An Enhanced Transformer 3D Target Tracking Method
By using the temporal convolutional hybrid embedding module and sliding window mechanism of the enhanced Transformer network, combined with training datasets of various maneuvering modes, the robustness problem of radar target tracking in complex scenarios is solved, achieving high-precision and real-time 3D target tracking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2026-03-25
- Publication Date
- 2026-05-26
Smart Images
Figure CN122085262A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar signal recognition technology, specifically relating to an enhanced Transformer three-dimensional target tracking method. Background Technology
[0002] Radar target tracking technology has broad application prospects in civil air traffic control, drone surveillance, meteorological monitoring, and environmental perception for autonomous driving. With the continuous improvement of the maneuverability of various aircraft and ground vehicles, target movements are characterized by suddenness and frequent changes in movement patterns, placing higher demands on the accuracy and robustness of radar tracking systems.
[0003] Traditional radar target tracking methods are primarily based on state-space models. They establish the target's kinematic equations and the radar's measurement equations, then use filtering algorithms to estimate the target's state. Kalman filtering and its derivatives are classic target tracking methods. However, Kalman filtering requires a linear system model and Gaussian noise distribution, making it difficult to apply directly to nonlinear maneuvering scenarios. To address this issue, extended Kalman filtering and unscented Kalman filtering have been proposed. These methods achieve state estimation by linearizing the nonlinear system or using deterministic sampling strategies. However, when the target undergoes violent maneuvers, linearization errors can significantly reduce filtering accuracy.
[0004] Interactive multi-model methods improve the algorithm's adaptability to maneuvering targets to some extent by running multiple basic motion models in parallel and using Markov chains to calculate the matching probabilities of each model for weighted fusion. However, this method has inherent limitations: First, the pre-set finite model set cannot cover all maneuvering modes of real targets, and model mismatch will occur when the target's motion mode exceeds the range of the model set; second, the update of model probabilities depends on the prior-set transition probabilities, and when the target undergoes continuous and violent maneuvers, the probability adjustment often lags behind the actual changes in the target's state, leading to increased tracking errors or even divergence. In addition, interactive multi-model methods require a high level of prior knowledge of noise statistical characteristics, which is difficult to obtain accurately in complex real-world environments.
[0005] In recent years, deep learning-based target tracking methods have attracted widespread attention. Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks have been explored for application in maneuvering target tracking due to their ability to process temporal data. These methods progressively pass temporal information through hidden state, enabling them to learn the target's motion patterns from historical observations. However, LSM networks employ a chain-like structure for sequential feature extraction, which suffers from gradient vanishing or exploding problems when processing long sequences, leading to a decline in long-range memory capacity. Furthermore, these networks lack the ability to perceive global information within a sequence, making it difficult to capture motion correlations between distant frames during target maneuvers. Their generalization ability and robustness in complex maneuvering scenarios need improvement.
[0006] The Transformer architecture, with its self-attention mechanism, demonstrates powerful global dependency capture capabilities in sequence modeling tasks. This architecture abandons the stepwise processing approach of recurrent neural networks, enabling parallel modeling of dependencies between any positions in a sequence and effectively solving the long-range memory problem. However, the standard Transformer has the following shortcomings when processing radar measurement data: First, it uses time-by-time independent linear projection for feature mapping of the input sequence, failing to fully utilize the local motion correlation between adjacent frames, making it difficult to effectively distinguish local motion features from measurement noise in the feature space; second, the self-attention mechanism needs to simultaneously handle local feature extraction and global dependency modeling. When measurement noise is strong, the calculation of attention weights is easily affected by noise, impacting the stability of feature representation and the accuracy of subsequent state estimation.
