A multi-object tracking method and system based on a cross-Transformer data association algorithm

Abstract

The present invention provides a multi-target tracking method and system based on a cross-Transformer data association algorithm, belonging to the field of artificial intelligence. To solve the problems that the existing traditional MTT algorithm based on radar data has high storage requirements and slow calculation speed, the learning-based algorithms need to constrain the number of targets, and the network design does not consider the typical features of position data, and it is impossible to distinguish between two crossing trajectories, which may result in missed detections or false detections. The present invention uses a cross-Transformer data association network and adopts a BLUE filter to obtain the best estimates in terms of the accuracy and efficiency of radar-based position measurements; in addition, a score-based trajectory management strategy is proposed to based on relevant historical measurements; Simulation experiments on various datasets show that the association and tracking methods proposed by the present invention are superior to existing methods in complex multi-target tracking scenarios, including trajectory crossing, a large number of false alarms and missed detections.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and more particularly, to a multi-object tracking method and system based on a cross-Transformer data association algorithm. Background Art

[0002] Multi-object tracking (MTT) or multi-object tracking (MOT) plays a crucial role in various fields, such as surveillance, autonomous driving, and robotics. The MTT problem refers to associating the relative measurements obtained by a sensor over time with multiple moving objects (targets) and estimating their trajectories and other attributes (speed, acceleration, etc.). A key problem in MTT is to solve the unknown correspondence between targets and measurements, and this task is called data association. In scenarios with occlusion, sensor noise, clutter, and targets that may enter or leave the scene, the MTT task is highly challenging.

[0003] To address these challenges, many algorithms and methods for multi-object tracking have been proposed, which can generally be divided into maximum likelihood data association methods and Bayesian data association methods. The maximum likelihood data association method is based on the likelihood ratio of the observed data for association. Classic algorithms include the joint maximum likelihood algorithm, integer programming, generalized association methods, etc. The maximum likelihood data association method is not applicable to cases with crossing targets, maneuvering targets, and dense multi-object scenarios, thus limiting its application in practice. Bayesian data association methods are based on the Bayesian criterion and mainly include the nearest neighbor (NN) algorithm, probability data association (PDA) algorithm, joint probability data association (JPDA) algorithm, and multiple hypothesis tracking (MHT) method, etc. These methods, such as JPDA, require the assumption of a constant number of targets. The computational complexity of the MHT method increases rapidly with the increase in the number of targets. Another class of MTT methods is based on random finite sets (RFS), which uses multi-object conjugate priors to derive a closed expression for the multiple object posterior and obtains the Bayesian optimal estimate. Typical algorithms include the probability hypothesis density (PHD) filter, cardinalized probability hypothesis density (CPHD) filter, generalized labeled multi-Bernoulli (GLMB) tracker. Methods based on RFS can transmit more information between different time steps, thus improving the tracking results. However, these methods usually rely on prior information about the true posterior density. The approximation of the true posterior density may lead to inaccurate data association. In addition, due to the use of a high-dimensional sampling process, the computational and storage requirements of these methods have increased.

[0004] In recent years, the use of deep learning-based MTT algorithms has increased significantly, especially in object tracking tasks in the field of computer vision. These methods involve optimizing models with a large number of parametric features by minimizing the experience on a dataset targeted at solving specific problems. The tracking bypass detection algorithm has achieved the SOTA algorithm by using a high-performance feature extraction network and an efficient learning-based association method. However, the improvement in tracking performance comes at the cost of increased computation and model complexity. At the same time, the tracking and correlation accuracy of these methods largely depends on extracting features from high-dimensional image data, which limits their applicability in cases where only the target location is available. In recent years, learning-based algorithms have been increasingly applied to radar-based MTT problems. Compared with the high-dimensional information obtained from image-based data, radar data can only provide limited position measurement information, which may hinder the distinction between two closely spaced trajectories. The above challenging scenarios will make the result of data association worse. The long short-term memory (LSTM) network has been used in MTT for simple scenarios, where the LSTM network is specifically designed to learn the probability of measurement-to-track association from noisy radar measurements and pre-existing tracks. The above algorithms solve the multi-object matching problem considering missing detections and the multi-measurement matching problem of data association. However, the current learning-based algorithms applied to radar-based MTT problems require constraints on the number of targets, have high storage requirements, and slow computational speed. In addition, the network design does not take into account the specific data characteristics and is combined with the tracking algorithm. And the network design does not consider the typical characteristics of position data and cannot distinguish between two crossing trajectories, which may result in missed detections or false detections. Summary of the Invention

[0005] The technical problem to be solved by the present invention is:

[0006] To solve the problems of the existing traditional MTT algorithms based on radar data having high storage requirements and slow computational speed, the learning-based algorithms requiring constraints on the number of targets, and the network design not considering the typical characteristics of position data and being unable to distinguish between two crossing trajectories, which may result in missed detections or false detections.

[0007] The technical solution adopted by the present invention to solve the above technical problems:

[0008] The present invention provides a multi-object tracking method based on a cross Transformer data association algorithm, including the following steps:

[0009] S100. Establish a dynamic model and a general measurement equation for a series of maneuvering dynamic targets, consider relative distance and relative angle measurements, and transform the target tracking problem into a posterior estimation of the target state for data association in step S200;

[0010] S200. Establish the correlation between the current measurement and the trajectory based on the architecture of the cross-Transformer data association network;

[0011] S300. Maintain each trajectory in step S200 by using the optimal linear unbiased estimation filter and perform state updates;

[0012] S400. Process unknown targets to be tracked and targets to be tracked with varying quantities by using a scoring-based track management strategy, and determine the initialization, update, and deletion of the trajectories.

