Multi-maneuvering target tracking method based on prediction update network and dynamic wave gate

Abstract

The application discloses a multi-maneuvering target tracking method based on a prediction updating network and a dynamic wave gate, and realizes the following steps: training a prediction network and an updating network by using a training set of the prediction network and the updating network; obtaining a prediction position vector of a target to be tracked by using the prediction network; obtaining an effective measurement of the target to be tracked by using the dynamic wave gate and judging whether the target to be tracked appears maneuvering or not; assigning a correlation measurement vector to the target to be tracked according to a global nearest neighbor (GNN) algorithm; and obtaining an updating position vector of the target to be tracked by using the trained updating network. The application combines the prediction updating network and the dynamic wave gate, gets rid of the limitation of preset models and transition probabilities between models, realizes self-adaptive tracking of multi-maneuvering targets, and improves tracking precision.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar, and more particularly to the technical field of multi-maneuvering target tracking based on a prediction update network and dynamic wave gates. The present application can be used for real-time dynamic adjustment of a motion model and high-precision target tracking when measuring data observed by a radar is used for multi-maneuvering target tracking. BACKGROUND

[0002] The main task of maneuvering target tracking is to accurately and reliably detect a maneuvering target from radar point track measurements obtained in the presence of various noises and clutters. With the rapid development of modern science and technology, a target may change its flight speed, turning angular speed, acceleration and other motion parameters at any time in order to avoid tracking or face the complex and changeable environment in low-altitude airspace. The difficulty of maneuvering target tracking lies in the complexity of different environments and targets, the type and time of target maneuvering being difficult to predict, and it is very difficult to establish an accurate target motion model in advance to reflect the real target motion. At present, there are a large number of research methods based on improved maneuvering target system models and using Kalman gain or innovation to detect maneuvering. These methods have the following problems: since they rely on preset models and transition probabilities between preset models, and it is difficult to obtain sufficient information in advance to define the models and rely on preset models and transition probabilities between preset models, they cannot timely and accurately adaptively track the maneuvering motion, affecting the tracking efficiency of the maneuvering target; due to the high dynamics of the maneuvering target motion and the uncertainty of the measurement noise, there are cases of model mismatch and inaccurate noise estimation, affecting the tracking accuracy of the maneuvering target.

[0003] Jingxian Liu et al. in their published paper "A deep learning maneuvering target-tracking algorithm based on bidirectional LSTM network" (Information Fusion 53 2020 289-304) disclose a deep learning maneuvering target tracking method based on bidirectional LSTM network. The specific steps of the method are, (1) constructing a LAST database to provide offline trajectory data for network training; (2) designing a DeepMTT network composed of three bidirectional long short-term memory layers, a filtering layer, a maxout layer and a linear output layer; (3) segmenting the trajectory, then tracking and predicting through the DeepMTT network, and finally fusing the processing results of each trajectory segment according to a certain proportion. The deficiency of this method is that when tracking the trajectory of a maneuvering target, the target needs to be segmented and predicted, and for multiple maneuvering trajectories, only a segment of the trajectory of a target can be processed each time, and multiple maneuvering target trajectories cannot be directly input into the network for tracking and prediction, affecting the tracking efficiency of the maneuvering target.

[0004] Xi'an University of Electronic Science and Technology in its applied patent document "Multi-maneuvering target tracking method based on LSTM network" (Patent application number 201911057969.1, authorized announcement number CN 110780290 B) discloses a maneuvering target tracking method applying LSTM network. The specific steps of the method are implemented as, (1) establishing a long short-term memory network LSTM; (2) generating a training data set; (3) training the long short-term memory network LSTM; (4) using the long short-term memory network LSTM for multi-maneuvering target resource allocation; (5) using Kalman filtering algorithm for multi-maneuvering target tracking. The deficiency of this method is that the predicted value and the observed value allocated through the LSTM network are directly used as the input of the Kalman filtering tracking algorithm, and due to the influence of clutter, there is a certain amount of error association data between the observed value and the predicted value, which causes the Kalman filtering algorithm to receive more invalid measurements, affecting the tracking accuracy of the maneuvering target. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies of the prior art, and to provide a multi-maneuvering target tracking method based on prediction update network and dynamic wave gate, aiming to solve the problem of being unable to accurately estimate the motion mode of multiple maneuvering targets and the high error rate of maneuvering target motion state prediction.

