A method for adaptive tracking of a maneuvering target

Abstract

The application provides a motorized target adaptive tracking method, and the implementation steps are as follows: obtaining a training sample set and a test sample set; building a motorized target adaptive tracking network model and iteratively training the same; and obtaining a motorized target adaptive tracking result.In the process of iteratively training the motorized target adaptive tracking network model, a long short-term memory neural network in a motion mode recognition module extracts time sequence information of a target coordinate sequence, a multi-head attention layer extracts context information in the time sequence information, and finally, a multilayer perception and a Softmax classifier are used to classify the motion mode of the motorized target, so that the adaptive recognition of the motion mode of the target is realized.Compared with the prior art, the tracking precision is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar tracking, and relates to a method for adaptively tracking a maneuvering target, which can be applied to the field of aircraft navigation. BACKGROUND

[0002] During the movement of a maneuvering target, the movement state of the target will change constantly, and it is difficult to accurately describe the movement of the target using a single movement model at this time, and an unmatched movement model will often lead to poor tracking performance. Therefore, it is more reasonable to use multiple target movement models to match the movement of the maneuvering target. One of the key technologies in the multi-model algorithm is the identification of the target movement mode, that is, the model and the target movement mode can achieve a reasonable matching probability, so that the model probability of the model that meets the target movement mode in the multi-model is larger, and the matching probability of the model that does not meet the target movement mode is smaller, so that the tracking algorithm has higher tracking accuracy. However, in the early static multi-model algorithm, the model probability is fixed, and there is a lack of necessary information interaction between the filters corresponding to each model, and when the target movement mode changes, the filter may diverge, which deteriorates the tracking performance. The appearance of the interacting multi-model algorithm overcomes this deficiency, and the algorithm can realize the identification of the target movement model and also transfer the model between the models through the Markov transition probability matrix. However, the update of the model probability in the interacting multi-model algorithm needs to calculate the innovation and innovation covariance of the filter algorithm, and other information, and cannot be updated separately from the filtering process, and in the case where the target movement mode does not change for a long time, unnecessary computational overhead is added, and the interacting multi-model algorithm needs prior information such as model probability and model transition probability, which is often difficult to obtain in actual scenarios, so that it cannot be applied to the tracking of a maneuvering target in an unknown scenario.

[0003] In order to overcome this defect, the patent application with the publication number CN 116736289A, entitled "Multi-maneuver target tracking method based on prediction update network and dynamic wave door", discloses a multi-maneuver target tracking method based on prediction update network and dynamic wave door. The invention trains the prediction network and the update network respectively by using the training set of the constructed prediction network and the update network. The prediction position vector of the target to be tracked is obtained by using the constructed prediction network; the effective measurement of the target to be tracked is obtained by using the dynamic wave door, and it is judged whether the target to be tracked appears maneuvering; the associated measurement vector of the target to be tracked is allocated according to the global nearest neighbor (GNN) algorithm; the update position vector of the target to be tracked is obtained by using the trained update network. The invention gets rid of the limitation of preset model and model transition probability by combining the prediction update network and the dynamic wave door, realizes the adaptive tracking of multi-maneuver target, and improves the tracking accuracy, but it cannot identify the motion mode of the maneuvering target in the tracking process, and uses the same network parameters to realize the prediction of the motion state of the target with different motion modes, which affects the further improvement of the target tracking accuracy. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the prior art and provide a maneuvering target adaptive tracking method to solve the problem of low tracking accuracy in the prior art.

[0005] To achieve the above purpose, the technical solution adopted by the present application includes the following steps:

[0006] (1) Obtain a training sample set and a test sample set:

[0007] Obtain a first data set G1 containing K target motion categories and S target coordinate sequences, and a second data set G2 containing S target coordinate sequences with noise interference. Each target coordinate sequence contains N coordinate values, and each coordinate value is normalized. Then, each target coordinate sequence is labeled, and the L target coordinate sequences in the normalized first data set and second data set and their corresponding labels are used to form a training sample set, and the remaining S-L target coordinate sequences are used as a test sample set, wherein K≥2, N≥10, S≥128,

[0008] (2) Build a maneuvering target adaptive tracking network model:

[0009] A maneuvering target adaptive tracking network model W is built, which includes a motion mode recognition module, a state prediction module and a denoising autoencoder. The motion mode recognition module includes a cascaded long short-term memory neural network, a residual connection and a regularization layer, a multi-head attention layer, two multi-layer perceptron (MLP) and a Softmax classifier.

