A twin target tracking method based on spatial-channel cross-correlation and centrality steering

Abstract

The application discloses a kind of twin target tracking methods based on space-channel cross-correlation and centrality guide, comprising the following steps: obtaining template and search area in image;Respectively, the template and search area are sent into deep feature extraction network and carry out feature extraction, and obtain template feature and search area feature;The template feature and search area feature are respectively sent into two space-channel cross-correlation modules, and the feature map R1 suitable for classification subnetwork and the feature map R2 suitable for regression subnetwork are obtained;The feature map R1 is sent into classification subnetwork to obtain classification map, and the feature map R2 is sent into regression subnetwork to obtain centrality map and regression map;Through loss function optimization, the bounding box of target is obtained according to the classification map, centrality map and regression map after optimization.The application effectively improves the precision and robustness of target tracking.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision target tracking, and in particular to a twin target tracking method based on space-channel cross-correlation and centrality guidance. BACKGROUND

[0002] Target tracking is a basic task of computer vision, which integrates knowledge from multiple fields such as machine learning, optimization and image processing, and has a wide range of applications in automatic monitoring, vehicle navigation, robot sensing, human-computer interaction and augmented reality. Although target tracking has made great progress, there are many factors affecting the robustness of visual tracking in complex and variable real scenes, such as light changes, occlusions, fast movements, deformations, scale changes and similar object interference.

[0003] In recent years, deep trackers based on twin networks have achieved a good balance between accuracy and speed, such as SiamFC, SiamRPN, DaSiamRPN, SiamRPN++, SiamCAR, etc., but there are still many shortcomings: SiamFC uses a simple naive correlation operation, which is essentially a convolution operation, directly using the template feature as the convolution kernel to correlate with the search feature, and the resulting feature map has only one channel; SiamRPN+ uses a deep cross-correlation operation, which independently correlates each channel of the input layer, so the number of channels of the resulting feature map is the same as that of the search feature, but it does not use the feature information of different channels at the same spatial position; some trackers first use a pixel-wise cross-correlation operation in the twin network, which is opposite to the deep cross-correlation, and it does not use the feature information of different spaces at the same channel position; and SiamCAR introduces a new centrality branch in the classification subnetwork to suppress the prediction box away from the center position, but if the centrality branch shares the same feature map with the classification branch, the prediction result may be similar to the classification prediction result, and the two branches cannot be well distinguished, and the centrality branch is not suitable for the classification subnetwork, as shown in (a) of Figure 1 Therefore, it is necessary to more effectively fuse the search region features and the template features, better utilize the characteristics of the centrality branch, and improve the high accuracy and high robustness of target tracking. SUMMARY

[0004] (I) Technical problems to be solved

[0005] Based on the above problems, the application provides a twin target tracking method based on space-channel cross-correlation and centrality guidance, which more effectively fuses search region features and template features, better utilizes the characteristics of the centrality branch, and solves the problems that the accuracy and robustness of target tracking need to be improved.

[0006] (II) Technical scheme

[0007] Based on the above technical problems, the application provides a twin target tracking method based on space-channel cross-correlation and centrality guidance, which more effectively fuses search region features and template features, better utilizes the characteristics of the centrality branch, and solves the problems that the accuracy and robustness of target tracking need to be improved.

[0008] S1, obtaining a template and a search region in an image;

[0009] S2, respectively feeding the template and the search region into a deep feature extraction network for feature extraction to obtain template features and search region features respectively as and The width x length x channel number size is Hz x Wz x C, Hx x Wx x C respectively.

[0010] S3, respectively feeding the template features and the search region features into two space-channel cross-correlation modules to obtain a feature map R1 suitable for a classification subnetwork and a feature map R2 suitable for a regression subnetwork.

[0011] S4, feeding the feature map R1 into the classification subnetwork to obtain a classification map, and feeding the feature map R2 into the regression subnetwork to obtain a centrality map and a regression map.

[0012] S5, optimizing the steps S3-S4 through a loss function, and obtaining a predicted target bounding box according to the optimized classification map, centrality map and regression map.