[0006] The idea for achieving the object of the present application is to construct a prediction network and an update network with the same structure and different parameter settings respectively. Since the input of the trained prediction network is a three-dimensional tensor composed of the number of to-be-tracked targets, the input time step and the state vector dimension of the to-be-tracked targets, the prediction network can simultaneously process the input position vectors from multiple to-be-tracked targets and learn the prediction position vectors of the to-be-tracked targets from various motion modes in the training set of the prediction network, thereby avoiding the problem of low tracking efficiency caused by the need for segmented tracking of maneuvering targets in the prior art. The present application uses the trained prediction network to obtain the prediction position vector of the to-be-tracked target, uses the innovation covariance and the survival weight to set the statistical distance between the prediction position vector and the measurement vector of the to-be-tracked target, sets the initial gate threshold and the adaptive gate threshold, and determines the size of the dynamic gate; uses the dynamic gate to obtain the effective measurement of the to-be-tracked target and judges whether the to-be-tracked target has maneuvered; and uses the trained update network to learn the noise characteristics from the effective measurement to update the position vector of the to-be-tracked target, thereby solving the problem of low tracking accuracy of maneuvering targets caused by excessive invalid measurements in the prior art.

[0007] The specific steps of the present application are as follows:

[0008] Step 1: build a prediction network composed of an input layer, a hidden layer, an output layer and a fully connected layer connected in series; set the input dimension of the input layer to 2, the depth to 3, and the input time step to 20, set the dimension of the hidden layer to 64, set the input dimension of the fully connected layer to 64, and set the output dimension to 2;

[0009] Step 2: build an update network with the same structure as the prediction network; set the input dimension of the input layer to 2, the depth to 4, and the input time step to 10, set the dimension of the hidden layer to 256, set the input dimension of the fully connected layer to 256, and set the output dimension to 2;

[0010] Step 3: generate the state vector of each moving target:

[0011] In the low-altitude plane scene with a flight height of 1000 meters or less, randomly select targets moving in constant speed CV and constant rotation speed CT, and select at least 256 moving targets from each mode to form a sample set, with a sampling interval T s = 1, uniformly sample the tracking duration of the moving targets in the sample set; use the state transition equation to obtain the state vector of each moving target at all sampling times; and combine the x-axis coordinate value and the y-axis coordinate value in the state vector of each moving target to form the position sequence of the moving target.

[0012] Step 4: generate the training set of the prediction network and the update network respectively:

[0013] Step 4.1, divide the position sequence of each moving target according to the sampling time, divide each position sequence into two subsequences of adjacent sampling time, and form a training sample of the moving target, taking the former subsequence as the input sequence and the latter subsequence as the label sequence, and form a training set of the prediction network by taking all training samples of the moving target;

[0014] Step 4.2, obtain the measurement sequence of each moving target at all sampling times by using the observation equation, divide the measurement sequence and the position sequence of each moving target according to the sampling time, form a training sample of the moving target by taking all position subsequences and measurement subsequences at the same sampling time, take the measurement subsequence as the input sequence and the position subsequence as the label sequence, and form a training set of the update network by taking all training samples of the moving target;

[0015] Step 5, training the prediction network and the update network:

[0016] The training set of the prediction network and the training set of the update network are input into the prediction network and the update network respectively, the mean square error MSE is used as the loss function, the Adam optimization algorithm and the cosine annealing algorithm are used to update the parameters in the network, and the training of the prediction network and the update network is completed until the loss functions of the prediction network and the update network converge respectively, and the trained prediction network and the update network are obtained;

[0017] Step 6, obtaining the predicted position vector of the to-be-tracked target by using the prediction network:

[0018] The target observed by the radar in two consecutive scanning periods is taken as the to-be-tracked target, the position sequence of each to-be-tracked target is normalized and input into the trained prediction network, and the last vector of the output position sequence of the prediction network is taken as the predicted position vector of the to-be-tracked target at the time corresponding to the vector;

[0019] Step 7, obtaining the effective measurement of the to-be-tracked target by using the dynamic wave gate:

