Method and device for tracking maneuvering target based on unet framework and multi-head attention mechanism
By employing the Unet framework and multi-head attention mechanism, the problem of large state estimation errors in maneuvering target tracking is addressed. Feature processing through encoder and decoder branches improves the performance and accuracy of tracking highly maneuvering targets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2023-07-12
- Publication Date
- 2026-07-14
Smart Images
Figure CN116879879B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of radar remote sensing application technology, and in particular to a method and apparatus for tracking maneuvering targets based on the Unet framework and multi-head attention mechanism. Background Technology
[0002] Maneuvering target tracking is an important task in fields such as airspace surveillance, traffic control, and automatic navigation. However, the uncertainty and variability of maneuvering target motion place higher performance requirements on maneuvering target tracking methods.
[0003] In related technologies, to improve the tracking performance of maneuvering targets, multiple motion models and different filters are used in an integrated framework to estimate the target state (e.g., IMM). IMM requires pre-defining the motion model before estimating the target state, and for highly maneuvering targets, it requires a long time to accumulate a relatively accurate motion model, thus causing significant state estimation delays and affecting tracking accuracy. While deep learning-based methods (e.g., DeepMTT) do not suffer from motion model estimation delays, they use traditional methods to process measurement data to obtain the model's input trajectory, which cannot provide a reliable input trajectory, resulting in large errors with the true trajectory and ultimately inaccurate model trajectory estimation. A maneuvering target tracking model (TrTNet) based on a complete transformer structure and state features effectively utilizes the correlation information between states at different times, but because the model input is a noisy measurement trajectory, the noise contained in the state features interferes with the extraction of the similarity matrix between states, thus affecting tracking performance.
[0004] Therefore, how to solve the problem of large target state estimation error in highly maneuverable target tracking missions and improve the performance of highly maneuverable target tracking missions is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of the above problems, embodiments of this application provide a method and apparatus for tracking maneuvering targets based on the Unet framework and multi-head attention mechanism, so as to overcome the above problems or at least partially solve the above problems.
[0006] A first aspect of this application discloses a maneuvering target tracking method based on the Unet framework and a multi-head attention mechanism, the method comprising:
[0007] The measurement trajectory is encoded by input and position to obtain encoded state features. The measurement trajectory is a sequence of multiple measurement data within a preset time period.
[0008] The encoded state features are input into the downsampling encoder branch for state feature encoding and similarity information extraction, resulting in multi-level encoder features and similarity matrices.
[0009] The multi-level encoder features and similarity matrix are input into the upsampling decoder branch to perform multi-level feature fusion and attention information fusion, resulting in decoder features that fuse different levels of features and attention information.
[0010] The residual trajectory is predicted based on the features of the decoder, and the residual trajectory is added to the measured trajectory to obtain the target trajectory.
[0011] Optionally, the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism is implemented through a pre-trained target tracking model. The training data of the target tracking model is: strong maneuvering target tracking data pairs constructed based on a constant turning uniform acceleration model. Each pair of strong maneuvering target tracking data pairs includes the actual trajectory and measured trajectory of the maneuvering target in two-dimensional space.
[0012] Optionally, the highly maneuverable target tracking data pair is constructed according to the following steps:
[0013] The constant turning uniform acceleration model can be transformed into different motion models by setting different control parameters, including: turning rate and acceleration;
[0014] The true trajectory is obtained based on the state equation and the state transition equation;
[0015] Based on the actual trajectory and observation equation, noisy measurement data is obtained;
[0016] The noisy measurement data is processed into the same form as the actual trajectory to obtain the measurement trajectory.
[0017] Optionally, the measurement trajectory is input-encoded and position-encoded to obtain the encoded state features, including:
[0018] The measurement trajectory is processed sequentially using a linear transformation layer and a global normalization layer to obtain initial encoded features. The global normalization layer uses the same pair of normalization parameters in both the time dimension and the feature dimension.
[0019] The absolute position encoding feature is added to the initial encoding feature to obtain the encoded state feature, which retains the difference information of the measurement trajectory in the time dimension and the feature dimension.
[0020] Optionally, the downsampling encoder branch includes N encoders, which are connected by downsampling; the encoded state features are input into the downsampling encoder branch for state feature encoding and similarity information extraction, resulting in a multi-level encoder feature and similarity matrix, including:
[0021] The first to Nth encoders are based on a multi-head attention mechanism, which sequentially extracts similarity information from the input features to obtain encoder features and similarity matrices from shallow to deep levels. The input features of the first encoder are the encoded state features, and the input features of the second to Nth encoders are the features obtained by downsampling the encoder features output by the previous encoder.
[0022] Optionally, the upsampling decoder branch includes N decoders, which are connected via upsampling. The first decoder is connected to the Nth encoder in the downsampling encoder branch through an intermediate module. The encoder features output by the first to Nth encoders and the similarity matrix correspond to the Nth to first decoders. The multi-level encoder features and similarity matrix are input into the upsampling decoder branch for multi-level feature fusion and attention information fusion to obtain decoder features that fuse features and attention information from different levels, including:
[0023] The first to Nth decoders are based on a multi-head attention mechanism, sequentially performing feature fusion and attention information fusion on the input features and similarity matrix to obtain the decoder features output by each decoder. The input features of the first decoder are the features obtained by upsampling the output features of the intermediate module and the features of the Nth level encoder. The similarity matrix input of the first decoder is the similarity matrix output by the Nth encoder. The input features of the second to Nth decoders are the features obtained by upsampling the decoder features output by the previous level decoder and the encoder features of the corresponding level. The similarity matrix input of the second to Nth decoders is the similarity matrix output by the encoder of the corresponding level.
[0024] The decoder features output by the Nth decoder are used as decoder features that fuse features from different levels and attention information.
[0025] Optionally, the similarity matrix of the multi-head attention layer in the decoder is obtained by fusing the similarity matrices generated by the multi-head attention layers in the encoder branch in an additive manner.
[0026] Optionally, both the encoder and the decoder use a noisy activation function and a global normalization layer to process the input features.
[0027] A second aspect of this application discloses a maneuvering target tracking device based on the Unet framework and a multi-head attention mechanism, the device comprising:
[0028] The trajectory encoding module is used to encode the input and position of the measurement trajectory to obtain the encoded state features. The measurement trajectory is a sequence of multiple measurement data within a preset time period.
[0029] The encoder module is used to input the encoded state features into the downsampling encoder branch for state feature encoding and similarity information extraction, thereby obtaining multi-level encoder features and similarity matrices.
