Model matching intelligent maneuvering target tracking algorithm

Abstract

The model matching intelligent maneuvering target tracking algorithm is suitable for an aerial carrier radar to track a target object, the radar comprises an interface control file and a GPS file, and the method comprises the steps that S101, network models of the target object are matched through a recurrent neural network and a full connection layer network; s102, training the network model, and determining verification data; s103, the target object is tracked based on the verification data, even at the maneuvering jump moment, the algorithm can still rapidly recognize correct target motion characteristics, the track interruption rate of the target at the maneuvering jump moment is greatly reduced, and accurate and stable situation information is provided for a pilot in the whole process.

Description

Technical Field

[0001] The invention belongs to the technical field of radar data processing, and in particular relates to a model matching intelligent maneuvering target tracking algorithm. Background Art

[0002] Target model selection is one of the most critical elements in the current maneuvering target tracking algorithm. The accuracy of target model selection is directly related to whether the target tracking system accurately describes the target motion law, which greatly affects the target tracking accuracy. The existing multi-model method is based on the constructed model library and calculates the likelihood probability of each model in the model library to achieve target model selection. This selection method is overly dependent on system parameters such as model set, Markov transfer matrix and measurement error, which leads to maneuver delay when the target maneuvers. The target tracking accuracy drops sharply during maneuver jumps, resulting in low target tracking performance. Summary of the invention

[0003] In view of this, the model matching intelligent maneuvering target tracking algorithm of the present invention solves the technical problem of low performance of existing methods for tracking targets.

[0004] A model matching intelligent maneuvering target tracking algorithm is suitable for matching and positioning of ground targets by airborne radar. The target matching method includes:

[0005] S101: Acquire a target image and a pre-stored reference image library, wherein the target image is a real-time image of a radar, and the reference image library includes reference images;

[0006] S102: extracting features of the target image and the reference image using a multi-scale feature pyramid network, and outputting a specific position of the target image in the reference image using a region perception and position regression model.

[0007] Beneficial Effects

[0008] The method of the present invention can greatly improve the accuracy of target motion model recognition by deeply mining a large amount of historical data feature information in an offline manner, thereby significantly improving the target tracking accuracy. Even at the moment of maneuver jump, the algorithm can still quickly identify the correct target motion characteristics, so that the target's track interruption rate at the moment of maneuver jump is greatly reduced, providing accurate and stable situation information to the pilot throughout the process, and can be widely applied to various target tracking platforms, with broad application prospects and huge application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1 Schematic diagram of the RNN model;

[0011] Figure 2 Schematic diagram of the fully connected layer network model;

[0012] Figure 3 is a flow chart of the method of the present invention;

[0013] Figure 4 To verify the northbound velocity tracking results in sample 1 scenario;

[0014] Figure 5 To verify the eastward velocity tracking results in sample 1 scenario. DETAILED DESCRIPTION

[0015] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0016] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0017] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0018] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure, and only show the components related to the present disclosure rather than being drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type may also be more complicated.

[0019] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, one skilled in the art will appreciate that aspects may be practiced without these specific details.

[0020] like Figures 1 to 5 The model matching intelligent maneuvering target tracking algorithm shown is suitable for tracking the target by the aircraft radar. The radar includes an interface control file and a GPS file. The method includes:

[0021] S101: Matching the network model of the target object by using a recurrent neural network and a fully connected layer network;

[0022] S102: The network model is trained and verification data is determined;

[0023] S103: Tracking the target object based on the verification data.

[0024] By deeply mining a large amount of historical data feature information in an offline manner, the accuracy of target motion model recognition can be greatly improved, thereby significantly improving target tracking accuracy. Even at the moment of maneuver jump, the algorithm can still quickly identify the correct target motion characteristics, greatly reducing the target's track interruption rate at the moment of maneuver jump, providing pilots with accurate and stable situation information throughout the process, and can be widely applied to various target tracking platforms, with broad application prospects and huge application value.