[0020] Calculating the initial wave gate threshold γ d , the statistical distance between the predicted position vector of the to-be-tracked target and the measurement vector Adaptive wave gate threshold Determining the size of the dynamic wave gate; comparing the size relationship of γ d , and , the measurement vector of or is taken as the effective measurement, and the to-be-tracked target of is taken as the maneuvering target occurring at the kth sampling time; the accumulated predicted position vector and the measurement vector of the maneuvering target are cleared, and the survival weight of the maneuvering target is reduced to 0.8;

[0021] Step 8, obtaining the updated position vector of the to-be-tracked target by using the update network:

[0022] Step 8.1, calculating the Euclidean distance between the predicted position vector of each to-be-tracked target at the current sampling time and each effective measurement, taking the Euclidean distance as the cost of the association between the to-be-tracked target and each effective measurement, using the global nearest neighbor (GNN) algorithm for data association for each effective measurement and the to-be-tracked target corresponding to the effective measurement, and assigning an associated measurement vector to each to-be-tracked target with an effective measurement;

[0023] Step 8.2, grouping all the associated measurement vectors of each to-be-tracked target into a measurement sequence, inputting the normalized measurement sequence of each to-be-tracked target into the trained update network respectively, and taking the last vector in each sequence output by the update network as the updated position vector of each to-be-tracked target at the current sampling time;

[0024] Step 8.3, determining whether the current tracking sampling time is the last sampling time, if yes, executing Step 9, otherwise, executing Step 6;

[0025] Step 9, outputting the tracking trajectory of the to-be-tracked target:

[0026] According to each updated position vector, determining the position of the to-be-tracked target at each sampling time, and connecting the positions of each to-be-tracked target at all sampling times in sequence to obtain the tracking trajectory of each to-be-tracked target.

[0027] Compared with the prior art, the present application has the following advantages:

[0028] First, since the present application respectively constructs a prediction network and an update network with the same structure and different parameter settings, the trained prediction network is used to learn the motion mode from the historical position vector of the to-be-tracked target to obtain the predicted position vector, and the trained update network is used to obtain the updated position vector of the to-be-tracked target, overcoming the deficiency of the prior art that needs to rely on a preset model and the transition probability between the preset models to track the maneuvering target, so that the present application realizes adaptive detection and tracking of the motion mode of multiple maneuvering targets.

[0029] Second, since the present application uses a dynamic gate to obtain the effective measurement of the to-be-tracked target, and uses the predicted state vector and the measurement vector to determine whether there is a measurement vector within the dynamic gate of each to-be-tracked target, for the target without a measurement vector within the dynamic gate, the gate threshold is dynamically adjusted and it is determined whether the to-be-tracked target has maneuvered, overcoming the defect of the prior art that the prediction error rate of the current state of the maneuvering target is high, so that the present application improves the tracking accuracy of multiple maneuvering targets. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 Flow chart of the present application;

[0031] Figure 2 Dynamic wave gate schematic diagram of the present application;

[0032] Figure 3 Simulation diagram of the present application. DETAILED DESCRIPTION

[0033] The present application will be further described in conjunction with the accompanying drawings and examples.

[0034] Referring to Figure 1 , the implementation steps of the embodiments of the present application will be further described.

[0035] Step 1, build a prediction network composed of an input layer, a hidden layer, an output layer and a full connection layer in series; set the input dimension of the input layer to 2, the depth to 3, and the input time step to 20, set the dimension of the hidden layer to 64, set the input dimension of the full connection layer to 64, and set the output dimension to 2.

[0036] Step 2, build an update network with the same structure as the prediction network; set the input dimension of the input layer to 2, the depth to 4, and the input time step to 10, set the dimension of the hidden layer to 256, set the input dimension of the full connection layer to 256, and set the output dimension to 2.

[0037] Step 3, generate the state vector of each moving target.

[0038] In the low-altitude plane scene with a flight height of 1000 meters, randomly select targets moving in constant speed CV and constant rotation speed CT, and select at least 256 moving targets from each mode to form a sample set, with a sampling interval T s = 1, uniformly sample the tracking duration of the moving targets in the sample set; use the state transition equation to obtain the state vector of each moving target at all sampling times; and combine the x-axis coordinate value and the y-axis coordinate value in the state vector of each moving target to form the position sequence of the moving target.

[0039] The tracking duration refers to the time length from the first appearance to the disappearance of each moving target in the radar scanning tracking scene diagram.

