A maneuvering target tracking optimization method based on deep learning

Abstract

The present invention discloses a maneuvering target tracking optimization method based on deep learning. Under the framework of Gaussian mixture probability hypothesis density filtering algorithm, the present invention firstly constructs a measurement model that introduces Doppler information, and performs Taylor series expansion at the predicted value to convert the nonlinear equation into a linear one, thereby improving the tracking performance while avoiding the increase in the amount of calculation; secondly, an adaptive turning rate estimation network is designed, and the multi-dimensional target motion state time series information is used to realize the real-time estimation of the turning rate at the current moment, so as to realize the update of the parameters of the collaborative turning filter model, improve the matching degree of the filter model, and reduce the maneuvering target tracking error.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile multi-target tracking, and in particular to a mobile target tracking optimization method based on deep learning. Background Art

[0002] As the autonomous capabilities and intelligence levels of drones continue to improve, they play an important role in many fields. They can perform military tasks such as search, reconnaissance, surveillance and strike, and can also be applied to civilian fields such as agricultural plant protection, geological exploration, and environmental testing. Therefore, the detection of drone targets is of great significance for ensuring air route safety and urban security. Radar is an important technical means for drone detection, but drone targets usually fly at low altitudes and have strong maneuverability. In addition, the detection environment is complex, which brings great challenges to the stable tracking of targets. Summary of the invention

[0003] In view of this, the present invention provides a maneuvering target tracking optimization method based on deep learning, which can achieve high-precision tracking of UAV targets in cluttered environments and improve track quality.

[0004] A deep learning-based maneuvering target tracking optimization method of the present invention comprises the following steps:

[0005] Construct a measurement model that introduces Doppler information, specifically: measure z k,j From state x k,i The measurement equation is

[0006] In the formula, is zero mean and covariance Gaussian white noise; They are position measurement and Doppler measurement values ​​respectively;

[0007] Based on the adaptive turning rate estimation network, the multi-dimensional target motion state time series is used to estimate the turning rate at the current moment and complete the optimization; wherein the input of the network is the feature matrix Fv containing the multi-dimensional target motion state time series information: Among them, the eigenvector at time k is expressed as k∈[1,K], where x k and k represents the target position estimate at time k, and represents the target velocity estimate at time k, Indicates Doppler measurement.

[0008] The measurement equation that introduces the Doppler measurement information is converted into a linear equation by performing Taylor series expansion at the predicted value.

[0009] Among them, the multi-dimensional target motion state time series information is used, and after being processed by the adaptive turning rate estimation network, the turning rate value of the target at the current moment is output in real time.

[0010] The turning rate parameter is fed back into the cooperative turning model, and the maneuvering target tracking optimization method is used to implement target tracking processing.

[0011] The state transfer matrix of the cooperative turning model is updated based on the estimated turning rate, and the target tracking filtering process is then implemented based on the linear Gaussian jump Markov-Gaussian mixture probability hypothesis density filter.

[0012] Among them, in the design of the adaptive turning rate estimation network, the time series related features of each feature in the input feature matrix Fv are extracted based on the long short-term memory network, and the output hidden state is expressed as HI={h1,h2,…,h K}, h k Represents the hidden state output by the LSTM network at time step k.

[0013] Among them, in the time pattern attention module, feature extraction is performed on the changing rules of each feature parameter in the time series dimension and the correlation between different features, as follows:

[0014] First, temporal pattern feature extraction is performed based on convolutional neural networks, using f one-dimensional CNN filters with a length of K-1 Extract features from the row vector of the hidden state HI and generate the matrix in, Represents the convolution value of the i-th row vector of HI and the j-th filter;

[0015] Then, H is calculated based on the attention mechanism C The weight and v of the row vector K ; Score function To assess the relevance, Yes H C The i-th row of Attention weight α i pass Based on α i For H C The row vectors of

[0016] Finally, v K With h K Concatenated to generate an estimate of the turn rate at the current moment:

[0017] h′ K=W h h K +W v v K

[0018]

[0019] Among them, h K ,

[0020] The loss function of this network is: in, is the turning rate estimate output by the network, ω k is the true value.

[0021] Beneficial effects:

[0022] The present invention introduces Doppler information into the measurement equation, and performs Taylor series expansion on the nonlinear equation between Doppler measurement and state variables at the predicted value to obtain its first-order approximate linear equation, thereby avoiding nonlinear filtering processing. The present invention designs an adaptive turning rate estimation network, and uses a multi-dimensional target motion state time series to realize the turning rate estimation of the current track, thereby improving the matching degree between the filtering model and the target motion, and thus improving the tracking accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The figure is an overall flow chart of the optimization method of the present invention.

[0024] Figure 2 Schematic diagram of the adaptive turning rate estimation network of the present invention.

