A deep-sea heterogeneous sonar fusion positioning method, system, device and storage medium

Abstract

The present invention discloses a deep-sea heterogeneous sonar fusion positioning method, system, device and storage medium, which relates to the technical field of hydroacoustics and hydroacoustic signal processing, in order to solve the problem of systematic defects caused by GMPHD filters. The deep-sea sonar fusion positioning method includes: collecting measurement information of each platform, using GMPHD filters to independently filter the measurement information of the platform to generate a local state estimate of the target; using a data association algorithm based on Mahalanobis distance as a criterion to associate the local state estimate of the target to obtain a correlation matrix between multiple local estimates; using a weighted convex combination fusion method to fuse the correlation matrix between the multiple local estimates to obtain a global state estimate of the target; using an adaptive feedback algorithm to analyze the global state estimate of the target at the previous moment to obtain a complete estimate of the global state of the target at the current moment. The present invention also provides a deep-sea heterogeneous sonar fusion positioning system, device and storage medium.

Description

Technical Field

[0001] The present invention relates to the technical field of hydroacoustics and hydroacoustic signal processing, and in particular to a heterogeneous sonar fusion positioning method, system, device and storage medium under deep-sea low detection probability conditions. Background Art

[0002] In recent years, multi-target localization has become an important research topic in the field of underwater target detection. The unknown association between measurement information and target true value has led to a lot of research on data association problems in multi-target localization algorithms.

[0003] Compared with the traditional association method, the probability hypothesis density (PHD) filter method is an emerging and practical method. This method regards the true value set of a single target as a set-valued state and the set of a single measurement value as a set-valued observation. Then, the set-valued state and the set-valued observation are used in the framework of Bayesian filtering. A method similar to the constant gain Kalman filter propagates the first-order moment of the single target state. The multi-target motion model is modeled in the presence of clutter and association uncertainty, and the association positioning of multiple targets is achieved after state estimation. Since the PHD filter equation has integral operations in the recursive process, it does not have an analytical solution in the general sense. Only in the case of linear Gaussian, the analytical expression of the equation can be obtained by using the properties of the Gaussian distribution, which derives the mixed Gaussian probability hypothesis density (GMPHD) filter.

[0004] Existing GMPHD filters often lose their estimate of targets in environments with high clutter density or low probability of detection; and the position prediction performance of GMPHD filters degrades when targets are close to / crossing each other; if there are missing measurements or measurement information is confused. GMPHD filters have these systematic defects because:

[0005] (1) This is mainly because the uncertainty of whether the measurement is from the target is not fully considered in the GMPHD filter algorithm. Specifically, the GMPHD filter generates new Gaussian components at each time step, which correspond to the measurement values ​​one by one. Regardless of whether the measurement value is from the target, its estimation error covariance is the same, so it can lose the estimate of the target when the clutter density is high or the detection probability is low;

[0006] (2) The GMPHD filter does not obey the one-to-one assumption, that is, it does not explicitly associate a measurement value with a target. From the filter algorithm structure, this filter implicitly considers the possible association between each measurement value and all targets. Therefore, in practical applications, the GMPHD filter is prone to violate the one-to-one assumption when targets are close to each other, such as when targets cross, causing the filter performance to deteriorate;

[0007] (3) Since the GMPHD filter does not provide a single target identity, the algorithm merges or removes estimated target locations through a merging / pruning process. Therefore, the target location being traced may be removed if the corresponding measurement is missing. Summary of the invention

[0008] The object of the present invention is to provide a multi-active heterogeneous sonar fusion positioning method under deep-sea low detection probability conditions, which is used to solve the systematic defects caused by the GMPHD filter.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] In a first aspect, the present invention provides a deep-sea heterogeneous sonar fusion positioning method, comprising:

[0011] Collecting measurement information from each platform, independently filtering the measurement information of the platform using a GMPHD filter to generate a local state estimate of the target;

[0012] Using a data association algorithm based on Mahalanobis distance as a criterion, the local state estimates of the target are associated to obtain an association matrix between multiple local state estimates;

[0013] Using a weighted convex combination fusion method to fuse the correlation matrices between the multiple local state estimates, to obtain a global state estimate of the target;

[0014] The adaptive feedback algorithm is used to analyze the global state estimation of the target at the previous moment to obtain a complete global state estimation of the target at the current moment.

