A method and system for tower clustering based on spatial relationship

Abstract

The application discloses a tower gathering method and system based on spatial relationship. The method comprises the following steps: generating a Planck coordinate code, fusing the Planck coordinate code with an original image to form a multi-channel image; performing feature extraction on the multi-channel image, adding a Patch code of the original image to a linearly mapped Planck coordinate code, and performing pre-training to extract high-order semantic features; extracting target region features from a feature map and fusing position information; modeling the target region features by alternately performing global attention and single-image attention to learn the relationship between targets and generate high-level features; calculating target region similarity based on the high-level features, and judging the target region attribution through the similarity to realize tower matching and deduplication. The application effectively models the relationship between targets in multi-view images, improves the target matching accuracy in complex scenes, the model can infer the potential position of a hidden target, makes up for the single-view blind area, and improves the integrity of single-image information.

Description

TECHNICAL FIELD

[0001] The present application relates to a tower gathering method in the field of power systems, in particular to a tower gathering method based on spatial relationship, and also relates to a tower gathering system based on spatial relationship. BACKGROUND

[0002] In power inspection, when a device such as a UAV takes pictures of a tower from multiple angles, due to the complexity of the target structure and the occlusion effect, a single image cannot completely capture the three-dimensional information of the tower. For example, the equipment on the back side of the tower or small elements (such as insulator strings) are not visible at a certain angle of view, resulting in information blind spots in single-image perception. In order to construct a complete three-dimensional model of the tower, multiple-view images must be fused to complement the spatial information, but traditional two-dimensional images lack effective cross-view correlation mechanisms.

[0003] The same target (such as equipment or defects) appears repeatedly in different angles, causing data redundancy and matching ambiguity. If multiple-view image information is simply superimposed, errors will accumulate due to target position offset and scale changes. For example, the lightning arrester at the top of the tower appears in different forms in the overhead and side-view images, and simple addition will destroy data consistency. Therefore, the existing tower gathering method has the problems of incomplete information and inaccurate matching. SUMMARY

[0004] To solve the technical problems of incomplete information and inaccurate matching in the existing tower gathering method, the present application provides a tower gathering method based on spatial relationship and a system thereof.

[0005] The present application adopts the following technical solutions: a tower gathering method based on spatial relationship, comprising the following steps:

[0006] S1: generating a Planck coordinate code based on the positioning coordinates of the camera and the tilt angle of the gimbal, fusing the Planck coordinate code with the original image through channel splicing to form a multi-channel image containing color and three-dimensional spatial information;

[0007] S2: performing feature extraction on the multi-channel image, adding the Patch code of the original image to the linearly mapped Planck coordinate code, and pre-training through a DINOv2 model to extract high-order semantic features;

[0008] S3: extracting target region features from the feature map extracted from the DINOv2 model and fusing position information;

[0009] S4: modeling the multi-view relationship of the target region features by alternately performing global attention and single-image attention to learn the relationship between targets and generate high-level features;

[0010] S5: Calculate the target region similarity based on the high-level features, and determine the target region attribution through the similarity to realize the tower matching and deduplication.

[0011] The application converts the camera coordinates and the gimbal angle into three-dimensional space encoding by means of the Planck coordinate coding fusion and the DINOv2 spatial feature adaptation, so that the two-dimensional image carries the spatial position information of the tower, and the DINOv2 model simultaneously analyzes the RGB texture and the three-dimensional structure through the coding addition mechanism, and can perceive the spatial relationship of the target in the occlusion scene (such as the equipment on the back side of the tower), so that the model can infer the potential position of the occluded target, compensate for the single-view blind area, and improve the integrity of the single-image information. The application calculates the cross-image target similarity through the alternate attention mechanism and the feature similarity matching to establish the spatial correlation, and at the same time, the local features of the targets in the same image are strengthened, the same target is determined through the similarity, and the repeated counting error is eliminated, so that the collection accuracy is obviously improved, thereby solving the technical problems of the existing tower collection methods that the information is incomplete and the matching is inaccurate.

[0012] As a further improvement of the above scheme, in step S1, the internal and external parameters of the camera are calculated based on the real-time dynamic differential positioning coordinates recorded when the camera shoots the tower and the pitch and roll angles, and then the internal and external parameters are converted into the Planck coordinate coding.