[0007] To address the aforementioned issues, there is an urgent need to develop a radar target tracking method that can maintain high-precision tracking performance and possess good robustness in complex maneuvering scenarios. Summary of the Invention
[0008] To address the aforementioned problems in the existing technology, this invention provides an enhanced Transformer 3D target tracking method. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides an enhanced Transformer 3D target tracking method, comprising: Based on radar systems deployed in real physical scenarios, radar measurement data containing three-dimensional moving target information is collected in real time to form a continuous data stream; By using a sliding window, a data segment of length L, ending at the current time, is extracted from the data stream to form an input sequence; The input sequence is input into the trained enhanced Transformer network. After local-global feature extraction and feature fusion are performed by the temporal convolutional hybrid embedding module inside the model, the internal Transformer encoder is used to perform global dependency modeling and target state estimation on the fused features, and outputs the target state estimate value at the current time. When the data for the next moment arrives, the data from the earliest moment in the sliding window is removed, and the data for the next moment is added to form an updated input sequence. The updated input sequence is then fed into the enhanced Transformer network for processing to achieve real-time continuous tracking of the target. In the model training phase, the enhanced Transformer network uses a training dataset jointly generated by a uniform linear motion model and a cooperative turning motion model, and superimposed with noise that matches the actual radar measurement noise characteristics, for iterative training.
[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the issue of poor robustness in target tracking under complex maneuvering scenarios in existing radar measurement data processing methods, this invention provides an enhanced Transformer 3D target tracking method. This method utilizes a temporal convolutional hybrid embedding module within a trained enhanced Transformer network to extract and fuse local and global features from the input sequence truncated by a sliding window. A Transformer encoder is then used to perform global dependency modeling and target state estimation on the fused features, achieving adaptive feature learning and high-precision state estimation for 3D maneuvering targets, thus avoiding the dependence on pre-defined motion models found in traditional methods. Simultaneously, a sliding window mechanism is employed to dynamically update the input sequence and recursively process it within the real-time data stream, ensuring the real-time and continuous nature of online tracking. Furthermore, the training dataset is jointly generated by a uniform linear motion model and a cooperative turning motion model, superimposed with noise matching the actual radar measurement noise characteristics. This allows the network to fully learn motion features under various maneuvering modes, significantly improving the model's generalization ability and tracking robustness in unknown maneuvering scenarios. Attached Figure Description
[0010] Figure 1 This is a flowchart illustrating the enhanced Transformer 3D target tracking method provided in an embodiment of the present invention; Figure 2 This is a network structure composition diagram of the temporal convolutional hybrid embedding module provided in an embodiment of the present invention; Figure 3 This is a graph showing the motion trajectory corresponding to number 1 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 4This is a graph showing the motion trajectory corresponding to number 2 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 5 This is a graph showing the motion trajectory corresponding to number 3 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 6 This is a graph showing the motion trajectory corresponding to number 4 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 7 This is a graph showing the motion trajectory corresponding to number 5 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 8 This is a graph showing the motion trajectory corresponding to number 6 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 9 This is a graph showing the motion trajectory corresponding to number 7 in the embodiments of the present invention and the average root mean square error curve of each model after processing; Figure 10 This is a graph showing the motion trajectory corresponding to number 8 in the embodiments of the present invention and the average root mean square error curve of the position after processing by each model. Detailed Implementation
[0011] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0012] The enhanced Transformer 3D target tracking method proposed in this invention will now be described in detail with reference to the accompanying drawings.
[0013] Figure 1 This is a flowchart illustrating the enhanced Transformer 3D target tracking method provided in an embodiment of the present invention. Figure 1 As shown, the method includes S1-S4; specifically: S1: Based on a radar system deployed in a real physical scene, radar measurement data containing three-dimensional moving target information is collected in real time to form a continuous data stream.
[0014] For example, a three-dimensional maneuvering target refers to an object that moves in three-dimensional space and possesses maneuverability. During its movement, the target's position changes over time along the x-axis (eastward), y-axis (northward), and z-axis (upward), and its motion state (including velocity direction, velocity magnitude, or trajectory curvature) may undergo sudden or regular changes. The information of this three-dimensional maneuvering target includes the target range, azimuth, and elevation angles output by the radar in real time, as well as optional Doppler velocity and signal-to-noise ratio parameters, used to characterize the target's position and motion state in three-dimensional space.
[0015] Here, before performing the sliding window extraction in S2, the method also includes converting the radar measurement data in polar coordinates into measurement data in Cartesian coordinates through an unbiased transformation.
[0016] S2: Using a sliding window, extract a data segment of length L from the data stream, ending at the current time, to form the input sequence.
