Generalized tag multi-Bernoulli tracking filtering method, system and device and storage medium

Abstract

The invention discloses a generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters, and relates to the field of multi-target tracking. At present, target tracking has the problem of too high operation complexity. The method comprises the following steps: generating new multi-target distribution required by a PHD filter and a GLMB filter according to a self-adaptive new model by using measurement data; measured data and a new target intensity function are sent to an unknown scene parameter estimator based on GM-PHD filtering, and estimation of detection rate and clutter intensity parameters is carried out; performing GLMB filtering by using the measured data and the scene parameters and new target distribution obtained by estimation to obtain a posterior probability density function of the iteration; and finally, extracting a multi-target tracking result from the posterior density. According to the technical scheme, efficient and accurate multi-target tracking under unknown scene parameters is realized by improving a beta Gaussian component merging process, combining an unknown clutter intensity filtering technology, combining a PHD filtering algorithm and a GLMB filter and other technical means.

Description

Technical Field

[0001] The present invention relates to the field of multi-target tracking, and in particular to a generalized label multi-Bernoulli tracking filtering method, system, device and storage medium. Background Art

[0002] Multi-target tracking technology can jointly estimate the number and state of time-varying targets, and has been widely used in machine vision, radar detection, navigation and positioning. In recent years, the multi-target tracking problem based on the Random Finite Set (RFS) theoretical framework has been studied by many scholars, and a variety of multi-target tracking filters have gradually been developed, such as Probability Hypothesis Density (PHD), Cardinalized PHD (CPHD), GLMB, etc. Among them, GLMB can output the label information corresponding to the state while estimating the target state, so it is more suitable for tracking scenarios that need to obtain target trajectory information. It is worth noting that the above filters all rely on a lot of prior information. However, in actual scenarios, the target detection probability, clutter intensity, and target new intensity information are usually not accurately known. If manually set based on experience, when the set parameters do not match the actual scene parameters, the target tracking effect will be deteriorated.

[0003] The paper "Tracking multiple targets from multi-static Doppler radar with unknown probability of detection" (Sensors, 2019, 19, 1672) proposes a GLMB tracking filter for unknown detection probability scenarios. Specifically, p d -CPHD filter is introduced to estimate the detection probability, and then the estimated value is sent to GLMB filter as a known detection probability to participate in tracking filtering. However, similarly, clutter intensity is also one of the scene parameters that cannot be known a priori. For this reason, the paper "Tracking multiple marine ships via multiple sensors with unknown backgrounds" (Sensors, 2019, 19, 5025) introduces λ-p which can simultaneously estimate clutter rate and detection probability. d–CPHD filter as an unknown scene parameter estimator. Furthermore, considering that the target regeneration intensity in the actual scene is also unknown, the document “Multi-object tracking with an adaptive generalized labeled multi-Bernoulli filter” (Signal Processing, 2022, 196, 108532) combines the data-driven adaptive regeneration model into the GLMB tracking filter. At this time, the tracking filter can realize the joint estimation of scene parameters and robust multi-target tracking in scenarios where the detection probability, clutter intensity and target regeneration intensity are unknown. The patent “A multi-target tracking method based on random finite set filter coupling” (application number: 201710124138, publication number: 106910205A) proposes a multi-target tracking filtering scheme that combines a robust multi-Bernoulli (RMB) filter with a GLMB filter. This method estimates the unknown clutter rate through the RMB filter, and then feeds the estimated result of the clutter rate into the GLMB filter to achieve robust online multi-target tracking. In fact, whether it is CPHD, MB filter or GLMB filter, they all have high computational complexity. Therefore, in order to improve the efficiency of the tracking filter and enhance its practicality, a GLMB filtering method for unknown scene parameters with less computational burden needs to be proposed. Summary of the invention

[0004] The technical problem to be solved and the technical task proposed by the present invention are to improve and perfect the existing technical solutions, and provide a generalized label multi-Bernoulli tracking filtering method with adaptive scene parameters, so as to adaptively adjust the parameters under the condition of unknown scene parameters, while maintaining the purpose of multi-target tracking with high computational efficiency and strong robustness. To this end, the present invention adopts the following technical solutions.

[0005] A generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters includes the following steps:

[0006] 1) Obtain the measurement data and generalized label multi-Bernoulli GLMB density of the previous iteration cycle, and use the measurement data and GLMB density of the previous iteration cycle to generate the new multi-target distribution required by the probability hypothesis density PHD filter and the GLMB filter according to the adaptive new model;

[0007] 2) The measurement data of the current iteration cycle and the new target distribution are sent to the unknown scene parameter estimator based on Gaussian mixture probability hypothesis density GM-PHD filtering to estimate the detection probability and clutter intensity parameters;

[0008] 3) Using the measured data and the estimated scene parameters and new target distribution, GLMB filtering is performed to obtain the posterior probability density function of this iteration;

[0009] 4) Extract the multi-target tracking results from the posterior density function.

[0010] This technical solution achieves efficient and accurate multi-target tracking under unknown scene parameters by improving the Beta-Gaussian component merging process, combining unknown clutter intensity filtering technology, PHD filtering algorithm and GLMB filter and other technical means. Specifically:

[0011] 1. Improve the Beta-Gaussian component merging process and enhance the accuracy of detection probability estimation: Under the condition of unknown detection probability, the present invention optimizes the Beta-Gaussian component merging process in the BGM-PHD filter, making the filter more accurate in estimating the detection probability, and effectively improving the target detection capability in complex environments.

[0012] Second, combined with the unknown clutter intensity filtering technology, comprehensive scene parameter estimation is achieved: In addition to the detection probability, the present invention also considers the key parameter of clutter intensity. By combining the BGM-PHD filter with the filtering technology of unknown clutter intensity, the scene parameter estimation values ​​of the detection probability and clutter intensity can be output simultaneously. This enables the filter to adapt to different working environments more comprehensively, improving the stability and reliability of the system.

