Distributed gm-lphd multi-target tracking method based on sc-lm

Abstract

The application discloses a decentralized GM-LPHD multi-target tracking method based on SC-LM. A decentralized multi-sensor multi-target tracking scene is constructed; the state and observation of the multi-target are modeled; a GM-LPHD filter is modeled; the posteriori Gaussian mixture set of all sensors is processed by SC-LM and then fused by using a decentralized fusion architecture; and the multi-target tracking is realized by using the GM-SC-LM-LPHD filter in the decentralized framework. The state clustering is used to match the label space of different sensor nodes, so that the labels of different sensors are consistent when the GCI fusion is performed, and the accurate tracking of the multi-target is realized. The application effectively suppresses the fusion performance decline problem caused by the label inconsistency between the decentralized sensors, and the tracking performance is improved compared with the original decentralized multi-sensor GM-LPHD filter, so that the application can be widely applied in multi-robot and unmanned aerial vehicle clusters.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of multi-sensor multi-target tracking, and relates to a decentralized GM-LPHD multi-target tracking method based on state clustering-label matching (SC-LM), which is used for solving the problem of fusion quality decline caused by inconsistent label space in a decentralized architecture, and improving the tracking quality of multiple targets in a detection space. BACKGROUND

[0002] Multi-target tracking refers to estimating the number, state and other information of targets from complex data obtained by sensors. Multi-sensor multi-target tracking methods are mainly divided into three categories: ① centralized fusion architecture; ② distributed fusion architecture; and ③ decentralized fusion architecture. The centralized and distributed architectures are both centralized, and the failure of the fusion center will cause the entire system to be paralyzed, and is not conducive to the access and exit of network nodes, and the network flexibility is limited. In order to solve these problems, the decentralized architecture without a center is proposed. In the decentralized network, there is no unified fusion center, each node can process local observation data, nodes can communicate with each other point-to-point, and the status between nodes is peer-to-peer. The advantages of this architecture compared with the centralized fusion framework are: ① strong survivability: the system performance does not depend on any node, and the failure of any node will not cause the paralysis of the entire network; ② flexibility: since the network is without a center and does not require global knowledge of the network topology, the system can flexibly add or lose nodes online; ③ scalability: by reducing the centralized computing bottleneck and the limitation of lack of communication bandwidth, the system has scalability. Due to the advantages of the decentralized architecture in many aspects, it has wide application in unmanned aerial vehicle clusters and multi-robot systems.

[0003] The following are some representative works based on the random finite set (RFS) tracking method:

[0004] In 2003, R. Mahler proposed a PHD filter based on RFS, which simplified the probability space modeling of state variables to the probability hypothesis space modeling of state variables, but did not obtain a closed-form solution. Subsequently, B. N. Vo, an Australian scholar, proposed a Gaussian mixture PHD (GM-PHD) filter and a sequential Monte Carlo PHD (SMC-PHD) filter. In 2007, R. Mahler proposed a CPHD filter. In 2009, Vo et al. proposed a tracker (LPHD) based on the GM-PHD filter to provide target label information, since the GM-PHD filter does not provide target identity information. Subsequently, Vo et al. proposed a distributed consensus fusion formula for label random finite set (LRFS).

[0005] The traditional LRFS-based generalized covariance fusion (GCI) is fused according to labels, and therefore, when the probability density of the LRFS is fused by using the GCI, perfect label consistency between different sensors is required, that is, the same track needs to have the same label in all sensor nodes. When the label consistency between the sensors is not met, the performance after the GCI fusion will be reduced. Therefore, the state clustering-label matching (SC-LM) is proposed, so that the label consistency between different sensors is achieved when the GCI fusion is performed, and the accurate tracking of the multiple targets is realized. SUMMARY

[0006] An object of the present application is to propose a decentralized GM-LPHD multi-target tracking method based on state clustering-label matching (SC-LM), referred to as GM-SC-LM-LPHD, to overcome the deficiencies of the prior art, and the specific content is as follows:

[0007] Step (1), a decentralized multi-sensor multi-target tracking scene is constructed, and relevant parameters are set, including the process noise of target motion and the observation noise of the sensor;

[0008] Step (2), the state and observation of the multiple targets are modeled;

[0009] Step (3), a single-sensor GM-LPHD filter is modeled

[0010] Step (4), based on steps (2) and (3), the SC-LM processing is performed on the posterior Gaussian mixture set of all sensors, and then the fusion is performed, by using the decentralized fusion architecture.

