Multi-target adaptive tracking method based on prediction network and update network

Abstract

A multi-target adaptive tracking method based on a prediction network and an update network, the implementation steps are as follows: using the generated training set of the prediction network and the update network, training the prediction network for learning the motion mode from the target historical position vector to predict the target position vector, and training the update network for updating the target position vector by learning the noise characteristics from the measurement. The trained prediction network is used to predict the position vector of the target; the global nearest neighbor GNN algorithm is used to assign the corresponding measurement to the target to be tracked; the trained update network is used to update the target position vector from the measurement. The present application predicts and updates the position of the target by two networks, which gets rid of the limitation of model and noise prior, so that the present application can adaptively process multiple motion modes while improving the tracking accuracy when tracking multiple targets.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar, and more particularly to the technical field of radar backend data processing, and relates to a multi-target adaptive tracking method based on a prediction network and an update network. BACKGROUND

[0002] The main task of multi-target tracking is to estimate the number and position of targets from a series of measurement data in a scene, so as to obtain the trajectory of each target. With the complication of radar application scenarios, multi-target tracking mainly faces the following challenges: the number of targets changes over time due to birth and death; the uncertainty of measurement noise and target motion mode reduces tracking accuracy; data association, detection uncertainty and false alarm also bring obstacles. At present, there are a large number of methods for multi-target tracking using random finite set (RFS). However, since these methods rely on motion models and measurement models when estimating target states, in unknown scenarios, due to the high dynamics of target motion and the uncertainty of measurement noise, there are cases of model mismatch and inaccurate noise estimation.

[0003] Harbin Institute of Technology discloses a tracking method applying multipath effect and an unscented Kalman filter in its applied patent document "Low-altitude target tracking filter method under multipath effect" (patent application number CN202010671798, application publication number CN111708015A). The specific steps of the method are: (1) establishing a system model based on single-pulse ratio observation; (2) applying a segmented dichotomy method for filter initialization; (3) applying an unscented Kalman filter for filtering; (4) applying a static multi-model estimator to obtain target height estimation. The deficiency of the method is that the tracking needs to rely on a preset motion model to predict the target state vector, and when the target motion mode cannot be accurately estimated, the tracking performance will be severely degraded.

[0004] In his published paper "Deep recurrent neural network for multi-target filtering." (International Conference on Multimedia Modeling, 2019), Emambakhsh M studied a target tracking method based on recurrent neural network. The specific steps of this method are: (1) establish a set of target state tuples; (2) use LSTM (Long Short Term Meomry) network for target state prediction; (3) filter according to the filtering algorithm; (4) update the surviving target and allocate a new target. The disadvantage of this method is that the allocated measurement vector is used as the estimated state vector of the target, and the tracking accuracy is affected by the measurement noise. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a multi-target adaptive tracking method based on prediction network and update network, which aims to solve the problem of inaccurate estimation of target motion mode and limited tracking accuracy by measurement noise.

[0006] The idea of achieving the purpose of the present application is that the present application uses the deep long short term memory network DLSTM (Deep Long Short Term Meomry) to effectively utilize the historical information, learns the motion mode from the historical position of the target to be tracked to predict the position of the target at the next time, so it is not necessary to estimate and pre-model the target motion mode, and the multi-target is adaptively tracked. In addition, the DLSTM network is used to learn the statistical characteristics of the noise from the historical measurement sequence of the target to update the target position, which will not be limited by the noise prior, so as to realize high-precision tracking of the target.

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

[0008] Step 1, build a prediction network:

[0009] A prediction network composed of a deep long short term memory DLSTM subnetwork and a fully connected subnetwork FCN is built, the depth of the DLSTM subnetwork of the prediction network is set to 2, the input dimension is set to 2, the input time step is set to 10, and the hidden layer dimension is set to 64; the input dimension of FCN is set to 64, and the output dimension is set to 2;

[0010] Step 2, build an update network:

[0011] An update network identical to the prediction network is built; the depth of the DLSTM subnetwork in the update network is set to 3, the input dimension is set to 2, the input time step is set to 20, and the dimension of the hidden layer is set to 128; the input dimension of the FCN is set to 128, and the output dimension is set to 2;

[0012] Step 3, generate the training set of the prediction network:

