Radar target tracking method based on attention mechanism

Abstract

The invention discloses a radar target tracking method based on an attention mechanism, and relates to the technical field of radar tracking, and the method comprises the steps: obtaining a radar observation sequence, each observation vector in the radar observation sequence comprising an azimuth angle and a distance of a target under polar coordinates; the radar observation sequence is converted into an observation sequence under a rectangular coordinate system, center maximum value normalization processing is carried out on a conversion result, and a normalized observation sequence is established; inputting the normalized observation sequence into a self-attention target tracking neural network, extracting global trajectory features through a multi-head self-attention mechanism, and outputting a normalized target state sequence; and carrying out reverse normalization processing on the target state sequence, and establishing an actual motion track of the target. The technical problem that global trajectory features are difficult to model effectively when a maneuvering target is tracked in the prior art is solved, the global trajectory features are extracted by introducing a multi-head self-attention mechanism, and the technical effect of improving the target tracking precision and stability is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar tracking, and particularly relates to a radar target tracking method based on an attention mechanism. BACKGROUND

[0002] In a target tracking task, traditional methods mostly rely on recursive models such as Kalman filtering, extended Kalman filtering or particle filtering, which have certain effects in processing linear systems and Gaussian noise. However, in the face of highly maneuverable targets with frequent trajectory changes, such methods can only use the current observation and limited historical states, and it is difficult to fully capture the movement rules of the target on a long time scale. The recurrent neural network and its variant LSTM introduce time series modeling capability, which to some extent alleviates the problem of short-term information dependence, but still mainly focuses on local time segments, and it is difficult to establish global connections across time steps, resulting in insufficient modeling capability of the overall trend of the trajectory, thereby affecting the tracking accuracy and robustness. SUMMARY

[0003] The present application provides a radar target tracking method based on an attention mechanism, which is used to solve the technical problem that the prior art is difficult to effectively model global trajectory features when tracking maneuverable targets.

[0004] In view of the above problems, the present application provides a radar target tracking method based on an attention mechanism.

[0005] The present application provides a radar target tracking method based on an attention mechanism, which comprises:

[0006] Obtaining a radar observation sequence, wherein each observation vector in the radar observation sequence comprises the azimuth angle and the distance of the target in polar coordinates; converting the radar observation sequence into an observation sequence in a rectangular coordinate system, and performing center maximum normalization processing on the conversion result to establish a normalized observation sequence; inputting the normalized observation sequence into a self-attention target tracking neural network, extracting global trajectory features through a multi-head self-attention mechanism, and outputting a normalized target state sequence; and performing inverse normalization processing on the target state sequence to establish an actual motion trajectory of the target.

[0007] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0008] The application obtains a radar observation sequence, wherein each observation vector in the radar observation sequence includes the azimuth angle and the distance of a target in polar coordinates; the radar observation sequence is converted into an observation sequence in a rectangular coordinate system, and the conversion result is subjected to center maximum normalization processing to establish a normalized observation sequence; the normalized observation sequence is input into a self-attention target tracking neural network, global trajectory features are extracted through a multi-head self-attention mechanism, and a normalized target state sequence is output; and the target state sequence is subjected to inverse normalization processing to establish a target actual motion trajectory. The application solves the technical problem that the prior art is difficult to effectively model global trajectory features when tracking a maneuvering target, and achieves the technical effect of improving target tracking precision and stability by introducing a multi-head self-attention mechanism to extract global trajectory features. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0010] Figure 1 A radar target tracking method based on an attention mechanism provided by the embodiment of the present application;

[0011] Figure 2 A self-attention target tracking neural network structure provided by the embodiment of the present application;

[0012] Figure 3 A local diagram of the self-attention target tracking neural network structure provided by the embodiment of the present application;

[0013] Figure 4 A trajectory sequence processing flowchart provided by the embodiment of the present application;

[0014] Figure 5 A diagram for identifying trajectory tracking in the X-Y plane;

[0015] Figure 6 A diagram for point-by-point position RMSE;

[0016] Figure 7 A diagram for point-by-point velocity RMSE. DETAILED DESCRIPTION

[0017] The radar target tracking method based on an attention mechanism is provided, and the technical problem that it is difficult to effectively model global trajectory features when tracking a maneuvering target in the prior art is solved. The global trajectory features are extracted by introducing a multi-head self-attention mechanism, so that the technical effects of improving target tracking accuracy and stability are achieved.

