Multi-extended-target joint tracking and classification method based on star convex RHM and LMB filters

Abstract

The invention discloses a multi-extended target joint tracking and classification method based on star convex RHM and LMB filters, and belongs to the field of radar target tracking. According to the method, an LMB parameter set is initialized by using target prior information, and then a sensor measurement set is divided through a mean shift algorithm; then, state prediction is achieved by combining survival target parameter updating and new target parameter sampling, and then measurement updating is completed through LMB-to-GLMB, GLMB updating and GLMB-to-LMB; and then trimming and fusing the LMB parameter set, estimating the number of targets, extracting state information, and circularly executing until observation is finished. According to the method, a star convex RHM modeling expansion state is adopted to reduce dimensions, the low detection probability / high clutter scene performance is improved based on an LMB framework, the tracking classification effect and the real-time performance are both considered, and the engineering application value is high.

Description

Technical Field

[0001] This invention belongs to the field of radar target tracking, specifically involving a multi-extended target joint tracking and classification method based on a star-convex random hypersurface model and a tag-based multi-Bernoulli filter. Background Technology

[0002] With the continuous development of high-resolution radar, a single target may acquire multiple measurements, and target description has gradually evolved from traditional point targets to extended targets (ET). Because ET can obtain richer information about the target, ET tracking methods can estimate not only the target's motion state but also its extended state and measurement rate state. Among these, the estimation of the extended state is of great significance for target classification.

[0003] In recent years, Joint Tracking and Classification (JTC) methods based on extended states have gradually become a focus of scholarly attention. In 2013, Lan Jian et al. proposed an ET JTC method based on a Random Matrix Model (RMM) (Lan J, Li X. Joint tracking and classification of extended object using random matrix[C] / / 16th International Conference on Information Fusion.IEEE,2013:1550-1557.). This method is the first ET JTC method based on extended states. It introduces the target's extended state as prior class information into the filtering process to achieve target classification and improve tracking performance, achieving good results. However, this method is only applicable to JTC of a single ET. For the multiple ET (MET) JTC problem, Ji Hongbing and other scholars proposed an ET probability hypothesis density (PHD) JTC method based on RMM (Hu Q, Ji H, Zhang Y. A standard PHD filter for joint tracking and classification of maneuvering extended targets using random matrix[J]. Signal Processing, 2018, 144:352-363.), namely the ET-JTC-GGIW-PHD filter.

[0004] However, since RMM can only estimate the extended state as an ellipse, the two methods mentioned above can only classify targets based on their size. When targets of different shapes are similar in size, classification errors are likely to occur. To address this, Wang Liping et al. proposed an ET-JTC-RHM-PHD filter based on the Random Hypersurface Model (RHM). This method can classify targets based on their different shapes, and it has better classification performance for MET targets with similar sizes but different shapes. However, this technical solution is based on the ET-PHD filter framework, which results in poor tracking performance under low detection probability or high clutter conditions. This leads to poor classification performance, thus weakening the corrective effect of the classification results on the tracking performance. Subsequently, Ji Hongbing and other scholars proposed a METJTC method based on Gaussian process (GP) and labeled multi-Bernoulli (LMB) filters (Cheng X, Ji H, Zhang Y. Multiple extended target joint tracking and classification based on GPs and LMB filter[J].IEEE Signal Processing Letters, 2024, 31: 1139-1143.), namely the ET-JTC-GP-LMB filter. This method has better tracking and classification performance than the ET-JTC-RHM-PHD filter, but it uses GP to model the extended state of the target, and the filtering process involves high-dimensional vector and matrix operations, resulting in high computational complexity and difficulty in engineering applications. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a multi-extended target joint tracking and classification method based on star-convex RHM and LMB filters. Compared to the ET-JTC-RHM-PHD filter, this invention is based on the ET LMB filtering framework, which can effectively improve tracking performance under low detection probability or high clutter conditions, thereby improving classification performance and enhancing the correction effect of classification results on tracking performance. Compared to the ET-JTC-GP-LMB filter, this invention uses star-convex RHM to model the target extended state, which can effectively reduce the dimensionality of the extended state vector and greatly reduce the computational load.

