Target tracking method and system based on grouping attention feature extraction network

Abstract

The invention provides a target tracking method and system based on a grouped attention feature extraction network, and the method comprises the steps: obtaining a template and search features based on a template and a search image through employing the grouped attention feature extraction network; learning local areas of the template target image features and the search image features, and aggregating to obtain target information; splicing the template with the search features to obtain the input of the self-attention encoder in the asynchronous interaction stage; the fused features are combined with a prediction head for calculation, a classification regression result is obtained, and a target is predicted according to the classification regression result; training the target tracking model in combination with the large-scale data set to obtain a trained target tracking model, and performing accurate target tracking on the target by using the trained target tracking model; according to the multi-scale feature extraction network based on Transform, sequential local and global information interaction in the windows and between the windows is promoted, and the optimal balance between performance and speed is achieved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and image processing technology, and in particular to a target tracking method and system based on a grouped attention feature extraction network. Background Art

[0002] In the field of computer vision and image processing, visual tracking is the core part of many basic applications, such as autonomous driving, video surveillance, traffic vehicle monitoring, and human-computer interaction. Therefore, with the enhancement of the practicality and real-time performance of tracking technology, more and more tracking-related applications are becoming more and more common in real life, which makes the research of video tracking technology more valuable.

[0003] Generally speaking, visual tracking remains a challenging task due to many reasons such as appearance changes, deformation, fast motion, and occlusion, and designing a robust tracker remains a challenging problem. In recent years, visual trackers based on convolutional neural networks have been widely used in the field of visual tracking. Specifically, feature information based on the convolutional neural network model is introduced into the template and search branches, and one or more response maps obtained through the correlation operation are used to decode and extract information. Among them, typical convolutional neural network models include AlexNet, VGG, and ResNet. At present, most convolutional neural network models mainly use multiple convolution operators, which can effectively capture spatial local information, thereby modeling the global relationship of image content. However, this approach weakens the model's ability to obtain important global contextual information when dealing with complex scenes.

[0004] In addition, the use of Transformer has led to rapid development in the field of computer vision. For many computer vision tasks, such as target tracking, how to improve the model's expressiveness and performance is still a very important issue, and Transformer itself comes at the cost of low efficiency, so it is not desirable to improve performance at the cost of lower efficiency. How to effectively explore the context between consecutive frames without losing local information, and improve the performance of the tracker while adding only a few or no parameters, and design a highly robust tracker, has become an important issue in target tracking. Summary of the invention

[0005] In view of the above situation, the main purpose of the present invention is to propose a target tracking method and system based on a group attention feature extraction network to solve the above technical problems.

[0006] The present invention proposes a target tracking method based on a group attention feature extraction network, the method comprising the following steps: Step 1: Based on the Transformer network and the group attention module, a dual-branch feature extraction network is constructed. The two branches of the dual-branch feature extraction network share weights. A feature fusion network is constructed based on the self-attention mechanism. The dual-branch feature extraction network, the feature fusion network and the prediction head constitute the target tracking model. Step 2: Input the template image and the search image into the target tracking model in sequence, and extract features of the template image and the search image in sequence through the group attention mechanism to obtain template features and search features; Step 3: Use the cache technology to cache the extracted template features, and splice and interact with the extracted search features through the splicing template to obtain the splicing result, and perform feature fusion on the splicing result through the multi-head self-attention mechanism to obtain the fusion result; Step 4: Send the fusion result to the center head for prediction to obtain the center score map. According to the position with the highest confidence value in the center score map, the bounding box is calculated through the corresponding regression coordinates to obtain the final prediction result. Step 5, repeating steps 2 to 4 in an iterative form to train the target tracking model, and stopping the training when a preset number of times is reached to obtain a trained target tracking model; Step 6: The trained target tracking model updates the final prediction result, and uses the target candidate block with the maximum similarity score in the updated final prediction result to perform accurate target tracking.

