A Target Tracking Method and System Based on Target Sequence Alignment and Prediction Box Modulation

Abstract

This invention proposes a target tracking method and system based on target sequence alignment and prediction box modulation. The method includes: constructing and initializing a target tracking model to obtain an initialized target tracking model; inputting a template image into the initialized target tracking model to obtain fixed template features; obtaining target prediction boxes and a response scoring map by searching the region image and template features; obtaining aligned template features using the response scoring map; updating the template based on the aligned template features, target prediction boxes, and fixed template features to obtain updated template features; using the updated template features as the current frame template features for the next frame and iterating to obtain continuous target prediction boxes, thus completing target tracking. This invention introduces a temporal memory mechanism based on dynamic template recursive fusion in the template update strategy, fundamentally different from the design of traditional target tracking methods that use a fixed-capacity memory or sliding window method to store historical template features.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a target tracking method and system based on target sequence alignment and prediction box modulation. Background Technology

[0002] Object tracking is a fundamental research task in computer vision. Its goal is to continuously and stably predict the position of a target in an image in subsequent frames, given the initial position of the target in the first frame of a video sequence. With the development of deep learning technology, object tracking methods based on convolutional neural networks and Transformer architectures have gradually become a research hotspot. In particular, Transformer-based methods, by representing images as a set of tokens and utilizing self-attention mechanisms to model global contextual relationships, exhibit strong feature modeling capabilities in complex scenes.

[0003] In existing Transformer-based target tracking frameworks, a template representation is typically constructed by feature encoding the target region in the first frame. In subsequent frames, the template features are matched with the search region features to achieve target localization. However, during continuous frame tracking, the target often undergoes displacement, scale changes, and viewpoint changes in the image space. Directly matching historical template features with current search features can easily lead to spatial misalignment between the template and search features, reducing the accuracy of feature matching and affecting the stability of the tracking results to some extent.

[0004] As the tracking process progresses, the target's appearance constantly changes due to factors such as illumination variations, deformation, and occlusion. Some existing methods incorporate search features from the current frame through online template update mechanisms to enhance the temporal representation capability of the template. However, in practical applications, without effective spatial constraints and region filtering mechanisms, features from background regions or interfering targets can easily be introduced into the template representation during the template update process. This causes the template to gradually deviate from the true target appearance, leading to tracking drift or tracking failure.

[0005] While existing methods typically output target prediction bounding boxes or response score maps during the target localization stage, these predictions are mostly used for the final target location regression and are not fully utilized in the template update or feature modulation stages. This results in a relative disconnect between the target's spatial location information and the feature update process, making it difficult to finely guide feature evolution at the token level. These problems are particularly pronounced in complex backgrounds, multi-target interference, or fast-moving scenarios, limiting the robustness and generalization ability of existing target tracking methods in practical applications. Summary of the Invention

[0006] In view of the above, the main objective of this invention is to propose a target tracking method based on target sequence alignment and prediction box modulation to solve the aforementioned technical problems.

[0007] This invention proposes a target tracking method based on target sequence alignment and prediction box modulation, the method comprising the following steps:

[0008] Step 1: Construct a target tracking model using a backbone feature extraction network and a head network that includes a template feature extraction branch and a template-search feature interaction branch;

[0009] Step 2: Load the pre-trained parameters into the target tracking model and initialize the backbone feature extraction network of the target tracking model to obtain the initialized target tracking model;

[0010] Step 3: Input the first frame template image into the initial target tracking model. Through the feature encoding process of the template feature extraction branch in the initial target tracking model, the initial template features are obtained and used as the fixed template features.

[0011] Step 4: Input the current frame search region image and the current frame template features into the initial target tracking model, and obtain the target prediction box and response score map of the current frame through template-search feature interaction branch processing;

[0012] Step 5: Based on the response scoring map, obtain the template features for the current frame alignment through target sequence temporal alignment processing;

[0013] Step 6: Update the template based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features for the current frame;

[0014] Step 7: Use the updated template features of the current frame as the template features of the next frame, repeat steps 4 to 6, and iterate the search region image of each subsequent frame to obtain a continuous target prediction box, so as to complete the tracking of the target in the video sequence.

[0015] This invention also proposes a target tracking system based on target sequence alignment and prediction box modulation, the system comprising:

[0016] The model building module is used for:

[0017] A target tracking model is constructed using a backbone feature extraction network and a head network that includes a template feature extraction branch and a template-search feature interaction branch;

[0018] The model initialization module is used for:

[0019] By loading pre-trained parameters into the target tracking model, the backbone feature extraction network of the target tracking model is initialized to obtain the initialized target tracking model.

[0020] The feature extraction and prediction module is used for:

[0021] The first frame template image is input into the initial target tracking model. The initial template features are obtained by feature encoding processing of the template feature extraction branch in the initial target tracking model. The initial template features are then used as fixed template features.

