A radar robust tracking method for maneuvering targets

Abstract

The present invention discloses a radar robust tracking method for maneuvering flight targets. It can achieve high-precision tracking of maneuvering targets in high-clutter environments and improve track quality. Under the framework of the Gaussian mixture probability hypothesis density filtering algorithm, multi-target state filtering results are obtained based on measurement information. First, the target state filtering results are input into the residual correction network, and the feature extraction of the time series is realized from the forward and reverse dimensions based on the bidirectional long short-term memory network; then, the outputs of different bidirectional long short-term memory network layers are weighted based on the multi-level attention mechanism. At the same time, the importance of the state of the input time series at different times to the network model training is captured based on temporal attention, thereby improving the feature extraction capability of the network and enabling it to adapt to the residual correction processing requirements of complex maneuvering targets. After processing by the residual correction network, the filter trajectory can be further smoothed, thereby improving the tracking accuracy of maneuvering targets in clutter environments.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile multi-target tracking, and in particular to a radar tracking method for mobile flying targets. Background Art

[0002] Detecting low-altitude targets such as birds and drones is crucial for ensuring air route safety and urban security. Approximately 21,000 bird strikes occur worldwide each year, resulting in economic losses estimated at $1.2 billion. The rapid development of low-altitude aircraft, such as rotary-wing drones, also poses a significant threat to flight safety within airport cleared airspaces. Several airports across China have experienced incidents of illegal drone flights disrupting flights. Radar provides a crucial technical means for effectively monitoring birds and drones. However, these targets typically fly at low altitudes and are highly maneuverable, creating a complex detection environment that makes high-precision and stable tracking difficult. Summary of the Invention

[0003] In view of this, the present invention provides a radar robust tracking method for maneuvering flight targets, which can achieve high-precision tracking of maneuvering targets in high-clutter environments and improve track quality.

[0004] A radar tracking method for a maneuvering flying target according to the present invention comprises the following steps:

[0005] The measurement points are filtered by LGJMS-GMPHD and then processed by a residual correction network to further smooth the filtering results. The input of the residual correction network is the target trajectory filter sequence obtained based on the GMPHD filter. When constructing the residual correction network, the features of the time series are first extracted from both the forward and reverse dimensions based on a bidirectional long short-term memory network. Then, the outputs of different BiLSTM layers are weighted based on a multi-level attention mechanism. At the same time, temporal attention is used to capture the importance of the state of the input time series at different times for network model training. Finally, the hidden state information is converted into the corrected filtering result output through processing in a fully connected layer.

[0006] Among them, the prediction and update of multi-target states are realized based on the probability hypothesis density filter under the linear Gaussian jump Markov system.

[0007] Among them, for the Bi-LSTM layer, the output vector h of the i-th layer of Bi-LSTM at time k is ik The output of the forward LSTM and the backward LSTM are jointly determined and used as the input of the next layer, h ik Expressed as:

[0008]

[0009] in, Represents the element and.

[0010] The multi-level attention is obtained by ik Weighting is performed to realize the feature extraction of the interaction relationship between layers. The processed hidden state matrix is expressed as The calculation method is as follows:

[0011]

[0012] Among them, N L Indicates the total number of Bi-LSTM layers, and each layer outputs h ik The weight value α i m k Calculated by the following method:

[0013] o ik =tanh(W i h ik +b i )

[0014]

[0015] Among them, W i , b i and u i Represents the weight parameter corresponding to the i-th layer.

[0016] Among them, in the temporal attention layer, H ML The feature weights are assigned to the hidden states at different time steps, and the weight value at time k is expressed as The processed hidden state matrix is expressed as The implementation steps are as follows:

[0017]

[0018]

[0019]

[0020] Among them, w k are the training parameters of the network.

[0021] Among them, after processing by the fully connected layer, the hidden state is converted into the desired corrected filter state output Among them, W FC and b FC denote the weight and bias matrices of the fully connected layer respectively.

[0022] Among them, the loss function of the residual correction network is:

[0023]

[0024] in, is the filtered correction result output by the network, is the true value of the trajectory.

[0025] Beneficial effects:

[0026] The present invention applies LGJMS-GMPHD filtering to the measurement traces, then processes the target state filtering results through a residual correction network to further smooth the filtering results. Within the framework of the Gaussian mixture probability hypothesis density filtering algorithm, multiple target state filtering results are derived based on the measurement information. The residual correction network is then used to smooth the multi-model filtering states, addressing the problem of decreased tracking accuracy caused by filter model matching delays during target maneuvers.

[0027] In this invention, the specific process of residual correction network processing is as follows: first, the target state filtering result is input into the residual correction network, and the bidirectional long short-term memory network is used to extract time series features from both the forward and reverse dimensions. Then, the outputs of different bidirectional long short-term memory network layers are weighted using a multi-level attention mechanism. At the same time, temporal attention is used to capture the importance of the input time series state at different moments for network model training, improving the network's feature extraction capabilities and making it suitable for residual correction processing of complex maneuvering targets. After processing by the residual correction network, the filtered trajectory can be further smoothed, thereby improving the tracking accuracy of maneuvering targets in clutter environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is an overall flow chart of the radar tracking method of the present invention.