[0006] According to a first aspect of the present invention, a method for joint tracking and classification of multiple extended targets based on star convex RHM and LMB filters is provided, the method comprising:

[0007] Step 1: Initialization

[0008] The tag's multi-Bernoulli LMB parameter set is initialized using the target's prior information to obtain the initial LMB parameter set. The LMB parameter set includes the assumed trajectory existence probability and the probability density mass function (PDMF) parameter set. The PDMF comprehensively considers the target's measurement rate state, motion state, expansion state, and class state.

[0009] Step 2: Measurement and Division

[0010] The mean-shift algorithm is used to divide the real measurement set obtained by the sensor into target measurements and clutter measurements.

[0011] Step 3: State Prediction

[0012] Based on the posterior PDMF parameter set of the multi-extended target MET at time k-1 and the PDMF parameter set of the newborn MET at time k, the predicted MET PDMF parameter set at time k is calculated. The prediction process includes the calculation of the surviving MET parameters and the sampling of the newborn MET parameters.

[0013] Step 4: Measurement Update

[0014] The three operations are executed sequentially: LMB to GLMB, GLMB update, and GLMB to LMB. During the GLMB update process, parameters need to be calculated separately for two scenarios: target missed detection and target detected.

[0015] Step 5: Trimming and Blending

[0016] The updated LMB parameter set is pruned and truncated, and the mixed components in the PDMF parameter set are pruned, merged, and weighted and normalized.

[0017] Step 6: State Estimation

[0018] The number of targets is estimated based on the pruned and fused LMB parameter set, and the pre-extraction... The system uses the state information of the trajectory with a high probability assumption to complete the estimation of target motion, measurement rate, expansion and class state;

[0019] Step 7: Loop and check

[0020] Check if the observation has ended. If not, return to step 3 and repeat the filtering process for the next time step. Before each loop, the measurement division operation in step 2 must be re-executed.

[0021] In some exemplary embodiments, the extended state in step 1 is divided into those that depend on the actual measurement. and depends on prior class information Both parts are modeled as Gaussian distributions; the measurement rate state is modeled as a gamma distribution, and the motion state is modeled as a Gaussian distribution, and the measurement rate state, motion state, and extended state are independent of each other.

[0022] In some exemplary embodiments, the survival MET parameter calculation in step 3 includes:

[0023] Update the target existence probability, the number of mixed component terms, and the weights based on the survival probability;

[0024] The measurement rate state parameters are updated using a forgetting factor.

[0025] The motion state parameters are updated by combining the motion state transition matrix and the process noise covariance, while carrying the auxiliary variables of the inverse Wissaud distribution to improve the estimation accuracy;

[0026] Update extended state parameters based on extended state transition matrix and process noise covariance;

[0027] Inherit the class state parameters directly at time k-1.

[0028] In some exemplary implementations, during GLMB update in step 4, a pseudo-measurement set composed of prior class information of the target extended state is introduced. The update of the measurement rate state parameter in the detection scenario needs to be combined with the number of measurements in the measurement partitioning unit. The extended state is updated nonlinearly using the unscented Kalman filter method, and the class state parameter is updated based on the Bayesian criterion combined with the pseudo-measurement information.

[0029] In some exemplary embodiments, step 5, trimming and merging, specifically includes:

[0030] Hypothetical trajectories with an existence probability below the pruning threshold are discarded, and when the number of trajectories exceeds the maximum number, they are truncated in descending order of existence probability. Mixed components with weights below the pruning threshold are discarded, and components whose motion state similarity satisfies the fusion threshold are merged. When the number of components exceeds the maximum number, they are truncated in descending order of weight. Finally, the component weights are normalized.

[0031] In some exemplary embodiments, during state estimation in step 6, the number of targets is determined by the maximum potential distribution of the LMB parameter set, the class state is determined by the category corresponding to the maximum class probability, and the extended state is obtained by combining the measurement update results and the class information correction results.

[0032] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the multi-extended target joint tracking and classification method based on star convex RHM and LMB filters described in the first aspect above.

[0033] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the multi-extended target joint tracking and classification method based on star convex RHM and LMB filters described in the first aspect above.

[0034] According to a fourth aspect of the present invention, an electronic device is provided, comprising:

[0035] Processor; and

[0036] Memory for storing the executable instructions of the processor;

[0037] The processor is configured to implement the multi-extended target joint tracking and classification method based on star convex RHM and LMB filters described in the first aspect above by executing the executable instructions.