[0025] Figure 3 It is a comparison diagram of tracking errors in a specific implementation case of the present invention. DETAILED DESCRIPTION

[0026] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0027] Aiming at the problem of poor track quality caused by difficulty in matching filter models for maneuvering target tracking in clutter environments, the present invention provides a maneuvering target tracking optimization method based on deep learning. The multi-target tracking algorithm based on random finite set (RFS) can avoid data association and directly estimate the number and state of targets, which is particularly suitable for solving complex scenes such as dense clutter and unknown number of targets. Among them, the Gaussian mixture probability hypothesis density (Gaussian Mixture Probability Hypothesis Density, GMPHD) filter transfers the first-order statistical moment of the posterior multi-target state - the posterior intensity, which has the advantage of low computational complexity and is widely used in engineering. The LGJMS-GMPHD algorithm, which is obtained by combining the Linear Gaussian Jump Markov System (LGJMS) with the GMPHD filter, has established a complete theoretical system and implementation scheme for maneuvering multi-target tracking. In practical applications, a set of constant velocity (CV) and coordinated turn (CT) models is usually used to describe target motion. However, the unknown motion parameters of maneuvering targets make it difficult to accurately match the filter model, which in turn leads to a decrease in tracking accuracy. Therefore, under the framework of the LGJMS-GMPHD algorithm, the present invention designs a deep learning-based filter optimization method. The overall process is as follows: Figure 1 Firstly, the Doppler information is introduced into the measurement equation to improve the tracking performance in clutter environment; then, an adaptive turning rate estimation network is designed to estimate the turning rate at the current moment by using the multi-dimensional target motion state time series to improve the matching degree of the filter model.

[0028] Specifically, the motion model and measurement model after the introduction of Doppler measurement are described as follows:

[0029] Given k-1 time multi-objective states X k-1 , then the multi-objective state X at time k k It is given by the union of the set of surviving and new targets:

[0030]

[0031] Where: B k is the new target state set; ∪ is the set union operation; S k|k-1 (x k-1,i ) is the target at time k-1 with survival probability p s (x k-1,i ) survives to time k and evolves to the target state set.

[0032] Among them, the target motion state can be expressed as:

[0033]

[0034] Where: x k,i ,y k,i is the target location; is the target movement rate, and N(k) is the number of targets in the space at time k.

[0035] Under the GM model, its state transfer function is:

[0036]

[0037] Where: is a Gaussian density with mean m and covariance P, F k-1 is the state transfer matrix, Q k-1 is the process noise covariance.

[0038] Multi-target measurements received by the sensor Z k It can be represented by the union of the measurements and clutter generated by the target:

[0039] Z k =θ(X k )∪K k (4)

[0040] In the formula, K k is the false alarm clutter measurement set; Θ(X k ) is the detected target measurement set, and the detection probability p D,k <1.

[0041] After introducing Doppler information, the measurement z k,j From state x k,i The measurement equation is:

[0042]

[0043] In the formula, is zero mean and covariance Gaussian white noise; They are position measurement and Doppler measurement values ​​respectively.

[0044] It can be seen from formula (6) that the observation vector cannot be linearly represented by the state vector. To simplify the calculation, this paper converts formula (6) into k,i Use Taylor series expansion at , ignore the terms above the second order, and get the observation matrix at this time:

[0045]

[0046] in,

[0047]

[0048] In the cooperative turning model, its state transfer matrix F k-1 The target turning rate parameter ω is included in the figure. Therefore, the present invention designs an adaptive turning rate estimation network. The network diagram is shown in FIG. Figure 2 As shown. The input of the network is the feature matrix Fv containing multi-dimensional target motion state time series information:

[0049]

[0050] Among them, the feature vector at time k is expressed as:

[0051]

[0052] Among them, x k and k represents the target position estimate at time k, and represents the target velocity estimate at time k, Indicates Doppler measurement.

[0053] Since the estimation accuracy of different features in the feature vector represented by equation (9) is different, for example, Doppler measurement has higher measurement accuracy than filter estimation, different features contribute differently to the network output during neural network training. In addition, the time dimension information of the feature parameters can also reflect the maneuvering motion characteristics of the target. Therefore, in the design of the adaptive turning rate estimation network, multi-angle and multi-level feature extraction is performed to fully mine the feature information and improve the turning rate estimation accuracy. The specific implementation steps are as follows:

[0054] 1) Based on the long short-term memory (LSTM) network, the time series related features of each feature in the input feature matrix Fv are extracted, and the output hidden state is expressed as HI={h1,h2,…,h K}, h k Represents the hidden state output by the LSTM network at time step k.

[0055] 2) In the Temporal Pattern Attention (TPA) module, the variation law of each feature parameter in the time series dimension and the correlation between different features are extracted. First, temporal pattern feature extraction is performed based on the Convolutional Neural Network (CNN), using f one-dimensional CNN filters with a length of K-1. Extract features from the row vector of the hidden state HI and generate the matrix in, Represents the convolution value of the i-th row vector of HI and the j-th filter. This step can be expressed by the formula:

[0056]

[0057] Then, H is calculated based on the attention mechanism C The weight and v of the row vector K . Define the following scoring function To assess relevance:

[0058]

[0059] in, Yes H C The i-th row of Attention weight α i Calculated by the following formula:

[0060]

[0061] Furthermore, based on α i For H C The row vectors of

[0062]

[0063] Finally, v K With h K Concatenated to generate an estimate of the turn rate at the current moment:

[0064] h′ K =W h h K +W v v K (14)

[0065]

[0066] Among them, h K ,

[0067] The loss function of this network is:

[0068]

[0069] in, is the turning rate estimate output by the network, ω k is the true value.

