A track missing value completion method based on spatio-temporal correlation perception and information alignment

Abstract

This invention provides a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment, comprising the following steps: the acquired track data is sequentially input into the first spatiotemporal information learning module and the Detach layer to obtain maskless spatiotemporal information features F. TF unmasked The acquired trajectory data is manually masked and then input into the second spatiotemporal information learning module. After passing through a linear layer, the masked spatiotemporal information features F are obtained. TF masked The spatiotemporal information feature F TF masked After performing the Replace operation, the track interpolation result X1' is obtained; the track interpolation result X1' is input into the track interpolation reconstruction module to obtain the track reconstruction result X2'; the track interpolation result X1' and the track reconstruction result X2' are input into the interpolation fusion module to obtain the fused interpolation result X3'; and after performing the Replace operation, the final interpolation result X is obtained. c '; Calculate and minimize the spatiotemporal alignment loss L based on the interpolation results. ALIGN Mask interpolation loss L MIT and observation reconstruction loss L ORT This invention significantly improves the accuracy of missing value completion.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment. Background Technology

[0002] Track missing value completion, also known as track interpolation, is an important technique for ensuring the continuity and integrity of trajectories in modern dynamic systems. It eliminates information gaps by dynamically completing discrete observation data, providing high-precision trajectory support for situational awareness, path prediction, and decision support. With the increasing demand for dynamic target tracking in complex environments, track interpolation technology has become a key enabling technology for improving system reliability and intelligence.

[0003] In existing technologies, traditional track interpolation methods mainly rely on linear interpolation, K-nearest neighbors, SAITS, or RNN-based models. These methods are effective in handling simple missing scenarios, but they also have significant limitations. For example, CN 121306586 A discloses a clinical multivariate time series data interpolation method based on generative adversarial networks. The generator module of this method uses the SAITS model, but this model still fails to fully extract spatiotemporal correlation information from the input information and fails to fully integrate the features extracted from mask interpolation and observation reconstruction tasks. This results in significant limitations in complex spatiotemporal patterns, long missing segments, and some situations where high interpolation accuracy is required.

[0004] Therefore, this invention proposes a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment. This method enhances feature consistency through a dual spatiotemporal information learning module and captures spatiotemporal dependencies using a diagonal mask self-attention module. The imputation and reconstruction results are dynamically weighted and fused to retain the original observations, and a joint feature alignment, mask imputation, and observation reconstruction loss optimization model is used.

[0006] To achieve the above objectives, the first aspect of the present invention provides a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment, which uses a missing track value completion model to complete track data;

[0007] The missing track value completion model includes: a first spatiotemporal information learning module, a second spatiotemporal information learning module, a track interpolation and reconstruction module, and an interpolation fusion module;

[0008] The first spatiotemporal information learning module and the second spatiotemporal information learning module use the same encoder, and the attention weight of the second spatiotemporal information learning module is used as the attention weight of the first spatiotemporal information learning module by weight copying.

[0009] The acquired trajectory data is sequentially input into the first spatiotemporal information learning module and the Detach layer to obtain the maskless spatiotemporal information features F. TF unmasked ;

[0010] The acquired trajectory data is manually masked and then input into the second spatiotemporal information learning module. After passing through a linear layer, the masked spatiotemporal information features F are obtained. TF masked ;

[0011] The spatiotemporal information feature F TF masked After performing the Replace operation, the track interpolation result X1' is obtained; the Replace operation is used to convert missing values ​​in the track data into predicted values;

[0012] The track interpolation result X1' is input into the track interpolation reconstruction module to obtain the track reconstruction result X2';

[0013] The track interpolation result X1' and the track reconstruction result X2' are input into the interpolation fusion module to obtain the fused interpolation result X3'.

[0014] After performing a Replace operation on the fused interpolation result X3', the final interpolation result X is obtained. c ';

[0015] When training the missing track value completion model, the following total loss function L is used:

[0016] Calculate the spatiotemporal information feature F TF unmasked and the spatiotemporal information feature F TF masked spatiotemporal alignment loss L ALIGN Its loss function formula is:

[0017] L ALIGN =CosineEmbeddingLoss(F TF masked , F TF unmasked )

[0018] Calculate the interpolation result X c 'Mask interpolation loss L between the actual value X and the true value X' MIT Its loss function formula is:

[0019] L MIT =MAE(X c ',X,I)

[0020] The formula for calculating MAE is:

[0021]

[0022] Where I represents the missing values ​​matrix of the true values, with 0 for missing values ​​X and 1 for non-missing values; D represents the feature dimension; and T represents the time length. It represents the Hadamardi (or Hadama) stack;

[0023] The observation reconstruction loss L is determined based on the track interpolation result X1', the track reconstruction result X2', and the fusion interpolation result X3'. ORT Its formula is:

[0024] L ORT =1 / 3(MAE(X1',X,M)+MAE(X2',X,M)+MAE(X3',X,M))

[0025] The total loss function L is composed of a weighted sum of the individual losses, and its definition is as follows:

[0026] L=L ALIGN +L MIT +L ORT

[0027] The goal of model training is to minimize the total loss function L. The training process continues to iterate until L converges, at which point the model training is complete.

