Single target tracking method based on adaptive feature fusion modeling
Through adaptive feature fusion modeling and improved Transformer architecture, the accuracy and robustness problems of existing visual target tracking methods in complex scenes are solved, stronger target perception and interference avoidance are achieved, and tracking performance is improved.
Patent Information
- Application Number
- CN202510733075.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-04
- Publication Date
- 2025-10-17
AI Technical Summary
Existing visual target tracking methods have difficulty in clearly perceiving the target and avoiding the influence of interference when faced with target deformation, occlusion, and background changes, resulting in reduced tracking accuracy.
An adaptive feature fusion modeling method is adopted to enhance target perception and reduce the influence of distractors through adaptive kernel reorganization and dynamic feature aggregation. The Transformer architecture is improved by combining convolution operations and attention mechanisms for local and global feature modeling.
It improves the robustness and accuracy of target tracking, can maintain stable association and recognition of targets in complex scenarios, and significantly improves tracking performance.
Smart Images
Figure CN120807966A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of video single target tracking, and particularly relates to a single target tracking method based on adaptive feature fusion modeling. BACKGROUND
[0002] Visual target tracking refers to locating any target in a video sequence under the condition of only giving an initial appearance, and object deformation, occlusion, scale change and the influence of interfering objects bring challenges to the tracker to distinguish the target and the background. Currently, mainstream tracking methods adopt transformer-based architecture for object modeling, such as Figure 1 As shown in (a), a double-flow tracking framework respectively models self-attention for a template image and a search image, and then interacts through cross-attention. This modeling method may lead to blurred recognition and false matching in the target tracking process due to the lack of perception of the target in the early stage since the search image has not been fused and interacted with the template when performing self-attention operation. Figure 1 As shown in (b), a single-flow tracking framework learns joint features of the template and the search image through self-attention operation. Although this method avoids the problem that the search image cannot perceive the target, the cluttered search region may introduce interfering objects to the template. This mixed information may reduce the accuracy of target positioning, especially when the target is deformed, occluded or the background changes, the single-flow framework is easily disturbed by irrelevant information, which affects the tracking effect. SUMMARY
[0003] The main purpose of the application is to overcome the shortcomings in the prior art, so that the model can not only clearly perceive the target, but also avoid the influence of interfering objects as much as possible. The application takes OStrack as the benchmark tracking, and provides a single target tracking method based on adaptive feature fusion modeling, which uses adaptive kernel reorganization and dynamic feature aggregation to enhance target perception and reduce similar object interference.
[0004] The design concept of the application is that the modeling method of the application is as shown in (c). Figure 1As shown in (c), the present invention fuses the features of the template image and the search area image through adaptive weighting and recombination operations. This method preserves the inherent characteristics of the target while effectively incorporating background information in the search area, ensuring that the target maintains strong relevance despite deformation, occlusion, and complex background changes. Specifically, the model first performs adaptive feature recombination on the features, generating convolution kernels by aggregating contextual information, and then recombining the features of the template and search area. This adaptive feature recombination mechanism combines features by weight and automatically generates weights based on content, thereby better capturing the relationships between local features and preparing for subsequent fusion. To further improve the performance of the model, AFTrack introduces a dynamic feature aggregation module, which bridges the fusion information and the backbone network. It processes the recombined features of the template and search area at the spatial and channel levels, and finally dynamically adjusts the fusion strategy based on feature changes, allowing the model to better focus on the key features of the target. The fused features guide the final modeling, enhancing the expressiveness of the features and the robustness of the model. Furthermore, the Transformer excels in effectively modeling global context but lacks the ability to extract local features. To address this issue, this paper improves the Transformer architecture by combining convolution operations with an attention mechanism. This improves feature modeling capabilities and increases the focus on targets, enabling the network to better identify and track objects, especially with greater robustness in complex scenarios. This significantly improves target tracking performance and has broad application prospects.
