Target tracking method based on attention aggregation and anchor point proposal

Abstract

The invention requests to protect a target tracking method of an attention aggregation and anchor proposal network. The method comprises the following steps: S1, a feature extraction network obtains a feature map of multi-scale information; s2, performing scale and channel attention weighting on the extracted multi-scale features, and fusing to obtain feature data of rich scale information; s3, performing attention cross-correlation operation by using the output of the multi-scale feature extraction network to obtain an adaptive-scale anchor point proposal network; s4, inputting the result of the feature fusion network into multi-classification and regression to obtain a proposal anchor point box; and S5, inputting a suggested anchor point result of the anchor point suggestion network into a multi-classification and regression result for fusion and maximum value selection to obtain a final tracking target frame, wherein the overall structure is as shown in figure 1. According to the method, the problem of tracking failure caused by rapid deformation, scale change and shielding of the object in the target tracking process can be effectively solved, and the accuracy and robustness of target tracking are improved.

Description

Technical Field

[0001] This invention belongs to the field of target tracking, specifically a target tracking method that employs attention aggregation and anchor proposal networks. Background Technology

[0002] Target tracking is an important branch of computer vision. It utilizes contextual information from video or image sequences to model the appearance and motion information of a target, thereby predicting the target's motion state and pinpointing its location. Target tracking integrates theories and algorithms from multiple fields such as image processing, machine learning, and optimization, and is a prerequisite and foundation for completing higher-level image understanding tasks.

[0003] In recent years, target tracking has received increasing attention in social production and daily life. With the development of modern information technology and science, various tracking systems based on target tracking technology have been implemented and widely applied in fields such as intelligent video surveillance, intelligent human-computer interaction, intelligent transportation, visual navigation, unmanned driving, unmanned autonomous flight, and battlefield situational reconnaissance. The purpose of target tracking is to predict the position of an object in subsequent frames based on its initial state. A significant difference between aerial tracking and general target tracking is that, due to the resource-constrained aerial platform, aerial tracking requires higher real-time speed and lower computational power. In addition, aerial tracking is affected by various challenging scenarios introduced by drones, such as fast camera movement, low-resolution images, and severely occluded targets. Based on the above characteristics of aerial tracking, a question naturally arises: can a good balance be found between efficiency and robustness, and can an efficient and effective aerial tracker be developed? Target tracking methods are mainly divided into two types: online trackers based on correlation filters and offline-trained trackers based on deep learning. Although the former has lower cost and lower energy consumption, its actual performance is still significantly lower than that of the latter, which utilizes offline training. However, the latter has very high energy consumption and computational costs. In order to achieve a satisfactory balance between performance and speed, researchers at home and abroad have tried to improve the performance of deep trackers while minimizing their redundancy to meet the conditions for their deployment and operation on terminals.

[0004] Therefore, how to rationally design and develop a target tracking system to maintain high tracking accuracy, strong algorithm robustness, strong real-time processing, and lightweight algorithm structure is of great significance for the implementation and promotion of target tracking systems in actual production and life.

[0005] A search revealed application publication number CN114820712B, which discloses an image method for accurate target tracking from the perspective of a UAV. The specific technical solution is as follows: an adaptive target bounding box optimization UAV tracking method, wherein the twin tracking network consists of a feature extraction backbone network module, an anchor point candidate module, a multi-scale feature extraction module, a feature fusion module, and an adaptive classification and regression module. On the one hand, this invention introduces a novel multi-scale channel attention mechanism—MSCA module—which improves the expressive power of template features used for anchor point generation. On the other hand, it introduces the idea of ​​deformable convolution, using anchor box information as an offset parameter to improve the information acquisition capability of the adaptive classification and regression network.

[0006] The MSCA module described in this patent primarily extracts scale features through a multi-scale channel attention mechanism to enhance the anchor accuracy of the anchor candidate module. However, the lack of attention to scale information in the feature fusion network may lead to an imbalance of feature information in the classification and regression branches during the backend classification and regression module calculations, thus limiting the accuracy and robustness of the final tracking results.

