Single target tracking method based on multi-scale feature fusion and channel attention mechanism

Abstract

The application discloses a single target tracking method based on multi-scale feature fusion and channel attention mechanism, and belongs to the technical field of deep learning and image processing, and the method comprises the following steps: acquiring an initial target template of a video sequence and a search area of a current frame; a template queue is constructed and maintained, and the template queue is used for storing multi-frame historical template data; multi-scale feature extraction is performed on template images in the template queue and images in the search area, and a multi-scale feature pyramid is constructed; a channel attention mechanism is used to perform weighted fusion on same-scale feature maps from the search area and the template queue; and based on the fused features, a score map ScoreMap of a target, a center point coordinate offset map OffsetMap and a target width and height information size map SizeMap are generated. The application solves the problem of linear expansion of memory by maintaining a fixed-size feature template queue and dynamically updating, and only retaining several latest and most relevant templates.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and image processing, and particularly relates to a single target tracking method based on multi-scale feature fusion and channel attention mechanism. BACKGROUND

[0002] In the field of single target tracking algorithm, its development process gradually evolves from the original discriminative method relying on hand-crafted features to the discriminative twin network tracking algorithm based on deep learning. In recent years, with the great breakthrough of the attention mechanism Transformer architecture in speech and text, this method has gradually achieved great success in the field of image processing, and has become the mainstream method in the field of single target tracking.

[0003] Due to the higher computational complexity of the Transformer architecture, when the input size of the image is too large, the inference speed of the model is seriously reduced, and when the input size is too small, the tracking performance of small targets is poor. In the case of long sequence input, the existing method model uses the initialized template for training, but lacks real-time dynamic updating of the template, causing the memory of the model to linearly expand with the increase of the frame number, which seriously affects the long-time tracking scene and is very easy to cause target loss. In addition, when the target is long-time occluded or completely occluded and then appears, the attention mechanism has unbalanced primary and secondary information, which is easy to regard the background with similar appearance as the target, causing permanent loss of the target. SUMMARY

[0004] One of the purposes of the present application is to provide a single target tracking method based on multi-scale feature fusion and channel attention mechanism to solve the problem of lack of real-time dynamic updating of the template and target loss in the prior art.

[0005] The present application is implemented by the following technical scheme, a single target tracking method based on multi-scale feature fusion and channel attention mechanism, comprising:

[0006] obtaining an initial target template of a video sequence and a search area of a current frame;

[0007] constructing and maintaining a template queue, wherein the template queue is used to store multi-frame historical template data;

[0008] performing multi-scale feature extraction on the template images in the template queue and the images in the search area to construct a multi-scale feature pyramid;

[0009] using a channel attention mechanism to perform weighted fusion on the same scale feature maps from the search area and each historical template in the template queue;

[0010] Based on the fused features, a score map ScoreMap of the target, a center point coordinate offset map OffsetMap and a target width and height information size map SizeMap are generated;

[0011] For each historical template in the template queue, a corresponding prediction score map is generated by inputting a prediction head network based on the corresponding fused features;

[0012] A predefined weighting coefficient is assigned to each template in the template queue, and the sum of all weighting coefficients is 1. According to the weighting coefficient of each template, the prediction score map corresponding to each template is fused to generate a weighted total score map;

[0013] Based on the weighted total score map, a final target bounding box is decoded;

[0014] According to a preset update condition, the template queue is dynamically updated.

[0015] The update condition includes:

[0016] The size deformation degree between the current frame target candidate box and the latest historical template in the template queue is calculated;

[0017] The maximum response value of the weighted total score map is calculated, and the maximum response value is taken as the target tracking confidence of the current frame;

[0018] When the size deformation degree is greater than a preset size deformation threshold, and the target tracking confidence of the current frame is less than a preset target confidence threshold, the queue template is updated, otherwise the current template queue remains unchanged.

[0019] Further, before updating the template queue, a reliability verification step is further included:

[0020] The average intersection over union of the new template and the target boxes of each template in the queue is calculated;

[0021] If the average intersection over union is greater than a preset threshold, the new template is considered reliable and is allowed to join the template queue;

[0022] If the average intersection over union is less than or equal to the preset threshold, the new template is discarded, and the template queue remains unchanged.

[0023] Further, the multi-scale feature extraction on the template images in the template queue and the images in the search area to construct a multi-scale feature pyramid includes:

[0024] The input image is transformed in three scales to form a three-layer feature pyramid:

[0025] Small target enhancement layer: the original image is enlarged to 2 times;

[0026] Original target layer: Preserves the original image size;

[0027] Large target adaptation layer: Reduces the original image to 0.5 times.

