Depth maneuvering target tracking method based on target state space projection

Through the method of projection in the target state state space and deep neural network processing, the problem of inaccurate target position estimation in maneuver target tracking is solved, the tracking accuracy is improved, and the performance losses caused by measurement noise are avoided.

CN119006523AActive Publication Date: 2024-11-22XIDIAN UNIV
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411101901.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-22
Estimated Expiration
2044-08-12

AI Technical Summary

Technical Problem

When the existing data-driven maneuvering target tracking algorithm uses position measurement information, the target position estimation is inaccurate, and performance losses are caused by insufficient utilization of measurement noise distribution information.

Method used

Using the deep maneuver target tracking method based on target state space projection, the target track data is projected and preprocessed by constructing the target state space, and a multi-channel image data set is generated, and a deep neural network is input to obtain the target detection probability, confidence and Estimate the position, filter the high probability area, calculate the target position estimation coordinates, and obtain the filtering result through reverse projection.

Benefits of technology

It effectively improves the target tracking accuracy in maneuverable target scenarios, avoids performance losses caused by insufficient utilization of measurement noise distribution information, and improves the accuracy of target position estimation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006523A_ABST
    Figure CN119006523A_ABST
Patent Text Reader

Abstract

The invention discloses a depth maneuvering target tracking method based on target state space projection, relates to the technical field of radars, and effectively solves the problem of inaccurate target position estimation during target tracking of a maneuvering target in the prior art. The method comprises the following steps: acquiring target track data of a radar after signal processing and data association, and converting target measurement data into multi-channel image data according to a sliding window mode; and inputting the multi-channel image data into the trained deep neural network to obtain an estimated position of the target in the image, thereby obtaining position estimation of the target in a specific space, and completing maneuvering target tracking. According to the method, performance loss caused by insufficient utilization of measurement noise distribution information is avoided, and the target tracking precision in a maneuvering target scene is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a deep maneuvering target tracking method based on target state space projection. Background Art

[0002] Traditional model-driven methods describe target motion through mathematical models, but the target maneuver model based on statistical theory is difficult to fully reflect various specific situations and is still affected by model mismatch. Data-driven methods simulate target motion models that match these data through a large amount of data. If its simulation ability is strong enough, it can better solve complex problems in practice.

[0003] Most existing data-driven maneuvering target tracking algorithms only utilize position measurement information, resulting in inaccurate target position estimation during target tracking.

[0004] The urgent problem to be solved is how to accurately estimate the target position when tracking a maneuvering target. Summary of the invention

[0005] The present invention adopts a deep maneuvering target tracking method based on target state space projection, which effectively solves the problem of inaccurate target position estimation during maneuvering target tracking in the prior art, avoids performance loss caused by insufficient utilization of measurement noise distribution information, and effectively improves target tracking accuracy in maneuvering target scenarios.

[0006] The present invention provides a deep maneuvering target tracking method based on target state space projection, the method comprising:

[0007] Constructing a target state space according to the target measurement data, projecting the target track data into the target state space and performing preprocessing to obtain a multi-channel image data set; wherein each multi-channel image data in the multi-channel image data set is data obtained through a windowing process;

[0008] Traversing the multi-channel image data in the multi-channel image data set;

[0009] Inputting the multi-channel image data into a trained deep neural network to obtain an output result set; wherein the output result set includes: output results of the deep neural network in multiple regions of the multi-channel image data, and the output result of each region includes: target detection probability, confidence and target estimated position;

[0010] Filter the target detection probabilities corresponding to multiple regions that are greater than the given detection threshold; the number of regions greater than the given detection threshold must be greater than or equal to 1, otherwise the confidence level in the output result set must be Set to zero;

[0011] Determine the area corresponding to the maximum detection probability in the network output result, and determine the confidence level of the area corresponding to the maximum value The estimated position of the target corresponding to the area Will and As the input of the target position estimation formula, the target position estimation coordinates are calculated;

[0012] The position estimation coordinates are back-projected to obtain the filtered result of the target's true position;

[0013] Until the traversal is completed, the filtered result set is obtained.

[0014] In a possible implementation, the training process of the deep neural network includes:

[0015] S10, constructing a maneuvering target track data set, and projecting the maneuvering target track data set into a target state space for preprocessing, to obtain a preprocessed maneuvering target track data set;

[0016] S11, constructing a deep neural network and determining a cost function of the deep neural network;

[0017] S12, setting the batch value of the mini-batch gradient descent method and the update step size of the deep neural network parameters;

[0018] S13, randomly selecting a plurality of preprocessed maneuvering target track data sets from the preprocessed maneuvering target track data sets to form a small batch of data according to the batch value, and inputting the small batch of data into the deep neural network to calculate the cost function value corresponding to the small batch of data;

[0019] S14, updating the parameters of the deep neural network using the cost function value and the update step size;

[0020] S15, looping S13 to S14 until the cost function converges to obtain a trained deep neural network.

