Deep maneuvering target tracking method based on target state space projection

Abstract

The application discloses a kind of depth maneuvering target tracking methods based on target state space projection, it is related to radar technical field, effectively solve the problem of inaccurate target position estimation when maneuvering target is tracked in prior art, the method comprises: obtaining the target track data after radar signal processing and after data association, target measurement data is converted into multi-channel image data according to the mode of sliding window;Multi-channel image data is input into the trained deep neural network, and the estimated position of target in image is obtained, so as to obtain the position estimation of target in specific space, complete maneuvering target tracking;The present method avoids the performance loss caused by insufficient utilization of measurement noise distribution information, effectively improves the target tracking accuracy in maneuvering target scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radar technology, and particularly relates to a deep maneuvering target tracking method based on target state space projection. BACKGROUND

[0002] The traditional model-driven method describes target motion through mathematical models, but the target maneuvering model based on statistical theory is difficult to comprehensively reflect various specific situations and is still affected by model mismatch. The data-driven method simulates a target motion model that fits the data through a large amount of data, and if the simulation capability is strong enough, it can solve complex problems in practice.

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

[0004] The problem to be solved at present is how to accurately estimate the target position during maneuvering target tracking. SUMMARY

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

[0006] The present application provides a deep maneuvering target tracking method based on target state space projection, which comprises the following steps:

[0007] Constructing a target state space according to target measurement data, projecting target track data into the target state space and 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 by one windowing processing;

[0008] Iterating 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 comprises the output results of the deep neural network in multiple regions of the multi-channel image data, and the output result of each region comprises target detection probability, confidence and target estimated position;

[0010] Screening regions with target detection probability greater than a given detection threshold from the target detection probability corresponding to multiple regions; wherein the number of regions with target detection probability greater than the given detection threshold needs to be greater than or equal to 1, otherwise the confidence in the output result set needs to be set to zero; ​

[0011] determining a region corresponding to a maximum value of the detection probability in the network output result, and determining a confidence degree of the region corresponding to the maximum value a target estimated position corresponding to the region taking and as an input of a target position estimation formula, calculating a target position estimation coordinate;

[0012] the position estimation coordinate is a filtering result of a target real position obtained through back projection;

[0013] until the end of the iteration, a filtering result set is obtained.

[0014] In a possible implementation, the training process of the deep neural network comprises the following steps.

[0015] S10, constructing a maneuvering target track data set, and projecting the maneuvering target track data set to 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 a batch value of a small batch gradient descent method and an update step length of a parameter of the deep neural network;

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

[0019] S14, using the cost function value and the update step length to update the parameter of the deep neural network;

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

[0021] In a possible implementation, the constructing a maneuvering target track data set comprises the following steps.

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

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

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

[0025] S1030-3, constructing a target track at the kth moment according to the initial position and the measurement value of the target at the kth moment z k , and determining whether a length of the target track at the kth moment is greater than or equal to a preset length;

[0026] S1030-4, if yes, outputting the target track at the kth moment, and if no, determining whether a current time of the target is greater than or equal to a time T s at which the target motion model is switched, if yes, repeating S1030-1 to S1030-3, and if no, executing S1030-2 to S1030-3, until the length of the target track at the kth moment is greater than or equal to the preset length, and outputting the target track at the kth moment;

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

[0028] In a possible implementation, the next time state of the target x k is specifically represented as:

[0029] x k = Fx k-1 + w k-1 , k = 1, …, N

[0030] wherein F represents a target state transition matrix corresponding to the target motion model, x k represents a target state at the kth moment, the target state includes a target position and a target velocity, w k represents process noise at the kth moment; and N represents a length of the maneuvering target track.

[0031] In a possible implementation, the pre-processing of the maneuvering target track data set to obtain a pre-processed maneuvering target track data set comprises:

[0032] determining a window length L of windowing, performing windowing processing on a target measurement sequence z 1:N in the maneuvering target track data set to obtain a windowed measurement sequence set;

[0033] sequentially calculating a windowed measurement sequence in the windowed measurement sequence set to obtain multi-channel image data corresponding to each windowed measurement sequence.