[0013] The step S3 includes:

[0014] S31, dividing the template features from the spatial dimension into a spatial kernel K1 including Hz x Wz small kernels, and the size of each small kernel is 1 x 1 x C; and dividing the template features from the channel dimension into a channel kernel K2 including C small kernels, and the size of each small kernel is 1 x 1 x HzWz.

[0015] S32, performing a pixel-by-pixel cross-correlation operation between the search region features and the spatial kernel K1 to obtain a feature map F1: ★ represents pixel-by-pixel cross-correlation;

[0016] S33, performing a pixel-by-pixel cross-correlation operation between the feature map F1 and the channel kernel K2 to obtain a feature map F2: F2=F1★K2.

[0017] S34, the feature map F1 and the feature map F2 are spliced, 1x1 convolution dimension reduction is performed, and the hot plug module SE-block is sent;

[0018] S35, the steps S31-S34 are repeated twice, and a feature map R1 suitable for a classification subnetwork and a feature map R2 suitable for a regression subnetwork are obtained, and the sizes of the feature maps are HzxWzxC.

[0019] Further, in the step S1, the template is an image of a specified pixel size that is cut out from a first frame image of a data set or a picture captured by a camera with a target as the center; and the search area is an image of a specified size that is cut out from an i+1th frame with the target position of an ith frame as the center in a tracking process.

[0020] Further, the step S4 comprises:

[0021] S41, the feature map R1 is sent to a classification subnetwork, the classification subnetwork is a CNN network, and there is only one classification branch, and a classification map is output. Each point of the classification map predicts the possibility that the point position is foreground or background.

[0022] S42, the feature map R2 is sent to a regression subnetwork, the regression subnetwork is a CNN network, and the regression subnetwork comprises a center degree branch and a bounding box regression branch, and a center degree map and a regression map are output. Each point of the center map predicts the possibility that the position is the center of the target, and each point of the regression map predicts the distance of the point from the upper, lower, left and right of the bounding box.

[0023] Further, in the step S5, the optimization by the loss function comprises: training the center degree branch by using a center degree loss, training the bounding box regression branch by using a center degree weighted regression loss, and training the classification branch by using a classification loss.

[0024] Further, the classification loss is:

[0025]

[0026] The center degree loss is:

[0027]

[0028] The center degree weighted regression loss is:

[0029]

[0030] wherein i and j represent coordinate positions of corresponding maps, and the true value of a classification label. a real value representing the center degree, N represents a result of network prediction, pos a number of positive samples, a set of positive samples, IoU represents an intersection over union of the two, B i,j , respectively represent a predicted bounding box and a real bounding box.

[0031] Further, in step S5, the predicted target bounding box obtained according to the optimized classification map, center degree map and regression map comprises:

[0032] a foreground part in the classification map is multiplied by the center degree map to obtain a point with the maximum response, that is, a predicted target center point; and distances of the point to four edges of a predicted target bounding box are obtained according to the regression map, and the predicted target center point is combined to obtain the predicted target bounding box.

[0033] Further, the specified size of the template cropping is 287*287 pixels, and the specified size of the search area cropping is 127*127 pixels.

[0034] Further, the width*length*channel size of the template feature is 13*13*256, and the width*length*channel size of the search area feature is 25*25*256.

[0035] The application also discloses a twin target tracking system based on spatial-channel cross-correlation and center degree guidance, comprising at least one processor and at least one memory connected with the processor in communication, wherein:

[0036] The memory stores program instructions executable by the processor, and the processor calling the program instructions can execute the twin target tracking method based on spatial-channel cross-correlation and center degree guidance, comprising the following functional modules:

[0037] The data acquisition module is used for executing the step S1.

[0038] The feature extraction module is used for executing the step S2.

[0039] The spatial-channel cross-correlation module is used for executing the step S3.

[0040] The classification regression module is used for executing the step S4.

[0041] The prediction module is used for executing the step S5.

[0042] The application also discloses a non-transitory computer readable storage medium storing computer instructions, which make the computer execute the twin target tracking method based on spatial-channel cross-correlation and centrality guidance.

