A multi-target tracking method, system, electronic device and storage medium

Abstract

The application discloses a multi-target tracking method and system, electronic equipment and a storage medium, and relates to the technical field of target tracking. The method comprises the following steps: determining each posterior target Gaussian component in which the weight of a local posterior multi-target intensity function of each sensor constructed satisfies a preset weight as a first screening Gaussian component; determining a plurality of target estimated state sets according to all first screening Gaussian components; determining a corresponding core Gaussian component based on all neighborhood sets corresponding to each target estimated state set, thereby determining a normal sensor identity tag set; constructing an initial fusion intensity function of each sensor based on the normal sensor identity tag set and each local posterior multi-target intensity function; and determining a final fusion intensity function based on a consistency fusion weight matrix and the initial fusion intensity function of each normal sensor, thereby determining the state of a plurality of tracked targets. The application reduces the error of multi-target tracking under the condition that the accuracy of part of sensors decreases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target tracking, in particular to a multi-target tracking method and system, an electronic device and a storage medium. BACKGROUND

[0002] The emergence and development of wireless sensor technology promote the development of distributed multi-target tracking method, and now the distributed multi-target tracking technology has been widely used in military and civil fields, such as target monitoring, air defense and precision guidance. At present, the commonly used data fusion methods for realizing distributed multi-target tracking mainly include consistency fusion, gossip fusion, time sequence fusion and diffusion fusion. Among them, the distributed multi-target tracking method based on consistency fusion has good expansibility and robustness, and is more applied in theoretical research and practical application.

[0003] Compared with single target tracking, the correspondence between observation and target in distributed multi-target tracking is uncertain, while the data association algorithm can correspond the observation and the target, but the multi-target tracking algorithm based on data association needs complex data association calculation, which increases the computational complexity of target tracking, and also causes the situation of "combination explosion". The emergence and development of random finite set theory provide a new implementation way for distributed multi-target tracking. For the distributed multi-target tracking technology based on random finite set, there are two keys: local filtering algorithm and distributed fusion algorithm. The development of local filtering algorithm based on random finite set is relatively mature, and according to different random finite set assumptions, various approximate filters of optimal multi-target Bayesian filter can be obtained, such as probability hypothesis density filter, multi-target multi-Bernoulli filter and generalized labeled multi-Bernoulli filter. At present, consistency fusion is also called average consistency, which has two forms: arithmetic average and geometric average of initial local value. However, the current multi-target tracking method focuses on the derivation of the fusion multi-target intensity formula of the posterior multi-target intensity of the local filter obtained by using different random finite set assumptions and various improved fusion methods to improve the fusion effect or reduce the computational burden, and does not consider the communication topology change caused by external interference, the fusion precision reduction caused by the precision reduction of part of sensors, and other situations, which further leads to the low tracking precision of the existing multi-target tracking method. SUMMARY

[0004] The purpose of the present application is to provide a multi-target tracking method, system, electronic device and storage medium, which reduces the error of multi-target tracking under the condition of precision reduction of part of sensors.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A multi-target tracking method applied to a wireless sensor network comprising a plurality of sensors and a plurality of tracked targets, the method comprising:

[0007] constructing, by means of an extended Gaussian mixture probability hypothesis density algorithm, a local posterior multi-target intensity function of each sensor at each sampling time;

[0008] determining, in real time, a state of the tracked targets based on the local posterior multi-target intensity function of each sensor at each sampling time;

[0009] wherein the determination process of the state of the plurality of tracked targets at any sampling time comprises:

[0010] determining each posterior target Gaussian component in the local posterior multi-target intensity function of each sensor as a first screened Gaussian component, where the weight of the each posterior target Gaussian component satisfies a preset weight;

[0011] determining a plurality of target estimated state sets according to all the first screened Gaussian components; the target estimated state sets comprise a plurality of second screened Gaussian components;

[0012] calculating a neighborhood set of each second screened Gaussian component;

[0013] determining a core Gaussian component corresponding to each target estimated state set based on all the neighborhood sets corresponding to the each target estimated state set;

[0014] determining a normal sensor identity tag set according to each core Gaussian component;

[0015] constructing an initial fusion intensity function of each sensor based on the normal sensor identity tag set and the local posterior multi-target intensity function of each sensor;

[0016] determining a final fusion intensity function of each sensor based on a consensus fusion weight matrix, the initial fusion intensity function of each sensor and a consensus fusion iteration number; the consensus fusion weight matrix is determined based on a communication topology in the wireless sensor network;

[0017] determining the state of the plurality of tracked targets according to the final fusion intensity function of each sensor.

[0018] A multi-target tracking system comprising:

[0019] a local posterior multi-target intensity function construction module configured to construct, by means of an extended Gaussian mixture probability hypothesis density algorithm, a local posterior multi-target intensity function of each sensor at each sampling time;

[0020] a tracking module configured to determine, in real time, a state of the tracked targets based on the local posterior multi-target intensity function of each sensor at each sampling time;

[0021] The determination process of the state of each tracked target at any sampling time comprises:

[0022] Each posterior target Gaussian component in the local posterior multi-target intensity function of each sensor is determined as a first screened Gaussian component, wherein the weight of each posterior target Gaussian component satisfies a preset weight.

[0023] A plurality of target estimated state sets are determined according to all the first screened Gaussian components; the target estimated state sets comprise a plurality of second screened Gaussian components.

[0024] A neighborhood set of each second screened Gaussian component is calculated.

[0025] A core Gaussian component corresponding to each target estimated state set is determined based on all the neighborhood sets corresponding to the target estimated state set.

[0026] A normal sensor identity tag set is determined according to the core Gaussian components.

[0027] An initial fusion intensity function of each sensor is constructed based on the normal sensor identity tag set and the local posterior multi-target intensity function of each sensor.

[0028] A final fusion intensity function of each sensor is determined based on a consensus fusion weight matrix, the initial fusion intensity function of each sensor and a consensus fusion iteration number; the consensus fusion weight matrix is determined based on a communication topology in the wireless sensor network.

[0029] The state of each tracked target is determined according to the final fusion intensity function of each sensor.

[0030] An electronic device comprises:

[0031] One or more processors;

[0032] A storage device having one or more programs stored thereon;

[0033] When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-target tracking method as described above.

[0034] A storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the multi-target tracking method as described above.

[0035] According to the embodiments of the present application, the following technical effects are provided:

[0036] The application discloses a multi-target tracking method and system, an electronic device and a storage medium, and relates to the technical field of multi-target tracking. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0038] Figure 1 A multi-target tracking method flowchart is provided for the embodiment 1 of the present application.

[0039] Figure 2 A distributed multi-target fault-tolerant tracking technology design flowchart is provided.

[0040] Figure 3 A two-dimensional plane diagram of a tracking scene during simulation is provided.

[0041] Figure 4 An estimated target number diagram of the DMTT algorithm based on DBSCAN, the AA-DGMPHD and the GA-DGMPHD is provided.

[0042] Figure 5 An OSPA diagram of the DMTT algorithm based on DBSCAN, the AA-DGMPHD and the GA-DGMPHD is provided. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0044] The application aims to provide a multi-target tracking method, system, electronic device and storage medium, which aims to reduce the error of multi-target tracking under the condition of part sensor accuracy decline.

[0045] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below in combination with the drawings and specific embodiments.

