Deep learning network-based track hierarchical classification method

Abstract

The invention relates to a track classification and recognition method in sea area situation awareness, in particular to a track hierarchical classification method based on a deep learning network, which comprises the following steps: firstly, converting a ship track into an image layer, and constructing a track image layer classification model based on a Swin-Transform network; secondly, for a track sequence layer, optimizing input of a track sequence based on a track compression algorithm of multi-dimensional information, and constructing a track sequence layer classification model based on a Gained-Transform-Network deep learning network; and finally, a fusion layer track classification model based on credibility is established, and the accuracy of track hierarchical classification is improved. Experiments prove that the average classification accuracy is 91.7%, the classification performance of the fusion classifier is averagely improved by 11% compared with other single classifiers, and the F1 score is averagely 0.82. The result shows that the newly proposed method and the new fusion classifier have a good classification effect on the ship track.

Description

Technical Field

[0001] The present invention relates to a track classification and identification method in maritime situation awareness, and in particular to a track hierarchical classification method based on a deep learning network. Background Art

[0002] As an important part of key target tracking and maritime situation awareness, the research on track classification and identification has received widespread attention. Track classification is an important research field. Track classification helps to improve maritime ship supervision and analyze whether ships have abnormal and suspicious behaviors. Analyzing historical tracks can reveal the behavior patterns of ships, improve the efficiency and accuracy of automatic identification and classification, and help predict future behaviors, thereby enhancing overall situation awareness and early warning capabilities. Fishing vessel track classification can strengthen the supervision of fishing behavior, and merchant ship track classification can further analyze and identify ship behavior and strengthen ship monitoring. In addition, the historical track information of ships contains rich information on ship activities. The classification of maritime target tracks can increase the perception of the maritime environment, improve the recognition and tracking accuracy of key targets, analyze the behavioral intentions of targets, improve battlefield situation estimation, and improve navigation safety in the sea.

[0003] At present, the specific application scenarios of ship trajectory classification methods at home and abroad mainly include the identification of ship types and the classification of ship motion patterns. The implementation process is divided into three parts: ship trajectory data cleaning preprocessing, ship track feature extraction and track classification model modeling. The track information volume is large and the geographical range is wide. The rich information volume also brings a lot of redundancy, which brings challenges to ship classification and identification. The track classification methods mainly include: 1) Clustering method based on the distance between tracks. Chen's AIS ship trajectory classification based on sparse representation classification algorithm improves the traditional classification model based on Lp-norm sparse representation to classify ship motion patterns. The proposed model can effectively classify ship motion patterns in inland waterways. Wang proposed an inland waterway ship trajectory classification algorithm based on width learning system (BLS). According to the selection of specific experimental data and experimental areas, the segmented cubic Hermite interpolation method is used to extract feature points from two angles of equal time interval or equal spacing to construct a trajectory feature matrix. The trajectory feature matrix is ​​substituted into the trained BLS to complete the automatic classification of ship trajectories. Tang improved FOP-OPTICS based on OPTICS to find ordered peaks to cluster sub-trajectory segments, which has obvious advantages in identifying the overall and local characteristics of ship trajectories, noise identification, and computational efficiency. 2) Classification method combining artificially extracted features and machine learning methods. Xu used the random forest algorithm to classify ship types by extracting the spatial geographical characteristics and target motion characteristics of ship trajectories. 3) Deep learning network classification method: mainly includes classification networks based on ship track maps and sea environment maps, and classification based on the temporal characteristics of ship trajectories combined with neural networks. Chen

[12] A new CNN-SMMC architecture is proposed to classify ships. Some scholars have combined LSTM and CNN to establish a hybrid classification model characterized by speed, acceleration, heading and curvature, using CNN to extract the spatial features of trajectory data and LSTM to extract the temporal features of trajectory data. Duan proposed a semi-supervised deep learning method to classify ship trajectories by integrating the knowledge of unlabeled data. The longitude and latitude data of ship trajectories such as GUO are converted into ship trajectory image data, and a deep convolutional neural network ship trajectory classification model is constructed using the ResNet residual network.