[0013] Further, in step S100, it includes:

[0014] Consider tracking a group of moving targets. Let the dynamic model of the i-th target (i ∈ {1, 2, …, N}) be

[0015] x i (k) = f(x i (k - 1), u i (k), η i (k)) (1)

[0016] where x i (k) and u i (k) are the state and input vectors at time k, and η i (k) is the process noise vector, assumed to be a zero-mean Gaussian process;

[0017] The general measurement equation is

[0018] z i (k) = h(x i (k), ν i (k)) (2)

[0019] where z i (k) is the measurement value, v i (k) is the observation noise, assumed to be a zero-mean Gaussian process;

[0020] Consider relative distance and relative angle measurements,

[0021]

[0022] where (x i (k), y i (k), z i (k) represents the position of target i at time k; r i (k) is the relative distance between the sensor and target i at time k, θ i (k) and φ i(k) are the azimuth angle and the elevation angle respectively; v i,r (k), v i,θ (k) and v i,φ (k) are respectively r i (k), θ i (k) and φ i (k) are the measurement noises of, and it is assumed that v i,r (k), v i,θ (k) and v i,φ (k) are all zero-mean Gaussian processes, that is and

[0023] Let X(k) = {x i (k)|i = 1, 2, …, N(k)} represent the states of all targets at k, where the x i (k) is the state of the i-th target, and N(k) is the number of targets; Let Z(k) = {z j (k)|j = 1, 2, …, M(k)} represent all the measurement values of the targets at k, where the z j (k) is the j-th measurement value, and M(k) is the number of measurement values; Let represent the set of all target states from the initial time to time step K, and represent the set of all measurement values, where K is the total number of time steps;

[0024] The multi-target tracking problem is formulated as obtaining a mapping as, transforming the target tracking problem into the posterior estimation of the target state,

[0025]

[0026] where, Θ represents the parameters of the MTT algorithm;

[0027] If new measurement results are obtained, the past tracking results will be retained;

[0028] Let Y(k) = {(y l (k), α l (k))|l = 1, 2, …, O(k)} represent the preliminary tracking estimation of the targets at the k-th moment obtained through association, y l (k) represents the preliminary tracking estimation of the trajectory l at the k-th moment, O(k) is the number of trajectories at the k-th moment, and α l (k) is the trajectory label used to label the attribution of the trajectory l;

[0029] The multi-target tracking problem is rewritten as,

[0030]

[0031] Furthermore, in step S200, it includes:

[0032] Assume that at time k, there are p track predictions and q measurements; let denote the association matrix with p rows and q columns provided by the data association network algorithm based on cross Transformer at time k; let be the p track predictions, and {z j (k)|j = 1, 2, …, q} are the q measurements; let a ij (k), i ∈ {1, 2, …, p}, j ∈ {1, 2, …, q} be the elements in the j-th row and the j-th column of A(k), and r i and c j be the i-th row and the j-th column vectors respectively;

[0033] Define the following three types of correlations:

[0034] (1) Measurements without associated tracks:

[0035] (2) Tracks without associated measurements:

[0036] (3) Measurement-track association pairs:

[0037] In the multi-object tracking problem, each measurement or prediction belongs to one of the three correlation cases.

[0038] Furthermore, in step S300, it includes:

[0039] Adopt the best linear unbiased estimator filter, and its calculation process is:

[0040] Prediction, let and P(k - 1|k - 1) be the state estimate and covariance at time k - 1; the state estimate and covariance propagation are,

[0041]

[0042] And,

[0043] P(k|k - 1) = F(P)P(k - 1|k - 1)F(k) T + G(k)Q k G(k) T (7)

[0044] Among them, P(k|k - 1) and P(k|k) respectively represent the prediction and estimation of the state covariance matrix at time k, and respectively represent the prediction and estimation of the state at time k; Q k represents the prior estimation matrix of the process noise;

[0045] Update, define and and

[0046] Let and

[0047] Let and

[0048] The update process of the state and covariance is

[0049]

[0050] and,

[0051] P(k|k) = (I - K(k)H(k))P(k|k - 1) (9)

[0052] where I is the identity matrix, The Kalman gain matrix K(k) is

[0053] K(k) = μ 1 P(k|k)JS -1 (10)

[0054] where J is the transformation matrix used to operate P(k|k)J to combine the columns of P(k|k) corresponding to x, y, and z into a new n×3 matrix; the matrix S = [s ij 3×3 The elements of are where, ∑ x 、∑ y and ∑ z respectively represent the variances of the estimates of x, y, and z; ∑ xy 、∑ yz and ∑ xy respectively represent the covariances between the estimates of x, y, and z.

[0055] Furthermore, in step S400, it includes:

[0056] S410. Obtain the trajectory estimate of the target using the best linear unbiased estimation filter related to the trajectory;

[0057] Let represent the label set, where ​is the set of positive integers; let denote two subsets of wherein, is the set of track labels being maintained, is the set of labels of candidate tracks to be confirmed;

[0058] Let S l (k) be the track score of the l-th track at time k, k ≥ 0; let ΔS l (k) be the increment of the track score at time k;

[0059] Allocate a track score with a truncated and decaying form,

[0060]

[0061] wherein, λ s ∈(0,1] is the decay factor, denotes the truncation parameter of the track score;

[0062] Let be the initialization time of the tracking track with label l; assume that for S l (k) = 0; according to the correlation classification, design ΔS l (k) as follows:

[0063] (1) For a measurement without an associated target , create a new candidate track by initializing the best linear unbiased estimator filter using a new label , wherein, and the label l will be added to The state of the target is initialized by the measurement and the prior, and the initial score of track l is given in the form of a log-likelihood ratio, defined as follows:

[0064]

[0065] wherein, P B denotes the birth probability of the target, P FA denotes the probability that a clutter point cannot be distinguished from a true target; P B and P FA are both calculated in advance according to the sensor characteristics;

[0066] (2) For track l without a measurement, update the estimate of the track by using the prediction,

[0067]

[0068] and,

[0069]

[0070] The increment of the corresponding trajectory score is

[0071]

[0072] where P D represents the target detection probability of the sensor;

[0073] (3) For a trajectory l with associated measurement value z j (k), the state and covariance of the trajectory are updated by using equations (8) and (9);

[0074] Calculate the increment of the corresponding trajectory score according to the position measurement and the estimate of trajectory l,

[0075]

[0076] where pr(·) represents the probability density function;

[0077] Given the predicted state at time k and covariance matrix Let have a probability density function that follows a Gaussian distribution

[0078]

[0079] Assume that clutter is uniformly distributed on the plane of region D, and the number of clutter N FA follows a Poisson distribution with parameter λ c being

[0080]

[0081] Substitute the distribution equations (17) and (18) into equation (16) to obtain

[0082]

[0083] where

[0084] S420. Let be the specified confirmation score threshold, and the score-based trajectory management strategy is

[0085] If then the trajectory l is confirmed; the label of trajectory l is deleted from L c and added to L m ;

[0086] If 0 ≤ S l (k) < S c, the trajectory l is still regarded as a candidate trajectory;

[0087] If S l (k) < 0, the trajectory l and the corresponding BLUE filter will be deleted;

[0088] For the trajectory with the label and in the maintenance state: If S l (k) ≥ 0, the trajectory l will be retained; otherwise, if S l (k) < 0, the trajectory l will be deleted and marked with the current label.