[0046] Embodiment 1

[0047] Figure 1 The multi-target tracking method flowchart provided for Embodiment 1 of the application is shown in the figure. Figure 1 As shown in the figure, the multi-target tracking method in this embodiment is applied to a wireless sensor network including multiple sensors and multiple tracked targets, and the method includes:

[0048] Step 1: using the extended Gaussian mixture probability hypothesis density algorithm, constructing a local posterior multi-target intensity function of each sensor at each sampling time.

[0049] As an optional implementation, using the extended Gaussian mixture probability hypothesis density algorithm, constructing a local posterior multi-target intensity function of the jth sensor at the tth sampling time, including:

[0050] Obtaining the state vector of multiple newborn target Gaussian components of the jth sensor at the tth sampling time, the covariance matrix of multiple newborn target Gaussian components at the tth sampling time, and the state vector and covariance matrix of multiple posterior target Gaussian components in the local posterior multi-target intensity function at the t-1th sampling time.

[0051] According to the state vector and covariance matrix of all the posterior target Gaussian components at the t-1th sampling time, constructing a survival multi-target intensity function at the tth sampling time. Specifically, it is realized by using formula (14).

[0052] According to the state vector of all the newborn target Gaussian components at the tth sampling time and the covariance matrix of all the newborn target Gaussian components, constructing a newborn multi-target intensity function at the tth sampling time. Specifically, it is realized by using formula (15).

[0053] According to the survival multi-target intensity function at the t th sampling moment and the new multi-target intensity function at the t th sampling moment, a local prior multi-target intensity function at the t th sampling moment is constructed; the local prior multi-target intensity function at the t th sampling moment includes a plurality of state vectors of prior target Gaussian components at the t th sampling moment and a plurality of covariance matrices of prior target Gaussian components at the t th sampling moment, and the prior target Gaussian components at the t th sampling moment include all survival target Gaussian components at the t th sampling moment and all new target Gaussian components at the t th sampling moment. Specifically, formula (13) is used to implement.

[0054] According to the state vector and the covariance matrix of the prior target Gaussian component at the t th sampling moment and the observation vector at the t th sampling moment, the state vector and the covariance matrix of the posterior target Gaussian component at the t th sampling moment are determined by using the Kalman filtering method. Specifically, formula (26) and formula (27) are used to implement.

[0055] According to the local prior multi-target intensity function at the t th sampling moment and the state vector and the covariance matrix of each posterior target Gaussian component at the t th sampling moment, a local posterior multi-target intensity function at the t th sampling moment is constructed. Specifically, formula (19) is used to implement.

[0056] Step 2: Real-time determination of the state of the tracked target based on the local posterior multi-target intensity function of each sensor at each sampling moment.

[0057] Among them, the determination process of the state of the plurality of tracked targets at any sampling moment includes:

[0058] Step 21: Each posterior target Gaussian component in the local posterior multi-target intensity function of each sensor whose weight satisfies the preset weight is determined as a first screening Gaussian component. Specifically, formula (30) is used to implement.

[0059] Step 22: According to all first screening Gaussian components, a plurality of target estimated state sets are determined. The target estimated state set includes a plurality of second screening Gaussian components. Formula (34) is used to implement.

[0060] Step 23: The neighborhood set of each second screening Gaussian component is calculated.

[0061] Step 24: Based on all neighborhood sets corresponding to each target estimated state set, a core Gaussian component corresponding to each target estimated state set is determined. Specifically, formula (34) is used to implement.

[0062] Step 25: A normal sensor identity tag set is determined according to each core Gaussian component. Specifically, formula (35) to formula (38) are used to implement.

[0063] Step 26: Constructing the initial fusion intensity function of each sensor based on the normal sensor identity tag set and the local posterior multi-target intensity function. Specifically, the formula (39) - formula (42) are used to achieve.

[0064] Step 27: Determining the final fusion intensity function of each sensor based on the consistency fusion weight matrix, the initial fusion intensity function of each sensor and the consistency fusion iteration number; the consistency fusion weight matrix is determined based on the communication topology in the wireless sensor network. The formula of the element in the consistency fusion weight matrix is shown in formula (9). The formula of the intermediate process of determining the final fusion intensity function is shown in formula (43).

[0065] Step 28: Determining the state of the multiple tracked targets according to the final fusion intensity function of each sensor.

[0066] As an optional implementation, constructing the survival multi-target intensity function at the tth sampling moment according to the state vector and the covariance matrix of all posterior target Gaussian components at the (t-1) th sampling moment, specifically including:

[0067] Determining the state vector of the survival target Gaussian component at the tth sampling moment corresponding to each posterior target Gaussian component at the (t-1) th sampling moment according to the state vector. Specifically, the formula (16) is used to achieve.

[0068] Determining the covariance matrix of the survival target Gaussian component at the tth sampling moment corresponding to each posterior target Gaussian component at the (t-1) th sampling moment according to the covariance matrix. Specifically, the formula (17) is used to achieve.

[0069] Constructing the survival multi-target intensity function at the tth sampling moment according to the state vector of all survival target Gaussian components at the tth sampling moment and the covariance matrix of all survival target Gaussian components. Specifically, the formula (14) is used to achieve.

[0070] As an optional implementation, using Kalman filtering method, determining the state vector and the covariance matrix of the posterior target Gaussian component at the tth sampling moment according to the state vector and the covariance matrix of the prior target Gaussian component at the tth sampling moment and the observation vector at the tth sampling moment, specifically including:

[0071] Determining the Kalman gain matrix at the tth sampling moment corresponding to each prior target Gaussian component at the tth sampling moment according to the covariance matrix. Specifically, the formula (28) is used to achieve.

[0072] According to the Kalman gain matrix of each sampling time, the state vector of the corresponding prior target Gaussian component at the tth sampling time, and the observation vector at the tth sampling time, the state vector of the corresponding posterior target Gaussian component at the tth sampling time is determined. Specifically, formula (26) is used to realize.

[0073] According to the Kalman gain matrix of each sampling time, the state vector of the corresponding prior target Gaussian component at the tth sampling time, and the observation vector at the tth sampling time, the state vector of the corresponding posterior target Gaussian component at the tth sampling time is determined. Specifically, formula (26) is used to realize.

[0074] As an optional implementation, the determination process of any current target estimation state set comprises:

[0075] The distance between each first screening Gaussian component of each non-current sensor in the wireless sensor network and the current first screening Gaussian component is calculated respectively, and is recorded as a first distance. The non-current sensor is a sensor other than the current sensor in all sensors. Specifically, formula (31) is used to realize.

[0076] The first screening Gaussian component corresponding to all non-current sensors of the first number of current first screening Gaussian components satisfying the preset number is determined as a second screening Gaussian component, and the set composed of the second screening Gaussian component is determined as the current target estimation state set. The first number is the number of target non-current sensors whose first distance satisfies the preset distance, and the target non-current sensor is a non-current sensor whose first distance satisfies the preset distance. Specifically, formula (32) is used to realize.

[0077] As an optional implementation, it specifically comprises: the determination process of the neighborhood set of any current second screening Gaussian component comprises:

[0078] The distance between the current second screening Gaussian component and each non-current second screening Gaussian component is calculated respectively, and is recorded as a second distance. The non-current second screening Gaussian component is a second screening Gaussian component other than the current second screening Gaussian component in all second screening Gaussian components of the target estimation state set in which the current second screening Gaussian component is located.

