A mamba-based maneuvering-aware target tracking method

Abstract

This invention discloses a maneuvering target tracking method based on a Mamba tracking network, comprising: 1. Preprocessing trajectory data, calculating the mean and standard deviation, and normalizing the input data to eliminate sequence non-stationarity; 2. Constructing a maneuvering factor generator to map statistical features into adjustment factors reflecting the target's maneuvering characteristics; 3. Constructing a target tracking network based on a Mamba tracking network, extracting long-range spatiotemporal dependence features of the trajectory using a selective scanning mechanism, and dynamically adjusting the network feature response through adjustment factors; 4. Performing denormalization on the network output to recover the target state prediction value at the physical scale. This invention, through the organic combination of the Mamba tracking network and adjustment factors, effectively compensates for the loss of maneuvering information during the normalization process, significantly improving the accuracy and generalization ability of tracking highly maneuvering targets while maintaining linear computational complexity, and has significant application value.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and target tracking technology, specifically to a deep learning method that integrates maneuvering factors and Mamba networks, suitable for accurate and real-time tracking of highly maneuverable targets. Background Technology

[0002] Maneuvering target tracking has significant applications in radar surveillance, missile defense, and autonomous driving. Traditional tracking methods rely on pre-set motion models (such as constant velocity model (CV), constant acceleration model (CA), and constant turning rate model (CTRV)) and filtering algorithms (such as Kalman filter (KF) and interactive multi-model (IMM)) to achieve trajectory tracking through recursive estimation of the target's motion state. However, these methods are poorly adaptable to highly maneuvering targets. When the target's motion pattern changes rapidly and drastically, the pre-set model struggles to accurately match the actual motion state, leading to significantly increased response delays and state estimation errors, making it difficult to cope with complex and ever-changing maneuvering scenarios.

[0003] With the rapid development of deep learning, the field of maneuvering target tracking has undergone a transformation, with neural network-based methods gradually replacing some traditional model-driven approaches. These neural network methods automatically learn the spatiotemporal evolution patterns in trajectory sequences through data-driven methods, improving adaptability in complex scenarios. For example, Recurrent Neural Networks (RNNs) and their variant, Long Short-Term Memory (LSTM), can recursively estimate the target state using historical temporal information, overcoming to some extent the dependence of traditional filtering methods on accurate motion models. Transformer networks, with their self-attention mechanism, model the global dependencies of sequences, further enhancing the model's ability to express complex maneuvering trajectories. Recently, the Mamba architecture, with its selective scanning mechanism, has demonstrated a significant advantage in linear time complexity for long sequence modeling, providing a new technical approach for efficiently processing long trajectory sequences.

