Intelligent tracking method for high maneuvering target in complex clutter environment

Abstract

The invention belongs to the technical field of target tracking, and discloses a high-maneuvering target intelligent tracking method in a complex clutter environment. Comprising the following steps: carrying out difference and normalization preprocessing on radar measurement; a spatial-temporal feature extraction network is constructed and trained, and the network integrates a convolutional neural network and a gating circulation unit and is used for extracting spatial-temporal features of target motion and predicting a state; a twinborn network based on VGG16 is constructed and trained for evaluating the similarity between a target and clutter distribution characteristics, and preliminary screening before association is realized; in the tracking stage, a space-time network is used for predicting the state, a twin network is used for matching and screening new measurement in a wave gate, and finally, optimal data association and track updating are completed through a Hungary algorithm. According to the method, state prediction and correlation judgment are respectively optimized through the deep learning network, and the tracking precision and robustness of the high-maneuvering target in a complex clutter environment are effectively improved.

Description

Technical Field

[0001] This application belongs to the field of target tracking technology, specifically relating to an intelligent tracking method for highly maneuverable targets in complex clutter environments. Background Technology

[0002] Radar target tracking technology is a core component of modern radar systems. It maintains stable and accurate estimates of the motion state of one or more targets during continuous observation, providing support for situational awareness and decision-making. Traditional tracking methods are typically based on a "filter-prediction-correlation" framework, and their performance largely depends on prior modeling of the target's motion patterns and accurate assumptions about the statistical characteristics of observation noise.

[0003] In practical applications, this technology faces two increasingly severe challenges. First, the maneuverability of targets has significantly increased. Modern high-speed aircraft, drones, and fast boats at sea can perform highly maneuverable actions such as rapid changes of direction, acceleration and deceleration, and evasive maneuvers, exhibiting strong nonlinear, non-stationary, and abrupt motion characteristics. Traditional state estimation algorithms based on fixed models (such as uniform speed, uniform acceleration, and coordinated turning models) (such as Kalman filtering and its extensions) are difficult to effectively match such complex dynamics, often leading to accumulated prediction biases due to model mismatch, ultimately resulting in decreased tracking accuracy or even target loss.

[0004] Secondly, radar operating environments are becoming increasingly complex, especially in coastal and shipborne maritime detection scenarios. The sea surface, ground features, and other background elements generate strong clutter that is time-varying, non-uniform, and non-Gaussian. After radar detection, this clutter forms numerous false alarm points that are similar to real targets in both spatial and velocity dimensions. Traditional data association methods (such as nearest neighbor and joint probability data association) largely rely on matching the geometric distance or statistical likelihood between the predicted target location and the current measurement. Under dense false alarm interference, these methods have limitations: firstly, the computational load of association increases exponentially with the number of false alarms, placing stringent demands on hardware computing power and resulting in high engineering implementation costs; secondly, association criteria relying solely on kinematic information are insufficient to effectively distinguish real targets from false alarms with similar characteristics, easily leading to false and missed associations, resulting in track confusion, the proliferation of false tracks, and breaks in the real track, disrupting tracking continuity.

[0005] Crucially, these two challenges are not isolated but rather coupled and mutually exacerbating. The unpredictability of the trajectory of highly maneuverable targets reduces the reliability of motion model predictions, leading to an expansion of the correlation gate based on predictions. This results in the inclusion of more false alarms, exacerbating the ambiguity of the correlation and the computational burden. In turn, the increase in false alarms interferes with the accurate learning and estimation of the true motion patterns of highly maneuverable targets, creating a vicious cycle.

[0006] Therefore, existing technologies lack a comprehensive solution that can coordinate the high maneuverability of targets with complex clutter interference, making it difficult to guarantee the tracking accuracy, stability, and real-time performance of radar systems in complex combat environments. Summary of the Invention

[0007] This application aims to address the problems of inaccurate predictions caused by motion model mismatch in traditional radar tracking methods when tracking highly maneuverable targets in complex clutter environments, as well as false correlations, high computational loads, and chaotic tracks caused by severe clutter interference. By integrating deep learning and optimization algorithms, this application provides an intelligent tracking method that can simultaneously improve state prediction accuracy and data correlation robustness, thereby achieving stable and accurate tracking of highly maneuverable targets.

