Hyperspectral target tracking method and system based on improved SiamGAT

Abstract

The invention relates to the technical field of image processing, in particular to a hyperspectral target tracking method and system based on improved SiamGAT, and the method comprises the steps: a tracker reads a first frame image of a hyperspectral image sequence as a template frame, and obtains an initial template feature map; reading a next frame of image as a current frame of image, and obtaining a search feature map; obtaining a target area of the current frame image according to the search feature map and the current accumulated template feature map; taking the current frame image as a template frame, and obtaining a template feature map of the current frame; and updating the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map. According to the hyperspectral target tracking method, the coping capacity of the hyperspectral target tracking method for appearance change is effectively enhanced, the situation that the appearance of the target in the hyperspectral image sequence changes remarkably can be effectively processed, and the accuracy of hyperspectral target tracking is improved.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a hyperspectral target tracking method and system based on improved SiamGAT. Background Art

[0002] Target tracking in hyperspectral image sequences is an important part of the field of hyperspectral image processing technology. It relies on early spectral imaging technology and combines imaging technology with spectral technology. Hyperspectral images have rich spectral bands and can distinguish objects with similar reflectivity in visible light bands. Hyperspectral target tracking technology has very wide practical applications in modern military, security, and remote sensing systems. Feature extraction based on hyperspectral images is crucial for subsequent image processing, target recognition and tracking tasks. For example, hyperspectral video target tracking is to achieve real-time, accurate, and stable tracking of targets that are difficult to distinguish in visible light by extracting features from hyperspectral images.

[0003] In the existing technology, the initial frame image of the hyperspectral image sequence is usually used as a fixed template frame image for tracking. Although the initial frame contains highly reliable prior information, it cannot adapt well to the deformation problem of the target in the later stage. Therefore, when using a simple fixed template image for feature extraction, it is easy to be disturbed by similar objects in the background, which makes the target easy to be lost in the later stage, resulting in reduced accuracy of the results when performing target detection, target recognition or target tracking based on the feature information. For example, the existing hyperspectral target tracking method usually directly uses a convolutional neural network to extract features from the template frame image and the search frame image respectively, and then performs a deep cross-correlation operation on the obtained template features and the search features to obtain the center feature map, classification feature map and regression feature map of the search frame to determine the target coordinate range of the current search frame image. However, when the appearance of the tracked target changes significantly, the saved template features cannot provide accurate target information, resulting in the model being unable to accurately locate the target position based on the extracted template features and search features, thereby affecting the accuracy of hyperspectral target tracking.

[0004] In summary, the existing hyperspectral target tracking methods are easily affected by target occlusion and deformation, which makes the template features unable to provide accurate target information, resulting in tracking deviation and reducing the accuracy of hyperspectral target tracking; and for the case of more background spots, that is, when there are too many similar objects, the tracking effect is poor. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the problem in the prior art that the accuracy of hyperspectral target tracking is reduced when the target appearance changes significantly, such as target occlusion and deformation.

[0006] In order to solve the above technical problems, the present invention provides a hyperspectral target tracking method based on improved SiamGAT, comprising: S1: The tracker reads the first frame of the hyperspectral image sequence as a template frame for preprocessing to obtain an initial template image, extracts features from the initial template image to obtain an initial template feature map, and uses the initial template feature map as an initial cumulative template feature map; S2: The tracker reads the next frame image of the hyperspectral image sequence as the current frame image, preprocesses the current frame image to obtain the current search frame image, and extracts features from the current search frame image to obtain a search feature map; S3: Obtain the target area of ​​the current frame image according to the search feature map and the current accumulated template feature map; S4: preprocessing the current frame image as a template frame to obtain a current frame template image; extracting features from the current frame template image to obtain a current frame template feature map; S5: after updating the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map, return to execute S2 to obtain the target area of ​​each frame image in the hyperspectral image sequence.

[0007] Preferably, obtaining the target area of ​​the current frame image according to the search feature map and the current accumulated template feature map includes: Input the search feature map and the current accumulated template feature map into the graph attention network for cross-correlation operation and output the aggregated features; The aggregated features are passed through the EMA multi-scale attention module to output the response map; The response map is passed through the classification regression head network to output the target area of ​​the current frame image.