[0089] A multi - target tracking system based on the cross - Transformer data association algorithm, the system has program modules corresponding to the above steps, and executes the steps in the above - mentioned multi - target tracking method based on the cross - Transformer data association algorithm when running.

[0090] A computer - readable storage medium, the computer - readable storage medium stores a computer program, and the computer program is configured to implement the steps of the multi - target tracking method based on the cross - Transformer data association algorithm when called by a processor.

[0091] Compared with the prior art, the beneficial effects of the present invention are:

[0092] The present invention relates to a multi - target tracking method and system based on the cross - Transformer data association algorithm, which is for the multi - target association and trajectory tracking problem based on radar position measurement. We consider a challenging scenario where the number of targets is unknown and changes dynamically over time. In addition, we also consider the cases of missed detections and false alarms. To solve these challenging tasks, a data association network using CTDA is used, and a BLUE filter is adopted to obtain the best estimates in terms of the accuracy and efficiency of radar - based position measurement. In addition, a score - based trajectory management strategy is proposed to based on relevant historical measurements. Simulation experiments on various data sets show that the association and tracking methods proposed in the present invention are superior to the existing methods in complex multi - target tracking scenarios (including trajectory intersections, a large number of false alarms and missed detections). Compared with the RFS - based method, which has better effects among the existing tracking algorithms, the computational amount and storage requirements are significantly reduced, realizing real - time tracking; compared with the learning - based tracking algorithms, the problem of the limitation on the number of targets is solved, and at the same time, a special design is carried out based on radar data, significantly improving the processing effect for false alarms and missed detections. Description of the Drawings

[0093] Figure 1 It is a flowchart of a multi - target tracking method based on the cross - Transformer data association algorithm in an embodiment of the present invention;

[0094] Figure 2 Flow chart of the data association method based on the cross-Transformer network architecture in the embodiment of the present invention;

[0095] Figure 3 Flow chart of data filling and chunking in the data preprocessing process in the embodiment of the present invention;

[0096] Figure 4 Encoder network structure diagram based on Transformer in the embodiment of the present invention;

[0097] Figure 5 Decoder network structure diagram based on Transformer in the embodiment of the present invention;

[0098] Figure 6 Data association network diagram based on cross-Transformer in the embodiment of the present invention;

[0099] Figure 7 Processing method diagram of the Kuhn-Munkres (KM) algorithm in the embodiment of the present invention;

[0100] Figure 8 Sample tracking result diagram of the test data set in the simulation experiment in the embodiment of the present invention;

[0101] Figure 9 Tracking result diagram of the x-y plane obtained by using this algorithm in the simulation experiment in the embodiment of the present invention;

[0102] Figure 10 Tracking result diagram of the x-axis varying with time obtained by using this algorithm in the simulation experiment in the embodiment of the present invention;

[0103] Figure 11 Tracking result diagram of the y-axis varying with time obtained by using this algorithm in the simulation experiment in the embodiment of the present invention;

[0104] Figure 12 Comparison curve diagram of the OSPA distance between this algorithm and PHD, CPHD, and GLMB in the simulation experiment in the embodiment of the present invention;

[0105] Figure 13 Comparison curve diagram of the position between this algorithm and PHD, CPHD, and GLMB in the simulation experiment in the embodiment of the present invention;

[0106] Figure 14 Comparison curve diagram of the cardinality result between this algorithm and PHD, CPHD, and GLMB in the simulation experiment in the embodiment of the present invention. Detailed implementation manner

[0107] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of specific embodiments of the present invention will be given with reference to the accompanying drawings.

[0108] Specific Embodiment 1: As shown in Figures 1 to 7 The present invention provides a multi-target tracking method based on a cross-Transformer data association algorithm, including the following steps:

[0109] Step S100: Consider the problem of tracking a group of moving targets, the number of which is preset to be unknown in advance. Therefore, the dynamic model of the i-th target (i ∈ {1, 2,..., N}) is given by the following formula,

[0110]

[0111] where, x i (k) and u i (k) are the state and input vectors at time k, and η i (k) is the process noise vector, assumed to be a zero-mean Gaussian process;

[0112] The general measurement equation is given by the following formula,

[0113]

[0114] where, z i (k) is the measurement value, and v i (k) is the observation noise, assumed to be a zero-mean Gaussian process;

[0115] Specifically, the present invention considers relative distance and relative angle measurements, that is,

[0116]

[0117] where, (x i (k), y i (k), z i (k) represents the position of target i at time k; r i (k) is the relative distance between the sensor and target i at time k, θ i (k) and φ i (k) are the azimuth angle and elevation angle respectively; v i,r (k), v i,θ (k) and v i,φ (k) are the measurement noises of r i (k), θ i (k) and φ i (k) respectively, and it is assumed that v i,r (k), vi,θ (k) and v i,φ (k) are both zero-mean Gaussian processes, i.e., and

[0118] Let \(X(k)=\{x i (k)|i = 1,2,\cdots,N(k)\}\) represent the states of all targets at \(k\), where the \(x i (k)\) is the state of the \(i\)-th target and \(N(k)\) is the number of targets; let \(Z(k)=\{z j (k)|j = 1,2,\cdots,M(k)\}\) represent all the measurement values of the targets at \(k\), where the \(z j (k)\) is the \(j\)-th measurement value and \(M(k)\) is the number of measurement values; let represent the set of all target states from the initial time to time step \(K\), and represent the set of all measurement values, where \(K\) is the total number of time steps;

[0119] The multi-target tracking problem can be formulated as obtaining a mapping: where, for each element in the data set , there is at most one corresponding element in the data set , and vice versa; in the actual scenario, the true state of the target cannot be obtained, so the target tracking problem is transformed into the posterior estimation of the target state, that is, to solve the following optimization problem:

[0120]

[0121] where \(\Theta\) represents the parameters of the MTT algorithm;

[0122] Since the MTT method runs in real time, once new measurement results are obtained, the past tracking results will be retained instead of being optimized;

[0123] Let \(Y(k)=\{(y l (k),\alpha l (k))|l = 1,2,\cdots,O(k)\}\) represent the preliminary tracking estimates of the targets at the \(k\)-th moment obtained through association, where \(y l (k)\) represents the preliminary tracking estimate of the trajectory \(l\) at the \(k\)-th moment, \(O(k)\) is the number of trajectories at the \(k\)-th moment, and \(\alpha l (k)\) is the trajectory label used to label the ownership of the trajectory \(l\);

[0124] According to the definition of tracking estimation, the MTT problem can be rewritten as,

[0125]

[0126] By re-expressing Equation (4) as Equation (5), the multi-target tracking process can be divided into two independent processes: a data association process and a filtering process;

[0127] The objective of the present invention is to obtain an association matrix for two sets of position data, taking into account false alarms and missed detections, and accurately tracking multiple target trajectories;

[0128] S200. Establish the correlation between the current measurement and the trajectory based on the architecture of the Cross Transformer-based Data Association (CTDA) network.