[0017] S3: Input the input sequence into the trained enhanced Transformer network. After performing local-global feature extraction and feature fusion using the temporal convolutional hybrid embedding module inside the model, the internal Transformer encoder performs global dependency modeling and target state estimation on the fused features, and outputs the target state estimate at the current time step.
[0018] It should be noted that this section only describes the operation of the enhanced Transformer network during the actual inference phase. The training process of the enhanced Transformer network will be described later.
[0019] Figure 2 This is a network structure diagram of the temporal convolutional hybrid embedding module provided in an embodiment of the present invention. For example... Figure 2 As shown, the temporal convolutional hybrid embedding module includes parallel linear mapping branches and temporal convolution branches, as well as a feature fusion submodule; the temporal convolutional hybrid embedding module is configured to perform the following operations: In the linear mapping branch, the input sequence is linearly transformed, and the output is a linear feature that preserves global information. In the temporal convolution branch, the input sequence is subjected to two cascaded one-dimensional temporal convolutions to aggregate the short-term motion information of the target in the local neighborhood and suppress measurement noise, and output local motion features. In the feature fusion submodule, the linear features and local motion features are element-wise added and layer normalized, and the fused features are output.
[0020] Specifically, the linear branch fully preserves the global information and numerical scale benchmark of the input residual sequence at the representation level, without performing any local aggregation or compression operations on the temporal structure of the sequence. This allows the subsequent self-attention mechanism to perform global interaction based on the original distribution features when modeling the dependencies between different time steps, providing a stable and undistorted reference benchmark for the calculation of attention weights and avoiding feature shifts or information loss that may be caused by local convolution operations. The temporal convolution branch performs two cascaded one-dimensional temporal convolutions on the same input sequence. The first convolution uses a convolutional layer with a kernel size of 3, performing local weighted aggregation on each time step and its immediate and next-to-immediate neighborhoods to extract short-term motion features; the second convolution also uses a convolutional kernel of size 3, further expanding the receptive field to 5 time steps to capture local dynamic information within a wider time window, while utilizing the smoothing properties of convolution to suppress high-frequency measurement noise. The local motion features output after the two convolutional layers have the same dimension as the linear features.
[0021] In one possible implementation, the linear mapping branch performs a linear projection transformation on the input sequence, mapping it to a high-dimensional network space. Let the input sequence be... The linear mapping branch is traversed by a learnable weight matrix. The linear transformation is expressed as follows: ;in, It is a linear feature. It is a bias vector that provides a learnable offset for each output dimension.
[0022] In one possible implementation, local motion features are processed by two cascaded one-dimensional temporal convolutions. The calculation expression is: ;in, It is the first convolutional layer. It is the second convolutional layer.
[0023] Finally, the fused features The calculation expression is: In the formula It is layer normalization.
[0024] Here, the Transformer encoder comprises multiple stacked coding layers, each consisting of a multi-head self-attention sublayer and a feedforward network sublayer, with residual connections and layer normalization applied after each sublayer; the Transformer encoder is configured to perform the following operations: Obtain a pre-configured position code with the same dimension as the fused feature, add the fused feature to the position code to obtain the encoder input sequence; Perform the following sub-steps at each coding layer: Multi-head self-attention computation is performed on the current input features to obtain the current attention features; the current input features are either the fused features or the output features of the previous encoding layer; the current attention features and the current input features are residually connected and layer normalized to obtain the first intermediate features; the first intermediate features are input into the feedforward network sublayer to obtain the feedforward features; the feedforward features and the first intermediate features are residually connected and layer normalized to obtain the output features of the current encoding layer; the output features of the last encoding layer are obtained as the global modeling features; the global modeling features are decoded to obtain the target state estimate at the current time step.
[0025] Here, positional encoding is a standard component in the Transformer architecture, its role being to provide temporal positional information to the input sequence. Positional encoding can be generated by a fixed function or by learning parameters to obtain an encoded vector with the same dimension as the fused features, used to inject temporal positional information before the input Transformer encoder.
[0026] It should be noted that the target state estimate at the current moment refers to the network's real-time inference of the target's motion state, which includes at least the target's three-dimensional spatial coordinates at the current moment. Depending on the task requirements, motion parameters such as the target's velocity, acceleration, or angular velocity can be further output for subsequent trajectory prediction, target recognition, or situational analysis.
