Multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation

Through the multi-sensor multi-objective elastic fusion tracking method based on Gaussian reliability propagation, the problems of high computational complexity and fault sensitivity of traditional algorithms are solved, and the multi-objective tracking with low complexity and high scalability are achieved, ensuring that the system is in failure Stable performance in the case.

CN118940215APending Publication Date: 2024-11-12HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411007361.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The traditional multi-sensor multi-objective information fusion algorithm has an exponential increase in the computational complexity when the number of sensors and targets increases, resulting in poor scalability and unable to effectively avoid system performance degradation caused by sensor failure.

Method used

Using a multi-sensor multi-objective elastic fusion tracking method based on Gaussian reliability propagation, the measurement data is allocated, the correlation vector is constructed, the fusion residual is calculated, fault detection and diagnosis is performed, and the calculation complexity is reduced and the system is improved. robustness.

Benefits of technology

It realizes low computational complexity and high scalability of multi-sensor multi-target tracking in cluttered environments, ensuring the system's fusion performance and target tracking accuracy in some sensor failures, and enhancing the robustness of the algorithm.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940215A_ABST
    Figure CN118940215A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation, and the method comprises the steps: detecting a plurality of detected targets in a detection region through a plurality of sensors in a clutter environment, verifying and screening original measurement, and executing data association, thereby completing the tracking of the plurality of detected targets. The distributed multi-sensor fault self-diagnosis is realized by using the fusion residual error, the fusion performance of the system is ensured on the premise that part of the sensors have faults, and the accuracy of target tracking is improved. Meanwhile, according to the tracking method, the multi-sensor multi-target tracking problem is described by using a reliability propagation algorithm and a probability graph model, the expandability of the algorithm is high, and the calculation complexity is low; while distributed multi-sensor information fusion is realized, the distributed nodes can perform fault self-diagnosis, and the robustness of target tracking is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of distributed multi-sensor information fusion, and in particular relates to a multi-sensor multi-target elastic fusion tracking method based on Gaussian credibility propagation. Background Art

[0002] With the continuous development of drones, unmanned boats and unmanned vehicles, future wars are characterized by flexibility, autonomy and distribution. This type of war is often referred to as "mosaic" warfare. In "mosaic" warfare, a large number of distributed sensor nodes with single functions are used to detect, locate and track multiple targets. However, as the types and number of sensor nodes increase, the probability of system damage due to sensor modeling errors, unexpected signal interference and undetected sensor failures also increases. Therefore, it is very important to develop efficient and reliable distributed multi-sensor elastic fusion and tracking technology.

[0003] The computational complexity of traditional multi-sensor multi-target information fusion algorithms tends to grow exponentially with the increase in the number of sensors and targets, so the scalability of the algorithm is poor. In battlefield environments, the operating environment of distributed multi-sensor multi-target tracking algorithms is generally in embedded systems. The resources of embedded systems are limited, especially in scenarios with massive sensors, the resources of embedded systems usually cannot meet the requirements. Therefore, while achieving elastic information fusion, it is of great practical value to develop a multi-target tracking algorithm with low computational complexity. At the same time, traditional multi-target information fusion algorithms cannot effectively avoid system damage when facing sensor erroneous measurements. When some sensor nodes fail and cause erroneous measurements to damage the system, it usually leads to a decrease in the performance of the multi-sensor information fusion algorithm. To this end, a multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation is proposed here. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and propose a multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation. While realizing elastic fusion of multi-sensor information, it reduces the computational complexity of the tracking algorithm and has excellent scalability, thereby avoiding the destruction of the system solution and the reduction of fusion performance due to partial faulty node measurements.

[0005] A multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation includes the following steps:

[0006] Step 1: Use the sensor to measure the target in the detection area to obtain the measurement data of the target Get the state vector of the k-1 measured target at the last moment and covariance Wherein, j=1,2,...,J; J is the number of measured targets; i=1,2,...,I; I is the number of sensors; n=1,2,...,M; M is the number of measurement data.

[0007] Step 2: Predict the state vector of target j at time k and covariance

[0008] Step 3: Allocate the measurement data measured by the sensor to different measured objects and use them as verification measurement data of the corresponding measured objects.