[0004] However, at present, most studies focus on the motion status of ships in fixed sea areas, while there are fewer studies on the classification of tracks with spatial differences and differences in motion characteristics. In addition, track classification research mainly converts tracks into images for classification, and then runs machine learning for modeling to judge new tracks. However, this method does not take into account the changing characteristics of track data sequences, resulting in insufficient utilization and analysis of track data. At the same time, machine learning has also achieved good results in this research field. It can model based on a large amount of historical data and then make judgments on new data. Compared with some methods that use specific algorithms to identify tracks, machine learning methods are based on a large amount of historical data for modeling, thereby improving the robustness of the model. Summary of the invention

[0005] In view of the problem that the existing track classification method cannot fully consider the time series characteristics and spatial structure characteristics of the track, resulting in a decrease in classification accuracy, the present invention proposes a track hierarchical classification method based on a deep learning network to classify the ship category to which the ship track belongs.

[0006] The present invention is implemented by adopting the following technical scheme: a track hierarchical classification method based on a deep learning network, firstly, the track image data is established through the track sequence, and the track image data is input into the trained Swin-Transformer model for classification and recognition to obtain the classification evidence body E 1 ; Secondly, the improved DP algorithm is used to compress the track sequence data, and the compressed track sequence data is input into the trained GTN network for classification and recognition to obtain the classification evidence body E 2 Finally, the improved DS evidence theory is used to classify the above two types of evidence bodies E 1 and E 2 Fusion processing is used to classify and identify the ship category to which the ship track belongs.

[0007] The process of establishing the track image data is as follows: mapping the longitude and latitude coordinates of the track points to the image pixel coordinates according to the scale, extracting the spatial features of the ship track data, and establishing the ship track image data set.

[0008] The process of compressing the track sequence data using the improved DP algorithm is as follows: calculate the distance between all track points in the middle of the starting point and the line AB connecting the starting point, and find the maximum value d max And the corresponding track points Calculate the distance between all track points in the starting point and the track points before and after it, calculate the average speed in the time domain based on the timestamp difference, and find the maximum speed V max And the corresponding track points Calculate the difference between the heading angles of all track points between the starting point and the heading angles before and after, and find the maximum difference H max And the corresponding track points Set the distance threshold d respectively th , speed threshold V th , heading angle threshold H th If any of the maximum distance, maximum speed, and maximum heading angle difference is greater than or equal to the corresponding threshold, the track is divided into two segments at the point corresponding to the maximum value for iterative track compression; if the maximum distance, maximum speed, and maximum heading angle difference are all less than the threshold, some intermediate track points of the track are discarded according to the compression ratio.

[0009] The specific process of the DS evidence theory to fuse the classified evidence body is as follows: i Self-certainty In the formula, M represents the number of events corresponding to the focal element A in the identification framework, i represents the i-th piece of evidence in the identification framework, and m i (A j ) represents the probability that the i-th piece of evidence is classified as the j-th category; by calculating the self-certainty of the two pieces of evidence, we can get two self-certainty values, which are expressed in vector form, and the self-certainty vector of the evidence body C = [c 1 ,c 2 ], the classification accuracy of the two classification models is used as the confidence of each evidence body, and the product of the self-certainty of the evidence body and the classification accuracy rate is used to obtain the relative credibility vector Rel = [Rel 1 ,Rel 2 ], and then judge the relative credibility vector, in represents the credibility threshold, which is expressed as Then the weight factor vector is ω=[ω 1 ,ω 2 ,...,ω N ], and normalized, Calculate the evidence body BPA function m′, Finally, the modified BPA function is subjected to one DS fusion to obtain the fusion result m and determine the output target category.