[0028] Based on the trained track missing value completion model, the newly input artificial mask data is sequentially processed by the second spatiotemporal information learning module for feature extraction, the track interpolation and reconstruction module for preliminary completion, and the interpolation fusion module for optimization, and finally the predicted interpolation result is obtained.

[0029] The above scheme, through a dual-path spatiotemporal information learning module and a dynamic fusion mechanism, enables the model to more effectively capture the spatiotemporal correlation and inherent patterns of track data, and significantly improve the accuracy of missing value completion.

[0030] A spatiotemporal alignment loss is introduced. The cosine similarity between the unmasked spatiotemporal information features output by the spatiotemporal information learning module and the masked spatiotemporal information features is calculated, and the loss L is then calculated. ALIGN It can force the model to learn deeper spatiotemporal correlation information, making the spatiotemporal information features extracted by the model when faced with missing values ​​closer to the spatiotemporal information features obtained when using complete track data.

[0031] The Replace operation is applied twice in the interpolation process to ensure that the original, unmissing observation data is not modified during the interpolation process, thus strictly preserving the authenticity and reliability of the data.

[0032] By combining the triple loss function of feature alignment, mask imputation and observation reconstruction, a comprehensive and stable optimization direction is provided for the model, enabling the model to accurately reconstruct known parts while filling in missing values, thus achieving higher learning efficiency.

[0033] In one possible embodiment, the interpolation fusion module includes: a first weighted fusion module and a second weighted fusion module;

[0034] The attention weights of the encoder in the second spatiotemporal information learning module and the attention weights of the encoder in the trajectory interpolation and reconstruction module are respectively input into the first weighted fusion module for weighted fusion. After processing by a linear layer and a Sigmoid activation function layer, the fused weight matrix is ​​obtained.

[0035] The trajectory interpolation result X1', the trajectory reconstruction result X2', and the fused weight matrix are respectively input into the second weighted fusion module for weighted fusion to obtain the fused interpolation result X3'.

[0036] Perform a Replace operation on the fused interpolation result X3' to obtain the final interpolation result X. c '.

[0037] In existing technologies, the weight matrix of the mask matrix and the weight matrix of the diagonal mask self-attention module are fused together. However, this scheme will fuse the output of the diagonal mask self-attention module in the spatiotemporal information learning module with the diagonal mask self-attention module in the track interpolation and reconstruction module. The weight obtained by the fusion of the two has two key capabilities: the inference capability learned from missing data and the reconstruction capability optimized for trajectory integrity, thereby significantly improving the comprehensiveness and discriminative power of weight generation.

[0038] To achieve the above objectives, a second aspect of the present invention provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0039] The memory is used to store computer programs; the processor is used to execute the programs stored in the memory to implement the steps of the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in the first aspect above.

[0040] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the steps of the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in the first aspect above.

[0041] To achieve the above objectives, a fourth aspect of the present invention provides a computer program product that, when executed by a processor, implements the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in the first aspect above.

[0042] The beneficial effects of this invention are as follows:

[0043] Throughout the training process, information spatiotemporal alignment is employed, and prediction loss is supervised by real data. This minimizes the alignment loss, mask interpolation loss, and observation reconstruction loss, avoiding feature distortion or overfitting that may result from optimizing only the task loss, as well as task irrelevance that may result from optimizing only the alignment loss. This achieves a joint improvement in feature quality and task performance. Attached Figure Description

[0044] Figure 1 This is a model structure block diagram provided in the embodiments of the present invention;

[0045] Figure 2 The image shows the trajectory interpolation effect on the test set provided by the embodiments of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0047] refer to Figure 1 This invention provides a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment, comprising:

[0048] During training, two spatiotemporal information learning modules are used for information alignment, enabling the model to learn more spatiotemporal correlation features. The original track data is artificially masked to create gaps, then input into a diagonal mask self-attention module after position encoding. This mechanism effectively captures temporal dependencies and spatial correlations by masking the self-attention at the current time step. Simultaneously, a second path is constructed by replicating parameters, and the information alignment loss L is obtained by calculating the cosine similarity loss using the features of the original and masked data. ALIGN Strengthen the learning of spatiotemporal feature consistency.