[0021] In a possible implementation, the step of constructing a maneuvering target track data set includes:

[0022] S1030-0, setting the maneuvering target track length N, the target motion model average switching time ΔT, and randomly generating the target initial position, initial speed, and initial signal-to-noise ratio;

[0023] S1030-1, determine the target, and randomly select a target motion model, process noise power q and target motion model switching time T for the target. s ;

[0024] S1030-2, calculating the next moment state of the target according to the target motion model, and calculating the target expected signal-to-noise ratio at time k according to the radar formula of the maneuvering target track; calculating the target measurement signal-to-noise ratio at time k according to the target fluctuation model and the target expected signal-to-noise ratio at time k; and calculating the target measurement value at time k according to the target measurement signal-to-noise ratio at time k;

[0025] S1030-3, based on the initial position and the target measurement value z at time k, k , construct the target track at time k, and determine whether the length of the target track at time k is greater than or equal to the preset length;

[0026] S1030-4, if yes, then output the target track at the time, if no, then determine whether the current time of the target is greater than or equal to the time T of switching the target motion model s If yes, loop S1030-1 to S1030-3, if no, execute S1030-2 to S1030-3 until the length of the target track at the time is greater than or equal to the preset length, and output the target track at the k time;

[0027] S1030-5, loop S1030-0 to S1030-4 to obtain a maneuvering target track data set.

[0028] In a possible implementation, the target state x at the next moment is k Specifically expressed as:

[0029] x k =Fx k-1 +w k-1 ,k=1,…,N

[0030] Among them, F represents the target state transfer matrix corresponding to the target motion model, x k represents the target state at time k, which includes the target position and target speed, w k represents the process noise at time k; N represents the track length of the maneuvering target.

[0031] In a possible implementation, preprocessing the maneuvering target track data set to obtain a preprocessed maneuvering target track data set includes:

[0032] Determine the window length L of the window, and measure the target sequence z in the maneuvering target track data set. 1:N Perform window processing to obtain a measurement sequence set within the window;

[0033] Calculating the in-window measurement sequences in the in-window measurement sequence set in sequence to obtain multi-channel image data corresponding to each in-window measurement sequence;

[0034] The multi-channel image data corresponding to the measurement sequences in each window are determined as the pre-processed maneuvering target track data set.

[0035] In a possible implementation, calculating the in-window measurement sequences in the in-window measurement sequence set to obtain multi-channel image data corresponding to each in-window measurement sequence includes:

[0036] S30, determining the measurement sequence within the window, and determining a first measurement error covariance matrix sequence corresponding to the measurement sequence within the window, and calculating a first channel image according to the measurement sequence within the window and the first measurement error covariance matrix sequence;

[0037] S31, determining the position corresponding to the first moment of the measurement sequence in the window, and determining the measurement signal-to-noise ratio corresponding to the first moment of the measurement sequence in the window; calculating a first filtering result according to the measurement sequence in the window and the interactive multiple model method; calculating the first filtering result, the measurement signal-to-noise ratio corresponding to the first moment, and the position corresponding to the first moment of the measurement sequence in the window, and calculating a second measurement error covariance matrix sequence; calculating a second channel image according to the second measurement error covariance matrix sequence and the first filtering result;

[0038] S32, using the measurement sequence in the window as the input of the time trajectory function of the target to obtain a second filtering result; and calculating a third measurement error covariance matrix sequence according to the second filtering result, the measurement signal-to-noise ratio corresponding to the first moment and the first moment corresponding position of the measurement sequence in the window; and calculating a third channel image according to the third measurement error covariance matrix sequence and the second filtering result;

[0039] S33, determining a second target position corresponding to the last moment of the measurement sequence in the window, and determining a first measurement error covariance matrix corresponding to the last moment of the first measurement error covariance matrix sequence, and calculating a fourth channel image using the second target position and the first measurement error covariance matrix;

[0040] S34, looping through steps S30 to S33 for the in-window measurement sequences in the in-window measurement sequence set to obtain multi-channel image data corresponding to each in-window measurement sequence.

[0041] In a possible implementation, the deep neural network includes: a backbone network, a neck network, and a head network;

[0042] The backbone network includes a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a first ELAN module, a first MP-1 module, a second ELAN module, a third MP-1 module, a third ELAN module, a fourth MP-1 module, and a third ELAN module connected in sequence.

[0043] The neck network includes a sequentially connected SPPCSPC module, a fifth CBS module, a first UPSample module, a first aggregation module, a first ELAN-W module, a sixth CBS module, a second UPSample module, a second aggregation module, a second ELAN-W module, a first MP-2 module, a third aggregation module, a third ELAN-W module, a second MP-2 module, a fourth aggregation module and a fourth ELAN-W module, the seventh CBS module is connected to the first aggregation module, and the eighth CBS module is connected to the second aggregation module;

[0044] The head network includes: a REP module and a CBM module connected in sequence;

[0045] The backbone network is interconnected with the neck network, and the head network is connected to the neck network;

[0046] The third ELAN module in the backbone network is connected to the SPPCSPC module in the neck network; the third ELAN module in the backbone network is connected to the seventh CBS module in the neck network; the first ELAN module in the backbone network is connected to the eighth CBS module in the neck network;

[0047] The first ELAN-W module in the neck network is connected to the third aggregation module.