[0008] Preferably, after obtaining the target area of ​​the current frame image according to the search feature map and the current cumulative template feature map, the method further includes: The confidence score of the target area of ​​the current frame image is calculated according to the response map. The formula is: ; in, represents the confidence score, Indicates the peak value of the current frame response graph, Represents the average value of the response graph of the previous frame, is a preset constant; i represents the frame index; If the confidence score is greater than the preset threshold, the current frame template feature map is obtained, and the cumulative template feature map is updated according to the initial template feature map, the current cumulative template feature map and the current frame template feature; if the confidence score is less than or equal to the preset threshold, the cumulative template feature map is not updated.

[0009] Preferably, updating the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map includes: After the initial template feature map, the current accumulated template feature map and the current frame template feature map are fused, they are combined with the initial template feature map. Figure 1 Then input the UpdateNet module for updating to obtain the updated cumulative template feature map.

[0010] Preferably, the initial template feature map, the current cumulative template feature map and the current frame template feature map are fused, and the formula is expressed as: ; in, represents feature extraction, represents the initial template feature map, Represents the current cumulative template feature map, Represents the current frame template feature map, Represents the fused feature map; i represents the frame index; Combine the fused feature map with the initial template feature Figure 1 Input the UpdateNet module to update and obtain the updated cumulative template feature map, which is expressed as: ; in, represents the UpdateNet module, Represents the updated cumulative template feature map; i represents the frame index.

[0011] Preferably, when updating the cumulative template feature map, the current cumulative template feature map is discarded every n frames, the initial template feature map is fused with the current frame template feature map to obtain a new cumulative template feature map, and the new cumulative template feature map is updated.

[0012] Preferably, the hyperspectral image sequence to be processed is input into the tracker after dimension reduction.

[0013] Preferably, the template frame is preprocessed, including: Acquire a target area in the template frame image, determine a cropping area according to the target area, crop the template frame image, and convert the cropped image to a preset template size as a template image; The cropping area is centered on the center of the target area, and its length and width are twice that of the target area respectively.

[0014] Preferably, the current search frame image is obtained after preprocessing the current frame image, including: Obtaining the target area of ​​the previous frame image of the current frame image, determining the search area according to the target area of ​​the previous frame image, cropping the search area and converting it to a preset search area size as the current search frame image; The search area is centered on the center position of the target area, and its length and width are twice that of the target area respectively.

[0015] The present invention also provides a hyperspectral target tracking system based on improved SiamGAT, comprising: An initial template feature map acquisition module is used for the tracker to read the first frame image of the hyperspectral image sequence as a template frame for preprocessing to obtain an initial template image, extract features from the initial template image to obtain an initial template feature map, and use the initial template feature map as an initial cumulative template feature map; A search feature map acquisition module is used for the tracker to read the next frame image of the hyperspectral image sequence as the current frame image, preprocess the current frame image to obtain the current search frame image, and extract features from the current search frame image to obtain a search feature map; A target region prediction module is used to obtain the target region of the current frame image according to the search feature map and the current accumulated template feature map; The current frame template feature map acquisition module is used to pre-process the current frame image as a template frame to obtain the current frame template image; extract features from the current frame template image to obtain the current frame template feature map; The cumulative template feature map updating module is used to update the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map, and then return to the search feature map acquisition module to obtain the target area of ​​each frame image in the hyperspectral image sequence.

[0016] The above technical solution of the present invention has the following beneficial effects compared with the prior art: The invention discloses a method for tracking a hyperspectral target based on improved SiamGAT, wherein template features are extracted for each frame image, and the cumulative template feature map is gradually updated using the initial template feature map, the current cumulative template feature map and the current frame template feature map, and the target area is predicted for the next search frame image. When the target changes greatly over time, the target features before and after will also differ greatly. The invention can still capture the current morphological features of the target by updating the cumulative template feature map, and the initial template feature map is always retained during the updating process, so that the complete morphology of the target can be obtained to the greatest extent, overcoming the disadvantage of the prior art that the first frame image is used as a fixed template image for tracking, and cannot adapt well to the deformation problem of the later target, and improves the tracking performance of the hyperspectral target under the conditions of significant changes in the target appearance, including deformation, occlusion and small targets.

[0017] Furthermore, the present invention determines whether to add the current frame template feature map to the cumulative template feature map by calculating the confidence score of the target area of ​​each frame image, thereby effectively suppressing the influence of the template feature map containing erroneous information on the tracking accuracy, overcoming the shortcomings of the prior art that the introduction of erroneous template information causes tracking deviation or failure, enhancing the stability of the cumulative template features, and further improving the tracking performance of hyperspectral targets in the case of significant changes in target appearance such as deformation, occlusion and small targets.