[0027] S4: When the data of the next moment arrives, remove the data of the earliest moment in the sliding window and add the data of the next moment to form an updated input sequence. The updated input sequence is then fed into the enhanced Transformer network for processing to achieve real-time continuous tracking of the target.
[0028] It should be noted that the radar system acquires radar measurement data in real time. The acquired data is temporarily stored in a buffer, and a sliding window obtains a continuous data stream from the buffer. After processing the input sequence at the current time (containing radar measurement data from time t-L+1 to time t), the sliding window moves forward, adding new data. The updated input sequence contains radar measurement data from time t-L+2 to time t+1. This updated input sequence is then fed back into the network for processing, outputting the target state estimate for the next time step. By continuously repeating the above recursive process of "acquisition—sliding window—inference—update," real-time continuous tracking of the target is achieved.
[0029] The aforementioned sliding window mechanism ensures that the input sequence always contains measurement information from the most recent L time moments, providing the network with sufficient historical context for motion feature extraction while also meeting the real-time processing requirements for data timeliness.
[0030] In the model training phase, the enhanced Transformer network uses a training dataset jointly generated by a uniform linear motion model and a cooperative turning motion model, and superimposed with noise that matches the actual radar measurement noise characteristics for iterative training.
[0031] Here, the training dataset used by the enhanced Transformer network during the model training phase is obtained in the following way: 1) Generate a first type of target motion trajectory based on the uniform linear motion model, and generate a second type of target motion trajectory based on the cooperative turning motion model; wherein, the second type of target motion trajectory includes turning motion trajectories that rotate around the x-axis, y-axis and z-axis respectively, and each turning motion trajectory moves at a constant speed and constant angular velocity in a plane perpendicular to the corresponding rotation axis, and the angular velocity is uniformly generated in the clockwise and counterclockwise directions within a preset range.
[0032] In this embodiment, the trajectory of the first type of target is generated based on a uniform linear motion model, which assumes that the target moves in a straight line in three-dimensional space with a constant speed and direction. However, considering the various unpredictable disturbances in the real environment (such as airflow and uneven road surfaces), the target speed will experience slight random fluctuations. Therefore, process noise is introduced into the model to simulate the uncertainty of real motion. The specific generation steps are as follows: The recursive form of the target state in the ideal uniform linear motion model in the discrete-time system is as follows: ; ; ; In the formula, , and Representing time respectively The target at time axis, Position components of the x-axis and z-axis, , and Indicates time The velocity component of the target's corresponding axis at that time. Indicates the sampling interval.
[0033] Here, process noise is added to the original model, and the uniform linear motion model is modeled as follows: ; ; ; In the formula, , and This represents the random variation of the target's velocity along the corresponding axis.
[0034] Therefore, the state equation of the target can be briefly summarized as follows: ; In the formula, express The state transition matrix of the constant-speed linear motion model. This represents the state vector of the uniform linear motion model at the corresponding moment. This represents the process noise at the corresponding time point.
[0035] Specifically, With a mean of zero and a covariance of The Gaussian white noise sequence, at this time The calculation expression is: In the formula, express Function. Furthermore, The calculation expression is further simplified to: In the formula, This represents the process noise distribution matrix at the corresponding time point. This represents the process noise vector.
[0036] Here, when the sampling interval is hour, and The expression is: ; .
[0037] Based on the above formula, repeat the generation process until the desired first-class target motion trajectory containing process noise is obtained.
[0038] Furthermore, in this embodiment, the trajectory of the second type of target is generated based on a cooperative turning motion model. This model assumes that the target is making uniform turning motion in three-dimensional space, that is, the speed (magnitude) remains constant, while rotating around a certain axis of rotation at a constant angular velocity. Considering that there are various unpredictable disturbances in the actual environment (such as airflow, turbulence, etc.), process noise is also introduced into the model to simulate the uncertainty of real motion.