[0048] In a possible implementation, the cost function of the deep neural network is specifically expressed as:

[0049] loss = λ d ·loss detecion +λ r ·loss regression +λ conf ·loss confidence +λ c ·loss constraint

[0050] Among them, λ d Represents the weight of the binary cross entropy loss function; loss detecion represents the binary cross entropy loss function; λ r Represents the weight of the mean square error loss function; loss regression represents the mean square error loss function; λconf Represents the weight of the entropy loss function; loss confidence represents the entropy loss function; λ c Represents the weight of the perplexity loss function; loss constraint represents the perplexity loss function.

[0051] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0052] The present invention preprocesses the target track data. Compared with the prior art, the multi-channel image data obtained by the present invention can convert the measurement sequence data into image data that is more conducive to network applications; through the trained deep neural network, the performance loss caused by insufficient utilization of measurement noise distribution information can be avoided, and the target tracking accuracy in the maneuvering target scenario can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A flowchart of the steps of a deep maneuvering target tracking method based on target state space projection provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of a deep neural network structure provided by an embodiment of the present invention;

[0055] Figure 3 This is a comparison result diagram of various tracking methods in the simulation embodiment provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The present invention provides a deep maneuvering target tracking method based on target state space projection, which includes the following steps S101 to S107.

[0058] S101, constructing a target state space according to target measurement data, projecting target track data into the target state space and performing preprocessing to obtain a multi-channel image data set; wherein each multi-channel image data in the multi-channel image data set is data obtained through a windowing process;

[0059] S102, traversing the multi-channel image data in the multi-channel image data set.

[0060] S103, input the multi-channel image data into the trained deep neural network to obtain an output result set; wherein the output result set includes: the output results of the deep neural network in multiple areas of the multi-channel image data, and the output results of each area include: target detection probability, confidence and target estimated position.

[0061] S104, screening the target detection probabilities corresponding to the multiple regions that are greater than the given detection threshold; wherein the number of regions greater than the given detection threshold must be greater than or equal to 1, otherwise the confidence level in the output result set must be Set to zero;

[0062] S105, determining the area corresponding to the maximum detection probability in the network output result, and determining the confidence level of the area corresponding to the maximum value The estimated position of the target corresponding to the region Will and As the input of the target position estimation formula, the target position estimation coordinates are calculated;

[0063] S106, the position estimation coordinates are back-projected to obtain a filtering result of the target's real position;

[0064] S107, until the traversal is completed, a filtering result set is obtained.

[0065] Specifically, in step S103, the training process of the deep neural network includes the following steps S1030 to S1035.

[0066] S1030, constructing a maneuvering target track data set, projecting the maneuvering target track data set into a target state space for preprocessing, and obtaining a preprocessed maneuvering target track data set;

[0067] S1031, constructing a deep neural network and determining a cost function of the deep neural network;

[0068] S1032, setting the batch value of the mini-batch gradient descent method and the update step size of the deep neural network parameters;

[0069] S1033, randomly selecting a plurality of preprocessed maneuvering target track data sets from the preprocessed maneuvering target track data set to form a small batch of data according to the batch value, and inputting the small batch of data into the deep neural network to calculate the cost function value corresponding to the small batch of data;

[0070] S1034, updating the parameters of the deep neural network using the cost function value and the update step size;

[0071] S1035, loop S1033 to S1034 until the cost function converges to obtain a trained deep neural network.

[0072] Specifically, in step S1030, a maneuvering target track data set is constructed, including the following steps S1030-0 to S1030-5.

[0073] S1030-0, setting the maneuvering target track length N, the target motion model average switching time ΔT, and randomly generating the target initial position, initial velocity, and initial signal-to-noise ratio;

[0074] S1030-1, determine the target, and randomly select a target motion model, process noise power q and target motion model switching time T for the target. s ;

[0075] S1030-2, calculate the state of the target at the next moment according to the target motion model; here, the state of the target at the next moment x k Specifically expressed as:

[0076] x k =Fx k-1 +w k-1 ,k=1,…,N (1)

[0077] Among them, F represents the target state transfer matrix corresponding to the target motion model, x k represents the target state at time k, which includes the target position and target speed, w k represents the process noise at time k; N represents the track length of the maneuvering target.

[0078] The expected signal-to-noise ratio of the target at time k is calculated according to the radar formula of the maneuvering target track; here, the expected signal-to-noise ratio is calculated by the radar equation, and the specific expected signal-to-noise ratio of the target at time k is expressed as:

[0079]

[0080] Among them, SNR k is the target expected signal-to-noise ratio at time k, ρ k is the distance between the target and the radar at time k.