[0004] However, the aforementioned deep learning methods still face their own challenges in maneuvering target tracking tasks: (1) RNN and LSTM suffer from gradient vanishing and long-term dependency decay when processing long trajectory sequences, making it difficult to effectively capture the correlation features between distant time steps; (2) Although Transformer networks have global modeling capabilities, the computational complexity of its self-attention mechanism increases quadratically with the sequence length, facing bottlenecks in resource consumption and processing speed in real-time tracking scenarios; (3) Although the Mamba structure can achieve efficient sequence modeling with linear time complexity, its original design does not fully consider the non-stationary characteristics of maneuvering target trajectory sequences, and lacks explicit perception and dynamic response capabilities for maneuvering information when the target motion state changes abruptly. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a maneuvering target tracking method based on a Mamba tracking network. The aim is to dynamically adjust the network feature response according to the target's maneuvering state, thereby achieving accurate tracking of highly maneuverable targets. This approach improves the accuracy and generalization ability of tracking highly maneuverable targets while maintaining linear computational complexity. To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for tracking a maneuvering sensing target based on a Mamba tracking network, characterized by the following steps: Step 1: Obtain a batch of radar observation trajectory sequences and the corresponding true trajectory sequence ,in, Indicates the first Observation trajectory, Indicates the first The state of the observed trajectory at time t, and They represent the first The observation trajectory at time Two-dimensional coordinates, and They represent the first The two-dimensional velocity components of the observed trajectory at time t; Indicates the first A real trajectory, Indicates the first The state of the real trajectory at time t, and They represent the first The real trajectory at any moment Two-dimensional coordinates, and They represent the first The real trajectory at any moment The two-dimensional velocity components; where B represents the number of trajectories in a batch, and M represents the time length of a single trajectory. The initial feature dimension representing the trajectory; Step 2, for Perform L-level discrete wavelet transform to obtain the first... The 4×L×2 dimensional wavelet statistical characteristics of the observed trajectory at time t, and compared with... After splicing, we get the first... The enhanced state vector of the observed trajectory at time t Thus, the feature-enhanced version of the first... observation trajectory This leads to the radar observation trajectory sequence with enhanced features. ; Calculate along the time dimension mean and standard deviation and utilize and right Normalization is performed to obtain the normalized i.e. observation trajectory Thus, the normalized radar observation trajectory sequence is obtained. ;in, This indicates the dimension of the observed trajectory after feature enhancement; Step 3: Construct a maneuvering factor generator and use three parallel multilayer perceptrons to analyze it. , and the standardized turning rate calculated from the two-dimensional velocity components Processing is performed to generate the representation of the first character. Scaling factor of the observed trajectory and the Offset factor of the observed trajectory ;in, The preset model feature dimensions; Step 4: Construct the feature encoding module and perform... Perform spatiotemporal feature extraction and output the first... Spatiotemporal characteristic sequence of observation trajectory ; Step 5: Constructing a Mamba tracking network with fused regulatory factors is... It consists of a series of stacked Mamba layers, and... Modeling and maneuver feature modulation processing are performed to obtain the first... Enhanced feature sequence of observed trajectories ; Step 6, The input is processed in a linear layer to obtain the first... Normalized predicted state sequence of observed trajectories And using the mean and standard deviation right Perform inverse normalization to obtain the k-th predicted trajectory. Thus, the predicted trajectory sequence is obtained. ;in, For the first The predicted trajectory at time The state vector; Step 7, for Perform L-level discrete wavelet transform using the same method as in step 2, and compare it with... By splicing, we obtain the feature-enhanced version of the first... Real trajectory Thus, the real trajectory sequence after feature enhancement is obtained. ;in, For the first The real trajectory at any moment The state vector; Step 8, based on and Construct the total loss function This is used to perform backpropagation training on a network consisting of a motor factor generator, a feature encoding module, a Mamba tracking network, and a mapping linear layer to update the network parameters until... The optimal maneuvering target tracking model is obtained by converging or reaching a preset number of training rounds, and is used to estimate the state and predict the trajectory of the input maneuvering target. The characteristic of the maneuverable sensing target tracking method described in this invention is that step 3 includes: Step 3.1: Use equations (1), (2), and (3) to respectively... Processing yields the first... Average turning rate of the observed trajectory Thus, to Standardization is performed to obtain the first Turning rate after standardization of the observed trajectory : (1) (2) (3) In equations (1), (2), and (3), Indicates the first The angle between the velocity direction of the observed trajectory at time t and the positive x-axis. Indicates the first The angle between the velocity direction of the observed trajectory at time t-1 and the positive x-axis. Indicates the first The instantaneous turning rate of the observed trajectory at time t; Step 3.2, using equation (4) to... Processing yields the first... Intensity modulation characteristics of the observed trajectory : (4) In equation (4), This represents the first multilayer perceptron; Indicates the activation function; Step 3.3, using equation (5) to... Processing yields the first... The first offset component of the observed trajectory : (5) In equation (5), This represents the second multilayer perceptron; Step 3.4, using equation (6) to... Processing yields the first... The second offset component of the observed trajectory : (6) In equation (6), This represents the third multilayer perceptron; Step 3.5: Using equations (7) and (8), obtain the first... Scaling factor of the observed trajectory and offset factor : (7) (8).