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0019] It should be noted that any variation of the terms "comprise" and "have" is intended to cover non-exclusive inclusion, for example, a process, method, system, product or server comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] As shown in the embodiments, the present application provides a radar target tracking method based on an attention mechanism, which comprises: Figure 1 and Figure 2 The radar target tracking method based on an attention mechanism is provided, and the technical problem that it is difficult to effectively model global trajectory features when tracking a maneuvering target in the prior art is solved. The global trajectory features are extracted by introducing a multi-head self-attention mechanism, so that the technical effects of improving target tracking accuracy and stability are achieved.

[0021] Step S100: obtaining a radar observation sequence, wherein each observation vector in the radar observation sequence comprises a bearing angle and a distance of a target in polar coordinates.

[0022] In the embodiments of the present application, in order to obtain the radar observation sequence, periodic scanning of the target is realized by the radar system at a fixed sampling frequency. Each scanning result forms an observation vector, and the observation vector comprises a bearing angle and a distance of the target in polar coordinates. The bearing angle refers to the angular deviation of the target relative to the radar reference direction (usually the north direction), and the unit of measurement is degree, which reflects the angular position of the target in the horizontal plane; the distance refers to the radial interval between the target and the radar antenna, and the unit is meter, which is used to represent the absolute position of the target in the radial direction.

[0023] In the process of obtaining the radar observation sequence, the radar system transmits electromagnetic waves and receives the signals reflected by the target, calculates the round-trip time of the electromagnetic waves, and combines the pointing angle of the antenna to obtain the bearing angle and distance observation value at each time point. Through multiple consecutive observations, the radar observation sequence is formed.

[0024] Step S200: converting the radar observation sequence into an observation sequence in a rectangular coordinate system, and performing center maximum value normalization processing on the conversion result to establish a normalized observation sequence.

[0025] In the embodiment of the present application, the radar observation sequence is first converted into an observation sequence in a rectangular coordinate system. Specifically, the radar observation sequence is composed of polar coordinate observation vectors at multiple time steps, and each observation vector includes a set of azimuth angles θ k and corresponding distances d k Each observation vector is converted from polar coordinate form to rectangular coordinate system form by using the conversion method of The resulting constitutes an observation sequence in a rectangular coordinate system, which is the conversion result.

[0026] The conversion result is then subjected to center maximum value normalization processing. In this process, a preset window length and a step threshold are first configured, and the conversion result is subjected to sliding segmentation processing to build multiple segmented observation sequences. Then, each segmented observation sequence is sequentially subjected to center maximum value normalization processing, and a normalized observation sequence is finally established.

[0027] Further, in the method provided by the application embodiment, before the conversion result is subjected to center maximum value normalization processing to establish a normalized observation sequence, the following steps are included:

[0028] A preset window length and a step threshold are configured; the segmented processing of the conversion result is performed by using the preset window length and the step threshold to establish a segmented observation sequence; and the segmented observation sequence is sequentially subjected to center maximum value normalization processing to establish a normalized observation sequence.

[0029] In the embodiment of the present application, a preset window length and a step threshold are first configured, wherein the preset window length is 10 and the step threshold is 5. Then, the segmented processing of the conversion result is performed by using the preset window length and the step threshold. In this process, a sliding window method is used to extract equal-length trajectory segments from the conversion result with the parameters of a window length of 10 and a step of 5, forming multiple segmented observation sequences with consistent structures.

[0030] Finally, the segmented observation sequences are sequentially subjected to center maximum value normalization processing to establish a normalized observation sequence. For each segmented observation sequence, the coordinates corresponding to the first time step thereof are taken as a center reference point, and the offsets of all observation points in the segment relative to the center are calculated. Then, the offsets are normalized according to a set maximum distance threshold to be scaled to a unified scale range, forming a normalized observation sequence.