[0007] The present invention also proposes a target tracking system based on a grouped attention feature extraction network, the system comprising: Building blocks for: A dual-branch feature extraction network is constructed based on the Transformer network and the group attention module. The two branches of the dual-branch feature extraction network share weights. A feature fusion network is constructed based on the self-attention mechanism. The dual-branch feature extraction network, feature fusion network and prediction head constitute the target tracking model. Extraction module for: The template image and the search image are sequentially input into the target tracking model, and the features of the template image and the search image are sequentially extracted through the group attention mechanism to obtain the template features and the search features; Fusion modules for: The extracted template features are cached using the cache technology, and are spliced ​​and interacted with the extracted search features through the splicing template to obtain the splicing result, and the splicing result is subjected to feature fusion through the multi-head self-attention mechanism to obtain the fusion result; Prediction module for: The fusion result is sent to the center head for prediction to obtain the center score map. According to the position with the highest confidence value in the center score map, the bounding box is calculated through the corresponding regression coordinates to obtain the final prediction result. Training modules for: Repeat the steps from the extraction module to the prediction module in an iterative form to train the target tracking model, and stop the training when the preset number of times is reached to obtain the trained target tracking model; Tracking module for: The trained target tracking model updates the final prediction result, and uses the target candidate block with the maximum similarity score in the updated final prediction result to perform accurate target tracking.

[0008] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention adopts a grouped attention method, which calculates the attention relationship between windows at a negligible cost through window marker embedding, establishes long-distance information interaction between multiple windows, captures more contextual information, and obtains better tracking performance.

[0009] 2. The present invention integrates the group attention technology into each encoder of the feature extraction module, so that the model can more accurately perform sequential local and global information interaction within and between windows when processing target images and search images, achieving the best balance between performance and speed.

[0010] 3. The present invention not only optimizes the feature extraction process, but also provides a more efficient and accurate solution for image recognition and search tasks.

[0011] Additional aspects and advantages of the present invention will be given in part in the following description, and in part will become apparent from the following description, or will be learned through embodiments of the present invention. Additional aspects and advantages of the present invention will be given in part in the following description, and in part will become apparent from the following description, or will be learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flowchart of the steps of the target tracking method based on the grouped attention feature extraction network proposed in the present invention.

[0013] Figure 2 This is a structural diagram of the target tracking framework in the target tracking method based on the grouped attention feature extraction network proposed in the present invention.

[0014] Figure 3 This is a schematic diagram of the overall group attention block in the target tracking method based on the group attention feature extraction network proposed in the present invention.

[0015] Figure 4 This is a schematic diagram of the grouped attention in the target tracking method based on the grouped attention feature extraction network proposed in the present invention.

[0016] Figure 5 This is a structural diagram of the target tracking system based on the grouped attention feature extraction network proposed in the present invention. DETAILED DESCRIPTION

[0017] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout are the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0018] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to provide some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0019] See also Figure 1 The embodiment of the present invention proposes a target tracking method based on a group attention feature extraction network, the method comprising the following steps: Step 1: Construct a dual-branch feature extraction network based on the Transformer network and the group attention module. The two branches of the dual-branch feature extraction network share weights. Construct a feature fusion network based on the self-attention mechanism. The dual-branch feature extraction network, feature fusion network and prediction head constitute the target tracking model.

[0020] In the step of building a dual-branch feature extraction network based on the Transformer network and the group attention module, the calculation process of the group attention module corresponds to the relationship: ; in, represents the feature connection operation, Indicates the number of layers, Representation layer normalization operation, represents the deep self-attention calculation, represents a feature slicing operation, Indicates that the self-attention mechanism can be separated. represents the multi-layer perceptron processing mechanism, represents the input features, represents the intermediate eigenvector, represents the feature vector of the final output, Represents the window characteristics, Represents the output of the previous layer.

[0021] It is necessary to add that The operation is used to concatenate the output of the previous layer with the window features to form a new input feature representation; operations to speed up training and stabilize models; The operation is used to slice the features after deep self-attention processing to obtain local features and window features; The operation is used to form residual connections, which facilitates information flow and gradient updates.