[0034] The multi-channel image data corresponding to each of the window measurement sequences is determined as the preprocessed maneuver target track data set.

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

[0036] S30, the window measurement sequence is determined, and a first measurement error covariance matrix sequence corresponding to the window measurement sequence is determined, and a first channel image is calculated according to the window measurement sequence and the first measurement error covariance matrix sequence.

[0037] S31, a first time corresponding position of the window measurement sequence is determined, and a first time corresponding measurement signal-to-noise ratio of the window measurement sequence is determined, a first filtering result is calculated according to the window measurement sequence and an interacting multiple model method, a second measurement error covariance matrix sequence is calculated according to the first filtering result, the first time corresponding measurement signal-to-noise ratio and the first time corresponding position of the window measurement sequence, and a second channel image is calculated according to the second measurement error covariance matrix sequence and the first filtering result.

[0038] S32, the window measurement sequence is taken as an input of a time trajectory function of a target to obtain a second filtering result, a third measurement error covariance matrix sequence is calculated according to the second filtering result, the first time corresponding measurement signal-to-noise ratio and the first time corresponding position of the window measurement sequence, and a third channel image is calculated according to the third measurement error covariance matrix sequence and the second filtering result.

[0039] S33, a second target position corresponding to a last time of the window measurement sequence is determined, and a first measurement error covariance matrix corresponding to a last time of the first measurement error covariance matrix sequence is determined, and a fourth channel image is calculated by using the second target position and the first measurement error covariance matrix.

[0040] S34, the steps S30 to S33 are circularly performed on the window measurement sequence in the window measurement sequence set to obtain the multi-channel image data corresponding to each of the window measurement sequences.

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

[0042] The backbone network comprises 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 comprises a 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 connected in sequence, and a seventh CBS module connected to the first aggregation module and an eighth CBS module connected to the second aggregation module.

[0044] The head network comprises a REP module and a CBM module connected in sequence.

[0045] The backbone network and the neck network are connected to each other, 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; and 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 represented as:

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

[0050] Wherein, λ d represents a binary cross-entropy loss function weight; loss detecion represents a binary cross-entropy loss function; λ r represents a mean square error loss function weight; loss regression represents a mean square error loss function; λconf denotes the entropy loss function weight; loss confidence denotes the entropy loss function; lambda c denotes the perplexity loss function weight; loss constraint denotes the perplexity loss function.

[0051] The one or more technical solutions provided in the application have at least the following technical effects or advantages:

[0052] The target track data is preprocessed in the application, and the multi-channel image data obtained by the application can convert the measurement sequence data into image data more suitable for network application, relative to the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A deep maneuvering target tracking method based on target state space projection is provided for the embodiments of the application.

[0054] Figure 2 A deep neural network structure diagram is provided for the embodiments of the application.

[0055] Figure 3 A comparison result diagram of various tracking methods in a simulation embodiment is provided for the embodiments of the application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

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

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

[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 comprises: output results of the deep neural network in multiple regions of the multi-channel image data, and each region's output result comprises: a target detection probability, a confidence, and a target estimation position.

[0061] S104, screen the regions whose target detection probabilities are greater than a given detection threshold; wherein the number of regions whose target detection probabilities are greater than the given detection threshold needs to be greater than or equal to 1, otherwise the confidence in the output result set needs to be set to zero.

[0062] S105, determine the region corresponding to the maximum detection probability in the network output result, and determine the confidence of the region corresponding to the maximum value and the target estimation position of the region as the input of the target position estimation formula, and calculate the target position estimation coordinates;

[0063] S106, the position estimation coordinates are used to obtain the filtering result of the target real position through back projection;

[0064] S107, until the end of the iteration, obtain the filtering result set.