[0040] The state transition equation is as follows:

[0041]

[0042] x qx x x x

[0043] The state vector is composed of five parameters of the x-axis coordinate value of the moving target, the movement speed in the x-axis direction, the y-axis coordinate value, the movement speed in the y-axis direction, and the rotation angular velocity.

[0044] Step 4, generate the training set of the prediction network and the update network respectively.

[0045] Step 4.1, divide the position sequence of each moving target according to the sampling time, and form the training sample of the moving target by two sub-sequences of adjacent sampling times after dividing each position sequence. The former sub-sequence is taken as the input sequence, and the latter sub-sequence is taken as the label sequence. All training samples of the moving target are used to form the training set of the prediction network.

[0046] Step 4.2, obtain the measurement sequence of each moving target at all sampling times by using the observation equation, and divide the measurement sequence and the position sequence of each moving target according to the sampling time. All position sub-sequences and measurement sub-sequences of the same sampling time are used to form the training sample of the moving target. The measurement sub-sequence is taken as the input sequence, and the position sub-sequence is taken as the label sequence. All training samples of the moving target are used to form the training set of the update network.

[0047] The observation equation is as follows:

[0048]

[0049] wherein z q x x x

[0050] Step 5, train the prediction network and the update network.

[0051] The training sets of the prediction network and the update network are input into the prediction network and update network respectively. The mean squared error (MSE) is used as the loss function. The parameters in the network are iteratively updated using the Adam optimization algorithm and the cosine annealing algorithm until the loss functions of the prediction network and the update network converge, respectively, thus obtaining the trained prediction network and update network.

[0052] The loss function of the prediction network is as follows:

[0053]

[0054] Where L represents the loss function of the prediction network, and M represents the total number of training samples in the training set of the prediction network. This indicates the i-th term in the training set of the prediction network. u The total number of vectors in the input sequence of each training sample. This indicates the i-th term in the training set of the prediction network. u The r-th vector in the label sequence of training samples, This indicates the i-th term in the training set of the prediction network. u The input sequence of training samples is the r-th vector in the output sequence of the prediction network, where ||·||2 represents the L2 norm operation.

[0055] The loss function of the updated network is as follows:

[0056]

[0057] Where L' represents the loss function for updating the network, and M' represents the total number of training samples in the training set for updating the network. This indicates that the i-th element in the training set of the updated network is... v The total number of vectors in the input sequence of each training sample. This indicates that the i-th element in the training set of the updated network is... v The s-th vector in the label sequence of training samples, This indicates that the i-th element in the training set of the updated network is... v The input sequence of training samples is used to update the s-th vector in the network output sequence.

[0058] Step 6: Use the prediction network to obtain the predicted position vector of the target to be tracked.

[0059] The targets observed by the radar within two consecutive scanning cycles are taken as the targets to be tracked; the position sequence of each target to be tracked is normalized and then input into the trained prediction network. The last vector of the position sequence output by the prediction network is taken as the predicted position vector of the target to be tracked at the corresponding time.

[0060] Step 7, get the effective measurement of the target to be tracked by using the dynamic gate.

[0061] Calculate the initial gate threshold γ d , the statistical distance between the predicted position vector and the measurement vector of the target to be tracked Adaptive gate threshold Determine the size of the dynamic gate; compare γ d , And The size relationship, the measurement vector of Or is recorded as the effective measurement, and the target to be tracked of is recorded as the maneuvering target that occurs at the kth sampling time; the accumulated predicted position vector and measurement vector of the maneuvering target are cleared, and the survival weight of the maneuvering target is reduced to 0.8.

[0062] The dynamic gate refers to taking the initial gate threshold or the adaptive gate threshold of the dynamic gate of the target to be tracked as the threshold, taking the position of the predicted position vector as the center point, and being used to determine the region where the measurement vector of the target to be tracked may appear at each sampling time.

[0063] The initial gate threshold γ d is obtained by the following formula:

[0064] γ d = -2ln(1-P d )

[0065] Wherein, γ d represents the initial threshold of the gate of the target to be tracked, P d represents the initial detection probability of the dynamic gate.

[0066] The statistical distance between the predicted position vector and the measurement vector of the target to be tracked is obtained by the following formula:

[0067]

[0068] Wherein, represents the statistical distance between the predicted position vector and the measurement vector of the qth target to be tracked at the kth sampling time, represents the innovation covariance of the qth target to be tracked at the kth sampling time, and the superscript T represents the transposition operation.