[0018] Therefore, the present invention effectively enhances the ability of the hyperspectral target tracking method to cope with appearance changes, and can effectively handle significant changes in target appearance in hyperspectral image sequences, including deformation, occlusion, and small targets, thereby improving the accuracy of hyperspectral target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of a hyperspectral target tracking method based on improved SiamGAT of the present invention; Figure 2 A schematic diagram of a search frame of a hyperspectral image after dimensionality reduction in an embodiment of the present invention; Figure 3 In the embodiment of the present invention Figure 2 Schematic diagram of the cropped image; Figure 4 In the embodiment of the present invention, Figure 3 Response graph of ; Figure 5 In the embodiment of the present invention Figure 2 Schematic diagram of labels and prediction boxes after tracking; Figure 6 The flowchart of template updating in the embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0021] Reference Figure 1 As shown, the present invention provides a hyperspectral target tracking method based on improved SiamGAT, comprising: S1: The tracker reads the first frame of the hyperspectral image sequence as a template frame for preprocessing to obtain an initial template image, extracts features from the initial template image to obtain an initial template feature map, and uses the initial template feature map as an initial cumulative template feature map.

[0022] The specific steps are as follows: S101: The tracker reads the first frame image of the hyperspectral image sequence as a template frame image, wherein the size of the first frame image is M×N×3.

[0023] S102: Acquire a target area in the template frame image, determine a cropping area according to the target area, crop the template frame image, and convert the cropped image to a preset template size as an initial template image.

[0024] Specifically, the target region in the template frame image is obtained according to the target label given by the first frame image, and the center position of the target region is used as the cropping center to determine a cropping region with a side length twice the width and height of the target region. The area of ​​the cropping region is four times the area of ​​the target region. The template frame image is cropped, and then the cropped image is converted to a preset template size as the initial template image. The initial template image of this embodiment refers to Figure 3 shown.

[0025] When the target is at the edge of the image, the cropping area determined according to the target area may exceed the image range. The global pixel mean can be used to fill the area beyond the image.

[0026] In this embodiment, the preset template size is 127*127*3.

[0027] S103: Use the GoogLeNet network to extract features from the initial template image to obtain an initial template feature map .

[0028] Preferably, before S1, the method further includes: reducing the dimension of the hyperspectral image sequence to be processed and inputting it into a tracker.

[0029] Specifically, the hyperspectral image sequence used in this embodiment is 25 bands, and the image size is M×N×25, where M×N is the scene size. A fine segmentation algorithm is used to divide the 25 bands of the hyperspectral image into three band groups, namely 1~9, 10~18, and 19~25, where highly correlated spectral bands are divided into one band group, and the correlation between the band groups is low. The band selection technology of the normalized filter is used to calculate the weight of each band in each band group, and the band with the largest weight is selected as the high-effectiveness band of the band group. The high-effectiveness bands in each band group are combined into a reduced-dimensional hyperspectral image of size M×N×3. The reduced-dimensional hyperspectral image of this embodiment refers to Figure 2 shown.

[0030] S2: The tracker reads the next frame image of the hyperspectral image sequence as the current frame image, obtains the current search frame image after preprocessing, and extracts features from the current search frame image to obtain a search feature map.

[0031] The specific steps are as follows: S201: The tracker reads the next frame image of the hyperspectral image sequence as the current frame image.

[0032] S202: Acquire the target area of ​​the previous frame image of the current frame image, determine the search area according to the target area of ​​the previous frame image, crop the search area and convert it to a preset search area size as the current search frame image.

[0033] Specifically, the target area of ​​the previous frame image of the current frame image is obtained, and the center position of the target area is used as the cropping center, and a search area with a side length twice the width and height of the target area is determined, and the area of ​​the search area is four times the area of ​​the target area. The remaining area outside the search area is the background image area. The search area is cropped and converted to a preset search area size as the search frame image.

[0034] When the target of the previous frame image is at the edge of the image, the search area determined according to the target area may exceed the image range, and the global pixel mean can be used to fill the area beyond the image.

[0035] In this embodiment, the preset search area size is 287*287*3.

[0036] S203: Use the GoogLeNet network to extract features of the current search frame image to obtain a search feature map.