[0005] The present invention is implemented through the following technical solutions: a single target tracking method based on adaptive feature fusion modeling, first, a training set is obtained from the video data set after data enhancement, and then a template image and a search area image are cropped from the training set, and the template image and the search area image are input into the fusion part for adaptive feature reorganization and dynamic feature aggregation. In the fusion part proposed by the present invention, the model aggregates context information to generate an adaptive kernel to reorganize the features of the template and the search area. This method effectively combines the inherent characteristics of the target with the information of the search area, guiding the model to maintain a stable association with the target in the face of challenges such as deformation and occlusion, clearly perceive the target and avoid interference; secondly, the reorganized features, template image and search area are input into the backbone network for feature extraction, and local and global feature modeling is performed; finally, the search area updated by the backbone network is sent to the tracking head to perform classification and regression tasks; and the classification results and regression results are combined to accurately locate the target. Specifically including the following steps:
[0006] S1. Crop template images from the enhanced training set and search area image By spatially segmenting and flattening, then linearly projecting and adding position encoding, the search region block embedding feature X0 and the initial template block embedding feature Z0 are obtained respectively;
[0007] S2, the cropped template picture and the search region picture are adaptively recombined and dynamically fused to obtain the final fused feature F0;
[0008] S3, the LTGFormer backbone network is composed of local blocks and global blocks, and the initial template block embedding feature Z0, the search region block embedding feature X0 and the final fused feature F0 are fed into the LTGFormer backbone network for joint feature learning and relationship modeling, and the output of the LTGFormer backbone network is taken as the input of the prediction head;
[0009] S4, the prediction head is composed of a classification branch and a regression branch, each branch includes a three-layer perceptron with hidden dimension d and a ReLU activation function; each vector of the feature map generated by the feature fusion network is predicted by the prediction head to obtain n=H z W z foreground / background classification results, and n=H z W z normalized coordinates of the search region, completing single target tracking based on adaptive feature fusion modeling.
[0010] Further, the step S1 includes the following steps:
[0011] S1-1, sample images from a video sequence (derived from a video dataset) to form training samples, and then process the training samples by using a data enhancement method (such as commonly used translation and brightness jitter) to expand the training set;
[0012] S1-2, in the first frame of the video sequence, the center of the target is taken as the center of the template picture, and the double length of the target side is taken as the side length of the template picture, the template picture is cropped into a square to obtain the template picture The content in the template picture includes the appearance information of the target and the surrounding local scene;
[0013] The center of the target in the previous frame is taken as the center of the search region picture, and the four times length of the target side is taken as the side length of the search region picture, the search region picture is cropped into a square to obtain the search region picture The content in the search region picture covers the possible movement range of the target;
[0014] S1-3, segment and flatten the template picture and the search region picture to obtain the initial template block sequence and search area block sequence in:
[0015]
[0016] In formula (1), N z Indicates the number of initial template blocks after the template image is segmented; in formula (2), N x Indicates the number of search area blocks after the search area image is divided; the resolution size of each image block is P×P;
[0017] S1-4. Apply the linearly projected block embedding layer to the initial template block sequence and the search area block sequence respectively, and obtain the initial template block embedding features and the search area block embedding features as follows:
[0018] Z0=P z W0+E x ; (3)
[0019] X0=P x W0+E z ; (4)
[0020] In formula (3) and formula (4), Z0 is the initial template block embedding feature, W0 is the learnable parameter of linear projection, X0 is the embedding feature of the search area block, D is the number of feature channels in the linear projection layer; E x and E z Both are encoding matrices with learnable positions, used to provide position prior information,