[0049] In the interpolation stage, spatiotemporal features are converted into interpolated values ​​through a linear layer, and the original observation data is preserved by a Replace operation. After the track interpolation results are reconstructed by the encoder, the diagonal mask self-attention modules in the spatiotemporal information learning module and the track interpolation and reconstruction module are weighted and fused using a learnable weight matrix. Finally, dynamic fusion weights are generated by Sigmoid activation. The track interpolation results and track reconstruction results are fused using the dynamic fusion weights, and then the original observation data is preserved by a Replace operation to obtain the final track interpolation result.

[0050] Example 1

[0051] This embodiment provides a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment, referencing... Figure 1 The missing value completion model is used to complete the track data;

[0052] The missing track value completion model includes: a first spatiotemporal information learning module, a second spatiotemporal information learning module, a track interpolation and reconstruction module, and an interpolation fusion module;

[0053] The first spatiotemporal information learning module and the second spatiotemporal information learning module use the same encoder, and the attention weight of the second spatiotemporal information learning module is used as the attention weight of the first spatiotemporal information learning module by weight copying.

[0054] The acquired trajectory data is sequentially input into the first spatiotemporal information learning module and the Detach layer to obtain the maskless spatiotemporal information features F. TF unmasked ;

[0055] Specifically, the first spatiotemporal information learning module includes an encoder, which consists of two layers of diagonal mask self-attention modules and a feedforward network; each diagonal mask self-attention module and feedforward network applies layer normalization and residual connections.

[0056] After receiving track data, the first spatiotemporal information learning module adds the data after position encoding. This input is then fed into a diagonal mask self-attention module, and a residual concatenation is performed with the output of the diagonal mask self-attention module. The result of the residual concatenation is then subjected to layer normalization. This normalized result is then passed through a feedforward network and, on the other hand, residually concatenated with the input of the feedforward network. The result of the residual concatenation is then subjected to layer normalization again to obtain the output of the first spatiotemporal information learning module. Since the first and second spatiotemporal information learning modules use the same encoder, their data processing flow is the same, but their input information differs.

[0057] The function of the Detach layer is to separate the unmasked spatiotemporal information features of the output from the current computation graph. The computation graph refers to all the computational steps from the input matrix to the output matrix. It can be understood that after the Detach operation, the gradient of the data matrix will not be propagated back to the upstream module that generated the data matrix during the subsequent backpropagation process.

[0058] The acquired trajectory data is manually masked and then input into the second spatiotemporal information learning module. After passing through a linear layer, the masked spatiotemporal information features F are obtained. TF masked ;

[0059] The spatiotemporal information feature F TF masked After performing the Replace operation, the track interpolation result X1' is obtained; the Replace operation is used to convert missing values ​​in the track data into predicted values;

[0060] The track interpolation result X1' is input into the track interpolation reconstruction module to obtain the track reconstruction result X2';

[0061] Specifically, the encoder of the track interpolation and reconstruction module adopts a diagonal mask self-attention module;

[0062] After adding the track interpolation result X1' to the position encoding information, the result is input into the encoder. The encoder output is then input into the linear layer, the ReLU activation function layer, and the linear layer in sequence to obtain the track reconstruction result X2'.

[0063] The track interpolation result X1' and the track reconstruction result X2' are input into the interpolation fusion module to obtain the fused interpolation result X3'.

[0064] Specifically, the interpolation fusion module includes: a first weighted fusion module and a second weighted fusion module;

[0065] The attention weights of the encoder in the second spatiotemporal information learning module and the attention weights of the encoder in the trajectory interpolation and reconstruction module are respectively input into the first weighted fusion module for weighted fusion. After processing by a linear layer and a Sigmoid activation function layer, the fused weight matrix is ​​obtained.

[0066] The trajectory interpolation result X1', the trajectory reconstruction result X2', and the fused weight matrix are respectively input into the second weighted fusion module for weighted fusion to obtain the fused interpolation result X3'.

[0067] Perform a Replace operation on the fused interpolation result X3' to obtain the final interpolation result X. c '.

[0068] It should be noted that the encoders of both the first and second spatiotemporal information learning modules use diagonal mask self-attention modules.