[0029] Figure 2 Schematic diagram of the residual correction network of the present invention.

[0030] Figure 3 This is a schematic diagram of the bidirectional long short-term memory network of the present invention.

[0031] Figure 4 It is a schematic diagram of a simulation scenario in a specific implementation case of the present invention.

[0032] Figure 5 This is a graph showing comparison of tracking accuracy results in a specific implementation case of the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0034] The multi-target tracking algorithm based on random finite set (RFS) can avoid data association and directly estimate the number and state of targets, and is particularly suitable for solving complex scenarios such as dense clutter and unknown number of targets. Among them, the Gaussian Mixture Probability Hypothesis Density filter (GMPHD) transmits 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 derived 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. However, in practical applications, the unknown motion parameters of the maneuvering target make it difficult to accurately match the filter model, which in turn leads to a decrease in tracking accuracy. Therefore, in order to solve the problem that the maneuvering target tracking filter model is difficult to match in a clutter environment and there is a delay resulting in poor track quality, the present invention proposes a radar tracking method for low-altitude maneuvering flight targets. The process is as follows: Figure 1 First, the measurement point traces are filtered using LGJMS-GMPHD, and then processed by the residual correction network to further smooth the filtering results to improve the target tracking accuracy and track quality.

[0035] Among them, the linear Gaussian jump Markov system is described as:

[0036] Consider a discrete, finite set of motion models Assumptions is the model index parameter, and the target motion model is switched according to the Markov chain. In the linear JM system, the model μ k The transition probability between each other is a constant, and the model transfer matrix formed is expressed as Each element of the matrix represents the transition probability between two models, that is,

[0037]

[0038] make and They represent the motion state (including target position and velocity) and measurement at time k respectively. When implementing multi-model tracking based on the JM system, the random state variables need to be generalized to discrete state variables containing the corresponding motion model, that is, x = (ξ, μ). The linear Gaussian jump Markov system is a JM system under the linear Gaussian model, that is, the model μ k As the condition, the state transition density and measurement likelihood are:

[0039]

[0040]

[0041] Where: F k-1 (μ k ) and H k (μ k ) are respectively based on the model μ k is the conditional transfer matrix and measurement matrix; Q k-1 (μ k ) and R k (μ k ) are respectively based on the model μ k is the covariance matrix of the process noise and measurement noise conditional on .

[0042] The closed-form PHD recursion of the multi-objective model under the LGJM system is as follows:

[0043] For the LGJM system multi-objective model, the predicted intensity v at time k k|k-1 is a Gaussian mixture form, that is

[0044] v k|k-1 (ξ,μ)=v S,k|k-1 (ξ,μ)+v γ,k (ξ,μ) (4)

[0045] Where, v γ,k (·) is the new strength at time k,

[0046]

[0047] J γ,k (μ), i=1,2,…,J γ,k (μ) is the Gaussian mixture density parameter of the newly generated target motion state at time k, conditioned on the model μ; is the strength of the new model at time k.

[0048]

[0049]

[0050]

[0051]

[0052] For the LGJM system multi-objective model, the intensity v is updated at time k. k Also in the form of Gaussian mixture,

[0053]

[0054] Jk (μ), i=1,2,…,J k (μ) is the Gaussian mixture density parameter of the posterior target motion state at time k, conditioned on the model μ.

[0055] When the target is maneuvering, there will be a certain delay in switching the filter model, which will lead to a sudden increase in the filtering error. Therefore, the present invention designs a residual correction network to further smooth the filtering state to improve the tracking accuracy of the maneuvering target. The network structure is as follows: Figure 2 shown.

[0056] The input of the network is the target trajectory filter sequence obtained based on the GMPHD filter:

[0057]

[0058] in, The state vector containing the target position estimate, whose changes along the time dimension can describe the target's maneuvering characteristics, is designed for residual correction. Therefore, in the designed residual correction network, a bidirectional long short-term memory (Bi-LSTM) network is first used to extract time series features from both the forward and reverse dimensions. Then, a multi-level attention mechanism (MLA) is used to weight the outputs of different BiLSTM layers. Simultaneously, temporal attention (TA) is used to capture the importance of the input time series states at different moments for network model training, improving the network's feature extraction capabilities and enabling it to adapt to the residual correction processing requirements of complex maneuvering targets. Finally, a fully connected layer is used to convert the hidden state information into a corrected filter output.

[0059] Specifically, for the Bi-LSTM layer: the Bi-LSTM network structure is as follows Figure 3 As shown, the output vector h of the i-th layer at time k is ik The output of the forward LSTM and the backward LSTM are jointly determined and used as the input of the next layer, h ik Expressed as:

[0060]

[0061] in, Represents the element and.