[0025] As a specific implementation method provided in this case, S101 includes:

[0026] Obtain parameter information fed back by the interface control file, the parameter information including distance information, angle measurement information and time series information, and obtain motion characteristics of the target object fed back by the GPS file;

[0027] The parameter information is used as an input of the network model and the motion feature is used as an output of the network model, and the network model is trained to determine the parameters of the network model, for example, the parameters include weights and biases, etc.;

[0028] The input of the network model is represented as [X1,X2,…X τ ], the output is expressed as [Y1,Y2,…Y τ ], the intermediate variable H of the network modeli And the final output Y i It is expressed as:

[0029]

[0030] Where W X , W H and B H They are the input weight parameter, hidden weight parameter and bias parameter of the recurrent neural network RNN ​​respectively; W Y and B Y is the input weight parameter and bias parameter of the fully connected layer, H i-1 is the output result at the i-1th moment, H i is the output result at the i-th moment, Y i Output result of the network; X i is the input at the i-th moment; e is a constant.

[0031] Furthermore, the target trajectory information can be characterized by the distance and angle information obtained by the radar. The present invention uses the distance, angle measurement information and time series as input information to train and obtain the current motion model of the target. When the time series length is 1, the output vector length can be set to 1 and the model result can be directly trained using a recurrent neural network (RNN). However, 1 frame of measurement information often cannot extract the target motion features, making it impossible for the network to judge the motion model. Therefore, multiple frames of measurement can be used for model matching during the training process.

[0032] Assuming that the length of the time series is τ and the dimension of the hidden layer is n, the total output dimension of the RNN is nτ. If the output dimension is o, a new network needs to be built to correct the output dimension. Therefore, after the RNN, a fully connected layer needs to be built to match the output information dimension. It is worth noting that for the model matching problem, the output dimension can be determined by the number of model sets. If the model matching is regarded as a classification problem, the output information can be regarded as vector information with a dimension of o. The input vector X = [X1, X2, ... X τ ], output vector Y = [Y1, Y2, ... Y τ ], the model recognition process of the recurrent neural network RNN ​​and the fully connected layer network is expressed as:

[0033]

[0034] As a specific implementation method provided in this case, S102 includes:

[0035] In the process of network model training, relevant work needs to be carried out in data preprocessing, data set division, and model parameter setting. In terms of data preprocessing, in order to make the data more suitable for RNN, the relevant information is linearly mapped to the interval between 0 and 1, and the differences between different sample sets are reflected through preprocessing;

[0036] The input vector X and the output vector Y are linearly mapped to the interval between 0 and 1 to obtain the processed input vector Output vector Respectively expressed as:

[0037]

[0038] Where, X total Input information sample set, Y total To output the information sample set, max() represents the maximum value of the set, min() represents the minimum value of the set, and the input vector Output vector Part of it is used as training data, and part of it is used as verification data. Generally, the data is divided in a ratio of 2:8.

[0039] In terms of data set division, the target will move in different maneuvering styles in different time periods, so that its full-cycle motion track often contains different motion categories. Therefore, identifying the current maneuvering characteristics of the target often only requires a small segment of historical data. In order to achieve this goal, the present invention first needs to determine the length of historical data required for identification, segment the target historical data based on the length, obtain the overall input and output sample set, and then divide the training set and verification set from the overall sample set according to a certain ratio for use by the deep learning network model. Generally, in the process of training data, parameter setting or matching is performed. For example, in order to ensure the efficiency of network training, it is necessary to reasonably set training parameters such as the number of iterations, optimization criteria, and learning rate. After the above work is completed, the parameter W in the designed network model can be adjusted. X , W H , B H , W Y and B Y Perform iterative training.

[0040] As a specific implementation method provided in this case, S103 includes:

[0041] Determine whether the length of the input vector X exceeds the tracking threshold length. If so, based on the output vector at time k+1 Weighted fusion is used to determine the tracking result of the target. o is the number of model sets. If not, the traditional interacting multi-model extended Kalman filter algorithm (I MM-EKF) is used to achieve target tracking.

[0042] Further, weighted fusion is used to determine the tracking results of the target object, including:

[0043] S601: Determine the output vector by the network model Model recognition probability satisfy:

[0044]

[0045] S602: The model recognition probability The prior information of the target object is obtained based on the Markov transfer matrix. The prior information includes the first state and the first covariance It is expressed as:

[0046]

[0047] In the formula, represents the transfer weight from the sub-model of the j-th target object to the sub-model of the i-th target object, π ij is the Markov transition probability from the jth target object sub-model to the i-th target object sub-model, and are the target state and covariance of the sub-model of the j-th target at time k;

[0048] S603: After obtaining the prior information, obtain the posterior information of each sub-model of the target object based on the extended Kalman filter (EKF), and the posterior information includes the second state and the second covariance It is expressed as:

[0049]