[0038] The beneficial effects of this invention are as follows:

[0039] The present invention provides a multi-extended target joint tracking and classification method based on star convex RHM and LMB filters, which has the following advantages compared with the prior art:

[0040] This invention not only enables tracking of METs, including estimating their motion state, measurement rate state, extended state, and track, but also classifies them and uses the classification results to correct the extended state, thus improving extended state estimation performance. Compared to the ET-JTC-RHM-PHD filter, the proposed method achieves better tracking and classification performance; compared to the ET-JTC-GP-LMB filter, the proposed method offers better real-time performance. Attached Figure Description

[0041] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0042] Figure 1This is a flowchart of the MET JTC method based on star convex RHM and LMB filters of the present invention;

[0043] Figure 2 This represents the actual trajectory of the target in the simulation experiment.

[0044] Figure 3 The average result of 50 Monte Carlo simulations: Figure 3 (a) presents the results of the target number estimation. Figure 3 (b) shows the OSPA distance in motion. Figure 3 (c) Shows the OSPA distance in measurement rate status. Figure 3 (d) shows the OSPA distance in the extended state. Figure 3 (e) shows the class recognition rate. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] like Figure 1 As shown, an embodiment of the present invention provides a MET JTC method based on star-convex RHM and LMB filters, comprising the following steps:

[0047] Step 1: Initialization. The LMB parameter set is initialized using prior target information to obtain the initial LMB parameter set. Where l0 represents the target label, which represents a hypothetical trajectory, and r0(l0) represents the probability of the existence of the hypothetical trajectory l0. This represents the parameter set of the probability density mass function (PDMF) of l0. This represents the label space at the initial moment. In this invention, the PDMF of l0 comprehensively considers the target's measurement rate state γ0 and motion state. Extended state The class state c0 is described below:

[0048] (1) γ0 is a scalar, modeled as a gamma distribution. Where α0 and β0 represent the shape parameter and the inverse scale parameter, respectively.

[0049] (2) Indicates the target location. Indicates the target speed. Modeled as Gaussian distribution in, and Let represent the mean and covariance, respectively.

[0050] (3) It is a product of n F The vector formed by the Fourier coefficients of order 2n, totaling 2n. F +1 dimension. In this invention, the target shape is modeled as a star-convex shape, and its boundary can be determined by a radial function. This radial function can be expanded into the following Fourier series form:

[0051]

[0052] In the formula,

[0053] Because in this invention, the extended state It is influenced by both actual measurements and prior information about the target; therefore, this invention will Divided into two parts, and This indicates an extended state that depends on actual measurements. These represent extended states that depend on prior class information, and they are all modeled as Gaussian distributions. The former can be represented as... The latter can be formally expressed as Where c = 1, ..., n c n c Indicates the number of target categories.

[0054] (4) Define class probability

[0055] In summary, the PDMF of a single target l0 can be represented as a gamma-Gaussian-Gaussian mixture of states with different weights:

[0056]

[0057] In the formula, J0(l0) represents the number of terms in the mixed component, w 0,j (l0) represents the weight of the j-th mixture component. Thus, the parameter set corresponding to the PDMF of l0 can be expressed as: In addition, it should be noted that formula (2) is based on the assumption that the target measurement rate state, motion state and extended state are independent of each other, while formula (3) is based on the assumption that the two parts constituting the extended state are independent of each other.

[0058] Step 2: Measurement Division. The mean-shift algorithm is used to divide the true measurement set Z obtained from the sensors. r,k Divide into Z. r,k It consists of target measurement and clutter measurement. In this invention, the target measurement is modeled as follows:

[0059]

[0060] In the formula, This represents the j-th measurement generated by the i-th ET at time k. Represents the motion state measurement matrix. This represents the motion state of the i-th ET at time k. This represents the set of measurements generated by the ET. This indicates the number of measurements generated by the ET. Indicates the scaling factor. Indicates measurement The corresponding angle, R represents measurement noise. k This represents the measurement noise covariance. The measurement model can ultimately be summarized in the following form:

[0061]

[0062] Step 3: State Prediction. Assume the MET posterior PDMF parameter set at time k-1 is... The set of parameters for the newly generated MET PDMF at time k is: Therefore, the set of MET PDMF parameters predicted at time k can be calculated as follows:

[0063]

[0064] In the formula, Indicates the live MET PDMF parameter set, Represents the label space at time k-1. Let k represent the label space of newly generated targets at time k.