[0006] Furthermore, step 4 includes: Step 4.1, using equation (9) to... Feature extraction is performed to obtain the intermediate feature sequence. : (9) In equation (9), , This represents the two weight matrices of the linear mapping layer. , This represents the two bias vectors of the linear mapping layer. Represents the instance normalization function. It is an activation function; Step 4.2, using equation (10) to... Injecting the mobility factor, we obtain the first Spatiotemporal characteristic sequence of observation trajectory : (10) In equation (10), Representation layer normalization.

[0007] Furthermore, each Mamba layer in step 5 includes: a bidirectional selection scan module and two layer normalization modules; Step 5.1, when When initializing the first Enhanced temporal features of the output of each Mamba layer ; Step 5.2, the first The first branch in the bidirectional selection scan module of the Mamba layer uses equation (11) to... Processing yields the first... Bidirectional scan output characteristics of the first branch of each Mamba layer : (11) In equation (11), and They represent the first The weight matrix and bias vector of the first branch of each Mamba layer, where DWConv is a convolution kernel of size ... Depth-separable convolution; , These represent forward and reverse scan operations, respectively. It is the activation function, and e is the dimensionality expansion index of the Mamba layer; The first The bidirectional selective scanning module of each Mamba layer uses equation (12) to perform a scan. Dynamic modulation is performed to obtain the first Features of the first branch modulation of the Mamba layer ; (12) The first The second branch in the bidirectional selection scanning module of the Mamba layer utilizes equation (13) to... Processing yields the first... Output characteristics of the second branch of the Mamba layer : (13) In equation (13), and They represent the first The weight matrix and bias vector of the second branch of each Mamba layer; The first The bidirectional selection scanning module of the Mamba layer uses equation (14) to obtain the first Mamba layer. Scan output characteristics of each Mamba layer : (14) Step 5.3, the The first layer normalization module in a Mamba layer Using equation (15) to obtain the first The first intermediate feature of the Mamba layer : (15) In equation (15), Indicates residual connection; Step 5.4, using equation (16) to... Dynamic modulation is performed to obtain the first The second intermediate feature of the Mamba layer ; (16) Step 5.5, the The second layer normalization module in the Mamba layer Using equation (17) to obtain the first Enhanced temporal features of the output of each Mamba layer ; (17) In equation (17), Residual connection, FFN is a noisy feedforward network, and has: (18) In equation (18), This represents the weight matrix of the first linear layer in a feedforward network with a noisy Tanh activation function. This represents the weight matrix of the second linear layer in a feedforward network with a noisy Tanh activation function. Let represent the noisy Tanh activation function, and we have: (19) In equation (19), N(0,1) is standard Gaussian noise; Step 5.6, Assign to Then, return to step 5.2 and execute sequentially until... Until then, thus obtaining the first Enhanced feature sequences output by each Mamba layer And as .

[0008] Furthermore, in step 8 It is established according to the following steps: Step 8.1: Construct the state prediction loss using equation (20). : (20) In equation (20), For the first The predicted trajectory at time The Dimensional state value, For the first The real trajectory at any moment The Dimensional state value; Step 8.2: Construct the feature reconstruction loss using equation (21). : (twenty one) Step 8.3: Construct the total loss function using equation (22). : (twenty two) In equation (22), and There are two hyperparameters.

[0009] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in executing the maneuvering target tracking method, and the processor is configured to execute the program stored in the memory.

[0010] The present invention provides a computer-readable storage medium on which a computer program is stored, characterized in that the computer program, when executed by a processor, performs the steps of the maneuvering target tracking method.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In the target tracking process, the present invention introduces the turning rate as a maneuver feature and combines the sequence mean and standard deviation to jointly model the target motion state, so that the model can not only characterize the overall trend and fluctuation characteristics of the trajectory, but also characterize the change of the target's turning behavior. This breaks through the limitation of existing methods that rely only on implicit feature learning, and enables the model to have the ability to explicitly perceive complex maneuver behavior. It has better adaptability and expressive ability in high maneuver and strong nonlinear motion scenarios.