[0031] Further, in the method provided by the application embodiment, the center maximum value normalization processing is calculated as follows:

[0032]

[0033] wherein, characterizing a normalized observation value, characterizing an observation value of a rectangular coordinate system at the kth time step, characterizing an observation value of a rectangular coordinate system at the first time step, D max characterizing a possible maximum moving distance threshold value of the sequence length K.

[0034] In the embodiments of the present application, for each segment of the segmented observation sequence, let the observation value of the rectangular coordinate system at the kth time step be the observation value at the first time step is In order to eliminate the influence of the initial position on the trajectory distribution, take as the reference center to calculate the relative offset of each time step, and use the possible maximum moving distance D max of the target in the segment of the trajectory to perform normalization. Wherein, D max is a reference scale preset by a technical expert. In the specific calculation, take as the reference point to calculate the relative offset between the observation value of each time step and it. Then scale the offset with the unified D max to obtain the normalized observation value Finally, the of all time steps constitute the normalized observation sequence of the segment of the trajectory

[0035] Step S300: input the normalized observation sequence into the self-attention target tracking neural network, extract global trajectory features through the multi-head self-attention mechanism, and output the normalized target state sequence.

[0036] In the embodiments of the present application, after inputting the normalized observation sequence into the self-attention target tracking neural network, first process the input sequence through the position encoding structure in the self-attention target tracking neural network, and introduce the position information of each observation value in the sequence. The position encoding structure adopts learnable position encoding, that is, each time step is mapped to a trainable position vector, so that the subsequent network module can identify the sequential relationship between different time steps.

[0037] After position encoding, the normalized observation sequence with position information is input into an N-layer stacked Transformer encoder. Each Transformer encoder is composed of a multi-head self-attention sublayer and a feed-forward fully connected sublayer, and is connected with a residual connection and a normalization operation after each sublayer. In the multi-head self-attention mechanism, the self-attention target tracking neural network maps the input sequence into query vectors, key vectors and value vectors respectively, calculates the correlation between each time step, aggregates the information of the whole sequence by weighting, realizes the global modeling of the target trajectory, and extracts the global trajectory features.

[0038] The global trajectory features output by the Transformer encoder are then input into a decoder module composed of two layers of 1-D full convolution. The decoder module maps the high-dimensional features into trajectory state representations through consecutive one-dimensional convolution operations, and outputs a normalized target state sequence consistent with the input time steps.

[0039] Further, the method provided by the application embodiment further comprises:

[0040] The self-attention target tracking neural network comprises a position encoding structure, an N-layer stacked Transformer encoder and a decoder module composed of two layers of 1-D full convolution, wherein each Transformer encoder comprises a multi-head self-attention sublayer, a feed-forward fully connected network sublayer, and two residual connection layers respectively arranged after the two sublayers.

[0041] In the application embodiment, as shown in Figure 3 The self-attention target tracking neural network is used for processing the normalized observation sequence and outputting a normalized target state sequence. The self-attention target tracking neural network comprises a position encoding structure, an N-layer stacked Transformer encoder and a decoder module composed of two layers of 1-D full convolution, wherein each Transformer encoder comprises a multi-head self-attention sublayer, a feed-forward fully connected network sublayer, and two residual connection layers respectively arranged after the two sublayers.

[0042] Specifically, the normalized observation sequence is first input into the position encoding structure. The function of the position encoding structure is to introduce the position information of each time step in the observation vector in the time sequence, so that the subsequent network can identify the relative or absolute sequence relationship between different time steps in the sequence. By using a learnable position encoding structure, a trainable position vector is assigned to each time step, which is embedded by adding the observation value and dynamically updated during the training process to adapt to the time sequence modeling requirements of specific tasks.

[0043] Then the sequence processed by the position encoding structure is input into an N-layer stacked Transformer encoder. Each layer of the Transformer encoder includes a multi-head self-attention sublayer and a feedforward fully connected network sublayer. In the multi-head self-attention sublayer, the input representation of each time step is mapped into a query vector, a key vector and a value vector, the attention weight is obtained by calculating the dot product between the query vector and the key vector and normalizing, and then the value vector is weighted and summed using the weight, so as to realize the aggregation modeling of the full sequence information. The feedforward fully connected network sublayer performs nonlinear mapping on the feature representation of each time step to enhance the representation ability. A residual connection layer and a normalization operation are arranged after each sublayer to ensure the stability of gradient propagation and the efficiency of model training. Through the stacking of N layers, the Transformer encoder can extract global trajectory features in the normalized observation sequence layer by layer.