[0039] Similar to the uniform linear motion model, the actual target motion is inevitably affected by environmental disturbances, leading to random fluctuations in velocity. Therefore, process noise needs to be superimposed after the state transition, and the cooperative turning motion model is modeled as follows: ; In the formula, ; ; ; ; in, Represents a 3x3 identity matrix. Represents the velocity rotation matrix. Represents the position integral matrix. Indicates the axis of rotation and , Let represent the cross product matrix of rotation axes; where the coefficients are . , .
[0040] Since a target can perform CT motion around multiple rotation axes in three-dimensional space, let's assume that the rotation axes corresponding to rotation around the x-axis, y-axis, and z-axis are respectively... , and Furthermore, according to the right-hand rule, when viewed from the positive direction of the rotation axis, when the angular velocity... When the target makes a counter-clockwise turn in a plane perpendicular to the axis; when At that time, the goal was to... Make a clockwise turn in a vertical plane. To cover all possible maneuvering modes, set the angular velocity within a preset range. ω max ω max Within the range, samples are randomly selected from a uniform distribution, ensuring a balance between clockwise (negative) and counterclockwise (positive) sample sizes. Based on the actual scenario, the angular velocity is set to a range of [-2π, 2π] radians / second, covering various maneuver intensities from slow turns to rapid turns.
[0041] Based on the above formula, repeat the generation process until the desired second-class target motion trajectory containing process noise is obtained.
[0042] 2) Mix the motion trajectories of the first type of target and the motion trajectories of the second type of target according to a preset ratio to form a basic trajectory dataset.
[0043] For example, the first type of target motion trajectory, the turning motion trajectory rotating around the x-axis, the turning motion trajectory rotating around the y-axis, and the turning motion trajectory rotating around the z-axis are mixed in a ratio of 1:100 / 3:100 / 3:100 / 3.
[0044] 3) Generate measurement noise that matches the actual radar measurement noise characteristics based on the performance parameters of the actual radar system; the performance parameters include radar bandwidth, beamwidth and signal-to-noise ratio.
[0045] 4) The measurement noise is superimposed on the basic trajectory dataset to form a training dataset, which includes multiple radar measurement sample data.
[0046] For example, the trajectory points at each moment in the basic trajectory dataset are converted into polar coordinates, then measurement noise is superimposed, and then the polar coordinate measurement values with superimposed measurement noise are converted back to rectangular coordinates to obtain noisy radar measurement samples.
[0047] Table 1 is a summary example of simulation parameters for the three-dimensional dataset provided in the embodiments of the present invention.
[0048] Table 1
[0049] Here, the enhanced Transformer network is obtained through the following steps: During the k-th training iteration: (1) Obtain the model updated after the (k-1)th training as the current model, and obtain the kth radar measurement sample data from the training dataset, as well as the corresponding kth supervision label representing the true state of the target, to form the kth sample pair; (2) Input the kth radar measurement sample data into the current model for forward propagation to obtain the kth target state estimate; (3) Using the kth target state estimate and the kth supervision label, calculate the error loss between the two to obtain the kth loss value; (4) The AdamW optimizer is adopted, and the learning rate scheduling strategy of linear preheating and cosine annealing restart is combined to update the model parameters in the current network based on the kth loss value, so as to obtain the model after the kth training update. (5) Determine whether the model after the kth training update has converged; (6) If not, continue with the (k+1)th training iteration; (7) If so, terminate the iteration and output the trained enhanced Transformer network.
[0050] Table 2 summarizes the main network training parameters provided in the embodiments of the present invention. These parameters collectively define the network's input and output formats, network structure and size, optimization strategies, and training details.
[0051] Table 2
[0052] Here, to verify the tracking performance of the trained enhanced Transformer network, Monte Carlo repeated simulations are used to evaluate its performance. The average RMSE performance of target tracking position is statistically analyzed through M repeated experiments. The specific steps are as follows: (1) Set the number of Monte Carlo experiments to M (e.g., M=100), and generate test trajectory data independently for each experiment. The test trajectory is generated in the same way as the training dataset, that is, the basic trajectory is generated by combining the uniform linear motion model and the cooperative turning motion model, and measurement noise that matches the actual radar measurement noise characteristics is superimposed to form radar measurement sample data for testing.
[0053] (2) For the first Experiment (1≤ ≤M), the test data is input into the trained enhanced Transformer network for processing to obtain the first... The target state estimate of the Monte Carlo experiment .