[0069] The adaptive gate threshold is obtained by the following formula:

[0070]

[0071] wherein, represents the adaptive gate threshold value of the qth to-be-tracked target at the kth sampling moment, represents the total length of the duration of the qth moving target with a survival weight of 1 between the 1st sampling moment and the k-1th sampling moment.

[0072] Referring to Figure 2 , the region size of the determined dynamic gate is further described.

[0073] Figure 2 The dashed line in the figure represents the predicted trajectory of the qth to-be-tracked target at the kth sampling moment after the predicted network, the solid curve represents the actual motion trajectory of the qth to-be-tracked target at the kth sampling moment, the solid line circle represents the dynamic gate region with the adaptive gate threshold value , the dashed line ellipse represents the dynamic gate region with the initial gate threshold value γ d . The solid circle represents the center point of the dynamic gate of the qth to-be-tracked target at each sampling moment, the solid square represents the intersection point of the dynamic gate of the qth to-be-tracked target at the k-1th sampling moment and the kth sampling moment, and the solid triangle represents the measurement vector position of the qth to-be-tracked target at the kth sampling moment. represents the predicted position vector of the qth to-be-tracked target at the kth sampling moment after the predicted network, represents a measurement vector falling within the dynamic gate of the to-be-tracked target q at the kth sampling moment, represents a measurement vector falling outside the dynamic gate of the to-be-tracked target q at the kth sampling moment, represents a measurement vector falling within the dynamic gate of the to-be-tracked target q at the k-1th sampling moment. Q1 represents the dynamic gate region of the qth to-be-tracked target at the k-1th sampling moment, and Q2 represents the dynamic gate region of the qth to-be-tracked target at the kth sampling moment.

[0074] In the embodiment of the present application, when the region where the to-be-tracked target is located is a highly fixed two-dimensional plane, P d is set to 0.98, and the initial gate threshold value γ d calculated thereby is 7.824.

[0075] In the embodiment of the present application, at the k-1th sampling moment, the statistical distance between the predicted position vector of the qth to-be-tracked target and the measurement vector Since , the initial gate threshold value γ d, the dynamic gate region of the qth target to be tracked at the k-1th sampling moment is generated, and the dynamic gate region is recorded as an ellipse Q1. At the kth sampling moment, the predicted position vector of the qth target to be tracked is calculated , and the statistical distance between the measurement vector and the predicted position vector is calculated. The adaptive gate threshold of the qth target to be tracked is calculated According to the adaptive gate threshold, the dynamic gate region of the qth target to be tracked at the kth sampling moment is generated, and the dynamic gate region is recorded as an ellipse Q2. Because the statistical distance between the measurement vector and the predicted position vector is calculated. , the measurement vector is recorded as the measurement vector falling within the dynamic gate region Q2, and the measurement vector falling within the dynamic gate region is recorded as the valid measurement of the qth target to be tracked at the kth sampling moment.

[0076] Step 8, the updated position vector of the target to be tracked is obtained by using the update network.

[0077] Step 8.1, the Euclidean distance between the predicted position vector of each target to be tracked at the current sampling moment and each valid measurement is calculated, the Euclidean distance is taken as the cost of the association between the target to be tracked and each valid measurement, the global nearest neighbor (GNN) algorithm is used for data association for each valid measurement and the target to be tracked corresponding to the valid measurement at each sampling moment of each target to be tracked, and an associated measurement vector is assigned to each target to be tracked with a valid measurement.

[0078] The Euclidean distance between the predicted position vector of each target to be tracked at the current sampling moment and each valid measurement is obtained by the following formula:

[0079]

[0080] wherein, represents the Euclidean distance between the predicted position vector of the qth target to be tracked at the kth sampling moment and each measurement vector, and D represents the dimension of the predicted position vector and the measurement vector, represents the dth dimension value in the predicted position vector

[0081] represents the dth dimension value in the measurement vector .

[0082] ​​​Step 8.2, all associated measurement vectors of each to-be-tracked target are combined into a measurement sequence, and each normalized measurement sequence of each to-be-tracked target is input into the trained updating network, and the last vector in each sequence output by the updating network is taken as the updating position vector of each to-be-tracked target at the current sampling moment.

[0083] Step 8.3, it is judged whether the current tracking sampling moment is the last sampling moment, if yes, step 9 is executed, otherwise, step 6 is executed.