[0044] Finally, the output result of the Transformer encoder is input into a decoder module composed of two layers of 1-D full convolution. The decoder module maps the high-dimensional global trajectory features extracted by the encoder into the target state representation consistent with the number of time steps through one-dimensional convolution operation in the continuous time sequence direction. Through the local receptive field and weight sharing characteristics of the convolution kernel, the modeling of the trend of the target state changing over time is realized. Finally, the decoder module outputs the normalized target state sequence consistent with the number of input time steps.

[0045] Further, the method provided in the application embodiment, the position encoding structure is a learnable position encoding, as follows:

[0046]

[0047] wherein s 1:K [i] represents the encoding result of the i-th subspace, is a normalized observation sequence, is a sine function, K represents the length of the sequence, i is the dimension number, and ω i and is a learnable parameter, and ω i is a learnable weight coefficient of the i-th dimension, is a bias term of the i-th dimension, and d e is the encoding dimension.

[0048] In the application embodiment, the position encoding structure adopts a learnable position encoding. Specifically, for the case where the dimension number is i=0, the normalized observation sequence is linearly weighted and a bias is added to form a linear encoding result.

[0049] For 1≤i≤d eIn the case of a linear combination, a sinusoidal function is introduced after the linear combination Non-linear transformation is performed to introduce periodic, multi-frequency and other complex coding features. By training ω i and jointly, the response ability of each dimension coding to different inputs is dynamically optimized, so as to improve the accuracy and robustness of position expression.

[0050] Further, the method provided by the application embodiment is used for:

[0051] The normalized observation sequence is mapped into query vectors, key vectors and value vectors through linear transformation; the query vectors, key vectors and value vectors are divided into M sub-sequences along the coding dimension, wherein each sub-sequence corresponds to an attention head, the attention head obtains attention weights through similarity calculation of vector elements of any two positions in the corresponding sequence, and M independent attention output results are generated in parallel; the attention output results of the M attention heads are spliced, and a multi-head self-attention output representation is generated through linear mapping.

[0052] In the application embodiment, first, the normalized observation sequence is linearly mapped to obtain query vectors Q, key vectors and value vectors V. Wherein, W Q , and are learnable matrices.

[0053] Then, the obtained Q, and V are divided into M sub-sequences (i.e. attention heads) along the coding dimension d e , and the dimension of each sub-sequence is corresponding to an independent attention head. For the mth attention head, the corresponding sub-vector is used to obtain attention output through scaled dot-product attention calculation. The process extracts attention features in the subspace by calculating the similarity of queries and keys at any two time steps, and weighting and aggregating the value vectors using the attention weights.

[0054] Finally, the output results of all M attention heads are spliced along the coding dimension, and the fused attention representation S M is obtained through splicing operation Concat (head1, …, head attention as the output of the current multi-head self-attention layer.

[0055] Further, the method provided by the application embodiment further comprises: obtaining the attention weight by performing similarity calculation on vector elements of any two positions in the corresponding sequence.

[0056]

[0057] wherein head m denotes an output of the mth attention head, Q m 、 V m denote a query matrix, a key matrix, and a value matrix respectively, wherein Q m 、 R denotes a real number set, d m denotes a vector dimension of each position, K denotes a sequence length, and M denotes a number of attention heads, denotes a similarity matrix, is a scaling factor.

[0058] In the application embodiment, in the multi-head self-attention mechanism, the core target of each attention head is to capture the dependency between positions in the input sequence, and the context semantic enhancement is achieved through attention weighting. To this end, the normalized observation sequence of the input is first subjected to linear mapping operation to generate a query matrix Q m , a key matrix , and a value matrix V m , wherein m denotes the mth attention head, and M is the total number of attention heads. The dimensions of these matrices are all wherein denotes the vector dimension in each attention head (obtained by equally dividing the original input dimension d e ), K is the length of the input sequence, and denotes the number of vectors contained in the sequence. R denotes a real number set, that is, the elements in all the above matrices are real numbers, which are values in the R domain, ensuring that the entire calculation process is performed on the real number domain.

