A radar error calibration method and system based on vision converter

Abstract

This invention discloses a radar error calibration method and system based on a vision transformer, belonging to the field of radar data processing and error calibration technology. The method includes: acquiring radar measurement data and AIS data as ground truth; transforming the geographic coordinates of the AIS data to a local Cartesian coordinate system centered on the radar position, and achieving time alignment with the radar measurement time through linear interpolation; calculating the range error sequence and azimuth error sequence based on the aligned data, and weighting the results with confidence levels by combining AIS position accuracy information and ship navigation status information; constructing a radar error calibration model based on a vision transformer architecture; using the trained radar error calibration model to predict errors in the radar data to be calibrated, and completing the calibration. This invention can effectively characterize the long-range dependence and nonlinear relationship in radar errors, improve the accuracy of radar ranging and direction finding error calibration, and has good generalization ability and interpretability.

Description

Technical Field

[0001] This invention relates to the field of radar data processing and error calibration technology, specifically to a method and system for accurately calibrating radar ranging and direction-finding system errors using deep learning models, particularly Vision Transformer networks. This invention is particularly applicable to maritime surveillance radar systems, used to correct system measurement errors of radar targets in the range and azimuth dimensions. Background Technology

[0002] Maritime surveillance radar systems are core technological equipment for ensuring maritime navigation safety, implementing maritime traffic management, and coastal defense monitoring, and are widely used in ship tracking, port scheduling, and maritime search and rescue. However, due to factors such as the radar equipment's own system deviations, antenna pointing errors, atmospheric refraction effects, and signal propagation delays, radar measurements of targets inherently contain systematic errors in both range and azimuth dimensions. If these errors are not effectively calibrated, they will directly lead to a decrease in target track accuracy and a deterioration in tracking continuity, thereby causing situational awareness errors and seriously threatening maritime safety.

[0003] In existing technologies, calibration methods for radar system errors can be mainly categorized as follows:

[0004] (1) Spatial partitioning calibration method: This method divides the radar monitoring area into several sub-regions based on the spatial distribution characteristics of the error, and establishes error correction models for each sub-region. For example, the non-uniform partitioning method proposed by Dong et al. belongs to this category. However, this type of method relies on manually designed partitioning strategies, and the determination of partition boundaries is highly subjective, making it difficult to adapt to the dynamic changes in the marine environment and the individual differences of different radars.

[0005] (2) Multi-station collaborative calibration method: This method utilizes the collaborative observation of the same target by multiple radar stations to solve the system error through geometric constraints. For example, Shang et al. studied the error correction problem of a bi-station fixed two-coordinate radar. However, multi-station systems are costly to deploy, have complex system structures, and have stringent requirements for inter-station time synchronization and data communication links, thus limiting their application scenarios.

[0006] (3) Data fusion and filtering method: Integrating data from multiple sources such as the Automatic Identification System (AIS) for error estimation and correction. For example, Sansot et al. proposed an automatic calibration method based on Kalman filtering, and Jiang et al. introduced a multi-target joint error estimation method based on real-time AIS data. However, methods such as Kalman filtering based on the linear Gaussian assumption have limited ability to handle the complex nonlinear modes existing in radar errors.

[0007] (4) Optimization-based methods: The error calibration problem is modeled as a parameter optimization problem, and the systematic error parameters are solved through iterative search. For example, Li et al. proposed a specific Iterative Closest Point (SICP) algorithm, and Liu et al. developed a calibration method based on an improved sparrow search algorithm. Such methods usually require linearization approximation of the nonlinear error equation or face the problem of low efficiency in searching high-dimensional parameter spaces, which may lead to a loss of accuracy.

[0008] In recent years, deep learning technology has provided new solutions for radar signal processing. However, traditional deep neural network (DNN) architectures, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), are limited by the inherent characteristics of their network structures and cannot fully capture the long-range time dependencies and complex spatiotemporal correlations in radar error sequences, thus encountering a bottleneck in further improving calibration accuracy.