[0013] As a further improvement of the above scheme, in step S2, the high-order semantic features of the image are extracted by inputting the multi-channel image into a feature extraction network, and the semantic information of the multi-channel image is extracted by the DINOv2 model to obtain the visual features.

[0014] As a further improvement of the above scheme, in step S3, according to the boundary box information obtained by the target detection based on the single image, the features of each target region are extracted from the feature map by using ROI-Align, and the position information features in the target region features are strengthened based on the position coding of the boundary box information.

[0015] As a further improvement of the above scheme, in step S4, the target region features are input into a Transformer network for processing, and a mask design graph neural network is designed by controlling the mask.

[0016] Further, the modeling method comprises the following steps: in the global attention stage, the self-attention mechanism of the Transformer network is configured to allow the features of all target regions of all images to interact with each other; and in the single-image attention stage, the self-attention mechanism of the Transformer network is configured to allow only the target region features within the same image to interact with each other.

[0017] Further, in step S5, first, the feature similarity between different target regions is calculated according to the features output by the Transformer network, and then it is judged whether the feature similarity reaches a preset similarity, yes or no, that is, whether different target regions belong to the same target or not.

[0018] As a further improvement of the above scheme, the alternately performing in step S4 includes the following steps:

[0019] (1) First, global attention is performed: the feature similarity of the target regions in all images is calculated;

[0020] (2) Then, single-image attention is performed: the feature similarity of the target regions in the same image is calculated;

[0021] (3) Steps (1) and (2) are alternately performed until the number of alternation rounds reaches a preset number.

[0022] As a further improvement of the above scheme, in step S1, the three channels of the original image and the six channels of the Planck coordinate encoding are spliced to generate a nine-channel image.

[0023] The application also provides a tower aggregation system based on spatial relationship, which applies any of the above-mentioned tower aggregation methods based on spatial relationship; the system comprises:

[0024] a three-dimensional space information fusion module for generating Planck coordinate encoding based on the positioning coordinates of the camera and the tilt angle of the gimbal, fusing the Planck coordinate encoding with the original image through channel splicing, and forming a multi-channel image containing color and three-dimensional space information;

[0025] a high-order semantic feature extraction module for extracting features from the multi-channel image, adding the Patch encoding of the original image to the linearly mapped Planck coordinate encoding, and pre-training through a DINOv2 model to extract high-order semantic features;

[0026] a target region feature extraction module for extracting target region features and fusing position information from the feature map extracted by the DINOv2 model;

[0027] a multi-view relationship modeling module for modeling the multi-view relationship of the target region features by alternately performing global attention and single-image attention, learning the relationship between targets, and generating high-level features;

[0028] a target matching and aggregation module for calculating the similarity of target regions based on high-level features, and judging the ownership of target regions through the similarity, realizing tower matching and deduplication.

[0029] Compared with the existing tower aggregation method, the tower aggregation method based on spatial relationship and the system thereof have the following beneficial effects:

[0030] 1. The tower aggregation method based on spatial relationship, by means of Planck coordinate coding fusion and DINOv2 spatial feature adaptation, converts camera coordinates and gimbal angles into three-dimensional space coding, so that the two-dimensional image carries the spatial position information of the tower, and the DINOv2 model simultaneously analyzes the RGB texture and three-dimensional structure through the coding addition mechanism, and can perceive the spatial relationship of the target in the occlusion scene (such as the equipment on the back side of the tower), so that the model can infer the potential position of the occluded target, make up for the single-view blind area, and improve the integrity of the single-image information, thereby solving the technical problem of incomplete information existing in the existing tower aggregation method.

[0031] 2. The tower aggregation method based on spatial relationship, by means of alternating attention mechanism and feature similarity matching, calculates the cross-image target similarity to establish spatial correlation, and at the same time strengthens the local features of the targets in the same image, determines the same target through similarity, eliminates the repeated counting error, so that the aggregation accuracy is obviously improved, thereby solving the technical problem of inaccurate matching existing in the existing tower aggregation method.