[0010] The track sequence data is preprocessed before being input into the two classification models, the Swin-Transformer model and the GTN network. The preprocessing includes eliminating duplicate points and track interpolation.

[0011] The method of the present invention first converts the ship track into an image layer, and constructs a track image layer classification model based on the Swin-Transformer model; secondly, for the track sequence layer, the track compression algorithm based on multi-dimensional information optimizes the input of the track sequence data, and constructs a track sequence layer classification model based on the Gained-Transformer-Network deep learning network. Finally, a fusion layer track classification model based on credibility is established to improve the accuracy of track hierarchical classification. After experimental verification, the average classification accuracy is 91.7%, and the classification performance of the fusion classifier is improved by an average of 11% compared with other single classifiers, and the average F1 score is 0.82. The results show that the newly proposed method and the new integrated classifier have a good classification effect on ship tracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a framework diagram of the method of the present invention.

[0013] Figure 2 This is the Swin-Transformer system diagram.

[0014] Figure 3 This is the structural diagram of the improved DP compression algorithm.

[0015] Figure 4 Schematic diagram of the track before and after DP algorithm compression.

[0016] Figure 5 This is the GTN network structure diagram.

[0017] Figure 6 It is the image layer Loss convergence curve.

[0018] Figure 7 This is the accuracy convergence curve.

[0019] Figure 8 Schematic diagram of the F1 score for image-level track data classification.

[0020] Fig. 9 It is the sequence layer loss convergence curve.

[0021] Fig.10 This is the convergence curve of the accuracy of the training set and the validation set.

[0022] Fig.11 Schematic diagram of the F1 score for classification of sequence-level track data.

[0023] Fig.12 Schematic diagram of the F1 score for classification of fusion layer track data.

[0024] Fig.13 Schematic diagram of the F1 score for classification of hierarchical track data.

[0025] Fig.14 This is the fusion result diagram. DETAILED DESCRIPTION

[0026] A track hierarchical classification method based on a deep learning network is used to classify track data with large temporal and spatial differences. In order to fully study the time series characteristics and spatial structure characteristics of the track, the present invention establishes a hierarchical classification model, starting from information from multiple angles, to improve the accuracy and robustness of track data classification and enhance the computational efficiency of the model. First, the abnormal data points of the AIS track are excluded, and the AIS track data is preprocessed, including eliminating duplicate points and track interpolation. The hierarchical classification model is divided into three layers. 1) Image layer: The image data of the track is established through the track sequence. According to the Swin-Transformer deep learning algorithm, the track classification model is trained using the swinv2 pre-trained model for classification and recognition, effectively capturing the spatial structural characteristics of the track. This method enables the model to recognize complex patterns and details; 2) Sequence layer: By processing the original track sequence data, for the redundant track sequence, the track is compressed using the improved DP algorithm, the sequence input of the GTN network is refined, and the track sequence is classified and recognized using the GTN deep learning network, so as to better capture the dynamic changes and trends in the time series and reduce the interference of redundant information on the classification results; 3) Fusion layer: By establishing a fusion layer based on credibility theory to fuse the results of multiple classifications, it is helpful to combine the advantages of different features and improve the accuracy and reliability of the final classification results.

[0027] Track image layer classification based on Swin-Transformer model

[0028] The Swin-Transformer model is an image classification model based on the Transformer model. Figure 2 As shown, Figure 2 (a) is the structure of the Swin-Transformer model. First, the input track image is divided into non-overlapping image blocks. The image features are projected to any dimension through a linear embedding layer. Then, the Swin-Transformer blocks are arranged in sequence and the block merging layer is used to form a stage layer. In order to produce a hierarchical representation, the number of image blocks in the merging layer will decrease as the network stage goes deeper. H is the height of the original track image, W is the width of the original track image, and C is the custom feature representation latitude. Figure 2(b) Two consecutive Swin-Transformer blocks. A Swin-Transforme module consists of a moving window-based self-attention module (MSA) module, followed by a two-layer multilayer perceptron (MLP) with GELU nonlinearity in between. A LayerNorm (LN) layer is applied before each MSA module and each MLP, and the residual connection is applied after each block. The modular structure and scalability of Swin-Transformer enable it to flexibly adapt to different types and sizes of track image data.