[0009] Step 4: Construct a correlation vector for the target being measured and the correlation vector for the measured data in, is the number of verified measurement data of the measured target j at sensor i at time k. and Get the correlation information from the measurement data to the measured object Correlation information from measurement data to the measured object Get data association information

[0010] Step 5: Associate information based on data Calculate measurement update information Calculate the state vector of target j at sensor i at the current moment k and covariance

[0011] Step 6: Obtain the fusion residual of the target And the covariance corresponding to the fusion residual According to the fusion residual of each sensor and covariance Perform fault detection on all sensors; combine the state vector of the normal sensor and covariance Get the state vector of target j at the current moment k and covariance

[0012] Step 7: Repeat steps 2 to 6 to obtain the state vector and covariance of the target j at different times to obtain the track of the target.

[0013] Preferably, in step 6, the fault detection method is: obtaining the fault diagnosis test variable of sensor i at time k Fault Diagnosis Test Variables The expression is:

[0014]

[0015] like The sensor i at time k is considered to be an abnormal sensor and does not participate in the target reliability calculation at time k; where α is the significance level; N k is the number of fusion residuals; D i is the dimension of the fused residual.

[0016] Preferably, in step 6, the sensor is diagnosed at regular intervals, and the diagnosis method is as follows:

[0017] Get the fault diagnosis test variable L of sensor i during this period of time i , fault diagnosis test variable L i The expression is:

[0018]

[0019] Where U is the accumulated time; is the innovation covariance, and its expression is:

[0020]

[0021] in, is the Jacobian matrix of sensor i with respect to target j at time k; is the measurement error covariance of sensor i at time k.

[0022] If L i > 2 (1-α,U×J×D i ), the corresponding sensor is placed in an offline state and does not participate in subsequent target tracking.

[0023] As a preferred method, in step 3, the method of allocating measurement data is: establishing a tracking gate according to the measured target The original measurement data is screened, and the measurement data falling into the overlapping area of ​​the tracking gates of different measured targets is assigned to the measured target that is closest. The measurement data falling into the tracking gate of measured target j is used as the verification measurement data of measured target j.

[0024] As a preferred embodiment, in step 6, the state vector and covariance The expression is:

[0025]

[0026] Among them, I k is the number of normal sensors at time k.

[0027] Preferably, in step 2, the predicted state vector and covariance The expressions are:

[0028]

[0029] Where F is the state transfer matrix; F T is the transpose of the state transfer matrix; Q k is the process noise error covariance.

[0030] Preferably, in the step 1, the generalized measurement model of the sensor constructed in the clutter environment is:

[0031]

[0032] in, is the nth measurement data obtained by sensor i at time k; is the measurement function of sensor i; is Gaussian white noise; It is the false alarm measurement data.

[0033] As a preferred embodiment, in the step 4, the correlation vector and The construction method is as follows:

[0034] Assume that each measured object j generates at most one measurement data at sensor i, and each measurement data at sensor i is generated by at most one measured object. for:

[0035]

[0036] Correlation vector for:

[0037]

[0038] As a preferred embodiment, in the step 5, the state vector and covariance The expression is:

[0039]

[0040] Among them, (h i ) -1 is the inverse function of the measurement function of sensor i; is the measurement error covariance of sensor i at time k; is the innovation covariance, and its expression is:

[0041]

[0042] in, is the Jacobian matrix of sensor i with respect to target j at time k.

[0043] As a preferred embodiment, in the step 4, the data association information The expression is:

[0044]

[0045] in, The expression is:

[0046]

[0047] Among them, P g is the threshold probability; is the detection probability of sensor i.

[0048] Correlation information from measurement data to the measured object The expression is:

[0049]

[0050] The expression is:

[0051]

[0052] As a preference, in said step 5, the measurement data update information The expression is:

[0053]

[0054] in, is the verification measurement data set of the measured target j at sensor i; is the state vector of the measured target j at time k; g(·) is the global likelihood function.

[0055] As a preferred embodiment, in step 6, the fusion residual The corresponding covariance The expression is:

[0056]

[0057] in, To update the weights; is the innovation covariance.

[0058] Update weights and innovation covariance The expressions are:

[0059]

[0060] in, is the Jacobian matrix of sensor i with respect to target j at time k.

[0061] The present invention has the following beneficial effects:

[0062] 1. The present invention uses multiple sensors to detect multiple targets in the detection area under a clutter environment, verifies and screens the original measurement data, performs data association, and completes the tracking of the targets. Distributed multi-sensor fault self-diagnosis is achieved by fusion residuals. Under the premise that some sensors fail, the fusion performance of the system is guaranteed, and the accuracy of target tracking is improved.