[0021] Furthermore, step S2 includes the following steps:
[0022] S2-1, template image Perform convolution operation and change the channel dimension to K 2 , get the image Z′;
[0023] Z′=Conv(Z); (5)
[0024] In formula (5),
[0025] Secondly, for image Z ′ Any position S = (a, b) is expanded at the channel to form a two-dimensional matrix, and the Softmax operation is used to generate the weight matrix W ′ S , as the adaptive kernel for subsequent calculations;
[0026] W ′S = Softmax(ψ(S, K)); (6)
[0027] In formula (6), ψ(S, K) represents flattening the channel dimension at position S into a KxK two-dimensional matrix;
[0028] S2-2, for multiple channels of input features, a channel-independent reorganization strategy is adopted, including the following steps:
[0029] First, taking position S=(a, b) as the center, a KxK size sub-region W SC of the cth channel is extracted, which is composed of the position S point and its surrounding pixel points, and is multiplied element by element with the adaptive weight matrix W ′ s generated at position S to generate a two-dimensional reorganized feature of the channel;
[0030] Secondly, the KxK region is compressed into a single channel feature value at position S by element-wise weighted summation;
[0031] Thirdly, after all channels complete the above operation in turn, the features of each channel are spliced along the channel dimension to form a complete reorganized feature vector at position S with a dimension of 1x1x3;
[0032] Finally, further reorganize all H z0 xW z0 space positions to integrate into a three-dimensional reorganized feature tensor with a dimension of H z0 xW z0 x3:
[0033]
[0034] In formula (7), Z''(S(a, b), c) represents the feature value of the cth channel after reorganization at position S of image Z ′ , represents the complete process of feature reorganization from S point feature reorganization to global three-dimensional feature construction of template picture Z;
[0035] S2-3, through adaptive feature reorganization, a reorganized template image Z
[0036] S2-4, repeat steps S2-2 to S2-3 above to reorganize the features of the search area picture to obtain a reorganized search area picture X
[0037] S2-5, reorganize the template image Z R and the reorganized search area picture XR Convolve the feature graph and add it element by element to get the preliminary fusion feature F:
[0038]
[0039] S2-6, the preliminary fusion feature F is calculated in parallel for channel attention and spatial attention. The channel attention consists of global average pooling and ReLU activation function. The first convolution layer is responsible for features to save calculation amount, and the second convolution layer is restored for feature dimension; spatial attention performs cross-channel splicing of feature maximum pooling and average pooling; the channel attention output is A c , the spatial attention output is A s ;
[0040]
[0041] S2-7, shuffle the channel: output A of the channel attention through the broadcast rule c and spatial attention output A s The features of are added and alternately arranged with the preliminary fusion feature F, and a 7×7 group convolution is applied. The final output is obtained by the Sigmoid function.
[0042] F mask =sigmoid(Conv 7×7 (Cs(concat(A c +A s )))); (11)
[0043] S2-8, F mask Multiply element-wise with the initial fusion feature F to obtain Finally, after 1×1 convolution and flattening, the final fusion feature is obtained
[0044]
[0045] F0=Conv 1×1 (F attn ). (13)
[0046] Furthermore, step S3 includes the following steps:
[0047] S3-1. The initial template block embedding feature Z0, the search area block embedding feature X0 and the final fusion feature F0 are linked as H0 = [Z0:X0:F0]. The re-parameterized convolution and feedforward neural network is used to extract local features and obtain feature H1:
[0048] H1=FFN(Rep3(H0))+H0;(14)
[0049] In formula (14), Rep3 represents a re-parameterized 3x3 convolution;
[0050] S3-2, in the global block, first, project H1 three times to obtain query, key and value; then, use H heads to perform matrix multiplication to extract features, each head is a matrix with a dimension of NxD / H, the attention matrix A i of the i-th head is:
[0051]
[0052] In formula (15), Q i and K i represent the query and key value of the i-th head respectively. At the same time, the V value is decomposed into H heads.
[0053] A=Concat(A1,…,A i ,…,A H );(16)
[0054] V=Concat(V1,…,V i ,…,V H );(17)
[0055] S3-3, multi-head self-attention MHSA calculation: by performing attention calculation, the model can focus on the most key information in the visual data;
[0056]
[0057] After a feedforward neural network and residual processing, the global feature extraction is completed:
[0058] H2=H1+MHSA(H1);(19)
[0059] H3=H2+FFN(H2);(20)
[0060] The updated feature H3 is taken as the input data of the next local block in the LTGFormer backbone network.