[0015] Compared with the prior art, the multi-active sonar fusion positioning method under deep-sea low detection probability conditions provided by the present invention has the following beneficial effects:

[0016] 1. Aiming at the problem that the detection probability of each platform is reduced, resulting in inaccurate estimation of the target position and number by the GMPHD filter, an adaptive feedback filter is proposed with reference to the physical meaning of PHD and the GMPHD update and estimation process. The feedback filter independently calculates the weight value of each measurement in the feedback information based on the platform's measurement, thereby detecting the target's measurement and avoiding complex correlation calculations;

[0017] 2. Compared with the simple convex combination fusion method, the present invention adopts a weighted convex combination fusion strategy based on the GMPHD filter to merge the estimation results of the adaptive feedback algorithm with the estimation results of the GMPHD filter, thereby solving the performance loss problem caused by the inability of the GMPHD filter to accurately estimate the target position and number when the detection probability decreases, and enhancing the performance of the local filter.

[0018] 3. Compared with the existing multi-array fusion method, the present invention makes full use of the detection information of each active platform, and solves the problem of decreased performance of underwater acoustic target position estimation when the target is missed by fusing local estimates of multiple platforms with the help of GMPHD filter, data association algorithm, data fusion algorithm and adaptive feedback filtering algorithm. At the same time, it also has the ability to solve the problem of interference from a large number of false targets, and realizes robust fusion positioning of targets under multiple active platforms.

[0019] In a second aspect, the present invention further provides a deep-sea heterogeneous sonar fusion positioning system, comprising:

[0020] A filtering module, for collecting measurement information of each platform, and independently filtering the measurement information of the platform using a GMPHD filter to generate a local state estimate of the target;

[0021] A data association module, using a data association algorithm based on Mahalanobis distance as a criterion, associates the local state estimates of the target to obtain an association matrix between multiple local estimates;

[0022] A data fusion module, which fuses the correlation matrices between the multiple local estimates using a weighted convex combination fusion method to obtain a global state estimate of the target;

[0023] The adaptive feedback module uses an adaptive feedback algorithm to calculate the global state estimate of the target at the previous moment and obtain a complete estimate of the global state of the target at the current moment.

[0024] In a third aspect, the present invention also provides a deep-sea heterogeneous sonar fusion positioning device, comprising a processor and a communication interface coupled to the processor; the processor is used to run a computer program or instruction to implement a multi-active sonar fusion positioning method under deep-sea low detection probability conditions as described in any one of the first aspects.

[0025] In a fourth aspect, the present invention further provides a computer storage medium having instructions stored therein, which, when executed, implements a multi-active sonar fusion positioning method under deep-sea low detection probability conditions as described in any one of the first aspects.

[0026] Compared with the prior art, the beneficial effects of the second to fourth aspects of the present invention are the same as the beneficial effects of the multi-active sonar fusion positioning method under deep-sea low detection probability conditions described in the first aspect, and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0029] Figure 2 A schematic diagram of the main assumptions of the multi-target motion model according to an embodiment of the present invention;

[0030] Figure 3 A schematic diagram of the main assumptions of the multi-target observation model according to an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of the principle of the nearest neighbor method according to an embodiment of the present invention;

[0032] Figure 5 This is a flow chart of an adaptive feedback algorithm according to an embodiment of the present invention;

[0033] Figure 6 It is a target motion situation diagram in the experiment of the embodiment of the present invention;

[0034] Figure 7 This is the measurement graph of platform 1 when the detection probability is 0.7;

[0035] Figure 8 This is the measurement graph of platform 2 when the detection probability is 0.7;

[0036] Fig. 9 This is the measurement diagram of platform 3 when the detection probability is 0.7;

[0037] Fig.10 The GMPHD filtered results are measured for platform 1 when the detection probability is 0.7;

[0038] Fig.11 The GMPHD filtered results for platform 2 when the detection probability is 0.7;

[0039] Fig.12 The GMPHD filtered results are measured for platform 3 when the detection probability is 0.7;

[0040] Fig.13 The processing result of the adaptive feedback algorithm is measured for platform 1 when the detection probability is 0.7;

[0041] Fig.14 The processing results of the adaptive feedback algorithm are measured for platform 2 when the detection probability is 0.7;

[0042] Fig.15 The processing results of the adaptive feedback algorithm are measured for platform 3 when the detection probability is 0.7;

[0043] Fig.16 This is the fusion result of three platforms based on the adaptive feedback algorithm when the detection probability is 0.7;

[0044] Fig.17aThis is the performance evaluation diagram of OSPA without feedback when the detection probability is 0.7;

[0045] Fig.17b This is the performance evaluation diagram of OSPA with feedback when the detection probability is 0.7;

[0046] Fig.18 This is a comparison chart of OSPA mean values ​​under the same measurement when the detection probability is 0.7. DETAILED DESCRIPTION

[0047] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

[0048] The present invention provides a deep-sea heterogeneous sonar fusion positioning method, which mainly adopts multiple GMPHD filters to generate local position estimation and uses data association, data fusion and adaptive feedback performance of GMPHD filters; it solves the problem of inaccurate position estimation when the target is close, and the position estimation and positioning accuracy of the target is significantly improved compared with the traditional GMPHD filter, which makes up for the performance loss of the GMPHD filter when the detection probability decreases.