[0008] To achieve the above technical objectives, this application specifically adopts the following technical solution: In one aspect of this application, a method for intelligent tracking of highly maneuverable targets in complex clutter environments is provided, comprising the following steps: S1. Obtain the true state of highly maneuverable targets after radar scanning. and its corresponding measurements Differential measurement sequences are obtained by performing differential processing on the measurement sequences. and the differential measurement sequence Perform normalization processing; compare the normalized differential measurement sequence with the corresponding differential target state. Composition of training samples ,in The difference between the normalized target true value and the normalized measurement. S2. Construct a spatiotemporal feature extraction network, and based on the training samples... The spatiotemporal feature extraction network is trained using a loss function; the spatiotemporal feature extraction network includes a spatial feature extraction module and a temporal feature extraction module connected in sequence; the spatial feature extraction module is a convolutional neural network used to extract the spatial features of the normalized differential test sequence, and the temporal feature extraction module is a gated recurrent unit network used to extract the temporal features of the normalized differential test sequence and output the predicted differential target state; S3. Obtain the distribution feature matrix of the target and clutter, and preprocess the distribution feature matrix to construct a distribution feature dataset; construct a Siamese network based on the VGG16 backbone network with two branches sharing weights; train the Siamese network based on the distribution feature dataset and the binary cross-entropy loss function, so that the Siamese network outputs a similarity value representing the similarity between two input distribution features; S4. For test time , with a length of Historical measurement sequence After performing differential and normalization processing, the results are input into the trained spatiotemporal feature extraction network to obtain the predicted differential target state. The time is obtained after inverse normalization and reduction. Target state prediction ; S5. At time For the new incremental measurement within the associated gate Measurements of the trajectory that were associated with the previous moment The corresponding distribution feature matrices are generated and input into the trained Siamese network; the similarity of the distribution features of the new measurement and the measurement of each track at the previous time is calculated, and the subset of new measurements with similarity higher than the preset threshold is selected. S6. For the selected subset of new incremental measurements and existing tracks, construct an association cost matrix based on the target state prediction value; apply the Hungarian algorithm to solve the association cost matrix to obtain the optimal matching relationship between the new measurements and tracks; based on the matching relationship, use the spatiotemporal feature extraction network to update the track state.

[0009] In one implementation, the differential processing in step S1 includes: processing the time period... Measurement sequence within Differential processing is performed to extract the target's dynamic features and obtain the differential measurement sequence. : .

[0010] In one implementation, the normalization process in step S1 uses the min-max normalization method to map the numerical range of the differential measurement sequence to the interval [-1, 1].

[0011] In one implementation, the loss function used in step S2 is: :

[0012] in, For the true differential target state, To predict the differential target state, This represents the number of samples.

[0013] In one implementation, the preprocessing of the distribution feature matrix in step S3 includes: first, normalizing the distribution feature matrix, and then adjusting the matrix to a preset size using a bicubic interpolation algorithm.

[0014] In one implementation, the loss function used in training the Siamese network in step S3 is... This is the binary cross-entropy loss function.

[0015] In one implementation, the preset threshold in step S5 is 0.9.

[0016] In one implementation, the associated cost matrix in step S6 elements in for:

[0017] in, For the filtered first One new measurement, H For the observation matrix, For the first The predicted status of the flight path.

[0018] In one implementation, the motion model of the highly maneuverable target includes a uniform speed model, a uniform acceleration model, a coordinated turning model, and any combination thereof.

[0019] In another aspect of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent tracking method for highly maneuverable targets in complex clutter environments.

[0020] In another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the intelligent tracking method for highly maneuverable targets in complex clutter environments.