[0022] The calculation process of the group attention module includes deep self-attention mechanism processing and separate self-attention mechanism processing; Among them, the corresponding relationship between the deep self-attention mechanism processing process is: ; in, represents the input feature vector, Represents the calculation process of the self-attention mechanism, represents matrix multiplication, Indicates that it is used for production query The linear transformation matrix of Indicates the key used to generate the query The linear transformation matrix of Indicates the value used to generate the query The linear transformation matrix of Represents the input element currently being processed by the model. Representation and Related information; represents the normalization function; express The dimension of the vector, Indicates transpose.

[0023] In the attention mechanism, Used with Comparisons are made to determine the degree of correlation between different elements. Representation and The element to match; Contains The corresponding data or features. and When matching, the corresponding will be extracted and used to construct the output. The input feature vector is transformed by three different linear transformations (i.e., multiplication by , , )generate , and , the generated query, key, and value are fed into the self-attention mechanism to compute the output. In addition, when computing the attention weights, and A dot product operation is performed to determine the degree of attention of each input element to other elements. These weights go through a scaling step to avoid the problem of vanishing or exploding gradients, that is, multiplying the result of the dot product by , and finally passed After the operation, the result is compared with Perform a dot product operation.

[0024] The computation process of the grouped attention module includes deep self-attention mechanism processing and separate self-attention mechanism; Among them, the relationship between the separation of the self-attention mechanism processing process and the existence of the corresponding equation is: ; in, represents the exponential linear unit activation function, Represents window characteristics.

[0025] In the present invention, given the input features and window features, a weighted feature representation is obtained through the self-attention mechanism, combined with the nonlinear transformation of the window features and the features after layer normalization. Among them, the feature representation captures the relationship between different parts of the input features and enhances the model's understanding of the input data. The final output feature representation is used in subsequent network layers or directly for specific prediction tasks.

[0026] Step 2: Input the template image and the search image into the target tracking model in sequence, and extract features of the template image and the search image in sequence through the group attention mechanism to obtain template features and search features.

[0027] In step 2, the template image and the search image are sequentially input into the target tracking model, and the features of the template image and the search image are sequentially extracted through the group attention mechanism. The step of obtaining the template features and the search features specifically includes the following sub-steps: The template image is input into the target tracking model, and the template image is divided into different windows. Each window of the template image is processed by the deep attention mechanism to obtain the weighted feature representation of the template image. The weighted feature representation of the template image is processed by the separable self-attention mechanism to obtain the template feature. The search image is input into the target tracking model, and the search image is divided into different windows. Each window of the search image is processed by a deep attention mechanism to obtain a weighted feature representation of the search image. The weighted feature representation of the search image is processed by a separable self-attention mechanism to obtain the search feature.

[0028] For further information, please refer to Figure 2 , Figure 3 and Figure 4 , in step 2, feature extraction is performed through the grouped attention encoder. The core idea of ​​grouped attention is to decompose the traditional self-attention operation into two steps: the first step is deep self-attention, which applies self-attention independently on each input channel or feature to process local information, capture local features within the channel, but does not interact information across channels; the second step is point self-attention, which performs information fusion between channels and establishes connections between channels. The point self-attention operation is a supplement to deep self-attention. After capturing local features, the model further integrates global context information. Through this decomposition, grouped attention can reduce computational complexity, maintain or improve model performance, and enable the model to focus more effectively on the key parts of the data, thereby improving the accuracy and efficiency of predictions.

[0029] The specific process is as follows: perform a window division on the input feature map. Each window is an input channel of the feature map, and different windows contain different information. A window is created for each window as a global representation, which is used to model the attention relationship in the point self-attention module; perform deep self-attention on all pixels in each window and their corresponding window features, and the window operation fuses the spatial information in each channel; Then, separable self-attention processing is performed. Separable self-attention fuses information across windows and obtains the final representation of the input feature map. Feature maps and window features are extracted from the output of deep self-attention. The window features model the attention relationship between windows and use a single feature to encapsulate the core information of each sub-window. The window features are initialized to a fixed zero vector or a learnable vector of zero. When passing through the deep self-attention module, there is information interaction between the window features and pixels in each window, so the window features learn the global representation of this window and model the attention relationship between windows with negligible computational cost; after the layer normalization layer and activation function, the attention map is generated, and the feature map is regarded as a branch of separable self-attention. No other additional operations are required. Attention calculations are performed between windows in the form of attention maps and feature maps to exchange global information.