[0032] 3. The tower aggregation method based on spatial relationship effectively integrates three-dimensional space information (camera internal and external parameters) into a two-dimensional image through Planck coordinate coding, providing rich three-dimensional space information for subsequent feature extraction and target matching, making up for the limitations of traditional two-dimensional images, enhancing the model's perception ability of the three-dimensional structure of the target, and improving the target matching accuracy in complex scenes. Moreover, in order to make DINOv2 adapt to the input image containing Planck coordinate coding, the structure and pre-training weight of DINOv2 are modified, and the patch coding of the original RGB image is added to the linearly mapped Planck coordinate coding. This modification not only retains the general feature extraction capability of the pre-training weight of DINOv2, but also enables the model to comprehensively utilize the three-dimensional space information in the RGB image and the Planck coordinate coding.

[0033] 4. The tower aggregation method based on spatial relationship, by controlling the mask of the Transformer network, realizes the design of the graph neural network (GNN), and effectively models the relationship between the targets in the multi-view image. By alternately performing the "global attention" and "single-image attention" stages, the model can simultaneously capture global and local relationships, thereby better understanding the target relationship in the multi-view image and improving the accuracy of target matching.

[0034] 5、The tower collection method based on spatial relationship, through the complete closed loop of fusion, feature extraction, relationship modeling and matching, automatically completes multi-view image alignment, target matching and information integration, and the processing efficiency is greatly improved compared with traditional manual verification, manpower cost is saved. Moreover, the algorithm can process hundreds of input images in parallel, meeting the needs of power corridor line inspection.

[0035] The tower collection system based on spatial relationship has the same beneficial effects as the tower collection method based on spatial relationship, and will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The flowchart of the tower collection method based on spatial relationship of the embodiment 1 of the application.

[0037] Figure 2 The collection framework diagram of the tower collection method based on spatial relationship of the embodiment 1 of the application. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0039] Embodiment 1

[0040] Please refer to Figure 1 , the embodiment provides a tower collection method based on spatial relationship, which can be designed into an algorithm to realize automatic tower matching and deduplication. The collection method of the embodiment is a robust multi-view matching algorithm, which can accurately identify and remove duplicate targets in power inspection images, integrate multi-view information, and perfect tower information. The overall framework is as follows Figure 2 , the figure shows the algorithm for collecting according to two pictures, which can be easily extended to the scene of multiple picture input. The collection method includes the following steps (S1-S5)

[0041] S1: Based on the positioning coordinates of the camera and the tilt angle of the holder, generate the Plank coordinate code, fuse the Plank coordinate code with the original image through channel splicing, form a multi-channel image containing color and three-dimensional space information, and the multi-channel image contains RGB color information and three-dimensional position information. This step provides accurate three-dimensional spatial information for subsequent feature extraction and target matching, makes up for the lack of spatial information in two-dimensional images, and enhances the overall framework's ability to perceive the three-dimensional structure of the target.

[0042] In this embodiment, the internal and external parameters of the camera are first calculated based on the real-time dynamic difference positioning coordinates (i.e., RTK (Real-time kinematic) coordinate information) recorded when the camera shoots the tower and the elevation angle, and then the internal and external parameters are converted into Plücker coordinates. This encoding is an encoding method that can effectively represent three-dimensional spatial information, which can map three-dimensional coordinates to high-dimensional space, so that the network can better learn and utilize this information, thus effectively representing the three-dimensional spatial position information of the pixel points in the image.

[0043] In some embodiments, the three channels of the original image can be spliced with the six channels of the Plücker coordinate encoding to generate a nine-channel image. Of course, other numbers of channels of the fused image can also be generated.

[0044] S2: Feature extraction is performed on the multi-channel image, the Patch encoding of the original image is added to the linearly mapped Plücker coordinate encoding, and a DINOv2 model is pre-trained to extract high-order semantic features. In this embodiment, the high-order semantic features of the image are extracted by inputting the multi-channel image into a feature extraction network, and the semantic information of the multi-channel image is extracted by the DINOv2 model to obtain visual features. DINOv2 realizes the generalization ability of visual features close to human level through automatic data construction, multi-task loss fusion and efficient distributed training.

[0045] After research, DINOv2 is selected as the pre-training network in this embodiment, because DINOv2 uses a self-supervised learning method to pre-train on a large-scale dataset for a long time. It can extract rich semantic information in the image and generate highly robust and general visual features, which perform well in various downstream tasks such as image classification, object detection, and semantic segmentation. Since the input image contains Plücker coordinate encoding, the model parameters of DINOv2 cannot be directly used. In the specific design, the Patch encoding part is modified, the Patch encoding of the original RGB image is added to the linearly mapped Plücker coordinate encoding, and then input into the subsequent DINOv2 network operation, which makes the model maintain the general feature extraction ability of the pre-training weight, and can also integrate the three-dimensional spatial information of the RGB image and the Plücker coordinate encoding, providing reliable feature representation for subsequent target matching.