[0129] Assume that at time k, there are p trajectory predictions and q measurements; let denote the association matrix with p rows and q columns provided by the CTDA algorithm at time k; let be the p trajectory predictions, and {z j (k)|j = 1, 2, …, q} be the q measurements; it should be noted that the association matrix A(k) is usually not full rank because there are missed detections and false alarms in the measurements; let a ij (k), i ∈ {1, 2, …, p}, j ∈ {1, 2, …, q} be the elements in the i-th row and j-th column of A(k), and r i and c j be the i-th row and j-th column vectors respectively; then the following three types of correlations can be defined:

[0130] (1) Measurements with no associated tracks:

[0131] (2) Tracks with no associated measurements:

[0132] (3) Measurement-track association pairs:

[0133] In the multi-target tracking problem, each measurement or prediction must belong to one of the three correlation cases;

[0134] S300. For each trajectory in step S200, maintain it by using the best linear unbiased estimator filter and perform state updates.

[0135] Assume that the motion and measurement equations for each target are described in Equations (1) and (2). Since Equation (2) is non-linear, any non-linear estimation method, such as the extended Kalman filter (EKF), unscented Kalman filter (UKF), and particle filter, can be used for state estimation; in the multi-target tracking problem considered, the measurement equation is constrained to the relative distance r, azimuth angle θ, and elevation angle φ with respect to the target; let (x, y, z) represent the position of the target and use the measurement equation defined in Equation (3); to provide the best estimate in terms of accuracy and efficiency under this measurement condition, a variant of the Kalman filter, namely the best linear unbiased estimator filter (BLUE), is adopted, and its calculation formula is:

[0136] Prediction: Let and P(k - 1|k - 1) be the state estimate and covariance at time k - 1; then the state estimate and covariance propagation are

[0137]

[0138] and

[0139] P(k|k - 1) = F(k)P(k - 1|k - 1)F(k) T + G(k)Q k G(k) T (7)

[0140] where P(k|k - 1) and P(k|k) represent the prediction and estimation of the state covariance matrix at time k respectively, and Q k represents the prior estimate matrix of the process noise, and represent the prediction and estimation of the state at time k respectively;

[0141] Update: For the sake of brevity, we define the following symbols and and

[0142] In addition, let

[0143] and Then we have and

[0144] The update process of the state and covariance is

[0145]

[0146] And,

[0147] P(k|k) = (I - K(k)H(k))P(k|k - 1) (9)

[0148] where I is the identity matrix, the Kalman gain matrix K(k) is given by

[0149] K(k) = μ 1 P(k|k)JS -1 (10)

[0150] where J is a transformation matrix that combines the columns of P(k|k) corresponding to x, y, and z into a new n×3 matrix by operating on P(k|k)J; the matrix S = [s ij 3×3 has elements: where Σ x 、Σ y and ∑ z represent the variances of the x, y, and z estimates, respectively; ∑ xy 、Σ yz and Σ xy represent the covariances between the x, y, and z estimates, respectively;

[0151] S400. Using a scoring-based track management strategy to process unknown targets to be tracked and targets to be tracked with varying quantities, and determining the initialization, update, and deletion of trajectories, including

[0152] Since each trajectory is treated as an independent entity and is effectively maintained by using the best linear unbiased estimator (BLUE) filter of step S300; the scoring strategy aims to assign a score to each trajectory, which is updated based on the association history of the trajectory; each trajectory is initialized by a single measurement and is deleted when its score drops below a certain threshold;

[0153] S410. To track multiple targets simultaneously, use a series of BLUE filters associated with the trajectories to obtain the trajectory estimates of the targets;

[0154] Let represent the set of labels, where is the set of positive integers; let represent two subsets of and is the set of trajectory labels being maintained, ​is the set of labels of candidate trajectories to be confirmed; note that candidate trajectories are not "true" trajectories, which means that their corresponding states are not considered tracking results until they are confirmed as maintained trajectories;

[0155] Let S l (k) be the trajectory score of the l-th trajectory at time k, k ≥ 0; Let ΔS l (k) be the increment of the trajectory score at time k, which can also be regarded as the evaluation score of the relationship between the measurement at time k and the trajectory with label l; To consider the association information of historical trajectories, we assign a trajectory scoring with a truncated and decaying form as follows,

[0156]

[0157] where λ s ∈(0,1] is the decay factor, represents the truncation parameter of the orbit scoring;

[0158] Let be the initialization time of the tracking trajectory with label l; We assume that for S l (k) = 0; According to the correlation classification, we design ΔS l (k) as follows:

[0159] (4) For measurements without associated targets , a new candidate trajectory is created by initializing the BLUE filter with a new label , where and the label l will be added to The state of the target is initialized by the measurement and the prior, and the initial score of the trajectory l is given in the form of a log-likelihood ratio (LLR), defined as follows:

[0160]

[0161] where P B represents the birth probability of the target, and P FA represents the probability that a clutter point cannot be distinguished from a true target; P B and P FA can both be calculated in advance according to the sensor characteristics;

[0162] (5) For the trajectory l without measurement, the estimate of the trajectory is updated by using the prediction, i.e.,

[0163]

[0164] and,

[0165]

[0166] At this time, the increment of the corresponding trajectory score is

[0167]

[0168] where P D represents the target detection probability of the sensor;

[0169] (6) For a trajectory l with associated measurement value z j (k), the state and covariance of the trajectory are updated by using equations (8) and (9); in this case, the increment of the corresponding trajectory score is calculated based on the position measurement and the estimate of trajectory l,

[0170]

[0171] where pr(·) represents the probability density function;