[0062] 1) Multi-level attention layer

[0063] Multi-level attention is achieved by focusing on the output h of each layer of Bi-LSTM network. ikWeighting is performed to realize the feature extraction of the interaction relationship between layers. The processed hidden state matrix is expressed as The calculation method is as follows:

[0064]

[0065] Among them, N L Indicates the total number of Bi-LSTM layers, and each layer outputs h ik The weight value Calculated by the following method:

[0066] o ik =tanh(W i h ik +b i ) (14)

[0067]

[0068] Among them, W i , b i and u i Represents the weight parameter corresponding to the i-th layer.

[0069] 2) Temporal Attention Layer

[0070] In the temporal attention layer, H ML The feature weights are assigned to the hidden states at different time steps, and the weight value at time k is expressed as The processed hidden state matrix is expressed as The implementation steps are as follows:

[0071]

[0072]

[0073]

[0074] Among them, w k are the training parameters of the network.

[0075] Finally, after processing in the fully connected layer, the hidden state is converted into the desired corrected filter state output:

[0076]

[0077] Among them, W FC and b FC denote the weight and bias matrices of the fully connected layer respectively.

[0078] The loss function of this network is:

[0079]

[0080] in, is the filtered correction result output by the network, is the true value of the trajectory.

[0081] The following is an example of the above algorithm, using a multi-target maneuvering scenario in a clutter environment as an implementation example. The clutter measurement is modeled as a Poisson random set, and its density is expressed as:

[0082]

[0083] in, Expressed as a uniform density covering the entire observation area, V = 6.25 × 10 6 m 2 is the area of the observation region, λ c =8×10 -6 m -2 Indicates the average number of clutter within a unit area. The trajectory parameters of the maneuvering target are shown in Table 1. The actual target motion trajectory scenario is as follows: Figure 4 As shown in Figure 1, it includes the five maneuver forms listed in Table 1.

[0084] Table 1 Target trajectory parameters

[0085]

[0086] The tracking performance is evaluated based on the Optimal Subpattern Assignment (OSPA) distance, which is defined as:

[0087]

[0088] in, n represents the number of measurement points, and m represents the number of targets. n Represents all permutations and combinations of the set {1,2,…,n}, and π is the true target state x(k) and the filtered value of each target selected at time k. One-to-one sorting. The OSPA distance can be decomposed into two parts: positioning error and potential error. The association sensitivity parameter c determines the relative weights of the potential error and positioning error parts; the distance sensitivity parameter p determines the sensitivity to outliers. In this embodiment of the application, p = 1 and c = 100 are set.

[0089] The OSPA distance comparison results are as follows: Figure 5 As shown in the figure, it can be seen that the tracking accuracy of GMPHD-RM after adding the residual correction network is improved compared with GMPHD, which verifies the effectiveness of the method proposed in the present invention in the maneuvering multi-target scenario under clutter environment.

[0090] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A radar tracking method for a maneuvering flying target, characterized in that: The steps include: The measurement points are filtered by LGJMS-GMPHD and then processed by a residual correction network to further smooth the filtering results. The input of the residual correction network is the target trajectory filter sequence obtained based on the GMPHD filter. When constructing the residual correction network, the features of the time series are first extracted from both the forward and reverse dimensions based on a bidirectional long short-term memory network. Then, the outputs of different Bi-LSTM networks are weighted based on a multi-level attention mechanism. At the same time, temporal attention is used to capture the importance of the states of the input time series at different times to the network model training. Finally, the hidden state information is converted into the corrected filtering result output through processing in a fully connected layer.

2. The method according to claim 1, wherein The prediction and update of multi-target states are realized based on the probability hypothesis density filter in the linear Gaussian jump Markov system.

3. The method according to claim 1, wherein For the Bi-LSTM network, the output vector h of the i-th layer of the Bi-LSTM network at time k is ik The output of the forward LSTM and the backward LSTM are jointly determined and used as the input of the next layer, h ik Expressed as: in, Represents the element and.

4. The method according to claim 3, wherein The multi-level attention mechanism is implemented by focusing on the output h of each layer of Bi-LSTM network. ik Weighting is performed to realize the feature extraction of the interaction relationship between layers. The processed hidden state matrix is expressed as The calculation method is as follows: Among them, N L Represents the total number of layers in the Bi-LSTM network, and each layer outputs h ik The weight value Calculated by the following method: o ik =tanh(W i h ik +b i ) Among them, W i , b i and u i Represents the weight parameter corresponding to the i-th layer.

5. The method according to claim 3, wherein In the temporal attention layer, H ML The feature weights are assigned to the hidden states at different time steps, and the weight value at time k is expressed as The processed hidden state matrix is expressed as The implementation steps are as follows: Among them, w k are the training parameters of the network.

6. The method according to claim 3, wherein After being processed by the fully connected layer, the hidden state is converted into the desired corrected filter state output Among them, W FC and b FC denote the weight and bias matrices of the fully connected layer respectively.

7. The method according to any one of claims 1 to 6, wherein: The loss function of the residual correction network is: in, is the filtered correction result output by the network, is the true value of the trajectory.

Images