[0009] Therefore, there is an urgent need in this field for a new radar error calibration method that can effectively capture complex nonlinear error modes, has good generalization performance, and does not rely on manual partitioning or complex physical modeling. Summary of the Invention

[0010] The purpose of this invention is to overcome the aforementioned shortcomings of the prior art and provide a radar error calibration method and system based on a vision transformer. This invention utilizes the powerful self-attention mechanism of the radar error calibration model to automatically learn the long-range dependence and nonlinear mapping relationships in the radar error sequence, achieving high-precision prediction and real-time calibration of radar ranging and direction-finding system errors, while providing good model interpretability.

[0011] The technical solution adopted in this invention is as follows:

[0012] A radar error calibration method based on a vision transformer includes the following steps:

[0013] S1: Acquire radar measurement data and AIS data as ground truth; the radar measurement data includes the target's timestamp, measurement distance, and measurement azimuth, and the AIS data is automatic identification system data, including the target's timestamp, geographic coordinates, position accuracy indication information, and navigation status information;

[0014] S2: Preprocess the radar measurement data and AIS data; the preprocessing includes spatial alignment and temporal alignment. Spatial alignment is: transforming the geographic coordinates of the AIS data to a local coordinate system centered on the radar position. Temporal alignment is: generating a synchronized AIS reference position for each radar measurement moment using an interpolation algorithm.

[0015] S3: Calculate the radar error sequence based on the preprocessed data, the radar error sequence including the range error sequence and the azimuth error sequence;

[0016] S4: Construct a radar error calibration model based on a vision transformer architecture. The radar error calibration model includes an embedding layer, a position encoding layer, multiple stacked transformer encoder layers, and an output layer connected in sequence.

[0017] S5: Using the preprocessed radar measurement data as input and the radar error sequence as supervision label, supervised learning training is performed on the radar error calibration model to obtain the trained radar error calibration model.

[0018] S6: Use the trained radar error calibration model to predict the error of the acquired radar measurement data, and subtract the prediction error from the original radar measurement value to obtain the calibrated radar measurement value.

[0019] Furthermore, spatial alignment in S2 specifically refers to:

[0020] Geographic coordinates in AIS data ( , Converted to radar geographic location Local Cartesian coordinates centered on The conversion formula is:

[0021]

[0022]

[0023] in, The radius is the Earth's radius, and all angular parameters are expressed in radians.

[0024] Furthermore, time alignment in S2 specifically involves:

[0025] For each radar measurement time Obtain two consecutive AIS location reports before and after that moment. , , )and( , , ), and satisfy ≤ ≤ The linear interpolation formula was used to calculate the result. Synchronization reference position at time ( , ):

[0026]

[0027]

[0028] In the formula, , For the time, ( , ),( , () represents the location;

[0029] And the interpolation time interval ( Data points exceeding a preset threshold are marked or removed.

[0030] Furthermore, the calculation of the radar error sequence in S3 specifically includes:

[0031] Calculate each radar measurement time using the following formula Corresponding distance error and azimuth error :

[0032]

[0033]

[0034] In the formula, and These are the distance and azimuth values ​​measured by radar, respectively. and These are the actual distance and azimuth values ​​derived from the synchronized reference position obtained by spatial and temporal alignment of AIS data.

[0035] Furthermore, in S4, the embedding layer in the radar error calibration model is used to map the input radar measurement data into a high-dimensional feature vector; the position encoding layer is used to add temporal position information to the feature vector using sinusoidal position encoding; the converter encoder layer stacks several layers, each containing a multi-head self-attention mechanism and a feedforward neural network for capturing long-range dependencies in radar error patterns, and applies layer normalization and residual connections after each sub-layer to output high-level semantic features; the output layer is used to map the high-level semantic features output by the converter encoder layer into the final range error prediction value and azimuth error prediction value.