[0063] 2. The present invention uses a belief propagation algorithm and a probabilistic graphical model to describe the multi-sensor multi-target tracking problem. The algorithm has strong scalability and low computational complexity. While realizing distributed multi-sensor information fusion, distributed nodes can perform fault self-diagnosis, and the algorithm has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a principle flow chart of the present invention. DETAILED DESCRIPTION

[0065] The present invention will be further described below in conjunction with the accompanying drawings.

[0066] like Figure 1 As shown, a multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation includes the following steps:

[0067] Step 1: Use a sensor to measure the target in the detection area, where the sensor is a sonar and the target is a ship; obtain the state vector of the target at the last moment Covariance and target reliability

[0068] In this embodiment, the current moment is recorded as moment k, and the previous moment is recorded as moment k-1.

[0069] Step 2: State prediction

[0070] Prior information on the state vector and covariance of the target j at time k for:

[0071]

[0072] in, is the state transition function of target j at time k; is the target reliability at time k-1; is the state vector of the measured target j at time k, and They represent the position and velocity of target j in the Cartesian coordinate system respectively; j=1,2,...,J; J is the number of targets being measured.

[0073] Under Gaussian conditions, equation (1) can be rewritten as:

[0074]

[0075] Among them, Z k-1 is the verification measurement data set at time k-1; N is the Gaussian distribution; F is the state transfer matrix; Q k is the process noise error covariance; is the predicted state vector at time k; is the actual state vector obtained at time k-1; is the predicted covariance at time k; is the actual covariance obtained at time k-1; ∝ is the proportional sign.

[0076] According to formula (2), the predicted state vector is obtained and the predicted covariance The expressions are:

[0077]

[0078] Step 3: Measurement Data Verification

[0079] 3-1. Constructing a generalized measurement model for sensors

[0080] The generalized measurement model of the sensor in a clutter environment is:

[0081]

[0082] in, is the nth measurement data obtained by sensor i at time k; is the measurement function of sensor i; is Gaussian white noise; It is the false alarm measurement data.

[0083] The measurement data generated for the measured object Assume that it is subject to zero-mean Gaussian white noise with known covariance Interference. For measurement data generated by false alarms From the probability density function Description, probability density function The expression is:

[0084]

[0085] in, is the tracking gate area of ​​target j at sensor i at time k.

[0086] 3-2. Measurement data segmentation

[0087] In a single target scenario, the sensor can directly obtain the target j tracking gate at time k. In multi-target scenarios, the sensor needs to establish a tracking gate based on the target prediction center. The original measurement data is screened, and the measurement data falling into the overlapping area of ​​the tracking gates of different measured targets is allocated to the measured target corresponding to the nearest target prediction center. In order to ensure the independence of residual samples between different targets, a measurement data segmentation method and allocation matrix are proposed to achieve measurement data allocation. After sensor i establishes a tracking gate for measured target j and obtains the verification measurement data of measured target j, the allocation matrix E established is as follows:

[0088]

[0089] Among them, e nj is a binary group, e nj ∈{0,1};e nj = 0 means that the nth measurement data does not fall into the tracking gate of the measured target j; e nj =1 indicates that the nth measurement data falls into the tracking gate of target j; n=1, 2, ..., M; M is the number of measurement data acquired by sensor i; i=1, 2, ..., I; I is the number of sensors.

[0090] For the measurement data that only falls into the tracking gate of a single measured target, that is, The measurement data is assigned to the corresponding measured target; for the measurement data falling into the overlapping area of ​​the tracking gates of different measured targets, that is, Assign the measurement data to the measured target j corresponding to the nearest target prediction center nearest , get the measured target j nearest The method is as follows:

[0091]

[0092] in, The tracking gate for the nth measurement data to the measured target j Distance from center.

[0093] After segmenting the measurement data of sensor i, the measurement data falling into the measured target j is used as the verification measurement data of the measured target j, and the tracking gate falling into the measured target j is obtained. Verification measurement data in, is the verification measurement data set of the measured target j at sensor i; is the mth verification measurement data of the measured target j at sensor i; is the number of verification measurement data of target j at sensor i,

[0094] 3-3. Constructing the association vector

[0095] Assume that each measured target j generates at most one measurement data at sensor i, and each measurement data at sensor i is generated by at most one measured target. In order to describe the association between the measured target and the measurement data, the association vector for the measured target is introduced and the correlation vector for the measured data The possible association situations occurring at sensor i at time k are represented by the association vector Description, association vector for:

[0096]

[0097] Correlation vector for measurement data Because the sensor measurement data is segmented, It can be further expressed as in for:

[0098]

[0099] Correlation vector and can fully and clearly describe the event occurring at sensor i at time k, so the association vector and They are equivalent in expressing related events.