[0012] (2) This invention embeds the maneuver modulation mechanism into multiple key stages such as feature extraction, state space modeling and nonlinear transformation, and deeply integrates it with the Mamba state space model, so that maneuver information runs through the entire time series modeling process. This enables the model to dynamically adjust the system response according to the maneuver state while modeling long sequence dependencies. Compared with the method of introducing condition information only at a single location, this significantly improves the model's ability to characterize complex dynamic behaviors and overall tracking performance. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 These are schematic diagrams of different modules; Figure 3a This is the result image of the first trajectory; Figure 3b This is the result image of the second trajectory; Figure 3c This is the result image of the third trajectory; Figure 3d This is the result image of the 4th trajectory; Figure 3e This is the result image of the 5th trajectory; Figure 3f This is the result image of the 6th trajectory; Figure 3g This is the result image for the 7th trajectory. Detailed Implementation

[0014] In this embodiment, a Mamba network-based maneuvering target tracking method incorporating maneuvering factors effectively compensates for the loss of maneuvering information during normalization through a collaborative design of maneuvering statistics (mean, standard deviation, and turning rate), multi-layer maneuvering factor modulation, feature encoding, and efficient Mamba sequence modeling. This ensures the method's real-time processing capability and cross-scenario generalization performance, overcoming the limitations of traditional methods and existing deep learning methods. Specifically, the overall process of this method is as follows: Figure 1 As shown, it includes the following steps: Step 1: Obtain a batch of radar observation trajectory sequences and the corresponding true trajectory sequence ,in, Indicates the first Observation trajectory, Indicates the first The state of the observed trajectory at time t, and They represent the first The observation trajectory at time Two-dimensional coordinates, and They represent the first The two-dimensional velocity components of the observed trajectory at time t; Indicates the first A real trajectory, Indicates the first The state of the real trajectory at time t, and They represent the first The real trajectory at any moment Two-dimensional coordinates, and They represent the first The real trajectory at any moment The two-dimensional velocity components; where B represents the number of trajectories in a batch, and M represents the time length of a single trajectory. This represents the initial feature dimension of the trajectory; in the specific implementation, B=64, M=100. =4.

[0015] In this embodiment, the original radar trajectory data is generated through simulation using a hybrid motion model (CV and CTRV), and the target distance range is set to... The speed range is The training set turning rate range is Radar sampling interval =0.1 s, the total length of a single trajectory segment is 10 s; in generating the real trajectory Then, measurement noise and process noise following a normal distribution with a mean of 0 were added, where the standard deviations of distance measurement noise, azimuth noise, and acceleration noise were within the range of [8 m, 13 m]. and [8 m / s 2 13 m / s 2 The data is randomly selected within a certain range, and then the noisy polar coordinate observation data is converted into rectangular coordinate system data, thus obtaining a noisy radar observation trajectory sequence. .

[0016] Step 2, for Perform L-level discrete wavelet transform, using the Daubechies 2 wavelet basis, to obtain the th... The 4×L×2 dimensional wavelet statistical characteristics of the observed trajectory at time t, and compared with... After splicing, we get the first... The enhanced state vector of the observed trajectory at time t Thus, the feature-enhanced version of the first... observation trajectory This leads to the radar observation trajectory sequence with enhanced features. ; Calculate along the time dimension mean and standard deviation and utilize and right Normalization is performed to obtain the normalized i.e. observation trajectory Thus, the normalized radar observation trajectory sequence is obtained. ;in, This represents the dimension of the observed trajectory after feature enhancement; in practice, L=2. =4×2×2+4=20.

[0017] Step 3, as follows Figure 2 As shown in (a), a maneuver factor generator is constructed, and three parallel multilayer perceptrons are used to analyze the data. , and the standardized turning rate calculated from the two-dimensional velocity components Processing is performed to generate the representation of the first character. Scaling factor of the observed trajectory and the Offset factor of the observed trajectory ;in, These are the preset model feature dimensions; in specific implementation, =128.

[0018] Step 3.1: Use equations (1), (2), and (3) to respectively... Processing yields the first... Average turning rate of the observed trajectory Thus, to Standardization is performed to obtain the first Turning rate after standardization of the observed trajectory ;: (1) (2) (3) In equations (1), (2), and (3), Indicates the first The angle between the velocity direction of the observed trajectory at time t and the positive x-axis. Indicates the first The angle between the velocity direction of the observed trajectory at time t-1 and the positive x-axis. Indicates the first The instantaneous turning rate of the observed trajectory at time t.