[0037] S3: Obtain the target area of ​​the current frame image according to the search feature map and the current accumulated template feature map, including: Input the search feature map and the current accumulated template feature map into the graph attention network for cross-correlation operation and output the aggregated features; The aggregated features are passed through the EMA multi-scale attention module to output a response graph; the EMA multi-scale attention module can extract features from inputs of different scales, while paying attention to local details and overall structure, and strengthening the aggregated features; the response graph refers to Figure 4 As shown; The response map is passed through the classification regression head network to output the target area of ​​the current frame image. Figure 5 The target area of ​​the current frame image includes target position and size information.

[0038] S4: preprocess the current frame image as a template frame, and the preprocessing method is the same as S2 to obtain the current frame template image; extract features of the current frame template image to obtain the current frame template feature map.

[0039] Use the GoogLeNet network to extract features from the current frame template image and obtain the current frame template feature map .

[0040] S5: after updating the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map, return to execute S2 to obtain the target area of ​​each frame image in the hyperspectral image sequence.

[0041] According to the initial template feature map, the current cumulative template feature map and the current frame template feature map, the cumulative template feature map is updated, including: After the initial template feature map, the current accumulated template feature map and the current frame template feature map are fused, they are combined with the initial template feature map. Figure 1 Then input the UpdateNet module for updating to obtain the updated cumulative template feature map.

[0042] Specifically, the initial template feature map, the current cumulative template feature map and the current frame template feature map are fused, and the formula is expressed as: ; in, represents feature extraction, represents the initial template feature map, Represents the current cumulative template feature map, Represents the current frame template feature map, Represents the fused feature map, and i represents the frame index.

[0043] Combine the fused feature map with the initial template feature Figure 1 Input the UpdateNet module to update and obtain the updated cumulative template feature map, which is expressed as: ; in, represents the UpdateNet module, Represents the updated cumulative template feature map; i represents the frame index.

[0044] Preferably, in S5, the current accumulated template feature map is discarded every n frames, and the initial template feature map is regenerated. The new cumulative template feature map is fused with the current frame template feature map, and then the template feature maps of subsequent frames are added in sequence to update the cumulative template feature map.

[0045] Specifically, whenever the number of templates fused in the current cumulative template feature map reaches n, the current cumulative template feature map is discarded, that is, the current cumulative template feature map is set as the initial template feature map. .

[0046] In the embodiment of the present invention, n is set to 10. Whenever the number of templates fused in the current cumulative template feature map reaches 10, the cumulative template feature map is reset to the initial template feature map. .

[0047] Specifically, after resetting the cumulative template feature map, the cumulative template feature map again only contains the initial template feature map. At this time, the template feature map of the subsequent frame is compared with the initial template feature map , the current cumulative template feature map (i.e. the initial template feature ) and then fuse the fused feature map with the initial template feature map They are sent together to the UpdateNet module for updating to obtain the updated cumulative template feature map.

[0048] Preferably, a template screening module is added after S3, and the template screening module calculates the confidence score of the target area of ​​the current frame according to the response map; if the confidence score is greater than a preset threshold, the current frame template feature map is obtained, and the cumulative template feature map is updated according to the initial template feature map, the current cumulative template feature map and the current frame template feature; if the confidence score is less than or equal to the preset threshold, the cumulative template feature map is not updated.

[0049] Specifically, the template screening module calculates the confidence score of the target area of ​​the current frame image according to the response map, and the formula is expressed as: ; in, represents the confidence score, Indicates the peak value of the current frame response graph, Represents the average value of the response graph of the previous frame, is a preset constant; i represents the frame index.

[0050] The peak value of the response graph of the current frame can be used as a measure of tracking confidence, but There are great differences in different environments. When the target appearance changes significantly, such as deformation, occlusion, and the target is too small, even if the target is tracked accurately, the peak of the response graph will still be very low, so in the embodiment of the present invention, the peak value of the response graph To standardize the process, Represents the confidence score after normalization.

[0051] Before updating the cumulative template feature map each time, first determine the confidence score of the target area of ​​the current frame image With preset threshold If the confidence score is greater than the preset threshold, the cumulative template feature map is updated according to the initial template feature map, the current cumulative template feature map and the current frame template feature map; if the confidence score is less than or equal to the preset threshold, it means that the current frame template feature map incorporates wrong information, causing the tracker to make a wrong judgment, so the current frame template feature map is removed, that is, the cumulative template feature map is not updated. The formula is expressed as: ; in, represents the preset threshold; i represents the frame number index.