[0100] Step 4: Measurement Data Evaluation

[0101] In order to obtain the joint posterior probability density function f(X, a|Z) of multiple targets, f(X, a|Z) is factorized into:

[0102]

[0103] Among them, f(X0) is the probability density function of the prior information of all measured targets at the initial moment; f(Z t |X t ,at ) is the likelihood function of all detected targets at time t; t=1,2,...,k; p(a t ) is the prior probability mass function of the correlation vector at time t.

[0104] The association vector at sensor i at time k The prior probability mass function The expression is:

[0105]

[0106] in, Exclude indicator function for the target.

[0107] Target Exclusion Indicator Function The expression is:

[0108]

[0109] Exclude indicator function by target Excluding the assumption that each measured object j generates at most one measurement data at sensor i, each measurement data at sensor i is generated by at most one measured object.

[0110] The expression is:

[0111]

[0112] Among them, P g is the threshold probability; is the detection probability of sensor i.

[0113] Because at sensor i, the target-oriented association vector and the correlation vector for the measured data There is a one-to-one correspondence between them, so the target excludes the indicator function can be formally replaced by Target Exclusion Indicator Function The expression is:

[0114]

[0115] Goal-Oriented Relevance Vector The prior probability mass function can be formally replaced by Prior probability mass function The expression is:

[0116]

[0117] The expression is:

[0118]

[0119] Step 5: Data Association

[0120] Through the message passing rules of the probabilistic graphical model, the association information from the measured data to the measured target is obtained; in the data association link, because the sensor measurement data is verified and segmented in step 2, the corresponding probabilistic graphical model is a tree structure, and the data association link only needs to be performed once.

[0121] Correlation information from measurement data to the measured object The expression is:

[0122]

[0123] According to the related information Get data association information Data association information The expression is:

[0124]

[0125] Step 6: Update measurement data

[0126] Getting data related information After that, according to the message passing rules of the probabilistic graphical model, the measurement data update information is obtained.

[0127]

[0128] Among them, g(·) is the global likelihood function.

[0129] Under Gaussian conditions, in, is the actual state vector obtained at time k; is the actual covariance obtained at time k.

[0130] According to formula (19), the state vector is obtained and covariance The expression is:

[0131]

[0132] Among them, (h i ) -1 is the inverse function of the measurement function of sensor i; is the measurement error covariance of sensor i at time k; is the innovation covariance, and its expression is:

[0133]

[0134] in, is the Jacobian matrix of sensor i with respect to target j at time k.

[0135] After getting the measurement data update information, get the associated vector The probability of m is used as the verification measurement data Update weights, update weights The expression is:

[0136]

[0137] The mth verification measurement data of the measured object j The corresponding residual The expression is:

[0138]

[0139] in, is the predicted measurement data of target j, and its expression is:

[0140]

[0141] Fusion residual The expression is:

[0142]

[0143] Assume that each residual follows an independent and identical distribution, that is, Therefore, the fusion residual The corresponding covariance The expression is:

[0144]

[0145] Step 7: Troubleshooting

[0146] According to the fusion residual and covariance Get the fault diagnosis test variable of sensor i at time k Fault Diagnosis Test Variables The expression is:

[0147]

[0148] like The sensor i at time k is considered to be an abnormal sensor and does not participate in the target reliability calculation at time k; where α is the significance level; N k is the number of fusion residuals; D i is the dimension of the fusion residual;2 (·) is the chi-square distribution.

[0149] Step 8: Obtain the reliability of the measured target j through normal sensors Marginalization of the joint posterior probability density function obtains the reliability of the measured target j Reliability of the measured target j The expression is:

[0150]

[0151] Among them, I k is the number of normal sensors at time k.

[0152] Under Gaussian conditions, the reliability of the measured target It can be further expressed as:

[0153]

[0154] According to formula (29), the state vector is obtained and covariance The expression is:

[0155]

[0156] Step 9: After accumulating a certain number of residual samples, diagnose the sensor. Sensors diagnosed as faulty will be placed in an offline state and will not participate in subsequent target state estimation and information fusion. The expression of the fault detection variable of sensor i is:

[0157]

[0158] If the following equation is true, the sensor is offline:

[0159] L i > 2 (1-α,U×J×D i ) (33)

[0160] Where U is the accumulated time.