[0036] The formula for calculating the multi-head self-attention mechanism is:

[0037]

[0038] Where Q is the query matrix, K is the key matrix, and V is the value matrix. The dimension of the key vector is denoted as . The multi-head self-attention mechanism projects the input features to multiple subspaces, calculates attention independently in each subspace, and concatenates the outputs of each subspace before performing a linear transformation to obtain the final output.

[0039] Furthermore, supervised learning training in S5 specifically includes:

[0040] The radar error sequence was divided into training set, validation set and test set according to the proportions.

[0041] Mean squared error is used as the loss function;

[0042] The Adam optimizer is used to iteratively update the parameters of the vision transformer model;

[0043] An early stopping strategy is implemented based on the validation set performance to prevent model overfitting.

[0044] Furthermore, the method also includes:

[0045] The attention weight matrix generated during the forward computation of the trained radar error calibration model is extracted and visualized to identify and analyze feature vectors or temporal positions where the attention weights are higher than a preset threshold, thereby providing interpretability of the model.

[0046] A radar error calibration system based on a vision converter includes:

[0047] The data acquisition module is used to acquire radar measurement data and AIS data as ground truth.

[0048] The data preprocessing module is used to perform spatial and temporal alignment on radar measurement data and AIS data;

[0049] An error calculation module is used to calculate a radar error sequence based on preprocessed data, the radar error sequence including a range error sequence and an azimuth error sequence;

[0050] The model building module is used to build a radar error calibration model based on a vision transformer architecture. The radar error calibration model includes an embedding layer, a position encoding layer, multiple transformer encoder layers, and an output layer. The transformer encoder layer includes a multi-head self-attention mechanism.

[0051] The model training module is used to train the radar error calibration model by taking radar measurement data as input and radar error sequence as training labels.

[0052] The error calibration module is used to predict errors using a trained radar error calibration model and output calibrated radar measurements.

[0053] Furthermore, the data preprocessing module includes:

[0054] Spatial alignment unit is used to transform the geographic coordinates of AIS data to a local Cartesian coordinate system centered on the radar location;

[0055] The time alignment unit is used to generate a synchronized AIS reference position for each radar measurement moment using an interpolation algorithm, and to evaluate and control the interpolation quality.

[0056] Furthermore, the model training module includes:

[0057] The dataset partitioning unit is used to divide the radar error sequence into training set, validation set and test set;

[0058] The loss calculation unit is used to calculate the mean square error loss between the prediction error and the actual error.

[0059] The parameter optimization unit is used to update model parameters using the Adam optimizer and implement an early stopping strategy based on the validation set loss.

[0060] Compared with the prior art, the present invention has the following significant advantages:

[0061] High calibration accuracy: Through the self-attention mechanism of the visual transformer, the long-range dependence and complex nonlinear relationships in radar error patterns can be effectively captured. Experiments show that the method of this invention can reduce the mean absolute range error (MAE) by 98.1% (from 514.76m to 9.93m) and the azimuth MAE by 86.3% (from 1.37° to 0.19°), and the calibration accuracy is significantly better than existing advanced methods such as STDNN, SVR, and XGBoost.

[0062] Strong generalization ability: As a data-driven end-to-end learning method, the model can automatically learn the inherent laws of error from a large amount of historical data without the need for manual division of spatial regions or the establishment of complex physical error models. It exhibits good adaptability and robustness to different types of radar and different sea conditions.

[0063] Low system cost: This invention only requires a single radar and a widely available AIS data source to achieve high-precision calibration, without the need to deploy a multi-station collaborative system, which greatly reduces the hardware cost and engineering complexity of calibration operations.

[0064] Good interpretability: By analyzing the attention weight map generated by the self-attention mechanism, the time-series data segments and features that the model focuses on when making error predictions can be intuitively revealed, providing an effective tool for understanding and diagnosing radar system errors and enhancing the transparency and credibility of the method. Attached Figure Description

[0065] Figure 1 A flowchart of a radar error calibration method based on a vision converter provided in an embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the radar error sequence acquisition process provided in an embodiment of the present invention.