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

[0066] S1030, construct a maneuvering target track data set, project the maneuvering target track data set to a target state space for preprocessing to obtain a preprocessed maneuvering target track data set;

[0067] S1031, construct a deep neural network and determine a cost function of the deep neural network;

[0068] S1032, set a batch value of the mini-batch gradient descent method and an update step length of the deep neural network parameters;

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

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

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

[0072] Specifically, in step S1030, the 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, randomly generating the target initial position, initial speed and initial signal-to-noise ratio;

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

[0075] S1030-2, calculating the next time state of the target according to the target motion model; here, the next time state of the target x k is specifically represented as:

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

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

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

[0079]

[0080] Wherein, SNR k is the target expected signal-to-noise ratio at time k, and ρ k is the target distance from the radar distance 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, it is assumed that the target fluctuation model in the application is Swerling I, and under the premise that the noise power is 1W, the measurement signal-to-noise ratio obeys an exponential distribution with a parameter β k =1+SNR k (β k >0):

[0082]

[0083] Here, the parameter β of the distribution is determined k Then, the measurement SNR at time k can be randomly sampled from the distribution Combining the measurement SNR And the radar measurement accuracy formula is:

[0084]

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

[0086]

[0087] Set the mean to 0, and the measurement noise of the covariance matrix obtained from equation (5) is v k , and 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, according to the initial position and the target measurement value z k at time k, the target track at time k is constructed, and it is determined whether the length of the target track at time k is greater than or equal to the preset length;

[0090] S1030-4, if yes, output the target track z 1:N at time k, where z 1:N is composed of the measurement z k at each time, k = 1, …, N, if not, it is determined whether the current time of the target is greater than or equal to the time T s switching the target motion model, if yes, the steps S1030-1 to S1030-3 are cycled, if not, the steps S1030-2 to S1030-3 are executed until the length of the target track at time k is greater than or equal to the preset length, and the target track at time k is outputted;

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

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

[0093] (1) determining a window length L of the windowing, and determining a target measurement sequence z in the target track data set of the maneuvering target 1:N performing the windowing to obtain a measurement sequence set in the window;

[0094] (2) sequentially calculating the measurement sequence in the measurement sequence set in the window to obtain the multi-channel image data corresponding to each measurement sequence in the window; here, the calculation of the measurement sequence in the measurement sequence set in the window to obtain the multi-channel image data corresponding to each measurement sequence in the window includes S30 to S34.

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

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

[0097] S32, taking the measurement sequence in the window as the input of the time trajectory function of the target to obtain the second filtering result; and calculating the third measurement error covariance matrix sequence according to the second filtering result, the first time corresponding measurement signal-to-noise ratio and the first time corresponding position of the measurement sequence in the window; calculating the 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 time of the measurement sequence in the window, and determining the first measurement error covariance matrix corresponding to the last time 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, repeating steps S30 to S33 for the measurement sequence in the measurement sequence set in the window to obtain the multi-channel image data corresponding to each measurement sequence in the window.

[0100] (3) determining the multi-channel image data corresponding to each measurement sequence in the window as the preprocessed maneuvering target track data set.

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

[0102]

[0103] The measurement sequence in the window is used The IMM filtering result is calculated The measurement noise covariance matrix is calculated by formula (4) and formula (5) using the target position in the filtering result and the measurement signal-to-noise ratio at the first time in the measurement sequence The second channel image is calculated using the above result:

[0104]

[0105] The measurement sequence in the window is used The filtering result of the target time trajectory function in the sliding window is calculated The measurement noise covariance matrix is calculated by formula (4) and formula (5) using the target position in the filtering result and the measurement signal-to-noise ratio at the first time in the measurement sequence Wherein N s The third channel image is calculated using the above result:

[0106]

[0107] The fourth channel image is calculated using the target position corresponding to the last time 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, that is, the preprocessing result of a maneuvering target track is completed;

[0110] 3) Repeat 2) several times until all tracks in the maneuvering target track data set are traversed

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

[0112] ​The backbone network comprises first, second, third, fourth CBS (Convoltional Layer-Batch Normalization-SiLU Activation Function) modules connected in sequence, 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.

[0113] The neck network comprises an SPPCSPC (Spatial Pyramid Pooling Cross Stage Partial Convolutional) module, a fifth CBS module, a first UP Sample (Upsampling Layer) module, a first aggregation module (cat), a first ELAN-W (Efficient Layer Aggregation Networks-Wide) module, a sixth CBS module, a second UP Sample 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 (cat), and a fourth ELAN-W module connected in sequence, wherein a seventh CBS module is connected to the first aggregation module (cat), and an eighth CBS module is connected to the second aggregation module (cat).