[0161] Step 10: Repeat steps 2 to 9 to obtain the state vector and covariance at different times, complete the tracking of the target, and obtain the track of the target.

Claims

1. A multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation, characterized by: The following steps are involved: Step 1: Use the sensor to measure the target in the detection area to obtain the measurement data of the target Get the state vector of the k-1 measured target at the last moment and covariance Wherein, j=1,2,...,J; J is the number of measured targets; i=1,2,...,I; I is the number of sensors; n=1,2,...,M; M is the number of measurement data; Step 2: Predict the state vector of target j at time k and covariance Step 3: Allocate the measurement data measured by the sensor to different measured objects and use them as verification measurement data of the corresponding measured objects; Step 4: Construct a correlation vector for the target being measured and the correlation vector for the measured data in, is the number of verified measurement data of the measured target j at sensor i at time k; according to the association vector and Get the correlation information from the measurement data to the measured object Correlation information from measurement data to the measured object Get data association information Step 5: Associate information based on data Calculate measurement update information Calculate the state vector of target j at sensor i at the current moment k and covariance Step 6: Obtain the fusion residual of the target And the covariance corresponding to the fusion residual According to the fusion residual of each sensor and covariance Perform fault detection on all sensors; combine the state vector of the normal sensor and covariance Get the state vector of target j at the current moment k and covariance Step 7: Repeat steps 2 to 6 to obtain the state vector and covariance of the target j at different times to obtain the track of the target.

2. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized in that: In step 6, the fault detection method is: obtaining the fault diagnosis test variable of sensor i at time k Fault diagnosis test variables The expression is: like The sensor i at time k is considered to be an abnormal sensor and does not participate in the target reliability calculation at time k; where α is the significance level; N k is the number of fusion residuals; D i is the dimension of the fused residual.

3. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized in that: In step 6, the sensor is diagnosed at regular intervals, and the diagnosis method is as follows: Get the fault diagnosis test variable L of sensor i during this period of time i , fault diagnosis test variable L i The expression is: Where U is the accumulated time; is the innovation covariance, and its expression is: in, is the Jacobian matrix of sensor i with respect to target j at time k; is the measurement error covariance of sensor i at time k; If L i >X 2 (1-α,U×J×D i ), the corresponding sensor is set to offline state and does not participate in subsequent target tracking; where α is the significance level; D i is the dimension of the fused residual.

4. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized in that: In the step 3, the method of allocating the measurement data is as follows: establishing a tracking gate according to the measured target The original measurement data is screened, and the measurement data falling into the overlapping area of ​​the tracking gates of different measured targets is assigned to the measured target closest to it; the measurement data falling into the tracking gate of measured target j is used as the verification measurement data of measured target j.

5. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized in that: In step 6, the fusion residual The corresponding covariance The expression is: in, To update the weights; is the innovation covariance; Update weights and innovation covariance The expressions are: in, is the Jacobian matrix of sensor i with respect to target j at time k; is the measurement error covariance of sensor i at time k.

6. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized by: In step 6, the state vector and covariance The expression is: Among them, I k is the number of normal sensors at time k.

7. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized by: In step 2, the predicted state vector and covariance The expressions are: Where F is the state transfer matrix; F T is the transpose of the state transfer matrix; Q k is the process noise error covariance.

8. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized by: In the step 1, the generalized measurement model of the sensor constructed in the clutter environment is: in, is the nth measurement data obtained by sensor i at time k; is the measurement function of sensor i; is Gaussian white noise; It is the false alarm measurement data.

9. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized by: In step 4, the correlation vector and The construction method is as follows: Assume that each measured object j generates at most one measurement data at sensor i, and each measurement data at sensor i is generated by at most one measured object; the association vector for: Correlation vector for:

10. The multi-sensor multi-target elastic fusion tracking method based on Gaussian belief propagation according to claim 1 is characterized by: In step 5, data association information The expression is: in, It is the association information from the measurement data to the measured object; The expression is: Among them, P g is the threshold probability; is the detection probability of sensor i; Correlation information from measurement data to the measured object The expression is: The expression is: Measurement data update information The expression is: in, is the verification measurement data set of the measured target j at sensor i; is the state vector of the measured target j at time k; g(·) is the global likelihood function; State Vector and covariance The expression is: Among them, (h i ) -1 is the inverse function of the measurement function of sensor i; is the measurement error covariance of sensor i at time k; is the innovation covariance, and its expression is: in, is the Jacobian matrix of sensor i with respect to target j at time k.