[0013] 3. Combining PHD filtering algorithm and GLMB filter to achieve efficient multi-target tracking: The present invention combines the scene parameter estimator based on PHD filtering with the adaptive new model and GLMB filter. The PHD filtering algorithm itself has the advantage of low computational complexity and can efficiently handle multi-target tracking problems. The GLMB filter can output the state and label information of the target, providing strong support for subsequent trajectory analysis and recognition. The combination of the two enables the present invention to achieve efficient and accurate multi-target tracking under unknown scene parameters.

[0014] 4. Improve computing efficiency: Based on the low computational complexity of the PHD filtering algorithm, the present invention can still maintain a high computing efficiency when scene parameters are unknown, so that the present invention performs well in application scenarios with high real-time requirements.

[0015] As a preferred technical means: in step 1), using the measurement data Z of the previous iteration cycle k-1 ={z 1,k-1 ,…,z M,k-1}, and the GLMB density of the previous iteration cycle, generating the required regeneration strength of the PHD filter and the GLMB filter according to the adaptive regeneration model, including the steps of:

[0016] 101) Calculate each measurement z i,k-1 , i∈{1,…,M} and the posterior GLMB density π of the previous iteration cycle k-1 The probability r of association of all tracks in (X) U (z i );

[0017] 102) According to each measurement z i The probability of association r U (z i ), calculate the probability of the corresponding new target

[0018] 103) According to the probability of each new student and measurement, generating LMB new multi-target distribution π B+ (X).

[0019] This technical solution uses the data from the previous iteration cycle to achieve iterative updates, allowing the system to adjust the strength of new targets in real time according to the current situation, allowing the system to quickly respond to changes and appearances of targets.

[0020] By calculating the association probability between the measurement and all tracks in the posterior GLMB density of the previous iteration cycle, and calculating the probability of new targets based on these association probabilities, the accuracy of the intensity estimation of new targets is improved, false detection and missed detection are avoided, and a stable target trajectory is maintained.

[0021] The adaptive new model automatically adjusts the strength of new targets according to changes in the current scene. It can remain robust in the face of complex and changing environments and maintain stable tracking performance through adaptive adjustments even when the number of targets, motion patterns or scene parameters change.

[0022] By combining the PHD filter and the GLMB filter, and using the LMB new multi-target distribution to generate new intensity, the computation is efficient. The PHD filter itself has a low computational complexity, while the GLMB filter can handle multi-target tracking problems. The two are combined with an optimization algorithm and take advantage of the existing filters to reduce the computational burden while maintaining high performance.

[0023] As a preferred technical means: in step 2), the measurement data Z of the current iteration cycle k and the new target distribution π B+ (X) is sent to the unknown scene parameter estimator based on GM-PHD filtering to perform detection probability p d And the estimation of clutter intensity parameter λ specifically includes the following steps:

[0024] 201) From the new target distribution π B+Extract the new target strength function v from (X) B+ (x,a);

[0025] 202) According to the new target strength function v B+ (x, a), and the intensity function of the previous iteration cycle, to obtain the predicted intensity function v + (x,a); and according to the number of clutter N in the previous iteration cycle c,k-1 , get the predicted clutter number N c,+ ;

[0026] 203) According to the measurement data Z k , we get the intensity function v k (x,a) and the number of clutter N c,k Update results;

[0027] 204) According to the posterior strength function v k (x,a) and the number of clutter N c,k , extract the detection probability p d,k and clutter intensity λ c,k The estimated results.

[0028] This technical solution does not need to rely on too much prior information. Instead, it realizes real-time estimation of detection probability and clutter intensity through the new target distribution generated by the adaptive new model and the measurement data of the current iteration cycle, so that it can adapt to different scenarios and changes, improving the robustness and adaptability of the algorithm.

[0029] By extracting the new target intensity function from the new target distribution and combining it with the intensity function and clutter quantity of the previous iteration cycle, the predicted intensity function and predicted clutter quantity are obtained. Then, the prediction result is updated according to the measurement data of the current iteration cycle, and finally the estimation result of the detection probability and clutter intensity is obtained, which effectively improves the accuracy of the estimation.

[0030] The GM-PHD filtering algorithm itself has low computational complexity. By applying it to the unknown scene parameter estimator, it can reduce the computational burden while ensuring the estimation accuracy. In practical applications, it has high real-time performance and computational efficiency, meeting the high performance requirements of the multi-target tracking system.

[0031] In addition to the detection probability and clutter intensity parameters, this technical solution can also output information such as the posterior intensity function and the amount of clutter, which can be used to further analyze the target state, detect anomalies, or make decisions, etc. This provides strong support for subsequent target tracking, trajectory analysis, and recognition tasks.

[0032] As a preferred technical means: In step 3), using the measurement data Z k And the estimated scene parameter λk and p d,k 、New target distribution π B+ (X), perform GLMB filtering to obtain the posterior probability density function π of this iteration k (X), comprising the steps of:

[0033] 301) According to the new target strength function v B+ (x), and the GLMB density π of the previous iteration k-1 (X), get the predicted multi-target distribution π + (X);

[0034] 302) According to the predicted multi-target distribution π + (X), measurement data Z k , detection probability estimate p d,k and the clutter strength estimate λ k , and obtain the posterior multi-target distribution π k (X).

[0035] By taking the estimated detection probability and clutter intensity as scene parameters and combining them with the new target distribution and measurement data, the GLMB filter can more accurately describe the uncertainty of multi-target states, which helps to reduce false detections and missed detections and improve the accuracy and reliability of target tracking.

[0036] This technical solution does not rely on specific target motion models or scene parameters, but updates the multi-target distribution based on real-time measurement data and scene parameter estimation results. This enables the algorithm to adapt to various complex and changing environments and scenes, and has strong adaptability and robustness.