[0021] The beneficial effects of this application are as follows: 1) By constructing a spatiotemporal feature extraction network to learn differential measurement sequences, the network can effectively capture the dynamic change features of highly maneuverable targets, achieve more accurate prediction of target state, and thus overcome the problem of mismatch between traditional fixed motion models and actual target maneuverability.

[0022] 2) By using a twin network based on distribution feature matching, the similarity of the distribution features of measurements and tracks can be measured. This can effectively distinguish between real targets and clutter false alarms before data association, reducing the probability of false alarms entering the subsequent association process and improving the reliability of association.

[0023] 3) The serial processing mechanism of first performing feature pre-association filtering by the Siamese network and then performing optimal allocation by the Hungarian algorithm is adopted. While ensuring the association accuracy, the number of candidate measurements that the Hungarian algorithm needs to process is significantly reduced, the overall computational complexity is reduced, and the real-time performance of the system is improved.

[0024] 4) The synergistic effect of the method in this application enables the tracking of highly maneuverable targets to achieve higher positioning accuracy, lower false track rate and better track continuity stability in strong clutter interference environment, with overall performance superior to traditional tracking methods. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the intelligent tracking method for highly maneuverable targets in complex clutter environments according to an embodiment of this application. Figure 2 This is the result of matching the target and target distribution features in the embodiments of this application; Figure 3 This is the result of matching target and false alarm distribution features in the embodiments of this application; Figure 4 This is a comparison of the tracking position accuracy of highly maneuverable targets with traditional methods in the embodiments of this application. Detailed Implementation

[0026] The technical solution of this application will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art will understand that the embodiments described below are only some embodiments of this application, not all embodiments, and are only used to illustrate this application, and should not be regarded as limiting the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] To address the problems of model mismatch, computational complexity, and high false alarm correlation rate in existing radar tracking technologies under complex clutter and highly maneuverable target scenarios, this application proposes an intelligent tracking method based on feature decoupling and dual-network collaboration.

[0028] Existing technologies employ a single, rigid model to simultaneously handle target state estimation and data association. In highly maneuverable situations, the target motion model struggles to accurately match, leading to large prediction errors. Furthermore, in dense clutter environments, association methods based on geometric distance or simple statistical characteristics are highly susceptible to false alarms, resulting in erroneous associations. These two factors are coupled, further amplifying the degradation in tracking performance.

[0029] To address the aforementioned issues, this application decouples the unified tracking task into two relatively independent and focused subtasks, and constructs dedicated neural networks for each subtask to optimize them. Finally, robust tracking is achieved through a collaborative process.

[0030] The first subtask focuses on high-maneuverability state prediction. Traditional methods rely on pre-defined motion models (such as CV and CA), which are difficult to adapt to sudden maneuvers. This application abandons the direct fitting of the overall motion model and instead learns the mapping rules of "state change quantities". Specifically, the measurement sequence is differentially processed to extract the dynamic change features of distance and velocity, and a spatiotemporal feature extraction network composed of convolutional neural networks (CNN) and gated recurrent units (GRU) is used for learning. This network directly establishes the mapping relationship from the historical measurement difference sequence to the current state prediction difference component. By learning from "change" rather than "absolute state", the network can more flexibly capture the non-stationary and nonlinear motion patterns of high-maneuverability targets, reduce the dependence on prior models, and improve the adaptability of state prediction to maneuvers.

[0031] The second subtask focuses on robust association in cluttered environments. Traditional association methods (such as nearest neighbor) are susceptible to interference in clutter. This application introduces image recognition concepts, transforming the association problem into a distribution feature matching problem. Specifically, the energy distributions of the target and measurements in the range-Doppler domain are constructed as a two-dimensional image, and deep features are extracted and similarity is calculated using a Siamese network based on the VGG16 architecture. This network, after training, can distinguish between the continuous distribution features of the target's true trajectory and the random distribution features of clutter false alarms. Before association, this Siamese network is used to filter candidate measurements, retaining only those with similar distribution features to existing trajectories for subsequent association steps. This effectively filters out most false alarms, reducing the computational burden and risk of misassociation in the data association stage.