[0050] In the formula, is the target state transfer matrix, Z k+1 is the radar measurement information, h() is the radar measurement function, and R k+1 are the process noise covariance and the measurement noise covariance, respectively. is the Jacobian matrix of the measurement function;

[0051] S604: After obtaining the posterior information of each sub-model of the target object, weighted fusion is performed in combination with the model recognition probability to obtain the final tracking result, which includes the third state X k+1k+1 and the third covariance P k+1k+1 , expressed as:

[0052]

[0053] The present invention determines the use conditions of the deep learning method in a logical judgment manner, and takes the tracking threshold as the distinguishing criterion. If the value is greater than the threshold, it means that the amount of input data has reached a certain value. By using formula 9-12, a higher tracking accuracy of the target object can be obtained. If the value does not reach a certain value, the extended Kalman filter algorithm (I MM-EKF) method is used to obtain the standard tracking accuracy of the target object.

[0054] For example, multiple sets of trajectory data from three different scenarios are used to train the model recognition network, and another three sets of trajectory data from the three scenarios are used to perform tracking tests to compare the performance of the method of the present invention with that of the traditional method.

[0055] 1. Data generation and network parameter setting

[0056] First, based on the Monte Carlo sample generation software, nine groups of trajectory data in three different scenarios are generated. The measurement noise covariance of the nine groups of data is consistent. Among them, the 1st to 3rd trajectories are uniform linear motion (CV) within 1-70 seconds, the 4th to 6th trajectories are uniform turning motion (CT) within 1-70 seconds, the 7th to 9th trajectories are uniform linear motion within 1-35 seconds, and uniform turning motion within 35.1s-70s. The other two trajectories are uniform motion in the first half and uniform turning motion in the second half. The radar sampling period is 0.01s. The measurement information of the trajectory data is used as input, and the motion model result is the output result. The model set of the target tracking system consists of CV model and CT model. When the target performs CV motion, its true output result is [1,0], and when the target performs CT motion, its true output result is [0,1]. Two of the three groups of trajectory data in different scenarios are taken as training sets, and another different trajectory data is taken as a validation set. The continuous 10-frame measurement is taken as a set of input information and brought into the network for training. The training set is divided into groups of 10 frames to obtain training samples. The training set and the validation set are standardized and preprocessed using equations (5) and (6). Based on the length of the training set sequence, the tracking threshold can be set to 10. The network model is optimized and trained based on the minimum variance criterion, and the parameter learning rate is set to 0.01. The number of iterative training is set to 1000 times.

[0057] 2. Network training

[0058] After the parameters are set, the training set data is brought into the network for training. After the network training is completed, the validation set data is brought into the network to obtain the model recognition results for use by the subsequent target tracking algorithm.

[0059] 3. Algorithm performance testing and comparison

[0060] The RNN recognition results are introduced into the EKF algorithm (RNN-EKF) to obtain the tracking results and compare them with the tracking results of the IMM-EKF algorithm. Figure 4-5 The north and east velocity tracking accuracy results and model recognition results of I MM-EKF and RNN-EKF for validation sample 1 (constant linear motion) are shown. It can be found that the model recognition accuracy and tracking accuracy of RNN-EKF are better than those of I MM-EKF, and the maneuvering delay of RNN-EKF is also smaller than that of I MM-EKF. From the tracking root mean square error (RMS) results, the velocity tracking accuracy of RNN-EKF in the north and east directions is improved by 4.3% and 6.2% respectively compared with I MM-EKF. In terms of model recognition accuracy, RNN-EKF is improved by 61.5% compared with I MM-EKF.

[0061] Based on the simulation software, the north and east speed tracking accuracy results and model recognition results of I MM-EKF and RNN-EKF for verification sample 2 (constant turning motion) are shown. It can be obtained that the model recognition accuracy and tracking accuracy of RNN-EKF are better than those of I MM-EKF, and the maneuvering delay of RNN-EKF is also smaller than that of I MM-EKF. From the tracking results, the speed tracking accuracy of RNN-EKF in the north and east directions is improved by 56% and 60.5% respectively compared with I MM-EKF. In terms of model recognition accuracy, RNN-EKF is improved by 55% compared with I MM-EKF.