[0046] S3: Extracting target region features from the feature maps extracted from the DINOv2 model and fusing position information. In this embodiment, according to the bounding box information obtained based on single-image target detection, the features of each target region are extracted from the feature maps using ROI-Align (Region of Interest Align), and the position information features in the target region features are enhanced based on the position coding of the bounding box information. ROI-Align eliminates quantization errors through "floating-point coordinates + bilinear interpolation", which is a key technology for improving target detection / segmentation accuracy. In power tower clustering, it is combined with position information enhancement and multi-view attention modeling to solve the problem of small target matching in complex scenes and provide a high-precision feature basis for three-dimensional reconstruction of towers.

[0047] S4: Modeling multi-view relationships of target region features by alternately performing global attention and single-image attention to learn the relationships between targets and generate high-level features. In this embodiment, the target region features are input into a Transformer network for processing, and a graph neural network (GNN) is designed by controlling the mask. The modeling method includes the following steps: in the global attention stage, the self-attention mechanism of the Transformer network is configured to allow the features of all target regions of all images to interact with each other; in the single-image attention stage, the self-attention mechanism of the Transformer network is configured to allow only the features of target regions within the same image to interact with each other.

[0048] In this embodiment, the Transformer network uses the self-attention mechanism to learn the relationships between target regions and extract higher-level feature representations. Moreover, the alternating "global attention-single image attention" form is performed in the Transformer. The specific way of the modeling method of this embodiment is that the alternating "global attention-single image attention" form is performed in the Transformer, first in the global attention stage, the self-attention mechanism of the Transformer is configured to allow the features of all target regions of all images to interact with each other, which helps to capture long-distance dependency relationships between different target regions, so that the model can understand the relationships between different targets in different images, for example, whether the targets in two images are the same target; in the single-image attention stage, the self-attention mechanism of the Transformer is configured to allow only the features of target regions within the same image to interact with each other, which helps to capture local relationships between target regions within the same image and extract more refined feature representations. By alternately performing the global attention and single-image attention stages, the model can capture both global and local relationships, so that the model can better understand the target relationships in multi-view images and improve the accuracy of target matching.

[0049] In other embodiments, the alternately performing in step S4 can include the following steps: (1) first performing global attention: calculating the feature similarity of the target region in all images; (2) then performing single-image attention: only calculating the feature similarity of the target region within the same image; (3) alternately performing steps (1) and (2) until the alternation round reaches a preset number of times.

[0050] S5: calculating the target region similarity based on the high-level features, and judging the target region attribution through the similarity to realize the tower matching and deduplication. Wherein, the feature similarity between different target regions is calculated by using the features output by the Transformer network, and whether the different target regions belong to the same target is judged according to the feature similarity, and finally the matching result is output. According to the matching result, the automatic matching and deduplication of the target in the multi-view image are realized, and finally the tower-level collection is completed.

[0051] In the embodiment, the specific judgment manner is that the feature similarity between different target regions is calculated according to the features output by the Transformer network, and then whether the feature similarity reaches a preset similarity is judged, yes, then it is determined that the different target regions belong to the same target, otherwise, the different target regions do not belong to the same target.

[0052] In summary, compared with the existing tower collection method, the tower collection method based on spatial relationship in the embodiment has the following beneficial effects:

[0053] 1. The tower collection method based on spatial relationship, by Plank coordinate coding fusion and DINOv2 spatial feature adaptation, converts the camera coordinates and gimbal angles into three-dimensional space coding, so that the two-dimensional image carries the spatial position information of the tower, and the DINOv2 model analyzes the RGB texture and three-dimensional structure through the coding addition mechanism, and can still perceive the spatial relationship of the target in the occlusion scene (such as the equipment on the back side of the tower), so that the model can infer the potential position of the occluded target, make up for the single-view blind area, improve the integrity of the single-image information, and solve the technical problem of incomplete information existing in the existing tower collection method.