[0007] The deformable convolution module described in this patent uses anchor point information to guide the convolution process, aiming to guide the generation of candidate regions and improve the information richness of the input feature map of the classification network. However, the additional convolution branches increase the computational cost of the network, which to some extent limits the real-time performance of the tracking algorithm. At the same time, it also destroys the independence of the input feature maps of the classification and regression branches, limiting the performance improvement of the classification and regression networks.

[0008] This invention uses a scale-plus-channel attention module in the feature fusion network. The two modules extract feature maps sequentially and fuse them into subsequent feature maps. The scale attention module provides richer spatial scale information to the channel attention module, which greatly improves the information richness of the input to the classification branch.

[0009] In the anchor point proposal network, this invention uses multi-layer feature maps as input information and uses a cross-correlation attention module for calculation, which preserves shallow and deep feature information, and achieves the fusion of spatial structure, splicing scale features and attention features, while taking into account the consistency of context and multi-scale information.

[0010] This invention also improves the regression branch of the classification regression module by using the EIoU loss calculation method to calculate the error of the proposed anchor point, making the regression branch calculation more accurate and convergent faster. Summary of the Invention

[0011] This invention aims to solve the problems of the prior art. It proposes a target tracking method based on attention aggregation and anchor point proposal. The technical solution of this invention is as follows:

[0012] A target tracking method based on attention aggregation and anchor point proposal includes the following steps:

[0013] Multi-scale feature extraction is performed on the input image to obtain an expression with rich semantic information;

[0014] The extracted multi-scale features are input into the feature fusion network, where attention feature weighting and splicing, and cross-correlation operations between the template and the search region are performed.

[0015] Multi-layer features from the feature extraction network are input into the anchor proposal network, and a cross-correlation attention mechanism is used to obtain more scale-robust target anchor boxes.

[0016] The cross-correlation results of the feature fusion network are input into the multi-classification and regression networks, and the quality and score of the proposed anchor boxes are evaluated using a three-classification structure.

[0017] The regression branch then concatenates the proposed anchor data from the anchor proposal network with the results from the multi-classification and regression networks and determines the maximum value to obtain the most likely target tracking location box.

[0018] Furthermore, the network has two input image branches: a template branch and a search branch. The inputs to these two branches are the template image z and the search region image x, respectively. The two branch images undergo multi-layer convolution calculations to extract feature information at different scales.

[0019] Furthermore, the feature fusion network achieves the fusion and perception of the scale features of the target object based on scale and channel attention modules. The scale attention module SC-att searches for relevant scale information at the spatial level, while the channel attention module CH-att focuses on the scale information between different channels.

[0020] Furthermore, the Scale Attention Module (SC-att) processing flow is as follows:

[0021] Input features Feature information derived from multi-feature correlation processing in the preceding feature fusion module is used as input features. Feature information from APN:

[0022] The input feature maps are first processed through global average pooling and maximum pooling before aggregation to balance spatial information and highlight features; the feature information from the two branches undergoes a similar processing structure to obtain self-focused features. and Then, after concatenating with the residual features of the original input, scale-attention-weighted feature information is obtained; self-attention features The calculation process involves queries. key value ,value All three parts are derived from the input feature map. The resulting final scale-channel autocorrelation attention feature map The specific calculation formula is as follows:

[0023]

[0024] In the formula, Obtained through 1x1 scaled convolution flattening; autocorrelation attention feature map The final scale attention feature map is obtained by weighted summation. :

[0025] .