[0052] When updating the cumulative template feature map, if the target is deformed, blocked, or other problems occur, directly passing each frame of the template feature map to the update network without necessary screening, the cumulative template feature map may incorporate a lot of erroneous information, resulting in a large amount of noise in the template, which may reduce the tracking effect. Therefore, in the embodiment of the present invention, the threshold is set. =0.95, when the confidence score When it is greater than 0.95, it means that the latest added current frame template feature map has correct feature information, and the update result of the cumulative template feature map is retained; when the confidence score When the result is less than or equal to 0.95, it means that the wrong feature information has been incorporated, and the latest added current frame template feature map is selected to be removed. The complete template update flow chart is as follows: Figure 6 shown.

[0053] In summary, the hyperspectral target tracking method based on improved SiamGAT described in the present invention extracts template features for each frame image, and gradually updates the cumulative template feature map using the initial template feature map, the current cumulative template feature map and the current frame template feature map, and predicts the target area for the next search frame image. When the target changes greatly over time, the target features before and after will also differ a lot. The present invention can still capture the current morphological features of the target by updating the cumulative template feature map, and always retains the initial template feature map during the updating process, so as to obtain the complete morphology of the target to the greatest extent, overcoming the disadvantage of using the first frame image as a fixed template image for tracking in the prior art, and not being able to adapt well to the deformation problem of the later target, and improving the tracking performance of the hyperspectral target under the conditions of significant changes in the target appearance, including deformation, occlusion and small targets.

[0054] Furthermore, the present invention also introduces a template screening module, which determines whether to add the current frame template feature map to the cumulative template feature map by calculating the confidence score of the target area of ​​each frame image, thereby effectively suppressing the influence of the template feature map containing erroneous information on the tracking accuracy, overcoming the shortcomings of the prior art of introducing erroneous template information, causing estimation errors, resulting in tracking offset or failure, enhancing the stability of the cumulative template features, and further improving the tracking performance of high-spectral targets under conditions of significant changes in target appearance such as deformation, occlusion and small targets.

[0055] Therefore, the present invention effectively enhances the ability of the hyperspectral target tracking method to cope with appearance changes, and can effectively handle significant changes in target appearance in hyperspectral image sequences, including deformation, occlusion, and small targets, thereby improving the accuracy of hyperspectral target tracking.

[0056] Based on the above-mentioned hyperspectral target tracking method based on improved SiamGAT, the present invention also provides a hyperspectral target tracking system based on improved SiamGAT, comprising: An initial template feature map acquisition module is used for the tracker to read the first frame image of the hyperspectral image sequence as a template frame for preprocessing to obtain an initial template image, extract features from the initial template image to obtain an initial template feature map, and use the initial template feature map as an initial cumulative template feature map; A search feature map acquisition module is used for the tracker to read the next frame image of the hyperspectral image sequence as the current frame image, preprocess the current frame image to obtain the current search frame image, and extract features from the current search frame image to obtain a search feature map; A target region prediction module is used to obtain the target region of the current frame image according to the search feature map and the current accumulated template feature map; The current frame template feature map acquisition module is used to pre-process the current frame image as a template frame to obtain the current frame template image; extract features from the current frame template image to obtain the current frame template feature map; The cumulative template feature map updating module is used to update the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map, and then return to the search feature map acquisition module to obtain the target area of ​​each frame image in the hyperspectral image sequence.

[0057] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0058] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0059] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0061] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A hyperspectral target tracking method based on improved SiamGAT, characterized in that: include: S1: The tracker reads the first frame of the hyperspectral image sequence as a template frame for preprocessing to obtain an initial template image, extracts features from the initial template image to obtain an initial template feature map, and uses the initial template feature map as an initial cumulative template feature map; S2: The tracker reads the next frame image of the hyperspectral image sequence as the current frame image, preprocesses the current frame image to obtain the current search frame image, and extracts features from the current search frame image to obtain a search feature map; S3: Obtain the target area of ​​the current frame image according to the search feature map and the current accumulated template feature map; S4: preprocessing the current frame image as a template frame to obtain a current frame template image; extracting features from the current frame template image to obtain a current frame template feature map; S5: after updating the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map, return to execute S2 to obtain the target area of ​​each frame image in the hyperspectral image sequence.