[0032] Ultimately, the two sub-tasks are integrated through process collaboration. First, a spatiotemporal feature extraction network is used to predict the state of existing tracks. Then, within the associated gate, a Siamese network is used to pre-screen new measurements based on their distribution characteristics. Finally, for the selected measurements and tracks, classic methods such as the Hungarian algorithm are used to achieve the final cost-optimal allocation. This sequential "prediction-screening-association" architecture allows the two networks to perform their respective functions, overcoming the challenges of high maneuverability and clutter interference, and jointly improving the accuracy, robustness, and efficiency of the overall tracking system in complex environments.

[0033] Example Reference Figure 1 As shown, the intelligent tracking method for highly maneuverable targets in complex clutter environments in this embodiment is implemented through the following steps: Step 1: Obtain the training and testing samples required for the spatiotemporal feature extraction network. (1.1) Creating the true state of a highly maneuverable target after radar scanning ,in This indicates that at time t, the first... The true state of the target on the track. Represents the distance components on the x and y axes. Represents the velocity components on the x and y axes; corresponding measurements Represented as ,in, Represented as an observation matrix, Represented as observation noise, This is represented by noise introduced by complex clutter. The created high-maneuverability target dataset includes constant velocity (CV), constant acceleration (CA), coordinated turn (CT), and arbitrary mixed motions to represent as many abrupt changes in high-maneuverability as possible.

[0034] (1.2) The measurement sequences of highly maneuverable targets are subjected to differencing and normalization. For time periods... Measurement sequence within Differential processing is performed to extract the target's dynamic features and obtain the differential measurement sequence. The solution is as follows: .

[0035] Differential operations can effectively suppress the interference of noise, while simultaneously decoupling features, separating the influence of initial distance and velocity, and achieving dynamic feature enhancement by extracting dynamic distance and velocity features. (Differential measurement sequence) The input differential measurement sequence contains different measurement features, such as distance and velocity, each with its own dimensions. The distance feature is significantly larger in magnitude than the velocity feature, causing crucial information contained in the velocity feature to be masked by the distance feature, ultimately negatively impacting the network's output. To effectively eliminate interference from different features due to their varying dimensions, it is necessary to perform differential subtesting on the input sequence. Perform data normalization preprocessing. Before training, the dataset contains... Each flight path The track can be divided into A length of The time period, that is Individual difference component measurement sequence Each differential measurement sequence There are a total of measurements If there are 1 difference subtest, then the dataset has 100,000 subtests. Individual difference component measurement Differential measurement set ,Right now for The first of them Individual difference measurements. The normalization criteria are as follows:

[0036]

[0037]

[0038]

[0039] in, , , , Represent the difference measurement sets respectively Corresponding to all distance and velocity difference components on the x and y axes; , , , They represent the first i The difference components measure the distance and velocity difference components of the sample on the x and y axes. , , , These represent the corresponding normalized distance and velocity difference components, respectively; This indicates the maximum value in this batch of data. This represents the minimum value in this batch of data. After min-max normalization, the difference set of measurements... The values ​​will be distributed between [-1, 1].

[0040] End-to-end operations complicate network mapping, increasing computational complexity and potentially reducing accuracy. Furthermore, directly learning a unified feature representation may prevent the model from capturing truly useful information, while decoupling features often yields more robust representations. Therefore, the desired output at the GRU output is the difference between the ground truth state of the target at time t and the corresponding measurement, expressed as follows:

[0041] Among them, the differential target state The normalized true state of the target at time t. Normalized measurement at time t The difference.

[0042] Normalized differential measurement sequences input at the same time Decoupling feature output differential target state Composition of training samples The expression is as follows:

[0043] Where time t is greater than time interval k, .

