of different algorithms in the embodiment of the present invention.

[0026] Figure 11 It is the XZ plane filtering result of different algorithms in the embodiment of the present invention.

[0027] Figure 12 It is the estimation result of the number of different algorithm targets in the embodiment of the present invention.

[0028] Figure 13 It is the comparison of the OSPA distance of different algorithms in the embodiment of the present invention. Specific implementation manner

[0029] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings.

[0030] The present invention discloses an improved trajectory extraction multi-maneuvering target tracking method. Aiming at the accurate tracking of multi-maneuvering targets in the complex environment of aircraft, the stability and tracking accuracy of the single motion model Gaussian mixture probability hypothesis density filtering algorithm are not ideal, and the problem of being unable to explicitly distinguish different target trajectories, a multi-model Gaussian mixture probability hypothesis density filtering algorithm with the labeled "nearest neighbor" trajectory management algorithm as the core is proposed to achieve the tracking of multi-maneuvering targets in complex environments. This method takes the labeled "nearest neighbor" trajectory management algorithm as the core of algorithm optimization, and it is mainly divided into two parts:

[0031] (1) Combining the advantages of the conventional labeled trajectory management algorithm and the "nearest neighbor" classification method, the labeled "nearest neighbor" trajectory management algorithm is proposed to complete the long-term and continuous trajectory extraction of various targets including derivative targets and newborn targets. This algorithm is designed for the state estimation process of the subsequent improved Gaussian mixture probability hypothesis filtering algorithm, and is used to obtain the multi-target estimation result from the filtered posterior parameters.

[0032] (2) Introduce the idea of jump Markov multi-model, form a multi-model Gaussian mixture probability hypothesis density filtering algorithm based on the jump Markov process, and based on the labeled "nearest neighbor" trajectory management algorithm and the multi-model Gaussian mixture probability hypothesis density filtering algorithm, form a labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm to achieve effective tracking and trajectory extraction of multiple maneuvering targets in complex environments.

[0033] The following is a detailed description of the specific implementation methods of the above steps:

[0034] Step 1: Labeled "nearest neighbor" trajectory management algorithm

[0035] In the hypothesis of the target random finite set by the conventional labeled GMPHD filtering algorithm, the existence of derivative targets is usually ignored or it is assumed that derivative targets do not exist. This kind of neglect often restricts the application scenarios of the labeled PHD filtering algorithm. In some scenarios, derivative targets cannot be ignored. To solve this problem and expand the application conditions of the labeled trajectory extraction algorithm, the original label management is improved to obtain the target state estimation and trajectory maintenance in the case of the existence of derivative targets. In addition, to solve the problems such as the easy omission of the designated label Gaussian term during the tracking process of high-maneuvering targets by the labeled GMPHD filtering algorithm, resulting in discontinuous trajectories and difficult trajectory maintenance, combined with the conventional "nearest neighbor" classification algorithm, a new labeled trajectory extraction method that can obtain long-time and continuous trajectories of various targets under high-dynamic motion conditions of the target is proposed.

[0036] Considering the hypothesis of derivative targets, the "nearest neighbor" labeled trajectory extraction process is described as follows:

[0037] In the initialization stage, assign labels to each Gaussian term to obtain the initial label set J0 represents the number of Gaussian terms at the initial moment, l i is the label information corresponding to Gaussian term i. Initialize the trajectory update set P Traj ={<p, l p >, n p}, l p =l i , p = i = 1,..., J0, that is, assign the trajectory p correspondence relationship to each Gaussian term label l i and save the mean value of the Gaussian term at this moment as the initial position of trajectory p in the set. n p in the set represents the number of times of filtering update of this trajectory. The initial number of times of trajectory update n p = 1. Initialize the set of missed tracking trajectories

[0038] In the prediction stage, the set of surviving target labels L in the prediction stage k-1 remains fixed, and a label is assigned to each newly born target j = 1, …, J γk , where J γk represents the number of newly born targets at time k. The set of target labels in the prediction stage is as follows:

[0039]

[0040] where is the set of labels for newly born targets, and L βk is the set of labels for derived targets; the number of label items after prediction (i.e., the number of Gaussian terms) is J k|k-1 .

[0041] In the update stage, using the set of observations Z k at the current time, for different observations z ∈ Z k in it, the mean of the same Gaussian term i = 1, …, J k|k-1 is updated. After the update, its label remains unchanged, that is, the label corresponding to the same Gaussian term is repeated (1 + |Z k |) times, and |·| represents the cardinality of the set.

[0042] Pruning and merging: After the update is completed, the set of labels corresponding to the updated Gaussian terms is denoted as L k . During the pruning and merging process, the label elements corresponding to different Gaussian terms change with the Gaussian terms. If the labels of two merged Gaussian terms are the same, then their same label is taken as the label of the merged Gaussian term. If they are different, then the label corresponding to the Gaussian term with the maximum weight is taken as the label of the new Gaussian term.

[0043] State estimation: The state estimation process uses the pruned Gaussian terms to obtain the multi-target state estimation value at this time. During the state extraction process, the original labeling management process, in order to ensure the uniqueness of the state quantity corresponding to the trajectory during trajectory extraction, only retains the Gaussian term with the maximum weight among the same labels. At this time, since the state of the derived target and its parent target that generated it belong to the same label and are discarded, resulting in incorrect estimation of the derived target, therefore, new label items need to be assigned to potential derived targets.

[0044] The improved state estimation process is as follows:

[0045] (1) Using the updated Gaussian term information L k to obtain the Gaussian term information that may be a target after filtering extraction. During the extraction process, it is sequentially determined whether the weight k of the Gaussian term corresponding to the elements in the updated label set L satisfies the extraction threshold w th , If satisfied, then Put the Gaussian term parameters into the target extraction label item set , where represents the number of targets determined by the Gaussian term weights. At this time, The Gaussian term labels are consistent, and their corresponding weights become the mean value All the label information corresponding to the target state quantity after extraction is saved in the extraction label set .

[0046] (2) Screen the Gaussian information corresponding to the extraction label set . For any element in , discuss it according to whether there are duplicate values.