[0030] The decoder module is used to input the multi-level encoder features and similarity matrix into the upsampling decoder branch to perform multi-level feature fusion and attention information fusion, so as to obtain decoder features that fuse different levels of features and attention information.
[0031] The trajectory prediction module is used to predict the residual trajectory based on the features of the decoder, and add the residual trajectory and the measurement trajectory to obtain the target trajectory.
[0032] A third aspect of this application discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps of the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism described in the first aspect of this application.
[0033] The embodiments of this application have the following advantages:
[0034] This application proposes a maneuvering target tracking method based on the Unet framework and multi-head attention mechanism. This method better captures the drastic state changes of highly maneuvering targets, addressing the problem of large target state estimation errors in highly maneuvering target tracking tasks and improving the performance of such tasks. The method involves input encoding and position encoding of the measured trajectory to obtain encoded state features. These encoded state features are then input to a downsampling encoder branch for state feature encoding and similarity information extraction, resulting in multi-level encoder features and a similarity matrix. These multi-level encoder features and similarity matrix are then input to an upsampling decoder branch for multi-level feature fusion and attention information fusion, yielding decoder features that fuse different levels of features and attention information. Finally, a residual trajectory is predicted based on the decoder features, and the residual trajectory is added to the measured trajectory to obtain the target trajectory.
[0035] By employing downsampling in the encoder branch and upsampling in the decoder branch, feature discarding and feature fitting in the temporal dimension of state features can be performed. This reduces the interference of noise from the measured trajectory on the extraction of true correlation information between states, which is beneficial for learning transition patterns. Simultaneously, by leveraging the U-structure of the Unet framework, the multi-level encoder features and similarity matrices output from the encoder branch are connected to the decoder branch. This achieves the fusion of shallow and deep features in the feature dimension and attention-enhanced fusion in the temporal dimension, increasing the feature dimension of the state features. Therefore, this method can better capture the drastic state change patterns of highly maneuverable targets, thereby improving the performance of highly maneuverable target tracking tasks. Attached Figure Description
[0036] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0037] Figure 1 This is a flowchart illustrating the steps of a maneuvering target tracking method based on the Unet framework and multi-head attention mechanism provided in an embodiment of this application.
[0038] Figure 2 This is a schematic diagram illustrating the process of obtaining normalization parameters for existing normalization layers and the global normalization layer provided in this application embodiment;
[0039] Figure 3 This is a framework diagram of a multi-head attention layer for encoder branches and a multi-head attention layer for decoder branches provided in an embodiment of this application;
[0040] Figure 4 This is a schematic diagram of the structure of a target tracking model provided in an embodiment of this application;
[0041] Figure 5 The target tracking results are compared between the traditional model and the maneuvering target tracking model provided in the embodiments of this application;
[0042] Figure 6 The target tracking results are shown in Figure 2, which compares the traditional model with the maneuvering target tracking model provided in the embodiments of this application.
[0043] Figure 7 The target tracking results are shown in Figure 3, which compares the traditional model with the maneuvering target tracking model provided in the embodiments of this application.
[0044] Figure 8 The target tracking results are shown in Figure 4, which compares the traditional model with the maneuvering target tracking model provided in the embodiments of this application.
[0045] Figure 9 This is a schematic diagram of the structure of a maneuvering target tracking device based on the Unet framework and multi-head attention mechanism provided in an embodiment of this application. Detailed Implementation
[0046] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. 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.
[0047] Reference Figure 1 As shown, Figure 1 This is a flowchart illustrating the steps of a maneuvering target tracking method based on the Unet framework and multi-head attention mechanism provided in an embodiment of this application. Figure 1 As shown, the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism provided in this application embodiment may specifically include steps S110 and S140:
[0048] Step S110: Input encoding and position encoding are performed on the measurement trajectory to obtain the encoded state features. The measurement trajectory is a sequence of multiple measurement data within a preset time period.
[0049] In this embodiment, the measured trajectory is obtained by converting observation data from a sensor (e.g., radar), which includes the observation angle and observation distance. The true trajectory of a target object in two-dimensional space is characterized by the target's coordinates in two-dimensional space and the velocity at the coordinate positions. Therefore, in order to more effectively extract the state features of the target trajectory and achieve accurate estimation of the target trajectory, the sensor's observation data is converted into a measured trajectory with the same form as the true target trajectory.
[0050] Furthermore, the measurement trajectory is a sequence of multiple measurement data points within a preset time period. To avoid tracking failure due to the accumulation of trajectory prediction errors over a long period, the preset time is set to the same duration as the trajectory used in the training process. Specifically, the measurement trajectory is obtained by acquiring long-series measurement data according to the target time interval, and then extracting a preset time length of measurement data from this long-series measurement data as the measurement trajectory. For example, if the target time interval is 0.1 seconds and the preset time is 10 seconds, the measurement trajectory is a sequence of measurement states at 100 moments. For example, the measurement trajectory can be represented as Z = [Z1, Z2, ..., Z...]. k ] T Z1, Z2, Zk These represent the first, second, and kth measurement states in the measurement trajectory, respectively.
[0051] In this embodiment of the application, in order to obtain the change pattern of the state in the measurement trajectory and improve the accuracy of the state estimation of the highly maneuvering target, the measurement trajectory is input encoded and position encoded to obtain the encoded state features. The encoded state features retain the differences of the trajectory features in the time dimension and feature dimension, which helps the downsampling encoder branch and upsampling decoder branch to extract more state feature information in subsequent steps, thereby improving the tracking performance.
[0052] In one optional implementation, the measurement trajectory is input-encoded and position-encoded to obtain encoded state features, including:
[0053] The measurement trajectory is processed sequentially using a linear transformation layer and a global normalization layer to obtain initial encoded features. The global normalization layer uses the same pair of normalization parameters in both the time and feature dimensions. The absolute position encoded features are added to the initial encoded features to obtain the encoded state features. The encoded state features retain the differences in the measurement trajectory in both the time and feature dimensions.