[0067] Figure 3 This is a schematic diagram of a radar error calibration model based on a vision converter architecture constructed for an embodiment of the present invention.

[0068] Figure 4 This is a schematic diagram illustrating the computational principle of the multi-head self-attention mechanism in an embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0070] Example 1

[0071] See Figure 1 This embodiment provides a radar error calibration method based on a vision converter, which specifically includes the following steps:

[0072] S1: Data Acquisition;

[0073] This embodiment uses a publicly available Maritime Target Detection and Tracking (MTDSP) dataset as the data source. This dataset contains approximately 80,000 pairs of radar-AIS synchronous observation data from seven different vessel types. Data acquisition covers a monitoring distance range of 0.5 to 25 nautical miles and an omnidirectional angle of 0 to 360 degrees, providing a sufficient data foundation for a comprehensive evaluation of the algorithm's performance. The data includes timestamps of radar measurements and measurement distances. , measuring azimuth This includes information such as the timestamp of the AIS report, the ship's latitude and longitude coordinates, and the position accuracy indicator factor.

[0074] S2: Data preprocessing;

[0075] like Figure 2 As shown, this step aims to unify radar measurement data and AIS data under the same spatiotemporal reference, and includes the following sub-steps:

[0076] (2.1) Spatial Alignment: The geographic coordinates in the AIS data are transformed to a local Cartesian coordinate system centered on the radar geographic location. A small-scale local planar coordinate transformation method based on a reference point is adopted, and the transformation formula is as follows:

[0077]

[0078]

[0079] in, The geographical coordinates of the radar station, ( , () represents the geographic coordinates in the AIS data. The Earth's average radius is taken as 6371km, and all angles are converted to radians before input.

[0080] (2.2) Time Alignment: Radar sampling frequencies are typically 2-10Hz, while AIS reporting intervals vary from 2 seconds to several minutes. To address the frequency mismatch issue, a linear interpolation algorithm is employed. For each radar measurement moment... Find the two AIS location reports that are immediately before and after it. , , )and( , , The interpolation position is calculated using the following formula:

[0081]

[0082]

[0083] At the same time, the interpolation time interval ( Data points older than 30 seconds are marked. These points will undergo additional verification or have their weight reduced during subsequent error calculations to eliminate interpolation uncertainties caused by long periods of missing AIS data.

[0084] S3: Radar error sequence calculation;

[0085] Based on the aligned data, each timestamp is calculated using the following formula. Corresponding distance error and azimuth error :

[0086]

[0087]

[0088] in, and These are the distance and azimuth values ​​measured by radar, respectively. and It is the true distance and bearing relative to the radar calculated from the interpolated AIS position.

[0089] S4: Construction of a radar error calibration model based on a vision converter architecture;

[0090] See Figure 3 The model constructed in this embodiment is a radar error calibration model based on a vision converter architecture, and its specific configuration is as follows:

[0091] Embedding layer: The input feature dimension is determined based on the radar measurement data, and the output embedding dimension is set to 256, which is used to map the input radar measurement data into a high-dimensional feature vector.

[0092] Position coding layer: Standard sinusoidal position coding is used to add temporal position information to the feature vector.

[0093] Transformer encoder layers: stacked 6 layers. Each transformer encoder layer takes the embedded feature representation as input and includes layer normalization, multi-head self-attention mechanism, residual connection, layer normalization, feedforward neural network and residual connection in sequence to extract the correlation features between different time positions in the input sequence and update the feature representation; among them, each multi-head self-attention mechanism contains 8 attention heads.

[0094] Multi-head self-attention mechanisms: such as Figure 4 As shown, the multi-head self-attention mechanism in this embodiment generates the query matrix Q, key matrix K, and value matrix V through three independent linear transformations. For each attention head, a scaled dot product attention formula is used. ,in This represents the dimension of the key vector. Multiple attention heads perform the above calculations in parallel in different feature subspaces to extract correlation features between different time positions in the input sequence; then, the outputs of each attention head are concatenated and linearly transformed to obtain the output of the multi-head self-attention mechanism.