[0081] According to the target fluctuation model and the target expected signal-to-noise ratio at time k, the target measurement signal-to-noise ratio at time k is calculated; here, the target fluctuation model is assumed to be Swerling I in the present invention, and the measurement signal-to-noise ratio is assumed to be 1W under the premise of assuming that the noise power is 1W. Subject to a parameter β k =1+SNR k (β k >0):

[0082]

[0083] Here, the distribution parameter β is determined k Then we can randomly sample from this distribution to get the measured signal-to-noise ratio at time k Combined measurement signal-to-noise ratio And the radar measurement accuracy formula:

[0084]

[0085] Where c is the speed of light, B is the bandwidth of the transmitted signal, and θ 3dB is the 3dB beam width. Through coordinate transformation, the measurement noise covariance matrix in the Cartesian coordinate system is obtained:

[0086]

[0087] The mean is set to 0, and the covariance matrix is ​​obtained by equation (5). The measurement noise of the covariance matrix is ​​v k , the measurement matrix in the Cartesian coordinate system is H. Through the target measurement equation, the measurement value at time k is obtained:

[0088] z k =Hx k +v k (6)

[0089] S1030-3, based on the initial position and the target measurement value z at time k k , construct the target track at time k, and determine whether the length of the target track at time k is greater than or equal to the preset length;

[0090] S1030-4, if yes, then output the target track z at the time 1:N , where z 1:N By measuring z at each moment k , k=1,…,N, if not, then determine whether the current target time is greater than or equal to the time T of switching the target motion model s If yes, then loop S1030-1 to S1030-3, if no, then execute S1030-2 to S1030-3 until the length of the target track at the time is greater than or equal to the preset length, and output the target track at the time k;

[0091] S1030-5, loop 1030-0 to 1030-4 to obtain the maneuvering target track data set

[0092] Specifically, in step S103, preprocessing the maneuvering target track data set to obtain a preprocessed maneuvering target track data set includes the following steps:

[0093] (1) Determine the window length L of the window, and calculate the target measurement sequence z in the maneuvering target track data set. 1:N Perform window processing to obtain a measurement sequence set within the window;

[0094] (2) The in-window measurement sequences in the in-window measurement sequence set are calculated in sequence to obtain multi-channel image data corresponding to each in-window measurement sequence; Here, the in-window measurement sequences in the in-window measurement sequence set are calculated to obtain multi-channel image data corresponding to each in-window measurement sequence, including S30 to S34.

[0095] S30, determining a measurement sequence within the window, and determining a first measurement error covariance matrix sequence corresponding to the measurement sequence within the window, and calculating a first channel image according to the measurement sequence within the window and the first measurement error covariance matrix sequence;

[0096] S31, determining the position corresponding to the first moment of the measurement sequence in the window, and determining the measurement signal-to-noise ratio corresponding to the first moment of the measurement sequence in the window; calculating a first filtering result according to the measurement sequence in the window and the interactive multiple model method; calculating the first filtering result, the measurement signal-to-noise ratio corresponding to the first moment, and the position corresponding to the first moment of the measurement sequence in the window, and calculating a second measurement error covariance matrix sequence; calculating a second channel image according to the second measurement error covariance matrix sequence and the first filtering result;

[0097] S32, using the measurement sequence in the window as the input of the time trajectory function of the target to obtain a second filtering result; and calculating a third measurement error covariance matrix sequence according to the second filtering result, the measurement signal-to-noise ratio corresponding to the first moment and the position corresponding to the first moment of the measurement sequence in the window; and calculating a third channel image according to the third measurement error covariance matrix sequence and the second filtering result;

[0098] S33, determining the second target position corresponding to the last moment of the measurement sequence in the window, and determining the first measurement error covariance matrix corresponding to the last moment of the first measurement error covariance matrix sequence, and calculating the fourth channel image using the second target position and the first measurement error covariance matrix;

[0099] S34, looping through steps S30 to S33 for the in-window measurement sequences in the in-window measurement sequence set to obtain multi-channel image data corresponding to each in-window measurement sequence.

[0100] (3) The multi-channel image data corresponding to the measurement sequence in each window is determined as the pre-processed maneuvering target track data set.