[0054] Specifically, the measurement trajectory can be represented by input encoding and position encoding as follows:
[0055] Z e =GNorm(ZW+b)+P
[0056] in, Represents the encoded state characteristics. and Let represent the weight matrix and bias of the linear transformation layer, respectively; P represents the absolute position encoding; and GNorm represents a global normalization layer provided in this embodiment. The normalization parameters of the global normalization layer GNorm are obtained as follows:
[0057]
[0058] Where x represents a tensor with L samples and a size of N×k×d, n represents the guide for a trajectory sample in a batch, and μ n and Let represent the mean and variance of the input data, respectively. The mean and variance are calculated based on all feature information of a single training sample. Specifically, the mean and variance are expressed as:
[0059]
[0060] Where t represents the t-th time step and l represents the l-th feature channel.
[0061] Figure 2 This diagram illustrates the process of obtaining normalization parameters for existing normalization layers and the global normalization layer provided in this embodiment. It can be seen that the global normalization layer obtains the mean and standard deviation parameters used for normalization from the features. The global normalization layer performs a global normalization operation on the state features. Because the addition of a normalization layer is essential for model convergence, especially for target tracking tasks where the range of positional variations in different sample measurement trajectories is large, the normalization operation can standardize the data to a range suitable for model learning.
[0062] Batch normalization calculates normalization parameters based on all features in the entire batch, which is unsuitable for tasks where the numerical values of different samples vary significantly. Layer normalization and instance normalization layers calculate normalization parameters for individual samples in the training batch. Layer normalization calculates normalization parameters in the time dimension, preserving the differences between feature subspaces, while instance normalization calculates normalization parameters in the feature dimension, preserving the differences between features in the time dimension. For the target tracking task in this embodiment, the differences in state features in the time dimension help extract the differences between states and facilitate the learning of state transition patterns. Therefore, the application of instance normalization layers is more suitable. However, it should be noted that the normalization operation of instance normalization layers first erases the differences between features in different feature subspaces, and then learns the characteristics of the original data itself by learning the inverse normalization parameters. Although learning the inverse normalization parameters can recover some of the characteristics of the original data, the normalization operation still erases some feature subspace differences that cannot be learned back.
[0063] In this embodiment, to obtain normalized features suitable for model learning while preserving the differences in features across the time and feature dimensions, a global normalization layer is proposed. This global normalization layer uses the same pair of normalization parameters in both the time and feature dimensions. For a single target trajectory sample, using the same pair of normalization parameters can normalize the data to a suitable range while preserving the differences in trajectory samples across the time and feature dimensions, thus helping the model learn more details.
[0064] Step S120: Input the encoded state features into the downsampling encoder branch to encode the state features and extract similarity information to obtain multi-level encoder features and similarity matrix.
[0065] In this embodiment, the encoder included in the downsampling encoder branch is an improvement on the traditional transformer structure encoder. The encoder primarily uses a multi-head attention mechanism to encode state features and extract similarity information from the input features. Specifically, the encoder encodes input features S∈R... k×4 (The input feature is a sequence of data) It is mapped to multiple key matrices through multiple linear transformation feature extraction layers. Query matrix (query) Value matrix In the process, similarity matrices are computed in parallel across different feature subspaces to further obtain enhanced attention information. Finally, a linear transformation layer is used to fuse the attention information from all feature subspaces. The extraction process of multi-head attention information can be represented as follows:
[0066]
[0067] Where MuAtten(·) represents the multi-head attention mechanism, S represents the input feature, Con represents concatenating features along the feature dimension, and F i (i = 1, ..., h) represents single-head self-attention information, and its extraction process can be represented as follows:
[0068]
[0069] Among them, A e The similarity matrix is obtained by the dot product of the query matrix and the key matrix, used to describe the similarity between state features at different time points; softmax represents the normalized exponential function; d h represents the feature dimension of the single-head attention layer, used to adjust the scale of the state feature data to avoid the softmax value from getting stuck in the region of minimal gradient; mask represents the mask superimposed on the similarity matrix, so that the prediction at the current time step can only see the input measurement data at the current time step and those before it.
[0070] In one optional implementation, the downsampling encoder branch includes N encoders connected by downsampling. The encoded state features are input to the downsampling encoder branch for state feature encoding and similarity information extraction to obtain multi-level encoder features and similarity matrices. This includes: the first to Nth encoders, based on a multi-head attention mechanism, sequentially extract similarity information from the input features to obtain encoder features and similarity matrices from shallow to deep levels. The input features of the first encoder are the encoded state features, and the input features of the second to Nth encoders are the encoder features output by the previous level encoder after downsampling.
[0071] The value of N is determined according to the actual situation. Specifically, the encoded state features in step S110 are input into the first encoder for state feature encoding and similarity information extraction to obtain the encoder features and similarity matrix of the first level. Then, the encoder features of the first level are downsampled and input into the second encoder for state feature encoding and similarity information extraction to obtain the encoder features and similarity matrix of the second level. Similarly, each subsequent encoder performs state feature encoding and similarity information extraction in sequence according to the above logic, and finally obtains the encoder features and similarity matrix from shallow level to deep level, that is, the encoder features and similarity matrix from the first level to the Nth level.
[0072] For example, the state feature encoding and similarity information extraction of the j-th encoder can be represented as follows:
[0073]
[0074]
[0075]
[0076] in, and These are the fixed parameters used in training, E j E is the output feature of the j-th encoder. j-1 This is the input of the j-th encoder, where j = 1, ..., N. When j = 1, the initial input of the encoder is... (i.e., the encoded state features), Down indicates an average downsampling operation in the time dimension, f ρ (·) is a noisy tanh activation function, which is obtained by adding some random noise to the original tanh nonlinear activation function. The noisy tanh activation function is expressed as:
[0077] f ρ (x)=f(x)+ρ·σ(x)=f(x)+ρ·(sigmoid(f(x)-x)-1) 2
[0078] Where x represents the input feature, f(·) represents the tanh nonlinear activation function, σ(·) represents the noise scaling function controlled by the input, ρ represents the Gaussian noise superimposed on the activation function, and sigmoid(·) represents the activation function. Compared to the classic ReLU activation function, the noisy tanh activation function has mean centering and negative values, making it more suitable for target tracking tasks. Furthermore, the target tracking model implemented in this application is based on fully connected layers and has a large number of parameters. To avoid the high-mobility target tracking model getting stuck in the saturation region of the activation function during training, the noisy tanh activation function is chosen to provide more randomness to the model training and accelerate the training process.
[0079] Step S130: Input the multi-level encoder features and similarity matrix into the upsampling decoder branch to perform multi-level feature fusion and attention information fusion to obtain decoder features that fuse different levels of features and attention information.