[0047] If there are duplicate values in the set , then accumulate the Gaussian term weights of the labels in the set that are , and use the derivative threshold w sp to judge whether there are derivative targets. If the sum of the accumulated Gaussian term weights is greater than the derivative threshold w sp , then there are derivative targets. Keep the Gaussian term label with the largest weight unchanged, and consider the remaining Gaussian terms as potential derivative targets and reassign new labels; if the sum of the Gaussian term accumulation is not greater than the derivative threshold w sp , then only keep the Gaussian term label with the largest weight under this label . If there are no duplicate values in the set , keep it directly.

[0048] (3) After completing the above screening work, the Gaussian term information retained in the extraction label set is all the estimated information of the target state at time k obtained through filtering calculation, and each Gaussian term has a unique label. Use the above information to perform the trajectory update work at the current moment.

[0049] During the trajectory update process, give priority to updating and maintaining the existing trajectories. First, use the label information corresponding to the state estimation information that may be the target at the current moment obtained and the trajectory update set P Traj = {<p, l p >, n p} at the previous moment to update the corresponding target trajectory. After the update is completed, delete it from the extraction label set . During the update process, count the trajectory update set at the previous moment but not updated at the current moment as P u . If the trajectory update set P uThere are elements in it, indicating that there are errors in the trajectory corresponding to the target in the previous moment, such as new births or missed detections. A new birth error occurs when the weight of the Gaussian term of the original label corresponding to a highly maneuverable target is too small after the target is lost in the labeling evolution process, and the observation information of this target is assigned a new label after being judged as a newly born target, resulting in the situation that the same target has multiple labels. At this time, if the label set is extracted The remaining elements are 0, that is it indicates that the trajectory has been lost. If the label set is extracted and the remaining elements are not 0, further detection is required to determine whether the remaining elements in it are newly born targets or newly born targets with errors.

[0050] Use the "nearest neighbor" clustering algorithm to obtain the set P of unupdated trajectories u The set of state quantities of the trajectory at time k-1 in it p ∈ P u When extracting the label set The remaining elements correspond to the mean of the Gaussian term The "nearest neighbor" clustering results within the set. If its distance between the corresponding clustering results satisfies the label merging threshold U T , it is considered as a newly born target with errors, which is essentially a new estimated value generated after a maneuvering target is missed. Take the mean of the clustering Gaussian term at this time as the updated value of the trajectory p at this moment and modify the trajectory update set P Traj ={<p, l p >, n p} The label value l tracked by the trajectory p in p is the label value of the clustering result, which is used as the subsequent tracking result of the trajectory. If the clustering distance does not meet the requirement, it is considered that the trajectory is missed at this moment.

[0051] After the clustering extraction is completed, if there are still elements in the extracted label set , it indicates that newly born targets or derivative targets have appeared. For the remaining elements in the extracted label set , establish a new trajectory update relationship and establish a new trajectory. The new trajectory filtering update count n p is initialized to 0.

[0052] (4) After the above steps are completed, the state information estimated by the filter at the current moment has been properly allocated and processed. It is also necessary to maintain the trajectories of short-term missed targets. For this purpose, construct a set of missed trajectories What is saved is the trajectory number p. The set of missed trajectories The elements in it are composed of the trajectory labels updated at the previous moment but not updated at the current moment and the short-term missed trajectory labels with the missed detection time less than the maximum missed detection time threshold.

[0053] (5) To eliminate the discrete trajectory quantity caused by incorrect estimation in the above trajectory extraction step, reduce the data storage pressure, and supplement the process of deleting and screening overly short trajectories. Every K np filtering step lengths, update the trajectory set P Traj ={ <p, l p >, n p} in which, the filtering update step number n p less than is deleted from the trajectory tracking set.

[0054] Step 2: Labeled Jump Markov Gaussian Mixture Probability Hypothesis Density Filtering

[0055] The labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm includes six steps: initialization, prediction, update, state pruning, state extraction, and newborn target creation. For the convenience of writing formulas, let r = JMS(r', Π) represent the jump Markov process decision-making process, where r is the motion mode at the current moment, r' is the motion mode at the previous moment, and Π is the one-step probability transition matrix. Figure 1 shows the flow of the labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm, where the processes with "*" represent the main modified parts of the present invention.

[0056] (1) Initialization

[0057] Set the filtering parameters of each initial Gaussian term and its corresponding label information according to empirical values, and establish trajectory information for each initial Gaussian term. Denote the filter initialization parameters as i = 1, …, J0, where, are the weights of the initial Gaussian terms, the means of the Gaussian terms, the process noise covariance matrix, the motion mode, the Gaussian term labels, and J0 is the number of initial Gaussian terms in sequence.

[0058] (2) Prediction

[0059] Denote the prediction parameters at time k as i = 1, …, J k|k-1 , where, are the weights of the Gaussian terms to be predicted, the means of the Gaussian terms, the process noise covariance matrix, the motion mode, the Gaussian term labels, and J k|k-1 = 0 at the initial moment. The parameters related to the filter state at time k - 1 are j = 1, …, J k-1 , where, J k-1 are the weights of the Gaussian terms at the previous moment, the means of the Gaussian terms, the process noise covariance matrix, the motion mode, the Gaussian term labels, and the number of Gaussian terms in sequence.

[0060] First, perform the newborn target prediction process. Denote the state parameters of the newborn target at time k as j = 1, …, J γ,k , where J γ,k are the weights of the Gaussian terms of the newborn target, the means of the Gaussian terms, the process noise covariance matrix, the motion mode, the labels of the Gaussian terms, and the number of Gaussian terms, respectively. Since there is no motion mode information of the previous time for the newborn target, set the motion mode of the newborn target to the default value. Denote as the label set of the newborn target. The newborn target prediction process satisfies the following formula:

[0061]

[0062] After the newborn target prediction is completed, J k|k-1 = J γ,k .