[0028] Furthermore, the step of utilizing channel attention mechanism to perform weighted fusion of feature maps of the same scale from the search region and each historical template in the template queue includes:

[0029] For the k-th layer features, the search region feature map and template feature map The pieces are stitched together along the channel dimension.

[0030] The concatenated features are input into an attention module consisting of convolutional layers and a sigmoid activation function to generate a channel weight vector. ;

[0031] The channel weight vector is used to analyze the feature map of the search region. and template feature map The weighted fusion formula is as follows: Where k takes values ​​of 1, 2, or 3. This indicates that the basic features of the template area should be preserved. This represents the fusion feature of the k-th layer.

[0032] Furthermore, the step of generating a target score map (ScoreMap), a center point coordinate offset map (OffsetMap), and a target width and height information size map (SizeMap) based on the fused features includes:

[0033] After the fused features are reshaped to a specified dimension, they are processed by three convolutional modules to output the corresponding score map (ScoreMap), center point coordinate offset map (OffsetMap), and target width and height information size map (SizeMap).

[0034] Further, decoding the final target bounding box based on the weighted total score map includes:

[0035] The maximum value point on the weighted total score graph is used as the target response center;

[0036] By combining the offset map of the center point coordinates corresponding to the center point of the target response and the size map of the target width and height information, the final target bounding box is obtained through coordinate decoding.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. The application provides a single target tracking method based on multi-scale feature fusion and channel attention mechanism, by maintaining a fixed size feature template queue and dynamically updating, only retaining the most recent and most relevant templates, solving the problem of linear expansion of memory, and changing the mode of traditional tracker relying on a single initial template, enabling the template to remember and utilize multiple states of the target in time sequence, thereby improving the feature extraction capability and anti-occlusion, anti-deformation capability, effectively preventing the introduction of error templates when the target is occluded (appearance changes dramatically) or tracking fails (low confidence), and at the same time, due to the constant amount of template memory, the stability and robustness of long-time tracking are improved.

[0039] 2. By performing multi-scale feature extraction on each template in the template queue and the image in the search area, a feature pyramid containing three scales is established, effectively enhancing the feature extraction capability for small targets.

[0040] 3. Through the channel attention mechanism, the search area features and the template features are adaptively weighted and fused, solving the redundancy and conflict problem of different scale features in the template and the search area. Through multi-template weighted fusion, multiple historical templates are jointly selected, even if the template matching of the current frame deviates, other correct historical templates can also make the result accurate through weighted fusion, thereby preventing permanent loss caused by single attention imbalance, enhancing the discrimination ability to background interference and similar objects, and when the target reappears after being occluded, the correct target can be quickly recaptured, reducing the tracking failure rate. BRIEF DESCRIPTION OF DRAWINGS

[0041] The drawings described herein are used to provide further understanding of the embodiments of the application, constitute a part of this application, and do not constitute a limitation on the embodiments of the application. In the drawings:

[0042] Figure 1 The schematic diagram of the single target tracking method based on multi-scale feature fusion and channel attention mechanism provided by the application.

[0043] Figure 2 The neural network architecture diagram for target tracking in the application.

[0044] Figure 3 The feature pyramid structure diagram in the application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below in conjunction with the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, not all the embodiments. The components of the embodiments of the application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0046] Existing Transformer-based tracking methods face serious computational efficiency challenges, mainly manifested as a sharp decline in inference speed when dealing with large-sized inputs, and poor tracking performance for small targets when dealing with small-sized inputs. The root cause of this problem lies in the quadratic computational complexity inherent in the Transformer self-attention mechanism, which grows in proportion to the square of the input sequence length. In the visual tracking task, the sequence length corresponds to the image resolution, and when dealing with high-resolution images, the computational burden becomes unbearable. At the same time, small targets only occupy a very small number of pixels in low-resolution inputs, and their feature representation ability is severely insufficient, making it difficult to provide enough discriminative information for accurate tracking.

[0047] At the same time, traditional trackers perform poorly in long-time tracking scenarios, specifically manifested as the inability to adapt to target appearance changes using fixed initial templates, and the linear expansion of model memory with the number of frames, ultimately leading to tracking failure. The technical root cause of this problem lies in the lack of effective dynamic template updating mechanisms and memory management strategies. Static templates cannot cope with real-world challenges such as changes in lighting, deformation, and occlusion, while the memory mechanism of the Transformer architecture indiscriminately accumulates all historical information, lacks the forgetting function of biological visual systems, and leads to the accumulation of redundant information, waste of computational resources, and reduced ability to focus on recent key information.