[0080] In this embodiment, the decoder included in the upsampling decoder branch is also a structure built from a multi-head attention layer and a feedforward network. Upsampling fitting is performed in the time dimension of the decoder branch, and multi-level feature fusion and attention information fusion are performed based on the multi-head attention mechanism.
[0081] In one optional implementation, the upsampling decoder branch includes N decoders connected via upsampling. The first decoder is connected to the Nth encoder in the downsampling encoder branch via an intermediate module. The encoder features output by the first to Nth encoders and the corresponding similarity matrix jump to the Nth to first decoders. The multi-level encoder features and similarity matrix are input into the upsampling decoder branch for multi-level feature fusion and attention information fusion to obtain decoder features that fuse different levels of features and attention information, including:
[0082] The first to Nth decoders are based on a multi-head attention mechanism, sequentially fusing the input features and the similarity matrix with feature fusion and attention information fusion to obtain the decoder features output by each decoder. The input features of the first decoder are the features obtained by upsampling the output features of the intermediate module and the features of the Nth level encoder. The similarity matrix input of the first decoder is the similarity matrix output by the Nth encoder. The input features of the second to Nth decoders are the features obtained by upsampling the decoder features output by the previous level decoder and the encoder features of the corresponding level. The similarity matrix input of the second to Nth decoders is the similarity matrix output by the encoder of the corresponding level. The decoder features output by the Nth decoder are used as the decoder features that fuse features from different levels and attention information.
[0083] Specifically, the encoder features of the Nth level are upsampled after passing through the intermediate module and concatenated with the encoder features of the Nth level to serve as the input features of the first decoder. The similarity matrix output by the Nth level encoder is also input into the first decoder. The input features of the first decoder are upsampled and concatenated with the encoder features of the (N-1)th level to serve as the input features of the second decoder. The similarity matrix output by the (N-1)th level encoder is also input into the second decoder. Similarly, each subsequent decoder performs feature fusion and attention information fusion sequentially according to the above logic, ultimately obtaining decoder features that fuse features from different levels and attention information.
[0084] For example, the i-th decoder can be represented as:
[0085] P i =Con(up(D) i-1 ),E N-i )
[0086]
[0087]
[0088] in, These are the parameters determined during training. D i-1 is the input of the i-th decoder, i = 1, ..., N. up is the time-dimension upsampling fit of the state features. It is the feature obtained by concatenating the upsampled feature in the i-th decoder with the Ni-th encoder feature in the feature dimension, thus increasing the feature dimension.
[0089] It is important to emphasize that both the encoder and the decoder use a noisy activation function and a global normalization layer to process the input features. Because the noisy activation function provides more randomness to the model training, and the global normalization layer can simultaneously preserve the differences in features across both the temporal and feature dimensions, the downsampling encoder branch and the upsampling decoder branch can learn more state feature information, further improving the tracking performance of moving targets.
[0090] Furthermore, the intermediate module also uses a noisy activation function and a global normalization layer. By processing the encoder features of the Nth level through the intermediate module, the dimensionality of the features can be better expanded. Specifically, the structure of the intermediate module can be represented as follows:
[0091]
[0092]
[0093] in, All are parameters determined during training, E N This represents the encoder feature output by the Nth encoder.
[0094] In one optional implementation, the similarity matrix of the multi-head attention layer in the decoder is obtained by fusing the similarity matrices generated by the multi-head attention layers in the encoder branch in an additive manner. By adding the skip-joining fusion of the similarity matrices, the state differences between the current state features and the state features at different times are fully learned. Specifically, in the process of obtaining the similarity matrix of each decoder, the similarity matrix generated by the corresponding multi-head attention layer in the encoder is fused with it in an additive manner to obtain the final decoder similarity matrix.
[0095] For example, Figure 3 This is a framework diagram of a multi-head attention layer for encoder branches and a multi-head attention layer for decoder branches provided in an embodiment of this application. Figure 3 The diagram illustrates the calculation process of the similarity matrix in the multi-head attention layer of the decoder. Specifically, the similarity matrix in the attention layer of the i-th decoder... It can be represented as follows:
[0096]
[0097] in, α is the similarity matrix of the self-attention layer in the encoder before softmax normalization, and α is a parameter that adjusts the superposition strength of the similarity matrix. It is set to a fixed value and the range is set to [0, 1]. In this embodiment, α can be set to 0.5.
[0098] Through steps S120 and S130, the downsampling process of the encoder branch and the upsampling process of the decoder branch, based on the Unet framework and multi-head attention mechanism, perform feature discarding and feature fitting in the temporal dimension of the state features. This increases the smoothing operation on the state features, providing a cleaner environment for the multi-head attention layers in the encoder and decoder, reducing the interference of noise on the extraction of true correlation information between states, and facilitating the learning of transition patterns. Furthermore, based on the characteristics of the Unet structure, the feature output and similarity matrix of the encoder branch are connected to the decoder branch. Features are concatenated in the feature dimension, while the similarity matrix is connected by addition, to achieve the fusion of shallow and deep features in the feature dimension and the attention-enhanced fusion in the temporal dimension.
[0099] Step S140: Based on the features predicted by the decoder, the residual trajectory is obtained, and the residual trajectory and the measurement trajectory are added together to obtain the target trajectory.
[0100] In this embodiment, considering the wide range of the target's true trajectory and the large numerical differences between different trajectories, the target tracking model is difficult to converge. However, learning the residual trajectory can significantly reduce the numerical differences between different trajectories. Therefore, the normalized mapping relationship between the measured trajectory and the residual trajectory is more conducive to model convergence and learning. Furthermore, instead of directly predicting the target's true trajectory, the residual trajectory between the true trajectory and the measured trajectory is predicted. After obtaining the residual trajectory, the target trajectory is calculated based on it. Specifically, a linear mapping layer is used to predict the residual trajectory from the decoder features output in step S130, and the predicted residual trajectory is added to the measured trajectory to obtain the target trajectory used for target tracking.
[0101] In actual target tracking, a sliding window with a fixed length of M sampling points (e.g., M=100) is used to track a long-term trajectory at a fixed preset sampling point K. p The overlapping sliding is performed, and the overlapping areas of the previous and subsequent trajectories are summed and averaged to obtain the final prediction result, where a fixed preset sampling point K is used. p The value can be flexibly selected based on the length of the long-term trajectory, such as setting K. p =50.