[0101] Exemplarily, 1) processing the maneuvering target measurement sequence z according to a sliding window method with a window length of L 1:N . By measuring the sequence in the window The corresponding measurement error covariance matrix sequence can be obtained by formula (4) and formula (5): in The measurement noise covariance matrix at each moment is composed of Σ k ,k=k1,…,k L , calculate the first channel image:

[0102]

[0103] By measuring the sequence in the window Calculate the filtering results of the interactive multi-model algorithm The measurement noise covariance matrix is ​​calculated by using the target position in the filtering result and the measurement signal-to-noise ratio at the first moment in the measurement sequence through formula (4) and formula (5): And use the above results to calculate the image of the second channel:

[0104]

[0105] By measuring the sequence in the window Calculate the filtering result of the time trajectory function of the target in the sliding window The measurement noise covariance matrix is ​​calculated by using the target position in the filtering result and the measurement signal-to-noise ratio at the first moment in the measurement sequence through formula (4) and formula (5): Where N s is the number of samples in the time period, and the above results are used to calculate the image of the third channel:

[0106]

[0107] The image of the fourth channel is calculated using the target position corresponding to the last moment in the window and the measurement noise covariance matrix:

[0108]

[0109] 2) Repeat 1) several times until the sliding window reaches the end of the maneuvering target measurement sequence, thus completing the preprocessing result of a maneuvering target track;

[0110] 3) Repeat 2) several times until the maneuvering target track data set is traversed All the tracks in the network training data set

[0111] Specifically, in step S103, the deep neural network includes: a backbone network, a neck network, and a head network;

[0112] The backbone network includes a first CBS (Convoltional Layer-Batch Normalization-SiLU Activation Function) module, a second CBS module, a third CBS module, a fourth CBS module, a first ELAN (Effective Layer Aggregation Network) module, a first MP-1 (Max Pooling-1) module, a second ELAN module, a third MP-1 module, a third ELAN module, a fourth MP-1 module, and a third ELAN module, which are connected in sequence.

[0113] The neck network includes a sequentially connected SPPCSPC (Spatial Pyramid Pooling Cross Stage Partial Convolutional) module, a fifth CBS module, a first UPSample (Upsampling Layer) module, a first aggregation module (cat), a first ELAN-W (Efficient Layer Aggregation Networks-Wide) module, a sixth CBS module, a second UPSample module, a second aggregation module (cat), a second ELAN-W module, a first MP-2 (Max Pooling-2) module, a third aggregation module (cat), a third ELAN-W module, a second MP-2 module, a fourth aggregation module and a fourth ELAN-W module, the seventh CBS module is connected to the first aggregation module (cat), and the eighth CBS module is connected to the second aggregation module (cat).

[0114] The head network includes: a REP (Reparameterizable Convolutional Layer) module and a CBM (Convolutional Layer-Batch Normalization-MishActivation Functinon) module connected in sequence.

[0115] The backbone network and the neck network are interconnected, the head network is connected to the neck network; the third ELAN module in the backbone network is connected to the SPPCSPC module in the neck network; the third ELAN module in the backbone network is connected to the seventh CBS module in the neck network; the first ELAN module in the backbone network is connected to the eighth CBS module in the neck network; the first ELAN-W module in the neck network is connected to the third aggregation module.

[0116] For example, Figure 2 As shown, a deep neural network consisting of a backbone network, a neck network, and a head network is constructed.

[0117] The backbone network adopts the CSPDarknet53 architecture including:

[0118] The first layer uses the first CBS module, with a convolution kernel size of 3 and a stride of 1;

[0119] The second layer uses the second CBS module with a convolution kernel size of 3 and a stride of 2;

[0120] The third layer uses the third CBS module with a convolution kernel size of 3 and a stride of 1;

[0121] The fourth layer uses the fourth CBS module with a convolution kernel size of 3 and a stride of 2;

[0122] The fifth layer uses the first ELAN module;

[0123] The sixth layer uses the first MP-1 module;

[0124] The seventh layer adopts the second ELAN module and outputs the first-level output of the feature map, such as Figure 2 As shown in ①;

[0125] The eighth layer uses the third MP-1 module;

[0126] The ninth layer adopts the third ELAN module and outputs the second level output of the feature map, such as Figure 2 As shown in ②;

[0127] The tenth floor uses the fourth MP-1 module;

[0128] The ninth layer adopts the third ELAN module and outputs the third level output of the feature map, such as Figure 2 As shown in ①.

[0129] The neck network combines the FPN and PANet structures, including:

[0130] The first layer uses the SPPCSPC module, which takes as input the third-level output of the backbone network feature map and also serves as the third-level output of the aggregation path, such as Figure 2 As shown in ④;

[0131] The second layer uses the fifth CBS module, with a convolution kernel size of 1 and a stride of 1;

[0132] The third layer uses the first UPSample module;

[0133] The fourth layer uses an aggregation module. The input of this layer is the second-level output of the backbone network feature map, which is passed through a CBS module with a convolution kernel size of 1 and a step size of 1 and the output of the third layer of the neck network;

[0134] The fifth layer uses the first ELAN-W module, which also serves as the second-level output of the aggregation path, such as Figure 2 As shown in ⑤;

[0135] The sixth layer often uses the sixth CBS module, with a convolution kernel size of 1 and a stride of 1;

[0136] The seventh layer uses the second UPSample module;

[0137] The eighth layer uses the second aggregation module. The input of this layer is the first-level output of the backbone network feature map, which is passed through a CBS module with a convolution kernel size of 1 and a step size of 1 and the output of the seventh layer of the neck network;