[0044] Step 2, construct the spatiotemporal feature extraction network: (2.1) Construct a convolutional network (CNN) module C for spatial feature extraction and a gated recurrent unit (GRU) module G for temporal feature extraction. The spatial feature extraction module C consists of convolutional layers, which process the normalized differential measurement sequence. Spatial feature extraction is mainly performed by convolution using convolution kernels.

[0045] Where tanh is the activation function, which compresses the range of the extracted data to [-1, 1]. For the first The filter in the first... The weights of each input channel, For the first The bias of each filter.

[0046] (2.2) The temporal feature extraction module G is composed of GRU modules, including reset gates and update gates. As time progresses, it can extract features from the normalized differential measurement sequence. Temporal feature extraction is performed. The internal calculations of the temporal feature extraction module G are as follows: Update Gate :

[0047] in, This indicates updating the gate output parameters when the output vector of the input convolutional network module C is updated. When (the feature vector output by the spatial feature extraction module at the current time step), the hidden state at the previous time step was: The updated gate weight matrix is ​​as follows: The corresponding bias is The updated gate parameters are obtained by performing the following calculations. The Sigmoid activation function has the following expression: ; The function takes its input value Compressing to the range [0, 1] is to determine historical information. The amount of data retained and the amount of new information introduced when updating the gate. The closer it gets to 1, the more historical information... The more you retain, the better.

[0048] The output of the reset gate at time t :

[0049] The weight matrix of the reset gate is as follows: The corresponding bias is This calculation aims to determine the impact of the loss or retention of historical information on candidate states, when... The closer it gets to 1, the more historical information... The more you retain, the more you discard.

[0050] Candidate state :

[0051] The weight matrix of the candidate states is: The corresponding bias is , For the Hadamard product, by resetting the gate With historical hidden state Perform Hadamard product calculation, select the required historical features, and filter out redundant features. The activation function has the following expression: ; The tanh activation function converts candidate hidden states. The range is compressed to [-1, 1].

[0052] Status Update: Hidden candidate state and historical hidden state Update the current state by updating the gate-weighted fusion. The calculation formula is as follows:

[0053] Updated status The target state is mapped to the desired output via a fully connected layer. This application focuses on analyzing the accuracy of the distance information predicted by the target state, so the output of the fully connected layer is the distance components of the x-axis and y-axis.

[0054] (2b) Constructing the loss function of the spatiotemporal feature extraction network :

[0055] In order to better utilize the normalized difference measurement sequence Fit to the difference target state loss function The Mean Squared Error (MSE) function is used, where For the true value, is the predicted value, and B is the sample size.

[0056] Step 3: Train the spatiotemporal feature extraction network: (3.1) The training sample dataset constructed in step 1 The data is sequentially input into a pre-defined spatiotemporal feature extraction network. This network processes the input samples layer by layer based on its internal multi-layer spatiotemporal convolutional structure (convolutional layers and GRU layers) to extract the spatiotemporal correlation features contained in the data.

[0057] (3.2) Parameter optimization: Based on the prediction results output by the network forward propagation and the corresponding true labels of the samples, the mean squared error loss function constructed in step 2 is used to calculate the current prediction error value. Subsequently, the calculated loss gradient is propagated back from the output layer to the input layer layer by layer through the error backpropagation algorithm, and the selected Adam optimization algorithm is applied to iteratively adjust and update all learnable parameters in the spatiotemporal feature extraction network model, aiming to minimize the value of the MSE loss function, thereby improving the prediction accuracy and generalization ability of the network model.

[0058] Step 4: Use the trained spatiotemporal feature extraction network to measure the target on the test set. Perform a tracking test. Use a measurement sequence of length k. Perform the difference processing and min-max normalization processing as in step (1b) to obtain the normalized difference sequence. The data is fed into a pre-trained spatiotemporal feature extraction network, and the differential prediction state of the target is obtained through indirect end-to-end hidden mapping. By inverse normalization: ; ; Obtain the differential prediction state of the original measurement The predicted state is obtained after restoration. The calculation is as follows: .