[0011] Step (5), the GM-SC-LM-LPHD filter constructed is used to realize the multi-target tracking in the decentralized framework.

[0012] Another object of the present application is to provide a computer readable storage medium, which stores a computer program, and when the computer program is executed in a computer, the computer is caused to execute the method described above.

[0013] Still another object of the present application is to provide a computing device, which includes a memory and a processor, the memory stores executable code, and when the processor executes the executable code, the method described above is realized.

[0014] The present application has the following beneficial effects:

[0015] The application provides a complete distributed multi-sensor multi-target tracking scheme, and state clustering is used to match the label space of different sensor nodes, so that the labels of different sensors are consistent when GCI fusion is performed, and accurate tracking of multi-target is realized. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flow chart for specifically implementing the core part of the method of the application;

[0017] Figure 2 is a target motion trajectory, a filtered track and a distributed sensor position map;

[0018] Figure 3 is a comparison chart of OSPA average values filtered by the method of the application and a traditional distributed GM-LPHD algorithm under 100 times of Monte Carlo;

[0019] Figure 4 is a comparison chart of target estimation number average values filtered by the method of the application and a traditional distributed GM-LPHD algorithm under 100 times of Monte Carlo; DETAILED DESCRIPTION

[0020] The application will be further analyzed in combination with the specific implementation mode.

[0021] The SC-LM-based distributed GM-LPHD multi-target tracking method comprises the following steps: step (1), constructing a distributed multi-sensor multi-target tracking scene, and setting related parameters, including process noise of target motion and observation noise of the sensor, and the specific steps are as follows:

[0022] The target moves in a two-dimensional plane, and the state is x=[p x ,v x ,p y ,v y ] T , wherein p x , p y are positions of the target in x and y directions respectively, v x , v y are velocities of the target in x and y directions respectively, and T represents transposition;

[0023] The process noise covariance of target motion is set as , wherein σ x (k) 2 represents the noise variance of the target in the x direction at the k time, and σ y (k)2 represents the noise variance of the environment to the target in y direction at time k;

[0024] Set the observation noise covariance of the sensor as where δ x 2 , δ y 2 respectively represent the noise variance of the sensor to the target x position and y position;

[0025] Step (2), modeling the state and observation of the target, specifically:

[0026] 2-1 Modeling of the state of the target

[0027] Use a labeled random finite set to represent the state of multiple targets at time k, N k represents the number of targets at time k, where represents the label space of the target at time k, represents the state of the i-th target at time k, l i represents the label of the i-th target;

[0028] The transition of the target from a certain state at time k-1 to a certain state at time k is called state transition, and the state transition probability density of each target is assumed to conform to a linear Gaussian model, which is represented as:

[0029]

[0030] where represents the spatial distribution of the target state , which is subject to a Gaussian distribution with a state mean of and a covariance of Q k-1 , F k-1 is a state transition matrix, and Q k-1 is a process noise covariance matrix;

[0031] 2-2 Modeling of observation

[0032] It is assumed that there are multiple sensors independent of each other in the tracking scene; if the target i is detected by the sensor s j at time k, the observation equation is represented as:

[0033]

[0034] where, represents the observation vector of the target i by the sensor s j at time k; H is an observation matrix;​ denotes the state of the i-th target at time k; denotes the sensor s j the observation noise covariance at time k;

[0035] the i-th target at time k; j In addition to the possible acquisition of the target's measurement, the environment clutter can also be acquired; the sensor s j The environment clutter obtained at time k is denoted by equation (3):