[0079] All non-current second screening Gaussian components whose second distance is less than the neighborhood radius are determined as the neighborhood set of the current second screening Gaussian component. Specifically, formula (33) is used to realize.

[0080] As an optional implementation, the construction process of any initial fusion strength function specifically comprises:

[0081] The correction matrix is established according to the normal sensor identity tag set. Specifically, formula (36) is used to realize.

[0082] The consistency fusion weight matrix is corrected based on the correction matrix to obtain a corrected consistency fusion weight matrix. Specifically, formula (40) is used to implement the correction.

[0083] The corrected consistency fusion weight matrix is normalized to obtain an initialization weight matrix. Specifically, formula (41) is used to implement the normalization.

[0084] An initial fusion intensity function is constructed according to the initialization weight matrix and the local posterior multi-target intensity function of each sensor. Specifically, formula (26) is used to implement the construction.

[0085] In fact, before the multi-target tracking method provided in Embodiment 1 is implemented, the models and formulas involved in Embodiment 1 need to be designed based on the model construction idea, as shown in the following formula (1) to (7). Figure 2 As shown in the following formula (1) to (7), the following design process is used to obtain a distributed multi-target fault-tolerant tracking technology that can implement Embodiment 1:

[0086] Step 1: Establish a multi-target tracking system model.

[0087] Consider a nonlinear discrete-time multi-target tracking system, where the motion equation of the multi-target is:

[0088] The observation equation of the multi-sensor is:

[0089]

[0090] wherein, xi(t) represents the state vector of the i-th target (target refers to any entity that needs to be tracked) at the t-th sampling time; Rn represents a real state space, and m1 represents the dimension of the state vector; fi(t) represents the state transition function of the i-th target from the (t-1)-th sampling time to the t-th sampling time; xi(t-1) represents the state vector of the i-th target at the (t-1)-th sampling time; yj(t) represents the observation vector of the i-th target by the j-th sensor at the t-th sampling time; j represents the identity tag of the j-th sensor; Rm represents a real state space, and m2 represents the dimension of the observation vector; hj(t) represents the observation function of the j-th sensor at the t-th sampling time; t and N s represent the number of targets and the number of sensors, respectively; xi(t-1) represents the uncertainty of the i-th target motion at the (t-1)-th sampling time, which is subject to a Gaussian distribution with a mean of 0 and a covariance of ; Let the noise observed by the j-th sensor at the t-th sampling time be represented by a mean of 0 and a covariance of 0. Gaussian distribution, Let represent a Gaussian distribution. Where the covariance is... and They are represented as follows:

[0091]

[0092] E[·] represents covariance; Let covariance represent the uncertainty of the motion of the i-th target at sampling time t-1; Let T represent the covariance of the noise observed by the j-th sensor at the t-th sampling time; T represents the transpose.

[0093] In the multi-objective state estimation problem, it is assumed that the state of the objective follows a state space. The linear Gaussian transition density is given by formula (4). The Markov process is represented. This represents the state of the i-th target at the t-th sampling time. The linear Gaussian transition density of state x at sampling time t-1, where, Let be the state transition matrix of the i-th target at the (t-1)-th sampling time. It can be derived from the multi-target motion equation, i.e., formula (1), in the state transition matrix. Local linearization is performed at x' to obtain x′ as a function The variable is the state x at the (t-1)th sampling time. t-1 At the t-th sampling time, it transitions to state x. t The probability is This state transition Markov process can be observed in the observation space. If it is observed, it can be expressed using the linear Gaussian probability density function, i.e., formula (5). To model. Let z represent the linear Gaussian probability density function of state x being observed as z, where z represents the observation vector. The observation matrix of the j-th sensor at the t-th sampling time can be obtained from the observation equation of the multi-sensor system, i.e., formula (2), with variable ξ = x We obtain ξ by performing local linearization at ξ, where ξ is a function. The variable. That is, for a state vector x t The corresponding observation vector is z. t The probability is Suppose that the measurement set consisting of the observation vectors of sensor j from the first sampling time to the t-th sampling time is... So the probability density of the target state x t at the tth sampling time is

[0094] Assume that the number of targets at the tth sampling time is N t , and the state vector of the target is The number of observation vectors of the jth sensor is The observation vector is Because the target state and the observation do not have a specific order, the Random Finite Set (RFS) theory can be naturally used to represent them.

[0095]

[0096] where X t is the multi-target state vector RFS at the tth sampling time; Zj,t represents the observation vector RFS of the jth sensor at the tth sampling time. In distributed multi-target tracking, the states of new-born targets, existing targets, and derived targets (derived targets are targets separated from existing targets, such as a missile dropped from an airplane, which is a derived target, and new-born targets have no relationship with existing targets and are new-born targets) can all be represented by RFS. Assume that the multi-target state vector RFS at the (t-1) th sampling time is X t-1 , and one target state vector x t-1 in X t-1 survives at the tth sampling time with a probability P s,t (x t-1 ) and dies at the tth sampling time with a probability 1-P s,t (x t-1 ), then the target state RFS generated by the state vector x t-1 at the tth sampling time is S t|t-1 (x t-1 ). When the target survives, S t|t-1 (x t-1 ) is {x t}, and when the target dies, S t|t-1 (x t-1 ) is an empty set. Zt|t-1 represents the process from the (t-1) th sampling time to the tth sampling time. Γt|t-1 is the state vector RFS of the derived target generated by the target in the multi-target state vector RFS X t-1 at the (t-1) th sampling time at the tth sampling time.t Let RFS represent the state vector of the newly formed target appearing at the t-th sampling time. Therefore, the multi-target state vector RFS X at the t-th sampling time in formula (6) is... t It can also be expressed as:

[0097]

[0098] Sensors in Wireless Sensor Networks (WSNs) may miss targets due to limitations in their detection capabilities and detection locations. Assume the multi-target state vector RFS at the t-th sampling time is X. t Target state vector x t ∈X t The target is based on probability P D,t (x t ) is observed by the sensor at the t-th sampling time, with 1-P D,t (x t The probability that a target state vector x is missed at the t-th sampling time is given. t ∈X t The resulting measurement vector RFS is θ t (x t When the target state vector x t ∈X t When detected, θ t (x t ) is {z t}, z t For x t ∈X t The observation vector at the t-th sampling time, when the target state vector x t ∈X t When a false negative is detected, θ t (x t () is an empty set. Simultaneously, environmental clutter will also be detected by the sensors; therefore, the observation vector RFS of the j-th sensor at sampling time t is... It can also be expressed as:

[0099]

[0100] in, Represents the clutter observation vector RFS of the j-th sensor at the t-th sampling time; Let X represent the multi-target state vector RFS at the t-th sampling time. t The observation vector RFS of the target in the middle.

[0101] Based on graph theory, we use G = {V(G), E(G)} to describe the communication relationships between sensors in a WSN, where G represents the communication topology of the WSN. denotes a set of sensors in a WSN, denotes a sensor, denotes a set of communication topology edges between sensors of a wireless sensor network, denotes a sensor may communicate with a sensor . denotes a sensor 's set of neighbor sensors, defined as the set of sensors and its set of neighbor sensors , Assume that the communication topology of the WSN is a undirected fully connected network, and the consensus fusion weight matrix is Metropolis matrix The design method is as follows:

[0102]

[0103] where, π a,b denotes the element in the a-th row and the b-th column of the consensus fusion weight matrix Π; |B a | denotes the number of elements in the set of neighbor sensors of the a-th sensor; B a denotes the set of neighbor sensors of the a-th sensor; |B b | denotes the number of elements in the set of neighbor sensors of the b-th sensor; B b denotes the set of neighbor sensors of the b-th sensor; π a,k denotes the element in the a-th row and the k-th column of the consensus fusion weight matrix Π.