[0059] Step 5: Construct a dataset associated with distribution feature data. (5.1) Range-Doppler data is obtained by preprocessing the echo data containing highly maneuvering targets and clutter collected through field measurements. A target with a measurement scale of 5×5 is constructed with the target and clutter as the center. and noise false alarm The distribution feature matrix is ​​used as input.

[0060] (5.2) In order to better extract the distribution features of the target and the false, the distribution feature matrix is... and Preprocessing is performed. First, normalization is applied:

[0061] Where M represents the input distribution feature matrix. or ; It is the value of the element at position (i,j) in the original matrix; It is a very small constant ( This linearly maps each value to the interval [0,1].

[0062] Let the normalized matrix be N (of size L×W), and the target image size be 105×105. Bicubic interpolation achieves scaling through a weighted average of neighboring pixels. The target location... The value is:

[0063] and These are the coordinates of the nearest neighbors in the source matrix (covering a 4×4 neighborhood). The bicubic kernel function W(s) is defined as:

[0064] parameter Take a fixed value of -0.5. It represents the normalized distance between grid points in the source image and interpolation points in the target image in a certain dimension.

[0065] By using bicubic weighted smoothing to enlarge the image, details and edges are preserved, resulting in the target distribution feature dataset T-chapter and the false alarm dataset C-chapter.

[0066] Step 6: Construct a Siamese network model based on distribution feature matching. (6.1) The backbone network for feature extraction adopts the VGG16 model, which demonstrates excellent performance in classification tasks. Transfer learning is used to freeze the weights of the backbone network. Each Siamese branch is a complete VGG16 structure and shares weights. Each branch outputs a one-dimensional feature vector. The feature vectors extracted from the two branches are combined, and their L1 norm is calculated to obtain the distance, thus determining their similarity. Finally, the similarity A is mapped to the [0,1] interval using the sigmoid function.

[0067] (6.2) Construct the loss function of the Siamese network V Loss function The BCEWithLogitsLoss loss function is used, and its expression is as follows:

[0068] in, This represents the raw logits (inactive values) output by the model. Indicates the target label (value is 0 or 1). It is the sigmoid activation function. Compress logits to the (0,1) interval. Calculate the log probability of the positive class prediction. Calculate the log probability of the negative class prediction by... and Weighted summation is used to calculate binary cross-entropy.

[0069] Step 7: The training method for the Siamese network V is similar to that in Step 3, and the loss function is adopted. The SGD optimization algorithm is used to iteratively adjust and update all learnable parameters in the Siamese network V model.

[0070] Step 8: Use the trained Siamese network V to match and verify the target features and false alarm features in the test set. After processing the target feature distribution matrix and the false feature distribution matrix as described in (5.2), feed them into the trained Siamese network V. The network outputs the similarity A between the two. Set a threshold; when... When the two are in a certain condition, they can be associated with each other; otherwise, they cannot be associated.

[0071] Step 9, Intelligent Tracking Algorithm Execution Process.

[0072] (9.1) At time t, based on the j track measurements accumulated from time tk to time t-1 Perform the difference and normalization operations of step 1 on the sample to obtain the processed normalized difference subsequence. After being fed into a spatiotemporal feature extraction network, the network calculates and outputs the differential prediction state of the corresponding sequence. After restoration, the predicted state of each track sequence is obtained. .

[0073] (9.2) First, perform distribution characteristic correlation. Measure the new increment within the correlation gate at time t. Where G represents the number of newly added measurements within the associated gate at time t, and g represents the g-th measurement. Measurements already associated with the track at time t-1 are... tr represents the corresponding track number, with a maximum value of TR, and is related to newly added measurements. Pairwise matching is performed to form a set to be processed. Then, after normalization and bitriangular interpolation as described in step 5, a 105×105 single-channel distribution image that conforms to the input of the Siamese network V is generated. This image is then fed into the Siamese network V, and the output of the Siamese network V shows similarity augmentation measurements. Measurements associated with the track If the similarity of the distribution characteristics is less than 0.9, no further association is performed, and the filtered new incremental measurement is obtained. .