[0026] In the formula,

[0027] Furthermore, the channel attention module CH-att, based on the self-attention and cross-attention of a large-scale channel attention network, realizes the perception and processing of inter-channel scale information, specifically as follows:

[0028] First, input the feature map. After pooling, this step includes global average pooling. and maximum pooling Then, the data is aggregated, summed element-wise, and subjected to a linear combination of 1x1 fast-scale convolutions, followed by a flattening operation to obtain the query result. key value , The specific calculation process is as follows:

[0029]

[0030] In the formula Represents a linear combination of 1x1 fast-scale convolutions Channel self-attention maps can be obtained through 1x1 scale convolution and flattening. It can be obtained in the following ways:

[0031]

[0032] In the formula, s is superscripted to represent self-attention, and the self-attention feature map is obtained by weighted summation to obtain the final channel attention feature map. :

[0033]

[0034] In the formula,

[0035] Furthermore, the anchor point proposal network structure includes a cross-ANN cross-correlation attention module and a back-end convolutional module. The cross-correlation attention uses the outputs of the last two convolutional layers as anchor point proposals, which are convolved with the kernel to adjust the features in the backbone. Then, a scale-channel cross-correlation layer is used to collaboratively generate a similarity score map, and adaptive anchor points are calculated. The penultimate layer of the feature extraction network outputs template information and search region information, respectively. and After cross-correlation calculation, we obtain After passing through a convolutional layer, we get The penultimate layer outputs template information and search region information respectively. and After cross-correlation calculation, we obtain After passing through a convolutional layer, we get Cross-correlation attention mainly uses and As input information for anchor point generation, Represents shallow feature information. Representing deep feature information, it uses complementary feature information from two depths, where the query... and key value All from To maximize the preservation of the spatial structural details of the target, the value From .

[0036] Furthermore, the cross-attention graph of the cross-correlation attention The calculation method is as follows:

[0037]

[0038] In the formula, c represents the calculation of cross-attention. By giving , and The result of concatenation and cross attention By adding weights, adaptive adjustments to spatial structure, splicing scale features, and attention features are achieved, thus refining the features. The calculation formula is:

[0039] ,

[0040] after The convolutional layers generate scale-aware anchor points, which then enter the multi-classification and regression structures to participate in the quality score evaluation and selection of the anchor points.

[0041] Furthermore, the multi-classification and regression network inputs the cross-correlation results of the feature fusion network into the multi-classification and regression network, and uses a three-classification structure to evaluate the quality and score of the proposed anchor boxes, specifically including:

[0042] The classification results of multiple branches are combined by weighted average to obtain the final score, which represents the overall probability that each candidate box belongs to the target region.

[0043] The three classification branches output classification maps as follows: , , The different classification loss functions after integration are:

[0044]

[0045] In the formula, The quality of the proposed reaction anchor point The focus is on the classification score of the anchor point. The focus is on evaluating the centrality of positive samples, therefore and All use Cross-entropy loss function, use Binary cross-entropy loss function; , , These represent the weight coefficients of the three branches.

[0046] The regression branch outputs a regression feature map. , representing the regression offset of each point relative to the target box. The proposed anchor box is obtained by transforming the original center and offset. The loss function of the regression branch consists of two parts: first, the error loss on the predicted box, using EIoU loss, which directly compares the difference between the width and height of the predicted box and the ground truth box, corresponding to the weights. Secondly, the difference loss between the predicted bias and the regression target is calculated using Smooth L1 Loss, with corresponding weights... The weighted summation calculation process is as follows:

[0047]

[0048] The regression branch loss is the overall loss function as follows:

[0049]

[0050] In the formula, and The coefficients are for the weighted classification and regression branch losses.

[0051] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements a target tracking method based on attention aggregation and anchor point proposal as described in any one of the claims.

[0052] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the target tracking method based on attention aggregation and anchor point proposal as described in any one of the claims.