[0104] step2: local filtering.

[0105] In order to obtain the closed-form solution of the probability hypothesis density filter, in addition to the assumptions represented by formula (4) and formula (5), the following assumptions need to be made.

[0106] Assumption 1: The survival probability and detection probability of the target are independent of the state of the target, that is:

[0107]

[0108] where, P s,t is the probability that the target at the t-1 sampling time still survives at the t sampling time, that is, the survival probability of the target at the t sampling time; P D,t is the probability that the target is detected by the sensor at the t sampling time, that is, the detection probability of the target at the t sampling time.

[0109] Assumption 2: To simplify the multi-target tracking problem, the derived targets are not considered here, and the intensities of the surviving targets and the newborn targets are approximated using Gaussian mixtures as:

[0110]

[0111] where c' (x) represents the intensity function of the surviving targets at the t-th sampling time from the multi-target at the (t-1)-th sampling time; s represents surviving; J s,t (x) represents the intensity function of the newborn targets at the t-th sampling time; i.e., the newborn multi-target intensity function; J s,t represents the number of the surviving targets at the t-th sampling time from the multi-target at the (t-1)-th sampling time. represents the weight of the i1-th surviving target Gaussian component. represents the state vector of the i1-th surviving target Gaussian component at the t-th sampling time. represents the covariance matrix of the i1-th surviving target Gaussian component at the t-th sampling time; b t (x) represents the intensity function of the newborn targets at the t-th sampling time; i.e., the newborn multi-target intensity function; J b,t represents the number of the newborn targets at the t-th sampling time; b represents newborn. represents the weight of the i2-th newborn target Gaussian component. represents the state vector of the i2-th newborn target Gaussian component. represents the covariance matrix of the i2-th newborn target Gaussian component; β t (x) represents the intensity function of the derived targets at the t-th sampling time from the multi-target at the (t-1)-th sampling time; x is the variable in the function.

[0112] Based on the above assumptions and the Gaussian mixture probability hypothesis density algorithm, it is assumed that the multi-target posterior intensity of the sensor j at the (t-1)-th sampling time is:

[0113]

[0114] where, represents the local posterior multi-target intensity function of the j-th sensor at the (t-1)-th sampling time. represents the estimated target Gaussian component number of the j-th sensor at the (t-1)-th sampling time. represents the weight of the i3-th target Gaussian component of the j-th sensor at the (t-1)-th sampling time. represents the state vector of the i3-th target Gaussian component of the j-th sensor at the (t-1)-th sampling time. represents the covariance matrix of the i3-th target Gaussian component of the j-th sensor at the (t-1)-th sampling time.

[0115] The local prior multi-target intensity function of the jth sensor at the tth sampling time is:

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] wherein, represents the local prior multi-target intensity function of the jth sensor at the tth sampling time; t|t-1 represents a one-step prediction process from the t-1th sampling time to the tth sampling time; represents the multi-target intensity function of the jth sensor at the tth sampling time that survives from the multi-target at the t-1th sampling time, i.e., the surviving multi-target intensity function at the tth sampling time; represents the new-born multi-target intensity function of the jth sensor at the tth sampling time; represents the number of Gaussian components of the jth sensor at the t-1th sampling time, i.e., the number of posterior target Gaussian components in the local posterior multi-target intensity function at the t-1th sampling time; represents the weight of the ith4 posterior target Gaussian component of the jth sensor at the t-1th sampling time; represents the state vector of the ith4 surviving target Gaussian component of the jth sensor at the tth sampling time; represents the covariance matrix of the ith4 surviving target Gaussian component of the jth sensor at the tth sampling time; represents the number of new-born target Gaussian components of the jth sensor at the tth sampling time; represents the weight of the ith5 new-born target Gaussian component of the jth sensor at the tth sampling time; represents the state vector of the ith5 new-born target Gaussian component of the jth sensor at the tth sampling time; represents the covariance matrix of the ith5 new-born target Gaussian component of the jth sensor at the tth sampling time; represents the state transition matrix of the ith4 surviving target Gaussian component of the jth sensor from the t-1th sampling time to the tth sampling time. represents the state vector of the ith4 posterior target Gaussian component of the jth sensor at the t-1th sampling time; denotes the covariance matrix of the i4th posterior target Gaussian component of the jth sensor at the (t - 1)th sampling instant; Q t-1 denotes the covariance matrix of the target motion uncertainty from the (t - 1)th sampling instant to the tth sampling instant.

[0123] denotes the set of observations of the jth sensor at the tth sampling instant. The local posterior multi-target intensity function can be obtained as:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135] wherein, denotes the local posterior multi-target intensity function of the jth sensor at the tth sampling instant; t|t denotes the measurement state update process at the tth sampling instant. denotes the local prior multi-target intensity function of the jth sensor at the tth sampling instant. denotes the number of prior target Gaussian components of the jth sensor at the tth sampling instant. denotes the weight of the i6th posterior target Gaussian component of the jth sensor at the tth sampling instant. denotes the state vector of the i6th posterior target Gaussian component of the jth sensor at the tth sampling instant. denotes the covariance matrix of the i6th posterior target Gaussian component of the jth sensor at the tth sampling instant. denotes the number of posterior target Gaussian components of the jth sensor at the tth sampling instant. denotes the weight of the i7th posterior target Gaussian component of the jth sensor at the tth sampling instant. This represents the state vector of the i7th posterior Gaussian component of the target at the t-th sampling time of the j-th sensor; Let represent the covariance matrix of the 7th posterior Gaussian component of the target at the t-th sampling time of the j-th sensor; This represents the number of posterior target Gaussian components at the (t-1)th sampling time of the j-th sensor; The number of observation vectors in the observation RFS of the j-th sensor at the t-th sampling time; This represents the weight of the 6th prior target Gaussian component at the t-th sampling time of the j-th sensor; express The probability that the observed vector is z; κ represents the state vector of the i-th and 6th prior Gaussian components of the target at the t-th sampling time of the j-th sensor; t This represents the clutter distribution at the t-th sampling time; This represents the weight of the k-th prior target Gaussian component at the t-th sampling time of the j-th sensor; express The probability that the observed vector is z; This represents the state vector of the k-th prior target Gaussian component at the t-th sampling time of the j-th sensor; The observation matrix represents the Gaussian component of the i6th prior target at the t-th sampling time of the j-th sensor; Let represent the covariance matrix of the 6th prior target Gaussian component at the t-th sampling time of the j-th sensor; Let represent the noise covariance matrix of the j-th sensor at the t-th sampling time; The Kalman gain matrix represents the Gaussian component of the i-th prior target at the t-th sampling time of the j-th sensor; Let ξ represent the observation function at the t-th sampling time of the j-th sensor; ξ is a variable in the observation function.

[0136] The Extended Gaussian Mixture Probability Assumption Density Filter (AMP) extends the Gaussian Mixture Probability Assumption Density Filter (AMP) to the field of nonlinear systems by performing local linearization on nonlinear multi-objective motion models and sensor observation models.

[0137] Step 3: Core DBSCAN clustering algorithm.