[0010] (3) Iterative training of the maneuvering target adaptive tracking network model:

[0011] Iterative training of the maneuvering target adaptive tracking network model by the training sample set, to obtain the trained maneuvering target adaptive tracking network model;

[0012] (4) Obtain the maneuvering target adaptive tracking result:

[0013] Forward propagation is performed on the test sample set as the input of the motion mode recognition module and the state prediction module in the trained maneuvering target adaptive tracking network model, to obtain the corresponding correlation measurement value of each test sample, and the denoising autoencoder is used to denoise the correlation measurement value, to obtain the maneuvering target adaptive tracking result.

[0014] Compared with the prior art, the present application has the following advantages:

[0015] In the process of iterative training of the maneuvering target adaptive tracking network model, the long short-term memory neural network in the motion mode recognition module extracts the time sequence information of the target coordinate sequence, the multi-head attention layer extracts the context information in the time sequence information, and finally the multi-layer perception and Softmax classifier are used to realize the classification of the motion mode of the maneuvering target, so as to realize the adaptive recognition of the target motion mode. Compared with the prior art, the tracking accuracy is effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the present application.

[0017] Figure 2 The structure and recognition principle diagram of the motion mode recognition module of the present application.

[0018] Figure 3 The OSPA distance simulation comparison diagram of the present application and the prior art. DETAILED DESCRIPTION

[0019] Referring to Figure 1 The present application comprises the following steps:

[0020] Step 1) Obtain the training sample set and the test sample set:

[0021] The first data set G1 and the second data set G2 are obtained, the first data set G1 contains K target motion categories and S target coordinate sequences, each target coordinate sequence contains N coordinate sequences, each coordinate data is normalized, the target motion category in each target coordinate sequence is labeled, L target coordinate sequences in the normalized first data set and the second data set and corresponding labels are used to form a training sample set, and the remaining S-L target coordinate sequences are used as a test sample set, wherein K is greater than or equal to 2, N is greater than or equal to 10, and S is greater than or equal to 128,

[0022]

[0023] In this embodiment, K is 3, N is 10, S is 128, and L is 96.

[0024] Step 2) Building a mobile target adaptive tracking network model:

[0025] A mobile target adaptive tracking network model W is built, which includes a motion pattern recognition module, a target state prediction module and a noise reduction autoencoder. The structure of the motion pattern recognition module is shown in Figure 2 The motion pattern recognition module includes a cascaded long short-term memory neural network LSTM, a residual connection and a regularization layer ADD&NORM, a multi-head attention layer, two multi-layer perceptrons MLP and a Softmax classifier.

[0026] The target state prediction module includes a plurality of encoding modules and a multi-layer perceptron MLP connected in cascade, wherein the encoding module includes a multi-head attention layer and a feedforward connection layer stacked, and a regularization layer connected in residual with the output of the multi-head attention layer and the feedforward connection layer.

[0027] The noise reduction autoencoder includes an encoder and a decoder connected in cascade.

[0028] The network structure and parameters of the motion recognition module in this example are as follows:

[0029] The long short-term memory neural network LSTM has 128 hidden layer units, 1 hidden layer, an input dimension of 2 and an output dimension of 2.

[0030] In the multi-head attention layer, the input dimension is 128, the number of heads is 8, and the output dimension is 512; in the first multi-layer perceptron, the number of units is 128, and in the second multi-layer perceptron, the number of units is 3.

[0031] The network structure and parameters of the target state prediction module in this example are as follows:

[0032] There are 3 parallel prediction networks in the target state prediction module: each network contains 4 encoding modules, the input data dimension is 512, the number of multi-head is 8, the number of feedforward layer units is 2048, the normalization method is layer normalization, and the number of multilayer perception units is 1;

[0033] The specific parameters of the denoising autoencoder network in the example are as follows:

[0034] The denoising autoencoder network: the input data dimension is 2, the number of hidden layer units is 16, and the output data dimension is 2;

[0035] Step 3) Iterative training of the maneuvering target adaptive tracking network model is implemented as follows:

[0036] (3a) Initialize the iteration number t, the maximum iteration number T, T≥250, and the learnable parameters of the maneuvering target adaptive tracking network model in the tthiteration as ω t , and let t=1;

[0037] (3b) The motion pattern recognition module classifies the motion pattern of each training sample to obtain the probability distribution p(k) of the membership motion pattern class of each training sample; the target state prediction module performs state prediction on each training sample to obtain the state prediction value The denoising autoencoder performs denoising processing on each training sample to obtain the state denoising value

[0038] The principle of classifying each training sample by the motion pattern recognition module is as shown in Figure 2 , and the implementation steps are as follows:

[0039] (3b1) The long short-term memory neural network extracts the time sequence information of each training sample to obtain N time step state outputs O=[o1,o2,...,o n ,...,o N ] and the hidden feature state h N of the last time step, where o n represents the output of the time step corresponding to the nth coordinate value.