[0053] The advantages and beneficial effects of this invention are as follows:

[0054] This invention discloses a target tracking method based on attention aggregation and anchor point proposal networks, belonging to the field of computer vision and intelligent tracking technology. This method combines a Siamese feature extraction network, an adaptive anchor box generation network, a feature fusion network, and a multi-branch classification and regression network. By dynamically generating adaptive anchor boxes during target tracking, it improves the robustness of the model in low-resolution, fast-moving, and occluded scenarios. This invention extracts features from template and search images using a Siamese network, and the anchor point proposal network generates target-adaptive anchor boxes based on mid-to-deep features. Each anchor box achieves target position and scale adaptation through regression of four offsets. Furthermore, the feature fusion network deeply fuses the positional information output by the anchor point proposal network with high-level semantic features, providing richer spatial and semantic information. Finally, the multi-branch classification and regression network uses a three-branch classification structure and a single-branch regression structure to perform refined classification and bounding box regression of candidate anchor boxes. Experiments show that this method significantly improves target tracking performance in low-resolution, scale-changing, fast-moving, and occluded scenarios, and can be widely applied in UAV monitoring, video security, autonomous driving, and other visual perception fields.

[0055] This invention incorporates scale and channel attention modules into the feature fusion network, corresponding to the steps and formulas described in claim 2. During the feature fusion stage, the scale changes of the target can be further learned, and the perception and processing of scale learning between channels can be realized. This enriches the input information of the classification branch in the classification and regression network, and effectively improves the tracking accuracy and robustness in fast-moving object scenarios.

[0056] This invention incorporates a cross-correlation attention module into the anchor proposal network, corresponding to the steps and formulas described in claim 3. In the anchor proposal stage, the advantages of shallow and deep feature information can be leveraged to maximize the retention of semantic information of the proposed anchor while also taking into account its spatial perception capabilities, ultimately generating more accurate anchors and improving the tracking performance of the entire tracking method.

[0057] This invention improves the loss function of the regression branch in the classification regression network. Corresponding to the steps and formulas described in claim 4, the EIoU loss method is used for the error calculation of the prediction box, making the calculation and convergence of the entire regression branch more accurate and faster. Attached Figure Description

[0058] Figure 1 This is a framework diagram of a target tracking method based on attention aggregation and anchor proposal network according to a preferred embodiment of the present invention;

[0059] Figure 2 This is a diagram of the scale attention computation framework for feature fusion in this invention;

[0060] Figure 3 This is a diagram of the channel attention calculation framework for feature fusion in this invention;

[0061] Figure 4 This is a framework diagram of the anchor point proposal network in this invention; Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0063] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0064] To solve the above technical problems, such as Figure 1 As shown, this invention proposes a target tracking method employing attention aggregation and anchor proposal networks, comprising the following steps:

[0065] Multi-scale feature extraction is performed on the input image to obtain an expression with rich semantic information;

[0066] The extracted multi-scale features are input into the feature fusion network, where scale and channel attention are processed, and feature splicing and cross-correlation operations between the template and the search region are performed.

[0067] Multi-layer features from the feature extraction network are fed into the anchor proposal network, and scale-dependent proposed anchor boxes are obtained using the cross-correlation attention module.

[0068] The cross-correlation results of the multi-feature fusion network are input into the multi-classification and regression network, and the three-classification structure is used to determine the possible target location in the cross-correlation feature map.

[0069] The regression branch then concatenates the proposed anchor data from the anchor proposal network with the results from the multi-classification and regression networks and determines the maximum value to obtain the most likely target tracking location box.

[0070] The multi-scale feature extraction method is characterized by having two input image branches: a template branch and a search branch. The inputs to these two branches are a template image (denoted by z) and a search region image (denoted by x), respectively. The two branch images undergo multi-layer convolution calculations to extract feature information at different scales.

[0071] The feature fusion network method described above is characterized by fusing and perceiving the scale features of the target object based on scale and channel attention modules. Specifically, the scale attention module SC-att searches for relevant scale information at the spatial level, while the channel attention module CH-att focuses on the scale information between different channels.