[0138] The ninth layer uses the second ELAN-W module, which also serves as the first-level output of the aggregation path, such as Figure 2 As shown in ⑥;

[0139] The tenth floor uses the first MP-2 module;

[0140] The eleventh layer uses the third aggregation module, and the input of this layer is the second-level output of the neck network aggregation path and the tenth layer input of the neck network;

[0141] The twelfth layer is the third ELAN-W module;

[0142] The thirteenth floor is the second MP-2 module;

[0143] The fourteenth layer is the fourth aggregation module, and the input of this layer is the third-level output of the neck network aggregation path and the thirteenth layer input of the neck network;

[0144] The 15th layer is the fourth ELAN-W module;

[0145] Independent seventh and eighth CBS modules.

[0146] The head network includes: the first layer adopts the REP module; the second layer adopts the CBM module.

[0147] Before determining the cost function of the deep neural network in step S1031, a composite cost function consisting of binary cross entropy, mean square error, entropy value and perplexity is used as the cost function loss(·) of the network. Specifically, the cost function consists of four parts, among which the first part:

[0148]

[0149] Where H'=H j ·r,W'=W j ·r, H j ,W j for j th The sample corresponds to the height and width, r is set in advance, here r = 1 / 32, N D is the number of training samples, is the detection probability output by the network, Output detection probability for the expected network;

[0150] Part II:

[0151]

[0152] Among them α As the expected network output confidence, Output confidence scores for the network:

[0153]

[0154] Part III:

[0155]

[0156] in is the desired output coordinate, for the fusion result The specific calculation method is as follows:

[0157]

[0158] in, is the local coordinate output by the network, i (m,j) is the global coordinate corresponding to region m, The filtering result of the interactive multi-model algorithm corresponds to the global coordinates on the image;

[0159] Part 4:

[0160]

[0161] in, |To find the cardinality of a set, is the global coordinate of the real target, d is the range of the constraint area, λe ,λ p is the weighting coefficient;

[0162] Specifically, in step S1031, the cost function of the deep neural network is specifically expressed as:

[0163] loss = λ d ·loss detecion +λ r ·loss regression +λ conf ·loss confidence +λ c ·loss constraint (17)

[0164] Among them, λ d Represents the weight of the binary cross entropy loss function; loss detecion represents the binary cross entropy loss function; λ r Represents the weight of the mean square error loss function; loss regression represents the mean square error loss function; λ conf Represents the weight of the entropy loss function; loss confidence represents the entropy loss function; λ c Represents the weight of the perplexity loss function; loss constraint represents the perplexity loss function.

[0165] For example, according to the network training data set and the cost function loss(·), and the deep neural network is trained using mini-batch stochastic gradient descent.

[0166] First, set the batch size B and the network parameter update step size η of the mini-batch gradient descent method;

[0167] Then, from the network training dataset Randomly select B training data to form a small batch, send it to the deep neural network and calculate the corresponding cost function loss(·);

[0168] Secondly, calculate the gradient g of the current deep neural network cost function loss(·) with respect to the network parameter θ;

[0169] Then, update the parameters of the deep neural network to θ-ηg;

[0170] Finally, the training process is repeated until the cost function loss(·) of the deep neural network tends to be constant, and a trained deep neural network is obtained.

[0171] In a specific embodiment provided by the present invention, target track data after signal processing and data association by radar is obtained;

[0172] According to S30 to S35, the target track data in the window is converted into multi-channel image data, and the corresponding coordinates of the filtering results of the interactive multi-model algorithm in the image are calculated.

[0173] Input the multi-channel image data into the trained deep neural network and get the network output:

[0174]

[0175] According to the probability of detection Determine whether the target is detected. If the detection probability is greater than the given threshold ∈, it is considered that the target exists, otherwise all Set to zero;

[0176] Under the premise that there is a target, select the area with the highest output detection probability:

[0177]

[0178] Use the confidence calculated by the network in this area and the estimated position of the target According to formula (15), the estimated target position on the image is calculated, and the distance Δd between the coordinate and the center of the image and the position measurement of the sequence head in the window are calculated. Add together to get k L The estimated result of the moment position;

[0179] Repeat several times until the sliding window reaches the end of the target track data, and the tracking filtering of the target track data is completed.

[0180] The effect of the present invention is further illustrated by the following simulation test:

[0181] (1) Simulation conditions

[0182] The hardware platform of the simulation experiment of the present invention: the processor is CPU AMD EPYC 7302 16-Core, the graphics card is GeForce RTX 3090, and the memory is 128GB; the software platform is: Ubuntu 22.04.1LTS, Python3.8;

[0183] The radar range resolution set in the simulation experiment of the present invention is 100m, the angle resolution is 1°, and the initial position of the target in the training scenario is assumed to be distributed at a distance of 150km to 400km from the radar, the initial azimuth is distributed in the range of 60° to 120°, the initial ground speed of the target is distributed in the range of 150m / s to 200m / s, the initial target motion direction is distributed in the range of 0° to 360°, and the initial target average signal-to-noise ratio is distributed in the range of 15dB to 25dB;

[0184] Assume that the target's scattering cross-sectional area fluctuation model is the Swerling I model, and assume that the process noise and measurement noise during simulation are Gaussian white noise;

[0185] Simulation 1 - Comparison of various tracking methods in a maneuvering target scenario.