[0037] The posterior probability density function output by the GLMB filter not only contains the state information of the target, but also the association information and label information between targets, so that it can process multiple targets at the same time and provide rich target tracking results, providing strong support for subsequent target recognition, trajectory analysis and other tasks.

[0038] A second object of the present invention is to provide a generalized label multi-Bernoulli tracking filter system for adaptive scene parameters, comprising:

[0039] Adaptive new module: used to generate new multi-target distributions for PHD filter and GLMB filter based on the measurement data of the previous iteration cycle and the generalized label multi-Bernoulli (GLMB) density;

[0040] Scene parameter estimation module: receives the measurement data and new target distribution of the current iteration cycle, and estimates the detection probability and clutter intensity parameters based on Gaussian mixture probability hypothesis density (GM-PHD) filtering;

[0041] GLMB filtering module: GLMB filtering is performed using the measured data, estimated scene parameters and the new target distribution to generate the posterior probability density function of this iteration;

[0042] Result extraction module: extracts multi-target tracking results from the posterior probability density function.

[0043] The adaptive new module can generate a new multi-target distribution based on the measurement data of the previous iteration cycle and the GLMB density, so that the system can flexibly respond to changes in the number, state and motion mode of targets, improving the robustness and adaptability of the system.

[0044] The scene parameter estimation module is based on GM-PHD filtering. It can use the measurement data of the current iteration cycle and the distribution of new targets to accurately estimate key scene parameters such as detection probability and clutter intensity, which helps to improve the accuracy and stability of target tracking.

[0045] The GLMB filter is efficient and accurate in dealing with multi-target tracking problems. It can process multiple targets simultaneously and maintain the association and label information between them.

[0046] The result extraction module extracts the multi-target tracking results from the posterior probability density function, including information such as the target's status, association and label, and can provide a comprehensive multi-target tracking solution.

[0047] Since both the GM-PHD filter and the GLMB filter have low computational complexity, they can achieve efficient multi-target tracking while ensuring accuracy.

[0048] As the preferred technical means: Adaptive newborn module includes:

[0049] Association probability calculation unit: calculates the association probability of each measurement with all tracks in the posterior GLMB density of the previous iteration period;

[0050] Newborn probability calculation unit: calculates the probability of the corresponding newborn target according to the associated probability of each measurement;

[0051] New target distribution generation unit: Generates LMB new multi-target distribution according to each new probability and measurement.

[0052] The adaptive new module can dynamically calculate the association probability of each measurement with the existing track based on the posterior GLMB density of the previous iteration cycle and the current measurement data, so that the system can identify new targets in real time and track them effectively.

[0053] By calculating the association probability of each measurement and the corresponding new target probability, the system can more accurately determine which measurements correspond to newly appeared targets, helping to reduce the error of misidentifying measurements of clutter or disappeared targets as new targets, thereby improving tracking accuracy.

[0054] The new target distribution generation unit directly generates LMB new multi-target distribution according to the calculated new probability and measurement data, which reduces unnecessary calculation steps and improves the processing efficiency of the system.

[0055] The adaptive new module can cope with complex and changing scenarios, including changes in the number of targets, motion patterns, and appearance times. By dynamically adjusting the probability and distribution of new targets, it can maintain stable tracking performance in different scenarios.

[0056] By accurately calculating the probability of association between measurements and existing tracks and judging new targets based on this, the adaptive new module can reduce the probability of missing new targets.

[0057] As a preferred technical means: the scene parameter estimation module includes:

[0058] New target intensity extraction unit: extracting new target intensity function from new target distribution;

[0059] Prediction intensity function calculation unit: obtains the prediction intensity function according to the new target intensity function and the intensity function of the previous iteration period, and obtains the prediction clutter quantity according to the clutter quantity of the previous iteration period;

[0060] Update intensity function and clutter quantity unit: obtain update results of intensity function and clutter quantity according to measurement data;

[0061] Detection probability and clutter intensity estimation unit: extracts the estimation results of detection probability and clutter intensity according to the posterior intensity function and the number of clutter.

[0062] The module estimates parameters based on real-time measurement data and new target distribution, rather than relying on fixed preset values ​​or empirical values. This makes the estimation results closer to the actual scene and improves the accuracy and reliability of parameter estimation.

[0063] Detection probability is one of the important scene parameters in target tracking, which directly affects the performance of the tracking algorithm. By calculating the posterior intensity function and the amount of clutter, the estimated result of the detection probability is extracted, which can provide accurate prior information for subsequent multi-target tracking filtering and improve the tracking accuracy and stability.

[0064] Clutter intensity is an important parameter that reflects the noise and interference level in the sensor measurement environment. By predicting and updating the intensity function and combining it with the measurement data, the scene parameter estimation module can dynamically estimate the clutter intensity, providing an effective noise suppression method for the subsequent filtering algorithm and reducing the possibility of false detection and missed detection.

[0065] This technical solution can adapt to the parameter estimation requirements in different scenarios. Regardless of the changes in the number of targets, motion patterns or clutter levels, the scene parameter estimation module can accurately estimate parameters through real-time data driving to maintain the tracking performance of the system.

[0066] Accurate scene parameter estimation can improve the robustness of the entire multi-target tracking system. Even in complex and changing environments, the system can maintain stable tracking performance through adaptive parameter estimation and reduce the impact of external interference on the tracking results.

[0067] As a preferred technical means: GLMB filtering module includes:

[0068] Predicted multi-target distribution generation unit: obtains predicted multi-target distribution according to the new target intensity function and the GLMB density of the previous iteration cycle;

[0069] Posterior multi-target distribution calculation unit: obtains the posterior multi-target distribution according to the predicted multi-target distribution, measurement data, detection probability estimation and clutter intensity estimation.