[0043] (III) Beneficial effects

[0044] The above technical solution of the application has the following advantages:

[0045] (1) The application realizes feature fusion in two dimensions of space and channel by sending the search region features and the template features into the spatial-channel cross-correlation module, utilizes feature information of different channels at the same spatial position and feature information of different spaces at the same channel position, realizes more effective fusion of the search region features and the template features, is beneficial to improving the accuracy and robustness of target tracking, and reduces the interference of similar objects and the calculation amount.

[0046] (2) The regression sub-network of the application includes a centrality branch and a regression branch, and the whole regression sub-network is guided by the centrality branch, in addition to the centrality target guiding the branch of the centrality, the centrality target is weighted multiplied to the IoU loss, the weight of the low-quality predicted target bounding box far from the center point is suppressed, and the accuracy of the predicted target bounding box is improved.

[0047] (3) Due to the lightness of the spatial-channel cross-correlation module for feature fusion, the application can repeatedly perform two operations to obtain two feature maps with different applicability, which are respectively used as the input of the classification sub-network and the regression sub-network to cope with different sub-tasks, better distinguish the two tasks, and improve the accuracy and robustness of target tracking. BRIEF DESCRIPTION OF DRAWINGS

[0048] The features and advantages of the application will be more clearly understood through the following detailed description with reference to the accompanying drawings, which are schematic and should not be understood as limiting the application, and in which:

[0049] Figure 1 It is a comparison schematic diagram of SiamCAR and the method of the embodiment of the application.

[0050] Figure 2 It is a whole principle schematic diagram of the twin target tracking method based on spatial-channel cross-correlation and centrality guidance of the embodiment of the application.

[0051] Figure 3 It is a principle schematic diagram of the spatial-channel cross-correlation module part of step S3 of the embodiment of the application.

[0052] Figure 4A schematic diagram of the principle of the center degree guided regression sub-network part in step S42 of the embodiment of the present application;

[0053] Figure 5 Performance comparison of the method of the embodiment of the present application and other methods on the OTB100;

[0054] Figure 6 Performance comparison of the method of the embodiment of the present application and other methods on the UAV123. DETAILED DESCRIPTION

[0055] The specific embodiments of the present application are described in further detail below in combination with the drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0056] The embodiment of the present application is a twin target tracking method based on spatial-channel cross-correlation and center degree guidance, as shown in (b) of FIG. 1 and FIG. 2, comprising the following steps: Figure 1 Figure 2 S1, obtaining a template and a search area of a specified size in an image:

[0057] S1, obtaining a template and a search area of a specified size in an image:

[0058] The first frame image of the dataset or the camera-captured picture is cropped to a specified pixel size as a template with the target as the center, and the i+1 frame is cropped to a set size as a search area with the target position of the i frame as the center in the tracking process; the specified size of the template and the search area is 287×287 pixels and 127×127 pixels, respectively.

[0059] S2, respectively sending the template and the search area into a deep feature extraction network for feature extraction to obtain template features and search area features, respectively and The size of the template features and the search area features is Hz×Wz×C and Hx×Wx×C, respectively, indicating the width×length×channel number of the template features or the search area features.

[0060] In the embodiment, the size is 13×13×256 and 25×25×256, respectively.

[0061] S3, respectively sending the template features and the search area features into two spatial-channel cross-correlation modules, i.e., SC3M, to obtain a feature map R1 suitable for a classification sub-network and a feature map R2 suitable for a regression sub-network, as shown in FIG. 3, comprising the following steps: Figure 3

[0062] S31, dividing the template features into a spatial kernel K1 including Hz×Wz small kernels in the spatial dimension, and the size of each small kernel is 1×1×C; dividing the search area features ​​Splitting the channel dimension into a channel kernel K2 including C small kernels, each of which has a size of 1x1xHzWz;

[0063] S32, performing a pixel-by-pixel cross-correlation operation on the search region feature with the spatial kernel K1 to obtain a feature map F1, which has a size of HxWxHzWz: ★ represents pixel-by-pixel cross-correlation;

[0064] This step fully fuses the search region feature with the template feature in the spatial dimension;

[0065] S33, performing a pixel-by-pixel cross-correlation operation on the feature map F1 and the channel kernel K2 to obtain a feature map F2, which has a size of HzWxC: F2=F1*K2;

[0066] This step fully fuses the search region feature with the template feature in the channel dimension;