[0172] Given the predicted state and covariance matrix at time k and assuming the probability density function of follows a Gaussian distribution i.e.,

[0173]

[0174] Assume that the clutter is uniformly distributed on the plane of region D, and the number of clutter N FA follows a Poisson distribution with parameter λ c being

[0175]

[0176] Substituting the distribution formulas (17) and (18) into formula (16) gives

[0177]

[0178] where

[0179] S420. Let be the specified confirmation score threshold, then we have the following score-based trajectory management strategy,

[0180] For a candidate trajectory with label

[0181] If then the trajectory l is confirmed; the label of trajectory l is deleted from L c and added to L m ;​

[0182] If 0 ≤ S l (k) < S c , then the trajectory l is still regarded as a candidate trajectory;

[0183] If S l (k) < 0, then the trajectory l and the corresponding BLUE filter will be deleted;

[0184] For a trajectory with a label of and in a maintained state: If S l (k) ≥ 0, then the trajectory l will be retained; otherwise, if S l (k) < 0, then the trajectory l will be deleted and marked with the current label.

[0185] In practice, it is not always necessary for all the parameters required to calculate the scores to be accurate. Providing parameters similar to those of the real - world scenario is usually sufficient. By simply adjusting the parameters λ s and λ t , the performance of the algorithm can be fine - tuned to better meet specific requirements.

[0186] Specific Example Two: As shown in combination with Figures 2 to 7 , in step S200, it includes,

[0187] The data association process aims to solve the problem of determining whether a set of data (which can be measurement data or estimated data) originates from the same target. In the MTT process, new target detection requires "measurement - measurement" data association across multiple sampling periods. To update the track and maintain the tracking estimate, "measurement - track" association is needed to identify new measurement data for track correction. The correlation between these associations, including "measurement - measurement" and "measurement - track" data associations, is described by an association matrix. Given two sets of data: and to be associated. Then the association matrix of the two sets of data is expressed as,

[0188]

[0189] where,

[0190]

[0191] It should be noted that each column and each row of the association matrix is allowed to have at most one non - zero element, that is, 1, while all other elements are 0. The purpose of this algorithm is to estimate the states of different numbers of targets by using position measurements from a noisy radar;

[0192] S210. Define a Relative Distance Feature (RDF) matrix to ensure the generalization of the normalized features of different data, including:

[0193] Let and D 1 and D 2 represent two sets of position data respectively, and these data can be measurements, estimates, and predictions from real targets or false alarms.

[0194] To characterize the correlation relationship features between the data sets D 1 and D 2 , m×n matrices regarding the relative distances in the x, y, and z directions are respectively defined, which are called Relative Distance Feature (RDF) matrices.

[0195] Among them, for the x-axis direction,

[0196] Let be the RDF matrix associated with the x direction as follows,

[0197]

[0198] Among them, and are the x-direction data from the data sets D 1 and the data set D 2 respectively; x bias is a small quantity with an appropriate size to avoid data errors caused by division by zero; x mid is the geometric median of the absolute values of all relative distances between the two data sets,

[0199]

[0200] For the RDF matrices in the y-axis direction and z-axis direction, they are respectively expressed as and and can be constructed in the same way;

[0201] In the RDF matrix, a value close to 1 indicates that the corresponding data points in the data sets D 1 and the data set D 2 are similar and likely to be correlated with each other; conversely, a value close to 0 indicates that the data points are significantly different and do not belong to the same target.

[0202] S220. Data preprocessing, introducing a chunking method and a padding method to augment the data in the data sets D 1 and the data set D 2 with additional data of appropriate dimensions and perform chunking processing, including:

[0203] In many tasks, there are significant variations in the number and dimensions of the acquired measurements; furthermore, even within a single task, the number of measurements at different times may vary;

[0204] In the case of tracking a limited number of targets, appropriate maximum padding values can be pre-allocated; however, when faced with a large number of targets and clutter, overly large padding values become computationally infeasible; in order to make the data association algorithm proposed by the present invention applicable to data sets of different sizes, a superior data partitioning technique needs to be introduced.

[0205] Let N pad be a fixed value, and N pad be less than or equal to m or n; let and be the integers of the rows and columns of the scaled padding matrix respectively, be the ceiling operation;

[0206] Let be the padding matrix associated with x after partitioning,

[0207]

[0208] To adapt to the input dimension of the CTDA network, further partitioning is performed as follows,

[0209]

[0210] where, and are the number of row and column partitions respectively; each partition sub-matrix is an N pad ×N pad matrix;

[0211] Using the partitioning defined in formula (25), the scaled padding matrix will be reshaped into a batch matrix as follows,

[0212]

[0213] The padding matrices related to directions y and z can be defined in the same way as direction x; subsequently, the enhanced matrix containing relative distance features is used as the input to the CTDA network;

[0214] Compared with the prior art, the data padding technique of the present invention has a clearer physical meaning, because it is equivalent to generating multiple false alarms at an infinite distance from the original data; therefore, during the entire padding and partitioning process, no additional information is introduced into the RDF matrix; it is ensured that the basic characteristics of the matrix after the padding process are not affected.

[0215] S230. Construct a data association network based on CrossTransformer, including

[0216] S231. Construct a Transformer encoder, including three integrated modules, namely, multi-head attention (MHA), feed-forward network (FFN), and normalization network, as Figure 4 shown, among which the attention module is particularly important, and it is the cornerstone of the MHA backbone;

[0217] The main goal of the self-attention neural network is to use the Query to retrieve the attention values corresponding to the Keys and assign them to the corresponding Values; then redistribute the Value according to the attention weights; let Attention represent the attention processing, and its form is as follows,

[0218]

[0219] where, and represent the Query and Key matrices respectively; represents the y matrix value, d is the dimension of Q, K, and V in the self-attention neural network, and Q, K, and V represent the same feature encoding;

[0220] In addition, by using the matrices W Q , W K and W V combined with the multi-head attention network, different types of relationships can be extracted from the data,

[0221] MultiHead(Q, K, V) = Concat(head 1 , …, head n )W O (28)

[0222] where, is each attention network head, and W O are trainable parameters, which are equivalent to the Q, K, and V tensors having the same shape;

[0223] S232. Construct a Transformer decoder, including two multi-head attention networks (MHA). Set a mask layer in the first multi-head attention network, and the mask layer can be configured as a matrix with all elements equal to 1, so as to be the same as the standard multi-head attention network; use Q, K, and V in the second multi-head attention network;