[0070] The GLMB filter module not only considers the new target intensity function when predicting the multi-target distribution, but also combines the GLMB density of the previous iteration cycle, which ensures the accuracy of the predicted distribution. At the same time, when calculating the posterior multi-target distribution, it uses measurement data, detection probability estimation and clutter intensity estimation, which further improves the accuracy of the filtering results.

[0071] The collaborative work between the predicted multi-target distribution generation unit and the posterior multi-target distribution calculation unit enables the entire filtering process to be carried out efficiently. The prediction step provides the necessary prior information for the posterior calculation, which is updated based on the prediction results and real-time measurement data, achieving fast and accurate multi-target tracking.

[0072] The GLMB filter module can adapt to the needs of multi-target tracking in different scenarios. Regardless of the number of targets, motion patterns, detection probability, and clutter intensity, the module can accurately filter and update through real-time data drive to maintain the system's tracking performance.

[0073] Since the GLMB filter module fully considers the impact of scene parameters such as detection probability and clutter intensity during the filtering process, it can maintain stable tracking performance in complex and changing environments. Even in the presence of a large amount of clutter or when targets frequently appear and disappear, the module can effectively suppress noise and interference, reducing the possibility of false detection and missed detection.

[0074] The third object of the present invention is to provide a computer device; the computer device includes one or more processors and one or more memories, and at least one program code is stored in the one or more memories. When the program code is executed by the one or more processors, the generalized label multi-Bernoulli tracking filter system with adaptive scene parameters is implemented.

[0075] The fourth object of the present invention is to provide a storage medium, in which at least one program code is stored. When the program code is executed by a processor, the steps of the generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters are implemented.

[0076] Beneficial effects: This technical solution achieves efficient and accurate multi-target tracking under unknown scene parameters by improving the Beta-Gaussian component merging process, combining unknown clutter intensity filtering technology, combining PHD filtering algorithm and GLMB filter and other technical means. Specifically:

[0077] 1. Improve the beta-Gaussian component merging process and improve the accuracy of detection probability estimation: Under the condition of unknown detection probability, the present invention optimizes the beta-Gaussian component merging process in the BGM-PHD filter. This improvement makes the filter more accurate in estimating the detection probability and effectively improves the target detection capability in complex environments.

[0078] Second, combined with the unknown clutter intensity filtering technology, comprehensive scene parameter estimation is achieved: In addition to the detection probability, the present invention also considers the key parameter of clutter intensity. By combining the BGM-PHD filter with the filtering technology of unknown clutter intensity, the present invention can simultaneously output the scene parameter estimation values ​​of detection probability and clutter intensity. This feature enables the filter to adapt to different working environments more comprehensively and improves the stability and reliability of the system.

[0079] 3. Combining PHD filtering algorithm and GLMB filter to achieve efficient multi-target tracking: The present invention combines the scene parameter estimator based on PHD filtering with the adaptive new model and GLMB filter. The PHD filtering algorithm itself has the advantage of low computational complexity and can efficiently handle multi-target tracking problems. The GLMB filter can output the state and label information of the target, providing strong support for subsequent trajectory analysis and recognition. This combination enables the present invention to achieve efficient and accurate multi-target tracking under unknown scene parameters.

[0080] 4. Improved computing efficiency: Based on the low computational complexity of the PHD filtering algorithm, the present invention can still maintain high computing efficiency when scene parameters are unknown. This advantage enables the present invention to perform well in application scenarios with high real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a principle block diagram of the present invention.

[0082] Figure 2 This is the multi-target tracking simulation scenario of the present invention.

[0083] Figure 3 It is the multi-target state estimation result of the present invention under unknown scene parameters.

[0084] Figure 4 It is the OSPA error of multi-target state estimation under unknown scene parameters of the present invention.

[0085] Figure 5 The multi-target state estimation OSPA under unknown scene parameters of the present invention (2) error.

[0086] Figure 6 It is the multi-target quantity estimation result of the present invention under unknown scene parameters.

[0087] Figure 7 It is the detection probability estimation result of the present invention under unknown scene parameters.

[0088] Figure 8 It is the clutter intensity estimation result of the present invention under unknown scene parameters. DETAILED DESCRIPTION

[0089] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0090] Embodiment 1:

[0091] like Figure 1 As shown, the present invention comprises the steps of:

[0092] S1: Get the measurement data and generalized label multi-Bernoulli GLMB density of the previous iteration cycle, and use the measurement data Z of the previous iteration cycle k-1 ={z 1,k-1 ,…,z M,k-1}, and the GLMB density of the previous iteration cycle, the new multi-objective distribution required by the PHD filter and the GLMB filter is generated according to the adaptive new model. The specific steps are as follows:

[0093] 1a: Calculate each measurement z i,k-1 , i∈{1,…,M} and the posterior GLMB density π of the previous iteration cycle k-1 The probability r of association of all tracks in (X) U (z i );

[0094] 1b: According to each measurement z i The probability of association r U (z i ), calculate the probability of the corresponding new target

[0095] 1c: Based on the probability of each new student and measurement, generating LMB new multi-target distribution π B+ (X).

[0096] S2: The measurement data Z of the current iteration cycle k and the new target distribution π B+ (X) is sent to the unknown scene parameter estimator based on GM-PHD filtering to perform detection probability p d And estimation of clutter intensity parameter λ: The specific steps include:

[0097] 2a: From the new target distribution π B+ Extract the new target strength function v from (X) B+ (x,a);

[0098] 2b: According to the new target strength function v B+ (x, a), and the intensity function of the previous iteration cycle, to obtain the predicted intensity function v + (x,a); and according to the number of clutter N in the previous iteration cycle c,k-1 , get the predicted clutter number N c,+ ;

[0099] 2c: According to the measured data Z k , we get the intensity function v k (x,a) and the number of clutter N c,k Update results;

[0100] 2d: According to the posterior strength function vk (x,a) and the number of clutter N c,k , extract the detection probability p d,k and clutter intensity λ c,k The estimated results.