[0138] From step 2, we can see that the local posterior multi-target intensity function obtained by the sensor after local filtering is in the form of a Gaussian mixture, i.e., formula (20). Although the target has P D,t The probability of N being detected by the sensor is not high, but for those with N... sThe probability that at least two nodes detect the same target is In the actual detection, P D,t will not be very small, so when N s is large, the probability that a target is detected by at least two sensors is large. For different sensors, the detected clutter is different, and after local filtering, the weight of the Gaussian component corresponding to the clutter observation is much smaller than that of the Gaussian component generated by the target observation. First, according to the local posterior multi-target intensity function of each sensor, the mean and covariance of the Gaussian component with a weight not less than a preset weight δ are extracted to form a state pair, which is likely to be the Gaussian component corresponding to the target. For the subsequent selection of the DBSCAN state pair, the calculation burden is reduced, which can be expressed as

[0139]

[0140] wherein, represents a set of state pairs (i.e., state vectors and covariance matrices) of the posterior target Gaussian components of the jth sensor at the tth sampling time; represents a state vector of the ith posterior target Gaussian component of the jth sensor at the tth sampling time; represents a state covariance matrix of the ith posterior target Gaussian component of the jth sensor at the tth sampling time; represents a weight of the ith posterior target Gaussian component of the jth sensor at the tth sampling time.

[0141] Because the mean vectors of the Gaussian components obtained by different sensors all belong to the state space of the target In the target state space, the state vectors belonging to the same target from different sensors are in an aggregated state in space, that is, the state vectors belonging to the same target are very close. The distance between the two state vectors in the two state pairs (x1, P1) and (x2, P2) is defined as:

[0142]

[0143] wherein, d(x1, x2) represents the distance between the state vectors x1 and x2; x1 and x2 are the state vectors of the state pairs (x1, P1) and (x2, P2) respectively; P1 and P2 are the covariance matrices of the state pairs (x1, P1) and (x2, P2) respectively. As can be seen from the foregoing analysis, there are multiple estimated states from different sensors around the target state with a large probability, so The estimated state set corresponding to each target in the is expressed as:

[0144]

[0145] in, This represents the set of estimated states of the u-th target at the t-th sampling time; This represents the state vector of the 9th posterior Gaussian component of the target at the t-th sampling time of the j-th sensor. Let represent the covariance matrix of the 9th posterior Gaussian component of the target at the t-th sampling time of the j-th sensor; Let G represent the set of state pairs of posterior target Gaussian components of all sensors in WSN except for the j-th sensor at sampling time t, where the distance between the posterior target Gaussian component of the j-th sensor and the posterior target Gaussian component of the j-th sensor is less than a preset distance γ. This represents the state vector of the k1-th posterior Gaussian component of the target at the t-th sampling time of the l-th sensor. Let represent the covariance matrix of the k1-th posterior Gaussian component of the target at the t-th sampling time of the l-th sensor. This represents the state vector of the i-th posterior Gaussian component of the target at the t-th sampling time of the j-th sensor. The state vector of the k1-th posterior Gaussian component of the target at the t-th sampling time of the l-th sensor. The distance between them; Let represent the set of state pairs of the posterior Gaussian components of the target at the t-th sampling time of the l-th sensor; The number of target estimated states at the t-th sampling time; |·| represents the number of elements in the set.

[0146] Each estimated state set Each can be considered as a clustered sample point set for DBSCAN clustering.

[0147] First, calculate the ε-neighborhood of each sample point:

[0148]

[0149] in, State vector The estimated state set of the u-th target at the t-th sampling time The ε-neighborhood set is the neighborhood set. Let j1 be the set of state pairs of the posterior target Gaussian component at the t-th sampling time of the j1-th sensor. The i-th 10 The state vector of the state pair of the posterior target Gaussian components, where j1 represents the identity label of the j1th sensor; This represents the set of state pairs representing the posterior Gaussian components of the target at the t-th sampling time of the l1-th sensor. state vector of the state pair of the k2th posterior target Gaussian component in Ω set of state pairs of the l1th sensor at the tth sampling time covariance matrix of the state pair of the k2th posterior target Gaussian component in Ω denotes the distance between and

[0150] Then, the core objects (i.e., core Gaussian components) are selected from the sample point set according to the ε-neighborhood using formula (34).

[0151]

[0152] where Ω u denotes the set of state pairs in Ω whose ε-neighborhood number is greater than MinPts; j2 denotes the identity label of the j2th sensor; denotes the state vector of the ith 11 posterior target Gaussian component of the j2th sensor at the tth sampling time; denotes the covariance matrix of the ith 11 posterior target Gaussian component of the j2th sensor at the tth sampling time; is the state vector in the target estimation state set Ω ; MinPts denotes the minimum neighborhood sample number threshold set when defining the core Gaussian component.

[0153] Different from the traditional DBSCAN clustering method, the core DBSCAN clustering proposed in the present application only clusters the core Gaussian components, and the specific iterative steps are as follows:

[0154]

[0155] where Ω is the set of state pairs in Ω whose ε-neighborhood number is greater than MinPts; Q denotes the set of core Gaussian components to be clustered; is the estimated state pair in Ω u denotes the state vector of the ith 12 posterior target Gaussian component of the j3th sensor at the tth sampling time; denotes the covariance matrix of the ith 12 ​​a covariance matrix of the k3th posterior target Gaussian component of the l2th sensor at the tth sampling time; l2 is the identity tag of the l2th sensor; Δ is the set of clustering core Gaussian components after clustering iteration ends; l3 is the identity tag of the l3th sensor; denotes a state vector in the set of target estimation state ε-neighborhoods; denotes a state vector of the k3th posterior target Gaussian component of the l2th sensor at the tth sampling time; denotes a covariance matrix of the k3th posterior target Gaussian component of the l2th sensor at the tth sampling time; l2 is the identity tag of the l2th sensor; Δ is the set of clustering core Gaussian components after clustering iteration ends; l3 is the identity tag of the l3th sensor; denotes a state vector of the k4th posterior target Gaussian component of the l3th sensor at the tth sampling time; denotes a covariance matrix of the k4th posterior target Gaussian component of the l3th sensor at the tth sampling time; denotes the uth set of target estimation state obtained initial normal sensor identity tag set.

[0156] Due to the influence of detection probability, not all sensors can observe the same target, so the state pair in the set of target estimation state is difficult to contain all sensors. The clustering label obtained after core DBSCAN clustering will not contain the missed sensor, but the missed sensor may be a normal sensor, so the missed sensor is also considered as a normal sensor to add to the clustering label set:

[0157]

[0158] wherein, denotes the uth set of target estimation state obtained normal sensor identity tag set including missed sensors; l4 is the identity tag of the l4th sensor; denotes a state vector of the k5th posterior target Gaussian component of the l4th sensor at the tth sampling time; denotes a covariance matrix of the k5th posterior target Gaussian component of the l4th sensor at the tth sampling time; denotes a set of state pairs of posterior target Gaussian components of the l4th sensor at the tth sampling time. The clustering label set obtained after core DBSCAN clustering of each set of target estimation state contains missed sensors; but the probability of the same sensor entering the clustering label set through missed detection is In actual MTT scenarios, P D,t It won't be very small; when When the value is relatively large, the sensor enters through missed detection. The probability of a cluster of label sets is almost zero. From this perspective, if a sensor has a value of not less than... If a node appears in a set of cluster labels, then the difference between the sensor's local estimated multi-target state and the actual target state is small; that is, the sensor has normal detection accuracy. The strategy for determining the normal detection accuracy of sensors is as follows:

[0159]

[0160]

[0161] in, Let be the set of normal sensor tags at the t-th sampling time (normal sensors are sensors with normal detection accuracy); Let be the set of identification labels of the j-th sensor that are determined to be normal sensors at the t-th sampling time. This is the minimum number of cluster labels to be defined.