[0013] Step 3.1, in the low-altitude plane scene of [-2000, 2000]m x [-2000, 2000]m, at least 128 targets moving in each motion mode are randomly selected, and the tracking time is uniformly sampled with the sampling interval of the period time of the radar scanning tracking scene to obtain multiple sampling time points;

[0014] Step 3.2, the state vector of each target is transferred from the previous sampling time point to the next sampling time point by using the state transition equation, and the state vector of the target at all sampling time points is obtained after m times of transfer, wherein the value of m is determined by the tracking time of each target to be tracked;

[0015] Step 3.3, the x-axis coordinate value and the y-axis coordinate value in the state vector of each target at each sampling time point are extracted to form a position vector, and the position vectors of all sampling time points of each target are combined to form the position sequence of the target;

[0016] Step 3.4, the position sequence of each target is divided into n-9 subsequences, and the two adjacent subsequences are combined to form a pair of training samples, the former subsequence is used as the input sequence, and the latter subsequence is used as the label sequence, and all training samples are combined to form the training set of the prediction network, wherein n represents the length of the position sequence, n = m + 1;

[0017] Step 4, generate the training set of the update network:

[0018] Step 4.1, the noisy measurement vector of each target at each sampling time point is calculated by using the observation equation, and the noisy measurement vectors of all sampling time points of each target are combined to form the measurement sequence of the target;

[0019] Step 4.2, the measurement sequence and the position sequence of each target are divided into n'-19 subsequences, the measurement subsequence and the position subsequence at the same starting point are combined to form a pair of training samples, the measurement subsequence is used as the input sequence, and the position subsequence is used as the label sequence, and all training samples are combined to form the training set of the update network, wherein n' represents the length of the measurement sequence, and the value is equal to n;

[0020] Step 5, train the prediction network:

[0021] Input the training set of the prediction network into the prediction network, and iteratively update the parameters in the prediction network using the Adam optimization algorithm and the simulated annealing algorithm until the loss function converges, to obtain the trained prediction network;

[0022] Step 6, training update network:

[0023] Input the training set of the update network into the update network, and iteratively update the parameters in the network using the Adam optimization algorithm and the simulated annealing algorithm until the loss function converges, to obtain the trained update network;

[0024] Step 7, using the prediction network and the update network to track multiple targets in the scene:

[0025] Step 7.1, the target observed continuously in the two scanning periods of the radar is taken as a target to be tracked;

[0026] Step 7.2, the position vector of each target to be tracked from the initial sampling time to the current sampling time is taken as a sequence, and the normalized position sequence of each target is input into the trained prediction network, and the last vector in the output sequence is taken as the predicted position vector of the target at the current sampling time;

[0027] Step 7.3, the Euclidean distance between the predicted position vector of each target to be tracked at the current sampling time and each noisy measurement vector is calculated as the cost of the association between the target and the measurement, and a global nearest neighbor GNN data association algorithm is used to assign an associated measurement vector to each target;

[0028] Step 7.4, all associated measurement vectors of each target to be tracked from the initial sampling time to the current sampling time are taken as a sequence, and the normalized measurement sequence of each target to be tracked is input into the trained update network, and the last vector in the output sequence is taken as the updated position vector of the target at the current sampling time;

[0029] Step 7.5, determine whether the current tracking sampling time is the last sampling time, if yes, execute step 8, otherwise, execute step 7.2;

[0030] Step 8, synthesize trajectory:

[0031] According to each updated position vector, the position of the target to be tracked at each time is determined, and the positions of each target to be tracked at all sampling times are connected to obtain the tracking trajectory of the target.

[0032] Compared with the prior art, the present application has the following advantages:

[0033] Firstly, the DLSTM sub-network and the fully connected network FCN (Fully Connected Network) are combined to form a prediction network, and the motion mode is learned from the historical position of the target through the network to predict the position of the target at the next moment, which overcomes the problem that the tracking needs to rely on the motion model to predict the target state in the prior art, and is driven by data and network, so that the present application can adaptively process various motion modes when tracking multiple targets.

[0034] Secondly, the DLSTM sub-network and the FCN are combined to form an updating network, and the statistical characteristics of the noise can be learned from the historical measurement of the target through the network to update the position of the target, which overcomes the problem that the noise statistical characteristics need to be relied on to accurately estimate the target position in the prior art, so that the present application can have higher precision when tracking multiple targets. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The flowchart of the present application;

[0036] Figure 2 The schematic diagram of the present application for dividing the position sub-sequence;

[0037] Figure 3 The simulation diagram of the present application. DETAILED DESCRIPTION

[0038] The present application will be further described below in combination with the drawings and examples.