[0049] It should be noted that, in the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0050] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0051] Reference Figure 1 :

[0052] Step S1: Collect measurement information of each platform, and use the GMPHD filter to independently filter the measurement information of the platform to generate a local state estimate of the target.

[0053] The platform is a plurality of active arrays used to detect targets and obtain measurement information.

[0054] First, the GMPHD filter is introduced.

[0055] For a single-target Bayesian filter, assume that the target state is in the state space It follows a Markov transition process, and Time has come The transfer density at time Let the likelihood function Indicated in The platform is in the observation space For example, for a given state , the measurement received by the platform The likelihood is Then, for a given platform measurement , its probability density can be described as a posterior form (or called filter density), namely:

[0056] (1)

[0057] in, is the filter density, is the target state, is the measurement information of the platform, For the moment.

[0058] If the initial density of the state is known , the posterior probability density of the target state at the current moment can be iteratively calculated using the following formula.

[0059] (2)

[0060] (3)

[0061] Therefore, according to the RFS (random finite set) theory, it can be extended to the multi-objective Bayesian filtering algorithm, that is, the optimal multi-objective Bayesian filter is obtained:

[0062] (4)

[0063] (5)

[0064] in, for A reference indicator of express The likelihood of multiple targets at a time.

[0065] It can be seen that all relevant targets are The information of the moment state is contained in its posterior probability density Among them, we can use methods such as minimum mean square error or maximum a posteriori probability to find the target The state of the moment.

[0066] For a multi-target tracking filtering problem, assume that the number and state of the targets are random and time-varying, and let and Respectively expressed in Always monitor the number of measurements and targets of the platform in the area. Then, according to the RFS (random finite set) theory, The multi-target states and platform measurement data at each moment can be modeled as a random finite set, namely:

[0067] (6)

[0068] (7)

[0069] in, and Single target measurement and status The set of all finite subsets of .

[0070] Assume that the survival probability of the target state is , Expressed as from Time has come Status at all times Derived target set The probability of Expressed as from Time has come The target set of the new state in the region at the moment Then, according to the multi-target motion model, we know that Time has come In addition to the target movement transfer, the number of targets may also change in the following three ways: new targets appear (birth), new targets spawn from old targets, and old targets die (death). In other words, the movement state of multiple targets can only occur Figure 2 Four variations are shown.

[0071] Then, according to the multi-target motion model theory in RFS (random finite set), we have:

[0072] (8)

[0073] in, , and Respectively expressed as Time to Always keep alive goals, derivative goals and The finite set of new targets at each moment is:

[0074] (9)

[0075] Similarly, considering that the platform may suffer from missed detection or false target interference, assume that the detection probability of the platform is , then according to the multi-target measurement model theory in RFS (random finite set), we have:

[0076] (10)

[0077] in, For the The platform error or false target interference measurement set at the moment, is represented as a finite set of target measurements.

[0078] In other words, there are only Figure 3 The three hypothetical situations described in .

[0079] The PHD (Probability Hypothesis Density) filter adopts a method similar to the constant gain Kalman filter to propagate the first-order moment (mean) of a single target state. It propagates the posterior first-order moment of the multi-target state (posterior probability density intensity) instead of the posterior density of multiple targets, thereby reducing the difficulty of calculating the multi-target Bayesian filter.

[0080] For one in A random finite set on , the probability distribution is , whose first-order moment is a Non-negative function on , called intensity. Then any region ,right The interval integral is equivalent to exist The expected number of targets in the region is:

[0081] (11)

[0082] Assuming that the PHD filter algorithm meets the following conditions:

[0083] A1. 1. The appearance or update of each target state is independent of other targets;

[0084] A1. 2. The false targets are Poisson-like and the measurements of the targets are independent of each other;

[0085] A1.3. From The multi-target random finite set predicted in also obeys Poisson distribution;

[0086] Then, we can combine the multi-target motion model and measurement model in RFS theory to obtain another expression form of equation (4) and equation (5):

[0087] (12)

[0088] (13)

[0089] in, for A new set of goals that appear at every moment Strength; Reason Moment-derived target set Strength; for False target interference set Strength; For Target status at all times arrive The probability of being alive at any time; for The detection probability at time.

[0090] It can also be noted from the above two equations that the recursive formula of this method involves multiple complex integral calculations, which usually do not have analytical solutions and are computationally intensive. Therefore, reasonable approximate calculations are necessary.