[0065] The parameters in the survival MET PDMF parameter set are calculated as follows:

[0066] (1) The existence probability, the number of mixed component terms, and the weights of the mixed components are calculated as follows:

[0067] r S,k|k-1 (l k )=η S,k|k-1 (l k )r k-1 (l k-1 (8)

[0068] JS,k|k-1 (l k ) = J k-1 (l k-1 (9)

[0069]

[0070] In the formula, p S,k This represents the survival probability of the target, which is assumed to be a constant independent of the target's state in this invention.

[0071] (2) The measurement rate status parameters are calculated as follows:

[0072]

[0073] In the formula, η γ This represents the forgetting factor.

[0074] (3) The motion state parameters are calculated as follows:

[0075]

[0076] In the formula, Represents the motion state transition matrix. This represents the noise covariance of the motion state process. Furthermore, in this invention, the motion state also carries an auxiliary variable during the recursive process. It is a symmetric positive definite random matrix that follows an inverse Wissaud distribution. The existence of this variable helps improve the accuracy of the target's motion state estimation. The prediction process for its parameters is as follows:

[0077]

[0078] In the formula, τ represents the attenuation factor, and d represents the dimension of the target motion space.

[0079] (4) The extended state parameters are calculated as follows:

[0080]

[0081] In the formula, This represents the extended state transition matrix. This represents the noise covariance of the extended state process.

[0082] (5) The class state parameters are calculated as follows:

[0083]

[0084] Furthermore, in the above formula, we have l k =l k-1 .

[0085] Finally, in this invention, the new MET PDMF parameter set is obtained by sampling based on the target prior information.

[0086] Step 4: Measurement Update. Measurement update includes three steps: converting LMB to Generalized LMB (GLMB), updating GLMB, and converting GLMB back to LMB.

[0087] (1) Convert LMB to GLMB

[0088] Assume the MET PDMF predicted at time k is a parameter set as follows The LMB random finite set (RFS) can then be transformed into a parameter set as... GLMB RFS, where, I represents the weights of the GLMB hypothesis components. k|k-1 This represents the set of target labels within the hypothetical component. Furthermore, Let h(·) be the exponential representation of the real-valued function, and let X represent a set.

[0089] (2) GLMB update

[0090] The parameter set updated by GLMB can be represented as in, This represents the weights of the hypothetical components in the updated GLMB model, and θ represents the association mapping relationship. Indicates Z r,k A measurement classification, This represents a pseudo-measurement set composed of prior information. In this invention, the introduced prior class information is the target extended state information, therefore g c This represents the extended state information of the c-th type of target. This represents the updated set of single-objective PDMF parameters under the association mapping θ. When the association mapping θ(l k When ) = 0, it means l k Missed detection, while when θ(l) k When ) > 0, it means l k It was detected. Therefore, The calculation needs to be discussed in two cases: missed detection and detected detection. The calculation method of the parameter set after GLMB update is introduced below.

[0091] The GLMB hypothesis assumes that the label set, weights, number of terms in a single target mixture component, and weight of the j-th term are calculated as follows:

[0092] I k =I k|k-1(twenty three)

[0093]

[0094]

[0095] In the formula, The j-th component of a single target mixture represents the unnormalized weight. Its calculation requires discussion of both missed and detected cases, and therefore will be placed in the parameter set later. The calculation is described in the section.

[0096] The parameter set is described below. The calculation method.

[0097] For θ(l) k ) = 0, parameter set The calculation is as follows:

[0098] ① The calculation methods for the number of terms in the mixed component and the weight of the j-th term are given in formulas (25)-(27), where The calculation is as follows:

[0099]

[0100] In the formula, p D,k This represents the detection probability. In this invention, it is assumed to be a constant independent of the target state.