[0039] Reference Figure 1 and examples, the implementation steps of the present application will be further described.

[0040] Step 1, constructing a prediction network.

[0041] Step 1.1, a prediction network composed of a DLSTM sub-network and a FCN in series is built, which is used to predict the position of the target at the next moment. The DLSTM sub-network is a network composed of multiple long short-term memory LSTM networks in the prior art. The parameters of the DLSTM sub-network are re-set as follows: the depth of the DLSTM sub-network is set to 2, the input dimension is set to 2, the input time step is set to 10, and the hidden layer dimension is set to 64. The FCN is a network in the prior art, and the parameters of the FCN network are re-set as follows: the input dimension of the FCN is set to 64, and the output dimension is set to 2.

[0042] Step 2, constructing an updating network.

[0043] Step 2.1, build an update network with the same structure as the prediction network, which is used to estimate the position of the target from the measurement. The parameter settings are as follows: the depth of the DLSTM subnetwork is set to 2, the input dimension is set to 2, the input time step is set to 20, and the dimension of the hidden layer is set to 128; the input dimension of the FCN is set to 128, and the output dimension is set to 2.

[0044] Step 3, generate the training set of the prediction network.

[0045] Step 3.1, randomly select 256 UAV targets in a two-dimensional plane tracking scene with a size of [-2000, 2000]m x [-2000, 2000]m. Since in the application scenario of radar multi-target tracking, targets are mostly moving in a constant velocity (CV) manner or a constant turn (CT) manner, in the embodiment of the present application, 128 targets are set to move in a CV manner and 128 targets are set to move in a CT manner. The tracking duration is set to 100s, and the radar scanning period T is assumed to be 1s. Uniformly sample the tracking duration at the radar scanning period interval to obtain 100 sampling time points. s

[0046] The state vector of the target moving in the CV manner is as follows:

[0047]

[0048] wherein, represents the state vector of the i-th target moving in the CV manner at the k-th sampling time point, represents the x-axis coordinate value of the i-th target moving in the CV manner at the k-th sampling time point, represents the moving speed of the i-th target moving in the CV manner along the x-axis direction, represents the y-axis coordinate value of the i-th target moving in the CV manner at the k-th sampling time point, represents the moving speed of the i-th target moving in the CV manner along the y-axis direction.

[0049] The state vector of the target moving in the CT manner is as follows:

[0050]

[0051] wherein, represents the state vector of the j-th target moving in the CT manner at the k-th sampling time point, represents the x-axis coordinate value of the j-th target moving in the CT manner at the k-th sampling time point, represents the moving speed of the j-th target moving in the CT manner along the x-axis direction at the k-th sampling time point,​ represents the y-axis coordinate value of the jth target moving in the CT mode at the kth sampling time, represents the moving speed of the jth target moving in the CT mode along the y-axis direction at the kth sampling time, Ω j represents the rotational angular velocity of the jth target moving in the CT mode.

[0052] Step 3.2, the state vector of each target is transferred from the kth sampling time to the kth sampling time by using the following state transition equation, and is transferred 100 times to obtain the state vector of each target at the 1st-100th sampling time:

[0053]

[0054] wherein, represents the state vector of the lth target at the kth sampling time, k = 1, 2,..., 100, represents the state transition matrix of the lth target at the k-1th sampling time, represents the state vector of the lth target at the k-1th sampling time, represents the noise vector of the lth target in the process of transferring from the k-1th sampling time to the kth sampling time, and the noise obeys a Gaussian distribution with a mean of 0 and a covariance of Q k-1 The state transition matrix for the target moving in the CV mode is The state transition matrix for the target moving in the CT mode is wherein sin(·) represents a sine operation, and cos(·) represents a cosine operation.

[0055] Step 3.3, the x-axis coordinate value and the y-axis coordinate value extracted from the state vector of each target at each sampling time constitute a position vector, and the position vectors of all sampling times of each target constitute a position sequence of the target;

[0056] Step 3.4, the position sequence is divided into sub-sequences, and the 1st-91st sampling time is taken as the starting point, and every 10 sampling times is taken as a division step, and each position sequence of the 256 targets moving in the two modes is divided into 91 sub-sequences.