[0072] The scale attention module method described above is characterized by integrating spatial information to lay the foundation for subsequent Ch-att. Different scales correspond to different relevance, allowing for further learning of the target's scale changes. The main processing flow is as follows: Figure 2 As shown, input features Feature information derived from multi-feature correlation processing in the preceding feature fusion module is used as input features. Feature information from APN:

[0073] The specific calculation process is as follows: the input feature maps are first processed by global average pooling and maximum pooling, and then aggregated to balance spatial information and highlight features. The feature information from the two branches is processed through a similar structure to obtain self-focused features. and This is then concatenated with the residual features of the original input to obtain scale-attention-weighted feature information. Self-attention features The calculation process involves queries. key value ,value All three parts are derived from the input feature map. The resulting scale-channel self-attention feature map The specific calculation formula is as follows:

[0074]

[0075] In the formula, Obtained through 1x1 scaled convolution flattening. The self-attention feature maps are then weighted and summed to obtain the final scaled attention feature map:

[0076]

[0077] In the formula,

[0078] The channel attention module method is characterized in that the proposed CH-att realizes the perception and processing of inter-channel scale information based on self-attention and cross-attention of a large-scale channel attention network. The main processing flow is as follows: Figure 3 As shown, input features Feature information from the previous stage of the feature fusion module after cross-correlation attention processing is used as input features. The feature information from the anchor proposal module after SC-att processing is used in the subsequent multiplication stage. Dark blue represents cross attention, and light blue represents self attention.

[0079] The specific calculation process is as follows: First, input the feature map. After pooling, this step includes global average pooling. and maximum pooling Then, the data is aggregated to balance spatial information and prominent features. Element-wise summation and a linear combination of 1x1 fast-scale convolutions are then performed to flatten the query result. key value , The specific calculation process is as follows:

[0080]

[0081] In the formula Represents a linear combination of 1x1 fast-scale convolutions Channel self-attention maps can be obtained through 1x1 scale convolution and flattening. It can be obtained in the following ways:

[0082]

[0083] In the formula, s is superscripted to represent self-attention, and the self-attention feature map is obtained by weighted summation to obtain the final channel attention feature map. :

[0084]

[0085] In the formula,

[0086] The anchor point proposal network method is characterized by a network structure comprising a cross-ANN cross-correlation attention module and a back-end convolutional module. The cross-correlation attention module uses the outputs of the last two convolutional layers as anchor point proposals, which are then convolved with the kernel to adjust the features in the backbone. Subsequently, a scale-channel cross-correlation layer is used to collaboratively generate a similarity score map, from which adaptive anchor points are calculated. (See below.) Figure 4 As shown. The penultimate layer of the feature extraction network outputs template information and search region information respectively. and After cross-correlation calculation, we obtain After passing through a convolutional layer, we get The penultimate layer outputs template information and search region information respectively. and After cross-correlation calculation, we obtain After passing through a convolutional layer, we get Cross-correlation attention mainly uses and As input information for anchor point generation, It represents shallow feature information, with high image resolution and rich details, but weak semantic information. Representing deep feature information, it has strong semantics and good anti-interference ability, but low spatial resolution. Using complementary feature information from two depths can maximize the preservation of semantic information while also taking into account spatial perception capabilities. (The query...) and key value All from To maximize the preservation of the spatial structural details of the target, the value From This ensures that the spatial location semantic information of the proposed data is more explicit and suppresses interference from negative background samples. In this way, the cross-correlation attention network can maintain spatial localization accuracy while obtaining richer semantic representation, thus generating more accurate anchor boxes.

[0087] Cross-attention graph of mutual attention The calculation method is as follows:

[0088]

[0089] In the formula, c represents the calculation of cross-attention. By giving , and The result of concatenation and cross attention By adding weights, adaptive adjustments are made to spatial structure, splicing scale features, and attention features, ensuring location information, strengthening the correlation between semantic information and scale, and simultaneously taking into account the consistency of context and multi-scale information, thus refining features. The calculation formula is:

[0090] ,

[0091] after Scale-aware anchors are generated through convolutional layers. These anchors then participate in the quality score evaluation and selection within multi-classification and regression structures.