[0101] ②The measurement rate status parameters are calculated as follows:

[0102]

[0103] ③ The motion state parameters are calculated as follows:

[0104]

[0105] ④ The extended state parameters are calculated as follows:

[0106]

[0107]

[0108] ⑤ The state parameters are calculated as follows:

[0109]

[0110] For θ(l) k ) > 0, parameter set The calculation is as follows:

[0111] ① The calculation methods for the number of terms in the mixed component and the weight of the j-th term are given in formulas (25)-(27), where The calculation is as follows:

[0112]

[0113] In the formula, κ represents the clutter intensity. Indicate l k After correlation mapping θ, the measurement is divided. The unit corresponding to the division unit in the middle. R represents the number of measurements in the partitioned unit, ρ represents the adjustment factor, and R represents the number of measurements in the partitioned unit. k Γ(·) represents the true measurement noise covariance, and Γ(·) represents the gamma function.

[0114] ②The measurement rate status parameters are calculated as follows:

[0115]

[0116] ③ The motion state parameters are calculated as follows:

[0117]

[0118]

[0119] ④ The extended state parameters are calculated as follows:

[0120] In this invention, the extended state The measurement updates are performed using an unscented Kalman filter. First, the extended state is augmented as follows: Its mean is covariance is Then, an unscented transformation is performed on the augmented state to obtain the sigma point and its weight coefficients:

[0121]

[0122] In the formula, Indicates the sigma point of the sample. and These represent the weights of the corresponding mean and covariance, respectively. The dimension of the augmented state is represented by the matrix root. Representation matrix The i-th column, and These are the parameters for the unscented transformation. Then, a nonlinear transformation is performed on the sigma point according to the measurement model shown in formula (6):

[0123]

[0124] In the formula, Indicates measurement division unit The m-th measurement in the dataset. The augmented extension state after measurement update using this measurement is calculated as follows:

[0125]

[0126] Formulas (55)-(62) are used The method for nonlinear transformation of the m-th measurement and measurement update of the augmented state is described. Since the range of values ​​for m is... Therefore, the above formula should be executed cyclically. Only after this time can it be used. Mean of augmented extended state after measurement update Covariance Note that when executing the above formula repeatedly, each iteration should be based on the result of the previous iteration. Finally, the desired result can be obtained. Extended status after measurement update:

[0127]

[0128]

[0129] In the formula, Indicates taking The first 2n F +1 dimension, Indicates taking The first 2n F +1 row, first 2n F +1 column.

[0130] Extended state The measurement update method is as follows:

[0131]

[0132] For simplicity, in formula (66) Abbreviated as

[0133] ⑤ The state parameters are calculated as follows:

[0134]

[0135] (3) Convert GLMB to LMB

[0136] The updated GLMB parameter set is achieved through first-order moment matching. It can be approximated as the LMB parameter set The calculation is as follows:

[0137]

[0138]

[0139] In the formula, This means that exactly Z r,k The measurement partitioning that divides the measurement into i units. Represents the associated mapping space, Represents parameter set The corresponding single-objective posterior PDMF.

[0140] Finally, the LMB parameter set It is represented in its parameterized form

[0141] Step 5: Pruning and Blending. After the measurement update is completed, the obtained LMB parameter set needs to be pruned and blended. The process involves trimming and merging. This step includes two aspects: first, trimming and truncating the LMB parameter set; and second, merging the parameter set... Trimming and merging are performed. These two aspects will be introduced separately below.

[0142] (1) Pruning and truncation of the LMB parameter set

[0143] By comparing the probability of existence of each hypothetical trajectory With pruning threshold r τ less than r τ The assumed trajectories will be discarded directly, while the remaining assumed trajectories will be retained. Then, the number of pruned LMB components and the preset maximum number of trajectories T will be calculated. max Compare them. If it is greater than the maximum number of trajectories T... max Then, based on their existence probabilities, the LMB components are sorted from largest to smallest, and the top T components are retained. max One LMB component, the rest discarded.

[0144] (2) For the parameter set Perform trimming and merging

[0145] ①Assign the weight of each blending component to the preset trimming threshold T p Compare and discard those with weights less than T. p The components. ② Calculate the similarity between any two components j and i using the mean and covariance of the motion states. If The value is less than the preset fusion threshold T u Then, these two components are merged. The weight of the merged component is the sum of the weights of the components before merging, and other parameters are the weighted average of the components before merging. ③ The number of the trimmed and merged mixed components is compared with the pre-set maximum number of components J for each hypothetical trajectory. max Compare them. If the number of components is still greater than J, then... maxThen sort the components by weight from largest to smallest, and keep the top J. max ④ Normalize the weights of the trimmed and merged mixed components.