[0061] Further, the step S4 comprises the following steps:
[0062] S4-1, according to the position of the maximum value in the classification result, the approximate position of the target is located, and the position is mapped back to the search area picture to obtain the center position of the target; according to the position of the maximum value in the classification result, the corresponding offset value of the target up, down, left and right relative to the center position in the regression result is selected;
[0063] S4-2, if there is a tracking failure or the classification score is lower than 0.7, update the overall model by the gradient descent method;
[0064] S4-3, draw the coordinate frame of the tracking target according to the center position of the target and the offset value, and obtain the single target tracking result.
[0065] The application has the beneficial effects that: the application introduces a new feature fusion method in the target tracking technology, enhances the perception ability of the target and reduces the interference of similar objects, so that the network can better identify and track the target object. The adaptive feature reorganization module enhances the interaction between the search area and the template, the improvement of the transformer improves the modeling ability of the model, significantly improves the target tracking performance, and has a wide application prospect. At the same time, the application has strong robustness in complex scenes and can effectively deal with interference, which brings positive effects to the target tracking field. BRIEF DESCRIPTION OF DRAWINGS
[0066] Figure 1 Comparison diagram of double-flow tracking framework, single-flow tracking framework and fusion feature modeling tracking framework of the application; Figure 1 In the figure, (a) is a double-flow tracking framework, (b) is a single-flow tracking framework, and (c) is a fusion feature modeling tracking framework of the application;
[0067] Figure 2 The overall framework diagram of the application;
[0068] Figure 3 The adaptive feature reorganization flowchart in the application;
[0069] Figure 4 The dynamic feature fusion flowchart in the application;
[0070] Figure 5 The LTGFormer backbone network flowchart in the application;
[0071] Figure 6 The comparison curve diagram of the tracking effect of the application and other trackers on the LaSOT dataset; Figure 6 In the figure, (a) is a success rate curve diagram, and (b) is a precision rate curve diagram;
[0072] Figure 7 The comparison curve diagram of the tracking effect of the application and other trackers on the GOT10k dataset;
[0073] Figure 8 The tracking visualization comparison diagram of the tracking effect of the application and other trackers on the LaSOT dataset;
[0074] Figure 9This is a tracking visualization comparison diagram of the tracking effect of the present invention and other trackers on the LaSOT_ext dataset. DETAILED DESCRIPTION
[0075] The present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0076] like Figure 2 A single target tracking method based on adaptive feature fusion modeling is shown, comprising the following steps:
[0077] S1. Crop template images from the enhanced training set and search area image By segmenting and flattening the space, linearly projecting it, and adding position encoding, the search area block embedding feature X0 and the initial template block embedding feature Z0 are obtained respectively; the following steps are included:
[0078] S1-1, sample images from the video sequence to form training samples, and then use data enhancement methods to process the training samples and expand the training set;
[0079] S1-2, in the first frame of the video sequence, take the center of the target as the center of the template image, take twice the length of the target side as the side length of the template image, crop the template image into a square, and obtain the template image The content of the template image includes the appearance information of the target and its surrounding local scene;
[0080] The center of the target in the previous frame is used as the center of the search area image, and the side length of the search area image is four times the side length of the target. The search area image is cropped into a square to obtain the search area image. The content in the search area image covers the possible movement range of the target;