[0048] When faced with target occlusion recurrence or similar background interference, existing methods are prone to attention dispersion, incorrectly identifying background information as the target, leading to permanent tracking loss. The essence of this problem lies in the lack of targeted selection ability of the Transformer's global attention mechanism for visual features. It uniformly processes features at all spatial locations, and cannot effectively highlight key target features and suppress background interference in complex scenes. In particular, when the target is completely occluded, the model can only rely on historical memory for inference, and if similar interference patterns exist in the memory, an incorrect association will be established, disrupting the temporal consistency of tracking.

[0049] Embodiment 1

[0050] A single-target tracking method based on multi-scale feature fusion and channel attention mechanism is provided in this embodiment to address the following problems in the prior art:

[0051] 1. When the image input size is too large, the model's inference speed is severely reduced, and when the input size is too small, the tracking performance for small targets is poor. 2. Lack of real-time dynamic template updating, the model's memory linearly expands with the number of frames, which seriously affects long-time tracking scenarios and easily leads to target loss. 3. When the target is occluded for a long time or completely occluded and then reappears, the attention mechanism has an imbalance between primary and secondary information, and similar backgrounds are easily mistaken for targets, causing permanent loss of the target.

[0052] Figure 1 A total flowchart of the scheme in the embodiment is shown, which can specifically include the following steps:

[0053] Step 1, obtaining an initial target template of a video sequence and a search region of a current frame;

[0054] Step 2, constructing and maintaining a template queue, which is used to store multi-frame historical template data.

[0055] Firstly, an initial target template is obtained from the video sequence, and a template queue structure for storing historical template data is initialized. For each frame processing, the system extracts a search region centered on the predicted target position. The template queue is dynamically updated according to the preset update condition; by maintaining a fixed-size feature template queue and dynamically updating, only the most recent and most relevant templates are retained, solving the problem of linear expansion of memory, and changing the mode of traditional trackers relying on a single initial template, so that the template can remember and utilize multiple states of the target in time series, thereby improving the feature extraction capability and anti-occlusion and anti-deformation capability of the model, effectively preventing the introduction of error templates when the target is occluded (appearance changes dramatically) or tracking fails (low confidence). At the same time, since the template memory size is constant, the stability and robustness of long-time tracking are improved.

[0056] Specifically, the system receives a target bounding box manually labeled or detected in the first frame of the video sequence, takes the target bounding box as the center to cut the template region, copies the initial template three times to construct the initial template queue Q = [T0, T0, T0], and establishes a fixed memory window with a length of 3. Traditional single-template tracking is prone to failure when the target appearance changes, and multi-template can provide rich appearance change samples. A fixed-length (3 frames) template queue is maintained as a sliding window memory mechanism, which not only retains the multi-state information of the target, but also covers the recent state changes of the target and prevents infinite expansion of memory. The fixed queue length ensures the predictability of computing resources and avoids excessive templates leading to high computational complexity and memory conflicts.

[0057] Step 3, performing multi-scale feature extraction on the template images in the template queue and the images in the search region to construct a multi-scale feature pyramid;

[0058] Multi-scale feature extraction is performed on each template in the template queue and the image in the search area, a feature pyramid containing three scales is established, and scale invariance is achieved by processing the features of the three scales (2 times magnification, original size, 0.5 times reduction) in parallel. No matter how far or near the target is, it will always have a scale of feature layer that can effectively express it, which significantly improves the model's adaptability to scale changes. In particular, the small target enhancement layer increases the effective pixels of small targets in the feature map by 2 times magnification, thereby solving the problem of weak small target features and easy loss, and improving the model's robustness to target scale changes and small target tracking performance.

[0059] Specifically, as shown in Figure 2 , the template area image is scaled to control the target size, the target area is enlarged by 0.5 times and scaled, and the original target size is used to form a three-layer multi-scale feature pyramid, which improves the model's feature extraction capability for small targets:

[0060] Small target enhancement layer: realizes =0.5 times magnification, the target size is enlarged to twice, and the small target feature expression capability is enhanced;

[0061] Original target layer: retains the original size and retains the original features of the target;

[0062] Large target adaptation layer: performs =2 times down-sampling reduction, reduces the resolution of large targets in the feature map;

[0063] The template feature pyramid under the i-th frame is:

[0064] ,

[0065] where Resize represents adjusting the size, S1 and S2 represent scaling factors, represents the top-left corner coordinates and the width and height of the target box coordinates of the i-th historical template in the template queue of the t-th frame, and .