[0091] In addition to the above assumptions A1.1-A1.3, for a linear Gaussian multi-target observation system, the GMPHD filter also needs to have the following assumptions:

[0092] A1.4. Each target follows a linear Gaussian dynamic model, and each platform follows a linear Gaussian measurement model; for example,

[0093] (14)

[0094] (15)

[0095] in, Represents a mean value of , the covariance is Gaussian density; is the Markov state transition matrix, is the covariance of the process noise, is the observation matrix, is the platform measurement noise covariance.

[0096] A1.5. The probability of the target surviving and being detected by the platform are independent, that is,

[0097] (16)

[0098] (17)

[0099] A1.6. The random finite set intensities of the new and derived targets are in the form of a Gaussian mixture; that is,

[0100] (18)

[0101] (19)

[0102] in, , , , and are parameters used to describe the intensity shape of the new target RFS. , , , , and Both are parameters used to describe the intensity shape of the derived target RFS.

[0103] For the linear Gaussian multi-objective model, the following two propositions not only explain how the Gaussian components of the posterior intensity are propagated in the GMPHD recursive process, but also give the closed solutions of the PHD (probability hypothesis density) recursive formulas (12) and (13).

[0104] Proposition 1: Assume that conditions A1. 4-A1. 6 are satisfied, and The posterior intensity at the moment is in the form of a Gaussian mixture, that is,

[0105] (20)

[0106] So, The posterior probability density intensity of the predicted target at the moment is also in the form of a Gaussian mixture:

[0107] (twenty one)

[0108] here, Same as the definition in equation (42). Then,

[0109] (twenty two)

[0110] (twenty three)

[0111] (twenty four)

[0112] (25)

[0113] (26)

[0114] (27)

[0115] Proposition 2: Assume that conditions A1. 4-A1. 6 are satisfied, and The prediction intensity at each moment is also in the form of a Gaussian mixture, i.e.

[0116] (28)

[0117] So, The posterior intensity at the moment is also in the form of a Gaussian mixture,

[0118] (29)

[0119] in,

[0120] (30)

[0121] (31)

[0122] (32)

[0123] (33)

[0124] (34)

[0125] (35)

[0126] Similarly, the following lemma can be applied to establish Proposition 1 and Proposition 2.

[0127] Lemma 1: Given , , , and ,and and If the positive definite condition is satisfied, then we have

[0128] (36)

[0129] Lemma 2: Given , , and ,and and If the positive definite condition is satisfied, then we have

[0130] (37)

[0131] here,

[0132] (38)

[0133] (39)

[0134] (40)

[0135] (41)

[0136] For a given Gaussian mixture intensity and , the corresponding target expectation number and It can be obtained from the sum of approximate intensities.

[0137] Under the premise of Proposition 1, the expectation of the number of predicted targets can be expressed as:

[0138] (42)

[0139] Under the premise of Proposition 2, the expectation of the updated target number can be expressed as:

[0140] (43)

[0141] Since the number of mixed Gaussian components will continue to increase with the iteration process, the amount of calculation of the GMPHD filter will continue to increase. Therefore, in order to ensure the calculation efficiency of GMPHD, it is necessary to reasonably prune and merge the mixed Gaussian components, that is, prune the components with smaller strength and merge the Gaussian components with similar distribution.

[0142] From formula (28), we can see that in order to express Current intensity of the moment , GMPHD filters usually require multiple Gaussian components as follows:

[0143] (44)

[0144] Therefore, the combined Gaussian distribution is:

[0145] (45)

[0146] (46)

[0147] in, , is the merging threshold. is the target state, is the covariance matrix.

[0148] The weight is used as the basis for judging the GMPHD detection target, and the detection threshold is set. (usually a value of 0.5) to determine the target to be extracted, that is, to extract all The target state is used as the local state estimate of the target output by GMPHD at the current moment.

[0149] Step S2: using a data association algorithm based on Mahalanobis distance as a criterion, the local state estimates of the target are associated to obtain an association matrix between multiple local state estimates.

[0150] Since the GMPHD filter has filtered out most of the false targets, the fusion center only needs to perform data association on a small number of local state estimates.

[0151] Mahalanobis distance is a method for calculating the similarity between unknown sample sets. This method not only considers the degree of association between states, but is also independent of the attribute type, that is, it is a dimensionless distance. In general, if the covariance matrix is ​​a unit matrix, then the Mahalanobis distance can be directly converted to the Euclidean distance; if the covariance matrix is ​​a diagonal matrix, then the Mahalanobis distance can also be called the normalized Euclidean distance.