[0019] Step 3.2, using equation (4) to... Processing yields the first... Intensity modulation characteristics of the observed trajectory : (4) In equation (4), This represents the first multilayer perceptron; This represents the activation function.

[0020] Step 3.3, using equation (5) to... Processing yields the first... The first offset component of the observed trajectory : (5) In equation (5), This represents the second multilayer perceptron.

[0021] Step 3.4, using equation (6) to... Processing yields the first... The second offset component of the observed trajectory : (6) In equation (6), This represents the third multilayer perceptron.

[0022] Step 3.5: Use equations (7) and (8) respectively to obtain the first... Scaling factor of the observed trajectory Translation factor : (7) (8) Step 4, as follows Figure 2 As shown in (b), a feature encoding module is constructed, and the feature encoding module is... Perform spatiotemporal feature extraction and output the first... Spatiotemporal characteristic sequence of observation trajectory .

[0023] Step 4.1, using equation (9) to... Feature extraction is performed to obtain the intermediate feature sequence. H 1 : (9) In equation (9), , This represents the two weight matrices of the linear mapping layer. , This represents the two bias vectors of the linear mapping layer. Represents the instance normalization function. It is an activation function.

[0024] Step 4.2, using formula (10) in Injecting the mobility factor, we obtain the first Spatiotemporal characteristic sequence of observation trajectory

[0025] (10) In equation (10), Representation layer normalization.

[0026] Step 5, as follows Figure 2As shown in (c), a Mamba tracking network with fused adjustment factors is constructed, consisting of L sequentially stacked Mamba layers, and... Modeling and maneuver feature modulation processing are performed to obtain the first... Enhanced feature sequence of observed trajectories Each Mamba layer includes: a bidirectional selection scanning module, such as... Figure 2 As shown in (d), there are two layer normalization modules; in the specific implementation, L=2.

[0027] Step 5.1, when When initializing the first Input features of each Mamba layer ;when At that time, the first The input feature of the Mamba is the first Output characteristics of each Mamba layer .

[0028] Step 5.2, the The first branch in the bidirectional selection scan module of the Mamba layer uses equation (11) to... Processing yields the first... Bidirectional scan output characteristics of the first branch of each Mamba layer : (11) In equation (11), and They represent the first The weight matrix and bias vector of the first branch of each Mamba layer, where DWConv is a convolution kernel of size ... Depth-separable convolution; , These represent forward and reverse scan operations, respectively. It is the activation function, where e is the dimensionality expansion exponent of the Mamba layer; in the specific implementation, e=3. .

[0029] No. The bidirectional selective scanning module of each Mamba layer uses equation (12) to perform a scan. Dynamic modulation is performed to obtain the first Features of the first branch modulation of the Mamba layer ; (12) No. The second branch in the bidirectional selection scanning module of the Mamba layer utilizes equation (13) to... Processing yields the first... Output characteristics of the second branch of the Mamba layer : (13) In equation (13), and They represent the first The weight matrix and bias vector of the second branch of each Mamba layer.

[0030] No. The bidirectional selection scanning module of the Mamba layer uses equation (14) to obtain the first Mamba layer. Scan output characteristics of each Mamba layer : (14) Step 5.3, the The first layer normalization module in a Mamba layer Using equation (15) to obtain the first The first intermediate feature of the Mamba layer : (15) In equation (15), This indicates a residual connection.

[0031] Step 5.4 Use equation (16) to... Dynamic modulation is performed to obtain the first The second intermediate feature of the Mamba layer : (16) Step 5.5, the The second layer normalization module in the Mamba layer Using equation (17) to obtain the first Enhanced temporal features of the output of each Mamba layer ; (17) In equation (17), Residual connection, FFN is a noisy feedforward network, and has: (18) In equation (18), This represents the weight matrix of the first linear layer of a feedforward network with a noisy Tanh activation function. Let denote the weight matrix of the second linear layer of the feedforward network with a noisy Tanh activation function. Let represent the noisy Tanh activation function, and we have: (19) In equation (19), N(0,1) is standard Gaussian noise.