[0224] Specifically, the decoder input is used as the query feature, while the encoder input is used as the Key and Value; this arrangement enables the decoder to utilize the encoded features obtained through multi-head attention as queries to retrieve information from the encoder output; subsequently, the resulting weighted input encoding is combined with the decoder input and passed through normalization and a feed-forward neural network to produce the decoder output;

[0225] S233. Construct a cross Transformer data association network (CTDA) by combining the Transformer encoder in step S231 and the Transformer decoder in step 232, including,

[0226] Construct a row-column cross Transformer network, which is applied to three channels of the input matrix; each channel is processed separately by a row Transformer encoder and a column Transformer encoder, considering the input rows and columns as feature vectors respectively; it should be noted that the input of each channel is processed independently; then, the output of the encoder is input to the row and column Transformer decoders in a cross mode;

[0227] Since the Transformer decoder includes two MHA modules, in the context of the row-column cross Transformer network, the outputs of the row and column encoders serve as the Key and Value inputs of the second MHA module in the row and column Transformer decoders respectively; meanwhile, the output of each encoder is cross-fed into the other Transformer decoder as the input of the first MHA module layer; the outputs of the row and column decoders (each channel consists of two matrices) are combined using a processing module; to meet different requirements, an element-wise method is introduced to improve the efficiency of this algorithm,

[0228]

[0229] where h e is the output of the mixing matrix, Reshape(·) represents converting the dimension of the matrix to a specified number, [·] represents the concatenation operation, W and b are trainable weights and biases, and h 1 and h 2 are the decoders;

[0230] A 3-channel cross Transformer network is constructed, and for the 3-channel encoded data obtained through bidirectional processing, a cross Transformer network structure is implemented; the output of the element-wise mixing processing module composed of 3-channel data is initially directed to three Transformer encoders to obtain encoded features; then, the output of each encoder corresponding to each channel is simultaneously fed into the first MHA module of the decoders of other channels; in addition, each output is simultaneously used as the key and value inputs of the second MHA module in the corresponding decoder of the channel; then, an output matrix is obtained through element-wise mixing.

[0231]

[0232] where, O 1 、O 2 and O 3 are the outputs of the decoders corresponding to the three channels;

[0233] The cross Transformer adopted in the 3-channel cross Transformer network is different from the corresponding cross Transformer in the row-column cross Transformer; firstly, compared with the row-column Transformer module, the 3-channel cross Transformer network contains fewer layers, and this adjustment is made by observing that the information fusion between different channels is often relatively less complex than the operations required by the row-column cross Transformer module, and using a deeper network in this specific case may lead to overfitting; secondly, the number of Transformers in the cross structure applied between each pair of channels varies with the dimension of the data rather than being fixed; in addition, the number of encoders and decoders can be different.

[0234] S234. Since the data association matrix exhibits sparsity and consists of binary values (0-1 matrix), the binary cross-entropy loss is used to calculate the relative association probability of the network output as the cost.

[0235]

[0236] where, and are the elements of the i-th row and j-th column predicted for the association matrix O pr and the true association matrix O gt respectively.

[0237] S240. Each element of the matrix obtained from the cross-Transformer-based data association network represents the association probability of a given input data pair. To further convert the network output into an association matrix, the Kuhn-Munkres (KM) algorithm, also known as the Hungarian matching algorithm, is used to find the optimal assignment. The specific steps are combined with Figure 7 as shown;

[0238] It should be noted that the KM algorithm is designed specifically to solve complex situations and is very useful when dealing with more challenging scenarios. In most cases, directly applying a threshold to the network output to generate the association matrix is sufficient.

[0239] The other combinations and connection relationships of this implementation scheme are the same as those of the first specific implementation scheme.

[0240] Specific implementation scheme three: A multi-target tracking system based on the cross-Transformer data association algorithm of the present invention. This system has program modules corresponding to the above steps and executes the steps in the above multi-target tracking method based on the cross-Transformer data association algorithm when running.

[0241] The other combinations and connection relationships of this implementation scheme are the same as those of the first or second specific implementation scheme.

[0242] Specific implementation scheme four: A computer-readable storage medium of the present invention. The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the multi-target tracking method based on the cross-Transformer data association algorithm when called by a processor.

[0243] The other combinations and connection relationships of this implementation scheme are the same as those of the first or second specific implementation scheme.

[0244] Simulation experiment

[0245] Generate and utilize a customized 3D dataset to evaluate the association accuracy of this algorithm, and adopt a challenging 2D multi-target tracking scenario to compare the multi-target tracking algorithm proposed by the present invention with several existing popular multi-target tracking algorithms. The existing popular multi-target tracking algorithms include PHD, CPHD, and GLMB. Monte Carlo simulation is used to evaluate the tracking effectiveness in this scenario.

[0246] To evaluate the results of different multi-target tracking algorithms, the Optimal Sub-Pattern Assignment (OSPA) distance is widely used. The OSPA distance is a metric that captures the difference between two sets by considering the difference in the number of points and the difference in the point values. It provides association accuracy and estimation accuracy. Define the set of target true positions as P = {p 1 , p2 , … p m}, and the set of estimated states of the targets is where m and n are the number of ground truths and the number of estimated targets respectively; p ∈ P, are the true target state vector and the estimated target state vector at a certain moment respectively. Then the calculation of the OSPA distance is

[0247]

[0248] where, Π n represents all combinations of m elements in the set X, and the number of combinations is represents finding the minimum distance using the following expression; r is the distance weight parameter, 1 ≤ r ≤ inf, is the cut-off distance between two positions, where c > 0 is the cut-off parameter;

[0249] The evaluation of localization and cardinality errors is given by

[0250]

[0251] and,

[0252]

[0253] It can be seen that and the values of are exactly thus providing the accuracy of localization and target number estimation.

[0254] On the other hand, when evaluating the results of multi-target association, the root mean square error (RMSE) is usually used for single-target evaluation. RMSE only focuses on the estimation accuracy of a single target.

[0255] The CTDA method proposed by the present invention can associate two-dimensional or three-dimensional trajectory data. The advantage is that it does not require prior quantitative information, such as false alarm probability and miss detection probability. In order to train the network in CTDA, a large amount of data is required to enhance the generalization ability of the network and prevent overfitting. However, obtaining real trajectory data may be challenging, and the available data is usually limited. To overcome these limitations, we developed a comprehensive multi-target simulation platform to generate a large amount of simulated data for trajectory measurement for network training.