[0101] S3: Using measurement data Z k And the estimated scene parameter λ k and p d,k 、New target distribution π B+ (X), perform GLMB filtering to obtain the posterior probability density function π of this iteration k (X), the specific steps include:

[0102] 3a: According to the new target intensity function v B+ (x), and the GLMB density π of the previous iteration k-1 (X), get the predicted multi-target distribution π + (X);

[0103] 3b: Based on the predicted multi-target distribution π + (X), measurement data Z k , detection probability estimate p d,k and the clutter strength estimate λ k , and obtain the posterior multi-target distribution π k (X).

[0104] S4: Extract multi-target tracking results from the posterior density function.

[0105] Compared with the prior art, the present invention has the following advantages:

[0106] First, the present invention improves the detection probability estimation accuracy of the BGM-PHD filter by improving the merging process of the Beta-Gaussian components under unknown detection probability, and combines the filter with the filtering technology of unknown clutter intensity to obtain a scene parameter estimator that can simultaneously output detection probability and clutter intensity. Second, the present invention combines the scene parameter estimator based on PHD filtering with the adaptive newborn model and GLMB filter, and uses the obtained detection probability and clutter intensity as scene parameters for multi-target tracking filtering. Thanks to the advantage of low computational complexity of the PHD filtering algorithm, the computational efficiency of the multi-target tracking filter under unknown scene parameters is improved.

[0107] In step S1, the measurement data of the previous iteration cycle is usually captured by sensors, cameras or other data acquisition devices in the real environment to describe the state or position of the target. In each iteration cycle, these measurement data are collected and provided as input to the tracking filter method. The GLMB density of the previous iteration cycle is calculated in the previous iteration step, which contains important information about the target state, such as the number, position, speed, etc. of the target. This density is used in step S1 to generate a new target distribution as the basis for subsequent iterations. The key to the iteration is that the output of step S3 (i.e., the posterior probability density function) will be used as the input of the GLMB density of the next iteration (i.e., the next iteration cycle). In the next iteration cycle, the updated posterior probability density function (or the information extracted from it) will be used as the initial value or input of the GLMB density, and then the process of steps 1) to 4) will be repeated. A closed-loop data processing flow is formed, which realizes the iterative update of multi-target tracking and the adaptive estimation of scene parameters.

[0108] The steps of the present invention are further described below:

[0109] Step 1: Use the measurement data Z from the previous iteration cycle k-1 ={z 1,k-1 ,…,z M,k-1}, as well as the GLMB density of the previous iteration cycle, the required regeneration intensity of the PHD and GLMB filters is generated according to the adaptive regeneration model.

[0110] (1.1) Assume that the posterior GLMB density π of the previous iteration k-1 (X) is written as

[0111] π(X)=Δ(X)∑ (I,ξ)∈F(L)×Ξ w (I,ξ) δ I (L(X))[p (ξ) ] X .

[0112] For each measurement z i,k-1 ,i∈{1,…,M}, and its association probability with each GLMB track is r U (z i ) is written as

[0113]

[0114] Among them, I,k It is the associated mapping set of a set of labels I in the current iteration cycle k. Indicator function 1 θ (z i ) ensures that the summation only considers the measurement z i Associated track assumptions.

[0115] (1.2) According to each measurement z i The probability of association r U (z i ), calculate the probability of the corresponding new target

[0116]

[0117] Among them, λ B+ is the expected number of new targets in the current iteration cycle, r B,max ∈[0,1] is the maximum probability of new birth.

[0118] (1.3) Multi-Bernoulli new distribution π of the current iteration cycle B+ (X) is given by

[0119]

[0120] in, Obey Gaussian distribution, weight is 1, mean

[0121] Covariance is set to a fixed value

[0122] Step 2: The measurement data Z of the current iteration cycle k and the new target distribution π B+ (X) is sent to the unknown scene parameter estimator based on GM-PHD filtering to perform detection probability p d and estimation of the clutter intensity parameter λ.

[0123] (2.1) From the new distribution π B+ Extract the new target strength function v from (X) B+ (x). The new target intensity function is represented by the following Gaussian mixture:

[0124]

[0125] In the parameter estimator based on GM-PHD filtering, the state space is augmented to x = [x, a], where a is the detection probability. Then

[0126]

[0127] in,

[0128] (2.2) According to the new target strength function v B+ (x,a), and the intensity function v of the previous iteration cycle k-1 (x,a) to get the predicted intensity function v + (x,a) is as follows

[0129]

[0130] Among them, p S (·) is the target survival probability, f + (x|·) is the Markov state transition density.

[0131] Secondly, according to the number of clutter N in the previous iteration cycle c,k-1 , calculate the number of clutter predictions N c,+

[0132]

[0133] in, is the expected number of new clutter, is the clutter survival probability.

[0134] (2.3) According to the measured data Z k , and the predicted intensity function v + (x, a), get the posterior strength function v of the current iteration cycle k (x,a)

[0135]

[0136] in, To monitor the uniform distribution density of clutter in the area, is the clutter detection probability.

[0137] Then, update the clutter number N c,k

[0138]

[0139] (2.4) According to the posterior strength function v k (x,a) and the number of clutter to estimate the detection probability p d,k and clutter intensity λ c,k .

[0140] Assume that the posterior strength function is expressed by

[0141]

[0142] To avoid underestimation of the detection probability, only the measured updated components are merged when merging the Beta-Gaussian components. The detection probability is calculated by the following formula

[0143]

[0144] Clutter intensity

[0145] Step 3: Use the measurement data Z k And the estimated scene parameter λ k and p d,k 、New target distribution π B+ (X), perform GLMB filtering to obtain the posterior probability density function π of this iteration k (X).