[0054] (3) Calculate the true position of the target at each time step, and calculate the root mean square error (RMSE) of the position at the corresponding time step. Wherein, the first... The root mean square error of the position at each sampling time is as follows: ; in, Indicates the first Time, passing Position error loss in the Monte Carlo experiment Indicates the first In the Monte Carlo experiment, the target was in the... The actual location at any given moment It is the first In the Monte Carlo experiment, the first The target state estimate at time t. It refers to the number of Monte Carlo experiments.
[0055] (4) The average root mean square error of position at all times is taken to obtain the overall average root mean square error of position. The calculation formula is as follows: ; in, Indicates the process The overall average root mean square error of position after several Monte Carlo experiments Indicates the first The set of true locations of the target in the Monte Carlo experiment. It is the first The set of target state estimates for the Monte Carlo experiment.
[0056] The Monte Carlo simulation described above can eliminate the influence of randomness in a single experiment and objectively evaluate the average tracking performance of the method of the present invention under different maneuvering scenarios.
[0057] To address the poor robustness of existing radar measurement data processing methods in target tracking under complex maneuvering scenarios, this invention provides an enhanced Transformer 3D target tracking method. Through a dual-branch parallel structure of a temporal convolutional hybrid embedding module, it effectively extracts local short-term motion features and suppresses measurement noise while preserving global information of the input sequence, thus enhancing the stability and discriminativeness of the input representation. A sliding window mechanism enables real-time data stream processing, ensuring continuous recursive target tracking. By jointly generating a training dataset based on a uniform linear motion model and a cooperative turning motion model, and superimposing noise matching the characteristics of actual radar measurement noise, the model can fully learn the motion characteristics of various maneuvering modes in 3D space, effectively improving its generalization ability to unknown maneuvering modes. Simulation results show that this invention achieves superior tracking accuracy compared to traditional methods under various maneuvering scenarios, exhibiting stronger robustness, especially in model mismatch and maneuver switching scenarios. It is suitable for 3D maneuvering target tracking tasks on various air-to-air, air-to-ground, and sea-surface radar platforms.
[0058] The performance of the proposed method is verified by combining simulation software. The tracking performance under different motion trajectories in three-dimensional space is analyzed. Considering that CT motion in three-dimensional space can move around different rotation axes, the experiment sets up 8 motion trajectories to cover 4 typical motion scenarios: (1) The target does not maneuver and the motion mode is consistent with the motion model preset by the Interacting Multiple Model (IMM); (2) The target does not maneuver, the motion mode is inconsistent with the motion model preset by the IMM and is outside the preset range; (3) The target maneuvers and the motion mode is consistent with the motion model preset by the IMM; (4) The target maneuvers, the motion mode is inconsistent with the motion model preset by the IMM and there are two situations: within the preset range and outside the preset range.
[0059] To ensure fairness in comparisons across different scenarios, 13 candidate motion models are pre-defined, including one uniform linear motion model and 12 cooperative turning motion models. The angular velocity of the cooperative turning motion model is taken as... and Rotation axis , and All the above simulations use a fixed radar signal bandwidth. Half beamwidth is The sampling interval was 0.1 s, and 100 Monte Carlo experiments were used for statistical evaluation. The specific settings for the key parameters of each motion trajectory are summarized in Table 3. Among them, numbers 1 and 2 represent motion scenario 1, numbers 3 and 4 represent motion scenario 2, numbers 5 and 6 represent motion scenario 3, and numbers 7 and 8 represent motion scenario 4.
[0060] Table 3
[0061] As shown in Table 1, target number 1 undergoes CV motion within 1-60s. Target number 2 undergoes CT motion within 1-30s, with an angular velocity of... The axis of rotation is Target number 3 undergoes CT motion within 1-30 seconds, with an angular velocity of... The axis of rotation is Target number 4 undergoes CT motion within 1-30 seconds, with an angular velocity of... The axis of rotation is Target number 5 undergoes CT motion within 1-20 seconds, with an angular velocity of... The axis of rotation is Perform CV motion within 21-40 seconds; perform CT motion within 41-70 seconds, with an angular velocity of... The axis of rotation is Target number 6 undergoes CT motion within 1-30 seconds, with an angular velocity of... The axis of rotation is Perform CV motion within 31-40 seconds; perform CT motion within 41-60 seconds, with an angular velocity of... The axis of rotation is Target number 7 undergoes CT motion within 1-30 seconds, with an angular velocity of... The axis of rotation is Perform CV motion within 31-40 seconds; perform CT motion within 41-60 seconds, with an angular velocity of... The axis of rotation is Target number 8 undergoes CT motion within 1-26 seconds, with an angular velocity of... The axis of rotation is Perform CV motion within 27-45 seconds; perform CT motion within 46-70 seconds, with an angular velocity of... The axis of rotation is .