2. The hyperspectral target tracking method based on improved SiamGAT according to claim 1, characterized in that: The target area of ​​the current frame image is obtained according to the search feature map and the current accumulated template feature map, including: Input the search feature map and the current accumulated template feature map into the graph attention network for cross-correlation operation and output the aggregated features; The aggregated features are passed through the EMA multi-scale attention module to output the response map; The response map is passed through the classification regression head network to output the target area of ​​the current frame image.

3. The hyperspectral target tracking method based on improved SiamGAT according to claim 2, characterized in that: After obtaining the target area of ​​the current frame image according to the search feature map and the current accumulated template feature map, it also includes: The confidence score of the target area of ​​the current frame image is calculated according to the response map. The formula is: ； in, represents the confidence score, Indicates the peak value of the current frame response graph, Represents the average value of the response graph of the previous frame, is a preset constant; i represents the frame index; If the confidence score is greater than the preset threshold, the current frame template feature map is obtained, and the cumulative template feature map is updated according to the initial template feature map, the current cumulative template feature map and the current frame template feature; if the confidence score is less than or equal to the preset threshold, the cumulative template feature map is not updated.

4. The hyperspectral target tracking method based on improved SiamGAT according to claim 1, characterized in that: According to the initial template feature map, the current cumulative template feature map and the current frame template feature map, the cumulative template feature map is updated, including: After the initial template feature map, the current cumulative template feature map and the current frame template feature map are fused, they are input into the UpdateNet module together with the initial template feature map for updating to obtain the updated cumulative template feature map.

5. The hyperspectral target tracking method based on improved SiamGAT according to claim 4, characterized in that: The initial template feature map, the current cumulative template feature map and the current frame template feature map are fused, and the formula is expressed as: ； in, represents feature extraction, represents the initial template feature map, Represents the current cumulative template feature map, Represents the current frame template feature map, Represents the fused feature map; i represents the frame index; The fused feature map and the initial template feature map are input into the UpdateNet module for updating to obtain the updated cumulative template feature map. The formula is expressed as: ； in, represents the UpdateNet module, Represents the updated cumulative template feature map; i represents the frame index.

6. The hyperspectral target tracking method based on improved SiamGAT according to claim 4, characterized in that: When updating the cumulative template feature map, the current cumulative template feature map is discarded every n frames, the initial template feature map is fused with the current frame template feature map to obtain a new cumulative template feature map, and the new cumulative template feature map is updated.

7. The hyperspectral target tracking method based on improved SiamGAT according to claim 1, characterized in that: The hyperspectral image sequence to be processed is input into the tracker after dimensionality reduction.

8. The hyperspectral target tracking method based on improved SiamGAT according to claim 1, characterized in that: Preprocess the template frame, including: Acquire a target area in the template frame image, determine a cropping area according to the target area, crop the template frame image, and convert the cropped image to a preset template size as a template image; The cropping area is centered on the center of the target area, and its length and width are twice that of the target area respectively.

9. The hyperspectral target tracking method based on improved SiamGAT according to claim 7, characterized in that: The current search frame image is obtained after preprocessing the current frame image, including: Obtaining the target area of ​​the previous frame image of the current frame image, determining the search area according to the target area of ​​the previous frame image, cropping the search area and converting it to a preset search area size as the current search frame image; The search area is centered on the center position of the target area, and its length and width are twice that of the target area respectively.

10. A hyperspectral target tracking system based on improved SiamGAT, characterized in that: include: An initial template feature map acquisition module is used for the tracker to read the first frame image of the hyperspectral image sequence as a template frame for preprocessing to obtain an initial template image, extract features from the initial template image to obtain an initial template feature map, and use the initial template feature map as an initial cumulative template feature map; A search feature map acquisition module is used for the tracker to read the next frame image of the hyperspectral image sequence as the current frame image, preprocess the current frame image to obtain the current search frame image, and extract features from the current search frame image to obtain a search feature map; A target region prediction module is used to obtain the target region of the current frame image according to the search feature map and the current accumulated template feature map; The current frame template feature map acquisition module is used to pre-process the current frame image as a template frame to obtain the current frame template image; extract features from the current frame template image to obtain the current frame template feature map; The cumulative template feature map updating module is used to update the cumulative template feature map according to the initial template feature map, the current cumulative template feature map and the current frame template feature map, and then return to the search feature map acquisition module to obtain the target area of ​​each frame image in the hyperspectral image sequence.

Images