[0030] See also Figure 3, the overall architecture is a grouped attention module, whose structure is as follows: the input window features are used to capture global context information; layer normalization is applied before and after generating the attention map to ensure the stability and consistency of the network layer input; the key part of the model is the grouped attention mentioned above, which allows information interaction within the local window and across windows to capture local and global features; residual connections are added between the input and the processed output to promote gradient propagation and help the network learn; the data processed by the grouped attention module is further extracted and nonlinearly transformed by a multi-layer perceptron; the data after all the processing is used as the output of the model; By combining deep self-attention and point-by-point self-attention, local and global features in the input data can be effectively captured, thereby improving the performance of the model. Deep self-attention focuses on the interaction between different parts of the input data, while point-by-point self-attention focuses on the feature adjustment of a single data point. This combination enables the model to understand and process the input data more comprehensively.

[0031] Step 3: Use caching technology to cache the extracted template features, and splice and interact with the extracted search features through the splicing template to obtain the splicing result, and perform feature fusion on the splicing result through the multi-head self-attention mechanism to obtain the fusion result.

[0032] In step 3, the extracted template features are cached by using the cache technology, and are spliced ​​and interacted with the extracted search features through the splicing template to obtain the splicing result, and the splicing result is subjected to feature fusion through the multi-head self-attention mechanism to obtain the fusion result. The specific steps include the following sub-steps: In the template branch, the target template image is input into the group attention encoder, and the template features are extracted through the nine-layer stacked encoder to obtain the feature representation of the template image. After the template features are extracted, the last layer of features is cached to the memory; the cache technology is mainly used to store template features to reduce computational redundancy. The template is usually the initial frame in the video sequence and remains unchanged, and its features are also constant. Caching this feature can eliminate repeated calculations during testing.

[0033] The cache is implemented by first extracting the template features separately, and then extracting the search features. After the template features are extracted, in the subsequent asynchronous interaction stage, the last layer of features is spliced ​​and interacted with the search features at a specific layer. In the feature extraction stage, the template image and the search image are converted into token sequences through the same patch embedding layer. In their respective hierarchical attention operations, the template feature extraction uses a stacked encoder layer, and the search feature only focuses on itself in this stage. This method only stores the last layer of the template features to save memory. By caching the template features, the amount of calculation is reduced and the tracking efficiency is improved.

[0034] In the search branch, the target search image is input into the group attention encoder, and the search image features are extracted through a six-layer stacked encoder to obtain the feature representation of the search image; The feature representation of the template image is concatenated with the feature representation of the search image, and the features are fused through a multi-head self-attention mechanism to obtain a fusion result.

[0035] In the step of concatenating the feature representation of the template image with the feature representation of the search image and fusing the features through the multi-head self-attention mechanism to obtain the fusion result, the calculation process of the multi-head self-attention mechanism for feature fusion corresponds to the relationship: ; in, Represents an image from a template and search images The output of the self-attention calculation after the common feature vector input, Represents the common feature vector input of the template image and the search image, Represents the self-search image The key vector transpose and template image The key vector transpose The output vector concatenation of The query vector representing the search features, Indicates from search image The query vector is Indicates from search image The value vector of and template image The value vector of The output vector concatenation of express The dimensions of the vector.

[0036] Furthermore, the target template image is input into the grouped attention encoder for independent feature extraction, which is completed through a serialized stacked encoder with nine layers in total.

[0037] The search image is fed into the same encoder with six layers. In the feature extraction stage, the features of the search image only participate in their own group attention operation and do not interact with the template features.

[0038] In the feature fusion stage of asynchronous interaction operation, the output of the spliced ​​template and the search feature is used as the input of the encoder in the asynchronous interaction stage. In the feature fusion stage, the encoder only uses multi-head self-attention.