[0162] In summary, the core DBSCAN clustering algorithm first extracts state pairs from the Gaussian components with larger weights in the local posterior intensity of each sensor. This operation removes most state pairs from clutter Gaussian components, thus reducing the computational burden of subsequent state pair association. The distance definition between two state vectors in state pair association introduces the covariance matrix of the estimated state vectors. Compared to the Euclidean distance definition, this definition corresponds to an ellipse in space, better reflecting the difference between the two estimated state vectors. The core DBSCAN clustering algorithm only clusters the core Gaussian components and clusters the estimated state sets of multiple targets separately. Compared to clustering the estimated state sets of multiple targets together, this further reduces the influence of estimated states from clutter Gaussian components.

[0163] Step 4: Improve the initial step of arithmetic mean fusion.

[0164] The set of normal sensor tags is obtained through the core DBSCAN clustering algorithm. After that, the local posterior intensities of normal sensors need to be fused. In order to eliminate the local posterior intensities of the sensor with low detection accuracy while achieving the consistency of distribution without changing the communication topology of the sensor, the initial step of the arithmetic average (AA) fusion method is improved in the improved arithmetic average (IAA) fusion method. In the initial step of the IAA fusion method, the sensor with low detection accuracy receives the local posterior intensities of its neighbor nodes with normal detection accuracy, and the local posterior intensities of the neighbor nodes with normal detection accuracy are fused by using the AA fusion method. At the same time, the sensor with normal detection accuracy receives the local posterior intensities of its neighbor nodes with normal detection accuracy, and the local posterior intensities of the neighbor nodes with normal detection accuracy and the local posterior intensity of this node are fused by using the AA fusion method. The multi-target intensity after the AA fusion is used as the initial fusion intensity of the distributed fusion. The initial step of the IAA fusion method can not change the communication topology of the WSN, but only needs to modify the fusion weight matrix.

[0165] Firstly, according to the normal sensor label set The modified matrix E t is established at the t th sampling moment

[0166]

[0167] wherein, The modified matrix E t The element in the a th row and the b th column of the modified matrix E

[0168] Secondly, the modified matrix E is used t The consistency fusion weight matrix Π is modified, and the modification formula is:

[0169]

[0170] wherein, The modified consistency fusion weight matrix at the t th sampling moment.

[0171] Thirdly, The sum of the rows is not 1, so the row normalization processing is needed to obtain the initial weight matrix The normalization formula is:

[0172]

[0173] wherein, The initial weight matrix at the t th sampling moment The element in the a th row and the b th column of the initial weight matrix at the t th sampling moment, which represents the weight value of the target state information of the b th sensor in the a th sensor in the initial fusion. This represents the corrected consistency fusion weight matrix at the t-th sampling time. The element in row a and column b.

[0174] Finally, use the initialized weight matrix. The initial fusion intensity function for each sensor is obtained as follows:

[0175]

[0176] in, Let f(j) represent the initial fusion intensity function of the j-th sensor at the t-th sampling time. This represents the initial weight matrix at the t-th sampling time. The j-th row, i-th 13 The elements of the column represent the i-th element during initial fusion. 13 The weight value for fusing the target state information of each sensor in the j-th sensor; Represents the i-th 13 The local posterior multi-object intensity function of a sensor at the t-th sampling time.

[0177] Step 5: Improve the iterative steps of arithmetic average fusion.

[0178] After initialization, the initial fusion strength of each sensor is independent of the local posterior strength of the node where the detection accuracy decreases. Therefore, during the consensus fusion iteration, the iterative formula for obtaining the final fusion strength function using the consensus fusion weight matrix Π is:

[0179]

[0180] in, Let f(x) represent the fusion intensity function of the j-th sensor at the t-th sampling time during the (11+1)-th fusion iteration. When the fusion iteration reaches the number of consistent fusion iterations, the final fusion intensity function is obtained. Indicates the i-th 13 The fusion intensity function of a sensor at the t-th sampling time and the ll-th fusion iteration; The consistency fusion weight matrix Π represents the i-th row of the j-th column. 13 The elements of the column represent the i-th element. 13 The weight value for fusing the target state information of each sensor to the j-th sensor.

[0181] Meanwhile, the fusion of Gaussian components at each sensor and the pruning of only the Gaussian components with weight greater than λ are performed for the next iteration of consistent fusion. Thus, a closed form of consistent fusion iteration is formed. Because only the Gaussian components with large weight values are retained and transmitted by the sensor during consistent fusion, the implementation here is "partial consistency", which only fuses and transmits the Gaussian components that are likely to come from the target, thereby reducing the computational burden and communication burden.

[0182] Specifically, in order to verify the advantages of the above method, simulation verification is also performed, and the process is as follows:

[0183] A multi-target tracking simulation scene with partial sensor detection accuracy degradation is constructed to demonstrate the effectiveness of the DBSCAN-based distributed multi-target tracking (DMTT) algorithm of the application in multi-target tracking with partial sensor detection accuracy degradation, and the tracking results are compared with those of the arithmetic average based distributed Gaussian mixture probability hypothesis density (AA-DGMPHD) algorithm and the geometric average based distributed Gaussian mixture probability hypothesis density (GA-DGMPHD) algorithm.

[0184] The tracking scene is a two-dimensional plane of [-1000m, 1000m] x [-1000m, 1000m] as shown in Figure 3 , wherein there are 6 targets with different appearance and disappearance times, and 10 distributed sensors. The initial state of the target is As shown in Table 1, x a is the initial x-axis position of the target, is the initial x-axis velocity of the target, y a is the initial y-axis position of the target, is the initial y-axis velocity of the target. The start and end times of the trajectory are shown in Table 2. Table 3 records the positions (x s , y s ) of the 10 sensors, x s is the x-axis position of the s-th sensor, and y s is the y-axis position of the s-th sensor. The communication topology between the sensors is a fully connected network, and the sensors can communicate with the sensors within a distance of 2000m, as shown in Figure 3 , Figure 3The middle dotted line is the communication topology between sensors. The total simulation sampling time is 1000s, and the sampling interval is T1=1s. The motion model of the target is a linear model of uniform turning as shown in equation (44).

[0185]

[0186] where x t is the state vector of the target at the tth sampling time, x t-1 is the state vector of the target at the (t-1)th sampling time, ω= -0.02s -1 is the turning rate, is the noise distribution, Q=diag([4m 2 ,0.2m 2 / s 2 ,4m 2 ,0.2m 2 / s 2 ]) is the covariance matrix of the noise distribution.

[0187] Table 1 Initial state table of the target

[0188]

[0189] Table 2 Start and end time table of the target trajectory

[0190]

[0191] Table 3 Sensor position table

[0192]

[0193] As shown in equation (45), the observation model of the sensor is a nonlinear model

[0194]

[0195] where z t is the observation vector at the tth sampling time. The noise of the sensor with normal detection accuracy is R n =diag([0.25m 2 ,3×10 - 6 rad 2 / s 2 ]), and the noise d of the sensor with reduced detection accuracy is 2 =diag([4m -5 ,4.87×10 2 rad 2 / s]) is the covariance matrix of the noise. The numbers of precision-degraded sensors and their corresponding time instants of precision degradation are shown in Table 4.