[0081] S1-3, split and flatten the template image and the search area image to obtain the initial template block sequence and search area block sequence in:
[0082]
[0083] In formula (1), N z Indicates the number of initial template blocks after the template image is segmented; in formula (2), N x Indicates the number of search area blocks after the search area image is divided; the resolution size of each image block is P×P;
[0084] S1-4. Apply the linearly projected block embedding layer to the initial template block sequence and the search area block sequence respectively, and obtain the initial template block embedding features and the search area block embedding features as follows:
[0085] Z0=P z W0+E x ; (3)
[0086] X0=P x W0+E z ; (4)
[0087] In formula (3) and formula (4), Z0 is an initial template block embedding feature, W0 is a learnable parameter of linear projection, X0 is a search region block embedding feature, D is the number of linear projection layer feature channels; E x and E z are encoding matrices with learnable positions, used to provide position prior information,
[0088] S2, the cropped template picture and the search region picture are subjected to adaptive feature reorganization (as shown in Figure 3 , used for feature reorganization of the template and the search region) and dynamic feature fusion (as shown in Figure 4 , used for sufficient fusion of the reorganized template and the search region) processing to obtain the final fusion feature F0; including the following steps:
[0089] S2-1, the template picture is subjected to convolution operation to change the channel dimension to K 2 , to obtain the picture Z';
[0090] Z'=Conv(Z); (5)
[0091] In formula (5),
[0092] Secondly, the picture Z ′ is unfolded at any position S=(a,b) in the channel to form a two-dimensional matrix, and a weight matrix W ′ S is generated by using Softmax operation as an adaptive kernel for subsequent calculation;
[0093]
[0094] In formula (6), ψ(S,K) represents that the channel dimension at position S is flattened into a K×K two-dimensional matrix;
[0095] S2-2, for multiple channels of input features, a channel-independent reorganization strategy is adopted, including the following steps:
[0096] First, extract the KxK size sub-region W corresponding to the cth channel with the position S=(a, b) as the center SC , which is composed of the position S point and its surrounding pixel points, and multiply it with the adaptive weight matrix W generated at the position S ′ s element by element to generate the two-dimensional reorganized feature of this channel;
[0097] Secondly, compress the KxK region into a single channel feature value at position S by element-wise weighted summation;
[0098] Thirdly, after all channels complete the above operation in turn, the features of each channel are spliced along the channel dimension to form a complete reorganized feature vector with a dimension of 1x1x3 at position S;
[0099] Finally, further reorganize all H z0 xW z0 space positions to integrate into a three-dimensional reorganized feature tensor with a dimension of H z0 xW z0 x3:
[0100]
[0101] In formula (7), Z″(S(a, b), c) represents the feature value of the cth channel of the image Z′ after reorganization at position S, which represents the complete process of feature reorganization from S point to global three-dimensional feature construction of the template picture Z;
[0102] S2-3, obtain the reorganized template image containing all positions and channels through adaptive feature reorganization
[0103] S2-4, repeat the above steps S2-2 to S2-3 to reorganize the features of the search area picture to obtain the reorganized search area picture
[0104] S2-5, convolve the feature maps of the reorganized template image Z R and the reorganized search area picture X R and add them element by element to obtain the preliminary fusion feature F:
[0105]
[0106] S2-6, the preliminary fusion feature F is subjected to parallel channel attention and spatial attention calculation, the channel attention is composed of global average pooling and ReLU activation function, the first convolutional layer is responsible for feature reduction to save calculation amount, and the second convolutional layer restores the feature dimension; the spatial attention is subjected to cross-channel splicing of feature maximum pooling and average pooling; the channel attention output is A c , and the spatial attention output is A s .