[0059] Next, the attention mechanism generates a KxK similarity matrix by performing matrix multiplication on the query matrix and the key matrix , which is used to represent the strength of the correlation between any two positions in the sequence. Since the dot product calculation may cause the value to be too large and thus affect the gradient stability, the similarity matrix is scaled in actual operation, that is, divided by to obtain the scaled similarity weight matrix. Then, the matrix is input into the softmax function for normalization processing, so that the sum of the attention weights of each row is 1, thereby forming a stable attention distribution.

[0060] After obtaining the normalized attention weight, it is combined with the corresponding value matrix V mThe weighted matrix multiplication is performed to obtain an output head of the mth attention head m The output represents the representation of each position in the input sequence after the global semantic information is fused under the current attention head. The above process is repeated in parallel to obtain the output results of the M attention heads. Finally, the outputs of the M attention heads are spliced in the feature dimension, thereby completing the construction and expression learning process of the entire multi-head self-attention sublayer.

[0061] Further, the method provided in the application embodiment further comprises:

[0062] The Transformer encoder in the self-attention target tracking neural network is 4 layers, and the number of attention heads of the multi-head self-attention sublayer is 8.

[0063] In the application embodiment, in the self-attention target tracking neural network, a deep structure composed of a 4-layer Transformer encoder stack is adopted, and each layer of the Transformer includes a group of multi-head self-attention mechanisms and a feedforward neural network structure. In each layer, the normalized sequence is first subjected to context interaction modeling by the multi-head self-attention mechanism. Specifically, the multi-head self-attention mechanism generates a query matrix Q m , a key matrix K and a value matrix V m , m = 1, 2,..., 8, indicating that 8 parallel attention heads are set in each layer of the network, and each attention head models the dependency relationship between the sequences in different representation subspaces.

[0064] In each attention head, the similarity between the query matrix and the key matrix is first calculated by dot product, and then input into the softmax function after processing by a scaling factor , to obtain the attention distribution corresponding to each position, and then multiplied by the value matrix to obtain a weighted representation head m . After the 8 attention heads output their corresponding context encoding results, the results are spliced in the feature dimension and mapped to the original dimension by linear transformation, thereby completing a complete multi-head self-attention processing.

[0065] The operation is sequentially performed in each layer of the Transformer encoder, and the accumulation of the 4-layer stack structure enables the model to have stronger hierarchical semantic extraction capability and deeper global dependency modeling capability, thereby providing a stable and discriminative feature basis for subsequent target representation fusion and matching.

[0066] Further, in the method provided in the application embodiment, the self-attention target tracking neural network is obtained by training through the following steps:

[0067] Step S1: constructing a motion model library, wherein a plurality of trajectory segments are integrated in the motion model library, each trajectory segment comprising real trajectory and noise data; Step S2: synchronizing the trajectory segments of the motion model library to a data preprocessing module to perform normalization processing; Step S3: sending the data after normalization processing to a self-attention target tracking neural network for training, calculating the root mean square error loss through real sequences, and propagating the error loss back to the network to update the parameters; and Step S4: performing iterative optimization, and when the convergence condition is met, the training is completed.

[0068] In the embodiments of the present application, in the training process of the self-attention target tracking neural network, a motion model library for training is first constructed. The motion model library is constructed based on historical motion trajectory data collected, and a plurality of representative trajectory segments are extracted from large-scale historical trajectory data through clustering analysis, trajectory segmentation and other methods. Each trajectory segment comprises a complete real trajectory and artificial superimposed noise data, which is used to simulate the motion change of the target under different observation conditions. The real trajectory is obtained by actual measurement or high-precision labeling, and the noise data is generated by applying Gaussian disturbance, random jitter or occlusion simulation to the real trajectory.

[0069] Subsequently, all trajectory segments in the motion model library are input to a data preprocessing module. In this stage, the center maximum value normalization method is used to standardize the spatial position sequence of each trajectory. Specifically, for the state sequence of each trajectory segment, the observation value at the first time step is taken as the reference center, and the position and velocity states at all time steps are offset. Subsequently, the offset trajectory state vector is divided by the preset maximum state scale vector X max , so as to realize normalization processing. The calculation method is , wherein x 1:K represents the original state sequence, is the normalized state sequence, represents the reference center composed of the observation value at the first time step, X max is the maximum value range of the state vector.