[0066] Step 4, using the channel attention mechanism, the same scale feature maps from the search area and the template queue are weighted and fused.

[0067] ​The channel attention mechanism is used for adaptively weighting and fusing the search region features and the template features, so that the redundancy and conflict problems of different scale features in the template and the search region are solved.

[0068] Specifically, as shown in the Atten Block module of the application, Figure 3 a weight vector is maintained to balance the same scale features of the template and the search region target, highlight the key feature channels of the target, and reduce background interference.

[0069] For the kth layer of the feature pyramid, the search region feature map and the template feature map are first spliced in the channel dimension to form a comprehensive feature representation, and then the spliced features are input into an attention module composed of a convolution layer and a Sigmoid activation function. The module generates a channel weight vector through learning, and each element value range [0, 1] represents the importance degree of the corresponding channel. Finally, the weight vector is used to weight and fuse the features according to the following formula to realize channel-level information filtering, and the formula is:

[0070]

[0071] wherein: represents the activation function sigmoid, || represents splicing, i.e. Concat operation, Conv Block represents convolution block operation, Avgpool represents average pooling operation, Maxpool represents maximum pooling operation, and the specific structure is shown in Figure 2

[0072] The kth layer fusion feature is:

[0073] wherein: k takes the value of 1, 2, 3, represents the basic feature of the template region.

[0074] Step 5, based on the fused features, a target score map ScoreMap, a center point coordinate offset map OffsetMap and a target width and height information size map SizeMap are generated.

[0075] ​​​First, the fused multi-scale features are reshaped to a unified specified dimension (such as 25x25x256) to adapt to subsequent standard convolution processing, and then the reshaped features are sequentially passed through three consecutive convolution modules (ConvBlock), each module containing a convolution layer, a batch normalization layer and a ReLU activation function. Finally, the network outputs three branches to generate: a score map ScoreMap representing the confidence of each position being the target center; a center point coordinate offset map OffsetMap used to compensate for the quantization error caused by downsampling; and a target width and height information size map SizeMap predicting the width and height of the target. This step provides accurate target position, confidence and size information, solving the problem of insufficient accuracy in mapping fused features to specific prediction tasks and limited positioning and size estimation accuracy.

[0076] Step 6, for each historical template in the template queue, a corresponding prediction score map is generated respectively.

[0077] To effectively utilize historical template information, the system performs an independent prediction process for each historical template in the template queue: first, each template and the current search area are respectively sent into the feature extraction network with shared weights, after the same multi-scale feature pyramid construction and channel attention fusion processing, the obtained fused features are respectively input into the prediction head network composed of three convolution modules, and finally each generates a corresponding prediction score map. The core purpose of this design is to preserve the appearance state features of the target at different historical moments through parallelized multi-template independent prediction, which provides diverse matching evidence for subsequent weighted fusion decision making, effectively utilizing the temporal change information of the target and significantly improving the decision reliability of the system in complex scenes such as target appearance change and local occlusion.

[0078] Step 7, fuse the prediction score maps corresponding to each template to generate a weighted total score map;

[0079] After obtaining the independent prediction score maps of each historical template, the system uses a weighted fusion strategy to generate the final total score map. Specifically, each template in the template queue is assigned a preset weighting coefficient W i , the sum of the coefficients is 1, and the newer templates are usually given higher weights (such as 0.4) and the older templates are given lower weights (such as 0.2), reflecting the difference in temporal importance. Then calculate the weighted total score map according to the formula S_total=W1M1+W2M2+W3M3, where: W1, W2, W3 represent the weighting coefficients of the first, second and third templates in the template queue, M1 represents the prediction score map of the first template, M2 represents the prediction score map of the second template, and M3 represents the prediction score map of the third template. This fusion method is essentially an ensemble learning strategy.

[0080] Multi-template fusion provides historical evidence when the target is occluded or similar objects interfere. The present application uses multi-template weighted fusion to make a collective decision among multiple historical templates. Even if the template matching of the current frame deviates, other correct historical templates can still make the result accurate through weighted fusion, thereby preventing permanent loss caused by single attention imbalance, enhancing the ability to distinguish background interference and similar objects, quickly recapturing the correct target when the target reappears after being occluded, and reducing the tracking failure rate.

[0081] Step 8, decoding the final target bounding box based on the weighted total score map.