[0102] For example, for a long-term trajectory sequence of 100 seconds (sampling interval of 0.1 seconds, totaling 1000 sampling points), a fixed preset sampling point is set to 50 (5 seconds), and sequence data within a 10-second time period is used as the measurement trajectory. During high-maneuverability target tracking, in the first sliding window, the estimated target trajectory for 0 to 10 seconds is calculated using the methods described in steps S110 to S150 above. In the second sliding window, the estimated target trajectory for 5 to 15 seconds is calculated using the same method. Then, by summing and averaging, the estimated target trajectory for the overlapping region for 5 to 10 seconds is calculated. This process is repeated to obtain the estimated target trajectory for target tracking. Similarly, the target trajectory for each time step is calculated sequentially according to the above logic to achieve target tracking.
[0103] In one alternative embodiment, the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism is implemented by a pre-trained target tracking model. The training data of the target tracking model is: strong maneuvering target tracking data pairs constructed based on a constant turning uniform acceleration model. Each pair of strong maneuvering target tracking data pairs includes the actual trajectory and measured trajectory of the maneuvering target in two-dimensional space.
[0104] Figure 4This is a schematic diagram of the structure of a target tracking model provided in an embodiment of this application. Specifically, the target tracking model includes a downsampling encoder branch, an upsampling decoder branch, an input module, and an intermediate module. Within the Unet framework, downsampling and upsampling processes reduce the interference of input noise on the similarity information between state features extracted by the multi-head attention layer. Simultaneously, the feature output and similarity matrix of the encoder branch are connected to the decoder branch. Features are concatenated along the feature dimension, while the similarity matrix is connected by addition, achieving fusion of shallow and deep features along the feature dimension and attention-enhanced fusion along the temporal dimension. Furthermore, to adapt to mobile target tracking tasks, the target tracking model uses a global normalization layer to extract more state feature information.
[0105] In this embodiment of the application, in order to adapt to the task of tracking highly maneuvering targets, a constant turn rate and acceleration (CTRA) model is used to build a training dataset for maneuvering target trajectories. The constant turn rate and acceleration model expands the complexity of maneuvering target trajectories, and a single short-time trajectory segment contains two different transition parameters, so that the trajectory segment contains transition points with different transition parameters, increasing the content that the target tracking model can learn.
[0106] In one alternative implementation, the highly maneuverable target tracking trajectory is constructed according to steps A1 to A4:
[0107] Step A1: By setting different control parameters, the constant turning uniform acceleration model is transformed into different motion models. The control parameters include: turning rate and acceleration.
[0108] Specifically, when the rotational speed is 0, the CTRA motion model degenerates into the CA uniform acceleration model; when the acceleration is 0, the CTRA motion model degenerates into the CT constant turning model; and when both the rotational speed and acceleration are 0, the CTRA motion model degenerates into the CV uniform motion model. Furthermore, a more maneuverable trajectory can be constructed using the constant turning uniform acceleration model.
[0109] Step A2: Obtain the true trajectory based on the state equation and the state transition equation.
[0110] In this embodiment, the state equation represents the state of the target at the current moment, while the state transition equation represents the state change of the target from the current moment to the next moment. Therefore, the true trajectory of the target can be obtained based on the state equation and the state transition equation.
[0111] For example, the state equations are expressed as follows:
[0112] X t-1 =[x,y,θ,v,ω,a] t-1
[0113] Where [x,y] represents the x-direction and y-direction position in two-dimensional space, θ is the yaw angle, which is the angle between the target and the x-axis in the current coordinate system. v is the radial velocity of the target, ω is the angular velocity of the target, and a is the acceleration of the target.
[0114] The state transition equation is expressed as:
[0115] X t =X t-1 +[Δx Δy ωΔt aΔt ω a]+n
[0116] Where Δt is the trajectory sampling interval, set to 0.1s. Δx and Δy are the changes in the target's position in the x and y directions within a time interval, specifically expressed as:
[0117]
[0118] Where n is the process noise, which can be simulated using Gaussian white noise, specifically expressed as:
[0119]
[0120] Where, n a and n ω These are acceleration noise and angular velocity noise, respectively. Because the influence of noise is considered in the state transition equation, the simulated trajectory more closely matches the actual target trajectory, thus ensuring the accuracy of model training.
[0121] Step A3: Based on the actual trajectory and observation equation, obtain noisy measurement data.
[0122] In this embodiment of the application, after obtaining the true trajectory (target state) of the target, the true trajectory is substituted into the observation equation of the sensor to obtain the noisy measurement data corresponding to the sensor.
[0123] For example, the observation equation is expressed as:
[0124] Z t =h(X) t )+m
[0125] Where m is noise, X t Z is the target state at time t, i.e., the actual state at time t; t These are the observation data corresponding to the target state at time t. The observation data can also be represented as Z. t =[r t ,θ t ],Right now:
[0126]
[0127] Where, r t and θ t These are the sensor observation distances r and r. t and angle θ t m r and m θ These are sensor observation range noise and observation angle noise, respectively. In practical implementation, to match real aircraft tracking scenarios, the sensor (such as radar) observation range, maneuvering target speed, maneuvering target rotation rate, transfer noise intensity, and sensor measurement noise intensity are set according to the parameters shown in Table 1.
[0128] Table 1 Trajectory Parameter Setting Indicators
[0129]
[0130] Step A4: Process the noisy measurement data into the same form as the real trajectory to obtain the measurement trajectory.
[0131] Specifically, to more effectively learn the characteristics of the target trajectory state, noisy measurement data is processed into a measurement trajectory in the same form as the real trajectory, and this measurement trajectory is used as the training input for the target tracking model. For example, the measurement trajectory Y... 1:k Represented as:
[0132] Y 1:k =[Y1,Y2,…,Y k ] T
[0133] in, Location It is calculated based on the sensor's observation distance and angle, specifically expressed as:
[0134]
[0135] In addition, speed The simulation is performed using Gaussian white noise with a standard deviation of 1 and a mean of 0.
[0136] In this embodiment, a high-maneuverability target tracking data pair is constructed based on a constant turning uniform acceleration model as training data. Simultaneously, to more effectively learn the characteristics of the target trajectory state, the noisy measurement data is processed into a measurement trajectory in the same form as the real trajectory. Therefore, the tracking model trained based on this training dataset is more adaptable to high-maneuverability target tracking tasks.