[0095] Through the aforementioned multi-head self-attention mechanism and feedforward neural network, the model can model the dependencies between different time positions in the radar measurement sequence, thereby more effectively extracting the temporal features related to range error and azimuth error prediction.

[0096] Feedforward neural network: consists of two linear transformation layers with the ReLU activation function in between.

[0097] Residual connections and layer normalization are set after the multi-head self-attention mechanism sub-layer and the feedforward neural network sub-layer, respectively.

[0098] Output Layer: A fully connected output layer with an output dimension of 2. It maps the high-level semantic features output from the transformer encoder layer to the final distance error predictions and orientation error predictions, respectively. and .

[0099] S5: Model training;

[0100] The radar error sequence dataset generated by S3 was randomly divided into training, validation, and test sets in a ratio of 70:15:15. During the division, it was ensured that continuous observation data of the same ship were not split into different sets, in order to verify the model's ability to generalize to unknown targets.

[0101] The training configuration is as follows:

[0102] Loss function: Mean Squared Error (MSE), formula is as follows In the formula, Here, represents the mean squared error loss function value, and N is the number of samples participating in the loss calculation for the current round. For sample index, For the first Predicted distance error for each sample. For the first The true value of the distance error for each sample. For the first Predicted orientation error for each sample For the first The true value of the orientation error for each sample.

[0103] Optimizer: Adam, initial learning rate 0.001, decay factor β1=0.9, β2=0.999.

[0104] Early stopping strategy: If the validation set loss does not decrease for 10 consecutive training epochs, then training is terminated.

[0105] Other parameters: Batch size is set to 32, and maximum number of training cycles is set to 100.

[0106] S6: Real-time error prediction and calibration;

[0107] After the model training is complete and the accuracy requirements are met, it is deployed for real-time radar data calibration. For the raw radar measurement data received in real time (… , First, the same data preprocessing procedure as S2 is performed. Then, the preprocessed sequence data is input into the trained radar error calibration model, and the model outputs the predicted system error. , Finally, the calibrated radar measurements were calculated as follows:

[0108]

[0109] .

[0110] Furthermore, to provide interpretability of the model, this embodiment also includes the following attention weight analysis steps:

[0111] The attention weight matrix generated by the multi-head self-attention mechanism during the forward computation of the trained radar error calibration model is extracted and visualized. Specifically, after softmax normalization, each element of the attention weight matrix represents the contribution weight of the corresponding input position to the current prediction, and satisfies the condition that the sum of all attention weights at the same query position is 1.

[0112] In this embodiment, the preset threshold is determined in the following way: for an input with a sequence length of T, the preset threshold is set to... That is, when the attention weight corresponding to a certain feature vector or temporal position is greater than If the location significantly contributes to the prediction result, then the preset threshold is determined to be the sum of the mean of the attention weights and k times the standard deviation, where k is a positive number, and in this embodiment, k=1. Through the above threshold filtering, input features or time locations that significantly contribute to radar error prediction can be identified and analyzed, thereby providing interpretability of the model.

[0113] Example 2

[0114] This embodiment provides a radar error calibration system based on a vision converter, corresponding to the method described above. The system adopts a modular design and includes:

[0115] Data acquisition module: Equipped with adapters for radar data interface and AIS data interface, supporting the reading of real-time data streams and offline historical data files.

[0116] Data preprocessing module: Includes embedded spatial alignment and temporal alignment units. The spatial alignment unit performs Mercator projection transformation, and the temporal alignment unit performs linear interpolation and quality control.

[0117] Error calculation module: Calculates distance and azimuth error sequences based on preprocessed data, and supports confidence weighting function.

[0118] Model building module: Provides a model hyperparameter configuration interface, allowing users to set the embedding dimension, number of encoder layers, number of attention heads, etc., and build the radar error calibration model accordingly.