[0107]
[0108] S2-7, a channel shuffling operation is performed: the features of the channel attention output A c and the spatial attention output A s are added by broadcasting rules, and are alternately arranged with the preliminary fusion feature F, a 7x7 grouped convolution is applied, a final output is obtained by a Sigmoid function, and F
[0109] F mask =sigmoid(Conv 7×7 (Cs(concat(A c +A s )))); (11)
[0110] S2-8, F mask and the preliminary fusion feature F are multiplied element by element to obtain F After 1x1 convolution and flattening, the final fusion feature F
[0111]
[0112] F0=Conv 1×1 (F attn )。 (13)
[0113] S3, as shown in Figure 5 , the LTGFormer backbone network is composed of local blocks and global blocks, the initial template block embedding feature Z0, the search region block embedding feature X0 and the final fusion feature F0 are fed into the LTGFormer backbone network for joint feature learning and relationship modeling, and the output of the LTGFormer backbone network is taken as the input of the prediction head; including the following steps:
[0114] S3-1, the initial template block embedding feature Z0, the search region block embedding feature X0 and the final fusion feature F0 are linked as H0=[Z0:X0:F0], and a re-parameterized convolution and a feedforward neural network are used to extract local features to obtain a feature H1:
[0115] H1=FFN(Rep3(H0))+H0;(14)
[0116] In formula (14), Rep3 represents the re-parameterized 3×3 convolution;
[0117] S3-2. In the global block, first, H1 is projected three times to obtain query, key, and value; then, matrix multiplication is performed using H heads to extract features. Each head is a matrix with a dimension of N×D / H. The attention matrix A of the i-th head is i for:
[0118]
[0119] In formula (15), Q i and K i Represent the query and key value of the i-th head respectively. At the same time, the V value is decomposed into H heads.
[0120] A=Concat(A1,…,A i ,…,A H );(16)
[0121] V=Concat(V1,…,V i ,…,V H );(17)
[0122] S3-3. Multi-head self-attention MHSA calculation:
[0123]
[0124] After a feedforward neural network and residual processing, global feature extraction is completed:
[0125] H2=H1+MHSA(H1);(19)
[0126] H3=H2+FFN(H2);(20)
[0127] The updated feature H3 is used as the input data of the next local block in the LTGFormer backbone network.
[0128] S4, the prediction head consists of a classification branch and a regression branch, each branch includes a three-layer perceptron with a hidden dimension d and a ReLU activation function; the feature map generated by the feature fusion network is generated by the prediction head Predict each vector of n=H z W z foreground / background classification results, and n = H z W zThe normalized coordinates of the search area are obtained to complete single target tracking based on adaptive feature fusion modeling; the following steps are included:
[0129] S4-1. Locate the approximate position of the target based on the position with the largest median value in the classification results, map that position back to the search area image, and obtain the center position of the target. Select the corresponding offset values of the target in the regression results relative to the center position, up, down, left, and right, based on the position with the largest median value in the classification results.
[0130] S4-2. If there is a tracking failure or the classification score is lower than 0.7, the overall model is updated using the gradient descent method;
[0131] S4-3. Draw a coordinate frame of the tracking target according to the center position and offset value of the target to obtain a single target tracking result.
[0132] We conducted preliminary testing of the proposed fusion feature decoding structure using the PyTorch framework on an NVIDIA TITANRTX4090 server. The fusion feature decoding structure was implemented using Python programming in PyCharm.
[0133] In this embodiment, the success rate and precision on the LaSOT dataset are first tested, and then the comparison of the real tracking results in the LaSOT dataset is tested. The results are analyzed as follows.
[0134] Figure 6 The success rate curve in (a) shows the percentage of overlap scores (OS) between the tracking boxes drawn by the tracking algorithm and the manually labeled tracking boxes that is greater than a given threshold. Figure 6 The accuracy curve in (b) shows the Euclidean distance between the center point of the tracking algorithm prediction box and the center point of the GroundTruth box. The threshold is usually 20 pixels, that is, tracking is considered successful if their Euclidean distance is within 20 pixels. Figure 6 AFTrack is the name of the tracker proposed in this paper. OStrack, Mixformer, CTTrack, ToMP, Swin-Track, KeepTrack, TrDimp, LTMU, and TransT are the names of trackers proposed by other researchers in recent years. The numbers following them represent the average success rate and average precision of the tracker.
[0135] Depend on Figure 6 It can be seen that on the general dataset LaSOT, compared with the advanced trackers in recent years, the tracking method Ours of the present invention has obvious improvements in accuracy and success rate.