[0040] (3b2) The residual connection and regularization layer performs residual connection and regularization processing on the N time step state outputs O and the hidden feature state h N to obtain the normalized output and the hidden feature state

[0041] (3b3) The multi-head attention layer performs nonlinear mapping on to obtain the query matrix Query, and performs nonlinear mapping on Different nonlinear mappings are performed to obtain the key matrix Key and the value matrix Value. Then, attention mechanism operations are performed on Query, Key, and Value to obtain H attention outputs. Finally, the H attention outputs are concatenated to obtain the multi-head attention output MultiHead.

[0042] (3b4) Two multilayer perceptrons (MLPs) perform nonlinear mapping on the multihead attention output MultiHead to obtain the category output results; the Softmax classifier transforms the category output results obtained by the nonlinear mapping to obtain the probability distribution p(k) of the motion pattern category of each training sample.

[0043] In the motion pattern recognition module for moving targets, a fixed-length N position sequence is used as the input to the network. The temporal information in the sequence is extracted by a long short-term memory neural network to obtain the result containing temporal information and attention mechanism operation. Then, the information is output through a perceptron and a softmax classifier to obtain the probability distribution of which motion pattern the fixed-length sequence belongs to, thereby realizing the recognition of the target motion pattern.

[0044] The steps for the target state prediction module to classify each training sample are as follows:

[0045] First, the position sequence in each training sample is encoded to obtain an encoding vector. Then, a corresponding position vector is added to the encoding vector to obtain the input of the encoding module. The position vector is encoded using triangular position encoding. Next, a multi-head self-attention mechanism is used in the encoding module to extract the correlation information in the input data. To make the data more stable and better propagation, residual connections and layer normalization operations are used for processing. Then, a feedforward layer is used for the output result. Similarly, to ensure data stability and propagation, residual connections and layer normalization operations are performed after the feedforward layer. The output of the resulting encoding module is used as the input of the next encoding module, and the above operations are repeated to obtain the state prediction value of each training sample.

[0046] In training the target prediction module, each type of target motion pattern training sample corresponds to a prediction network. This ensures that a network model can only learn the motion trend of one type of maneuvering target motion pattern, thereby achieving better prediction of the target state.

[0047] (3c) via p(k), and The loss value of the computer moving target adaptive tracking network model W is calculated using the Adam optimizer and cosine annealing algorithm, and the loss is applied to the learnable parameters ω. t The model is updated to obtain the adaptive tracking network model W for the maneuvering target after this iteration.t wherein:

[0048] Loss = aLoss1 + bLoss2 + gLoss3

[0049]

[0050]

[0051]

[0052]

[0053] wherein, Loss1, Loss2, Loss3 represent the loss values of the motion pattern recognition module, the target state prediction module, and the denoising autoencoder respectively, a, b, and g represent weight parameters, and a+b+g=1, ∑ represents a summation operation, p(k) represents a probability of belonging to the kth category, η represents a gradient descent parameter, and ω t+1 represents the update result of ω t . represents a partial derivative operation, wherein η=0.0001.

[0054] (3d) determining whether t=T is true, if so, obtaining the trained maneuvering target adaptive network model W * , otherwise, setting t=t+1, W t =W, and executing step (3c).

[0055] Step 4) obtaining a maneuvering target adaptive tracking result:

[0056] taking the test sample set as the input of the motion pattern recognition module and the state prediction module in the trained maneuvering target adaptive tracking network model, the motion pattern recognition module classifies the motion pattern of each test sample to obtain the probability distribution of the motion pattern category to which each test sample belongs, the target state prediction module predicts the state of each test sample to obtain K state prediction values of each test sample, and the K state prediction values are weighted and summed according to the probability distribution to obtain a target state prediction value, then a global nearest neighbor method is used to obtain an associated measurement value according to the target state prediction value, and the denoising autoencoder is used to denoise the associated measurement value to obtain the maneuvering target adaptive tracking result.

[0057] The technical effects of the present application will be further described below in combination with simulation results:

[0058] 1. Simulation conditions and contents:

[0059] The processor used in the simulation experiment is an Intel i7 13700H CPU with a frequency of 2.4 GHz and a memory of 16 GB. The software platform is: Windows 10 operating system and Python 3.8.