[0029] Track compression algorithm based on multi-dimensional information

[0030] The track data is redundant and the length of track sequence samples is inconsistent. In order to optimize the input of the GTN network, the present invention uses an improved DP algorithm to compress the track. The traditional DP algorithm only considers the distance between the track point and the starting and ending points, and ignores other dimensional information in the track data. The present invention adds the speed and heading angle difference to the threshold dimension of track compression, and sets the compression ratio to optimize the track sequence data. The compression ratio is: the ratio of the number of track points before and after compression N:n. The structure of the improved DP algorithm is as follows: Figure 3 The settings of various threshold parameters are shown in Table 1.

[0031] After the preprocessed compressed track, calculate the distance between all track points in the middle of the starting point and the line AB connecting the starting point, and find the maximum value d max And the corresponding track points Calculate the distance between all track points in the starting point and the track points before and after it, calculate the average speed in the time domain based on the timestamp difference, and find the maximum speed V max And the corresponding track points Calculate the difference between the heading angles of all track points between the starting point and the heading angles before and after, and find the maximum difference H max And the corresponding track points Set the distance threshold d respectively th , speed threshold V th , heading angle threshold H th If any of the three-dimensional information of the maximum distance, the maximum speed, and the maximum heading angle difference is greater than or equal to the corresponding threshold, the track to be compressed is divided into two segments corresponding to the maximum value for iterative track compression. If the three-dimensional information is less than the threshold, some intermediate track points of the track are discarded according to the compression ratio.

[0032] Table 1 DP algorithm threshold settings

[0033]

[0034] DP compression algorithm such as Figure 3As shown, the trajectory before compression is Figure 4 As shown in (a), the number of track points is 36082, and the compressed track diagram is as follows Figure 4 As shown in (b), (c), and (d), the number of track points is 1697, 2702, and 3685, respectively, and the compression ratios are 21:1, 13:1, and 9.7:1, respectively. Figure 4 It can be seen that the improved DP algorithm can effectively compress the number of track points to the expected number of track points, retain the original track characteristics, and further optimize the input of the subsequent deep learning network.

[0035] Track classification network structure based on GTN network

[0036] GTN (Gated-Transformer-Network) is a multi-dimensional time series classification network based on Transformer. The GTN network structure is as follows Figure 5 As shown in the figure, track sequences often involve complex relationships in time and space. GTN uses the attention mechanism and the self-attention mechanism of Transformer to effectively capture long-distance dependencies in the sequence without being restricted by the sequence length, thereby better modeling the spatiotemporal dynamic characteristics of the track sequence. The GTN network can better capture the channel characteristics of track sequence information in multiple dimensions. Track sequences may contain features at different scales and time scales. The GTN network can effectively integrate and utilize feature information at different scales through multi-head attention mechanism and hierarchical representation learning to improve classification accuracy.

[0037] Classification of track fusion layers based on credibility

[0038] DS evidence theory is integrated with the reasoning method used to deal with uncertainty and incomplete information. The image layer and sequence layer of the track belong to different dimensions, and different layers themselves have uncertainties. By assigning basic probability distribution functions BPA to different layers and comprehensively considering the degree of conflict between them, the comprehensive trust level of the track category can be obtained, thereby improving the accuracy of track classification.

[0039] The degree of certainty quantifies the characteristics of the evidence body through certain rules, reflecting the degree of aggregation of the probability distribution function and the distribution recognition of the evidence body for its own BPA. The higher the degree of certainty, the higher the degree of aggregation and distribution recognition. i The degree of self-certainty c i It can be calculated according to the following formula:

[0040]

[0041] Among them, M represents the number of events corresponding to the focal element A in the identification frame Θ, i represents the i-th piece of evidence in the identification frame, and mi (A j ) represents the probability that the i-th piece of evidence is classified as the j-th category.