[0063] Subsequently, predict and update the state parameters of the derivative target. The state parameters of the derivative target at time k - 1 are j = 1, …, J β,k , where J β,k is the number of derivative targets at time k. The target prediction parameters generated by the derivative target are not only related to the derivative target but also related to the surviving parent targets. For the derivative target, the birth potential motion mode of the derivative target is the same as that of the parent target, and the motion state is the same as that of the parent target. However, the motion mode at the prediction time of the derivative target follows an independent JMS process. To maintain the independence of the target labels, new labels are re - assigned to each derivative target during the derivative target prediction process. Denote as the newly assigned label of the Gaussian term corresponding to the derivative target.

[0064] After the derivative target is updated, its state parameters are also incorporated into the prediction parameters. Its update process is:

[0065]

[0066] In the formula, π k|k-1 represents the motion mode probability transition matrix of the derivative target; represents the probability that the derivative target jumps from the motion mode to during the prediction process; is the correction vector of the derivative target state transition process; is the motion state transition matrix of the derivative target; is the noise covariance matrix of the derivative target state transition process; is the system noise covariance matrix of the corresponding parent target of the derivative target; J β,k is the number of derivative targets during the prediction process; J k-1is the total target quantity at the previous moment;

[0067] After the derivative target prediction, J k|k-1 = J γ,k + J β,k × J k-1 .

[0068] Finally, the previous step's surviving target is predicted according to the surviving target prediction formula. According to the Jump Markov multi-target model assumption, in addition to considering the survival probability p S,k , the state transition probability t k|k-1 (r|r') also needs to be considered. After the surviving target is updated, its state parameters are also incorporated into the prediction parameters. After the surviving target is predicted, J k|k-1 = J γ,k + J β,k × J k-1 + J k-1 . The surviving target prediction formula is:

[0069]

[0070] where t k|k-1 represents the probability transfer matrix of the surviving target motion mode; is the probability from motion mode to motion mode ; p S,k is the target survival probability; is the state transition matrix of the surviving target under the current motion mode; Q k-1 is the process noise covariance matrix of the surviving target;

[0071] During the prediction process, the label set prediction formula is:

[0072] L k|k-1 = L k-1 ∪ L γ,k ∪ L β,k .

[0073] (3) Update

[0074] Suppose the parameters after the k - moment filter prediction are i = 1, …, J k|k-1 , the set of observed quantities is Z k . After introducing the motion mode state quantity, the parameter update process formula is:

[0075]

[0076] where are the innovation of the corresponding Gaussian term, innovation covariance matrix, Gaussian term gain matrix, and process noise matrix; is the measurement matrix under the corresponding motion mode; is the measurement noise covariance matrix; after the parameter update is completed, the sets of various parameters are denoted as j = 1, …, J k|k-1 .

[0077] Let the set of posterior parameters of the filter after the state update at time k be i = 1, …, J k , where J k are the weights of the Gaussian terms, the means of the Gaussian terms, the process noise covariance matrix, the motion mode, the labels of the Gaussian terms, and the number of Gaussian terms after the update, respectively. Before the update, the number of posterior parameters J k = 0. In the state quantity update process, the weights of the predicted Gaussian terms are first scaled, that is:

[0078]

[0079] In the formula, p D,k is the probability that the target is detected, that is, the detection probability; after the weight scaling is completed, J k = J k|k-1 .

[0080] Subsequently, the posterior intensity of the target is updated using the measurement information , where z( l ), l = 1, …, J z represents the observation information of a single target in the target group, and J z represents the number of target observation information obtained by the cluster. In the update process, the motion mode corresponds to the Gaussian terms one by one and does not change with the observation information. Taking the update of a single element z k in Z (l) , l = 1, …, J z as an example, the update formula is:

[0081]

[0082] Among them, represents the Gaussian likelihood density of the observation vector z (l) with a mean of and a covariance of ;

[0083] Considering the environmental clutter interference, assuming that the environmental clutter density satisfies κ k (·) function, then the weights of the updated Gaussian terms need to be processed, that is:

[0084]

[0085] Among them, κ k (z (l) ) is the Gaussian likelihood density at the observation vector z (l)The environmental clutter density at; successively perform the above update process on the elements in Z k and finally complete the filter update step. At this time, the posterior parameter set is i = 1, …, J k , where J k = J z × J k|k-1 + J k|k-1 .

[0086] (4) State pruning

[0087] To limit the number of Gaussian components propagated by the filter over time and prevent its unlimited growth, prune the updated Gaussian terms. Prune the Gaussian term weights of the smaller part to obtain an approximate solution of the algorithm. The pruning process is carried out by discarding the components with weights lower than a preset threshold and merging the information of Gaussian terms with closer mean distances.

[0088] During the pruning process, if the distance between the means of two Gaussian terms is less than the merging threshold, merge them to generate new Gaussian term information. During the merging process, if the labels of the two Gaussian terms to be merged are the same, take their same label as the label of the merged Gaussian term; if different, take the label corresponding to the Gaussian term with the largest weight as the label of the new Gaussian term.

[0089] In the state pruning step of the labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm, due to the introduction of a new state variable r, it is necessary to add a pruning step for the motion mode r. The pruning principle is that the motion mode of the target Gaussian term after pruning and merging is the motion mode of the target Gaussian term with the largest weight. The target state parameters after state pruning are denoted as i = 1, …, l. The state pruning process is shown in Table 1.

[0090] Table 1 State pruning algorithm

[0091]

[0092] (5) State estimation

[0093] Refer to the labeled "nearest neighbor" trajectory management algorithm designed in step 1 to design the state estimation process. The state estimation process uses the Gaussian terms after filter update to obtain each target state quantity and the target motion trajectory at the current moment. The process is shown in Table 2.

[0094] Table 2 State extraction algorithm

[0095]

[0096]

[0097]

[0098] The present invention will be described in detail below in conjunction with embodiments.

[0099] Embodiment

[0100] Based on numerical simulation to verify the contents of the claims of the present invention, the following designed experiments are carried out for computational simulation to fully demonstrate the method flow and the universality of the law.