[0039] In the encoder of the structure, since the number of channels of the feature and the dimension of the converter label in this embodiment remain unchanged in the encoder, the template feature can interact with the search feature of any intermediate encoder layer. The present invention uses the ninth layer output of the template feature and the sixth layer output of the search feature to splice as the input of the feature fusion module. Using the cache technology, the template feature can be cached like the two-stage tracker in the test phase, thereby maintaining memory efficiency and reducing repeated calculations.

[0040] Step 4: Send the fusion result to the center head for prediction to obtain the center score map. According to the position with the highest confidence value in the center score map, the bounding box is calculated through the corresponding regression coordinates to obtain the final prediction result.

[0041] Furthermore, in step 4, the extracted target image features are fed into the center head for prediction. The center head consists of three convolution branches, which are used for center classification, offset regression, and size regression. The center classification branch outputs a center score map, where each score represents the confidence that the center of the target is located at the corresponding position. The prediction value of the offset regression branch is the discretization error for the center; the size regression branch predicts the height and width of the target. The position with the highest confidence value in the center score map is selected as the target position, and the corresponding regression coordinates are used to calculate a bounding box as the final prediction.

[0042] Step 5: Repeat steps 2 to 4 in an iterative manner to train the target tracking model. When the preset number of times is reached, the training is stopped to obtain the trained target tracking model.

[0043] In step 5, the target tracking model is trained in the following specific steps: Using the weighted focal loss function, the classification loss is defined for the tracking target of the tracking model. The classification loss calculation process has the following relationship: ; in, represents the weighted focal loss, Represents the weight coefficient for balancing positive and negative samples, represents the probability that the model predicts the positive category, represents the parameter that adjusts the model's attention to difficult and easy samples, Represents the natural logarithm with base e.

[0044] Weighted focal loss is a method to solve the class imbalance problem. It is mainly used in classification tasks by reducing the attention paid to easy-to-classify samples and increasing the attention paid to difficult-to-classify samples.

[0045] Using the global loss function, the target tracking model optimizes the prediction of the target position and calculates the loss of the bounding box regression. The calculation process of the loss of the bounding box regression has the following relationship: ; in, express of loss, represents the intersection-over-union ratio between the predicted bounding box and the true bounding box, represents the predicted bounding box, represents the true bounding box, represents the area of ​​the bounding box, represents the minimum closed area of ​​the predicted bounding box and the true bounding box, represents the diagonal length of the minimum enclosed area that encloses two bounding boxes, Represents a small constant.

[0046] Using the regression loss function, the distance between the predicted value and the true value of the bounding box regression loss is calculated. The calculation process has the following relationship: ; in, represents a commonly used regression loss, represents the true value, represents the predicted value, represents the number of samples, An index representing a single sample in the dataset.

[0047] Among them, the weighted focal loss function is used for classification. In terms of positioning, combined with loss and global loss are used as training objectives.

[0048] According to the calculated classification loss, bounding box regression loss and the distance between the predicted value and the true value of the bounding box regression loss, the overall loss is calculated. The calculation process of the overall loss has the following relationship: ; in, represents the overall loss, and Represents the weight coefficient.

[0049] It should be noted that Used to adjust the contribution of global loss in the total loss. This coefficient is used to control the relative importance of different loss terms; For adjustment The contribution of the loss to the total loss. This coefficient is used to control the relative importance of different loss terms.

[0050] Combined with large-scale datasets, steps 1 to 4 are iteratively repeated to train the target tracking model. The training datasets used are: GOT-10k, laSOT, COCO and TrackingNet, and the test datasets are: GOT-10k, TrackingNet, LaSOT, NFS, TNL2K, UAV123.

[0051] Step 6: The trained target tracking model updates the final prediction result, and uses the target candidate block with the maximum similarity score in the updated final prediction result to perform accurate target tracking.