[0070] After that, the trajectory data after completing the normalization processing is input to the self-attention target tracking neural network for training. The neural network structure comprises a plurality of Transformer encoders and embeds a multi-head self-attention mechanism to model the time sequence relationship and position dependence within the trajectory sequence. In the training stage, the supervised learning method is adopted, the trajectory predicted by the network is compared with the original real trajectory point by point, and the root mean square error therebetween is calculated as the loss function. The loss function reflects the overall deviation between the model prediction result and the true value, and the error result is used for the back propagation algorithm to optimize the weight parameters in the network and gradually adjust the prediction ability of the model.

[0071] The training process is performed in an iterative manner, in each iteration, the network parameters are constantly updated, and the loss function value is gradually reduced. When the set convergence condition is met, such as the loss function change of a plurality of consecutive rounds is less than the set threshold, the training process is automatically terminated. At this time, the training of the self-attention target tracking neural network is completed.

[0072] Step S400: performing reverse normalization processing on the target state sequence to establish a target actual motion trajectory.

[0073] Further, the method provided in the application embodiment is executed as follows:

[0074]

[0075] wherein, The actual state sequence representing the rth trajectory segment, P=5 is the step length, is a preset window length, is the normalized target state sequence, represents element-wise multiplication, X max = [D max , D max , V max , V max ], V max is the maximum speed of the simulation target, is the position component of the state vector x1.

[0076] In the application embodiment, for the normalized target state sequence Reverse normalization is achieved by Hadamard product (element-wise multiplication) with the maximum state range vector X max to obtain the actual state sequence at time point 1+rP wherein, is a preset window length, P=5 is the step length.

[0077] is the position component of the state vector x1. X max = [D max , D max , V max , V max ], V max is the maximum speed of the simulation target.

[0078] After the above operation, all the reverse normalized state segment sequences Perform segment splicing. There is an overlapping area of ​​KP in adjacent state segments, which needs to be fused to ensure the continuity of the state sequence. and the r+1th state sequence Merge into a line of length Splicing state sequence The overlapping part of the two state sequences and The fusion is performed using a weighted average method, and the fusion result is The above fusion operation is performed between each pair of adjacent actual state segments, with the aim of constructing a numerically continuous transition zone at the state connection of adjacent time periods to eliminate the mutation or discontinuity problems caused by segmented prediction.

[0079] After the fusion is completed, the first (P-1) state of the r-th state sequence Status of the fused area And the r+1th state sequence Status Splice in sequence to get a complete splicing state sequence This iteration is carried out, such as Figure 4 As shown, the target state trajectory is finally constructed.

[0080] Based on the aforementioned radar target tracking method based on the attention mechanism, the tracking performance was tested through experiments. The test scope is shown in Table 1.

[0081] Table 1 Test range

[0082] Distance range 1 km ~ 10 km Angle range 0°～360° Speed range -300 m / s ~ 300 m / s Turning rate -10 ° / s ~ 10 ° / s

[0083] During the experiment, 200,000 tracks were generated based on SSM as a data set, with distance and angle ranges of 1 to 10 km and 0 to 360 degrees, and the speed and turning rate of the maneuvering target were randomly sampled in the range of -300 to 300 m / s and -10 to 10 degrees / s. 2 , 8m / s 2 ]The standard deviation σ of the random sampling acceleration noise within the range α The standard deviation of the azimuth and distance noises are randomly sampled in the ranges [5m, 8m] and [0.1°, 0.3°]. In addition, assuming that D max =3km, V max =300m / s, observe the target every 1s.

[0084] Meanwhile, the self-attention object tracking neural network consists of four encoder layers with eight attention heads. eThe dimension of is 512. The output dimensions of the 1D convolutional layer in the decoder are 64 and 4 respectively. The model is trained using the Adam optimizer with the corresponding parameters β1 = 0.9, β2 = 0.98 and ε = 10 -9 The learning rate is linearly warmed up in the first 10 training cycles and then dynamically adjusted with policy decay. We train for 300 epochs on a single NVIDIA TITAN Xp GPU with a batch size of 64.