[0119] Model training module: This module integrates dataset partitioning, loss function calculation, Adam optimizer, and early stopping strategy control functions to drive the model training process.

[0120] Error calibration module: Loads the trained model parameter file, performs error prediction on the input real-time or batch radar data, and outputs high-precision calibrated results.

[0121] Experimental Results and Analysis:

[0122] To objectively evaluate the effectiveness of the method of this invention, experimental verification was conducted on the MTDSP dataset, and the results were compared with uncalibrated raw data and various baseline methods (STDNN, SVR, XGBoost). The mean absolute error (MAE) of distance and orientation was used as the evaluation metric, and 10-fold cross-validation was employed to ensure the statistical significance of the results.

[0123] The experimental results are shown in the table below:

[0124] method Distance to MAE (m) Azimuth MAE (°) Uncalibrated 514.76 1.37 ST-DNN 18.00 0.30 SVR 25.32 0.45 XGBoost 20.15 0.38 Method of the present invention 9.93 0.19

[0125] As can be seen from the data in the table, the calibration errors of the method of this invention are significantly lower than those of all comparative methods in both the distance and azimuth dimensions. Compared with uncalibrated data, the distance MAE is reduced by 98.1%, and the azimuth MAE is reduced by 86.3%; compared with the best baseline method STDNN, the distance MAE is improved by 44.8%, and the azimuth MAE is improved by 36.7%. The cross-validation results (distance MAE: 10.10±0.32m, azimuth MAE: 0.19±0.01°) further demonstrate the robustness and stability of the method of this invention.

[0126] Industrial applicability:

[0127] This invention can be widely applied to the following technical scenarios:

[0128] Maritime surveillance radar system: used to improve ship tracking accuracy and enhance maritime traffic situational awareness.

[0129] Port Vessel Traffic Management System (VTS): Used to optimize vessel scheduling and improve port operational efficiency and safety.

[0130] Maritime search and rescue and emergency command system: used to improve the accuracy of locating distressed targets and shorten emergency response time.

[0131] Coastal defense and shore-based surveillance system: used to enhance the ability to identify, track and collect evidence of maritime targets.

[0132] Shipborne navigation radar system: used to improve the accuracy and reliability of ship autonomous navigation.

Claims

1. A radar error calibration method based on a vision converter, characterized in that, Includes the following steps: S1: Acquire radar measurement data and AIS data as ground truth; The radar measurement data includes the target's timestamp, measurement distance, and measurement azimuth. The AIS data is automatic identification system data, including the target's timestamp, geographic coordinates, position accuracy indication information, and navigation status information. S2: Preprocess the radar measurement data and AIS data; the preprocessing includes spatial alignment and temporal alignment. Spatial alignment is: transforming the geographic coordinates of the AIS data to a local coordinate system centered on the radar position. Temporal alignment is: generating a synchronized AIS reference position for each radar measurement moment using an interpolation algorithm. S3: Calculate the radar error sequence based on the preprocessed data, the radar error sequence including the range error sequence and the azimuth error sequence; S4: Construct a radar error calibration model based on a vision transformer architecture. The radar error calibration model includes an embedding layer, a position encoding layer, multiple stacked transformer encoder layers, and an output layer connected in sequence. S5: Using the preprocessed radar measurement data as input and the radar error sequence as supervision label, supervised learning training is performed on the radar error calibration model to obtain the trained radar error calibration model. S6: Use the trained radar error calibration model to predict the error of the acquired radar measurement data, and subtract the prediction error from the original radar measurement value to obtain the calibrated radar measurement value.

2. The radar error calibration method based on a vision converter according to claim 1, characterized in that, Spatial alignment in S2 is specifically as follows: Geographic coordinates in AIS data ( , Converted to radar geographic location Local Cartesian coordinates centered on The conversion formula is: ； ； in, The radius is the Earth's radius, and all angular parameters are expressed in radians.