[0084] Step 9, the tracking trajectory of the to-be-tracked target is output.

[0085] According to each updating position vector, the position of each to-be-tracked target at each sampling moment is determined, and the positions of each to-be-tracked target at all sampling moments are sequentially connected to obtain the tracking trajectory of each to-be-tracked target.

[0086] The effect of the present application is further illustrated in combination with a simulation experiment.

[0087] 1. Simulation experiment conditions:

[0088] The hardware platform of the simulation experiment of the present application is: the processor is Intel i7 13700H CPU, the main frequency is 2.4 GHz, and the memory is 16 GB.

[0089] The software platform of the simulation experiment of the present application is: Windows 10 operating system and Python 3.8.

[0090] 2. Simulation content and result analysis:

[0091] In the simulation experiment of the present application, the radar and the to-be-tracked target are in rectangular coordinates, the radar is located at [0, 0, 0] m, the target is located at a fixed height and below 1000 meters, and the range is [1200, 1200] m x [1200, 1200] m.

[0092] There are three to-be-tracked targets in the simulation scene, to-be-tracked target 1 makes CV motion from 1st to 10th second, CT motion with different rotation angular velocities from 11th to 20th second and from 21st to 30th second. To-be-tracked target 2 makes CV motion from 20th to 35th second, CT motion with different rotation angular velocities from 36th to 49th second and from 50th to 60th second. To-be-tracked target 3 makes CV motion from 40th to 50th second, and CT motion from 51st to 60th second.

[0093] In the simulation experiment of the present application, the radar continuously observes the to-be-tracked target for 100 sampling moments, the sampling interval T s = 1, the covariance of the measurement noise in the scene is set to The mean value of the clutter number is set to λc = 10, the detection probability of the target is set as P d = 0.98, the survival probability of the target is set as p s = 0.99.

[0094] The simulation experiment of the application is performed on the three tracking targets described above at 60 consecutive sampling times by using the method of the application and one prior art, and the actual motion trajectories of the three tracking targets in the simulation experiment are plotted into Figure 3 (a) shown, the predicted trajectories of the three tracking targets tracked by using the method of the application are plotted into Figure 3 (b) shown.

[0095] In order to verify the simulation effect of the application, the comprehensive error between the number and position of the tracking targets estimated by the two methods and the actual motion values is calculated by using the OSPA error formula, and Figure 3 (c) is obtained.

[0096] In the simulation experiment, the prior art used refers to:

[0097] The application discloses a method for tracking a mobile target by using an LSTM network in the patent literature "Multi-mobile target tracking method based on an LSTM network" (patent application number 201911057969.1, authorized announcement number CN 110780290 B) applied by Xi'an University of Electronic Science and Technology.

[0098] The effect of the application will be further described below in combination with the simulation diagram of Figure 3 .

[0099] Figure 3 (a) is the real trajectory diagram of the three tracking targets in the simulation experiment of the application, the horizontal coordinate represents the length of the simulation scene area, the unit is m, and the vertical coordinate represents the width of the simulation scene area, the unit is m, Figure 3 The three curves in (a) represent the actual motion trajectories of the three tracking targets.

[0100] Figure 3 (b) is the predicted trajectory diagram of the three tracking targets tracked by using the method of the application, the horizontal and vertical coordinates are the same as those set in Figure 3 (a), Figure 3 The curve marked with a circle in (b) represents the predicted trajectory of each tracking target at each time obtained by using the method proposed in the application.

[0101] It can be seen from Figure 3 (b) that the method proposed in the application can accurately estimate the target motion trajectory.

[0102] Figure 3(c) is a graph of OSPA error obtained by the method of the present application and a prior art, the abscissa represents the sampling time, in s, and the ordinate represents the OSPA error, in m. The curves marked with circles and squares represent the OSPA error curves at each time obtained by the method proposed in the present application and the prior art, respectively.

[0103] From ​ (c) It can be seen that the OSPA error obtained by the method of the present application has certain fluctuations at the 10th, 20th and 30th sampling times when the target to be tracked appears to be born and die; except for the time when the number of targets to be tracked changes, the OSPA error obtained by the method of the present application is significantly lower than that of the prior art method, and at the 10th, 20th, 30th and 35th sampling times when the maneuvering phenomenon appears, the OSPA errors of the present application and the prior art method do not appear large fluctuations, and pass through the maneuvering stage smoothly. The method of the present application uses a prediction update network and a dynamic wave gate to track multiple maneuvering targets, and has high tracking accuracy.