[0062] The trajectories in Table 4 were processed using IMM-UKF, the standard Transformer, and the enhanced Transformer (HTEC-Transformer) network used in this invention, respectively. Table 2 lists the average root mean square error of position obtained from each model processing. Figure 3-10These are the motion trajectories corresponding to each serial number and the average root mean square error curves of each model after processing, provided in the embodiments of the present invention. Among them, (a) in each figure is the motion trajectory corresponding to each serial number, and (b) is the average root mean square error curve of each model after processing.
[0063] Table 4
[0064] Based on Table 4 and Figure 3-10 It can be seen that in motion scenario 1, without any maneuvering, the enhanced Transformer, standard Transformer, and IMM-UKF can all stably estimate CV and CT moving targets. Among them, the Transformer-based method has a smaller overall estimation error and better performance. In motion scenario 2, the IMM performance degrades significantly because the angular velocity and rotation axis of the target motion model exceed the range of the IMM preset model. The Transformer-based method can still maintain high estimation accuracy and shows stronger robustness. In motion scenario 3, when a maneuver occurs, the IMM needs to dynamically adjust the probabilities of each model. Affected by the lag in model probability updates, it is difficult to match the target motion state in time. In contrast, the Transformer-based network can learn the target motion state features through training, thereby obtaining higher estimation accuracy. In motion scenario 4, when the real motion model is within the candidate model set, the IMM can still effectively estimate the motion of non-preset models. However, when the angular velocity and rotation axis of the target motion model exceed the range of the IMM preset model, similar to the situation in scenario (2), the performance degrades significantly. The Transformer-based method can maintain good estimation accuracy. The proposed HTCE-Transformer has a lower estimation error than the standard Transformer and IMM-UKF, thus verifying the effectiveness and advantages of the proposed method.
[0065] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. An enhanced Transformer 3D target tracking method, characterized in that, include: Based on radar systems deployed in real physical scenarios, radar measurement data containing three-dimensional moving target information is collected in real time to form a continuous data stream; By using a sliding window, a data segment of length L, ending at the current time, is extracted from the data stream to form an input sequence; The input sequence is input into the trained enhanced Transformer network. After local-global feature extraction and feature fusion are performed by the temporal convolutional hybrid embedding module inside the model, the internal Transformer encoder is used to perform global dependency modeling and target state estimation on the fused features, and outputs the target state estimate value at the current time. When the data for the next moment arrives, the data from the earliest moment in the sliding window is removed, and the data for the next moment is added to form an updated input sequence. The updated input sequence is then fed into the enhanced Transformer network for processing to achieve real-time continuous tracking of the target. In the model training phase, the enhanced Transformer network uses a training dataset jointly generated by a uniform linear motion model and a cooperative turning motion model, and superimposed with noise that matches the actual radar measurement noise characteristics, for iterative training.
2. The enhanced Transformer 3D target tracking method according to claim 1, characterized in that, The temporal convolutional hybrid embedding module includes parallel linear mapping branches and temporal convolution branches, as well as a feature fusion submodule; The temporal convolutional hybrid embedding module is configured to perform the following operations: In the linear mapping branch, the input sequence is linearly transformed, and the output is a linear feature that retains global information; In the temporal convolution branch, the input sequence is subjected to two cascaded one-dimensional temporal convolutions to aggregate short-term motion information of the target in the local neighborhood and suppress measurement noise, and output local motion features. In the feature fusion submodule, the linear features and the local motion features are subjected to element-wise addition and layer normalization, and the fused features are output.