3. The radar error calibration method based on a vision converter according to claim 1, characterized in that, The time alignment in S2 is specifically as follows: For each radar measurement time Obtain two consecutive AIS location reports before and after that moment. , , )and( , , ), and satisfy ≤ ≤ The linear interpolation formula was used to calculate the result. Synchronization reference position at time ( , ): ； ； In the formula, , For the time, ( , ),( , () represents the location; And the interpolation time interval ( Data points exceeding a preset threshold are marked or removed.

4. The radar error calibration method based on a vision converter according to claim 1, characterized in that, The calculation of the radar error sequence in S3 specifically includes: Calculate each radar measurement time using the following formula Corresponding distance error and azimuth error : ； ； In the formula, and These are the distance and azimuth values ​​measured by radar, respectively. and These are the actual distance and azimuth values ​​derived from the synchronized reference position obtained by spatial and temporal alignment of AIS data.

5. The radar error calibration method based on a vision converter according to claim 1, characterized in that, In S4, the radar error calibration model includes an embedding layer that maps the input radar measurement data into a high-dimensional feature vector; a position encoding layer that adds temporal position information to the feature vector using sinusoidal position encoding; a converter encoder layer that stacks several layers, each containing a multi-head self-attention mechanism and a feedforward neural network for capturing long-range dependencies in radar error patterns, and applying layer normalization and residual connections after each sub-layer to output high-level semantic features; and an output layer that maps the high-level semantic features output by the converter encoder layer into the final range error prediction value and azimuth error prediction value. The formula for calculating the multi-head self-attention mechanism is: ； Where Q is the query matrix, K is the key matrix, and V is the value matrix. The dimension of the key vector is denoted as . The multi-head self-attention mechanism projects the input features to multiple subspaces, calculates attention independently in each subspace, and concatenates the outputs of each subspace before performing a linear transformation to obtain the final output.

6. The radar error calibration method based on a vision converter according to claim 1, characterized in that, Supervised learning training in S5 specifically includes: The radar error sequence was divided into training set, validation set and test set according to the proportions. Mean squared error is used as the loss function; The Adam optimizer is used to iteratively update the parameters of the vision transformer model; An early stopping strategy is implemented based on the validation set performance to prevent model overfitting.

7. The radar error calibration method based on a vision transducer according to claim 1, characterized in that, The method further includes: The attention weight matrix generated during the forward computation of the trained radar error calibration model is extracted and visualized to identify and analyze feature vectors or temporal positions where the attention weights are higher than a preset threshold, thereby providing interpretability of the model.

8. A radar error calibration system based on a vision transducer, characterized in that, include: The data acquisition module is used to acquire radar measurement data and AIS data as ground truth. The data preprocessing module is used to perform spatial and temporal alignment on radar measurement data and AIS data; An error calculation module is used to calculate a radar error sequence based on preprocessed data, the radar error sequence including a range error sequence and an azimuth error sequence; The model building module is used to build a radar error calibration model based on a vision transformer architecture. The radar error calibration model includes an embedding layer, a position encoding layer, multiple transformer encoder layers, and an output layer. The transformer encoder layer includes a multi-head self-attention mechanism. The model training module is used to train the radar error calibration model by taking radar measurement data as input and radar error sequence as training labels. The error calibration module is used to predict errors using a trained radar error calibration model and output calibrated radar measurements.

9. The radar error calibration system based on a vision converter according to claim 8, characterized in that, The data preprocessing module includes: Spatial alignment unit is used to transform the geographic coordinates of AIS data to a local Cartesian coordinate system centered on the radar location; The time alignment unit is used to generate a synchronized AIS reference position for each radar measurement moment using an interpolation algorithm, and to evaluate and control the interpolation quality.

10. The radar error calibration system based on a vision converter according to claim 8, characterized in that, The model training module includes: The dataset partitioning unit is used to divide the radar error sequence into training, validation, and test sets. The loss calculation unit is used to calculate the mean square error loss between the prediction error and the actual error. The parameter optimization unit is used to update model parameters using the Adam optimizer and implement an early stopping strategy based on the validation set loss.

Images