[0036]

[0037] where ρ(n k ) denotes the probability density function of the Poisson distribution with parameter λ; n k denotes the expected number of clutters at time k; f(c) denotes the probability density function of the clutter c in the observation space, and V is the volume of the observation space;

[0038] Therefore, at time k, the observation data of the sensor s j is denoted as where the sensor s j The i-th measurement value monitored at time k may be a target or a clutter; step (3), the single-sensor GM-LPHD filter modeling, is as follows:

[0039] 3-1 Single-sensor GM-LPHD prediction

[0040] Assume that the posterior Gaussian mixture set at time k-1 is where denote the weight, the state mean, and the covariance of the i-th GM-LPHD component at time k-1, respectively, and J k-1 denotes the number of posterior Gaussian components at time k-1, then the prediction of the i-th GM-LPHD at time k is denoted by equations (4)-(6)

[0041]

[0042]

[0043]

[0044] In the formula, denotes the prediction weight of the i-th GM-LPHD component at time k; p s,k denotes the probability of the target surviving at time k; denotes the state mean prediction of the i-th GM-LPHD component at time k; F k-1 denotes the state transition matrix at time k-1; denotes the covariance prediction of the ith GM-LPHD component at time k; Q k-1 denotes the process noise covariance at time k-1; superscript T denotes transpose;

[0045] 3-2 Single-sensor GM-LPHD update

[0046] Based on the GM-LPHD prediction step, the predicted LPHD components of the multi-target at time k are obtained, and the LPHD is further updated according to equations (7)-(11):

[0047]

[0048]

[0049]

[0050]

[0051]

[0052] wherein denotes the update weight of the ith GM-LPHD component at time k, P d,k denotes the probability that the target is detected by the sensor at time k, c k (z) denotes the intensity of the clutter, J k|k-1 denotes the number of predicted Gaussian components at time k, denotes the probability density function of the observation prediction of the ith target at time k, H k is the observation matrix of the sensor at time k, R k is the measurement noise covariance at time k, denotes the state mean update of the ith GM-LPHD component at time k, denotes the Kalman gain of the ith target at time k, z denotes the observation value of the sensor to , denotes the covariance update of the ith GM-LPHD component at time k, I is the identity matrix;

[0053] Step (4), based on steps (2)-(3), the posterior Gaussian set of all sensors is first processed by SC-LM using a decentralized fusion architecture, and then fused; the details are as follows:

[0054] 4-1 At time k, based on steps (2)-(3), the sensor s j is filtered by GM-LPHD to obtain the labeled posterior Gaussian mixture set wherein denote the posterior Gaussian set of the sensor s jThe weight, state mean, and covariance of the i-th GM-LPHD component after step (3) filtering, s j The number of GM-LPHD components after step (3);

[0055] 4-2 State clustering (SC)

[0056] Assume any two sensors s q , s q The label posterior Gaussian set of s According to formula (12), the distance d(i,j) between any two label posterior Gaussian components i and j in s q , s p is calculated as follows:

[0057]

[0058] Where ||·||2 represents the 2-norm;

[0059] According to formula (12), the cost matrix D between s q and s p is as follows:

[0060]

[0061] Where N(s q ), N(s p ) represent the number of label posterior Gaussian components in s q , s p .

[0062] 4-3 Label matching (LM)

[0063] The matching relationship between the posterior states of s q and s p is obtained by formula (14) from the cost matrix D obtained by 4-2 clustering:

[0064]

[0065] Where π * represents the matching relationship between the Gaussian components in s q and the Gaussian components in s p , and π * (i) = j means that the i-th label Gaussian component in s q matches the j-th label Gaussian component in s p in state.

[0066] For all matched pairs (i, j), their labels are paired when the condition of equation (15) is satisfied:

[0067] d(i, π * (i)) < c (15)

[0068] where c is a threshold, which is set by human.