[0152] Assume that the two states are represented as and ,use and Respectively represent the error covariance matrix corresponding to the two states, then the Mahalanobis distance between the two states can be calculated as:

[0153] (47)

[0154] According to the principle of the nearest neighbor association algorithm, if the Euclidean distance or Mahalanobis distance between two targets is less than or equal to the association threshold , and at the same time satisfy that its value is the smallest, then the two states can be considered to be related to each other and are the most related, otherwise they can be considered to be unrelated.

[0155] or (48)

[0156] Here is a brief introduction to the nearest neighbor association algorithm. Its basic meaning is: uniquely select the platform measurement that falls within the relevant wave gate and is closest to the target predicted position, and use this set of measurements and predicted positions as the association object. The meaning of "nearest" is not only the closest statistical distance, but also the largest error covariance probability density. Figure 4 The figure shows the principle of the nearest neighbor method. The measurement information of the two targets is determined based on the historical measurements of the active sonar (black solid circles), and the predicted positions of the two targets at the current moment are predicted based on the measurement information (black hollow circles). If the measurement error of the active sonar does not exceed a certain range (black dotted circle), then the nearest neighbor method can be used to determine which target the new measurement belongs to. Figure 4 If the new measurement is located at the four-pointed star (within the prediction gate of measurement 2), it is considered that the measurement is the measurement of target 2; if the new measurement is located at the five-pointed star (within the prediction gate of measurement 1), it is considered that the measurement is the measurement of target 1; if the new measurement is located at the six-pointed star (not in the prediction gate of the position of any target), it is considered that the measurement may be the measurement of a new target.

[0157] By using the target position information and error covariance matrix provided by the GMPHD filter and combining it with equation (47), we can realize the correlation between the local state estimates of different platforms and obtain the correlation matrix between multiple local state estimates of different platforms. represents the target state vector, Represents the corresponding error covariance matrix. At this moment, suppose platform 1 generates local state estimates, assuming that platform 2 generates local state estimates, then the association matrix can be expressed as:

[0158] (50)

[0159] (51)

[0160] like If both equation (48) and equation (51) are satisfied, then the first The local state estimate is related to the platform 2 The local state estimates are the most relevant, and multi-platform data fusion is performed.

[0161] (52)

[0162] in, To obtain the minimum function, if (or ) do not satisfy equation (50), then it is considered that the first local state estimate (or the first local state estimates) are unconnected to any local state estimates on other platforms.

[0163] Step S3: Use the weighted convex combination fusion method to fuse the correlation matrices between multiple local state estimates to obtain the global state estimate of the target.

[0164] Since the weighted convex combination fusion method borrows the geometric interpretation from the standard Kalman fusion method, it usually has two assumption constraints:

[0165] A1.7. The observation processes between platforms are independent of each other, and the covariances of the local state estimation errors generated are also independent of each other;

[0166] A1. 8. The observation of the target by each platform meets the consistency requirement.

[0167] Assume Platform and Platform The local state estimation and error covariance matrix for the same target are and ,in Then, we have

[0168] (53)

[0169] in, and are the state estimation errors of the two platforms respectively, and the two are independent of each other.

[0170] According to the prior state expectation and measurement The resulting posterior state expectation estimate for:

[0171] (54)

[0172] make and Respectively represent the platform and Platform Prior information, then

[0173] (55)

[0174] (56)

[0175] Among them, the error The mean of is 0 and are independent of each other, and the covariance matrix is , then the terms in formula (54) can be expressed as:

[0176] (57)

[0177] The corresponding covariance matrix is: (58)

[0178] Then, formula (54) can be specifically expressed as:

[0179] (59)

[0180] When the local state estimation errors are uncorrelated, we have:

[0181] (60)

[0182] Then, the fusion result can be simplified as:

[0183] (61)

[0184] When the number of platforms When, and all the estimation errors When they are uncorrelated with each other, then equation (61) can be expanded to:

[0185] (62)

[0186] It can be seen that when the algorithm can meet the assumptions, the optimal fusion estimate can be obtained by using equation (61) or equation (62) based on the correlation between the local state estimates.

[0187] Although the convex combination fusion method has better performance, it does not consider the problems of multi-target, platform missed detection and false target interference. Therefore, the convex combination fusion method mainly adopts a weighted fusion strategy, including the following:

[0188] ① If the local state estimates of multiple platforms can be correlated with each other and the correlation is minimal, the convex combination fusion method is used to fuse the local state estimates of multiple platforms to obtain the global state estimate;

[0189] ② If the local state estimates of multiple platforms cannot be associated, all estimation results are merged into the global state estimate at the current moment.