[0069] When training the missing track value completion model, the following total loss function L is used:

[0070] Calculate the spatiotemporal information feature F TF unmasked and the spatiotemporal information feature F TF masked spatiotemporal alignment loss L ALIGN Its loss function formula is:

[0071] L ALIGN =CosineEmbeddingLoss(F TF masked , F TF unmasked )

[0072] Calculate the interpolation result X c 'Mask interpolation loss L between the actual value X and the true value X' MIT Its loss function formula is:

[0073] L MIT =MAE(X c ',X,I)

[0074] The formula for calculating MAE is:

[0075]

[0076] Where I represents the missing values ​​matrix of the true values, with 0 for missing values ​​X and 1 for non-missing values; D represents the feature dimension; and T represents the time length. It represents the Hadamardi (or Hadama) stack;

[0077] The observation reconstruction loss L is determined based on the track interpolation result X1', the track reconstruction result X2', and the fusion interpolation result X3'. ORT Its formula is:

[0078] L ORT =1 / 3(MAE(X1',X,M)+MAE(X2',X,M)+MAE(X3',X,M))

[0079] The total loss function L is composed of a weighted sum of the individual losses, and its definition is as follows:

[0080] L=L ALIGN +L MIT +L ORT

[0081] It should be noted that two spatiotemporal information learning modules are used during training for information alignment, enabling the model to learn more spatiotemporally relevant features. These two modules are only activated simultaneously during training; during inference, since L is not required to be calculated... ALIGN Therefore, the spatiotemporal information learning module on the left side of the above figure is disabled.

[0082] The goal of model training is to minimize the total loss function L. The training process continues to iterate until L converges, at which point the model training is complete.

[0083] Based on the trained track missing value completion model, the newly input artificial mask data is sequentially processed by the second spatiotemporal information learning module for feature extraction, the track interpolation and reconstruction module for preliminary completion, and the interpolation fusion module for optimization, and finally the predicted interpolation result is obtained.

[0084] To verify the effectiveness of the method provided in this embodiment, a validation experiment was conducted on a civil aviation and shipbuilding dataset, containing 35,244 ADS-B data entries and 32,847 AIS data entries. 80% of the data was used for training, 10% was used as the validation set, and the final 10% was used as the test set. (Reference) Figure 2 The image shows the trajectory interpolation effect on the test set. In the experiment, Adam was selected as the optimizer, the learning rate was set to 0.001, and the number of training epochs was 150. The accuracy of linear interpolation, K-nearest neighbors, SAITS, and the method proposed in this embodiment were verified and compared under the conditions of 10% missing, 20% missing, and 30% missing, and the results are shown in the table below.

[0085] Table 1. Accuracy comparison of various methods when trajectory data is 10% missing.

[0086] Linear interpolation 0.00793 0.00674 7.83 K-Nearest Neighbors (KNN) 0.00665 0.00546 6.45 SAITS 0.00486 0.00427 4.47 Ours 0.00354 0.00391 2.93

[0087] Table 2. Accuracy comparison of various methods when 20% of trajectory data is missing.

[0088] Linear interpolation 0.00815 0.00697 8.06 K-Nearest Neighbors 0.00684 0.00563 6.68 SAITS 0.00498 0.00439 4.59 Ours 0.00377 0.00412 3.17

[0089] Table 3. Accuracy comparison of various methods when trajectory data is 30% missing.

[0090] Linear interpolation 0.00856 0.00728 8.49 K-Nearest Neighbors (KNN) 0.00727 0.00594 6.92 SAITS 0.00524 0.00458 4.76 Ours 0.00436 0.00445 3.45

[0091] It should be noted that RMSE represents root mean square error, MAE represents mean absolute error, and MAPE represents mean absolute percentage error.

[0092] Based on the test results of the three sets above, it can be seen that as the data missing rate increases from 10% to 30%, the error indicators of the four methods all show a slight and steady upward trend. However, the proposed method consistently outperforms the linear interpolation, KNN, and SAITS methods. Specifically, the RMSE of the proposed method remains between 0.00354 and 0.00436, the MAE between 0.00391 and 0.00445, and the MAPE between 2.93% and 3.45%. Among the comparative methods, linear interpolation has the largest error, followed by KNN. SAITS performs well but is still inferior to the proposed method. This fully demonstrates that the proposed method has good stability and reliability under different levels of data missing scenarios and has stronger adaptability.

[0093] The effect of using the proposed track missing value completion method based on spatiotemporal correlation perception and information alignment to complete the track is shown in the figure below. Figure 2 As shown, it can be seen that the method proposed in this embodiment can achieve good interpolation results for both aircraft with relatively high speeds and ships with relatively gentle movements, and the interpolated track has a high degree of overlap with the real track.

[0094] Example 2

[0095] This application embodiment discloses a computer device. The computer device includes a processor and a memory. The processor, memory, and input / output interface are connected via a system bus, and a communication interface, display unit, and input device are connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides steps for the operation of the operating system and computer programs in the non-volatile storage medium using a method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment.