[0069] The matched label Gaussians and are processed as follows:

[0070]

[0071] where l f denotes the label of the fused label Gaussian;

[0072] According to equation (16), the label Gaussians and are updated to and

[0073] 4-4 Decentralized geometric average fusion

[0074] After SC-LM processing of the label posterior Gaussians in sensor s q and sensor s p , geometric average (GA) fusion is performed, where the formula of GA is as follows:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] where ω1, ω2 denote the fusion coefficients of GA fusion, denotes the fusion weight of the i-th and j-th label posterior Gaussians after SC-LM processing, denotes the fusion covariance of the i-th and j-th label posterior Gaussians after SC-LM processing; denotes the fusion mean of the i-th and j-th label posterior Gaussians after SC-LM processing.

[0081] 4-5 judge whether the fusion times of the distributed geometric mean fusion reach the consistency fusion times requirement, if not, jump to step 4-2 to continue to execute, if yes, output the filtering result.

[0082] Step (5), using the constructed GM-SC-LM-LPHD filter to realize multi-target tracking in a distributed framework.

[0083] The target tracking effect of the application is shown in Figure 2 、 Figure 3 、 Figure 4 ; Figure 2 For tracking scenes, real trajectories and filtering trajectories of 11 targets are contained, and 16 distributed sensors with different fields of view are contained; Figure 3 、 Figure 4 The target estimation number average comparison chart and the OSPA average comparison chart after 100 times of Monte Carlo simulation are shown in Figure 3 、 Figure 4 It can be seen that the ordinary distributed GM-LPHD filter has very low precision after fusion due to the inconsistency of the label space, and the method GM-SC-LM-LPHD of the application has very obvious improvement in the target number and OSPA precision; therefore, compared with the traditional distributed GM-LPHD algorithm, the application can overcome the fusion quality decline problem caused by the inconsistency of the label space of the distributed sensor, and can track multiple targets in real time and accurately.

Claims

1. A distributed GM-LPHD multi-target tracking method based on SC-LM, characterized in that... Includes the following steps: Step (1): Construct a distributed multi-sensor multi-target tracking scenario and set relevant parameters, including the process noise of target motion and the observation noise of sensors; Step (2): Model the state and observation of the target; Step (3): Modeling a single-sensor GM-LPHD filter; Step (4): Based on steps (2)-(3), using a distributed fusion architecture, perform SC-LM processing on the posterior Gaussian sets of all sensors before fusion; specifically as follows: 4-1 in At any given time, based on steps (2)-(3), the sensor can be determined. The posterior Gaussian mixture set of the label is obtained after GM-LPHD filtering. ,in , , They represent Time sensor The first result obtained after filtering in step (3) The weights, state mean, and covariance of each GM-LPHD component. Indicates sensor The number of GM-LPHD components obtained after step (3); 4-2 State Clustering Assuming any two sensors , The posterior Gaussian sets of the labels are respectively , The sensor is calculated according to equation (12). , The Middle The distance between the j-th and j-th GM-LPHD components : ； in Represents the 2-norm; The sensor is obtained according to equation (12). With sensors Cost matrix between as follows: ； 4-3 tag matching The cost matrix obtained after state clustering in step 4-2 The sensor is obtained through equation (14). and sensors Matching relationships between posterior states: ； in Indicates sensor Gaussian components in sensors Matching relationship of Gaussian components. mean The Middle Each label Gaussian component and The Middle Each label Gaussian component is matched in state; For all matching pairs When the condition of equation (15) is met, their labels are paired: ； Where c1 is the threshold; Complete the matching of the label Gaussian components and The tags should be processed as follows: ； in The label represents the fused Gaussian component; Therefore, according to formula (16), the label Gaussian component and Updated to and ; 4-4 Distributed Geometric Mean Fusion sensor With sensors After the posterior Gaussian components of the labels are processed by SC-LM, geometric mean GA fusion is performed, as shown in the following formula: ； ； ； ； ； in , This represents the fusion coefficient of GA fusion. This represents the th after SC-LM processing. With the The fusion weights of the posterior Gaussian components of each label This represents the th after SC-LM processing. With the Fusion covariance of posterior Gaussian components of each label; This represents the th after SC-LM processing. With the The fusion mean of the posterior Gaussian components of each label; 4-5 Determine whether the number of fusions for the distributed geometric mean fusion has reached the consistency fusion number requirement. If not, jump to step 4-2 to continue execution. If it has, output the filtering result. Step (5): Use the constructed GM-SC-LM-LPHD filter to achieve multi-target tracking in a distributed framework.