[0032] Step 5.6, Assign to Then, return to step 5.2 and execute sequentially until... Until then, thus obtaining the first Enhanced feature sequences output by each Mamba layer And as .

[0033] Step 6, The input is processed in a linear layer to obtain the first... Normalized predicted state sequence of observed trajectories And using the mean and standard deviation right Perform inverse normalization to obtain the k-th predicted trajectory. ;in, For the first The predicted trajectory at time The state vector, For the first The predicted trajectory at time The Dimensional state values; for all trajectories within a batch, repeat steps 2 to 6 to obtain the predicted trajectory sequence. .

[0034] Step 7, for Perform L-level discrete wavelet transform using the same method as in step 2, and compare it with... By splicing, we obtain the feature-enhanced version of the first... Real trajectory Thus, the real trajectory sequence after feature enhancement is obtained. ;in, For the first The real trajectory at any moment The state vector, For the first The real trajectory at any moment The Dimensional state value; Step 8, based on and Construct the total loss function This is used to perform backpropagation training on a network consisting of a motor factor generator, a feature encoding module, a Mamba tracking network, and a mapping linear layer to update the network parameters until... The training continues until the target converges or reaches the preset number of training rounds, thus obtaining the optimal maneuvering target tracking model, which is used to estimate the state and predict the trajectory of the input maneuvering target. In this example, the initial learning rate is set to 0.001, the ADAMW optimizer is used, the decay factor is 0.01 per round, and the training period is set to 100.

[0035] Step 8.1: Construct the state prediction loss using equation (20). : (20) Step 8.2: Construct the feature reconstruction loss using equation (21). : (twenty one) Step 8.3: Construct the total loss function using equation (22). : (twenty two) In equation (23), and There are two hyperparameters. In this example, and Take values ​​of 0.8 and 0.2 respectively.

[0036] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0037] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0038] In this embodiment, the present invention provides 7 trajectory tracking results with different maneuverability, and the specific parameter settings for each trajectory are shown in Table 1.

[0039]

[0040] In this invention, the standard deviation of distance, azimuth angle, and acceleration noise for all trajectories is set to 10 m. 10 m / s 2 Each trajectory lasts for 100 seconds, with a sampling interval of 0.1 seconds. The state of each trajectory is divided into three stages. To test the generalization ability of the method, the turning rate of some trajectories is set to exceed the range of the training set. 100 Monte Carlo experiments were performed on each trajectory, and the mean and deviation of the root mean square error of trajectory position, velocity, and overall state were calculated to evaluate the tracking performance of the method.

[0041] This invention verifies the effectiveness of the Mamba network with fused adjustment factors in maneuvering target tracking from both quantitative and qualitative perspectives. Figures 3a-3f Table 2 shows a comparison of the results of this invention with other classic maneuvering target tracking methods under different test trajectories. Specifically, this invention is compared with previous target tracking methods IMM (Interacting Multiple Model), TBN (Transformer-Based Network), and TrMTT (Transformer-based Maneuvering Target Tracking). Tables 2 and 3 show the average and standard deviation of the RMSE results of position and velocity tracking for trajectories 1-7 under 100 Monte Carlo experiments, respectively. Table 4 shows the overall performance comparison results for each trajectory.

[0042] Table 2. Results (locations) of 100 Monte Carlo experiments for trajectories 1-7.

[0043] Table 3 shows the results (velocity) of 100 Monte Carlo experiments with trajectories 1-7.

[0044] Table 4 shows the results of 100 Monte Carlo experiments for trajectories 1-7 (all states).

[0045] The results show that the method of this invention achieves optimal performance on all test trajectories. Its errors are significantly lower than those of the comparative methods in both position and velocity dimensions, and its fluctuation range is smaller. Particularly in complex trajectories involving multi-stage maneuvering transitions, this invention maintains stable and accurate tracking performance, demonstrating its stronger modeling capability for non-stationary sequences. Experimental results fully demonstrate that under trajectory conditions with mixed maneuvering modes and frequent state changes, the method of this invention exhibits good tracking accuracy and stability, verifying the effectiveness and advancement of the proposed maneuvering factor modeling and Mamba network fusion method in complex maneuvering target tracking tasks, and possessing practical application value.