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

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

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

[0117] The backbone network adopts a CSPDarknet53 architecture, which includes:

[0118] The first layer adopts a first CBS module, and the convolution kernel size is 3 and the step is 1;

[0119] The second layer adopts a second CBS module, and the convolution kernel size is 3 and the step is 2;

[0120] The third layer adopts a third CBS module, and the convolution kernel size is 3 and the step is 1;

[0121] The fourth layer adopts a fourth CBS module, and the convolution kernel size is 3 and the step is 2;

[0122] The fifth layer adopts a first ELAN module;

[0123] The sixth layer adopts a first MP-1 module;

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

[0125] The eighth layer adopts a third MP-1 module;

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

[0127] The tenth layer adopts a fourth MP-1 module;

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

[0129] The neck network combines FPN and PANet structures, which includes:

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

[0131] The second layer adopts a fifth CBS module, with a convolution kernel size of 1 and a step size of 1;

[0132] The third layer adopts a first UP Sample module;

[0133] The fourth layer adopts an aggregation module, and the input of the layer is the second-level output of the backbone network feature map after 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 adopts a first ELAN-W module, which is also the second-level output of the aggregation path, as shown in (5) of the description; Figure 2

[0135] The sixth layer commonly adopts a sixth CBS module, with a convolution kernel size of 1 and a step size of 1;

[0136] The seventh layer adopts a second UP Sample module;

[0137] The eighth layer adopts a second aggregation module, and the input of the layer is the first-level output of the backbone network feature map after 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 adopts a second ELAN-W module, which is also the first-level output of the aggregation path, as shown in (6) of the description; Figure 2

[0139] The tenth layer adopts a first MP-2 module;

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

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

[0142] The thirteenth layer is a second MP-2 module;

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

[0144] The fifteenth layer is a fourth ELAN-W module;

[0145] An independent seventh CBS module and an eighth CBS module.

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

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

[0148]

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

[0150] The second part is:

[0151]

[0152] wherein o α is the expected output confidence, is the output confidence of the network:

[0153]

[0154] The third part is:

[0155]

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

[0157]

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

[0159] The fourth part is:

[0160]

[0161] wherein, is the set cardinality operation, is the global coordinate of the real target, d is the range of the constraint region, and λe ,λ p is a weighting coefficient;

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

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

[0164] Wherein, λ d represents the binary classification cross-entropy loss function weight; loss detecion represents the binary classification cross-entropy loss function; λ r represents the mean square error loss function weight; loss regression represents the mean square error loss function; λ conf represents the entropy loss function weight; loss confidence represents the entropy loss function; λ c represents the perplexity loss function weight; loss constraint represents the perplexity loss function.

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

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

[0167] Then, B training data are randomly selected from the network training data set to form a mini-batch, which is sent into the deep neural network and the corresponding cost function loss(·) is calculated;

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

[0169] Then, the parameters of the deep neural network are updated 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 the trained deep neural network is obtained.

[0171] In one specific embodiment provided by the present application, target track data after signal processing and data association of the radar are obtained;

[0172] The process, including steps S30 to S35, involves converting the target trajectory data within the window into multi-channel image data and calculating the corresponding coordinates of the filtering result of the interactive multi-model algorithm in the image.

[0173] Multi-channel image data is input into a trained deep neural network to obtain the network output:

[0174]

[0175] Based on detection probability To determine whether a target has been detected, if the detection probability is greater than a given threshold ∈, then the target is considered to exist; otherwise, all targets are excluded. Set to zero;

[0176] Given a target, select the region with the highest output detection probability:

[0177]

[0178] The confidence level calculated by the network in this area is compared with the estimated location of the target. The target position estimate on the image is calculated according to formula (15), and the distance Δd between the coordinate and the image center is calculated, along with the position measurement of the sequence head within the window. Add them together to get k L Time and position estimation results;

[0179] Repeat this process several times until the sliding window reaches the end of the target track data, thus completing the target track data tracking and filtering.