[0057] In combination with Figure 2 , the process of dividing into sub-sequences is further described, Figure 2 wherein the horizontal coordinate represents the sampling time, and the vertical coordinate represents the serial number of all targets moving in the two modes. Figure 2Each small rectangle represents the position vector of each target at each time, and the rectangular frame represents the sub-sequence composed of the position vectors in the frame. The position sequence of each target is divided into 91 sub-sequences with the 1st-91st time as the starting point and every 10 time as a division step in a sliding form. The data of two adjacent sub-sequences is overlapped at the middle 9 times. Taking the 1st and 2nd sub-sequences of the 1st target as an example, the 1st sub-sequence contains the position vectors of the 1st target at the 1st-10th time, and the 2nd sub-sequence contains the position vectors of the 1st target at the 2nd-11th time. The data of the 1st and 2nd sub-sequences is overlapped at the 2nd-10th time.

[0058] The first sequence in the two adjacent sub-sequences after dividing the sequence is taken as the input sequence, and the second sequence is taken as the label sequence to form a pair of training samples. All the training samples are used to form the training set of the prediction network.

[0059] Step 4, generating the training set of the update network.

[0060] Step 4.1, using the following observation equation, the noisy measurement vector of each target at each time is calculated, and the noisy measurement vector of each target at all times is used to form the measurement sequence of the target:

[0061]

[0062] wherein, represents the noisy measurement vector of the lth target at the kth sampling time, H k represents the conversion matrix of the state vector from the state space to the observation space, represents the noise vector in the process of converting the state vector into the measurement vector of the lth target at the kth sampling time, which is subject to a Gaussian distribution with a mean of 0 and a covariance of R k .

[0063] Step 4.2, the measurement sequence and the position sequence of each target are divided into 81 sub-sequences respectively. The position sequence and the measurement sequence of each target are divided into 81 sub-sequences respectively with the 1st-81st time as the starting point and every 20 time as a division step in a sliding form. The measurement sub-sequence and the position sub-sequence at the same starting time are used to form a pair of training samples. 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 are used to form the training set of the update network.

[0064] Step 5, training the prediction network.

[0065] Each pair of training samples is input into the prediction network. The loss function value between the output sequence of the prediction network and the label sequence in the training samples is calculated. The parameters of the prediction network are iteratively updated using the Adam optimization algorithm and simulated annealing algorithm until the loss function converges, resulting in a trained prediction network. In the embodiment of this invention, the loss function converges after 10,000 iterations of network parameter updates.

[0066] The loss function L is as follows:

[0067]

[0068] Where M represents the total number of training samples in the training set of the prediction network. Indicates the i-th p The total number of vectors contained in the input sequence of each training sample. Indicates the i-th p The r-th vector in the label sequence of training samples, Indicates the i-th p The input sequence of training samples is the r-th vector in the output sequence of the prediction network, and ||·||2 represents the calculation of the L2 norm.

[0069] Step 6: Train and update the network.

[0070] Input the training set of the update network into the update network. The training method of the update network is the same as that in step 5.

[0071] The loss function L′ is as follows:

[0072]

[0073] Where M′ represents the total number of training samples in the training set for updating the network. Indicates the i-th u The total number of vectors contained in the input sequence of each training sample. Indicates the i-th u The r′-th vector in the label sequence of the training samples Indicates the i-th u The input sequence from the training samples is used to update the r′-th vector in the network output sequence.

[0074] Step 7: Use the prediction network and the update network to track multiple targets in the scene.

[0075] Step 7.1: The targets continuously observed by the radar within two scanning cycles are taken as the targets to be tracked. In this embodiment of the invention, there are 5 UAV targets to be tracked. The radar scanning cycle is 1 second and the tracking time is 100 seconds.

[0076] Step 7.2, a sequence is formed by the position vectors of each to-be-tracked target from the initial sampling time to the current sampling time, and each normalized position sequence of each target is input into the trained prediction network, and five position sequences are output, and the last vector in the position sequence is taken as the predicted position vector of the UAV target at the current sampling time.