2. The method according to claim 1, characterized in that... Step (1) is as follows: The target moves in a two-dimensional plane, and its state is: ,in The target is in The position of direction, The target is in velocity in direction, Indicates transpose; The noise covariance of the target motion process is set as follows: ,in express The noise variance of the environment on the target in the X direction at any given time. express The noise variance of the environment at any given time in the Y direction relative to the target; Set the observation noise covariance of the sensor to be... ,in , These represent the noise variance of the sensor for the target in the X and Y directions, respectively.

3. The method according to claim 2, characterized in that... Step (2) specifically involves: 2-1 Target State Modeling Use a label to create a random finite set To represent multiple targets in The state at any given moment, express The number of targets at any given time, of which , express The label space in which the target is located at any given moment. express Time of the first The status of each target; The goal is from A certain state at a certain moment Transferred to A certain state at a given moment This is called state transition, and we assume the state transition probability density for each objective. It conforms to a linear Gaussian model, expressed as: ； in Indicates the target state The spatial distribution of follows a state mean of . The covariance is Gaussian distribution, Here is the state transition matrix. Let k be the process noise covariance matrix at time k-1; 2-2 Observation Modeling Assume there are multiple independent sensors in the tracking scenario; if the first... One goal is Constantly by sensors It was detected, and its observation equation is expressed as: ； in, Indicates sensor exist Time for the first The observation vector of each target; It is the observation matrix; express Time of the first The status of each target; Indicates sensor exist The observation noise covariance at any given time; exist At any moment, the sensor In addition to acquiring measurements of the target, the sensor also acquires environmental clutter; exist The environmental clutter obtained at time t is represented by equation (3): ； in, The parameter is The probability density function of the Poisson distribution; express The expected number of clutter items at any given time; Indicates clutter The probability density function in the observation space. The volume of the observation space; Therefore, in At any moment, the sensor The observation data is represented as Among them, sensors exist The first time monitored Individual measurement value For target or clutter.

4. The method according to claim 3, characterized in that... Step (3) is as follows: 3-1 Single-sensor GM-LPHD prediction Assumption The posterior Gaussian mixture set at time t is ,in , , They represent Time of the first The weights, state mean, and covariance of each GM-LPHD component. express The number of posterior Gaussian components at time t, then at Time of the first The predictions for each GM-LPHD are represented by (4)-(6): ； ； ； In the formula, Indicates the first One GM-LPHD component in Prediction weights at different times; express The probability of the target surviving at any given moment; Indicates the first One GM-LPHD component in Prediction of the mean state at any given time; express The state transition matrix at time t; Indicates the first One GM-LPHD component in Covariance prediction at time points; 3-2 Single Sensor GM-LPHD Update Based on the GM-LPHD prediction steps, the following is obtained: The predicted LPHD components of the multi-target time step are further updated according to equations (7)-(11): ； ； ； ； ； in express Time of the first Update weights of each GM-LPHD component. express The probability that a target is detected by a sensor at any given moment. Indicates the intensity of clutter. express Predict the number of Gaussian components at any given time. express Time of the first The probability density function of the predicted target observations. For sensors Time-based observation matrix for Measurement noise covariance at time, express Time of the first State mean update for each GM-LPHD component express Time of the first Kalman gain for each target Indicates sensor pair The observed values, express Time of the first Covariance update of each GM-LPHD component It is an identity matrix.

5. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method according to any one of claims 1-4.

6. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-4.

Images