[0070] The following is an example of the target trajectory of various maneuvering forms to illustrate and analyze the above algorithm. Different initial speed values ​​and turning rates are designed to construct the maneuvering target trajectory data set. The specific parameters are shown in Table 1. Two groups of turning rate change intervals are set, ω1∈[-8° / s,8° / s], ω2∈[-15° / s,-8° / s]∪[8° / s,15° / s]. The uniform motion frame number range is 20-25, the turning rate ω1 motion frame number range is 10-50, the turning rate ω2 motion frame number range is 8-20, and the frame interval T is 1s. In the simulation results, the LGJMS-GMPHD algorithm is abbreviated as GMPHD, and after the introduction of Doppler measurement, it is recorded as GMPHDwD, and after further adding adaptive turning rate estimation processing, it is recorded as GMPHDwD-AT.

[0071] Table 1 Trajectory kinematic parameters

[0072]

[0073] Based on the five trajectory parameters in Table 1, 100 trajectories are constructed to test the performance of the adaptive turning rate network. The tracking accuracy evaluation index is the root mean square error (RMSE):

[0074]

[0075] in, and Represents the output of the residual correction network, x k and k represents the true value of the trajectory, and M represents the number of simulations.

[0076] The comparison results of the root mean square error of the five groups of trajectories are as follows: Figure 3 As shown in the figure, it can be seen that compared with GMPHDwD, the tracking error of GMPHDwD-AT is reduced, indicating that the adaptive turning rate estimation network can obtain more accurate turning rate estimation, improve the matching degree between the filter model and the target motion model, and thus improve the tracking accuracy and tracking performance. In addition, the method can maintain good tracking performance under different turning rates and motion model switching modes, indicating that the proposed method has certain robustness to different maneuvering forms and different measurement noises.

[0077] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A maneuvering target tracking optimization method based on deep learning, characterized in that: The steps include: Construct a measurement model that introduces Doppler information, specifically: measure z k,j From state x k,i The measurement equation is In the formula, is zero mean and covariance Gaussian white noise; They are position measurement and Doppler measurement values ​​respectively; Based on the adaptive turning rate estimation network, the multi-dimensional target motion state time series is used to estimate the turning rate at the current moment and complete the optimization; wherein the input of the network is the feature matrix Fv containing the multi-dimensional target motion state time series information: Among them, the eigenvector at time k is expressed as Among them, x k and k represents the target position estimate at time k, and represents the target velocity estimate at time k, Indicates Doppler measurement.

2. The method according to claim 1, characterized in that The measurement equation that introduces Doppler measurement information is converted from a nonlinear equation into a linear equation by performing Taylor series expansion at the predicted value.

3. The method according to claim 1, characterized in that By using the multi-dimensional target motion state time series information and processing it through an adaptive turning rate estimation network, the turning rate value of the target at the current moment is output in real time.

4. The method according to any one of claims 1 to 3, characterized in that: The turning rate parameter is fed back into the cooperative turning model, and the maneuvering target tracking optimization method is used to implement target tracking processing.

5. The method according to claim 4, characterized in that The state transfer matrix of the cooperative turning model is updated based on the turning rate estimation value, and then the target tracking filtering processing is realized based on the linear Gaussian jump Markov-Gaussian mixture probability hypothesis density filter.

6. The method according to claim 1, 2, 3 or 5, characterized in that In the design of the adaptive turning rate estimation network, the time series related features of each feature in the input feature matrix Fv are extracted based on the long short-term memory network, and the output hidden state is expressed as HI={h1,h2,…,h K }, h k Represents the hidden state output by the LSTM network at time step k.

7. The method according to claim 6, characterized in that In the time pattern attention module, feature extraction is performed on the changing rules of each feature parameter in the time series dimension and the correlation between different features, as follows: First, temporal pattern feature extraction is performed based on convolutional neural networks, using f one-dimensional CNN filters with a length of K-1 Extract features from the row vector of the hidden state HI and generate the matrix in, Represents the convolution value of the i-th row vector of HI and the j-th filter; Then, H is calculated based on the attention mechanism C The weight and v of the row vector K ; Score function g: To assess the relevance, Yes H C The i-th row of Attention weight α i pass Based on α i For H C The row vectors of Finally, v K With h K Concatenated to generate an estimate of the turn rate at the current moment: h′ K =W h h K +W v v K Among them, h K , The loss function of this network is: in, is the turning rate estimate output by the network, ω k is the true value.

Images