[0101] Design an experimental scenario where two aircraft circle at the same altitude in a cooperative positioning scenario. The initial parameters of aircraft 1 are X r1 = [-100m, -600m, 600m, 40m / s, 0m / s, 0m / s] T , and the initial parameters of aircraft 2 are X r2 = [-100m, -910m, 800m, 62.5m / s, 0m / s, 0m / s] T . The cruising radius of aircraft 1 is 630m, and the cruising radius of aircraft 2 is 1000m. The distributed tracking system takes aircraft 1 as the research aircraft to carry out subsequent experiments. Assume that each aircraft carries a strapdown inertial navigation system and a satellite navigation system for its own navigation and positioning. Each aircraft in the aircraft group is equipped with the same sensors, and the real-time position information of the aircraft group is simulated by constructing a simple navigation simulation system. The surveillance area of the two aircraft is S = {(x, y, z)}, where x ∈ [-2500m, 2500m], y ∈ [-2500m, 2500m], and z ∈ [0, 500m].

[0102] To properly estimate the target motion process, the target state vector is defined as:

[0103]

[0104] Select the uniform motion, uniformly accelerated motion, and coordinated turn models as the multi-models of the jump Markov process, which are the motion modes r = 1, 2, 3 in sequence. Their motion equations are as follows.

[0105] The uniform motion model is used to describe the uniform motion of the target with zero acceleration. The uniform motion model takes the target position and velocity as the state, regards the target acceleration as a random perturbation input, and uses Gaussian white noise with a mean of zero for modeling. Assume the target state quantity is Then the continuous-time equation is:

[0106]

[0107] where w(t) is the system noise term; F is the continuous-time state transition matrix; G is the continuous-time noise drive matrix.

[0108] The uniform acceleration model regards the motion state of the target as a constant acceleration motion, and its motion state is described by position, velocity, and acceleration. Considering the actual tracking environment, the change in the target's acceleration cannot be zero, so the change in the target's acceleration is set to a random perturbation that satisfies Gaussian white noise. Let the target state quantity be Then the continuous-time equation is:

[0109]

[0110] where, F is the continuous-time state transition matrix; G is the continuous-time noise driving matrix.

[0111] The coordinated turn model assumes that the target moves in a circular motion in a plane at a constant rate, that is, the velocity and acceleration are perpendicular. Its motion schematic is as Figure 2 shown. Let the target state quantity Then the continuous-time equation is:

[0112]

[0113] where, F is the continuous-time state transition matrix; G is the continuous-time noise driving matrix.

[0114] The switching between motion modes is given by the Markov probability transition matrix, that is:

[0115]

[0116] For simplicity, assume that in each motion mode, the target survival probability p S,k|k-1 is 0.9 for all. Secondly, ignoring the change in the target detection probability caused by sensor characteristics, signal interference, etc. during the angle detection process, assume that in each motion mode, the target detection probability p D,k is a constant value, and the detection probability of each aircraft is 0.98, then p D,k = 0.9604.

[0117] During the tracking process, the aircraft cluster directly observes the position information Z of the target group as the filter input, denoted as z = [x (l) , y (l) , z (l) ∈ Z, l = 1, …, |Z|. Let the filter measurement equation be:

[0118] z = H(r)X + v r , v r ~ N(0, R k (r))

[0119] where, H(r) is the measurement matrix in motion mode r; X is the target motion state quantity; v r is the system observation noise, satisfying a mean of 0 and a covariance of R kGaussian white noise of (r).

[0120] Then the system observation matrix and the system measurement noise covariance matrix satisfy:

[0121]

[0122] R(r = 1, 2) = diag([4900m, 4900m, 10000m])

[0123] R(r = 3) = diag([10000m, 10000m, 10000m])

[0124] Considering the characteristics of multi - interference in the combat scenario, 5 randomly - appearing interference targets are set, and their positions at each moment appear randomly. The interference targets are used to describe the random interferences in the environment that may provide incorrect angle measurement observations, such as smoke and dust reflection, etc. The position distribution of the random interference targets at any moment conforms to the random distribution within the surveillance area S = {(x, y, z)}.

[0125] A total of 5 tracking targets are set in the full scenario, including 2 long - term existing targets, 2 derivative targets, and 1 newly - born target. Among them, target 2 is the derivative target of target 1 at 20s, and target 5 is the derivative target of target 4 at 40s. To simulate the scenario where the tracked target scatters bait, after derivation, the derivative target moves in the opposite direction to the parent target and gradually moves away from the parent target. Target 3 is the newly - born target. The speed and position of each target are continuous during the movement process, but the acceleration is not continuous. The target parameters are set as follows:

[0126] (1) Target 1

[0127] Set its initial position as (1000m, 500m, 0m), initial velocity as (0m / s, 0m / s, 5m / s), and the initial acceleration is 0. The movement process of target 1 is divided into three stages: from 0s to 30s is a uniformly - accelerated motion, and the acceleration during this process is (-2m / s 2 , 2m / s 2 , 0m / s 2 ); from 30s to 80s is a coordinated turning process, and its combined acceleration is 6m / s 2 ; from 80s to 100s is a uniform motion.

[0128] (2) Target 2

[0129] Target 2 is the derivative target of target 1, and its derivation time is 20s. After derivation, the movement process of target 2 is divided into three stages: from 20s to 35s is a quasi - coordinated turning motion. During this stage, its acceleration is composed of the coordinated turning acceleration with a combined acceleration of 10m / s 2 and the acceleration of 1m / s in the Z - axis direction 2The additional acceleration component, whose acceleration and velocity vector do not satisfy the coordinated turn assumption; from 35 s to 55 s is a uniformly accelerated motion with an acceleration of (-2 m / s 2 , -1 m / s 2 , 0 m / s 2 ); from 55 s to 80 s is a coordinated turn motion process, and its resultant acceleration is 5 m / s 2 . Setting a quasi-coordinated turn motion for the first stage of Target 2 is to explore whether the jump Markov process can effectively envelope the situations that do not belong to the set motion model.

[0130] (3) Target 3

[0131] Target 3 is a newly born target with its initial position (-2000 m, -500 m, 300 m), and both its initial velocity and initial acceleration are 0. The new birth time of Target 2 is 5 s. From 5 s to 35 s is a uniformly accelerated motion with an acceleration of (2 m / s 2 , 1 m / s 2 , 0.1 m / s 2 ); from 35 s to 50 s is a uniform motion; from 50 s to 70 s is a coordinated turn process, and its resultant velocity is 5 m / s 2 .