[0136] Figure 8 and Figure 9 The comparison of the tracking method of the present application and the real tracking results of advanced trackers such as OSTrack, Mixformer and TransT in the LSO and LaSOT_ext datasets in recent years is shown in the following table: Figure 8 and Figure 9 As can be seen from the above table, the bounding box generated by the tracking method of the present application is more accurate than other trackers when similar targets appear in the search area and when long-time tracking is performed; and the present application can also avoid the influence of interfering objects more than other trackers when encountering similar object interference.
[0137] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any changes or replacements easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A single target tracking method based on adaptive feature fusion modeling, characterized in that: The following steps are involved: S1. Crop template images from the enhanced training set and search area image By segmenting and flattening the space, linearly projecting it, and adding position encoding, we can obtain the search area block embedding feature X0 and the initial template block embedding feature Z0 respectively; S2. The cropped template image and search area image Perform adaptive feature recombination and dynamic feature fusion processing to obtain the final fusion feature F0; S3, the LTGFormer backbone network consists of local blocks and global blocks. The initial template block embedding feature Z0, the search area block embedding feature X0 and the final fusion feature F0 are fed into the LTGFormer backbone network for joint feature learning and relationship modeling. The output of the LTGFormer backbone network is used as the input of the prediction head. S4, the prediction head consists of a classification branch and a regression branch, each branch includes a three-layer perceptron with a hidden dimension d and a ReLU activation function; the feature map generated by the feature fusion network is generated by the prediction head Predict each vector of n=H z W z foreground / background classification results, and n = H z W z The normalized coordinates of the search area are obtained to complete single target tracking based on adaptive feature fusion modeling.
2. A single target tracking method based on adaptive feature fusion modeling according to claim 1, characterized in that: The step S1 comprises the following steps: S1-1, sample images from the video sequence to form training samples, and then use data enhancement methods to process the training samples and expand the training set; S1-2, in the first frame of the video sequence, take the center of the target as the center of the template image, take twice the length of the target side as the side length of the template image, crop the template image into a square, and obtain the template image The content of the template image includes the appearance information of the target and its surrounding local scene; The center of the target in the previous frame is used as the center of the search area image, and the side length of the search area image is four times the side length of the target. The search area image is cropped into a square to obtain the search area image. The content in the search area image covers the possible movement range of the target; S1-3, split and flatten the template image and the search area image to obtain the initial template block sequence and search area block sequence in: In formula (1), N z Indicates the number of initial template blocks after the template image is segmented; in formula (2), N x Indicates the number of search area blocks after the search area image is divided; the resolution size of each image block is P×P; S1-4. Apply the linearly projected block embedding layer to the initial template block sequence and the search area block sequence respectively, and obtain the initial template block embedding features and the search area block embedding features as follows: Z0=P z W0+E x ; (3) X0=P x W0+E z ; (4) In formula (3) and formula (4), Z0 is the initial template block embedding feature, W0 is the learnable parameter of linear projection, X0 is the embedding feature of the search area block, D is the number of feature channels in the linear projection layer; E x and E z Both are encoding matrices with learnable positions, used to provide position prior information, 3. The single target tracking method based on adaptive feature fusion modeling according to claim 1, characterized in that: The step S2 comprises the following steps: S2-1, template image Perform convolution operation and change the channel dimension to K 2 , get the image Z′; Z′=Conv(Z); (5) In formula (5), Secondly, any position S = (a, b) of the image Z′ is expanded at the channel to form a two-dimensional matrix, and the Softmax operation is used to generate the weight matrix W′ S , as the adaptive kernel for subsequent calculations; W′ S =Softmacx(ψ(S,K)); (6) In formula (6), ψ(S,K) represents the channel dimension at position S flattened into a K×K two-dimensional