[0146] Step 6: State estimation. Assume the trimmed and merged LMB parameter set is as follows: Therefore, the number of ETs can be estimated as follows: Where ρ k (n) represents the potential distribution of the LMB parameter set. Then, the parameter set... All hypothetical trajectories are arranged in descending order according to their respective existence probabilities, starting from the first... State extraction is performed on each hypothetical trajectory. The estimated ET state set is then obtained. The specific calculations are as follows:

[0147]

[0148] Step 7: Determine if the process has ended. Check if the observation has ended. If so, end the filtering process; otherwise, proceed to Step 3 to begin the filtering process for the next time step. It should be noted that before starting the filtering process for the next time step, the measurement set for the next time step should be divided, i.e., the operation in Step 2 should be performed.

[0149] In practice, the timing of the measurement in step 2 is quite flexible; it only needs to be executed before step 4 begins. Therefore, it can be executed before or after step 3. In this embodiment, it is executed before step 3 begins.

[0150] The technical effects of the present invention will be further described below in conjunction with simulation experiments:

[0151] 1. Simulation conditions

[0152] The monitored area measures [-1000, 1000]m × [-1000, 1000]m, and two ETs appear within this area. The actual movement trajectories of the targets are as follows: Figure 2 As shown, the black circle represents the target's starting position, and the black triangle represents the target's ending position. Target 1 is a rectangle, 22m long and 16m wide. Target 2 is an ellipse, with a major axis of 18m and a minor axis of 16m. The survival time of both targets is 1-100s. The target motion model is a constant velocity model, with a sampling interval of 1s and a process noise standard deviation of 1m / s. 2The observation duration for the entire monitoring area was 100 seconds. Clutter was uniformly distributed throughout the monitoring area, with an average value of 10. The target survival probability was set to 0.99, the detection probability to 0.9, and the standard deviation of measurement noise to 0.1 m. Tracking results were evaluated using Optimal Sub-Pattern Assignment (OSPA) distance, with the cutoff distance and sensitivity parameters set to c = 50 and p = 1, respectively. Classification results were evaluated using class recognition rate.

[0153] The methods used for comparison include the ET-RHM-LMB filter, the ET-JTC-RHM-PHD filter, the ET-JTC-GP-LMB filter, and the ET-JTC-RHM-LMB filter. Among these, the ET-RHM-LMB filter represents the version of the present invention considered only for tracking. The ET-JTC-RHM-PHD filter and the ET-JTC-GP-LMB filter are the two most advanced MET JTC methods currently available; for details, please refer to the background section. The ET-JTC-RHM-LMB filter is the technical solution proposed in this invention.

[0154] 2. Simulation Results and Analysis

[0155] The average result of 50 Monte Carlo simulations is as follows Figure 3 As shown. Figure 3 (a) presents the results of the target number estimation. Figure 3 (b) shows the OSPA distance in motion. Figure 3 (c) Shows the OSPA distance in measurement rate status. Figure 3 (d) shows the OSPA distance in the extended state. Figure 3 (e) shows the class recognition rate.

[0156] analyze:

[0157] (1) From Figure 3 As can be seen, compared with the ET-RHM-LMB filter, the proposed technical solution can not only track MET, but also classify MET, and can use the classification results to correct the extended state estimation performance, thus having a better extended state estimation effect.

[0158] (2) From Figure 3 It can also be seen that, compared with the ET-JTC-RHM-PHD filter, the proposed technical solution has better tracking and classification performance, including more accurate target number estimation, motion state estimation, measurement rate state estimation, extended state estimation and higher class recognition rate.

[0159] (3) From Figure 3The results also show that, compared to the ET-JTC-GP-LMB filter, the proposed solution achieves the same performance in target number, motion state, and measurement rate state estimation, while its extended state estimation performance is slightly worse, and its class recognition rate convergence is slightly slower. However, statistics show that in terms of runtime, the ET-JTC-GP-LMB filter takes 21.0169 seconds, while the proposed solution takes only 9.4359 seconds, reducing the runtime to approximately 44.90% of that of the ET-JTC-GP-LMB filter, significantly reducing computational load. Overall, while meeting tracking and classification performance requirements, the proposed solution has better engineering application value.