[0190] The essential idea of ​​the convex combination fusion method is to construct a weight relationship for each platform, and then use the size of the weight to determine the influence of the local state estimation of the platform in the fusion algorithm.

[0191] Step S4: Analyze the global state estimation of the target at the previous moment using an adaptive feedback algorithm to obtain a complete global state estimation of the target at the current moment.

[0192] Among them, the adaptive feedback algorithm constructs a feedback filter according to the detection rule of the GMPHD filter, and makes the feedback filter and the GMPHD filter run independently. Its structure is as follows: Figure 5 shown.

[0193] Assuming that the target states in the global state estimation are independent of each other and the error covariance follows a Gaussian distribution, then the complete estimate of the global state of the target at the current moment can be achieved by the following steps:

[0194] Step S41: Analyze the global state estimation at the previous moment according to the Markov transfer matrix method to obtain the global state prediction data at the current moment.

[0195] set up The global state estimation state of the fusion center at the moment is expressed as , then according to the RFS (random finite set) theory, the entire global state estimate can be regarded as a random finite set ,Right now:

[0196] (63)

[0197] in, for The estimated number of global states at a given moment.

[0198] Then, according to the Markov state transfer matrix , we can get Global prediction status at time and the corresponding error covariance matrix .

[0199] (64)

[0200] (65)

[0201] in, , is the process noise covariance matrix.

[0202] Since the global state estimation state is a random finite set, its global state prediction data is also a random finite set, that is:

[0203] (66)

[0204] Step S42: Feedback the global state prediction data at the current moment to the feedback filter, and calculate the weight value of each platform measurement information through the feedback filter.

[0205] Referring to the physical meaning of PHD (probability hypothesis density) and the updating and estimation process of GMPHD filter, the process of calculating weight value through feedback filter is as follows: The feedback filter calculates the joint probability density function of each measurement in the platform according to the feedback information. In expectation of , the variance is The weight value of the probability density intensity corresponding to the probability density function .

[0206] (67)

[0207] in, is the probability density function.

[0208] Step S43: When the weight value is greater than the detection threshold, the global state prediction data at the current moment is estimated and updated using the Kalman filter theory to obtain the local state estimation that is missing from the GMPHD filter at the current moment.

[0209] The Kalman filter theory is the Kalman gain, which can be obtained by the error covariance matrix of the global prediction state It is concluded that:

[0210] (68)

[0211] in, is the Kalman gain; Representative feedback; for the moment; Label the estimated points; is the error covariance; To use Always The predicted value at the moment; is the Markov state transfer matrix; is the observation noise.

[0212] That is, the weight is greater than the detection threshold ,Right now , then we can determine the measurement The predicted state corresponding to the largest weight value They are interrelated, and the Kalman filter theory is used to estimate and update the global state prediction data at the current moment.

[0213] (69)

[0214] (70)

[0215] in, is the estimated value of position; is the index of the number of Gaussian components in the current feedback algorithm; is the real measurement information corresponding to the estimated value; The index of platform measurement; is the unit diagonal matrix.

[0216] Formula (69) is the missing local state estimate of the GMPHD filter at the current moment, and formula (70) is the updated error covariance. Here, if we assume that the platform Moment Share A measurement, then and They are the index of the platform measurement and the index of the number of Gaussian components in the current feedback algorithm, respectively.

[0217] Step S44: Associating the local state estimate with the missing local state estimate of the GMPHD filter at the current moment to obtain an updated association matrix, and fusing multiple updated association matrices using a weighted fusion method to obtain a complete estimate of the global state of the target at the current moment.

[0218] Among them, equations (47) and (48) are used for data association. The difference between the weighted fusion method here and the convex combination fusion method in step S3 is that the fusion strategy of the weighted fusion method is: only when the feedback filter generates a target state estimate and the GMPHD filter does not have a corresponding target state estimate, multiple updated association matrices are fused; in other cases, the estimation result of the GMPHD filter is used as the main one, and no data fusion is performed. Its purpose is to minimize the bias effect of the adaptive feedback algorithm on the local state estimation.

[0219] In summary, the local state estimation supplemented by the adaptive feedback filtering algorithm makes up for the performance loss of the GMPHD filter when the detection probability decreases, solves the problem of target estimation quantity and position deviation, and minimizes the bias influence problem brought by the adaptive feedback filter, thereby enhancing the detection capability of the local filter and the entire detection system.