[0082] After obtaining the weighted total score map, the system first locates the maximum response point in the map as a rough estimate of the target center, then reads the fine offset in the offset map corresponding to the center point position, reads the target size in the size map, and finally decodes to obtain the final target bounding box. The core reason for this design is to solve the quantization error problem caused by network downsampling and the smoothing size prediction constraint problem by separating response positioning, sub-pixel offset compensation, and size regression. Its important role is to achieve high-precision mapping from discrete score map to continuous image coordinates, achieving sub-pixel level positioning accuracy through offset compensation, and ensuring that the output bounding box meets the physical size constraints of the actual target through exponential function and prior size guidance, thereby maintaining stable and accurate tracking frame output in complex scenes.

[0083] Step 9, dynamically updating the template queue according to the preset update condition.

[0084] The traditional method lacks an effective template update mechanism, which can easily lead to tracking failure due to false updates. In the present application, the template queue is dynamically updated by a preset update condition, thereby improving the ability of the model to resist occlusion and deformation. As shown by Qt in Figure 1 , the update strategy is: when the size deformation of the target in the search area and the template in the template area is greater than the preset deformation threshold, and the target tracking confidence is less than the preset confidence threshold, do not update the template; otherwise, update the template queue.

[0085] Specifically, the size deformation degree of the current target candidate box and the latest historical template (i.e. the template of the frame) in the template queue is calculated, which is used to measure the magnitude of the target appearance change , and the calculation formula is:

[0086] ,

[0087] wherein: represents the left upper corner coordinate and the width-height coordinate of the target candidate frame in the t-th frame search area ;

[0088] Through the relative changes of width, height, area and quantization size deformation, the limitation of single dimension (such as width-height ratio) is avoided. The greater the target appearance change range is, the more significant the target size and shape change is.

[0089] When the update condition is: 1, the size deformation degree is greater than the preset size deformation threshold; 2, the target tracking confidence of the current frame is less than the preset target confidence threshold; the template update condition formula is:

[0090] ,

[0091] Among them: represents the size deformation threshold (if the threshold is exceeded, it is determined that the target appearance change is too large), which is currently set to 0.5, represents the confidence threshold of the target, and if it is lower than the threshold, it is determined that the current target is unreliable, and the current setting is 0.72, represents the tracking confidence of the tracking model to the current target candidate frame .

[0092] The size deformation threshold and the confidence threshold constitute the basic standard of system decision. By setting the size deformation threshold, normal deformation and abnormal mutation can be effectively distinguished; by setting the confidence threshold, a balance between tracking reliability and update frequency can be achieved, avoiding being too conservative or aggressive; by setting the intersection-over-union threshold, the new template and the historical target can have high spatial consistency, preventing target identity confusion.

[0093] In the above template update condition, when one of the two conditions is not established, the template queue is not updated, and the original template queue is continued to be used, and the formula is: ,

[0094] If both are established, the template is updated, and the update rule is as follows:

[0095] ,

[0096] Among them: represents the template queue of the t-th frame; represents the left upper corner coordinate and the width-height coordinate of the target frame coordinate of the i-th historical template in the t-th frame template queue, and is represented as: .

[0097] ​In the template updating decision process, the system performs reliability verification by calculating the average intersection over union (IoU) of the candidate new template and all existing template target boxes in the template queue, thereby improving the stability of tracking and preventing positioning deviation caused by single template error update. The update formula of a single template is:

[0098]

[0099] wherein, represents the current target candidate box; represents the top-left corner coordinates and the width and height of the i-th historical template in the template queue in the t-th frame. represents the updated single template.

[0100] By judging whether the intersection of the current frame template and the historical template is greater than a certain threshold, which is set to 0.8, when the average value is greater than the preset threshold 0.8, it is determined that the template is reliable and is allowed to join the queue, otherwise it is discarded.

[0101] The final multi-template weighted total score The calculation formula is:

[0102]

[0103] wherein, W1, W2 and W3 represent the weighting coefficients of the first, second and third templates in the template queue, respectively.

[0104] By setting a threshold to judge whether the weighted total score is the current correct tracking target, the purpose is to ensure the temporal continuity of the target identity through the quantitative evaluation of spatial overlap, and its important role is to establish an effective safety barrier, which can intelligently identify and filter abnormal templates caused by false detection, target mutation or severe occlusion, thereby preventing the template queue from being contaminated from the root cause, and significantly improving the stability and robustness of the system in the long-time tracking process.