[0160] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A multi-extent object joint tracking and classification method based on star-convex RHM and LMB filters, characterized in that, The method comprises the following steps: Step 1: initialization The target prior information is used to initialize the label multi-Bernoulli (LMB) parameter set, and an initial LMB parameter set is obtained, which comprises a hypothesis track existence probability, a probability density mass function (PDMF) parameter set, and the PDMF comprehensively considers the measurement rate state, the motion state, the extension state and the class state of the target; Step 2: measurement division The real measurement set obtained by the sensor is divided by using a mean shift algorithm, and the real measurement set is composed of target measurements and clutter measurements; Step 3: state prediction Based on the posterior PDMF parameter set of the multi-extended target (MET) at k-1 and the PDMF parameter set of the new MET at k, the predicted MET PDMF parameter set at k is calculated, and the prediction process comprises survival MET parameter calculation and new MET parameter sampling; Step 4: measurement update The operations of LMB to GLMB, GLMB update and GLMB to LMB are sequentially performed, and the GLMB update process needs to distinguish between target missed detection and detection to calculate parameters respectively; Step 5: pruning and fusion The LMB parameter set after measurement update is pruned and truncated, and the mixed components in the PDMF parameter set are pruned, merged and weight-normalized; Step 6: state estimation The number of targets is estimated based on the pruned and fused LMB parameter set, and the pre-extraction... The system uses the state information of the trajectory with a high probability assumption to complete the estimation of target motion, measurement rate, expansion and class state; Step 7: loop judgment It is checked whether the observation is ended, and if not, the next time filtering process is repeatedly executed, and the measurement division operation of step 2 needs to be re-executed before each loop.

2. The method of claim 1, wherein, The extended state in step 1 is divided into two parts, which are dependent on real measurement and prior class information , both of which are modeled as Gaussian distribution; the measurement rate state is modeled as Gamma distribution, the motion state is modeled as Gaussian distribution, and the measurement rate state, the motion state and the extended state are independent of each other.

3. The method of claim 1, wherein, The survival MET parameter calculation in step 3 comprises: The target existence probability, the number of mixed component items and the weight are updated based on the survival probability; The measurement rate state parameter is updated in a manner with a forgetting factor; The motion state parameter is updated in combination with a motion state transition matrix and a process noise covariance, and an auxiliary variable of an inverse Wishart distribution is carried to improve the estimation accuracy; The extension state parameter is updated based on an extension state transition matrix and a process noise covariance; The class state parameter at k-1 is directly inherited.

4. The method of claim 1, wherein, In the GLMB update in step 4, a pseudo measurement set composed of target extension state prior class information is introduced, the measurement rate state parameter update in the detection scenario needs to be combined with the number of measurements in the measurement division unit, the extension state is updated by using an unscented Kalman filter method for nonlinear measurement, and the class state parameter is updated based on the Bayesian rule in combination with the pseudo measurement information.

5. The method of claim 1, wherein, The pruning and fusion in step 5 specifically comprise: Hypothesis tracks with existence probabilities lower than a pruning threshold are discarded, and tracks exceeding a maximum track number are truncated in a descending order of existence probability; Mixed components with weights lower than a pruning threshold are discarded, components with similar motion states and satisfying a fusion threshold are merged, components exceeding a maximum component number are truncated in a descending order of weight, and finally the component weights are normalized.

6. The method of claim 1, wherein, In the state estimation in step 6, the number of targets is determined by the maximum value of the potential distribution of the LMB parameter set, the class state is determined by the class corresponding to the maximum class probability, and the extension state is obtained in combination with the measurement update result and the class information correction result.

7. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by a processor to realize the multi-extended target joint tracking and classification method based on star-convex RHM and LMB filter according to any one of claims 1 to 6.

8. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method of claim 1 to 6 for joint tracking and classification of multiple extended objects based on star-convex RHM and LMB filters.

9. An electronic device, comprising: Comprise: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to implement the method of claim 1 to 6 for joint tracking and classification of multiple extended objects based on star-convex RHM and LMB filters via executing the executable instructions.

Images