[0052] See also Figure 5 , an embodiment of the present invention further provides a target tracking system based on a group attention feature extraction network, the system comprising: Building blocks for: A dual-branch feature extraction network is constructed based on the Transformer network and the group attention module. The two branches of the dual-branch feature extraction network share weights. A feature fusion network is constructed based on the self-attention mechanism. The target tracking model consists of the dual-branch feature extraction network, the feature fusion network and the prediction head. Building blocks for: A dual-branch feature extraction network is constructed based on the Transformer network and the group attention module. The two branches of the dual-branch feature extraction network share weights. A feature fusion network is constructed based on the self-attention mechanism. The dual-branch feature extraction network, feature fusion network and prediction head constitute the target tracking model. Extraction module for: The template image and the search image are sequentially input into the target tracking model, and the features of the template image and the search image are sequentially extracted through the group attention mechanism to obtain the template features and the search features; Fusion modules for: The extracted template features are cached using the cache technology, and are spliced ​​and interacted with the extracted search features through the splicing template to obtain the splicing result, and the splicing result is subjected to feature fusion through the multi-head self-attention mechanism to obtain the fusion result; Prediction module for: The fusion result is sent to the center head for prediction to obtain the center score map. According to the position with the highest confidence value in the center score map, the bounding box is calculated through the corresponding regression coordinates to obtain the final prediction result. Training modules for: Repeat the steps from the extraction module to the prediction module in an iterative form to train the target tracking model, and stop the training when the preset number of times is reached to obtain the trained target tracking model; Tracking module for: The trained target tracking model updates the final prediction result, and uses the target candidate block with the maximum similarity score in the updated final prediction result to perform accurate target tracking.

[0053] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0054] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0055] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A target tracking method based on a group attention feature extraction network, characterized in that: The method comprises the following steps: Step 1: Based on the Transformer network and the group attention module, a dual-branch feature extraction network is constructed. The two branches of the dual-branch feature extraction network share weights. A feature fusion network is constructed based on the self-attention mechanism. The dual-branch feature extraction network, the feature fusion network and the prediction head constitute the target tracking model. Step 2: Input the template image and the search image into the target tracking model in sequence, and extract features of the template image and the search image in sequence through the group attention mechanism to obtain template features and search features; Step 3: Use the cache technology to cache the extracted template features, and splice and interact with the extracted search features through the splicing template to obtain the splicing result, and perform feature fusion on the splicing result through the multi-head self-attention mechanism to obtain the fusion result; Step 4: Send the fusion result to the center head for prediction to obtain the center score map. According to the position with the highest confidence value in the center score map, the bounding box is calculated through the corresponding regression coordinates to obtain the final prediction result. Step 5, repeating steps 2 to 4 in an iterative form to train the target tracking model, and stopping the training when a preset number of times is reached to obtain a trained target tracking model; Step 6: The trained target tracking model updates the final prediction result, and uses the target candidate block with the maximum similarity score in the updated final prediction result to perform accurate target tracking.

2. The target tracking method based on the group attention feature extraction network according to claim 1 is characterized in that: In the step of building a dual-branch feature extraction network based on the Transformer network and the group attention module, the calculation process of the group attention module corresponds to the relationship: ； in, represents the feature connection operation, Indicates the number of layers, Representation layer normalization operation, represents the deep self-attention calculation, represents a feature slicing operation, Indicates that the self-attention mechanism can be separated. represents the multi-layer perceptron processing mechanism, represents the input features, represents the intermediate eigenvector, represents the feature vector of the final output, Represents the window characteristics, Represents the output of the previous layer.

3. The target tracking method based on the grouped attention feature extraction network according to claim 2 is characterized in that: The calculation process of the group attention module includes deep self-attention mechanism processing and separate self-attention mechanism processing; Among them, the corresponding relationship between the deep self-attention mechanism processing process is: ； in, represents the input features, Represents the calculation process of the self-attention mechanism, represents matrix multiplication, Indicates that it is used for production query The linear transformation matrix of Indicates the key used to generate the query The linear transformation matrix of Indicates the value used to generate the query The linear transformation matrix of Represents the input element currently being processed by the model. Representation and The elements to match, Representation and Related information, represents the normalization function, express The dimension of the vector, Indicates transpose.