[0105] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.​​​

Claims

1. A single-target tracking method based on multi-scale feature fusion and channel attention mechanism, characterized in that, include: Obtain the initial target template and the search region of the current frame in the video sequence; Construct and maintain a template queue, which is used to store multiple frames of historical template data; Multi-scale feature extraction is performed on the template images in the template queue and the images in the search area to construct a multi-scale feature pyramid; Using a channel attention mechanism, the feature maps of the same scale from the search region and each historical template in the template queue are weighted and fused respectively. Based on each fused feature, a score map (ScoreMap), a center point coordinate offset map (OffsetMap), and a target width and height information size map (SizeMap) are generated respectively. For each historical template in the template queue, based on the corresponding fusion features, it is input into the prediction head network to generate the corresponding prediction score map; Assign a predefined weighting coefficient to each template in the template queue, and the sum of all weighting coefficients is 1. Based on the weighting coefficients of each template, fuse the prediction score maps corresponding to each template to generate a weighted total score map. The process of fusing the predicted score maps corresponding to each template to generate a weighted total score map includes: Assign a preset weighting coefficient W to each template in the template queue. i The sum of the coefficients is 1. Newer templates are usually assigned higher weights, while older templates are assigned lower weights, reflecting the differences in the importance of time series. The weighted total score map is calculated according to the formula S_total=W1M1+W2M2+W3M3, where: W1, W2, and W3 represent the weighting coefficients of the first, second, and third templates in the template queue, respectively; M1 represents the predicted score map of the first template; M2 represents the predicted score map of the second template; and M3 represents the predicted score map of the third template. The final target bounding box is decoded based on the weighted total score map; The template queue is dynamically updated according to preset update conditions; The update conditions include: Calculate the degree of dimensional deformation between the current frame's target candidate box and the latest historical template in the template queue; Calculate the maximum response value of the weighted total score map, and use the maximum response value as the target tracking confidence of the current frame; The queue template is updated when the degree of size deformation is greater than the preset size deformation threshold and the target tracking confidence of the current frame is less than the preset target confidence threshold; otherwise, the current template queue remains unchanged.

2. The single-target tracking method based on multi-scale feature fusion and channel attention mechanism according to claim 1, characterized in that, Before updating the template queue, a reliability verification step is also included: Calculate the average intersection-union ratio (IoU) between the new template and the target boxes of all existing templates in the queue; If the average intersection-union ratio is greater than a preset threshold, the new template is considered reliable and allowed to be added to the template queue. If the average intersection-union ratio is less than or equal to a preset threshold, the new template is discarded, and the template queue remains unchanged.

3. The single-target tracking method based on multi-scale feature fusion and channel attention mechanism according to claim 1, characterized in that, The step of extracting multi-scale features from the template images in the template queue and the images in the search region, and constructing a multi-scale feature pyramid, includes: The input image is transformed at three scales to form a three-layer feature pyramid: Small target enhancement layer: Magnifies the original image by 2 times; Original target layer: Preserves the original image size; Large target adaptation layer: Reduces the original image to 0.5 times.

4. The single-target tracking method based on multi-scale feature fusion and channel attention mechanism according to claim 1, characterized in that, The method of utilizing channel attention to perform weighted fusion of feature maps of the same scale from the search region and each historical template in the template queue includes: For the k-th layer features, the search region feature map and template feature map The pieces are stitched together along the channel dimension. The concatenated features are input into an attention module consisting of convolutional layers and a sigmoid activation function to generate a channel weight vector. ; The channel weight vector is used to analyze the feature map of the search region. and template feature map The weighted fusion formula is as follows: Where k takes values ​​of 1, 2, or 3. This indicates that the basic features of the template area should be preserved. This represents the fusion feature of the k-th layer.

5. The single-target tracking method based on multi-scale feature fusion and channel attention mechanism according to claim 1, characterized in that, The process of generating a target score map (ScoreMap), a center point coordinate offset map (OffsetMap), and a target width and height information size map (SizeMap) based on the fused features includes: After the fused features are reshaped to a specified dimension, they are processed by three convolutional modules to output the corresponding score map (ScoreMap), center point coordinate offset map (OffsetMap), and target width and height information size map (SizeMap).

6. The method according to claim 1, characterized in that, Decoding the final target bounding box based on the weighted total score map includes: The maximum value point on the weighted total score graph is used as the target response center; By combining the offset map of the center point coordinates corresponding to the center point of the target response and the size map of the target width and height information, the final target bounding box is obtained through coordinate decoding.

Images