[0092] The multi-classification and regression network method is characterized in that the classification results of multiple branches are combined through weighted summation to obtain the final score, which represents the comprehensive probability that each candidate box belongs to the target region. The advantages of multi-branch classification are: improved classification accuracy, as different branches focus on different classification dimensions, comprehensively considering the IoU, region location, and center distance of candidate boxes, reducing misclassification; handling imbalanced samples, as each branch classifies according to different sample distributions (such as IoU, intra-region points, and center distance), avoiding the insufficient performance of a single branch in imbalanced scenarios; and enhanced model robustness, as the collaboration between branches can more accurately distinguish between the target and the background when the target appearance is similar to the background.

[0093] The three classification branches output classification maps as follows: , , The different classification loss functions after integration are:

[0094]

[0095] In the formula, The quality of the proposed reaction anchor point The focus is on the classification score of the anchor point. The focus is on evaluating the centrality of positive samples, therefore and All use Cross-entropy loss function, use Binary cross-entropy loss function. , , These represent the weight coefficients of the three branches.

[0096] The regression branch outputs a regression feature map. , representing the regression offset of each point relative to the target box. The proposed anchor box is obtained by transforming the original center and offset. The loss function of the regression branch consists of two parts. The first is the error loss on the predicted box, using EIoU loss, which directly compares the difference between the width and height of the predicted box and the ground truth box, making the regression more accurate and converging faster. The corresponding weights... Secondly, the difference loss between the predicted bias and the regression target is calculated using Smooth L1 Loss, with corresponding weights... The weighted summation calculation process is as follows:

[0097]

[0098] The regression branch loss is the overall loss function as follows:

[0099]

[0100] In the formula, and The coefficients are for the weighted classification and regression branch losses.

[0101] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0102] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A target tracking method based on attention aggregation and anchor point proposal, characterized in that, Includes the following steps: Multi-scale feature extraction is performed on the input image to obtain an expression with rich semantic information; The extracted features are input into the feature fusion network, and scale and channel attention feature weighting and splicing, template and search region cross-correlation operations are performed. The shallow and deep features output by the feature extraction network are input into the anchor point proposal network, and the cross-correlation attention mechanism is used to obtain more scale-robust target anchor points. The cross-correlation results of the feature fusion network are input into the multi-classification and regression networks, and the quality and score of the proposed anchor boxes are evaluated using a three-classification structure. The regression branch then concatenates the proposed anchor data from the anchor proposal network with the results from the multi-classification and regression networks and determines the maximum value to obtain the most likely target tracking location box.

2. The target tracking method based on attention aggregation and anchor point proposal according to claim 1, characterized in that, The network has two input image branches: a template branch and a search branch. The inputs to these two branches are the template image z and the search region image x, respectively. The two branch images undergo multi-layer convolution calculations to extract feature information at different scales.

3. The target tracking method based on attention aggregation and anchor point proposal according to claim 1, characterized in that, The feature fusion network achieves the fusion and perception of the scale features of the target object based on scale and channel attention modules. The scale attention module SC-att searches for relevant scale information at the spatial level, while the channel attention module CH-att focuses on the scale information between different channels.