[0054] 2. The tower collection method based on spatial relationship, by alternately attention mechanism and feature similarity matching, calculates the cross-image target similarity to establish the spatial correlation, and at the same time strengthens the local features of the target within the same image, determines the same target through the similarity, eliminates the repeated counting error, so that the collection accuracy is obviously improved, and the technical problem of inaccurate matching existing in the existing tower collection method is solved.

[0055] 3、The tower aggregation method based on spatial relationship effectively integrates three-dimensional spatial information (camera internal and external parameters) into two-dimensional images through Plenoptic coordinate encoding, providing rich three-dimensional spatial information for subsequent feature extraction and target matching, making up for the limitations of traditional two-dimensional images, enhancing the model's perception of target three-dimensional structure, and improving the accuracy of target matching in complex scenes. Moreover, in order to adapt the DINOv2 to input images containing Plenoptic coordinate encoding, the structure and pre-training weights of DINOv2 are modified, and the Patch encoding of the original RGB image is added to the linearly mapped Plenoptic coordinate encoding. This modification not only retains the general feature extraction capability of the DINOv2 pre-training weights, but also enables the model to utilize the three-dimensional spatial information in both RGB images and Plenoptic coordinate encoding.

[0056] 4、The tower aggregation method based on spatial relationship realizes the design of Graph Neural Network (GNN) by controlling the mask of the Transformer network, effectively modeling the relationships between targets in multi-view images. By alternating between "global attention" and "single-image attention" stages, the model can capture both global and local relationships, better understanding the target relationships in multi-view images and improving the accuracy of target matching.

[0057] 5、The tower aggregation method based on spatial relationship automatically completes multi-view image alignment, target matching and information integration through a complete closed loop of fusion, feature extraction, relationship modeling and matching, with significantly improved processing efficiency compared to traditional manual verification, saving labor costs. Moreover, the algorithm can handle hundreds of input images in parallel, meeting the needs of power corridor line inspection.

[0058] Embodiment 2

[0059] This embodiment provides a tower aggregation system based on spatial relationship, which applies the tower aggregation method based on spatial relationship in embodiment 1 and specifically includes a three-dimensional spatial information fusion module, a high-order semantic feature extraction module, a target region feature extraction module, a multi-view relationship modeling module, and a target matching and aggregation module.

[0060] The three-dimensional space information fusion module is configured to generate a Planck coordinate code based on the camera-based positioning coordinates and the gimbal tilt angle, fuse the Planck coordinate code with the original image through channel splicing, and form a multi-channel image containing color and three-dimensional space information. The high-order semantic feature extraction module is configured to extract features from the multi-channel image, add the patch code of the original image to the linearly mapped Planck coordinate code, and pre-train through a DINOv2 model to extract high-order semantic features. The target region feature extraction module is configured to extract target region features from the feature map extracted by the DINOv2 model and fuse the position information. The multi-view relationship modeling module is configured to model the multi-view relationship of the target region features by alternately performing global attention and single-image attention, learn the relationship between targets, and generate high-level features. The target matching and collection module is configured to calculate the similarity of target regions based on the high-level features, determine the ownership of the target regions through the similarity, and realize the matching and deduplication of towers.

[0061] Embodiment 3

[0062] The embodiment provides a computer terminal including a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor implements the steps of the tower collection method based on spatial relationship of embodiment 1 when executing the program.

[0063] The method of embodiment 1 can be applied in the form of software, such as designed as a program running independently, installed on a computer terminal, which can be a computer, a smart phone, a control system, and other Internet of Things devices, etc. The method of embodiment 1 can also be designed as an embedded program running, installed on a computer terminal, such as installed on a single-chip microcomputer.

[0064] Embodiment 4

[0065] The embodiment provides a computer-readable storage medium having a computer program stored thereon. The program is executed by a processor to implement the steps of the tower collection method based on spatial relationship of embodiment 1.

[0066] The method of embodiment 1 can be applied in the form of software, such as designed as a program running independently on a computer-readable storage medium, which can be a U disk, designed as a U disk, and designed as a program starting the entire method through external triggering through the U disk.