[0220] This method is verified by experiments, and the simulation parameters are as follows:

[0221] The target is detected for 50 minutes at intervals of 30 seconds, and the 100 frames of data obtained by simulation are analyzed and verified. The active platform is set to be interfered by at most 100 random false targets, and the measurement error is 25m. The starting positions of the three targets are (-500, -2500), (2000, -1500), and the position of target 1 after 15 minutes of movement; except for target 3 which starts to move at the 15th minute, targets 1 and 2 move continuously during the detection time; the motion states of the three targets are linear motion at a speed of 5m / s along the 45° direction, linear motion at a speed of 3m / s along the 0° direction, and linear motion at a speed of 2m / s along the 135° direction. The parameter table is as follows:

[0222]

[0223] Figure 6 is the target's motion situation. When the detection probability of each platform is 0.7, the target situation simulation is as follows: Figure 7~Figure 9 ; Using the traditional GMPHD filter-based position estimation method, the results are as follows Figure 10~Figure 12 As shown; the positioning results based on the adaptive feedback algorithm are fused as shown Figure 13~Figure 16 shown.

[0224] Here, we use the optimal sub-pattern assignment (OSPA) distance to evaluate the performance of the positioning estimation results under each detection probability. The results are as follows: Fig.17a and Fig.17b As shown in Figure 2, the OSPA mean values ​​under the same measurement are compared. Fig.18 As shown in Figure 2, it can be seen that the adaptive feedback algorithm can effectively improve the detection ability of the local filter for targets with low weight values. The local estimation effect after the adaptive feedback algorithm is better and the fusion result is more accurate.

[0225] The embodiment of the present invention also discloses a deep-sea heterogeneous sonar fusion positioning system, comprising:

[0226] A filtering module, for collecting measurement information of each platform, and independently filtering the measurement information of the platform using a GMPHD filter to generate a local state estimate of the target;

[0227] A data association module, using a data association algorithm based on Mahalanobis distance as a criterion, associates the local state estimates of the target to obtain an association matrix between multiple local estimates;

[0228] A data fusion module, which fuses the correlation matrices between the multiple local estimates using a weighted convex combination fusion method to obtain a global state estimate of the target;

[0229] The adaptive feedback module uses an adaptive feedback algorithm to calculate the global state estimate of the target at the previous moment and obtain a complete estimate of the global state of the target at the current moment.

[0230] Among them, the adaptive feedback module includes:

[0231] A prediction unit, used to analyze the global state estimation at the previous moment according to the Markov transfer matrix method to obtain the global state prediction data at the current moment;

[0232] The feedback unit feeds back the global state prediction data at the current moment to the feedback filter, and calculates the weight value of each platform measurement information through the feedback filter;

[0233] The updating unit, when the weight value is greater than the detection threshold, estimates and updates the global state prediction data at the current moment using the Kalman filter theory to obtain the local state estimation missing at the current moment of the GMPHD filter;

[0234] The merging unit associates the local state estimate with the missing local state estimate of the GMPHD filter at the current moment to obtain an updated association matrix, and fuses multiple updated association matrices using a weighted fusion method to obtain a complete estimate of the global state of the target at the current moment.

[0235] Furthermore, the fusion strategy of the weighted convex combination fusion method includes:

[0236] If the local state estimates of multiple platforms can be correlated with each other and the correlation is minimal, the convex combination fusion method is used to fuse the local state estimates of multiple platforms to obtain the global state estimate;

[0237] If the local state estimates of multiple platforms cannot be associated, all estimation results are merged into the global state estimate at the current moment.

[0238] The fusion strategies of the weighted fusion method include:

[0239] The multiple updated correlation matrices are fused only when the feedback filter produces a target state estimate and the GMPHD filter does not have a corresponding target state estimate.

[0240] An embodiment of the present invention also discloses a deep-sea heterogeneous sonar fusion positioning device, including a processor and a communication interface coupled to the processor; the processor is used to run a computer program or instruction to implement a multi-active sonar fusion positioning method under deep-sea low detection probability conditions as described above.

[0241] The embodiment of the present invention further discloses a computer storage medium, in which instructions are stored. When the instructions are executed, the aforementioned multi-active sonar fusion positioning method under deep-sea low detection probability conditions is implemented.

[0242] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instruction is loaded and executed on a computer, the process or function described in the embodiment of the present invention is executed in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, a terminal, a user device or other programmable device. The computer program or instruction may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer program or instruction may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired or wireless means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, a hard disk, or a tape; it may also be an optical medium, such as a digital video disc (DVD); it may also be a semiconductor medium, such as a solid state drive (SSD).

[0243] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in a claim. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0244] Although the present invention has been described in conjunction with specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present invention. Accordingly, this specification and the accompanying drawings are merely exemplary illustrations of the present invention as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present invention. Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, the present invention is intended to include such modifications and variations if they fall within the scope of the claims of the present invention and their equivalents.