[0096] Example 3

[0097] Based on the above embodiments, this embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of track missing value completion based on spatiotemporal correlation perception and information alignment as described in Embodiment 1.

[0098] Example 4

[0099] Based on the above embodiments, this embodiment provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in any of the embodiments in 1.

[0100] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment, characterized in that, The missing value completion model is used to complete the track data; The missing track value completion model includes: a first spatiotemporal information learning module, a second spatiotemporal information learning module, a track interpolation and reconstruction module, and an interpolation fusion module; The first spatiotemporal information learning module and the second spatiotemporal information learning module use the same encoder, and the attention weight of the second spatiotemporal information learning module is used as the attention weight of the first spatiotemporal information learning module by weight copying. The acquired trajectory data is sequentially input into the first spatiotemporal information learning module and the Detach layer to obtain the maskless spatiotemporal information features F. TF unmasked ; The acquired trajectory data is manually masked and then input into the second spatiotemporal information learning module. After passing through a linear layer, the masked spatiotemporal information features F are obtained. TF masked ; The spatiotemporal information feature F TF masked After performing the Replace operation, the track interpolation result X1' is obtained; the Replace operation is used to convert missing values ​​in the track data into predicted values; The track interpolation result X1' is input into the track interpolation reconstruction module to obtain the track reconstruction result X2'; The track interpolation result X1' and the track reconstruction result X2' are input into the interpolation fusion module to obtain the fused interpolation result X3'. After performing a Replace operation on the fused interpolation result X3', the final interpolation result X is obtained. c '; When training the missing track value completion model, the following total loss function L is used: Calculate the spatiotemporal information feature F TF unmasked and the spatiotemporal information feature F TF masked spatiotemporal alignment loss L ALIGN Its loss function formula is: L ALIGN =CosineEmbeddingLoss(F TF masked , F TF unmasked ) Calculate the interpolation result X c 'Mask interpolation loss L between the actual value X and the true value X' MIT Its loss function formula is: L MIT =MAE(X c ',X,I) The formula for calculating MAE is: Where I represents the missing values ​​matrix of the true values, with 0 for missing values ​​X and 1 for non-missing values; D represents the feature dimension; and T represents the time length. It represents the Hadamardi (or Hadama) stack; The observation reconstruction loss L is determined based on the track interpolation result X1', the track reconstruction result X2', and the fusion interpolation result X3'. ORT Its formula is: L ORT =1 / 3(MAE(X1',X,M)+MAE(X2',X,M)+MAE(X3',X,M)) The total loss function L is composed of a weighted sum of the individual losses, and its definition is as follows: L=L ALIGN +L MIT +L ORT The goal of model training is to minimize the total loss function L. The training process continues to iterate until L converges, at which point the model training is complete. Based on the trained track missing value completion model, the newly input artificial mask data is sequentially processed by the second spatiotemporal information learning module for feature extraction, the track interpolation and reconstruction module for preliminary completion, and the interpolation fusion module for optimization, and finally the predicted interpolation result is obtained.

2. The method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment according to claim 1, characterized in that, The interpolation fusion module includes: a first weighted fusion module and a second weighted fusion module; The attention weights of the encoder in the second spatiotemporal information learning module and the attention weights of the encoder in the trajectory interpolation and reconstruction module are respectively input into the first weighted fusion module for weighted fusion. After processing by a linear layer and a Sigmoid activation function layer, the fused weight matrix is ​​obtained. The trajectory interpolation result X1', the trajectory reconstruction result X2', and the fused weight matrix are respectively input into the second weighted fusion module for weighted fusion to obtain the fused interpolation result X3'. Perform a Replace operation on the fused interpolation result X3' to obtain the final interpolation result X. c '.

3. A method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment according to claim 1 or 2, characterized in that, The encoders of both the first and second spatiotemporal information learning modules employ diagonal mask self-attention modules.

4. A method for completing missing track values ​​based on spatiotemporal correlation perception and information alignment according to claim 1 or 2, characterized in that, The encoder of the track interpolation and reconstruction module adopts a diagonal mask self-attention module; After adding the track interpolation result X1' to the position encoding information, the result is input into the encoder. The encoder output is then input into the linear layer, the ReLU activation function layer, and the linear layer in sequence to obtain the track reconstruction result X2'.

5. A computer device, characterized in that: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; the processor is used to execute the programs stored in the memory to implement the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in any one of claims 1 to 4.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the track missing value completion method based on spatiotemporal correlation perception and information alignment as described in any one of claims 1 to 4.

Images