[0077] Step 7.3, the Euclidean distance between the state vector of the pth target at the current sampling time and the bth noisy measurement vector is calculated as the cost of the association between the target and the measurement, and all costs of the association between the targets and the measurements are obtained, and a global nearest neighbor (GNN) algorithm is used to assign an associated measurement vector to each target:

[0078]

[0079] wherein x p represents the position vector of the pth target at the current sampling time, z q represents the qth noisy measurement vector at the current sampling time, D represents the dimension of the state vector and the noisy measurement vector, x p (d) represents the value of the dth dimension in x p (d) represents the value of the dth dimension in x q (d) represents the value of the dth dimension in x q (d) represents the value of the dth dimension in x

[0080] Step 7.4, a sequence is formed by all associated measurement vectors of each UAV target from the initial sampling time to the current sampling time, and each normalized measurement sequence of each to-be-tracked target is input into the trained update network, and five position sequences are output, and the last vector in the position sequence is taken as the updated position vector of the UAV target at the current sampling time.

[0081] Step 7.5, it is judged whether the current tracking sampling time is the last sampling time, if yes, step 8 is executed, otherwise, step 7.2 is executed.

[0082] Step 8, trajectory synthesis:

[0083] According to each updated position vector, the position of each to-be-tracked target at each time is determined, and the positions of each to-be-tracked target at all sampling times are connected to obtain the tracking trajectory of the target.

[0084] The effect of the application will be further described below in combination with a simulation experiment.

[0085] 1. Simulation experiment conditions:

[0086] The hardware platform of the simulation experiment of the application is: the processor is Intel i7 4790 CPU, the main frequency is 3.6 GHz, and the memory is 16 GB.

[0087] The software platform of the simulation experiment of the application is: Windows 10 operating system and Python 3.8.

[0088] 2. Simulation content and result analysis:

[0089] In the simulation experiment of the application, the radar and the unmanned aerial vehicle target are in rectangular coordinates, the radar is located at [0, 0, 0] m, and the target is located at a fixed height, and the range is a low-altitude plane of [-2000, 2000] m x [0, 2000] m. There are a total of 5 unmanned aerial vehicle targets to be tracked in the simulation scene, of which 1 target moves in a CV manner and 4 targets move in a CT manner. In the simulation experiment of the application, the radar continuously observes the target for 100 scanning periods, and each scanning period is 1 s. The covariance of the measurement noise in the scene is set to The clutter rate is set to λ c = 10, the survival probability of the target is set to p s = 0.99, and the detection probability of the target is set to p d = 0.98.

[0090] In the simulation experiment of the application, the method of the application and two prior arts are used to track the above-mentioned 5 unmanned aerial vehicle targets at 100 consecutive sampling times, and the target positions and numbers estimated by the three methods at each time are obtained. The moving trajectories of the 5 targets in the simulation experiment are plotted into Figure 3 (a). Figure 3 (b) is plotted by the number of targets estimated by the three methods at each sampling time.

[0091] In order to verify the simulation effect of the application, the position error between the target positions estimated by the three methods and the true positions is calculated by using the Euclidean distance formula, and Figure 3 (c) is obtained, the comprehensive error between the target positions and numbers estimated by the three methods and the true values is calculated by using the OSPA error formula, and Figure 3 (d) is obtained.

[0092] In the simulation experiment, the two prior arts used are:

[0093] Prior art 1 refers to a tracking method applying multipath effect and unscented Kalman filter disclosed in the patent document “Low-altitude target tracking filter method under multipath effect” (patent application number CN202010671798, application publication number CN111708015A) applied by Harbin Institute of Technology.

[0094] The prior art 2 refers to the target tracking method based on recurrent neural network proposed by Emambakhsh M in the published paper "Deep recurrent neural network for multi-target filtering." (Conference Proceedings of the International Conference on Multimedia Modeling, 2019).

[0095] The effects of the present application will be further described below in combination with Figure 3 the simulation figures.

[0096] Figure 3 (a) is a moving track diagram of 5 targets, the horizontal coordinate represents the length of the simulation scene area, and the unit is m, and the vertical coordinate represents the width of the simulation scene area, and the unit is m. Figure 3 The straight line in (a) represents the moving track of one target moving in the CV mode, and the remaining four curves represent the moving tracks of four targets moving in the CT mode.

[0097] Figure 3 (b) is a target estimation number curve diagram obtained by the method of the present application and two prior arts, the horizontal coordinate represents the sampling time, and the vertical coordinate represents the estimated target number. Figure 3 The curve marked with a circle in (b) represents the target estimation number curve at each time obtained by the method proposed in the present application, the curve marked with a square represents the target estimation number curve at each time obtained by the prior art 1, and the curve marked with a triangle represents the target estimation number curve at each time obtained by the prior art 2.