[0067] S34, splicing the feature map F1 and the feature map F2, performing 1x1 convolution dimension reduction, and sending to a hot plug module SE-block;

[0068] The 1x1 convolution and the hot plug module SE-block in this step can obtain a feature map suitable for different biases through optimization; the hot plug module SE-block does not change the size of the input feature, and its function is to obtain global information and the correlation between channels;

[0069] S35, repeating the steps S31-S34 twice, since the two repeated training optimizations have different focuses, a feature map R1 suitable for a classification subnetwork and a feature map R2 suitable for a regression subnetwork are obtained, and the sizes of R1 and R2 are both HzWxC, and are the same;

[0070] Due to the lightweight feature fusion of the steps S31-S34, the operation can be repeated twice; in SiamCAR, since the feature fusion calculation is large, it can only be calculated once to obtain a feature map for all subtasks.

[0071] S4, sending the feature map R1 to a classification subnetwork to obtain a classification map, and sending the feature map R2 to a regression subnetwork to obtain a centrality map and a regression map;

[0072] S41, sending the feature map R1 to a classification subnetwork, and the classification subnetwork is a CNN network, which has only one classification branch and outputs one classification map The classification map predicts the possibility of each point being foreground or background.

[0073] S42, the feature map R2 is sent to a regression sub-network, the regression sub-network is a CNN network, including a center degree branch and a bounding box regression branch, respectively outputting a center degree map and a regression map The center map predicts the possibility of each point being the center of the target, and the regression map predicts the distance of each point from the upper, lower, left and right of the bounding box, as shown in Figure 4 .

[0074] Among them, the classification branch, the center degree branch and the bounding box regression branch are all three full convolution layers, only the number of output channels is different.

[0075] S5, the predicted target bounding box is obtained by optimizing the steps S3-S4 through a loss function;

[0076] S51, the center degree branch is trained with a center degree loss, the bounding box regression branch is trained with a center degree weighted regression loss, and the classification branch is trained with a classification loss, wherein the classification loss adopts a CE (cross entropy) loss function, and the classification loss:

[0077]

[0078] The center degree loss adopts a BCE (binary cross entropy) loss function, and the center degree loss:

[0079]

[0080] The center degree weighted regression loss:

[0081]

[0082] Wherein, i, j represent the coordinate positions of the corresponding map, represents the true value of the classification label, represents the true value of the center degree, represents the result predicted by the network, N pos represents the number of positive samples, represents the set of positive samples, and IoU represents the intersection over union of the two, i,j , respectively represent the predicted bounding box and the real bounding box. We weight the true value of the center degree to the IoU regression loss, which can suppress the weight of the low-quality predicted target bounding box far from the center point.

[0083] S52, obtaining a predicted target bounding box according to the classification map, the center map and the regression map after optimization:

[0084] The classification map predicts the possibility of the position of each point being foreground and background, the center map predicts the possibility of the position of each point being the center of the target, and the regression map predicts the distance of each point from the upper, lower, left and right of the bounding box; therefore, the foreground part of the classification map is multiplied by the center map to obtain the point with the maximum response, that is, the predicted target center point; then, the distance of the point to the four edges of the predicted bounding box is obtained according to the regression map, and the predicted target center point is combined to obtain the predicted target bounding box.