[0062] Based on the simulation software, the north and east speed tracking accuracy results and model recognition results of I MM-EKF and RNN-EKF for verification sample 3 (uniform linear motion from 1 to 35 seconds and uniform turning motion from 35.1 seconds to 70 seconds) are shown. It can be found that the model recognition accuracy and tracking accuracy of RNN-EKF are better than those of I MM-EKF, and the maneuvering delay of RNN-EKF is also smaller than that of I MM-EKF. From the tracking results, the speed tracking accuracy of RNN-EKF in the north and east directions is improved by 52.9% and 46.1% respectively compared with I MM-EKF. In terms of model recognition accuracy, RNN-EKF is improved by 87.1% compared with I MM-EKF.

[0063] Therefore, compared with traditional methods, the deep learning model method proposed in the present invention can effectively improve the accuracy of target model recognition and target tracking accuracy.

[0064] The above is only a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present disclosure should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A model matching intelligent maneuvering target tracking algorithm, suitable for tracking a target by an aircraft radar, the radar including an interface control file and a GPS file, characterized in that: The methods include: S101: Matching the network model of the target object by using a recurrent neural network and a fully connected layer network; S102: The network model is trained and verification data is determined; S103: Tracking the target object based on the verification data.

2. The model matching intelligent maneuvering target tracking algorithm according to claim 1, characterized in that: S101 includes, Obtain parameter information fed back by the interface control file, the parameter information including distance information, angle measurement information and time series information, and obtain motion characteristics of the target object fed back by the GPS file; The parameter information is used as the input of the network model and the motion feature is used as the output of the network model, and the network model is trained to determine the parameters of the network model; The input of the network model is represented as [X1,X2,…X τ ], the output is expressed as [Y1,Y2,…Y τ ], the intermediate variable H of the network model i And the final output Y i It is expressed as: Where W X , W H and B H They are the input weight parameter, hidden weight parameter and bias parameter of the recurrent neural network RNN ​​respectively; W Y and B Y is the input weight parameter and bias parameter of the fully connected layer, H i-1 is the output result at the i-1th moment, H i is the output result at the i-th moment, Y i Output result of the network; X i is the input at the i-th moment; e is a constant.

3. The model matching intelligent maneuvering target tracking algorithm according to claim 2, characterized in that: Definition: When the time series length is τ and the hidden layer dimension is n, the total output dimension of the recurrent neural network RNN ​​is nτ, and the output dimension is o. The output information is regarded as vector information with dimension o. The input vector X = [X1, X2, ... X τ ], output vector Y = [Y1, Y2, ... Y τ ], the model recognition process of the recurrent neural network RNN ​​and the fully connected layer network is expressed as:

4. The model matching intelligent maneuvering target tracking algorithm according to claim 3, characterized in that: S102 includes: The input vector X and the output vector Y are linearly mapped to the interval between 0 and 1 to obtain the processed input vector Output vector Respectively expressed as: Where, X total Input information sample set, Y total To output the information sample set, max() represents the maximum value of the set, and min() represents the minimum value of the set; Input Vector Output vector Part of it is used as training data, and part of it is used as validation data.

5. The model matching intelligent maneuvering target tracking algorithm according to claim 4, characterized in that: S103 includes: Determine whether the length of the input vector X exceeds the tracking threshold length. If so, based on the output vector at time k+1 Weighted fusion is used to determine the tracking result of the target. o is the number of model sets. If not, the interacting multi-model extended Kalman filter algorithm is used to achieve target tracking.

6. The model matching intelligent maneuvering target tracking algorithm according to claim 5, characterized in that: Weighted fusion is used to determine the tracking results of the target, including: S601: Determine the output vector by the network model Model recognition probability satisfy: S602: The model recognition probability The prior information of the target object is obtained based on the Markov transfer matrix. The prior information includes the first state and the first covariance It is expressed as: In the formula, represents the transfer weight from the sub-model of the j-th target object to the sub-model of the i-th target object, π ij is the Markov transition probability from the jth target object sub-model to the i-th target object sub-model, and are the target state and covariance of the sub-model of the j-th target at time k; S603: After obtaining the prior information, obtain the posterior information of each sub-model of the target object based on the extended Kalman filter, and the posterior information includes the second state and the second covariance It is expressed as: In the formula, is the target state transfer matrix, Z k+1 is the radar measurement information, h() is the radar measurement function, and R k+1 are the process noise covariance and the measurement noise covariance, respectively. is the Jacobian matrix of the measurement function; S604: After obtaining the posterior information of each sub-model of the target object, weighted fusion is performed in combination with the model recognition probability to obtain the final tracking result, which includes the third state X k+1k+1 and the third covariance P k+1k+1 , expressed as:

Images