[0180] The effects of this invention are further illustrated by the following simulation experiments:

[0181] (1) Simulation conditions

[0182] The hardware platform for the simulation experiment of this invention is: CPU AMD EPYC 7302 16-Core, graphics card GeForce RTX 3090, and memory 128GB; the software platform is: Ubuntu 22.04.1LTS and Python 3.8.

[0183] In the simulation experiment of this invention, the radar range resolution is set to 100m and the angular resolution to 1°. In the training scenario, it is assumed that the initial position of the target is distributed at a distance of 150km to 400km from the radar, the initial azimuth angle is distributed in the range of 60° to 120°, the initial ground velocity of the target is distributed in the range of 150m / s to 200m / s, the initial target motion orientation is distributed in the range of 0° to 360°, and the initial average signal-to-noise ratio of the target is distributed in the range of 15dB to 25dB.

[0184] Assuming that the fluctuation model of the scattering cross-section area of the target is a Swerling I model, and assuming that the process noise and the measurement noise during simulation are Gaussian white noise;

[0185] Simulation 1 - Comparison of various tracking methods in the maneuvering target scene.

[0186] According to the aforementioned related parameters, a maneuvering target measurement sequence simulation is performed, 100 times of Monte Carlo tests are respectively performed using the method of the present application, a traditional interacting multiple model method, a deep LSTM method based on Bayesian filtering, a deep Bi-LSTM method, and a Transformer method, tracking result graphs of each method are obtained, and the results are as shown in Figure 3 The horizontal axis is the time, 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 scene, the position estimation error of the method of the present application is lower than that of other methods in most of the time, indicating that the maneuvering target tracking precision of the method of the present application is higher.

[0188] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment mainly explains the difference from other embodiments. The whole or part of the present application 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, small computers, large computers, distributed computing environments including any of the above systems or devices, and the like.

[0189] The above embodiments are only used to illustrate the technical solutions of the present application, and are not limited to the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the present application.

Claims

1. A deep maneuvering target tracking method based on target state space projection, characterized in that, The method comprises the following steps: According to the target measurement data, a target state space is constructed, target track data is projected to the target state space and preprocessed to obtain a multi-channel image data set; each multi-channel image data in the multi-channel image data set is data obtained through one windowing processing; the preprocessing includes determining a window length of windowing The target measurement sequence in the target track data is subjected to windowing processing to obtain a windowed measurement sequence set; the windowed measurement sequences in the windowed measurement sequence set are sequentially calculated to obtain multi-channel image data corresponding to each windowed measurement sequence; the calculation includes: S30, determining the windowed measurement sequence and determining a first measurement error covariance matrix sequence corresponding to the windowed measurement sequence, calculating a first channel image according to the windowed measurement sequence and the first measurement error covariance matrix sequence; S31, determining a first time corresponding position of the windowed measurement sequence and determining a first time corresponding measurement signal-to-noise ratio of the windowed measurement sequence; calculating a first filtering result according to the windowed measurement sequence and an interacting multiple model method; calculating the first filtering result, the first time corresponding measurement signal-to-noise ratio and the first time corresponding position of the windowed measurement sequence to calculate 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, taking the windowed measurement sequence as an input of a target time trajectory function to obtain a second filtering result; and calculating a third measurement error covariance matrix sequence according to the second filtering result, the first time corresponding measurement signal-to-noise ratio and the first time corresponding position of the windowed measurement sequence; 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 a last time of the windowed measurement sequence and determining a first measurement error covariance matrix corresponding to a last time of the first measurement error covariance matrix sequence, calculating a fourth channel image using the second target position and the first measurement error covariance matrix; S34, repeatedly performing steps S30 to S33 on the windowed measurement sequences in the windowed measurement sequence set to obtain multi-channel image data corresponding to each windowed measurement sequence; The multi-channel image data corresponding to each window measurement sequence is determined as a pre-processed maneuvering target track data set, and the multi-channel image data in the multi-channel image data set is traversed; The multi-channel image data is input into the trained deep neural network to obtain an output result set; wherein the output result set comprises the output results of the deep neural network in multiple regions of the multi-channel image data, and the output result of each region comprises a target detection probability, a confidence and a target estimation position; Screening the regions corresponding to the target detection probability greater than the given detection threshold; wherein the number of regions greater than the given detection threshold needs to be greater than or equal to 1, otherwise the confidence of the output result set needs to be set to 0 Set to zero; determining a region corresponding to a maximum detection probability in a network output result, and determining a confidence of the region corresponding to the maximum and an estimated position of the target corresponding to the region , as an input of a target position estimation formula, calculating an estimated coordinate of the target position and as an input of a target position estimation formula, calculating an estimated coordinate of the target position The position estimation coordinates are obtained by back projection to obtain a filtering result of the target real position; Until the traversal is completed, a filtering result set is obtained.