[0098] It can be seen from Figure 3 (b) that: the three technologies can accurately estimate the target number.

[0099] Figure 3 (c) is a position error curve diagram obtained by the method of the present application and two prior arts, the horizontal coordinate represents the sampling time, and the vertical coordinate represents the position error, and the unit is m. Figure 3 The curves marked with a circle, a square and a triangle in (c) respectively represent the position error curve at each time obtained by the method proposed in the present application, two prior arts.

[0100] It can be seen from Figure 3 (c) that: in the first about 10 sampling time of tracking, the position error obtained by the method of the present application is not much different from the two prior arts, and after about 10 sampling time, the position error obtained by the method of the present application is significantly lower than the two prior arts, which shows that the method of the present application using the method based on the prediction network and the update network can more accurately estimate the target position.

[0101] Figure 3 (d) The OSPA error curves of the method of the present application and the two prior arts, with the abscissa representing the sampling time and the ordinate representing the OSPA error in meters. The curves marked with circles, squares and triangles represent the OSPA error curves at each time obtained by using the method of the present application, the two prior arts respectively.

[0102] From ​ (d) It can be seen that: in the first few sampling time of starting tracking, the OSPA error obtained by the method of the present application is not much different from the two prior arts, but after the 20th sampling time, the OSPA error obtained by the method of the present application is significantly lower than the two prior arts. The present application uses a prediction network and an update network for tracking, and has relatively high tracking accuracy.

Claims

1. A multi-target adaptive tracking method based on a prediction network and an update network, characterized in that, A prediction network is constructed to predict the target's position vector by learning motion patterns from the target's historical position vectors. An update network is constructed to update the target's position vector by learning noise characteristics from measurements. Multi-target adaptive tracking is achieved by combining this network with a Global Nearest Neighbor (GNN) data association algorithm. The specific steps of this method are as follows: Step 1, Construct the prediction network: A prediction network is constructed by concatenating a deep long short-term memory (DLSTM) subnetwork and a fully connected subnetwork (FCN). The depth of the DLSTM subnetwork is set to 2, the input dimension is set to 2, the input time step is set to 10, and the hidden layer dimension is set to 64. The input dimension of the FCN is set to 64, and the output dimension is set to 2. Step 2, Build and update the network: Build an update network with the same structure as the prediction network; set the depth of the DLSTM subnetwork in the update network to 3, the input dimension to 2, the input time step to 20, and the dimension of the hidden layer to 128; set the input dimension of the FCN to 128 and the output dimension to 2. Step 3, generate the training set for the prediction network: Step 3.1: In a low-altitude plane scene of [-2000, 2000]m × [-2000, 2000]m, randomly select at least 128 targets moving in each motion mode, and uniformly sample the tracking time with the period of radar scanning and tracking scene as the sampling interval to obtain multiple sampling times. Step 3.2: Using the state transition equation, the state vector of each target is transferred from the previous sampling time to the next sampling time. After m transfers, the state vector of the target at all sampling times is obtained, where the value of m is determined by the tracking duration of each target to be tracked. Step 3.3: Extract the x-axis coordinates and y-axis coordinates from the state vector of each target at each sampling time to form a position vector, and combine the position vectors of each target at all sampling times to form the position sequence of the target; Step 3.4: Divide the position sequence of each target into n-9 subsequences, and form a pair of training samples by combining two adjacent subsequences. The first subsequence is used as the input sequence and the second subsequence is used as the label sequence. All training samples are combined into the training set of the prediction network, where n represents the length of the position sequence, n = m + 1. Step 4: Generate a training set for the updated network: Step 4.1: Using the observation equation, calculate the noisy measurement vector of each target at each sampling time, and combine the noisy measurement vectors of each target at all sampling times to form the measurement sequence of the target; Step 4.2: Divide the measurement sequence and position sequence of each target into n′-19 subsequences. Combine the measurement subsequence and position subsequence at the same starting time into a pair of training samples. Use the measurement subsequence as the input sequence and the position subsequence as the label sequence. Combine all training samples into the training set for updating the network. Here, n′ represents the length of the measurement sequence and takes a value equal to n. Step 5, train the prediction network: The training set of the prediction network is input into the prediction network, and the parameters in the prediction network are iteratively updated using the Adam optimization algorithm and the simulated annealing algorithm until the loss function converges, thus obtaining the trained prediction network. Step 6, train and update the network: The training set of the update network is input into the update network, and the parameters in the network are iteratively updated using the Adam optimization algorithm and the simulated annealing algorithm until the loss function converges, thus obtaining the trained update network. Step 7: Use the prediction network and the update network to track multiple targets in the scene: Step 7.1: Select the targets continuously observed by the radar during two scanning cycles as the targets to be tracked; Step 7.2: Form a sequence of position vectors for each target to be tracked from the initial sampling time to the current sampling time. Input the normalized position sequence of each target into the trained prediction network. Take the last vector in each output sequence as the predicted position vector of the target at the current sampling time. Step 7.3: Using the Euclidean distance calculation formula, calculate the Euclidean distance between the predicted position vector of each target to be tracked at the current sampling time and each noisy measurement vector. This distance serves as the cost of associating the target and the measurement. Using the Global Nearest Neighbor (GNN) data association algorithm, assign an associated measurement vector to each target. Step 7.4: Form a sequence of all associated measurement vectors of each target to be tracked from the initial sampling time to the current sampling time. Input the normalized measurement sequence of each target to be tracked into the trained update network. Use the last vector in each output sequence as the update position vector of the target at the current sampling time. Step 7.5: Determine whether the current tracking sampling time is the last sampling time. If yes, proceed to step 8; otherwise, proceed to step 7.