[0196] Table 4 Numbers of precision-degraded sensors and their corresponding time instants of precision degradation

[0197]

[0198] In Table 4, T start denotes the start time; T end denotes the end time.

[0199] The survival probability of the target is P s = 0.99, and the detection probability of different sensors is P D = 0.6. The parameters in the fusion step are β = 30, λ = 0.8. The parameters of DBSCAN clustering are ε = 15, MinPts = 3. The number of clutters in the tracking scenario obeys a Poisson distribution with a mean of 2. Monte Carlo simulation is repeated 100 times. In order to compare and measure the effects of the three algorithms, the number of target estimates and the optimal sub-patten assignment (OSPA) distance are used as two measurement means, and the calculation formula of OSPA is:

[0200]

[0201] where D p,c (X, Y) is the OSPA distance between the random finite sets X and Y, X = [x1, x2, …, x m ] represents the estimated multi-target state RFS, m represents the number of state vectors of X, Y = [y1, y2, …, y n ] is the true multi-target state RFS, n represents the number of state vectors in Y, represents the Euclidean distance between x and y, Π n represents the set of permutations of n elements taken from the set [y1, y2, …, y n ], p represents the order, and c represents the minimum cut-off distance, p = 10, c = 2. The estimated target numbers and OSPA of the proposed DBSCAN-based DMTT algorithm, AA-DGMPHD and GA-DGMPHD are shown in Figure 4 and Figure 5 .

[0202] From the estimated target number simulation results of the three DMTT algorithms, it can be seen that the effect of the DBSCAN-based DMTT algorithm is better than that of the AA-DGMPHD and GM-DGMPHD. Before the situation of the partial sensors appearing the detection precision reduction, the estimated target number of the DBSCAN-based DMTT algorithm is almost the same as that of the AA-DGMPHD, and the difference between the two increases after the partial sensors appear the detection precision reduction, because the local estimation strength corresponding to the sensor detection precision reduction is quite different from the local estimation strength of the normal precision node, introducing these local estimation strengths into the distributed consistency fusion is equivalent to reducing the weight of the Gaussian term corresponding to the real target, and the target estimation number obtained after the fusion and pruning is lower than that of the IAA fusion method of deleting the local estimation strength of the detection precision reduction node. Compared with the DBSCAN-based DMTT algorithm and the AA-DGMPHD, the estimated target number obtained by the GM-DGMPHD is quite different from the real target number, and most of the time is lower than the real target number, because the characteristics of the GA fusion method determine that the same part of the estimation strength is enhanced and the different part is greatly reduced, which leads to the fact that the strength of the missed target is greatly reduced after the GA fusion, thereby leading to the fact that the estimated number is less than the real value. The performance of the three algorithms in the OSPA is basically the same as the estimated target number, and the details are not given here. Figure 3 It can be seen that whether the sensor detection precision reduces or not has almost no effect on the OSPA of the DBSCAN-based DMTT algorithm, and the OSPA of the DBSCAN-based DMTT algorithm almost does not change in the whole tracking process. The OSPA of the AA-DGMPHD and the GM-DGMPHD is affected to different degrees, and the AA-DGMPHD is less affected than the GM-DGMPHD, because the estimated target number of the AA-DGMPHD is closer to the real value than that of the GM-DGMPHD, and the AA fusion loses fewer targets.

[0203] In summary, the DBSCAN-based DMTT algorithm of the present application can keep good MTT tracking effect compared with the AA-DGMPHD and GM-DGMPHD algorithms in the case of the partial sensors appearing the detection precision reduction.

[0204] The present application has the following advantages:

[0205] 1. The Gaussian mixture probability hypothesis density algorithm is extended to the nonlinear field.

[0206] 2. The local posterior multi-target probability distribution of the neighbor sensor with normal detection precision is used to replace the local posterior multi-target probability distribution of the sensor with reduced detection precision in the initial step of the consistency fusion.

[0207] 3. The local posterior multi-target probability distribution of the communication topology can be kept unchanged while the detection accuracy of the sensor is removed during consistency fusion.

[0208] 4. The accuracy of the distributed multi-target tracking can be improved under the condition of the accuracy of part of the sensors.

[0209] Embodiment 2

[0210] The multi-target tracking system in the embodiment comprises:

[0211] The local posterior multi-target intensity function construction module is configured to construct the local posterior multi-target intensity function of each sensor at each sampling time by using the extended Gaussian mixture probability hypothesis density algorithm.

[0212] The tracking module is configured to determine the state of the tracked target in real time based on the local posterior multi-target intensity function of each sensor at each sampling time.

[0213] The determination process of the state of the plurality of tracked targets at any sampling time comprises:

[0214] Each posterior target Gaussian component in the local posterior multi-target intensity function of each sensor whose weight satisfies the preset weight is determined as a first screening Gaussian component.

[0215] A plurality of target estimated state sets are determined according to all the first screening Gaussian components; the target estimated state set comprises a plurality of second screening Gaussian components.

[0216] A neighborhood set of each second screening Gaussian component is calculated.

[0217] A core Gaussian component corresponding to each target estimated state set is determined based on all the neighborhood sets corresponding to the target estimated state set.

[0218] A normal sensor identity tag set is determined according to each core Gaussian component.

[0219] An initial fusion intensity function of each sensor is constructed based on the normal sensor identity tag set and the local posterior multi-target intensity function of each sensor.

[0220] A final fusion intensity function of each sensor is determined based on a consistency fusion weight matrix, the initial fusion intensity function of each sensor and the number of iterations of the consistency fusion; the consistency fusion weight matrix is determined based on the communication topology in the wireless sensor network.

[0221] The state of the plurality of tracked targets is determined according to the final fusion intensity function of each sensor.

[0222] Embodiment 3

[0223] An electronic device comprises:

[0224] one or more processors.

[0225] a memory device storing one or more programs.

[0226] The one or more programs, when executed by the one or more processors, enable the one or more processors to implement the multi-target tracking method in embodiment 1.

[0227] Embodiment 4

[0228] A storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the multi-target tracking method in embodiment 1.

[0229] The various embodiments described in the specification are presented in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0230] The principles and implementation manners of the present application are described by using specific examples in the present application. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges can be changed. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A multi-target tracking method applied to a wireless sensor network including a plurality of sensors and a plurality of tracked targets, characterized by, The method comprises: constructing a local posterior multi-target intensity function of each sensor at each sampling time by using an extended Gaussian mixture probability hypothesis density algorithm; determining the state of a tracked target in real time based on the local posterior multi-target intensity function of each sensor at each sampling time; wherein the determination process of the state of a plurality of tracked targets at any sampling time comprises: determining each posterior target Gaussian component in the local posterior multi-target intensity function of each sensor as a first screening Gaussian component, wherein the weight of each posterior target Gaussian component meets a preset weight; determining a plurality of target estimated state sets according to all first screening Gaussian components; the target estimated state set comprises a plurality of second screening Gaussian components; calculating a neighborhood set of each second screening Gaussian component; determining a core Gaussian component corresponding to each target estimated state set based on all neighborhood sets corresponding to each target estimated state set; determining a normal sensor identity tag set according to each core Gaussian component; constructing an initial fusion intensity function of each sensor based on the normal sensor identity tag set and the local posterior multi-target intensity function of each sensor; determining a final fusion intensity function of each sensor based on a consistency fusion weight matrix, the initial fusion intensity function of each sensor and a consistency fusion iteration number; the consistency fusion weight matrix is determined based on a communication topology in a wireless sensor network; determining the state of a plurality of tracked targets according to the final fusion intensity function of each sensor.