Claims

1. A multi-maneuvering target tracking method based on a prediction update network and dynamic wave gates, characterized in that, The prediction network and the update network with the same structure and different parameter settings are constructed, and the effective measurement of the target to be tracked is obtained by using the dynamic wave gate; the specific steps of the target tracking method are as follows: Step 1, a prediction network composed of an input layer, a hidden layer, an output layer and a fully connected layer is built; the input dimension of the input layer is set to 2, the depth is set to 3, and the input time step is set to 20; the dimension of the hidden layer is set to 64; the input dimension of the fully connected layer is set to 64, and the output dimension is set to 2; Step 2, an update network with the same structure as the prediction network is built; the input dimension of the input layer is set to 2, the depth is set to 4, and the input time step is set to 10; the dimension of the hidden layer is set to 256; the input dimension of the fully connected layer is set to 256, and the output dimension is set to 2; Step 3, the state vector of each moving target is generated: In the low altitude plane scene within the flight height 1000 meters, randomly select constant velocity CV and constant rotation speed CT two ways of moving target, select at least 256 moving targets from each way to form a sample set, sample interval T s =1, uniformly sample the tracking duration of the moving targets in the sample set; The state vector of each moving target at all sampling times is obtained by using the state transition equation; the x-axis coordinate value and the y-axis coordinate value in the state vector of each moving target are combined to form the position sequence of the moving target; Step 4, the training set of the prediction network and the update network is generated respectively: Step 4.1, the position sequence of each moving target is divided according to the sampling time; after the division, the two sub-sequences of adjacent sampling times of each position sequence form the training sample of the moving target; the former sub-sequence is taken as the input sequence, and the latter sub-sequence is taken as the label sequence; all the training samples of the moving targets form the training set of the prediction network; Step 4.2, the measurement sequence of each moving target at all sampling times is obtained by using the observation equation; the measurement sequence and the position sequence of each moving target are divided according to the sampling time; all the position sub-sequences and the measurement sub-sequences of the same sampling time of each moving target form the training sample of the moving target; the measurement sub-sequence is taken as the input sequence, and the position sub-sequence is taken as the label sequence; all the training samples of the moving targets form the training set of the update network; Step 5, the prediction network and the update network are trained: The training set of the prediction network and the training set of the update network are input into the prediction network and the update network respectively; the mean square error MSE is used as the loss function; the Adam optimization algorithm and the cosine annealing algorithm are used to update the parameters in the network iteratively until the loss functions of the prediction network and the update network converge respectively; the trained prediction network and the update network are obtained respectively; Step 6, the prediction position vector of the target to be tracked is obtained by using the prediction network: The target observed by the radar in two consecutive scanning periods is taken as the target to be tracked; the position sequence of each target to be tracked is normalized and input into the trained prediction network; the last vector of the output position sequence of the prediction network is taken as the prediction position vector of the target to be tracked at the time corresponding to the vector; Step 7, the effective measurement of the target to be tracked is obtained by using the dynamic wave gate: Compute initial gate threshold γ d , the statistical distance between the predicted position vector of the target to be tracked and the measurement vector Adaptive gate threshold Determine the region size of dynamic gate; compare γ d , And The size relationship, the measurement vector of Or Is recorded as the effective measurement, and the target to be tracked of Is recorded as the maneuvering target that occurs maneuvering at the kth sampling time; the accumulated predicted position vector and measurement vector of the maneuvering target are emptied, and the survival weight of the maneuvering target is reduced to 0.8; Step 8, the update position vector of the target to be tracked is obtained by using the update network: Step 8.1, calculate the Euclidean distance between the predicted position vector of each to-be-tracked target at the current sampling time and each valid measurement, take the Euclidean distance as the cost of the association between the to-be-tracked target and each valid measurement, use the global nearest neighbor (GNN) algorithm for data association for each valid measurement and the to-be-tracked target corresponding to the valid measurement at each sampling time of each to-be-tracked target, and assign an associated measurement vector to each to-be-tracked target with a valid measurement; Step 8.2, group all associated measurement vectors of each to-be-tracked target into a measurement sequence, input the normalized measurement sequence of each to-be-tracked target into the trained update network respectively, and take the last vector in each sequence output by the update network as the updated position vector of each to-be-tracked target at the current sampling time; Step 8.3, determine whether the current tracking sampling time is the last sampling time, if yes, execute step 9, otherwise, execute step 6; Step 9, output the tracking trajectory of the to-be-tracked target: determine the position of the to-be-tracked target at each sampling time according to each updated position vector, and connect the positions of each to-be-tracked target at all sampling times in turn to obtain the tracking trajectory of each to-be-tracked target.

2. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The tracking duration in step 3 refers to the time length from the first appearance of each moving target in the radar scanning tracking scene graph to the disappearance.

3. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The state transition equation in step 3 is as follows: wherein x q represents the state vector of the qth moving target at all sampling time points, n represents the total number of sampling at all sampling time points of the moving target, k represents the serial number of the sampling number of the moving target, represents the state transition matrix of the qth moving target at the kth sampling time point, represents the state vector of the qth moving target transferred from the kth sampling time point to the k+1th sampling time point, represents the process noise vector of the qth moving target transferred from the kth sampling time point to the k+1th sampling time point.

4. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The state vector in step 3 is composed of five parameters: x-axis coordinate value, movement speed in x-axis direction, y-axis coordinate value, movement speed in y-axis direction, and angular velocity.

5. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 3, characterized in that, The observation equation in step 4.2 is as follows: where z q represents the measurement sequence of the qth moving target at all sampling times, represents the transition matrix of the qth moving target mapped from the state space to the observation space at the kth sampling time, represents the measurement noise vector of the qth moving target at the kth sampling time.

6. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The loss function of the prediction network in step 5 is as follows: Where L represents the loss function of the prediction network, and M represents the total number of training samples in the training set of the prediction network. This indicates the i-th term in the training set of the prediction network. u The total number of vectors in the input sequence of each training sample. This indicates the i-th term in the training set of the prediction network. u The r-th vector in the label sequence of training samples, This indicates the i-th term in the training set of the prediction network. u The input sequence of training samples is the r-th vector in the output sequence of the prediction network, where ||·||2 represents the L2 norm operation.

7. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The loss function of the update network in step 5 is as follows: Where L' represents the loss function for updating the network, and M' represents the total number of training samples in the training set for updating the network. This indicates that the i-th element in the training set of the updated network is... v The total number of vectors in the input sequence of each training sample. This indicates that the i-th element in the training set of the updated network is... v The s-th vector in the label sequence of training samples, This indicates that the i-th element in the training set of the updated network is... v The input sequence of training samples is used to update the s-th vector in the network output sequence.

8. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The dynamic wave gate in step 7 refers to taking the initial wave gate threshold or the adaptive wave gate threshold of the to-be-tracked target dynamic wave gate as the threshold, and taking the position of the predicted position vector as the center point to determine the region where the measurement vector of the to-be-tracked target may appear at each sampling time.

9. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The initial wave gate threshold in step 7 is obtained by the following formula: gamma d = -2 ln(1 - P d ) where γ d represents the initial threshold of the gate for the target to be tracked, P d represents the initial detection probability of the dynamic gate.

10. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, the statistical distance between the predicted position vector of the target to be tracked and the measurement vector in step 7 is obtained from the following equation: wherein, represents the predicted position vector of the qth target to be tracked at the kth sampling time represents the statistical distance between the measurement vector represents the innovation covariance of the qth target to be tracked at the kth sampling time, and the superscript T represents the transposition operation.​ 11. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 1, characterized in that, The adaptive wave gate threshold described in step 7 is obtained from the following equation: wherein, represents the adaptive gate threshold value of the qth target to be tracked at the kth sampling moment, represents the total length of the duration in which the qth moving target has a survival weight of 1 between the 1st sampling moment and the k-1th sampling moment.

12. The multi-maneuvering target tracking method based on prediction update network and dynamic wave gate according to claim 10, characterized in that, The Euclidean distance between the predicted position vector of each to-be-tracked target at the current sampling time and each valid measurement in step 8.1 is obtained by the following formula: wherein, represents the Euclidean distance between the predicted position vector of the qth target to be tracked at the kth sampling time and each measurement vector, D represents the dimension of the predicted position vector and the measurement vector, represents the dth dimension value in the predicted position vector , represents the dth dimension value in the measurement vector .

Images