[0085] To verify the technical effect of the embodiment, the above method is verified on multiple authoritative data sets such as VOT2018, OTB100, UAV123, GOT-10k, etc. Figure 5 The performance comparison of the method of the present application and other methods on the OTB100 is shown, wherein Figure 5 the (a) graph in (a) represents the precision graph on the OTB2015 data set, Figure 5 the (b) graph in (b) represents the success rate graph on the OTB2015 data set; Figure 6 The performance comparison of the method of the present application and other methods on the UAV123 data set is shown, wherein Figure 6 the (a) graph in (a) represents the precision graph of the method of the present application and other methods on the UAV123 data set, Figure 6Fig. b represents the success rate diagram of the method of the present application and other methods on the UAV123 dataset; the abscissa of the accuracy diagram represents the threshold value, and the ordinate represents the percentage of video frames in which the distance between the center point of the target position (bounding box) estimated by the tracking algorithm and the center point of the target marked by the ground truth is less than a given threshold value, the abscissa of the success rate diagram represents the threshold value, and the ordinate represents the percentage of frames in which the OS is greater than a set threshold value among all frames, the performance of the present application is better in the accuracy diagram and the success rate diagram; Table 1 shows the performance comparison of the method of the present application and other methods on the VOT2018 dataset, the larger the accuracy value is, the higher the accuracy is, the larger the robustness value is, the worse the stability is, EAO (abbreviation of Expected Average Overlap) represents the expected non-reset overlap, the larger the value is, the better the performance is, from the values in the table, the accuracy of the method of the present application is only inferior to SiamRCNN, the performance of EAO is the best, the robustness is also better, and the overall performance is excellent; Table 2 shows the performance comparison of the method of the present application and other methods on the GOT-10k dataset, AO represents the average overlap between all estimated bounding boxes and ground truth boxes, SR0.5 represents the success tracking frame rate of the overlap exceeding 0.5, SR0.75 represents the frame of the overlap exceeding 0.75, FPS represents the frame rate per second, the AO value of the method of the present application is only inferior to SiamGAT++, the SR 0.50 value is only inferior to SiamGAT++ and RBO, the SR 0.75 value is the best, the FPS value is also higher, and the overall performance of target tracking is better.

[0086] Table 1

[0087]

[0088] Table 2

[0089]

[0090] The embodiment of the present application also provides a twin target tracking system based on spatial-channel cross-correlation and centrality guidance, which can realize the twin target tracking method based on spatial-channel cross-correlation and centrality guidance, comprising a processor and a storage medium, the storage medium is used for storing instructions; the processor executes the twin target tracking method based on spatial-channel cross-correlation and centrality guidance, which comprises the following functional modules: a data acquisition module for executing step S1; a feature extraction module for executing step S2; a spatial-channel cross-correlation module for executing step S3; a classification regression module for executing step S4; and a prediction module for executing step S5.

[0091] The twin target tracking method described above can be converted into software program instructions, which can be implemented by using a twin target tracking system including a processor and a memory, or by computer instructions stored in a non-transitory computer readable storage medium. The integrated unit implemented in the form of a software function unit described above can be stored in a computer readable storage medium. The software function unit described above is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of the steps of the method described in various embodiments of the application. And the foregoing storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and various program code storage media.

[0092] In summary, through the twin target tracking method based on space-channel cross-correlation and centrality guidance described above, the following advantages are obtained:

[0093] (1) The present application sends the search region features and the template features into the space-channel cross-correlation module for feature fusion, realizes the fusion of features in the space and channel dimensions, utilizes the feature information of different channels at the same spatial position, and also utilizes the feature information of different spaces at the same channel position, realizes more effective fusion of search region features and template features, is conducive to improving the accuracy and robustness of target tracking, reduces the interference of similar objects, and effectively reduces the amount of calculation;

[0094] (2) The regression sub-network of the present application includes a centrality branch and a regression branch, and the entire regression sub-network is guided by the centrality branch. In addition to using the centrality target to guide the branch of the centrality, the centrality target is also weighted multiplied to the IoU loss, the weight of the low-quality predicted target bounding box far from the center point is suppressed, and the accuracy of the predicted target bounding box is improved;

[0095] (3) The lightness of the space-channel cross-correlation module for feature fusion enables the operation to be repeated twice, obtaining two feature maps with different applicability, which are respectively used as the input of the classification sub-network and the regression sub-network to cope with different sub-tasks, better distinguishing the two tasks, and improving the accuracy and robustness of target tracking.