[0146] (3.1) According to the new target distribution π B+ (X), and the GLMB density π of the previous iteration k-1 (X), get the predicted multi-target distribution π + (X)

[0147]

[0148] in,

[0149]

[0150] (3.2) According to the predicted multi-target distribution π + (X), measurement data Z k , detection probability estimate p d,k and the clutter strength estimate λ k , and obtain the posterior multi-target distribution π k (X)

[0151] π k (X) = Δ(X)∑ (I,ξ)∈F(L)×Ξ ∑ θ∈Θ(I) w (I,ξ,θ) (Z)δ I (L(X))[p (ξ,θ) (x|Z)] X ,

[0152] in,

[0153]

[0154] In the formula,

[0155]

[0156] Step 4: From the posterior density π of this iteration cycle k (X) Extract the multi-target tracking results.

[0157] The present invention is further described below in conjunction with specific simulations.

[0158] 1.Simulation scenarios and parameters:

[0159] As Figure 2Taking the multi-target tracking scenario shown in the figure as an example, the adaptive GLMB filter in the present invention is used to track multiple targets. Assuming that the target moves according to the uniform velocity (CV) model, the target state vector is And the motion state is transferred according to the following formula

[0160]

[0161] in is the Kronecker product, I2 is the unit matrix of dimension 2, T s = 1. The process noise is subject to Zero-mean Gaussian white noise with standard deviation s w =5m / s 2 . The measurement is z = [x, y] T +v is given, the process noise is subject to v~N(0,s 2 I2) is a zero-mean Gaussian white noise with a standard deviation of s=10m.

[0162] Other relevant parameters in the present invention are shown in Table 1.

[0163] Table 1 Simulation parameters

[0164]

[0165]

[0166] 2. Simulation content and result analysis:

[0167] Simulation 1: Under the above simulation scenario and simulation parameters, a single tracking and filtering process is performed on the measurements in the monitoring area, and the following is obtained: Figure 3 The tracking results shown in Figure 3 It can be seen that the target track estimation results are basically normal, and there is no obvious increase in false alarms or missed detections caused by the mismatch between the detection probability and the clutter rate. From the output target trajectory, there is no false trajectory caused by clutter, nor is there any track interruption caused by missed detection, and it can be regenerated in a short time after the target appears.

[0168] Simulation 2: Under the above simulation scenario and simulation parameters, 100 Monte Carlo experiments were performed to record the average OSPA error, OSPA(2) error and target number estimation of the tracking filter results, as shown in Figures 4 to 6 As shown in the figure, it can be seen that both OSPA and OSPA(2) show high error peaks when the target is born or disappears. In particular, when the number of targets increases, the number of targets is underestimated. This is because the adaptive new model requires several cycles to gradually increase the existence probability value, which leads to missed detection for several cycles after the target is born.

[0169] Simulation 3: Under the above simulation scenario and simulation parameters, 100 Monte Carlo experiments are performed to record the average detection probability and clutter intensity estimation of the tracking filter results, such as Figure 7 , 8 As shown in the figure, it can be seen that the detection probability is underestimated when the target is born, but as the number of iterations increases, the estimated detection probability gradually approaches the true value. The clutter intensity shows a slightly underestimated trend as a whole, and is closer to the true value when the target is born.

[0170] Embodiment 2:

[0171] A generalized label multi-Bernoulli tracking filter system with adaptive scene parameters, comprising:

[0172] 1. Adaptive new module: used to generate new multi-target distributions for PHD filter and GLMB filter based on the measurement data of the previous iteration cycle and the generalized label multi-Bernoulli (GLMB) density;

[0173] include:

[0174] 1. Association probability calculation unit: calculates the association probability of each measurement with all tracks in the posterior GLMB density of the previous iteration period;

[0175] 2. New probability calculation unit: calculates the probability of the corresponding new target according to the associated probability of each measurement;

[0176] 3. New target distribution generation unit: Generate LMB new multi-target distribution according to each new probability and measurement.

[0177] The adaptive new module can dynamically calculate the association probability of each measurement with the existing track based on the posterior GLMB density of the previous iteration cycle and the current measurement data, so that the system can identify new targets in real time and track them effectively.

[0178] By calculating the association probability of each measurement and the corresponding new target probability, the system can more accurately determine which measurements correspond to newly appeared targets, helping to reduce the error of misidentifying measurements of clutter or disappeared targets as new targets, thereby improving tracking accuracy.

[0179] The new target distribution generation unit directly generates LMB new multi-target distribution according to the calculated new probability and measurement data, which reduces unnecessary calculation steps and improves the processing efficiency of the system.

[0180] The adaptive new module can cope with complex and changing scenarios, including changes in the number of targets, motion patterns, and appearance times. By dynamically adjusting the probability and distribution of new targets, it can maintain stable tracking performance in different scenarios.

[0181] By accurately calculating the probability of association between measurements and existing tracks and judging new targets based on this, the adaptive new module can reduce the probability of missing new targets.

[0182] 2. Scene parameter estimation module: Receives the measurement data and new target distribution of the current iteration cycle, estimates the detection probability and clutter intensity parameters based on Gaussian mixture probability hypothesis density (GM-PHD) filtering;

[0183] include:

[0184] 1. New target intensity extraction unit: extracting new target intensity function from new target distribution;

[0185] 2. Prediction intensity function calculation unit: obtains the prediction intensity function according to the new target intensity function and the intensity function of the previous iteration period, and obtains the predicted clutter quantity according to the clutter quantity of the previous iteration period;

[0186] 3. Update intensity function and clutter quantity unit: obtain the update results of intensity function and clutter quantity according to the measurement data;

[0187] 4. Detection probability and clutter intensity estimation unit: extracts the estimation results of detection probability and clutter intensity based on the posterior intensity function and the number of clutter.