[0042] According to the above formula, the self-certainty of N (N=2) evidence bodies can be calculated respectively, and N self-certainty values ​​can be obtained, and expressed in vector form to obtain the self-certainty vector C of the evidence body of the system, as follows:

[0043] C=[c 1 ,c 2 ,...,c N ]

[0044] The present invention uses the training classification accuracy of the classification model as the confidence of the evidence body. Since the GTN network is prone to fall into the local optimum, its own support is retained at the original 30%.

[0045] By calculation, we get the relative credibility vector Rel = [Rel 1 ,Rel 2 ,...,Rel N ], in order to further reduce the impact of evidence with poor credibility on fusion, the relative credibility vector is judged as shown in Eq.

[0046]

[0047] in represents the credibility threshold, which is expressed as

[0048]

[0049] The modified weight factor vector is ω=[ω 1 ,ω 2 ,...,ω N ] and normalized them.

[0050]

[0051] In the identification framework of multiple evidence bodies, credibility is used as the correction coefficient of the evidence body, and the BPA of the corresponding evidence body is weighted averaged to obtain the corrected evidence BPA function m′.

[0052]

[0053] Then, the modified BPA function is subjected to one DS fusion to obtain the fusion result m and determine the output target category.

[0054] According to the AIS track data set, the present invention conducted experimental research on track image data classification based on Swin-Transformer and track sequence data classification based on GTN network, and analyzed the results to verify the effect of the classification model.

[0055] The data set of the present invention is derived from the track classification subject-specific data set published in the Journal of Radar in the "Golden Dolphin" Cup competition. The data set is prepared by collating global open source AIS data. The data set contains two major categories of track data, which are subdivided into 14 small categories of data sets. Each track contains a varying number of track points, which include timestamp, longitude, latitude, speed, and heading angle information.

[0056] The present invention selects 5 types of data from the training data set for research. As shown in Table 2, after the original track is preprocessed, the samples with a large time domain span are first segmented. Secondly, in order not to affect the time domain spatial structure of the track, the preprocessed track is compressed at different compression ratios using an improved DP algorithm to optimize the data sample.

[0057] Table 2 Dataset structure and sample size after preprocessing

[0058]

[0059] The Swin-Transformer model is used to classify the image data of the track, and the five types of track data in the training data set are used for classification, and the results are analyzed. The tracks are classified based on the track image data. Since the track information used does not belong to the same area, the longitude and latitude are quite different. In order to avoid the influence of the background island image information on the track classification training, the present invention uses the Mercator projection based on the latitude and longitude values ​​of the track information to calculate the projection point coordinates of the longitude and latitude in the plane coordinates. When making the image data set, the island information of the original image background is discarded, and only the track point and line map is retained. The resolution of each ship track image is 224*224. The total number of track image data sets is 989, of which the training set, validation set, and test set account for 0.8, 0.1, and 0.1 respectively.

[0060] The experiment is a five-category classification based on the Swin-Transformer network. The input layer size parameter of the model is set to 224*224*3, the mini-batch size is set to 64, and the output layer categories are set to 5. The swinv2-tiny pre-trained model is used, the epoch is set to 300, and the maximum learning rate is 1.25e -3 , the minimum learning rate is 1.25e -5 The training set LOSS convergence curve is as follows: Figure 6 As shown in (a), the LOSS convergence curve of the test set is as follows Figure 6(b) is shown. The accuracy convergence curve is shown in Figure 7 As shown in (a), the maximum accuracy convergence curve is as follows Figure 7 (b) as shown.