[0137] In one optional embodiment, after obtaining the training data for the target tracking model (i.e., pairs of highly maneuverable target tracking data, each pair including: a true trajectory and a measured trajectory), the target tracking model is trained based on the training data. To enable the target tracking model to learn the mapping relationship from the measured trajectory to the residual trajectory, during training, the predicted residual trajectory output by the model is compared with the true residual trajectory, and a loss function is calculated. Specifically, for N pairs of training data, the root mean square error is calculated as the loss function for the target tracking model, expressed as:
[0138]
[0139] Where k represents the duration of the trajectory, O is the predicted residual trajectory of the target tracking model, and R is the true residual trajectory between the actual trajectory and the measurement trajectory input to the target tracking model.
[0140] Furthermore, through the backpropagation process during the target tracking model training, the loss function is minimized, causing the high-maneuverability target tracking model to converge, resulting in a well-trained target tracking model. This well-trained target tracking model has the ability to predict the residual trajectory between the real target trajectory and the measured trajectory. Therefore, in actual target tracking, the residual trajectory between the real target trajectory and the measured trajectory can be directly predicted, and the target trajectory can then be calculated based on the residual trajectory to achieve target tracking.
[0141] This application proposes a maneuvering target tracking method based on the Unet framework and multi-head attention mechanism. Through downsampling in the encoder branch and upsampling in the decoder branch, feature discarding and feature fitting in the temporal dimension of state features are achieved, reducing the interference of measurement trajectory noise on the extraction of true correlation information between states and facilitating the learning of transition patterns. Simultaneously, by leveraging the U-structure of the Unet framework, multi-level encoder features and similarity matrices output from the encoder branch are connected to the decoder branch, achieving fusion of shallow and deep features in the feature dimension and attention-enhanced fusion in the temporal dimension. Therefore, this method can better capture the drastic state change patterns of highly maneuvering targets, thereby improving the performance of highly maneuvering target tracking tasks.
[0142] Furthermore, to verify the effectiveness of the maneuvering target tracking model (i.e., the AUMTT model) in this embodiment, five ablation experiments were designed to evaluate the effects of using a global normalization layer, skipping the encoder similarity matrix to the decoder, using a noisy activation function, and using downsampling and upsampling processes in the AUMTT model. This embodiment uses a test trajectory to compare the results of the five ablation experiments with the AUMTT model; the parameters of the test trajectory are shown in Table 2. All six experiments (i.e., the AUMTT model and the five ablation experiments) used the same training dataset and parameter settings, and were trained for 100 epochs each. Table 3 lists the mean and standard deviation of the root mean square error of position tracking and the root mean square error of velocity tracking for the five ablation experiments and the AUMTT model when tracking the test trajectory.
[0143] Table 2 Parameter settings for the test trajectory
[0144]
[0145] Table 3 Comparison of target tracking results between ablation experiments and the AUMTT model.
[0146]
[0147] The results are analyzed as follows:
[0148] (1) Specific impact of using different normalization layers on tracking results: Ablation experiments 1 and 2 respectively show the model using instance normalization layers and layer normalization layers, while the AUMTT model uses the newly proposed global normalization layer. These two ablation experiments demonstrate the effectiveness of the newly proposed global normalization layer on tracking results. From the experimental results of the test trajectories in Table 3, it can be seen that while using instance normalization layers can guarantee performance in tracking position, it fails to converge at all in terms of velocity. When layer normalization layers are used in the model, neither position performance nor velocity performance can be guaranteed, and the test trajectory cannot be correctly tracked. Compared to the global normalization layer proposed in this embodiment, AUMTT can guarantee both position and velocity performance, greatly improving tracking performance.
[0149] (2) Specific impact of using the encoder similarity matrix to the decoder jump connection on the tracking results: In ablation experiment 3, the encoder-decoder similarity matrix jump connection was removed. According to the tracking performance results in Table 3, jumping the encoder branch similarity matrix to the decoder can improve position performance and the velocity tracking performance is improved even more.
[0150] (3) Specific impact of using a noisy activation function on tracking results: The ReLU activation function was used in ablation experiment 4. The classic transformer encoder structure only uses the ReLU activation function in the first fully connected layer of the fully connected network to introduce nonlinearity, while the AUMTT model proposed in this application uses a noisy tanh activation function instead of the ReLU function. The noisy tanh activation function adds random noise to the original tanh nonlinear activation function, providing more randomness for model training and accelerating the training process. As can be seen from Table 3, using a noisy activation function can improve tracking performance.
[0151] (4) Specific impact of upsampling and downsampling processes on tracking results: In ablation experiment 5, downsampling and upsampling operations in the time dimension were removed. For the tracking task, this embodiment uses residual mapping. The model itself outputs a residual trajectory. The residual position in the residual trajectory is the position noise after subtracting the true position from the measured position, while the residual velocity in the residual trajectory is the value after subtracting the true velocity from a small amount of Gaussian noise, which changes over time. The downsampling and upsampling processes can partially discard and fit noisy state features, thus retaining more time-dependent feature subspaces with less noise, which is important for recovering the velocity trajectory that changes over time. According to the tracking results in Table 3, the downsampling and upsampling processes can significantly improve velocity performance.
[0152] Furthermore, to evaluate the performance of the AUMTT model, four maneuvering target tracking scenarios (trajectories 1-4) were set up, each containing five different motion models, with trajectory parameters shown in Table 4. To demonstrate the effectiveness of the AUMTT model, IMM, DeepMTT, and TrTNet models were used as comparison methods, and 100 Monte Carlo simulations were performed to track target trajectories. This application will compare the target tracking results of these four methods and compare the tracking performance under strong maneuvering conditions. In addition, in these 100 simulations, the mean and bias of the root mean square error of the tracking results of the 100 Monte Carlo experiments were compared, and the performance of IMM, DeepMTT, TrTNet, and AUMTT was evaluated respectively. The DeepMTT and TrTNet models used the default parameter settings from the original paper. Furthermore, for ease of comparison, the three deep learning models, DeepMTT, TrTNet, and AUMTT, were all trained and tested using the training dataset constructed above.
[0153] Table 4 Trajectory Design Parameters
[0154]
[0155] Figure 5-8The images show a comparison of target tracking results for four maneuver estimations. Two local regions in each of the four comparison images have been magnified to better observe the tracking results of different methods. Observing the tracking results of the four maneuvering targets, it can be found that the traditional IMM method performs worse than the other three deep learning tracking models, regardless of whether the maneuvering phase is strong or weak. For the Deepmtt model, the TrTNet model, and the AUMTT model proposed in this application, these three deep learning models all achieved good tracking results when tracking the trajectories of the four maneuvering targets, demonstrating the effectiveness of deep learning models in learning different trajectory states in the dataset. Furthermore, local segments 1 (46s-56s) and 2 (68s-78s) of trajectory 1, and local segments 1 (36s-46s) and 2 (57s-67s) of trajectory 3, all contain two trajectory segments with different motion models. For such trajectory segments with varying motion models, the AUMTT model proposed in this application can still track accurately and unaffected, with a smaller error compared to the other two deep learning models. Moreover, for different motion models within the same trajectory, whether it is a strong maneuvering motion model or a weak maneuvering motion model, the AUMTT model outperforms traditional methods and the other two deep learning models.