Claims

1. A method for tracking a maneuvering sensing target based on a Mamba tracking network, characterized in that, Includes the following steps: Step 1: Obtain a batch of radar observation trajectory sequences and the corresponding true trajectory sequence ,in, Indicates the first Observation trajectory, Indicates the first The state of the observed trajectory at time t, and They represent the first The observation trajectory at time Two-dimensional coordinates, and They represent the first The two-dimensional velocity components of the observed trajectory at time t; Indicates the first A real trajectory, Indicates the first The state of the real trajectory at time t, and They represent the first The real trajectory at any moment Two-dimensional coordinates, and They represent the first The real trajectory at any moment The two-dimensional velocity components; where B represents the number of trajectories in a batch, and M represents the time length of a single trajectory. The initial feature dimension representing the trajectory; Step 2, for Perform L-level discrete wavelet transform to obtain the first... The 4×L×2 dimensional wavelet statistical characteristics of the observed trajectory at time t, and compared with... After splicing, we get the first... The enhanced state vector of the observed trajectory at time t Thus, the feature-enhanced version of the first... observation trajectory This leads to the radar observation trajectory sequence with enhanced features. ; Calculate along the time dimension mean and standard deviation and utilize and right Normalization is performed to obtain the normalized i.e. observation trajectory Thus, the normalized radar observation trajectory sequence is obtained. ;in, This indicates the dimension of the observed trajectory after feature enhancement; Step 3: Construct a maneuvering factor generator and use three parallel multilayer perceptrons to analyze it. , and the standardized turning rate calculated from the two-dimensional velocity components Processing is performed to generate the representation of the first character. Scaling factor of the observed trajectory and the Offset factor of the observed trajectory ;in, The preset model feature dimensions; Step 4: Construct the feature encoding module and perform... Perform spatiotemporal feature extraction and output the first... Spatiotemporal characteristic sequence of observation trajectory ; Step 5: Constructing a Mamba tracking network with fused regulatory factors is... It consists of a series of stacked Mamba layers, and... Modeling and maneuver feature modulation processing are performed to obtain the first... Enhanced feature sequence of observed trajectories ; Step 6, The input is processed in a linear layer to obtain the first... Normalized predicted state sequence of observed trajectories And using the mean and standard deviation right Perform inverse normalization to obtain the k-th predicted trajectory. Thus, the predicted trajectory sequence is obtained. ;in, For the first The predicted trajectory at time The state vector; Step 7, for Perform L-level discrete wavelet transform using the same method as in step 2, and compare it with... By splicing, we obtain the feature-enhanced version of the first... Real trajectory Thus, the real trajectory sequence after feature enhancement is obtained. ;in, For the first The real trajectory at any moment The state vector; Step 8, based on and Construct the total loss function This is used to perform backpropagation training on a network consisting of a motor factor generator, a feature encoding module, a Mamba tracking network, and a mapping linear layer to update the network parameters until... The optimal maneuvering target tracking model is obtained by converging or reaching a preset number of training rounds, and is used to estimate the state and predict the trajectory of the input maneuvering target.