[0132] (4) Target 4

[0133] The initial position is (-1000 m, 500 m, 450 m), the initial velocity is (15 m / s, -20 m / s, 0 m / s), and the initial acceleration is zero. Target 4 maintains a uniform linear motion from 0 s to 40 s, and is a uniformly accelerated motion from 40 s to 70 s with an acceleration of (0.5 m / s 2 , -0.5 m / s 2 , -0.25 m / s 2 ), and from 70 s to 100 s is still a uniformly accelerated motion, but the acceleration is modified to (2 m / s 2 , -2 m / s 2 , 1 m / s 2 ).

[0134] (5) Target 5

[0135] Target 5 is a derivative target of Target 4, and the derivative time is 40 s. Its motion is divided into two stages: from 40 s to 65 s is a coordinated turn motion, and its resultant acceleration is 1 m / s 2 , and from 65 s to 100 s is a uniform motion.

[0136] Figures 3 to 5 They are respectively the three-dimensional schematic diagram of the target motion trajectory, the XY-direction schematic diagram of the target motion trajectory, and the XZ-direction schematic diagram of the target motion trajectory.

[0137] During the simulation process, the tracking target and the interference target together constitute the target set. To simulate the angle measurement values of each aircraft for the target, the same method as described in Experiment 4.4.2 is used to generate the true angle measurement values for each aircraft, and Gaussian white noise with a mean of zero and a variance of σ a = 0.1° is added as the angle measurement value. Figure 6 Shows the hovering trajectories of each aircraft and the true target positions in the scene, as well as the position observation results of the aircraft cluster.

[0138] It is assumed that the observation clutter intensity function satisfies

[0139]

[0140] where represents the regional uniform distribution density, V represents the regional volume, and λ c represents the clutter density. In the simulation, the regional volume is V = 2.5×10 10 m 3 , and the clutter density λ c = 4×10 -10 m -3 . z represents the position observation vector.

[0141] The intensity function of the newly born target is

[0142]

[0143] where π k (r) is the motion mode distribution of the newly born target; m γ is the newly born position; P γ is the covariance matrix of the newly born process noise, and the parameter settings are as follows:

[0144] m γ = [-2000, -500, 300, 0, 0, 0, 0, 0, 0] T

[0145] P γ = diag([2500, 2500, 2500, 25, 25, 25, 4, 4, 4])

[0146] [π k (r)] = [0.95 0.05 0.05].

[0147] In addition, an intensity function is established for the derived target, and the intensity function of the derived target is:

[0148]

[0149] where ξ', r' are the motion state quantities and motion modes of the parent target; ξ, r are the motion state quantities and motion modes of the derived target; Qβ is the noise covariance matrix; π k|k-1 (r|r') is the derivative target mode transition matrix; the noise covariance matrix and the mode transition matrix are set as follows:

[0150] Q β = diag([2500, 2500, 2500, 25, 25, 25, 1, 1, 1])

[0151]

[0152] Set the simulation step size to 1 s, the simulation time to 100 s, and the pruning weight threshold T = 10 -5 , the target merging threshold U = 1, the maximum number of Gaussian components J max = 100, the state extraction threshold E = 0.5, the derivative target judgment threshold w sp = 1.5, the clustering merging threshold The maximum missed detection time threshold K miss = 5, the interference trajectory deletion threshold K np = 100, the minimum number of trajectory updates The present invention uses the OSPA distance to evaluate the tracking performance of the algorithm. During the experiment, the OSPA parameters p = 1 and c = 200 m are selected.

[0153] The simulation results are as follows: Figure 7 、 Figure 8 shows the target trajectory results tracked by the algorithm of the present invention.

[0154] Tracking trajectory 1 and tracking trajectory 4 are the actual trajectory tracking results of target 1 and its derivative target 2. From the tracking results, the process of "nearest neighbor" labeling well distinguishes the derivative target and the parent target. At the initial stage of tracking trajectory 4, due to the approximate trajectories of the derivative target and the parent target, there is partial overlap between tracking trajectory 1 and tracking trajectory 4, and the algorithm well tracks this process, which indicates that the algorithm still maintains good trajectory resolution ability when the target trajectories are close. Since the observation quantity does not identify which target the observation comes from, the trajectories of the derivative target and its parent target by the algorithm are not exactly in accordance with the simulation setting. From the image, the trajectory marked by tracking trajectory 4 is more like the derivative target trajectory. It should be noted that tracking trajectory 1 maintains a good tracking effect during the middle section class coordination turning process, verifying that the jump Markov process can effectively envelope the motion outside the motion model set. However, the end trajectory diverges along a straight line, which is the result of the short-term trajectory extrapolation by the algorithm after target 2 disappears. The design of targets 1 and 2 and their tracking results show that the algorithm has good tracking effects for targets with large maneuvers, such as turning in circles and S-turns.

[0155] Tracking trajectory 2 and tracking trajectory 5 are the tracking results of target 4 and its derivative target 5. The rear end of the trajectory of target 4 is designed for large maneuvers in the height direction, and target 5 is designed for constant-height planar maneuvers. The tracking effect of the above target maneuver mode verification algorithm in the height direction is verified. The results show that the algorithm can effectively track the maneuvers of the target at different heights.

[0156] Tracking trajectory 3 is the tracking result of the new target 3. In the front section of the trajectory tracking, since the initial motion model of the new target is defaulted to uniform motion, while the actual motion of target 3 is uniformly accelerated motion, the tracking trajectory effect is not good. With the accumulation of filtering, the estimation of the target state quantity becomes gradually accurate, and the effect of the middle and rear sections of tracking trajectory 3 is better. Figure 9 It is the three-dimensional tracking result image of the target tracking trajectory.