[0188] The module estimates parameters based on real-time measurement data and new target distribution, rather than relying on fixed preset values ​​or empirical values. This makes the estimation results closer to the actual scene and improves the accuracy and reliability of parameter estimation.

[0189] Detection probability is one of the important scene parameters in target tracking, which directly affects the performance of the tracking algorithm. By calculating the posterior intensity function and the amount of clutter, the estimated result of the detection probability is extracted, which can provide accurate prior information for subsequent multi-target tracking filtering and improve the tracking accuracy and stability.

[0190] Clutter intensity is an important parameter that reflects the noise and interference level in the sensor measurement environment. By predicting and updating the intensity function and combining it with the measurement data, the scene parameter estimation module can dynamically estimate the clutter intensity, providing an effective noise suppression method for the subsequent filtering algorithm and reducing the possibility of false detection and missed detection.

[0191] This module can adapt to the parameter estimation needs in different scenarios. Regardless of the changes in the number of targets, motion patterns, or clutter levels, the scenario parameter estimation module can accurately estimate parameters through real-time data driving to maintain the tracking performance of the system.

[0192] Accurate scene parameter estimation can improve the robustness of the entire multi-target tracking system. Even in complex and changing environments, the system can maintain stable tracking performance through adaptive parameter estimation and reduce the impact of external interference on the tracking results.

[0193] 3. GLMB filtering module: GLMB filtering is performed using the measured data, estimated scene parameters and new target distribution to generate the posterior probability density function of this iteration;

[0194] include:

[0195] 1. Prediction multi-target distribution generation unit: According to the new target intensity function and the GLMB density of the previous iteration cycle, the prediction multi-target distribution is obtained;

[0196] 2. A posteriori multi-target distribution calculation unit: The a posteriori multi-target distribution is obtained according to the predicted multi-target distribution, measurement data, detection probability estimation and clutter intensity estimation.

[0197] The GLMB filter module not only considers the new target intensity function when predicting the multi-target distribution, but also combines the GLMB density of the previous iteration cycle, which ensures the accuracy of the predicted distribution. At the same time, when calculating the posterior multi-target distribution, it uses measurement data, detection probability estimation and clutter intensity estimation, which further improves the accuracy of the filtering results.

[0198] The collaborative work between the predicted multi-target distribution generation unit and the posterior multi-target distribution calculation unit enables the entire filtering process to be carried out efficiently. The prediction step provides the necessary prior information for the posterior calculation, which is updated based on the prediction results and real-time measurement data, achieving fast and accurate multi-target tracking.

[0199] The GLMB filter module can adapt to the needs of multi-target tracking in different scenarios. Regardless of the number of targets, motion patterns, detection probability, and clutter intensity, the module can accurately filter and update through real-time data drive to maintain the system's tracking performance.

[0200] Since the GLMB filter module fully considers the impact of scene parameters such as detection probability and clutter intensity during the filtering process, it can maintain stable tracking performance in complex and changing environments. Even in the presence of a large amount of clutter or when targets frequently appear and disappear, the module can effectively suppress noise and interference, reducing the possibility of false detection and missed detection.

[0201] 4. Result extraction module: extract multi-target tracking results from the posterior probability density function.

[0202] In summary, it can be seen that the adaptive new module can generate a new multi-target distribution based on the measurement data of the previous iteration cycle and the GLMB density, so that the system can flexibly respond to changes in the number, state and motion mode of targets, and improve the robustness and adaptability of the system.

[0203] The scene parameter estimation module is based on GM-PHD filtering. It can use the measurement data of the current iteration cycle and the distribution of new targets to accurately estimate key scene parameters such as detection probability and clutter intensity, which helps to improve the accuracy and stability of target tracking.

[0204] The GLMB filter is efficient and accurate in dealing with multi-target tracking problems. It can process multiple targets simultaneously and maintain the association and label information between them.

[0205] The result extraction module extracts the multi-target tracking results from the posterior probability density function, including information such as the target's status, association and label, and can provide a comprehensive multi-target tracking solution.

[0206] Since both the GM-PHD filter and the GLMB filter have low computational complexity, they can achieve efficient multi-target tracking while ensuring accuracy.

[0207] Embodiment three:

[0208] The functions that can be achieved by the generalized label multi-Bernoulli tracking filter system with adaptive scene parameters can all be completed by a computer device, which includes one or more processors and one or more memories, and at least one program code is stored in the one or more memories. The program code is loaded and executed by the one or more processors to implement the wireless signal processing method.

[0209] The processor takes out instructions from the memory one by one, analyzes the instructions, and then completes the corresponding operations according to the instruction requirements, generating a series of control commands, so that the various parts of the computer can automatically, continuously and coordinately move to become an organic whole, realize the input of programs, the input of data, and the calculation and output of results. The arithmetic operations or logical operations generated in this process are all completed by the operator; the memory includes a read-only memory (ROM), which is used to store computer programs, and a protection device is provided outside the memory.

[0210] Exemplarily, the computer program may be divided into one or more modules, one or more modules are stored in a memory and executed by a processor to implement the present invention. One or more modules may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in a terminal device.

[0211] Those skilled in the art will understand that the description of the above service equipment is merely an example and does not constitute a limitation on the terminal equipment. It may include more or fewer components than described above, or a combination of certain components, or different components, for example, it may include input and output devices, network access devices, buses, etc.

[0212] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire user terminal.