[0067] The above is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for collecting towers based on spatial relationships, characterized in that, It includes the following steps: S1: Based on the camera's positioning coordinates and the gimbal's elevation angle, Planck coordinate codes are generated. The Planck coordinate codes are then fused with the original image through channel stitching to form a multi-channel image containing color and three-dimensional spatial information. S2: Perform feature extraction on the multi-channel image, add the Patch encoding of the original image to the Planck coordinate encoding after linear mapping, and pre-train it through a DINOv2 model to extract high-order semantic features; S3: Extract target region features from the feature map extracted by the DINOv2 model and fuse position information: Based on the bounding box information obtained from target detection based on a single map, use ROI-Align to extract features of each target region from the feature map, and strengthen the position information features in the target region features based on the position encoding of the bounding box information; S4: Multi-view relationship modeling of the target region features is performed by alternately executing global attention and single-image attention to learn the relationships between targets and generate high-level features; wherein, the target region features are input into a Transformer network for processing, and a graph neural network is designed by controlling the mask; the modeling method includes the following steps: in the global attention stage, the self-attention mechanism of the Transformer network is configured to allow features of all target regions in all images to interact with each other; in the single-image attention stage, the self-attention mechanism of the Transformer network is configured to allow features of target regions within the same image to interact with each other; S5: Calculate the similarity of the target region based on advanced features, and determine the belonging of the target region through the similarity to achieve tower matching and deduplication.

2. The pole and tower aggregation method based on spatial relationships as described in claim 1, characterized in that, In step S1, the intrinsic and extrinsic parameters of the camera are first calculated based on the real-time dynamic differential positioning coordinates and the elevation angle recorded when the camera photographs the tower, and then the intrinsic and extrinsic parameters are converted into Planck coordinate codes.

3. The pole and tower aggregation method based on spatial relationships as described in claim 1, characterized in that, In step S2, the high-order semantic features of the image are extracted by inputting the multi-channel image into a feature extraction network, and the semantic information of the multi-channel image is extracted by the DINOv2 model to obtain visual features.

4. The pole and tower aggregation method based on spatial relationships as described in claim 1, characterized in that, In step S5, the feature similarity between different target regions is first calculated based on the features output by the Transformer network. Then, it is determined whether the feature similarity reaches a preset similarity. If it does, the different target regions are determined to belong to the same target; otherwise, the different target regions do not belong to the same target.

5. The pole and tower aggregation method based on spatial relationships as described in claim 1, characterized in that, The alternating execution described in step S4 includes the following steps: (1) First, perform global attention: calculate the feature similarity of the target region in all images; (2) Perform single-image attention again: calculate the feature similarity of target regions within the same image only; (3) Alternately execute steps (1) and (2) until the number of alternations reaches a preset number.

6. The pole and tower aggregation method based on spatial relationships as described in claim 1, characterized in that, In step S1, the three channels of the original image are concatenated with the six channels of Planck coordinate encoding to generate a nine-channel image.

7. A pole collection system based on spatial relationships, characterized in that, Its application is the pole and tower aggregation method based on spatial relationships as described in any one of claims 1-6; the system includes: The three-dimensional spatial information fusion module is used to generate Planck coordinate codes based on the camera's positioning coordinates and the gimbal's elevation angle. The Planck coordinate codes are then fused with the original image through channel stitching to form a multi-channel image containing color and three-dimensional spatial information. The high-order semantic feature extraction module is used to extract features from the multi-channel image by adding the Patch encoding of the original image to the Planck coordinate encoding after linear mapping, and pre-training it through a DINOv2 model to extract high-order semantic features. The target region feature extraction module is used to extract target region features from the feature map extracted by the DINOv2 model and fuse position information: based on the bounding box information obtained from target detection based on a single map, the module uses ROI-Align to extract features of each target region from the feature map, and strengthens the position information features in the target region features based on the position encoding of the bounding box information. A multi-view relationship modeling module is used to model the multi-view relationships of the target region features by alternately executing global attention and single-image attention, so as to learn the relationships between targets and generate high-level features. The target region features are input into a Transformer network for processing, and a graph neural network is designed by controlling a mask. The modeling method includes the following steps: in the global attention stage, the self-attention mechanism of the Transformer network is configured to allow features of all target regions in all images to interact with each other; in the single-image attention stage, the self-attention mechanism of the Transformer network is configured to allow features of target regions within the same image to interact with each other only. The target matching and aggregation module is used to calculate the similarity of target regions based on high-level features, and to determine the belonging of target regions through the similarity, thereby realizing pole matching and deduplication.

Images