2. The target state space projection based deep maneuvering target tracking method according to claim 1, characterized in that, The training process of the deep neural network comprises the following steps: S1030, a maneuvering target track data set is constructed, and the maneuvering target track data set is projected to a target state space for pre-processing to obtain a pre-processed maneuvering target track data set; S1031, a deep neural network is constructed, and a cost function of the deep neural network is determined; S1032, a batch value of a small batch gradient descent method and an update step length of the parameters of the deep neural network are set; S1033, according to the batch value, a plurality of pre-processed maneuvering target track data sets are randomly selected from the pre-processed maneuvering target track data set to form a small batch data, and the small batch data is input into the deep neural network to calculate a cost function value corresponding to the small batch data; S1034, the parameters of the deep neural network are updated by using the cost function value and the update step length; S1035, S1033 to S1034 are cycled until the cost function converges, and a trained deep neural network is obtained.

3. The deep maneuvering target tracking method based on target state space projection according to claim 2, characterized in that, The construction of the maneuvering target track data set comprises the following steps: S1030-0, set the length of the maneuvering target track , average switching time of target motion model , randomly generate initial position, initial velocity and initial signal-to-noise ratio of the target S1030-1, determine a target, and randomly select a target motion model, process noise power for the target and switch target motion model time ; S1030-2, calculating the next time state of the target according to the target motion model, and calculating the target state according to the radar formula of the maneuvering target track the target expected SNR at the time t; calculating the target expected SNR at the time t according to the target fluctuation model and the target expected SNR at the time t, calculating the target expected SNR at the time t according to the target fluctuation model and the target measurement SNR at the time t; calculating the target measurement SNR at the time t according to the target fluctuation model and the target measurement SNR at the time t, calculating the target measurement SNR at the time t according to the target fluctuation model and the target measurement value at the time t; calculating the target measurement value at the time t according to the target fluctuation model and S1030-3, based on the initial position and Target measurement value at time 1 , build Record the target trajectory at any time and determine Is the length of the target track at any given time greater than or equal to the preset length? S1030-4, if yes, output the target track at the time, if no, determine whether the target current time is greater than or equal to the time for switching the target motion model , 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 time S1030-5, S1030-0 to S1030-4 are cycled 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, characterized in that, the target next time state is specifically represented as: wherein, denotes a target state transition matrix corresponding to a target motion model, denotes a target state at time k, the target state including a target position and a target velocity, denotes process noise at time k; denotes a length of a maneuvering target track.

5. The target state space projection based deep maneuvering target tracking method according to claim 1, wherein, The deep neural network comprises a backbone network, a neck network and a head network; The backbone network comprises 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 and a fourth MP-1 module which are connected in sequence; The neck network comprises a 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 which are connected in sequence, and a seventh CBS module is connected to the first aggregation module, and an eighth CBS module is connected to the second aggregation module; The head network comprises a REP module and a CBM module which are connected in sequence. The third ELAN module in the backbone network is connected with the SPPCSPC module in the neck network; the third ELAN module in the backbone network is connected with the seventh CBS module in the neck network; the first ELAN module in the backbone network is connected with the eighth CBS module in the neck network; The first ELAN-W module in the neck network is connected with the third aggregation module.

6. The target state space projection based deep maneuvering target tracking method according to claim 1, wherein, The cost function of the deep neural network is specifically represented as: wherein, represents a binary cross-entropy loss function weight; represents a binary cross-entropy loss function; represents a mean squared error loss function weight; represents a mean squared error loss function; represents an entropy loss function weight; represents an entropy loss function; represents a perplexity loss function weight; represents a perplexity loss function.

Images