[0157] Figures 10 to 13 It is the comparison result between different filtering algorithms in the same tracking scenario. Figure 10 、 Figure 11 It is the comparison of the filtering state quantities of different algorithms. When derivative targets 2 and 5 appear, the LGMPHD algorithm has missed detections. This is because in the state estimation process, only the Gaussian term with the largest weight is taken as the single label estimation result, which causes the algorithm to be unable to estimate the state quantities of the derivative target and the parent target simultaneously. Compared with the algorithm of the present invention and the WSP-LGMPHD filtering algorithm, the LGMPHD has the worst effect on target state estimation. And compared with the algorithm of the present invention, the WSP-LGMPHD filtering algorithm has poor estimation accuracy in the XZ plane position. Figure 12 In the target quantity estimation image, the algorithm of the present invention tracks the change of the target quantity more timely, and the WSP-LGMPHD filtering algorithm is slower in tracking the change of the target quantity, especially at the moment of target derivation. This is because the WSP-LGMPHD filtering algorithm only considers the situation where the same labels are extracted simultaneously in the state estimation process, while ignoring the situation where the Gaussian terms are merged when the initial motion stage of the derivative target is close to the position of the parent target. The algorithm of the present invention compensates for this defect. Generally speaking, the LGMPHD algorithm has the worst target quantity estimation effect. Figure 13 The OSPA error distances between the algorithm of the present invention, the WSP-LGMPHD algorithm and the LGMPHD algorithm are given. In the initial stage of filtering, the three algorithms all have small OSPA distances. At 20 s, due to the appearance of derivative targets, the LGMPHD algorithm misses detections of the derivative targets, resulting in its OSPA distance being much larger than that of the algorithm of the present invention. At 40 s, due to the defect that the WSP-LGMPHD filtering algorithm estimates the number of targets inaccurately in the initial stage when derivative targets appear, its OSPA distance is larger than that of the algorithm of the present invention.

[0158] Table 3 shows the average single - operation time of three algorithms in multiple trials. In the simulation environment where the calculation step of the algorithm of the present invention is 1 s, its operation time only accounts for 3%, and the increase in operation time compared with the conventional algorithm is small, meeting the real - time requirement of the algorithm.

[0159] Table 3 Single - operation time of different algorithms

[0160]

[0161] According to the above comparison, from the whole process of tracking, the algorithm of the present invention has a better target continuous tracking effect and can meet the requirements of multi - maneuvering target tracking.

Claims

1. An improved trajectory extraction multi-maneuvering target tracking method, characterized in that: Includes steps: S1. Use the labeled nearest neighbor trajectory management algorithm to complete the extraction of continuous time and continuous trajectories of various targets including derived targets and new targets; S2. Introduce the idea of ​​jump Markov multi-model and propose a multi-model Gaussian mixture probability hypothesis density filtering algorithm based on jump Markov process; based on the labeled nearest neighbor trajectory management algorithm and the multi-model Gaussian mixture probability hypothesis density filtering algorithm, design a labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm to achieve effective tracking and trajectory extraction of multiple maneuvering targets in complex environments.

2. The improved trajectory extraction multi-maneuvering target tracking method according to claim 1 is characterized in that: In step S1, the labeled nearest neighbor trajectory management algorithm specifically includes: Considering the derived target hypothesis, the nearest neighbor labeled trajectory extraction process is described as follows: Step 1: Initialization phase: assign a label to each Gaussian item to obtain the initial label set J0 represents the number of Gaussian terms at the initial moment, l i is the label information corresponding to Gaussian item i; initialize the trajectory update set P Traj ={ <p,l p >,n p },l p = l i ,p=i=1,…,J0, that is, each Gaussian item label l i Assign the corresponding relationship of trajectory p and take the mean of Gaussian term at that moment Save as trajectory p initial moment position n in the collection p Indicates the number of times the trajectory filter is updated, the number of initial trajectory updates n p =1; Initialize the missing track set Step 2: Prediction phase: the set of surviving target labels L in the prediction phase k-1 Fixed, assign a label to each new target Among them J γk represents the number of new targets at time k, and the target label set in the prediction stage is: in, is the new target label set; L βk is the derived target label set; the number of predicted label items, i.e. the number of Gaussian items, is J k|k-1 ; Step 3: Update phase, using the current observation set Z k Different observations z∈Z k For the same Gaussian term, the mean Update, after updating its label unchanged, that is, the same Gaussian term corresponds to repeated labels (1+|Z k |) times, |·| represents the cardinality of the set; Step 4: Pruning and merging. After the update is completed, update the label set corresponding to the Gaussian item to L k , the label elements corresponding to different Gaussian items change with the Gaussian items during pruning and merging; if the labels of the two merged Gaussian items are the same, then their same labels are taken as the labels of the merged Gaussian items; if they are different, then the label corresponding to the Gaussian item with the largest weight is taken as the label of the new Gaussian item; Step 5: State estimation. The state estimation process uses the pruned Gaussian term to obtain the estimated value of the multi-target state at that moment. In the state extraction process, the original label management process only retains the Gaussian term with the largest weight in the same label to ensure the uniqueness of the state quantity corresponding to the trajectory during trajectory extraction. At this time, the derived target state is discarded because it has the same label as the parent target that generates it, resulting in an error in the derived target estimation. Therefore, new label items need to be assigned to potential derived targets. The improved state estimation process is as follows: (1) Using the updated Gaussian information L k Get the Gaussian item information that may be the target after filtering and extraction, and judge and update the label set L in turn during the extraction process k Medium Element Corresponding Gaussian item weight Whether the extraction threshold w is met th , If satisfied, then Gaussian item parameters are put into the target extraction label item set Among them Represents the number of targets determined by the Gaussian item weights. The labels of the Gaussian items remain consistent, and their corresponding weights become the mean After extraction, the corresponding labels of the target state information are saved in the extracted label set. middle; (2) Extracting label sets The corresponding Gaussian information is filtered. Any element in The discussion is divided into cases according to whether there are repeated values; If it is in the collection If there are duplicate values ​​in the set The label is The Gaussian item weights are accumulated and the derived threshold w is used sp Determine whether there is a derived target. If the sum of the Gaussian item weights is greater than the derivative threshold w sp , then there is a derived target, the Gaussian item with the largest weight is retained unchanged, and the remaining Gaussian items are considered potential derived targets and new labels are assigned; if the cumulative sum of the Gaussian items is not greater than the derivative threshold w sp , then only keep this tag The Gaussian item label with the maximum weight under There are no duplicate values ​​in the , so they are retained directly; (3) After completing the above screening work, extract the label set The Gaussian item information retained in is all the estimated information of the target state at time k obtained by filtering calculation. Each Gaussian item has a unique label, and the above information is used to update the trajectory at the current moment; During the trajectory update process, the existing trajectory is updated and maintained first; first, the label information corresponding to the state estimation information of the current moment that may be the target is used to update the set P of trajectories at the previous moment. Traj ={ <p,l p >,n p } Update the corresponding target trajectory, and extract it from the label set after the update is completed In the updating process, the trajectory set that was updated at the last moment but not updated at the current moment is counted as P u ; If the trajectory set P is not updated u If there are elements in , it means that the target corresponding to the trajectory at the previous moment has an error or missed detection; at this time, if the label set is extracted The remaining elements are 0, that is This indicates that the track is missed; if the label set is extracted If the remaining elements are not 0, further testing is required The remaining elements in are new targets or wrong new targets; Use the nearest neighbor clustering algorithm to obtain the unupdated trajectory set P u The state set of the trajectory at time k-1 p∈P u In extracting the tag set Remaining elements Corresponding Gaussian term mean The nearest neighbor clustering results within the set, If The distance between the corresponding clustering results meets the label merging threshold U T , it is considered to be an erroneous new target; the mean of the clustered Gaussian term at this time is used as the updated value of the trajectory p at this moment And modify the trajectory update set P Traj ={ <p,l p >,n p }The label value l tracked by trajectory p p is the clustering result label value, which is used as the subsequent tracking result of the trajectory. If the clustering distance is not satisfied, the trajectory is considered to be missed at this moment; After cluster extraction is completed, if the label set is extracted If there are still elements in , it indicates that a new target or a derived target has appeared, which is to extract the label set The remaining elements in establish a new trajectory update relationship and establish a new trajectory. The new trajectory filter updates the count n p Initialized to 0; (4) After the above steps are completed, the state information estimated by the filter at the current moment is distributed and processed, and the trajectory of the short-term missed target needs to be maintained; for this purpose, a missed track set is constructed. What is saved is the trajectory number p; the missed trajectory set The middle element is composed of the trajectory labels that were updated at the last moment but not updated at the current moment and the short-term missed tracking trajectory labels whose missed detection time is less than the maximum missed detection time threshold; (5) In order to eliminate the discrete trajectory caused by the wrong estimation in the above trajectory extraction step, the short trajectory deletion and screening process is supplemented; every K np The filter step size updates the trajectory set P Traj ={ <p,l p >,n p }, the number of filter update steps n p Less than The trajectory tracking collection is deleted.