2. The multi-target tracking method of claim 1, wherein, constructing a local posterior multi-target intensity function of the jth sensor at the tth sampling time by using an extended Gaussian mixture probability hypothesis density algorithm, comprising: obtaining a state vector of a plurality of new target Gaussian components at the tth sampling time of the jth sensor, a covariance matrix of the plurality of new target Gaussian components at the tth sampling time, and a state vector and a covariance matrix of a plurality of posterior target Gaussian components in the local posterior multi-target intensity function at the t-1th sampling time; constructing a survival multi-target intensity function at the tth sampling time according to the state vector and the covariance matrix of all posterior target Gaussian components at the t-1th sampling time; constructing a new multi-target intensity function at the tth sampling time according to the state vector of all new target Gaussian components at the tth sampling time and the covariance matrix of all new target Gaussian components; constructing a local prior multi-target intensity function at the tth sampling time according to the survival multi-target intensity function at the tth sampling time and the new multi-target intensity function at the tth sampling time; the local prior multi-target intensity function at the tth sampling time comprises a state vector of a plurality of prior target Gaussian components at the tth sampling time and a covariance matrix of the plurality of prior target Gaussian components at the tth sampling time, wherein the prior target Gaussian component at the tth sampling time comprises all survival target Gaussian components at the tth sampling time and all new target Gaussian components at the tth sampling time; determining a state vector and a covariance matrix of a posterior target Gaussian component at the tth sampling time by using a Kalman filtering method according to the state vector and the covariance matrix of the prior target Gaussian component at the tth sampling time and an observation vector at the tth sampling time; According to the local prior multi-target intensity function at the tth sampling moment and the state vector and the covariance matrix of each posterior target Gaussian component at the tth sampling moment, a local posterior multi-target intensity function at the tth sampling moment is constructed.

3. The multi-target tracking method of claim 2, wherein, According to the state vector and the covariance matrix of all the posterior target Gaussian components at the t-1th sampling moment, a surviving multi-target intensity function at the tth sampling moment is constructed, specifically including: According to the state vector of each posterior target Gaussian component at the t-1th sampling moment, a state vector of a surviving target Gaussian component at the tth sampling moment corresponding to the posterior target Gaussian component is determined; According to the covariance matrix of each posterior target Gaussian component at the t-1th sampling moment, a covariance matrix of a surviving target Gaussian component at the tth sampling moment corresponding to the posterior target Gaussian component is determined; According to the state vector of all the surviving target Gaussian components at the tth sampling moment and the covariance matrix of all the surviving target Gaussian components at the tth sampling moment, a surviving multi-target intensity function at the tth sampling moment is constructed.

4. The multi-target tracking method of claim 2, wherein, According to the state vector and the covariance matrix of each prior target Gaussian component at the tth sampling moment and the observation vector at the tth sampling moment, the state vector and the covariance matrix of each posterior target Gaussian component at the tth sampling moment are determined by using the Kalman filtering method, specifically including: According to the covariance matrix of each prior target Gaussian component at the tth sampling moment, a Kalman gain matrix at the tth sampling moment corresponding to the prior target Gaussian component is determined; According to the Kalman gain matrix at the tth sampling moment, the state vector of the corresponding prior target Gaussian component at the tth sampling moment and the observation vector at the tth sampling moment, the state vector of the corresponding posterior target Gaussian component at the tth sampling moment is determined; According to the Kalman gain matrix at the tth sampling moment and the covariance matrix of each prior target Gaussian component at the tth sampling moment, the covariance matrix of the corresponding posterior target Gaussian component at the tth sampling moment is determined.

5. The multi-target tracking method of claim 1, wherein, The determination process of any current target estimation state set includes: The distance between each first screening Gaussian component of each non-current sensor in the wireless sensor network and the current first screening Gaussian component is calculated respectively, and is denoted as a first distance; the non-current sensor is a sensor other than the current sensor in all sensors; All the first screening Gaussian components of all the non-current sensors corresponding to the first number of current first screening Gaussian components satisfying the preset number are determined as second screening Gaussian components, and a set composed of the second screening Gaussian components is determined as the current target estimation state set; the first number is the number of target non-current sensors whose first distances satisfy the preset distance, and the target non-current sensor is a non-current sensor whose first distance satisfies the preset distance.

6. The multi-target tracking method of claim 1, wherein, Specifically, the determination process of any current second screening Gaussian component neighborhood set includes: The distance between the current second screening Gaussian component and each non-current second screening Gaussian component is calculated respectively, and is denoted as a second distance; the non-current second screening Gaussian component is a second screening Gaussian component other than the current second screening Gaussian component in all the second screening Gaussian components of the target estimation state set in which the current second screening Gaussian component is located; All non-current second screening Gaussian components with a second distance less than the neighborhood radius are determined as a neighborhood set of the current second screening Gaussian component.

7. The multi-target tracking method of claim 1, wherein, The construction process of any initial fusion intensity function specifically includes: establishing a correction matrix according to a normal sensor identity tag set; correcting a consensus fusion weight matrix based on the correction matrix to obtain a corrected consensus fusion weight matrix; normalizing the corrected consensus fusion weight matrix to obtain an initialization weight matrix; constructing an initial fusion intensity function according to the initialization weight matrix and the local posterior multi-target intensity function of each sensor.

8. A multi-target tracking system, characterized by The system comprises: a local posterior multi-target intensity function construction module configured to construct the local posterior multi-target intensity function of each sensor at each sampling time by using an extended Gaussian mixture probability hypothesis density algorithm; a tracking module configured to determine the state of a tracked target in real time based on the local posterior multi-target intensity function of each sensor at each sampling time; wherein the determination process of the state of each tracked target at any sampling time comprises: determining each posterior target Gaussian component with a weight satisfying a preset weight in the local posterior multi-target intensity function of each sensor as a first screening Gaussian component; determining a plurality of target estimated state sets according to all first screening Gaussian components; the target estimated state set comprises a plurality of second screening Gaussian components; calculating a neighborhood set of each second screening Gaussian component; determining a core Gaussian component corresponding to each target estimated state set based on all neighborhood sets corresponding to the target estimated state set; determining a normal sensor identity tag set according to each core Gaussian component; constructing an initial fusion intensity function of each sensor based on the normal sensor identity tag set and the local posterior multi-target intensity function of each sensor; determining a final fusion intensity function of each sensor based on the consensus fusion weight matrix, the initial fusion intensity function of each sensor and a consensus fusion iteration number; the consensus fusion weight matrix is determined based on a communication topology in a wireless sensor network; determining the state of a plurality of tracked targets according to the final fusion intensity function of each sensor.

9. An electronic device, comprising: comprise: one or more processors; a storage device having one or more programs stored thereon; when the one or more programs are executed by the one or more processors, the one or more processors implement the multi-target tracking method according to any one of claims 1 to 7.

10. A storage medium, characterized by a computer program is stored thereon, wherein the computer program is executed by a processor to implement the multi-target tracking method according to any one of claims 1 to 7.

Images