matrix; S2-2. For multiple channels of input features, a channel-independent recombination strategy is adopted, including the following steps: First, with position S = (a, b) as the center, extract the K×K size sub-region W corresponding to the cth channel SC , the sub-region consists of the pixel points at position S and its surroundings, and combines it with the adaptive weight matrix W′ generated at position S s Perform element-by-element multiplication to generate the two-dimensional reconstructed features of the channel; Secondly, the K×K region is compressed into a single channel feature value at position S by element-by-element weighted summation; Again, after all channels complete the above operations in sequence, the features of each channel are spliced along the channel dimension to form a complete recombined feature vector with a dimension of 1×1×3 at position S; Finally, further for all H z0 ×W z0 This operation is performed iteratively over the spatial positions, integrating into dimension H z0 ×W z0 ×3 three-dimensional reorganized feature tensor: In formula (7), Z”(S(a, b), c) represents the eigenvalue of the image Z′ after the cth channel is reorganized at position S. Represents the complete process of template image Z from S-point feature reconstruction to global 3D feature construction; S2-3, through adaptive feature recombination, obtain the recombined template image containing all positions and channels S2-4, repeat the above steps S2-2 to S2-3, and search the area image. The features of the reorganized search area are reorganized to obtain the reorganized search area image S2-5, reconstruct the template image Z R and reorganize the search area image X R Convolve the feature graph and add it element by element to get the preliminary fusion feature F: S2-6, the preliminary fusion feature F is calculated in parallel for channel attention and spatial attention. The channel attention consists of global average pooling and ReLU activation function. The first convolution layer is responsible for features to save calculation amount, and the second convolution layer is restored for feature dimension; spatial attention performs cross-channel splicing of feature maximum pooling and average pooling; the channel attention output is A c , the spatial attention output is A s ; S2-7, shuffle the channel: output A of the channel attention through the broadcast rule c and spatial attention output A s The features of are added and alternately arranged with the preliminary fusion feature F, and a 7×7 group convolution is applied. The final output is obtained by the Sigmoid function. F mask =sigmoid(Conv 7×7 (Cs(concat(A c +A s )))); (11) S2-8, F mask Multiply element-wise with the initial fusion feature F to obtain Finally, after 1×1 convolution and flattening, the final fusion feature is obtained F0=Conv 1×1 (F attn )。 (13) 4. The single target tracking method based on adaptive feature fusion modeling according to claim 1, characterized in that: The step S3 comprises the following steps: S3-1. The initial template block embedding feature Z0, the search area block embedding feature X0 and the final fusion feature F0 are linked as H0 = [Z0:X0:F0]. The re-parameterized convolution and feedforward neural network is used to extract local features and obtain feature H1: H1=FFN(Rep3(H0))+H0; (14) In formula (14), Rep3 represents the re-parameterized 3×3 convolution; S3-2. In the global block, first, H1 is projected three times to obtain query, key, and value; then, matrix multiplication is performed using H heads to extract features. Each head is a matrix with a dimension of N×D / H. The attention matrix A of the i-th head is i for: In formula (15), Q i and K i Represent the query and key value of the i-th head respectively. At the same time, the V value is decomposed into H heads. A=Concat(A1,…,A i ,…,A H ); (16) V=Concat(V1,…,V i ,…,V H ); (17) S3-3. Multi-head self-attention MHSA calculation: After a feedforward neural network and residual processing, global feature extraction is completed: H2=H1+MHSA(H1); (19)H3=H2+FFN(H2); (20) The updated feature H3 is used as the input data of the next local block in the LTGFormer backbone network.
5. The single target tracking method based on adaptive feature fusion modeling according to claim 1, characterized in that: The step S4 comprises the following steps: S4-1. Locate the approximate position of the target based on the position with the largest median value in the classification results, map that position back to the search area image, and obtain the center position of the target. Select the corresponding offset values of the target in the regression results relative to the center position, up, down, left, and right, based on the position with the largest median value in the classification results. S4-2. If there is a tracking failure or the classification score is lower than 0.7, the overall model is updated using the gradient descent method; S4-3. Draw a coordinate frame of the tracking target according to the center position and offset value of the target to obtain a single target tracking result.