4. The target tracking method based on attention aggregation and anchor point proposal according to claim 3, characterized in that, The Scale Attention Module (SC-att) processing flow is as follows: Input features Feature information derived from multi-feature correlation processing in the preceding feature fusion module is used as input features. Feature information from APN: The input feature maps are first processed through global average pooling and maximum pooling before aggregation to balance spatial information and highlight features; the feature information from the two branches undergoes a similar processing structure to obtain self-focused features. and Then, after concatenating with the residual features of the original input, scale-attention-weighted feature information is obtained; self-attention features Queries involved in the calculation process key value ,value All three parts are derived from the input feature map. The resulting scale-channel self-attention feature map The specific calculation formula is as follows: In the formula, The scaled attention feature map is obtained by flattening the feature maps using 1x1 convolutions; the self-attention feature maps are then weighted and summed to obtain the final scaled attention feature map. : ； In the formula, 5. The target tracking method based on attention aggregation and anchor point proposal according to claim 4, characterized in that, The channel attention module CH-att, based on the self-attention and cross-attention of a large-scale channel attention network, realizes the perception and processing of inter-channel scale information, specifically as follows: First, input the feature map. After pooling, this step includes global average pooling. and maximum pooling Then, the data is aggregated, summed element-wise, and subjected to a linear combination of 1x1 fast-scale convolutions, followed by a flattening operation to obtain the query result. key value , The specific calculation process is as follows: In the formula Represents a linear combination of 1x1 fast-scale convolutions Channel self-attention maps can be obtained through 1x1 scale convolution and flattening. It can be obtained in the following ways: In the formula, s is superscripted to represent self-attention, and the self-attention feature map is obtained by weighted summation to obtain the final channel attention feature map. : In the formula, 6. The target tracking method based on attention aggregation and anchor point proposal according to claim 1, characterized in that, The anchor point proposal network structure includes a cross-ANN cross-correlation attention module and a back-end convolutional module. The cross-correlation attention uses the outputs of the last two convolutional layers as anchor point proposals, which are then convolved with the kernel to adjust the features in the backbone. Subsequently, a scale-channel cross-correlation layer is used to collaboratively generate a similarity score map, from which adaptive anchor points are calculated. The penultimate layer of the feature extraction network outputs template information and search region information, respectively. and After cross-correlation calculation, we obtain Then, after passing through a convolutional layer, we get... The penultimate layer outputs template information and search region information respectively. and After cross-correlation calculation, we obtain Then, after passing through a convolutional layer, we get... Cross-correlation attention mainly uses and As input information for anchor point generation, Represents shallow feature information. Representing deep feature information, it uses complementary feature information from two depths, where the query... and key value All from To maximize the preservation of the spatial structural details of the target, the value From .

7. The target tracking method based on attention aggregation and anchor point proposal according to claim 6, characterized in that, The cross-attention graph of the cross-correlation attention The calculation method is as follows: In the formula, c represents the calculation of cross-attention. By giving , and The result of concatenation and cross attention By adding weights, adaptive adjustments to spatial structure, splicing scale features, and attention features are achieved, thus refining the features. The calculation formula is: , after Scale-aware anchor points are generated through convolutional layers; It participates in the quality score evaluation and selection of anchor points in multi-class and regression structures.

8. The target tracking method based on attention aggregation and anchor point proposal according to claim 1, characterized in that, The multi-classification and regression network inputs the cross-correlation results of the feature fusion network into the multi-classification and regression network, and uses a three-classification structure to evaluate the quality and score of the proposed anchor boxes, specifically including: The classification results of multiple branches are combined by weighted average to obtain the final score, which represents the overall probability that each candidate box belongs to the target region. The three classification branches output classification maps as follows: , , The different classification loss functions after integration are: In the formula, The quality of the proposed reaction anchor point The focus is on the classification score of the anchor point. The focus is on evaluating the centrality of positive samples, therefore and All use Cross-entropy loss function, use Binary cross-entropy loss function; , , The weight coefficients representing the three branches; The regression branch outputs a regression feature map , which represents the regression offset of each point to the target box. The proposed anchor box is obtained by transforming the original center and offset; the loss function of the regression branch consists of two parts. First, it is the error loss of the predicted box. The EIoU loss is used to directly compare the differences in width and height between the predicted box and the ground truth box, corresponding to the weight ; second, it is the difference loss between the predicted offset and the regression target, using Smooth L1 Loss, corresponding to the weight . The calculation process of weighted summation is as follows: The regression branch loss is the overall loss function as follows: In the formula, and The coefficients are for the weighted classification and regression branch losses.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the target tracking method based on attention aggregation and anchor point proposal as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the target tracking method based on attention aggregation and anchor point proposal as described in any one of claims 1 to 8.

Images