[0256] First, a 3D dynamic model and a measurement model are established to generate training data. Select the state vector as x = [x, y, z, v, ζ, ψ] T and the measurement z = [r, θ, φ], we have

[0257]

[0258] where x, y, and z represent the position of the target; v, ζ, and ψ represent the velocity, pitch angle, and yaw angle, respectively; θ and φ are the azimuth angle and elevation angle, respectively; u D , u L and u G are the controls for velocity, pitch angle, and yaw angle, respectively; η is the process noise and measurement noise vector. The measurement equation is described in Equation (3). The parameters of the customized multi-target dataset are shown in Table 1 below.

[0259] Table 1

[0260]

[0261]

[0262] Using these parameters, a total of 2000 trajectories were generated. Among them, 1800 trajectories were used for training, and the remaining 200 trajectories were reserved for testing the network performance.

[0263] The network was trained for 8 epochs on the training dataset with a batch size of 256. The Adam optimizer was adopted with a learning rate of 10 -4 . Since the training process was relatively short, learning rate decay was not used.

[0264] In the CTDA network of this algorithm, the selected data dimension is 40, which corresponds to the size of the input and output matrices. The multi-head attention (MHA) module in CTDA uses 4 heads. A dropout rate of 0.3 was selected for regularization. During testing and filtering, a threshold of 0.1 (λ a = 0.1) was selected.

[0265] The sample tracking results of the test dataset are as Figure 8 shown. The gray plots depict all the measurement results, while the trajectories of multiple targets are distinguished by different colors. The results presented illustrate the accurate estimation of trajectories from noisy measurements, as well as the successful identification of trajectory birth and disappearance events.

[0266] Next, multiple simulation experiments will be conducted to verify and evaluate the proposed association and multi-target tracking algorithms. In addition, we will compare the proposed algorithm with RFS-based tracking algorithms (including PHD, CPHD, and GLMB) to demonstrate the advantages of the proposed MTT algorithm.

[0267] For a fair comparison, the parameter settings used in the experiments are the same as those in. The experiments study a variable number of objects (up to 10) over time. These objects undergo birth, death, and crossing. The kinematics of each individual object uses the state vector is described, where [x, y] is the planar position, is the velocity, and ω is the turning rate. The sampling period is set to 1 second. In discrete time, the state equation for each target is,

[0268] x k+1 = F ω x k + v ω (23)

[0269] Let,

[0270]

[0271] where, is the process noise, σ ω = 15 m / s 2 and σ u = (π / 180) rad / s are the process noise standard deviations.

[0272] The measurement results are noisy 2D azimuth and range detections as,

[0273]

[0274] where the range is limited to a disk with a radius of 2000 m, and the noise standard deviations are σ r = 5 m and σ θ = (π / 360) rad. According to the experimental settings in, the targets are set to be born near four positions. The training parameters are shown in Table 2 below.

[0275] Table 2

[0276]

[0277] The tracking results are as Figure 9 , Figure 10 , Figure 11 shown. The gray plots represent all the measurements, and the trajectories corresponding to multiple targets are marked with different colors. The tracking results demonstrate the effectiveness of the algorithm, successfully tracking multiple targets in real time and accurately assigning a unique label to each target. This can be clearly seen from the different colors assigned to the targets. Most importantly, the proposed algorithm can handle trajectory intersections with false alarms and missed detections, and can also resume tracking after encountering adverse situations. In particular, the algorithm ensures that the trajectory labels do not switch throughout the tracking process, which is very beneficial for practical applications.

[0278] To further demonstrate the tracking performance of the proposed method, we compare it with PHD, CPHD, and GLMB. We perform 100 Monte Carlo simulations on the test dataset, and the average OSPA distances of each method over time are shown in the figure. Figure 12 (Select c = 100 and r = 1). Accordingly, the separate position and cardinality results are shown in Figure 13 and Figure 14 respectively. Table 3 summarizes the OSPA distance, position, and cardinality results obtained from the proposed method and the comparative methods.

[0279] Table 3

[0280]

[0281] As shown in Table 3, the proposed method outperforms all other methods in terms of the OSPA distance. The proposed method ranks second in terms of the OSPA position, while achieving similar results to the other methods in terms of the OSPA cardinality. From Figure 13 and Figure 14 , it can be seen that the proposed method produces similar tracking accuracy to PHD, outperforms the tracking accuracy of CPHD and GLMB, and performs better than PHD in terms of the OSPA cardinality. In addition, Figure 14 shows the fast response of our method when a target is born or disappears. It can be seen that at the main nodes where a target is born or disappears, including time nodes such as 10 s, 60 s, 80 s, etc., our method can drop to the normal tracking state faster.

[0282] It is worth mentioning that since there is no sample algorithm and the number of filters under the maintained scale is linearly related to the number of targets, the proposed tracking method has a relatively low complexity. In addition, the running time of the complete tracking process on our platform (Intel i7-8700K) is 0.3 s, which is suitable for real-time tracking.

[0283] Although the present invention is disclosed as above, the scope of protection of the present invention is not limited thereto. Those skilled in the art of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will all fall within the scope of protection of the present invention.

Claims

1. A multi-target tracking method based on a cross-Transformer data association algorithm, characterized in that: The following steps are involved: S100, establishing a dynamic model and a universal measurement equation for a series of maneuvering dynamic targets, considering relative distance and relative angle measurement, and converting the target tracking problem into a posteriori estimation of the target state for data association in step S200; S200, the architecture of the data association network based on the cross Transformer, establishes the correlation between the current measurement and the trajectory; S300, maintaining each trajectory in step S200 by using an optimal linear unbiased estimation filter and performing a state update; including: The optimal linear unbiased estimation filter is used, and the calculation process is: Forecast, set and P(k-1|k-1) are the state estimates and covariances at time k-1; the state estimates and covariance propagation are, and, P(k|k-1)=F(k)P(k-1|k-1)F(k) T +G(k)Q k G(k) T (7) in, P(k|k-1) and P(k|k) represent the prediction and estimation of the state covariance matrix at time k, respectively. and They represent the prediction and estimation of the state at time k respectively; Q k Represents the prior estimation matrix of process noise; Update, define and and set up and set up and The updating process of state and covariance is, and, P(k|k)=(IK(k)H(k))P(k|k-1) (9) Where I is the identity matrix, The Kalman gain matrix K(k) is, K(k)=μ1P(k|k)JS -1 (10) Where J is a transformation matrix used to operate P(k|k)J to merge the columns of P(k|k) corresponding to x, y, and z into a new n×3 matrix; the matrix S = [s ij ] 3×3 The elements of are, Among them, ∑ x ,∑ y and∑ z Represent the variance of x, y and z estimates respectively; ∑ xy ,∑ yz and∑ xy denote the covariance between the estimates of x, y, and z, respectively; S400: Using a scoring-based track management strategy to process unknown targets to be tracked and a changing number of targets to be tracked, determine initialization, update and deletion of tracks.