2. The mobile sensing target tracking method according to claim 1, characterized in that, Step 3 includes: Step 3.1: Use equations (1), (2), and (3) to respectively... Processing is performed to obtain the first... Average turning rate of the observed trajectory Thus, to Standardization is performed to obtain the first Turning rate after standardization of the observed trajectory : (1) (2) (3) In equations (1), (2), and (3), Indicates the first The angle between the velocity direction of the observed trajectory at time t and the positive x-axis. Indicates the first The angle between the velocity direction of the observed trajectory at time t-1 and the positive x-axis. Indicates the first The instantaneous turning rate of the observed trajectory at time t; Step 3.2, using equation (4) to... Processing is performed to obtain the first... Intensity modulation characteristics of the observed trajectory : (4) In equation (4), This represents the first multilayer perceptron; Indicates the activation function; Step 3.3, using equation (5) to... Processing is performed to obtain the first... The first offset component of the observed trajectory : (5) In equation (5), This represents the second multilayer perceptron; Step 3.4, using equation (6) to... Processing is performed to obtain the first... The second offset component of the observed trajectory : (6) In equation (6), This represents the third multilayer perceptron; Step 3.5: Using equations (7) and (8), obtain the first... Scaling factor of the observed trajectory and offset factor : (7) (8)。 3. The method according to claim 1, characterized in that, Step 4 includes: Step 4.1, using equation (9) to... Feature extraction is performed to obtain the intermediate feature sequence. : (9) In equation (9), , This represents the two weight matrices of the linear mapping layer. , This represents the two bias vectors of the linear mapping layer. Represents the instance normalization function. It is an activation function; Step 4.2, using equation (10) to... Injecting the mobility factor, we obtain the first Spatiotemporal characteristic sequence of observation trajectory : (10) In equation (10), Representation layer normalization.

4. The method according to claim 1, characterized in that, Each Mamba layer in step 5 includes: a bidirectional selection scanning module and two layer normalization modules; Step 5.1, when When, initialize the first Enhanced temporal features of the output of each Mamba layer ; Step 5.2, the first The first branch in the bidirectional selection scan module of the Mamba layer uses equation (11) to... Processing is performed to obtain the first... Bidirectional scan output characteristics of the first branch of each Mamba layer : (11) In equation (11), and They represent the first The weight matrix and bias vector of the first branch of each Mamba layer, where DWConv is a convolution kernel of size ... Depth-separable convolution; , These represent forward and reverse scan operations, respectively. It is the activation function, and e is the dimensionality expansion index of the Mamba layer; The first The bidirectional selective scanning module of each Mamba layer uses equation (12) to perform a scan. Dynamic modulation is performed to obtain the first Features of the first branch modulation of the Mamba layer ; (12) The first The second branch in the bidirectional selection scanning module of the Mamba layer utilizes equation (13) to... Processing is performed to obtain the first... Output characteristics of the second branch of the Mamba layer : (13) In equation (13), and They represent the first The weight matrix and bias vector of the second branch of each Mamba layer; The first The bidirectional selection scanning module of the Mamba layer uses equation (14) to obtain the first Mamba layer. Scan output characteristics of each Mamba layer : (14) Step 5.3, the The first layer normalization module in a Mamba layer Using equation (15) to obtain the first The first intermediate feature of the Mamba layer : (15) In equation (15), Indicates residual connection; Step 5.4, using equation (16) to... Dynamic modulation is performed to obtain the first The second intermediate feature of a Mamba layer ; (16) Step 5.5, the The second layer normalization module in the Mamba layer Using equation (17) to obtain the first Enhanced temporal features of the output of each Mamba layer ; (17) In equation (17), Residual connection, FFN is a noisy feedforward network, and has: (18) In equation (18), This represents the weight matrix of the first linear layer in a feedforward network with a noisy Tanh activation function. This represents the weight matrix of the second linear layer in a feedforward network with a noisy Tanh activation function. Let represent the noisy Tanh activation function, and we have: (19) In equation (19), N(0,1) is standard Gaussian noise; Step 5.6, Assign to Then, return to step 5.2 and execute sequentially until... Until then, thus obtaining the first Enhanced feature sequences output by each Mamba layer And as .

5. The method according to claim 1, characterized in that, In step 8 It is established according to the following steps: Step 8.1: Construct the state prediction loss using equation (20). : (20) In equation (20), For the first The predicted trajectory at time The Dimensional state value, For the first The real trajectory at any moment The Dimensional state value; Step 8.2: Construct the feature reconstruction loss using equation (21). : (21) Step 8.3: Construct the total loss function using equation (22). : (22) In equation (22), and There are two hyperparameters.

6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the maneuvering target tracking methods of claims 1-5, and the processor is configured to execute the programs stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the maneuvering target tracking method according to any one of claims 1-5.

Images