4. The target tracking method based on the grouped attention feature extraction network according to claim 3 is characterized in that: The computation process of the grouped attention module includes deep self-attention mechanism processing and separate self-attention mechanism; Among them, the relationship between the separation of the self-attention mechanism processing process and the existence of the corresponding equation is: ； in, represents the exponential linear unit activation function, Represents window characteristics.

5. The target tracking method based on the grouped attention feature extraction network according to claim 4 is characterized in that: In step 2, the template image and the search image are sequentially input into the target tracking model, and the features of the template image and the search image are sequentially extracted through the group attention mechanism. The step of obtaining the template features and the search features specifically includes the following sub-steps: The template image is input into the target tracking model, and the template image is divided into different windows. Each window of the template image is processed by the deep attention mechanism to obtain the weighted feature representation of the template image. The weighted feature representation of the template image is processed by the separable self-attention mechanism to obtain the template feature. The search image is input into the target tracking model, and the search image is divided into different windows. Each window of the search image is processed by a deep attention mechanism to obtain a weighted feature representation of the search image. The weighted feature representation of the search image is processed by a separable self-attention mechanism to obtain the search feature.

6. The target tracking method based on the grouped attention feature extraction network according to claim 5 is characterized in that: In step 3, the extracted template features are cached by using a cache technology, and are spliced ​​and interacted with the extracted search features through a splicing template to obtain a splicing result, and the splicing result is subjected to feature fusion through a multi-head self-attention mechanism to obtain a fusion result. Specifically, the step includes the following sub-steps: In the template branch, the target template image is input into the group attention encoder, and the template features are extracted through the nine-layer stacked encoder to obtain the feature representation of the template image. After the template features are extracted, the last layer of features is cached into the memory; In the search branch, the target search image is input into the group attention encoder, and the search image features are extracted through a six-layer stacked encoder to obtain the feature representation of the search image; The feature representation of the template image is concatenated with the feature representation of the search image, and the features are fused through a multi-head self-attention mechanism to obtain the fusion result.

7. The target tracking method based on the grouped attention feature extraction network according to claim 6 is characterized in that: In the step of concatenating the feature representation of the template image with the feature representation of the search image and fusing the features through the multi-head self-attention mechanism to obtain the fusion result, the calculation process of the multi-head self-attention mechanism for feature fusion corresponds to the relationship: ； in, Represents an image from a template and search images The output of the self-attention calculation after the common feature vector input, Represents the common feature vector input of the template image and the search image, Represents the self-search image The key vector transpose and template image The key vector transpose The output vector concatenation of The query vector representing the search features, Indicates from search image The query vector is Indicates from search image The value vector of and template image The value vector of The output vector concatenation of express The dimensions of the vector.

8. A target tracking system based on a group attention feature extraction network, characterized in that: The system applies the target tracking method based on the group attention feature extraction network according to any one of claims 1 to 7, and the system comprises: Building blocks for: A dual-branch feature extraction network is constructed based on the Transformer network and the group attention module. The two branches of the dual-branch feature extraction network share weights. A feature fusion network is constructed based on the self-attention mechanism. The dual-branch feature extraction network, feature fusion network and prediction head constitute the target tracking model. Extraction module for: The template image and the search image are sequentially input into the target tracking model, and the features of the template image and the search image are sequentially extracted through the group attention mechanism to obtain the template features and the search features; Fusion modules for: The extracted template features are cached using caching technology, and spliced ​​and interacted with the extracted search features through the splicing template to obtain the splicing result, and the splicing result is fused through the multi-head self-attention mechanism to obtain the fusion result; Prediction module for: The fusion result is sent to the center head for prediction to obtain the center score map. According to the position with the highest confidence value in the center score map, the bounding box is calculated through the corresponding regression coordinates to obtain the final prediction result. Training modules for: Repeat the steps from the extraction module to the prediction module in an iterative form to train the target tracking model, and stop the training when the preset number of times is reached to obtain the trained target tracking model; Tracking module for: The trained target tracking model updates the final prediction result, and uses the target candidate block with the maximum similarity score in the updated final prediction result to perform accurate target tracking.

Images