[0156] In addition to comparing the trajectory tracking results, this application embodiment further provides the mean and deviation of the tracking mean square error for 100 Monte Carlo results of the above four maneuvering targets. Table 5 shows the mean and deviation of the tracking mean square error at different positions, and Table 6 shows the mean and deviation of the tracking mean square error at different speeds, so as to quantitatively compare the tracking performance of the above four methods.
[0157] Table 5. Mean and deviation of tracking mean square error at different segment positions.
[0158]
[0159] Table 6. Mean and deviation of the mean square error for speed tracking at different speed segments.
[0160]
[0161] Tables 5 and 6 show that the three deep learning models consistently outperform the traditional IMM method in position tracking. Compared to the DeepMTT model without a normalization layer and the TrTNet model using two normalization layers, the model proposed in this application demonstrates superior position tracking performance. Compared to DeepMTT, the position RMSE decreases by almost 50%, and compared to the TrTNet model, it decreases by approximately 25%, thus demonstrating the AUMTT model's superior tracking performance. Even under drastic changes in motion, the AUMTT model maintains stable tracking performance. Furthermore, for maneuvering targets with high initial speeds, such as the third maneuvering target, the DeepMTT and TrTNet models show a significant performance drop compared to trajectories with lower initial speeds (2 and 4), while the AUMTT model maintains relatively stable tracking performance. It can handle both high-speed and low-speed trajectories, as well as trajectories with high and low turn rates, exhibiting better generalization ability. Therefore, compared to traditional methods and the other two deep learning maneuvering target tracking models, the AUMTT model is more suitable for maneuvering target tracking and can handle both strong and weak maneuvering targets.
[0162] Furthermore, the computational complexity of the above four methods is discussed, using the computation time of a single iteration for each method to illustrate this. The same Intel(R) Xeon(R) Platinum 8280 CPU, 2.70GHz, and 1TB RAM were used to test the IMM, DeepMTT, TrTNet, and AUMTT models. In one iteration of the tracking process, the IMM method, DeepMTT (2 layers), and the new TrTNet (4 layers) model consumed 1.45ms, 8.08ms, and 1.54ms respectively, while the AUMTT (4 layers) model consumed 1.22ms. The AUMTT model proposed in this application has advantages in terms of time consumption and can be applied to real-time target tracking tasks.
[0163] This application also provides a maneuvering target tracking device based on the Unet framework and multi-head attention mechanism, referring to... Figure 9 As shown, Figure 9 This is a schematic diagram of a maneuvering target tracking device based on the Unet framework and multi-head attention mechanism provided in an embodiment of this application. The device includes:
[0164] The trajectory encoding module 910 is used to perform input encoding and position encoding on the measurement trajectory to obtain the encoded state features. The measurement trajectory is a sequence of multiple measurement data within a preset time period.
[0165] The encoder module 920 is used to input the encoded state features into the downsampling encoder branch for state feature encoding and similarity information extraction, so as to obtain multi-level encoder features and similarity matrix.
[0166] Decoder module 930 is used to input the multi-level encoder features and similarity matrix into the upsampling decoder branch to perform multi-level feature fusion and attention information fusion, so as to obtain decoder features that fuse different levels of features and attention information;
[0167] The trajectory prediction module 940 is used to predict the residual trajectory based on the features of the decoder, and add the residual trajectory and the measurement trajectory to obtain the target trajectory.
[0168] In one alternative embodiment, the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism is implemented by a pre-trained target tracking model. The training data of the target tracking model is: strong maneuvering target tracking data pairs constructed based on a constant turning uniform acceleration model. Each pair of strong maneuvering target tracking data pairs includes the actual trajectory and measured trajectory of the maneuvering target in two-dimensional space.
[0169] In an optional embodiment, the apparatus further includes a dataset module for generating the highly maneuverable target tracking trajectory, the dataset module comprising:
[0170] The first data sub-module is used to transform the constant turning uniform acceleration model into different motion models by setting different control parameters, including: turning rate and acceleration;
[0171] The second dataset submodule is used to obtain the true trajectory based on the state variable equation and the state transition equation;
[0172] The third dataset submodule is used to obtain noisy measurement data based on the real trajectory and observation equation;
[0173] The fourth data set module is used to process the noisy measurement data into the same form as the real trajectory to obtain the measurement trajectory.
[0174] In one optional embodiment, the trajectory encoding module includes:
[0175] The first trajectory encoding submodule is used to process the measurement trajectory sequentially using a linear transformation layer and a global normalization layer to obtain initial encoding features. The global normalization layer uses the same normalization parameter in both the time dimension and the feature dimension.
[0176] The second trajectory encoding submodule is used to add the absolute position encoding feature to the initial encoding feature to obtain the encoded state feature, and the encoded state feature retains the difference information of the measurement trajectory in the time dimension and the feature dimension.
[0177] In one optional embodiment, the downsampling encoder branch includes N encoders, which are connected via downsampling. The encoder module includes:
[0178] The first encoder submodule includes a mechanism for the first to Nth encoders to sequentially extract similarity information from the input features based on a multi-head attention mechanism, thereby obtaining encoder features and similarity matrices from shallow to deep levels. The input features of the first encoder are the encoded state features, and the input features of the second to Nth encoders are the features obtained by downsampling the encoder features output by the previous level encoder.
[0179] In one optional embodiment, the upsampling decoder branch includes N decoders, which are connected via upsampling. The first decoder is connected to the Nth encoder in the downsampling encoder branch through an intermediate module. The encoder features output by the first to Nth encoders and the similarity matrix correspond to the Nth to first decoders. The decoder module includes:
[0180] The first decoder submodule is used by the first to Nth decoders to sequentially perform feature fusion and attention information fusion on the input features based on a multi-head attention mechanism, so as to obtain the decoder features output by each decoder. The input features of the first decoder are the features obtained by upsampling the output features of the intermediate module and the features of the Nth level encoder. The similarity matrix input of the first decoder is the similarity matrix output by the Nth encoder. The input features of the second to Nth decoders are the features obtained by upsampling the decoder features output by the previous level decoder and the encoder features of the corresponding level. The similarity matrix input of the second to Nth decoders is the similarity matrix output by the encoder of the corresponding level.