[0060] The optimal sub-pattern allocation (OSPA) distance of the present application and the existing multi-maneuver target tracking method based on predictive update network and dynamic wave gate were compared and simulated, and the results are shown in Figure 3

[0061] 2. Analysis of simulation results:

[0062] Referring to Figure 3 , the abscissa represents the sampling time of 60s, and the ordinate represents the OSPA distance, with the unit of m. It can be seen from the figure that at the 20th and 40th sampling time when the target motion mode changes, the OSPA distances obtained by the present application and the prior art are improved to a certain extent, and the OSPA distances at other times do not appear large fluctuations, and the OSPA distances of the present application at all times are less than those of the prior art, so the present application has higher tracking accuracy compared with the prior art.​

Claims

1. An adaptive tracking method for maneuvering targets, characterized in that, Includes the following steps: (1) Obtain the training sample set and the test sample set: All of the above are included. Types of target motion categories and including The first dataset of segment target coordinate sequence and includes noise interference The second dataset of segment target coordinate sequences Each target coordinate sequence contains Each coordinate value is normalized, and then each target coordinate sequence is labeled. Finally, the normalized values ​​from the first and second datasets are compared. The training sample set consists of the sequence of target coordinates and their corresponding labels, and the remaining... The segment target coordinate sequence is used as the test sample set, where, , , , ; (2) Construct an adaptive tracking network model for maneuvering targets: Construct an adaptive tracking network model for maneuvering targets, comprising a motion pattern recognition module, a state prediction module, and a noise reduction autoencoder. The motion pattern recognition module includes a cascaded long short-term memory neural network, residual connections and regularization layers, a multi-head attention layer, two multilayer perceptrons (MLPs) and a Softmax classifier. (3) Iteratively train the adaptive tracking network model for maneuvering targets: The adaptive tracking network model for maneuvering targets is iteratively trained using a training sample set to obtain a well-trained adaptive tracking network model for maneuvering targets. (4) Obtain adaptive tracking results for maneuvering targets: The test sample set is used as the input to the motion pattern recognition module and state prediction module in the trained maneuvering target adaptive tracking network model for forward propagation to obtain the associated measurement value corresponding to each test sample. The denoising autoencoder denoises the associated measurement value to obtain the maneuvering target adaptive tracking result.

2. The method according to claim 1, characterized in that, The adaptive tracking network model for maneuvering targets described in step (2) includes: The motion pattern recognition module, including its multi-head attention layer, comprises... Head attention layer; Target state prediction module, including The prediction network has the same structure, and each network contains multiple cascaded coding modules and a multilayer perceptron (MLP). The coding modules include stacked multi-head attention layers and feedforward connection layers, as well as regularization layers that are residually connected to the outputs of the multi-head attention layers and feedforward connection layers, respectively. Noise-reducing autoencoders, including cascaded encoders and decoders.

3. The method according to claim 2, characterized in that, The iterative training of the adaptive tracking network model for maneuvering targets described in step (3) is implemented as follows: (3a) Initialize the number of iterations to be The maximum number of iterations is , , No. The learnable parameters of the next iteration of the adaptive tracking network model for maneuvering targets are: and order ; (3b) The motion pattern recognition module classifies the motion pattern of each training sample and obtains the probability distribution of the motion pattern category to which each training sample belongs. ; The target state prediction module performs state prediction for each training sample, obtaining the state prediction value for each training sample. ; The denoising autoencoder denoises each training sample to obtain the denoised state value for each training sample. ; (3c) Through , and Computer-aided adaptive tracking network model for moving targets loss value and through For learnable parameters The model is updated to obtain the adaptive tracking network model for maneuvering targets after this iteration. ; (3d) Judgment If true, then a well-trained adaptive network model for maneuvering targets is obtained. Otherwise, let , And perform step (3c).

4. The method according to claim 3, characterized in that, The adaptive tracking network model for maneuvering targets described in step (3c) loss value The calculation formula is: ； ； ； ； in, , , These represent the loss values ​​of the motion pattern recognition module, the target state prediction module, and the noise reduction autoencoder, respectively. , and These represent the weight parameters, and , This indicates a summation operation.

5. The method according to claim 3, characterized in that, The step (3c) described for the learnable parameters The update is performed using the following formula: ； in, Represents the gradient descent parameters. express The update results This indicates the partial derivative operation.

6. The method according to claim 3, characterized in that, The method for obtaining the associated measurement value corresponding to each test sample mentioned in step (4) is as follows: The motion pattern recognition module classifies the motion pattern of each test sample, obtaining the probability distribution of the motion pattern category to which each test sample belongs; the target state prediction module predicts the state of each test sample, obtaining the probability distribution of the target state of each test sample. One state prediction value; And according to probability distribution The predicted state values ​​are weighted and summed to obtain the predicted target state value. Then, the global nearest neighbor method is used to obtain the associated measurement value based on the predicted target state value.

Images