[0085] The TBN+CM model, the IMM algorithm, and the LSTM+MM model were then compared. For comparison, an LSTM+CM model was constructed. The LSTM network consists of four hidden layers and has a dimension of 128. The aforementioned networks were trained using the same training set. CM stands for center-max normalization, and MM stands for min-max normalization.

[0086] In the experiments, we first compare the performance of LSTM+MM, LSTM+CM, and TBN+CM on a test set containing 20,000 segmented trajectories. The tracking results are listed in Table 2.

[0087] Table 2. Numerical results of several segmented trajectory tracking methods

[0088]

[0089] In Table 2, the LSTM+CM model achieves smaller position and velocity RMSEs than the LSTM+MM model, demonstrating that center-max normalization improves the network's tracking capabilities by reducing the complexity of trajectory learning. Meanwhile, the TBM+CM model achieves the smallest position and velocity RMSEs. Therefore, it can be concluded that the TBN model performs better than the LSTM model when tracking segmented trajectories.

[0090] Then, a target with an initial state of [2 km, 2 km, 50 m / s, 0 m / s] and a turning angle rate of 0° was simulated and a Monte Carlo simulation was performed to generate an 80-time-step trajectory, which is denoted as A1. The turning angle rates of the target at the 10th and 50th steps were -1° and 3°, respectively. TBN+CM, LSTM+MM, LSTM+CM, and IMM algorithms were evaluated on trajectory A1. The tracking results are listed in Tables 3 and 4. Figure 5 、 Figure 6 and Figure 7 middle.

[0091] Table 3 Numerical results of several methods for tracking A1 trajectory

[0092]

[0093] in, Figure 5 ~ Figure 6The results of tracking the maneuvering target with TBN+CM, LSTM+MM, LSTM+CM and IMM algorithms are shown. Figure 5 Identify the trajectory tracking in the X-Y plane. Figure 6 The point-by-point position RMSE. Figure 7 The point-by-point velocity RMSE. In the listed figures, Figure 5 Show the effect of the algorithm tracking the trajectory. Figure 6 ~ Figure 7 Show the point-by-point RMSE of trajectory A1. Figure 5 TBN+CM represented by the red dot has the smallest error most of the time. The results in bold in Table 3 also show that the tracking error of TBN+CM is the smallest, indicating the superiority of TBN+CM in tracking maneuvering targets.

[0094] In addition, the initial position of trajectory A1 is moved to [12km, 12km] and [15km, 15km] to obtain trajectories A2 and A3. Through the generalization experiment on trajectories A2 and A3, the results show that TBN+CM can be generalized to trajectories beyond the preset distance. However, LSTM+MM fails to track due to its fixed normalization mechanism.

[0095] In the embodiments of the present application, in summary, the embodiments of the present application have at least the following technical effects:

[0096] The radar observation sequence is obtained, wherein each observation vector in the radar observation sequence includes the azimuth angle and the distance of the target in polar coordinates; the radar observation sequence is converted into an observation sequence in a rectangular coordinate system, and the conversion result is subjected to center maximum value normalization processing to establish a normalized observation sequence; the normalized observation sequence is input into a self-attention target tracking neural network, global trajectory features are extracted through a multi-head self-attention mechanism, and a normalized target state sequence is output; and the target state sequence is subjected to inverse normalization processing to establish a target actual motion trajectory. The present application solves the technical problem that the prior art is difficult to effectively model global trajectory features when tracking a maneuvering target, and through the introduction of a multi-head self-attention mechanism to extract global trajectory features, the technical effects of improving target tracking accuracy and stability are achieved.

[0097] It should be noted that the above sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above describes specific embodiments of the present application. The processes depicted in the drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0098] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0099] The specification and drawings are only exemplary and illustrative of the present application and are to be considered within the scope of the present application. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and equivalent technology, the present application is intended to include these modifications and variations.

Claims

1. A radar target tracking method based on attention mechanism, characterized in that: The method comprises: Acquire a radar observation sequence, wherein each observation vector in the radar observation sequence includes an azimuth and a distance of a target in polar coordinates; Converting the radar observation sequence into an observation sequence in a rectangular coordinate system, and performing central maximum normalization processing on the conversion result to establish a normalized observation sequence; The normalized observation sequence is input into the self-attention target tracking neural network, global trajectory features are extracted through the multi-head self-attention mechanism, and the normalized target state sequence is output; The target state sequence is denormalized to establish the target actual motion trajectory.