[0074] (9.3) If the similarity is higher than 0.9, the filtered... The newly added measurements and existing TR tracks are correlated using the Hungarian algorithm.

[0075] (9.3.1) Construction of cost matrix first, With track The associated costs are The formula is as follows: ; This cost reflects the confidence level of the match, and typically requires G'=TR, meaning that the number of new measurements matches the number of trajectories. Therefore, when there is a mismatch, the maximum value of the two is used for completion.

[0076] in: It is a measurement The predicted state vector. H is the observation matrix. The cost matrix C is... A matrix whose elements express and Matching cost; smaller This indicates a better match.

[0077] (9.3.2) Using the cost matrix C as input, apply the Hungarian algorithm to solve for the minimum cost allocation. The algorithm steps are as follows: 1) Line rules: For each line Calculate the minimum value among all elements in the row. Then, for each element in that row... Subtract the minimum value:

[0078] Ensure that each row contains at least one zero element.

[0079] 2) Column reduction: For each column tr, calculate the minimum value among all elements in that column. Then each element in that column Subtract the minimum value:

[0080] Further increase the number of zero elements.

[0081] 3) Cover all zero elements: Cover all zero elements using the minimum number of rows (e.g., by row or column marking). If the coverage number equals tr, the optimal match is found; otherwise, adjust the matrix. Find the smallest uncovered element δ.

[0082] Subtract δ from the uncovered rows, and add δ to the covered columns:

[0083] Repeat this process until a perfect match is found.

[0084] 4) Output matching results The algorithm outputs an allocation matrix X, where each row of matrix X corresponds to a new incremental index after filtering. The column corresponds to the existing track index tr. Elements in the matrix The value is defined as: if the algorithm decides to measure... Assigned to track tr, then Otherwise, it is 0. The total minimum cost is determined by... Provided.

[0085] (9.4) Post-processing of association After matching is completed, the system updates the track status: if a measurement is assigned, a spatiotemporal feature extraction network is used for tracking filtering. Unassigned measurements may be initialized as new tracks, and tracks that have not been updated may be terminated.

[0086] To verify the actual effect of this application, the following experiments were conducted, and the results were analyzed.

[0087] 1. Experimental verification conditions The experimental data was collected from a phased array radar deployed in a coastal environment. This environment is complex, with strong clutter interference caused by sea surface reflection. The targets selected for tracking in the experiment were mainly small targets with high maneuverability, such as fast boats and low-altitude UAVs. Their movement patterns included rapid changes of direction, acceleration, and deceleration, which effectively tested the performance of the tracking algorithm in complex scenarios.

[0088] 2. Experimental Results and Analysis Validation of distribution feature matching effectiveness: Figure 2 This paper demonstrates the results of the Siamese network in this application for matching the distribution features of continuous track points belonging to the same real target. In the figure, two input grayscale images represent the energy distribution of the target in the range-Doppler domain at two adjacent time points, with the grayscale depth representing energy intensity. The Siamese network calculates a similarity score A close to 1. This indicates that for continuous measurements originating from the same target, their deep distribution features are highly consistent, and the network can accurately identify them.

[0089] Verification of false alarm filtering capability: Figure 3The results of matching the distribution features of a real target measurement and a false alarm measurement caused by clutter, as demonstrated by the Siamese network of this application, are shown. As shown in the figure, there are significant differences in the energy distribution patterns of the two. The similarity score A calculated by the network is close to 0. This proves that the trained Siamese network can effectively distinguish between the real features of the target and the random features of clutter, thereby filtering out such false alarms before association.