3. The improved trajectory extraction multi-target tracking method according to claim 2 is characterized in that: In step S2, the labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm specifically includes: The labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm includes six steps: initialization, prediction, update, state pruning, state extraction, and new target creation. The jump Markov process decision process is represented by r = JMS (r', Π), where r is the current motion mode, r' is the previous motion mode, and Π is the one-step probability transfer matrix. Step 1: Initialization According to the empirical value, the filter parameters of each initialized Gaussian item and its corresponding label information are set, and the trajectory information is established for each initialized Gaussian item. The filter initialization parameters are recorded as in, The following are the initialization of Gaussian item weights, Gaussian item mean, process noise covariance matrix, motion mode, Gaussian item label, and J0 initialization of the number of Gaussian items; Step 2: Prediction The prediction parameter at time k is in, They are the weight of the Gaussian item to be predicted, the mean of the Gaussian item, the process noise covariance matrix, the motion mode, the Gaussian item label, and the initial time J. k|k-1 =0; the filter state related parameters at time k-1 are in, J k-1 They are the Gaussian item weight at the previous moment, the Gaussian item mean, the process noise covariance matrix, the motion mode, the Gaussian item label, and the number of Gaussian items; First, the new target prediction process is carried out, and the new target state parameter at time k is recorded as in, J γ,k They are the Gaussian item weight of the new target, the mean of the Gaussian item, the process noise covariance matrix, the motion mode, the Gaussian item label, and the number of Gaussian items. Since the new target has no motion mode information at the previous moment, the motion mode of the new target is set. is the default value; is the new target label set, and the new target prediction process satisfies the following formula: After the new target prediction is completed, J k|k-1 =J γ,k ; Then the derived target state parameters are predicted and updated. The derived target state parameters at time k-1 are: Among them J β,k is the number of derived targets at time k. The target prediction parameters generated by the derived targets are not only related to the derived targets, but also to the surviving parent targets. The motion mode of the derived targets at birth is consistent with that of the parent targets, and the motion state is consistent with that of the parent targets. However, the motion mode of the derived targets at the time of prediction follows an independent JMS process. In order to maintain the independence of the target labels, the derived target prediction process reallocates new labels to each derived target. New assigned labels for the Gaussian terms corresponding to the derived targets; After the derived target is updated, its state parameters are also incorporated into the prediction parameters. The update process is: In the formula, π k|k-1 represents the derived target motion pattern probability transfer matrix; Derived from the motion pattern during prediction Jump to The probability of Correction vector for the derived target state transition process; is the derived target motion state transfer matrix; is the noise covariance matrix of the derived target state transfer process; is the noise covariance matrix of the derived target corresponding to the parent target system; J β,k The number of targets derived from the prediction process; J k-1 is the total target quantity at the previous moment; After the derived target prediction, J k|k-1 =J γ,k +J β,k ×J k-1 ; Finally, the survival target of the previous step is predicted according to the survival target prediction formula. According to the JMS multi-objective model assumption, the survival target weight is considered in addition to the survival probability p. S,k In addition, we also need to consider the state transition probability t k|k-1 (r|r'); After the survival target is updated, its state parameters are also incorporated into the prediction parameters; After the survival target is predicted, J k|k-1 =J γ,k +J β,k ×J k-1 +J k-1 ; The survival target prediction formula is: Among them, t k|k-1 Represents the probability transfer matrix of the motion mode of the surviving target; From sport mode To Sport Mode The probability of S,k is the target survival probability; is the state transfer matrix of the surviving target in the current motion mode; Q k-1 is the noise covariance matrix of the survival target process; The label set prediction formula in the prediction process is: L k|k-1 =L k-1 ∪L γ,k ∪L β,k ; Step 3: Update Assume that the filter prediction parameters at time k are The set of observations is Z k , after introducing the motion mode state quantity, the parameter update process formula is: in, Corresponding to Gaussian item innovation, innovation covariance matrix, Gaussian item gain matrix, process noise matrix; is the measurement matrix in the corresponding motion mode; is the measurement noise covariance matrix; after the parameter update is completed, the various parameter sets are recorded as Suppose the filter posterior parameter set after the state is updated at time k is in, J k They are the weight of the updated Gaussian item, the mean of the Gaussian item, the process noise covariance matrix, the motion mode, the Gaussian item label, and the number of Gaussian items; before the update, the number of posterior parameters J k =0; the state quantity update process first scales the weight of the predicted Gaussian term, that is: In the formula, p D,k is the probability of the target being detected, that is, the detection probability; after the weight scaling is completed, J k =J k|k-1 ; Then use the cluster measurement information collection where z (l) ,l=1,…,J z represents the observation information of a single target in the target group, J z Indicates the number of target observation information obtained by the cluster; updates the target posterior strength. During the update process, the motion mode corresponds to the Gaussian term one by one and does not change with the observation information. k A single element z (l) ,l=1,…,J z Taking updating the prediction parameters as an example, the update formula is: in, Represents the observation vector z (l) At the mean The covariance is Gaussian likelihood density under ; Considering the environmental clutter interference, it is assumed that the environmental clutter density satisfies κ k (·) function, the weight of the updated Gaussian item needs to be processed, that is: Among them, κ k (z (l) ) is the observation vector z (l) The ambient clutter density at Z k The elements in the filter perform the above update process and finally complete the filter update step. At this time, the posterior parameter set is Among them J k =J z ×J k|k-1 +J k|k-1 ; Step 4: State Pruning Prune the updated Gaussian term; prune the Gaussian term weights The algorithm approximate solution is obtained by merging the parts below the preset threshold; in addition, the clipping process also prunes by merging the Gaussian items whose mean distance is less than the preset merging distance; During the pruning process, if the distance between the means of two Gaussian items is less than the merging threshold, they are merged to generate new Gaussian item information. During the merging process, if the labels of the two merged Gaussian items are the same, their same labels are taken as the merged Gaussian item labels. If they are different, the label corresponding to the Gaussian item with the largest weight is taken as the new Gaussian item label. In the state pruning step of the labeled JMS-GMPHD filter, due to the introduction of a new state variable r, it is necessary to add a pruning step for the motion mode r. The pruning principle is that the motion mode of the target Gaussian item after pruning and merging is the motion mode of the target Gaussian item with the maximum weight; the target state parameter after state pruning is recorded as Step 5: State Estimation The labeled nearest neighbor trajectory management algorithm is used to design the state estimation process; the state estimation process uses the Gaussian terms after filtering and updating to obtain the state quantity of each target and the target motion trajectory at the current moment.