2. According to claim 1, a multi-target tracking method based on a cross-Transformer data association algorithm is characterized in that: In step S100, it includes: Consider tracking a group of moving targets, and assume that the dynamic model of the i-th target (i∈{1,2,…,N}) is, x' i (k)=f(u' i (k-1),u i (k),η i (k)) (1) Among them, x' i (k) and u i (k) is the state and input vector at time k, η i (k) is the process noise vector, Assume a zero-mean Gaussian process; The general measurement equation is, z' i (k)=h(x' i (k),v i (k)) (2) Among them, z' i (k) is the measured value, v i (k) is the observation noise, Assume a zero-mean Gaussian process; Considering relative distance and relative angle measurements, Among them, (x i (k),y i (k),z i (k) represents the position of target i at time k; r i (k) is the relative distance between the sensor and target i at time k, θ i (k) and φ i (k) are azimuth and elevation respectively; v i,r (k), v i,θ (k) and v i,φ (k) are r i (k),θ i (k) and φ i (k) measurement noise, and assuming that ν i,r (k), v i,θ (k) and ν i,φ (k) are all zero-mean Gaussian processes, that is and Let X(k) = {x' i (k)|i=1,2,…,N(k)} represents the state of all targets at k, where the x' i (k) is the state of the i-th target, N(k) is the number of targets; let Z(k) = {z' j (k)|j=1,2,…,M(k)} represents all the measured values ​​of the target at k, where z' j (k) is the jth measurement value, M(k) is the number of measurements; let represents the set of all target states from the initial moment to time step K, and represents the set of all measurements, where L is the total number of time steps; The multi-target tracking problem is formulated as obtaining a mapping as, Convert the target tracking problem into a posterior estimation of the target state. Where, Θ represents the parameters of the MTT algorithm; If new measurements are obtained, past tracking results will be retained; Let Y(l) = {(y l (k),α l (k))|l=1,2,…,O(k)} represents the initial tracking estimate of the target at the kth moment obtained by association, y l (k) represents the initial tracking estimate of trajectory l at the kth moment, O(k) is the number of trajectories at the kth moment, and α l (k) is the trajectory label, which is used to mark the ownership of trajectory l; The multi-target tracking problem can be rewritten as:

3. The multi-target tracking method based on the cross Transformer data association algorithm according to claim 2 is characterized in that: In step S200, it includes: Assume that there are p trajectory predictions and q measurements at time k; represents the association matrix with p rows and q columns provided by the cross-Transformer-based data association network algorithm at time k; let is p trajectory prediction, {z j (k)|j=1,2,…,q} is q measurement values; let a ij (k), i∈{1,2,…,p},j∈{1,2,…,q} is the j-th row and j-th column element in A(k), r i and c j are the i-th row and j-th column vectors respectively; The following three types of dependencies are defined: (1) Measurement without associated track: (2) Tracks without associated measurements: (3) Measurement-track correlation pair: In the multi-target tracking problem, each measurement or prediction belongs to one of three relevant cases.

4. The multi-target tracking method based on the cross-Transformer data association algorithm according to claim 3 is characterized in that: In step S400, it includes: S410, obtaining a trajectory estimate of the target using an optimal linear unbiased estimation filter associated with the trajectory; set up represents a label set, where is a set of positive integers; let express two subsets of in, is the set of trajectory labels being maintained, is the label set of candidate trajectories to be confirmed; Let S l (k) is the trajectory score of the lth trajectory at time k, k≥0; let ΔS l (k) is the increment of the trajectory score at time k; Assign trajectory scores with truncation and decay forms, Among them, λ s ∈(0,1] is the attenuation factor, represents the cutoff parameter for track scoring; set up is the initialization time of the tracking trajectory with label l; assuming that for S l (k) = 0; according to the correlation classification, design ΔS l (k) are as follows: (1) For targets without any association By using the new label Initialize the best linear unbiased estimation filter to create new candidate trajectories, where And the label l will be added to The state of the target is initialized by measurements and priors, and the initial score of the trajectory l is given in the form of log-likelihood ratio, defined as follows: Among them, P B represents the probability of the target being born, P FA It indicates the probability that the clutter point cannot be distinguished from the real target; P B and P FA All are calculated in advance based on the sensor characteristics; (2) For a trajectory l without measurement, update the trajectory estimate by using the prediction, and, The corresponding trajectory score increment is, Among them, P D represents the target detection probability of the sensor; (3) For the relevant measurement value z j (k) The trajectory l, the state and covariance estimates of the trajectory are updated by using formula (8) and formula (9); The increment of the corresponding trajectory score is calculated based on the position measurement and the estimate of trajectory l, Where pr(·) represents the probability density function; Given the predicted state at time k and the covariance matrix set up The probability density function of Assume that the clutter is evenly distributed on a plane with an area of ​​D, and the number of clutter N FA It follows a Poisson distribution with parameter λ c for, Substituting distribution formula (17) and formula (18) into formula (16), we obtain: in, S420, Order For the specified confirmation score threshold, the score-based trajectory management strategy is, if Then the track l is confirmed; the label of track l is changed from L c Remove and add to L m middle; If 0≤S l (k) c , then trajectory l is still considered as a candidate trajectory;​ If S l (k)<0, then the trajectory l and the corresponding BLUE filter will be deleted; For labels And the trajectory in maintenance state: If S l (k) ≥ 0, then the trajectory l will be retained; otherwise, if S l (k)<0, the track l will be deleted and marked as the current label.

5. A multi-target tracking system based on a cross-Transformer data association algorithm, characterized by: The system has a program module corresponding to the steps of any one of claims 1 to 4 above, and executes the steps in the above-mentioned multi-target tracking method based on the cross-Transformer data association algorithm when running.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the multi-target tracking method based on the cross-Transformer data association algorithm described in any one of claims 1 to 4 when called by a processor.

Images