[0061] We performed convergence tests to find a good baseline learning rate for training models for image-level track data classification. The figure shows the results of 200 training cycles at different learning rates for a model using pre-trained swin-v2 weights. In this experiment, we used a constant learning rate schedule. -3 and 1e -2 At a learning rate of , Swin-Transformer quickly reached a verification accuracy of more than 80%, with a convergence rate of around 0.75, and the final verification set accuracy stabilized at 87%.

[0062] Five types of track information were selected as experimental data. To verify the effectiveness of the model, the present invention set up three groups of experiments. Each group randomly selected 50 track data from the five types of data as the data set for this experiment. The F1 (classification regression evaluation index) score and mean of each category are shown in Table 3. Figure 8 It can be seen that the total mean F1 score of image layer track data classification is 0.6492, which achieves a good F1 score effect at an accuracy of 87% in the validation set, verifying the effectiveness of image layer track data classification based on the Swin-Transformer network.

[0063] Table 3 F1 score of image layer track data classification

[0064]

[0065] The GTN network is used to classify the sequence data of the track. The experimental data selects five types of track information in the training data set, and the results are analyzed. The track sequence data training retains all dimensional information in the track data. Since the number of track points in the track sequence data in the data set is different and varies greatly, the present invention uses an improved DP algorithm to compress the number of track points in the track sequence to 1500. For track sequences with less than 1500 track points, the cubic spline method is used to interpolate the track data to optimize the input of the Transformer network. The training set, validation set, and test set in the track sequence data set account for 0.5, 0.3, and 0.2 respectively.

[0066] The experiment is divided into five categories. Based on the GTN network, the mini-batch size is set to 4, and the output sequence categories are set to 5. The epoch is set to 100 and the maximum learning rate is 1.25e -4 , the minimum learning rate is 1.25e -5 . Fig. 9The figure shows the loss convergence curve of this experimental training, and the loss converges to around 0.53. Fig.10 The results of the deep learning experiment are shown, with the highest accuracy of 89.56% on the training set and 80.2% on the test set.

[0067] Fig.10 Results are shown for 100 training epochs at different learning rates using the pre-trained model. In this experiment, a constant learning rate schedule was used. -3 and 1e -2 At the learning rate of , the verification accuracy of more than 70% was quickly reached, the convergence rate was around 0.75, and the final verification set accuracy was stabilized at 78%. Due to the overfitting phenomenon of the GTN network, the confidence ratio based on the GTN network should be reduced compared to that based on the Swin-Transformer network in the subsequent fusion.

[0068] The experimental data selected the same five types of track information as the test data set. To verify the effectiveness of the model, the present invention sets up three groups of experiments, each of which randomly selects 50 track data from the five types of data as the data set for this experiment. The F1 score and mean of each category are shown in Table 4. Fig.11 It can be seen that the total mean F1 score of the sequence-level track data classification is 0.6, and a good F1 score effect is achieved at an accuracy of 78% in the validation set, which verifies the effectiveness of the sequence-level track data classification based on the GTN network.

[0069] Table 4 F1 score of sequence layer track data classification

[0070]

[0071] By randomly selecting multiple test sets to verify the fusion experiment based on DS evidence theory, the classification results based on the Swin-Transformer model and the GTN network model were fused. The experimental results show that the fusion effectively improves the accuracy of the classification results. The experimental data selected the same five categories of track information in the test data set. To verify the effectiveness of the model, 50 track data were randomly selected from each of the 5 categories of data as the data set for this experiment. To verify the effectiveness of the algorithm, the fusion experiment randomly selected 3 groups of data from the 5 categories of track data as the samples of the experiment. The F1 score of the fusion layer track data classification is shown in Table 5. The macro F1 score reached an excellent score of 0.82. Fig.12 Shows a clearer and more intuitive observation of the F1 scores of different classes represented by the fusion layer. Fig.13The comparison of F1 scores of image layer, sequence layer and fusion layer is shown. In the five types of tasks, the fusion layer shows excellent F1 scores, which proves that the fusion layer track data classification improves the reliability of the classification algorithm based on the image layer and sequence layer track classification.