[0181] The second decoder submodule is used to take the decoder features output by the Nth decoder as decoder features that fuse features from different levels and attention information.
[0182] In one optional embodiment, the similarity matrix of the multi-head attention layer in the decoder is obtained by fusing the similarity matrices generated by the multi-head attention layers in the encoder branch in an additive manner.
[0183] In one alternative embodiment, both the encoder and the decoder process the input features using a noisy activation function and a global normalization layer.
[0184] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor performs steps to implement the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism described in this application.
[0185] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0186] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.
[0187] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.
[0188] The foregoing has provided a detailed description of a maneuvering target tracking method and apparatus based on the Unet framework and multi-head attention mechanism provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A maneuvering target tracking method based on the Unet framework and multi-head attention mechanism, characterized in that, The method includes: The measurement trajectory is encoded by input and position to obtain encoded state features. The measurement trajectory is a sequence of multiple measurement data within a preset time period. The encoded state features are input into the downsampling encoder branch for state feature encoding and similarity information extraction, resulting in multi-level encoder features and similarity matrices. The multi-level encoder features and similarity matrix are input into the upsampling decoder branch to perform multi-level feature fusion and attention information fusion, resulting in decoder features that fuse different levels of features and attention information. The residual trajectory is predicted based on the features of the decoder, and the residual trajectory is added to the measured trajectory to obtain the target trajectory; The process of input encoding and position encoding of the measurement trajectory to obtain encoded state features includes: The measurement trajectory is processed sequentially using a linear transformation layer and a global normalization layer to obtain initial encoded features. The global normalization layer uses the same pair of normalization parameters in both the time dimension and the feature dimension. The absolute position encoding feature is added to the initial encoding feature to obtain the encoded state feature, which retains the difference information of the measurement trajectory in the time dimension and the feature dimension.
2. The method according to claim 1, characterized in that, The maneuvering target tracking method based on the Unet framework and multi-head attention mechanism is implemented through a pre-trained target tracking model. The training data of the target tracking model is: strong maneuvering target tracking data pairs constructed based on a constant turning uniform acceleration model. Each strong maneuvering target tracking data pair includes the actual trajectory and measured trajectory of the maneuvering target in two-dimensional space.
3. The method according to claim 2, characterized in that, The highly maneuverable target tracking data is constructed according to the following steps: The constant turning uniform acceleration model can be transformed into different motion models by setting different control parameters, including: turning rate and acceleration; The true trajectory is obtained based on the state equation and the state transition equation; Based on the actual trajectory and observation equation, noisy measurement data is obtained; The noisy measurement data is processed into the same form as the actual trajectory to obtain the measurement trajectory.
4. The method according to claim 1, characterized in that, The downsampling encoder branch includes N There are several encoders connected via downsampling. The encoded state features are input to the downsampling encoder branch for state feature encoding and similarity information extraction, resulting in a multi-level encoder feature and similarity matrix, including: The first to the second N Each encoder, based on a multi-head attention mechanism, sequentially extracts similarity information from the input features, resulting in encoder features and similarity matrices from shallow to deep levels. The input features of the first encoder are the encoded state features, and the features of the second to third encoders are... N The input features of each encoder are obtained by downsampling the encoder features output by the previous stage encoder.
5. The method according to claim 4, characterized in that, The upsampling decoder branch includes N There are several decoders, which are connected via upsampling. The first decoder is connected to the second decoder in the downsampling encoder branch via an intermediate module. N The encoders are connected, the first to the second N Jump to the encoder features output by the encoder and the similarity matrix corresponding to the first encoder. N One to the first decoder; The multi-level encoder features and similarity matrix are input into the upsampled decoder branch for multi-level feature fusion and attention information fusion, resulting in decoder features that fuse features and attention information from different levels, including: The first to the second N Each decoder is based on a multi-head attention mechanism, sequentially fusing input features and a similarity matrix to obtain decoder features output by each decoder. The input features of the first decoder are the features obtained by upsampling the output features of the intermediate module and the features of the second decoder. N The features of the first-level encoder, the similarity matrix input of the first decoder is the first-level encoder feature. N The similarity matrix output by the encoder, from the 2nd to the 3rd... N The input features of each decoder are the features obtained by upsampling the decoder features output from the previous level decoder and the encoder features of the corresponding level. (The second to the third level...) N The similarity matrix input for each decoder is the similarity matrix output by the encoder at the corresponding level; The first N The decoder features output by each decoder are used as decoder features that fuse features from different levels and attention information.
6. The method according to claim 5, characterized in that, The similarity matrix of the multi-head attention layer in the decoder is obtained by fusing the similarity matrices generated by the multi-head attention layers in the encoder branch in an additive manner.
7. The method according to claim 4 or 5, characterized in that, Both the encoder and the decoder use a noisy activation function and a global normalization layer to process the input features.
8. A maneuvering target tracking device based on the Unet framework and multi-head attention mechanism, characterized in that, The device includes: The trajectory encoding module is used to encode the input and position of the measurement trajectory to obtain the encoded state features. The measurement trajectory is a sequence of multiple measurement data within a preset time period. The encoder module is used to input the encoded state features into the downsampling encoder branch for state feature encoding and similarity information extraction, thereby obtaining multi-level encoder features and similarity matrices. The decoder module is used to input the multi-level encoder features and similarity matrix into the upsampling decoder branch to perform multi-level feature fusion and attention information fusion, so as to obtain decoder features that fuse different levels of features and attention information. The trajectory prediction module is used to predict the residual trajectory based on the features of the decoder, and add the residual trajectory and the measurement trajectory to obtain the target trajectory; The trajectory encoding module is further configured to process the measurement trajectory sequentially using a linear transformation layer and a global normalization layer to obtain initial encoding features. The global normalization layer uses the same pair of normalization parameters in both the time and feature dimensions. The absolute position encoding features are added to the initial encoding features to obtain the encoded state features. The encoded state features retain the difference information of the measurement trajectory in both the time and feature dimensions.
9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the steps of the maneuvering target tracking method based on the Unet framework and multi-head attention mechanism as described in any one of claims 1-7.