[0090] Overall tracking accuracy comparison: Figure 4 This paper presents a comparison of the target tracking and positioning accuracy of the complete intelligent tracking method provided by this invention with that of traditional tracking methods (traditional method 1-KF, traditional method 2-EKF, and traditional method 3-LSTM) under the same experimental data. The curves in the figure show that throughout the tracking period, the root mean square error of the position estimation of the trajectory obtained by the method of this invention is consistently lower than that of the traditional methods. This result demonstrates that the strategy of this invention, which adapts to highly maneuverable motion through a spatiotemporal feature extraction network and combines it with a Siamese network for pre-association filtering, can significantly improve the tracking and positioning accuracy of highly maneuverable targets in complex clutter environments.

[0091] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and these are all within the scope of protection of this application.

Claims

1. A method for intelligent tracking of highly maneuverable targets in complex clutter environments, characterized in that, Includes the following steps: S1. Obtain the true state of highly maneuverable targets after radar scanning. and its corresponding measurements ; Differential measurement sequences are obtained by performing differential processing on the measurement sequences. and the differential measurement sequence Perform normalization processing; The normalized differential measurement sequence is compared with the corresponding differential target state. Composition of training samples ,in The difference between the normalized target true value and the normalized measurement. S2. Construct a spatiotemporal feature extraction network, and based on the training samples... The spatiotemporal feature extraction network is trained using a loss function; the spatiotemporal feature extraction network includes a spatial feature extraction module and a temporal feature extraction module connected in sequence; the spatial feature extraction module is a convolutional neural network used to extract the spatial features of the normalized differential test sequence, and the temporal feature extraction module is a gated recurrent unit network used to extract the temporal features of the normalized differential test sequence and output the predicted differential target state; S3. Obtain the distribution feature matrix of the target and clutter, and preprocess the distribution feature matrix to construct a distribution feature dataset; construct a Siamese network based on the VGG16 backbone network with two branches sharing weights; train the Siamese network based on the distribution feature dataset and the binary cross-entropy loss function, so that the Siamese network outputs a similarity value representing the similarity between two input distribution features; S4. For test time , with a length of Historical measurement sequence After performing differential and normalization processing, the results are input into the trained spatiotemporal feature extraction network to obtain the predicted differential target state. The time is obtained after inverse normalization and reduction. Target state prediction ; S5. At time For the new incremental measurement within the associated gate Measurements of the trajectory that were associated with the previous moment The corresponding distribution feature matrices are generated and input into the trained Siamese network; the similarity of the distribution features of the new measurement and the measurement of each track at the previous time is calculated, and the subset of new measurements with similarity higher than the preset threshold is selected. S6. For the selected subset of new incremental measurements and existing tracks, construct an association cost matrix based on the target state prediction value; apply the Hungarian algorithm to solve the association cost matrix to obtain the optimal matching relationship between the new measurements and tracks; based on the matching relationship, use the spatiotemporal feature extraction network to update the track state.

2. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, The differential processing in step S1 includes: processing the time period Measurement sequence within Differential processing is performed to extract the target's dynamic features and obtain the differential measurement sequence. : ； The normalization process uses the min-max normalization method to map the numerical range of the differential measurement sequence to the interval [-1, 1].

3. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, The loss function used in step S2 is: : in, For the true differential target state, To predict the differential target state, This represents the number of samples.

4. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, Step S3 involves preprocessing the distribution feature matrix, which includes: first, normalizing the distribution feature matrix, and then adjusting the matrix to a preset size using a bicubic interpolation algorithm.

5. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, In step S3, the loss function used in training the Siamese network is... This is the binary cross-entropy loss function.

6. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, The preset threshold mentioned in step S5 is 0.

9.

7. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, The associated cost matrix mentioned in step S6 elements in for: in, For the filtered first One new measurement, H For the observation matrix, For the first The predicted status of the flight path.

8. The intelligent tracking method for highly maneuverable targets in complex clutter environments according to claim 1, characterized in that, The motion model of the highly maneuverable target includes the uniform speed model, the uniform acceleration model, the coordinated turning model, and any combination thereof.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent tracking method for highly maneuverable targets in complex clutter environments as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the intelligent tracking method for highly maneuverable targets in complex clutter environments as described in any one of claims 1 to 8.

Images