3. The enhanced Transformer 3D target tracking method according to claim 1, characterized in that, The Transformer encoder comprises multiple stacked coding layers, each including a multi-head self-attention sublayer and a feedforward network sublayer, with residual connections and layer normalization applied after each sublayer; the Transformer encoder is configured to perform the following operations: Obtain a pre-configured position code with the same dimension as the fused feature, and add the fused feature to the position code to obtain the encoder input sequence; The following sub-steps are performed at each coding layer: Multi-head self-attention computation is performed on the current input features to obtain the current attention features; the current input features are either the fused features or the output features of the previous coding layer. The current attention feature and the current input feature are subjected to residual connection and layer normalization to obtain the first intermediate feature; The first intermediate feature is input into the feedforward network sub-layer to obtain the feedforward feature; The feedforward features and the first intermediate features are residually connected and layer normalized to obtain the output features of the current coding layer. The output features of the last encoding layer are used as global modeling features; The global modeling features are decoded to obtain the target state estimate at the current moment.
4. The enhanced Transformer 3D target tracking method according to claim 1, characterized in that, The training dataset used by the enhanced Transformer network during the model training phase was obtained in the following way: The first type of target motion trajectory is generated according to the uniform linear motion model, and the second type of target motion trajectory is generated according to the cooperative turning motion model; wherein, the second type of target motion trajectory includes turning motion trajectories that rotate around the x-axis, y-axis and z-axis respectively, and each turning motion trajectory moves at a constant speed and constant angular velocity in a plane perpendicular to the corresponding rotation axis, and the angular velocity is uniformly generated in the clockwise and counterclockwise directions within a preset range; The motion trajectories of the first type of target and the motion trajectories of the second type of target are mixed according to a preset ratio to form a basic trajectory dataset; Based on the performance parameters of the actual radar system, measurement noise is generated that matches the measurement noise characteristics of the actual radar; the performance parameters include radar bandwidth, beamwidth, and signal-to-noise ratio. The measurement noise is superimposed on the basic trajectory dataset to form the training dataset, which includes multiple radar measurement sample data.
5. The enhanced Transformer 3D target tracking method according to claim 4, characterized in that, The first type of target motion trajectory, the turning motion trajectory rotating around the x-axis, the turning motion trajectory rotating around the y-axis, and the turning motion trajectory rotating around the z-axis are mixed in a ratio of 1:100 / 3:100 / 3:100 / 3.
6. The enhanced Transformer 3D target tracking method according to claim 4, characterized in that, The range of the angular velocity is [-2π, 2π] radians / second.
7. The enhanced Transformer 3D target tracking method according to claim 1, characterized in that, The enhanced Transformer network is obtained through the following steps: During the k-th training iteration: The model updated after the (k-1)th training is obtained as the current model, and the kth radar measurement sample data and the corresponding kth supervision label representing the true state of the target are obtained from the training dataset to form the kth sample pair. The kth radar measurement sample data is input into the current model for forward propagation to obtain the kth target state estimate; Using the kth target state estimate and the kth supervision label, the error loss between the two is calculated to obtain the kth loss value; The AdamW optimizer is used, and a learning rate scheduling strategy of linear preheating and cosine annealing restart is combined to update the model parameters in the current network based on the k-th loss value, so as to obtain the model after the k-th training update. Determine whether the model after the k-th training update has converged; If not, continue with the (k+1)th training iteration; If so, terminate the iteration and output the trained enhanced Transformer network.
8. The enhanced Transformer 3D target tracking method according to claim 7, characterized in that, The trained enhanced Transformer network was subjected to M repeated experiments using Monte Carlo simulation. The overall average root mean square error (RMSE) was calculated for performance evaluation. The formula for calculating the overall average RMSE is as follows: ； in, Indicates the process The overall average root mean square error of position after several Monte Carlo experiments Indicates the first The set of true locations of the target in the Monte Carlo experiment. The first one is the one mentioned The set of target state estimates for the Monte Carlo experiment.
9. The enhanced Transformer 3D target tracking method according to claim 1, characterized in that, Prior to using sliding window extraction, the method further includes: Radar measurement data in polar coordinates is converted into measurement data in Cartesian coordinates through unbiased transformation.