[0186] According to the above-mentioned relevant parameters, the maneuvering target measurement sequence simulation is carried out, and 100 Monte Carlo experiments are carried out using the method of the present invention, the traditional interactive multi-model method, the deep LSTM method based on Bayesian filtering, the deep Bi-LSTM method and the Transformer method respectively, and the tracking result diagrams of each method are obtained. The results are shown in FIG. Figure 3 As shown, where: the horizontal axis is the time, starting at 10s to exclude the influence of initialization of each tracking method, and the vertical axis is the position estimation error of each method;

[0187] As can be seen from the figure, in the maneuvering target tracking scenario, the position estimation error of the method proposed in the present invention is lower than that of other methods most of the time, indicating that the maneuvering target tracking accuracy of the method of the present invention is higher.

[0188] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. All or part of the present invention can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0189] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the present invention.

Claims

1. A deep maneuvering target tracking method based on target state space projection, It is characterized in that include: Constructing a target state space according to the target measurement data, projecting the target track data into the target state space and performing preprocessing to obtain a multi-channel image data set; wherein each multi-channel image data in the multi-channel image data set is data obtained through a windowing process; Traversing the multi-channel image data in the multi-channel image dataset; Inputting the multi-channel image data into a trained deep neural network to obtain an output result set; wherein the output result set includes: output results of the deep neural network in multiple regions of the multi-channel image data, and the output result of each region includes: target detection probability, confidence and target estimated position; Filter the target detection probabilities corresponding to multiple regions that are greater than the given detection threshold; the number of regions greater than the given detection threshold must be greater than or equal to 1, otherwise the confidence level in the output result set must be Set to zero; Determine the area corresponding to the maximum detection probability in the network output result, and determine the confidence level of the area corresponding to the maximum value The estimated position of the target corresponding to the area Will and As the input of the target position estimation formula, the target position estimation coordinates are calculated; The position estimation coordinates are back-projected to obtain the filtered result of the target's true position; Until the traversal is completed, the filtering result set is obtained.

2. The deep maneuvering target tracking method based on target state space projection according to claim 1, It is characterized in that The training process of the deep neural network includes: S1030, constructing a maneuvering target track data set, and projecting the maneuvering target track data set into a target state space for preprocessing, to obtain a preprocessed maneuvering target track data set; S1031, constructing a deep neural network, and determining a cost function of the deep neural network; S1032, setting a batch value of a small batch gradient descent method and an update step size of the deep neural network parameters; S1033, randomly selecting a plurality of preprocessed maneuvering target track data sets from the preprocessed maneuvering target track data sets to form a small batch of data according to the batch value, and inputting the small batch of data into the deep neural network to calculate the cost function value corresponding to the small batch of data; S1034, updating the parameters of the deep neural network using the cost function value and the update step size; S1035, loop S1033 to S1034 until the cost function converges to obtain a trained deep neural network.

3. The deep maneuvering target tracking method based on target state space projection according to claim 2, It is characterized in that The step of constructing a maneuvering target track data set includes: S1030-0, setting the maneuvering target track length N, the target motion model average switching time ΔT, and randomly generating the target initial position, initial speed, and initial signal-to-noise ratio; S1030-1, determine the target, and randomly select a target motion model, process noise power q and target motion model switching time T for the target. s ; S1030-2, calculating the next moment state of the target according to the target motion model, and calculating the target expected signal-to-noise ratio at time k according to the radar formula of the maneuvering target track; calculating the target measurement signal-to-noise ratio at time k according to the target fluctuation model and the target expected signal-to-noise ratio at time k; and calculating the target measurement value at time k according to the target measurement signal-to-noise ratio at time k; S1030-3, based on the initial position and the target measurement value z at time k, k , construct the target track at time k, and determine whether the length of the target track at time k is greater than or equal to the preset length; S1030-4, if yes, then output the target track at the time, if no, then determine whether the current time of the target is greater than or equal to the time T of switching the target motion model s If yes, loop S1030-1 to S1030-3, if no, execute S1030-2 to S1030-3 until the length of the target track at the time is greater than or equal to the preset length, and output the target track at the k time; S1030-5, loop S1030-0 to S1030-4 to obtain a maneuvering target track data set.