[0072] Table 5 F1 score of fusion layer track data classification

[0073]

[0074]

[0075] The fusion results are shown in Table 6. When the single classification result has a low accuracy of 75.2%, the accuracy after fusion is still improved to 87.6%. Fig.14 It can be seen that the fusion layer track data classification effectively improves the accuracy of track data classification on the basis of image layer and sequence layer track classification, and verifies that the fusion classification algorithm based on the improved DS evidence theory effectively improves the accuracy of track classification and recognition compared with a single classification algorithm, and improves the robustness of the track classification model.

[0076] Table 6 Fusion results

[0077]

Claims

1. A track hierarchical classification method based on a deep learning network, characterized by: Firstly, the image data of the track is established through the track sequence. The image data of the track is input into the trained Swin-Transformer model for classification and recognition to obtain the classification evidence body E1. Secondly, the improved DP algorithm is used to compress the track sequence data, and the compressed track sequence data is input into the trained GTN network for classification and recognition to obtain the classification evidence body E2; Finally, the improved DS evidence theory is used to fuse the above two types of classification evidence bodies E1 and E2 to achieve the classification and identification of the ship category to which the ship track belongs.

2. The track hierarchical classification method based on deep learning network according to claim 1, characterized in that: The process of establishing the image data of the track is as follows: mapping the latitude and longitude coordinates of the track points to the image pixel coordinates according to the scale, extracting the spatial features of the ship track data, and establishing the ship track image data set.

3. The track hierarchical classification method based on deep learning network according to claim 1 or 2, characterized in that: The process of compressing the track sequence data using the improved DP algorithm is as follows: calculate the distance between all track points in the middle of the starting point and the line AB connecting the starting point, and find the maximum value d max And the corresponding track points Calculate the distance between all track points in the starting point and the track points before and after it, calculate the average speed in the time domain based on the timestamp difference, and find the maximum speed V max And the corresponding track points Calculate the difference between the heading angles of all track points between the starting point and the heading angles before and after, and find the maximum difference H max And the corresponding track points Set the distance threshold d respectively th , speed threshold V th , heading angle threshold H th If any of the maximum distance, maximum speed, and maximum heading angle difference is greater than or equal to the corresponding threshold, the track is divided into two segments at the point corresponding to the maximum value for iterative track compression; if the maximum distance, maximum speed, and maximum heading angle difference are all less than the threshold, some intermediate track points of the track are discarded according to the compression ratio.

4. The track hierarchical classification method based on deep learning network according to claim 1 or 2, characterized in that: The specific process of DS evidence theory to integrate classified evidence is as follows: i Self-certainty In the formula, M represents the number of events corresponding to the focal element A in the identification framework, i represents the i-th piece of evidence in the identification framework, and m i (A j ) represents the probability that the i-th piece of evidence is classified as the j-th category; by calculating the self-certainty of the two pieces of evidence, we can get two self-certainty values, which are expressed in vector form, and the self-certainty vector of the evidence body C = [c1, c2] is obtained. The classification accuracy of the two classification models is used as the confidence of each piece of evidence. The product of the self-certainty of the evidence body and the classification accuracy is obtained to obtain the relative credibility vector Rel = [Rel1, Rel2]. Then the relative credibility vector is judged. in represents the credibility threshold, which is expressed as Then the weight factor vector is ω=[ω1,ω2,...,ω N ], and normalized, Calculate the evidence body BPA function m′, Finally, the modified BPA function is subjected to one DS fusion to obtain the fusion result m and determine the output target category.

5. The track hierarchical classification method based on deep learning network according to claim 1 or 2, characterized in that: The track sequence data is preprocessed before being input into the two classification models, the Swin-Transformer model and the GTN network. The preprocessing includes eliminating duplicate points and track interpolation.

Images