[0213] The memory can be used to store computer programs and / or modules. The processor can realize various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as an information collection template display function, a product information release function, etc.); the data storage area can store data created according to the use of the berth status display system (such as product information collection templates corresponding to different product types, product information that different product providers need to release, etc.). In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0214] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable medium. Based on this understanding, the present invention implements all or part of the modules / units in the above-mentioned embodiment system, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable medium, and the computer program can realize the functions of the above-mentioned various system embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The storage medium can be a computer-readable medium, and the computer-readable medium can include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0215] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters, characterized in that The following steps are involved: 1) Obtain the measurement data and generalized label multi-Bernoulli GLMB density of the previous iteration cycle, and use the measurement data and GLMB density of the previous iteration cycle to generate the new multi-target distribution required by the probability hypothesis density PHD filter and the GLMB filter according to the adaptive new model; 2) The measurement data of the current iteration cycle and the new target distribution are sent to the unknown scene parameter estimator based on Gaussian mixture probability hypothesis density GM-PHD filtering to estimate the detection probability and clutter intensity parameters; 3) Using the measured data and the estimated scene parameters and new target distribution, GLMB filtering is performed to obtain the posterior probability density function of this iteration; 4) Extract the multi-target tracking results from the posterior density function.

2. The generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters according to claim 1, characterized in that: In step 1), the measurement data Z of the previous iteration cycle is used k-1 ={z 1,k-1 ,…,z M,k-1 }, and the GLMB density of the previous iteration cycle, generating the required regeneration strength of the PHD filter and the GLMB filter according to the adaptive regeneration model, including the steps of: 101) Calculate each measurement z i,k-1 , i∈{1,…,M} and the posterior GLMB density π of the previous iteration cycle k-1 The probability r of association of all tracks in (X) U (z i ); 102) According to each measurement z i The probability of association r U (z i ), calculate the probability of the corresponding new target 103) According to the probability of each new student and measurement, generating new multi-target distribution of LMB 3. The generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters according to claim 2, characterized in that: In step 2), the measurement data Z of the current iteration cycle k and the new target distribution π B+ (X) is sent to the unknown scene parameter estimator based on GM-PHD filtering to perform detection probability p d And the estimation of clutter intensity parameter λ specifically includes the following steps: 201) From the new target distribution π B+ Extract the new target strength function v from (X) B+ (x,a); 202) According to the new target strength function v B+ (x, a), and the intensity function of the previous iteration cycle, to obtain the predicted intensity function v + (x,a); and according to the number of clutter N in the previous iteration cycle c,k-1 , get the predicted clutter number N c,+ ; 203) According to the measurement data Z k , we get the intensity function v k (x,a) and the number of clutter N c,k Update results; 204) According to the posterior strength function v k (x,a) and the number of clutter N c,k , extract the detection probability p d,k and clutter intensity λ c,k The estimated results.

4. The generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters according to claim 3, characterized in that: In step 3), using the measurement data Z k And the estimated scene parameter λ k and p d,k 、New target distribution π B+ (X), perform GLMB filtering to obtain the posterior probability density function π of this iteration k (X), comprising the steps of: 301) According to the new target strength function v B+ (x), and the GLMB density π of the previous iteration k-1 (X), get the predicted multi-target distribution π + (X); 302) According to the predicted multi-target distribution π + (X), measurement data Z k , detection probability estimate p d,k and the clutter strength estimate λ k , and obtain the posterior multi-target distribution π k (X).

5. A generalized label multi-Bernoulli tracking filter system with adaptive scene parameters, characterized in that include: Adaptive new module: used to generate new multi-target distributions for PHD filter and GLMB filter based on the measurement data of the previous iteration cycle and the generalized label multi-Bernoulli (GLMB) density; Scene parameter estimation module: receives the measurement data and new target distribution of the current iteration cycle, and estimates the detection probability and clutter intensity parameters based on Gaussian mixture probability hypothesis density (GM-PHD) filtering; GLMB filtering module: GLMB filtering is performed using the measured data, estimated scene parameters and the new target distribution to generate the posterior probability density function of this iteration; Result extraction module: extracts multi-target tracking results from the posterior probability density function.

6. The generalized label multi-Bernoulli tracking filter system for adaptive scene parameters according to claim 5, characterized in that: Adaptive Freshman Modules include: Association probability calculation unit: calculates the association probability of each measurement with all tracks in the posterior GLMB density of the previous iteration period; Newborn probability calculation unit: calculates the probability of the corresponding newborn target according to the associated probability of each measurement; New target distribution generation unit: Generates LMB new multi-target distribution according to each new probability and measurement.

7. The generalized label multi-Bernoulli tracking filter system for adaptive scene parameters according to claim 6, characterized in that: The scene parameter estimation module includes: New target intensity extraction unit: extracting new target intensity function from new target distribution; Prediction intensity function calculation unit: obtains the prediction intensity function according to the new target intensity function and the intensity function of the previous iteration period, and obtains the prediction clutter quantity according to the clutter quantity of the previous iteration period; Update intensity function and clutter quantity unit: obtain update results of intensity function and clutter quantity according to measurement data; Detection probability and clutter intensity estimation unit: extracts the estimation results of detection probability and clutter intensity according to the posterior intensity function and the number of clutter.

8. The generalized label multi-Bernoulli tracking filter system for adaptive scene parameters according to claim 7, characterized in that: GLMB filter module: Predicted multi-target distribution generation unit: obtains predicted multi-target distribution according to the new target intensity function and the GLMB density of the previous iteration cycle; Posterior multi-target distribution calculation unit: obtains the posterior multi-target distribution according to the predicted multi-target distribution, measurement data, detection probability estimation and clutter intensity estimation.

9. A computer device, characterized in that: The device includes one or more processors and one or more memories, wherein at least one program code is stored in the one or more memories. When the program code is executed by the one or more processors, a generalized label multi-Bernoulli tracking filter system with adaptive scene parameters as described in claim 6 or 7 is implemented.

10. A storage medium, wherein at least one program code is stored in the storage medium, characterized in that: When the program code is executed by a processor, the steps of a generalized label multi-Bernoulli tracking filtering method for adaptive scene parameters as described in any one of claims 1 to 5 are implemented.

Images