Claims

1. A deep-sea heterogeneous sonar fusion positioning method, characterized in that: include: Collecting measurement information from each platform, independently filtering the measurement information of the platform using a GMPHD filter to generate a local state estimate of the target; Using a data association algorithm based on Mahalanobis distance as a criterion, the local state estimates of the target are associated to obtain an association matrix between multiple local state estimates; Using a weighted convex combination fusion method to fuse the correlation matrices between the multiple local state estimates, to obtain a global state estimate of the target; The global state estimation of the target at the previous moment is analyzed using the adaptive feedback algorithm to obtain a complete global state estimation of the target at the current moment, including: The global state estimation at the previous moment is analyzed according to the Markov transfer matrix method to obtain the global state prediction data at the current moment; Feedback the global state prediction data at the current moment to the feedback filter, and calculate the weight value of each platform measurement information through the feedback filter; When the weight value is greater than the detection threshold, the Kalman filter theory is used to estimate and update the global state prediction data at the current moment, and the local state estimation missing at the current moment of the GMPHD filter is obtained; The local state estimate is associated with the missing local state estimate of the GMPHD filter at the current moment to obtain an updated association matrix, and multiple updated association matrices are fused using a weighted fusion method to obtain a complete estimate of the global state of the target at the current moment.

2. A deep-sea heterogeneous sonar fusion positioning method according to claim 1, characterized in that: The fusion strategy of the weighted convex combination fusion method includes: If the local state estimates of multiple platforms can be correlated with each other and the correlation is minimal, the convex combination fusion method is used to fuse the local state estimates of multiple platforms to obtain the global state estimate; If the local state estimates of multiple platforms cannot be associated, all estimation results are merged into the global state estimate at the current moment.

3. A deep-sea heterogeneous sonar fusion positioning method according to claim 1, characterized in that: The fusion strategy of the weighted fusion method includes: The multiple updated correlation matrices are fused only when the feedback filter produces a target state estimate and the GMPHD filter has no corresponding target state estimate.

4. A deep-sea heterogeneous sonar fusion positioning system, characterized in that: include: A filtering module, for collecting measurement information of each platform, and independently filtering the measurement information of the platform using a GMPHD filter to generate a local state estimate of the target; A data association module, using a data association algorithm based on Mahalanobis distance as a criterion, associates the local state estimates of the target to obtain an association matrix between multiple local state estimates; A data fusion module, which uses a weighted convex combination fusion method to fuse the correlation matrices between the multiple local state estimates to obtain a global state estimate of the target; The adaptive feedback module uses an adaptive feedback algorithm to calculate the global state estimate of the target at the previous moment and obtain a complete estimate of the global state of the target at the current moment.

5. A deep-sea heterogeneous sonar fusion positioning system according to claim 4, characterized in that: The fusion strategy of the weighted convex combination fusion method includes: If the local state estimates of multiple platforms can be correlated with each other and the correlation is minimal, the convex combination fusion method is used to fuse the local state estimates of multiple platforms to obtain the global state estimate; If the local state estimates of multiple platforms cannot be associated, all estimation results are merged into the global state estimate at the current moment.

6. A deep-sea heterogeneous sonar fusion positioning system according to claim 4, characterized in that: The adaptive feedback module comprises: A prediction unit, used to analyze the global state estimation at the previous moment according to the Markov transfer matrix method to obtain the global state prediction data at the current moment; The feedback unit feeds back the global state prediction data at the current moment to the feedback filter, and calculates the weight value of each platform measurement information through the feedback filter; The updating unit, when the weight value is greater than the detection threshold, estimates and updates the global state prediction data at the current moment using the Kalman filter theory to obtain the local state estimation missing at the current moment of the GMPHD filter; The merging unit associates the local state estimate with the missing local state estimate of the GMPHD filter at the current moment to obtain an updated association matrix, and fuses multiple updated association matrices using a weighted fusion method to obtain a complete estimate of the global state of the target at the current moment.

7. A deep-sea heterogeneous sonar fusion positioning system according to claim 6, characterized in that: The fusion strategy of the weighted fusion method includes: The multiple updated correlation matrices are fused only when the feedback filter produces a target state estimate and the GMPHD filter does not have a corresponding target state estimate.

8. A deep-sea heterogeneous sonar fusion positioning device, characterized in that: It comprises a processor and a communication interface coupled to the processor; the processor is used to run a computer program or instruction to implement a deep-sea heterogeneous sonar fusion positioning method as described in any one of claims 1-3.

9. A computer storage medium, characterized in that The computer storage medium stores instructions, and when the instructions are executed, the deep-sea heterogeneous sonar fusion positioning method described in any one of claims 1 to 3 is implemented.

Images