2. The radar target tracking method based on the attention mechanism according to claim 1, characterized in that The self-attention target tracking neural network includes a position encoding structure, an N-layer stacked Transformer encoder and a decoder module consisting of two layers of 1-D full convolution, wherein each Transformer encoder includes a multi-head self-attention sublayer, a feedforward fully connected network sublayer, and two residual connection layers respectively arranged after the two sublayers.

3. The radar target tracking method based on the attention mechanism according to claim 2, characterized in that The position encoding structure is a learnable position encoding, as follows: Among them, s 1:K [i] represents the encoding result of the i-th subspace, is the normalized observation sequence, is a sine function, K represents the sequence length, i is the dimension number, ω i and is a learnable parameter, ω i is the learnable weight coefficient of the i-th dimension, is the bias term of the i-th dimension, d e The encoding dimension.

4. The radar target tracking method based on the attention mechanism according to claim 2, characterized in that The multi-head self-attention sublayer is used to: The input normalized observation sequence is mapped into query vector, key vector and value vector respectively through linear transformation; Split the query vector, key vector, and value vector into M subsequences along the encoding dimension, where each subsequence corresponds to an attention head. The attention head calculates the similarity between vector elements at any two positions in the sequence to obtain an attention weight, and generates M independent attention output results in parallel. The attention output results of M attention heads are spliced ​​together, and a multi-head self-attention output representation is generated through linear mapping.

5. The radar target tracking method based on the attention mechanism according to claim 4, characterized in that: The attention head obtains the attention weight by performing similarity calculation on the vector elements at any two positions in the corresponding sequence, which is calculated as follows: Among them, head m Represents the output of the mth attention head, Q m 、 V m Represent the query matrix, key matrix, and value matrix respectively, where Q m 、 R represents the set of real numbers, d m The dimension of the vector representing each position, K represents the sequence length, M is the number of attention heads, Representation similarity matrix, is the scaling factor.

6. The radar target tracking method based on the attention mechanism according to claim 2, characterized in that The Transformer encoder in the self-attention target tracking neural network has 4 layers, and the number of attention heads in the multi-head self-attention sublayer is 8.

7. The radar target tracking method based on the attention mechanism according to claim 1, characterized in that The conversion results are normalized to the central maximum value before the normalized observation sequence is established, including: Configure preset window length and step threshold; Performing segmented processing on the conversion results using the preset window length and step threshold to establish a segmented observation sequence; The segmented observation sequences are sequentially subjected to central maximum normalization processing to establish a normalized observation sequence.

8. The radar target tracking method based on the attention mechanism according to claim 1, characterized in that The self-attention target tracking neural network is trained by the following steps: Step S1: constructing a motion model library, wherein the motion model library integrates multiple trajectory segments, wherein each trajectory segment includes a real trajectory and noise data; Step S2: Synchronize the trajectory segments of the motion model library to the data preprocessing module and perform normalization processing; Step S3: Send the normalized data to the self-attention target tracking neural network training, calculate the root mean square error loss through the real sequence, and backpropagate the error loss back to the network to update the parameters; Step S4: Perform iterative optimization. When the convergence condition is met, the training is completed.

9. The radar target tracking method based on the attention mechanism according to claim 1, characterized in that The center maximum normalization process is calculated as follows: in, represents the normalized observation value, Represents the observation value of the rectangular coordinate system at the kth time step, Represents the observation value of the rectangular coordinate system at the first time step, D max Characterizes the maximum possible moving distance threshold under sequence length K.

10. The radar target tracking method based on the attention mechanism according to claim 1, characterized in that: The denormalization process is performed as follows: in, Characterize the actual state sequence of the rth segment trajectory, P = 5 is the step length, =10 is the preset window length, is the normalized target state sequence, ⊙ represents the element-by-element multiplication, X max =[D max , D max , V max , V max ],V max is the maximum speed of the simulation target, is the position component of the state vector x1.

Images