[0096] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the embodiments of the present application are described in conjunction with the drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A twin target tracking method based on spatial-channel cross-correlation and centrality guidance, characterized in that, Includes the following steps: S1. Obtain the template and search region from the image; S2. The template and the search region are respectively fed into a deep feature extraction network for feature extraction, resulting in template features and search region features, respectively. and The width × length × number of channels are respectively Hz × Wz × C and Hx × Wx × C; S3. The template features and search region features are respectively sent to two spatial-channel cross-correlation modules to obtain feature map R1 suitable for the classification sub-network and feature map R2 suitable for the regression sub-network. S4. The feature map R1 is fed into the classification sub-network to obtain the classification map, and the feature map R2 is fed into the regression sub-network to obtain the centrality map and the regression map; S5. Optimize steps S3-S4 using a loss function, and obtain the predicted target bounding box based on the optimized classification map, centrality map, and regression map. Step S3 includes: S31, The template features The space is divided into spatial kernels K1, each consisting of Hz×Wz mini-kernels, with each mini-kernel having a size of 1×1×C; the template features are then... The channel is divided into a channel core K2 consisting of C small cores, each with a size of 1×1×HzWz. S32, The features of the search area A pixel-by-pixel cross-correlation operation is performed with the spatial kernel K1 to obtain the feature map F1: ★ indicates pixel-wise cross-correlation; S33. Perform a pixel-by-pixel cross-correlation operation between the feature map F1 and the channel kernel K2 to obtain the feature map F2: F2 = F1 ★ K2; S34. The feature map F1 and feature map F2 are concatenated, 1×1 convolution is performed to reduce the dimensionality, and the concatenation is sent to the hot-swappable module SE-block. S35. Steps S31-S34 are repeated twice to obtain feature map R1, which is suitable for the classification sub-network, and feature map R2, which is suitable for the regression sub-network, both with a size of Hz×Wz×C.

2. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 1, characterized in that, In step S1, the template is an image of a specified pixel size cropped from the first frame of the dataset or camera capture image, centered on the target; the search area is an image of a specified size cropped from the (i+1)th frame during the tracking process, centered on the target position in the i-th frame.

3. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 1, characterized in that, Step S4 includes: S41. The feature map R1 is fed into the classification sub-network, which is a CNN network with only one classification branch, and outputs a classification map. Each point in the classification chart predicts the probability that its location is either foreground or background. S42. The feature map R2 is fed into the regression sub-network, which is a CNN network, including a centrality branch and a bounding box regression branch, which output the centrality map respectively. Regression plot The center plot predicts the probability that each point is the center of the target, while the regression plot predicts the distance of each point from the bounding box in all directions.

4. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 1, characterized in that, In step S5, the optimization using the loss function includes: training the centrality branch using centrality loss, training the bounding box regression branch using centrality-weighted regression loss, and training the classification branch using classification loss.

5. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 4, characterized in that, The classification loss: The centrality loss: The centrality-weighted regression loss: Where i and j represent the coordinate positions in the corresponding graph. This represents the actual value of the category label. The true value representing centrality, N represents the result of network prediction. pos Represents the number of positive samples. Represents the set of positive samples, IoU represents the intersection-union ratio of the two samples within parentheses, and B i,j , These represent the predicted bounding box and the ground truth bounding box, respectively.

6. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 1, characterized in that, In step S5, obtaining the predicted target bounding box based on the optimized classification map, centrality map, and regression map includes: The foreground portion of the classification map is multiplied by the centrality map to obtain the point with the largest response, which is the predicted target center point; then, the distance from the point to the four sides of the predicted bounding box is obtained according to the regression map, and combined with the predicted target center point, the predicted target bounding box is obtained.

7. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 2, characterized in that, The specified size for template cropping is 287×287 pixels, and the specified size for search area cropping is 127×127 pixels.

8. The twin target tracking method based on spatial-channel cross-correlation and centrality guidance according to claim 1, characterized in that, The template feature has a width × length × number of channels of 13 × 13 × 256, and the search region feature has a width × length × number of channels of 25 × 25 × 256.

9. A twin target tracking system based on spatial-channel cross-correlation and centrality guidance, characterized in that, Includes at least one processor; and At least one memory communicatively connected to the processor, wherein: The memory stores program instructions executable by the processor. The processor, by calling the program instructions, can execute the twin target tracking method based on spatial-channel cross-correlation and centrality guidance as described in any one of claims 1 to 8, including the following functional modules: The data acquisition module is used to perform step S1. The feature extraction module is used to perform step S2. The spatial-channel correlation module is used to perform step S3. The classification and regression module is used to perform step S4. The prediction module is used to perform step S5.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the twin target tracking method based on spatial-channel cross-correlation and centrality guidance as described in any one of claims 1 to 8.

Images