2. Step 8, Synthesize the trajectory: The position of the target to be tracked at each time step is determined based on each updated position vector, and the tracking trajectory of the target is obtained by connecting the positions of each target to be tracked at all sampling times.

2. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The state transition equation described in step 3.2 is as follows: in, Let k represent the state vector of the l-th target at the k-th sampling time, where k = 1, 2, ..., n. This represents the state transition matrix of the l-th target at sampling time k-1. This represents the state vector of the l-th target at the (k-1)-th sampling time. This represents the noise vector during the process of the l-th target transitioning from the (k-1)-th sampling time to the k-th sampling time.

3. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The parameters included in the state vector in step 3.2 are as follows: the target's x-axis coordinate, the target's moving speed along the x-axis, the target's y-axis coordinate, the target's moving speed along the y-axis, the target's rotational angular velocity, the target's acceleration in the x-axis, and the target's acceleration in the y-axis.

4. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The step 3.4, which describes dividing the position sequence of each target into n-9 subsequences, means dividing the position sequence of each target into n-9 subsequences in a sliding manner, starting from each moment in each position sequence and taking a step size of 10 moments.

5. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The observation equation described in step 4.1 is as follows: in, H represents the noisy measurement vector of the l-th target at the k-th sampling time. k The transformation matrix represents the mapping of the state vector from the state space to the observation space. This represents the noise vector during the process of the l-th target being transformed from a state vector to a measurement vector at the k-th sampling time.

6. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The step 4.2, which describes dividing the measurement sequence and position sequence of each target into n′-19 subsequences, means dividing the position sequence and measurement sequence of each target into n′-19 subsequences in a sliding manner, starting from each time point in each sequence and using every 20 sampling times as a division step.

7. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The loss function described in step 5 is as follows: Where M represents the total number of training samples in the training set of the prediction network. Indicates the i-th p The total number of vectors contained in the input sequence of each training sample. Indicates the i-th p The r-th vector in the label sequence of training samples, Indicates the i-th p The input sequence of training samples is the r-th vector in the output sequence of the prediction network, and ||·||2 represents the calculation of the L2 norm.

8. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The loss function described in step 6 is as follows: Where M′ represents the total number of training samples in the training set for updating the network. Indicates the i-th u The total number of vectors contained in the input sequence of each training sample. Indicates the i-th u The r′-th vector in the label sequence of the training samples Indicates the i-th u The input sequence from the training samples is used to update the r′-th vector in the network output sequence.

9. The multi-target adaptive tracking method based on a prediction network and an update network according to claim 1, characterized in that: The formula for calculating the Euclidean distance mentioned in step 7.3 is as follows: Where, x p Let z represent the position vector of the p-th target at the current sampling time. q Let x represent the q-th noisy measurement vector at the current sampling time, where D represents the dimension of the state vector and the noisy measurement vector, and x represents the dimension of the state vector and the noisy measurement vector. p (d) represents x p The value of the d-th dimension, z q (d) represents z q The value of the d-th dimension.

Images