4. The improved trajectory extraction multi-target tracking method according to claim 3 is characterized in that: The state estimation process of the labeled jump Markov Gaussian mixture probability hypothesis density filtering algorithm includes: Given k time pruning related parameters State extraction weight threshold E, previous moment trajectory label set The last moment missed track set Trajectory set P Traj ={ <p,l p >,n p }, clustering threshold U T ; Initialization: Update set at current moment The current missed detection set Step 1: Get the target information set to be extracted At the initial moment q=0; The elements are judged, if Then elements are assigned During the assignment process The rest of the information remains the same; Step 2: Extract target information sets one by one To filter the elements, if If there are duplicate values ​​in the set, calculate the weights and Then w l ≥w sp , then the label of the Gaussian item with the largest weight is retained unchanged, and the labels of the remaining Gaussian items are reallocated; w l <w sp , then only the Gaussian item information with the largest weight is retained, and the remaining Gaussian item information is deleted from the extraction target set. After the extraction is completed, each item in the set has a unique label; Step 3: Set the collection Update the existing trajectory information and find the trajectory update set P in turn Traj ={ <p,l p >,n p Each track p in the} corresponds to a label l p Extracting target information set If it exists, it will correspond to Gaussian information. As the current trajectory p position, n p =n p +1, put the track number p into and delete it from the collection; if it does not exist, the record is not updated and the track number p is put into middle; Step 4: If And the extraction target set is not empty, then extract The position set of the previous moment in the trajectory Matching using the "nearest neighbor" algorithm Mean distance of medium-Gaussian term The element closest to each element in the array, and determine whether the closest distance between the two is less than U T , if less than, The "nearest neighbor" element in the extracted information set corresponding to the element in is used as the current trajectory p position, and its label is updated n p =n p +1, put the track number p into and delete it from the extracted information set; if it is greater than, it is ignored; if Go to step 5; Step 5: If If it is still not empty, then establish new trajectory correspondences for the elements in turn and update the trajectory correspondence set P Traj ={ <p,l p >,n p }; Assign an initial value to the newly created trajectory and put the newly created trajectory label p into If it is empty, proceed to step 6; Step 6: Construct the missed track set, first check the last moment The flag bits corresponding to each element in Is the difference between the current time k and the maximum missed detection time threshold K? miss If it is less than Put in label set Secondly, update the label set of each trajectory one by one Check the elements in the collection, if they appear in the current trajectory update collection If it does not appear, it is placed in the missed track set. In order to maintain The track identification is unique, when it is put in, There is already a track identification flag n in miss No update is performed, if there is no label, the flag bit n is recorded miss = k; Finally, check Whether the element appears in If it appears, delete it to prevent repeated updates of the trajectory; Step 7: Use the state update equation and the trajectory of the previous moment to update the corresponding trajectory in the missed trajectory set. When updating, the motion state transfer matrix is ​​still selected using the jump Markov process; Output the extracted trajectory information, P Traj ={ <p,l p >,n p }, missed track set Track Update Set 5. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 4 are implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images