4. The deep maneuvering target tracking method based on target state space projection according to claim 3, It is characterized in that The next state of the target x k Specifically expressed as: x k =Fx k-1 +w k-1 ,k=1,…,N Among them, F represents the target state transfer matrix corresponding to the target motion model, x k represents the target state at time k, which includes the target position and target speed, w k represents the process noise at time k; N represents the track length of the maneuvering target.

5. The deep maneuvering target tracking method based on target state space projection according to claim 2, It is characterized in that The preprocessing of the maneuvering target track data set to obtain the preprocessed maneuvering target track data set includes: Determine the window length L of the window, and measure the target sequence z in the maneuvering target track data set. 1:N Perform window processing to obtain a measurement sequence set within the window; sequentially calculating the in-window measurement sequences in the in-window measurement sequence set to obtain multi-channel image data corresponding to each in-window measurement sequence; The multi-channel image data corresponding to the measurement sequences in each window are determined as the pre-processed maneuvering target track data set.

6. The method for deep maneuvering target tracking based on target state space projection according to claim 5, It is characterized in that The calculating the in-window measurement sequences in the in-window measurement sequence set to obtain multi-channel image data corresponding to each in-window measurement sequence includes: S30, determining the measurement sequence within the window, and determining a first measurement error covariance matrix sequence corresponding to the measurement sequence within the window, and calculating a first channel image according to the measurement sequence within the window and the first measurement error covariance matrix sequence; S31, determining the position corresponding to the first moment of the measurement sequence in the window, and determining the measurement signal-to-noise ratio corresponding to the first moment of the measurement sequence in the window; calculating a first filtering result according to the measurement sequence in the window and the interactive multiple model method; calculating the first filtering result, the measurement signal-to-noise ratio corresponding to the first moment, and the position corresponding to the first moment of the measurement sequence in the window, and calculating a second measurement error covariance matrix sequence; calculating a second channel image according to the second measurement error covariance matrix sequence and the first filtering result; S32, using the measurement sequence in the window as the input of the time trajectory function of the target to obtain a second filtering result; and calculating a third measurement error covariance matrix sequence according to the second filtering result, the measurement signal-to-noise ratio corresponding to the first moment and the first moment corresponding position of the measurement sequence in the window; and calculating a third channel image according to the third measurement error covariance matrix sequence and the second filtering result; S33, determining a second target position corresponding to the last moment of the measurement sequence in the window, and determining a first measurement error covariance matrix corresponding to the last moment of the first measurement error covariance matrix sequence, and calculating a fourth channel image using the second target position and the first measurement error covariance matrix; S34, looping through steps S30 to S33 for the in-window measurement sequences in the in-window measurement sequence set to obtain multi-channel image data corresponding to each in-window measurement sequence.

7. The deep maneuvering target tracking method based on target state space projection according to claim 1, It is characterized in that The deep neural network includes: a backbone network, a neck network, and a head network; The backbone network includes a first CBS module, a second CBS module, a third CBS module, a fourth CBS module, a first ELAN module, a first MP-1 module, a second ELAN module, a third MP-1 module, a third ELAN module, a fourth MP-1 module, and a third ELAN module connected in sequence. The neck network includes a sequentially connected SPPCSPC module, a fifth CBS module, a first UPSample module, a first aggregation module, a first ELAN-W module, a sixth CBS module, a second UPSample module, a second aggregation module, a second ELAN-W module, a first MP-2 module, a third aggregation module, a third ELAN-W module, a second MP-2 module, a fourth aggregation module and a fourth ELAN-W module, the seventh CBS module is connected to the first aggregation module, and the eighth CBS module is connected to the second aggregation module; The head network includes a REP module and a CBM module connected in sequence; The third ELAN module in the backbone network is connected to the SPPCSPC module in the neck network; the third ELAN module in the backbone network is connected to the seventh CBS module in the neck network; the first ELAN module in the backbone network is connected to the eighth CBS module in the neck network; The first ELAN-W module in the neck network is connected to the third aggregation module.

8. The method for deep maneuvering target tracking based on target state space projection according to claim 1, It is characterized in that The cost function of the deep neural network is specifically expressed as: loss=λ d ·loss detecion +λ r ·loss regression +λ conf ·loss confidence +λ c ·loss constraint Among them, λ d Represents the weight of the binary cross entropy loss function; loss detecion represents the binary cross entropy loss function; λ r Represents the weight of the mean square error loss function; loss regression represents the mean square error loss function; λ conf Represents the weight of the entropy loss function; loss confidence represents the entropy loss function; λ c Represents the weight of the perplexity loss function; loss constraint represents the perplexity loss function.

Citation Information

Patent Citations

  • Before-radar-target-detecting tracking method based on track enhancing

    CN104297748A

  • Circumferential SAR moving target tracking method based on GMPHD filter

    CN109934840A

  • Space high-maneuverability target tracking method based on depth correlation

    CN114063073A

  • Radar target tracking algorithm in maneuvering target scene and electronic equipment

    CN115685128A

  • Maneuvering target tracking method based on long-term and short-term memory network

    CN116449360A