[0022] The current frame search region image and the current frame template features are input into the initial target tracking model. The target prediction box and response score map of the current frame are obtained through template-search feature interaction branch processing.

[0023] The sequence alignment module is used for:

[0024] Based on the response scoring map, the template features for the current frame alignment are obtained through temporal alignment of the target sequence.

[0025] The template update module is used for:

[0026] The template is updated based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features of the current frame;

[0027] The iteration and tracing module is used for;

[0028] The updated template features of the current frame are used as the template features of the next frame, and the search region image of each subsequent frame is iterated to obtain a continuous target prediction box, thus completing the tracking of the target in the video sequence.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1. This invention introduces a temporal memory mechanism based on dynamic template recursive fusion in the template update strategy, which is fundamentally different from the design of storing historical template features in traditional target tracking methods using a fixed-capacity memory bank or sliding window method. This approach takes into account both short-term appearance changes and long-term stability, and reduces the interference of introducing outdated features on the current tracking.

[0031] 2. This invention integrates historical dynamic template features into the current template representation frame by frame in a recursive manner. This eliminates the reliance on an explicit historical feature storage structure during template updates, instead creating a continuously evolving dynamic template representation. In this process, the discriminative information from the latest frame is continuously introduced, while the feature information from earlier frames naturally decays during multiple recursive updates. This effectively reduces the negative impact of distant historical frames on current tracking decisions and avoids target drift caused by long-term accumulated errors.

[0032] 3. This invention combines a temporal sequence alignment mechanism to perform spatial alignment processing on historical features before dynamic template recursive updates, ensuring a consistent target reference position in the feature space during the fusion process, thereby preventing feature misalignment and superposition problems caused by target movement or viewpoint changes. Through a segmentation and modulation mechanism guided by prediction boxes, target region constraints are applied to features participating in recursive fusion, making template updates more focused on the target itself and reducing the accumulation of background information in temporal memory.

[0033] 4. This invention introduces an implicit memory mechanism with dynamic recursive fusion and temporal decay characteristics, achieving continuous updating and effective preservation of target appearance information without explicitly building and maintaining a large-scale historical template library. This mechanism can gradually introduce the discrimination information of recent frames during the tracking process, while naturally weakening the influence of distant historical frames, thereby avoiding interference from redundant historical information on the current tracking decision. Attached Figure Description

[0034] Figure 1 This is a flowchart of the target tracking method based on target sequence alignment and prediction box modulation proposed in this invention;

[0035] Figure 2 This is a framework diagram of the target tracking method based on target sequence alignment and prediction box modulation proposed in this invention;

[0036] Figure 3 This is a schematic diagram of the template feature branch of the target tracking method based on target sequence alignment and prediction box modulation proposed in this invention;

[0037] Figure 4 This is a schematic diagram of the template-search feature interaction branch of the target tracking method based on target sequence alignment and prediction box modulation proposed in this invention.

[0038] Figure 5 This is a schematic diagram of the target tracking system based on target sequence alignment and prediction box modulation proposed in this invention. Detailed Implementation

[0039] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0040] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0041] Please see Figure 1 and Figure 2 This invention proposes a target tracking method based on target sequence alignment and prediction box modulation, which includes the following steps:

[0042] Step 1: Construct a target tracking model using a backbone feature extraction network and a head network that includes a template feature extraction branch and a template-search feature interaction branch.

[0043] For further details, please refer to Figure 3 The template feature extraction branch in the target tracking model constructed in this step includes a block segmentation module, a position encoding module, and a multi-layer attention feature extraction module connected in sequence.

[0044] Specifically, the template image is first divided into multiple non-overlapping or partially overlapping image blocks by a block segmentation module, and each image block is mapped to a corresponding feature representation. The location information is encoded and added to the image block features to obtain a block-level feature representation containing spatial location information.

[0045] The block-level features containing location information are sequentially input into a multi-layered stacked attention module for feature modeling. The attention module is used to model the global and local correlations between different image blocks, and gradually enhances the discriminative ability of template features through multi-layer feature interaction.

[0046] Through the forward computation of the above template feature extraction branch, the template feature representation corresponding to the template image is obtained. The template features include at least fixed template features that remain unchanged throughout the tracking process, which are used as the initial appearance representation of the target and provide a reference for subsequent search region feature matching and template updating.

[0047] Step 2: Load the pre-trained parameters into the target tracking model and initialize the backbone feature extraction network of the target tracking model to obtain the initialized target tracking model.

[0048] Step 3: Input the first frame template image into the initial target tracking model. Through the feature encoding process of the template feature extraction branch in the initial target tracking model, the initial template features are obtained and used as the fixed template features.

[0049] Step 4: Input the current frame search region image and the current frame template features into the initial target tracking model, and obtain the target prediction box and response score map of the current frame through template-search feature interaction branch processing;

[0050] In step 4, the current frame search region image and the current frame template features are input into the initial target tracking model. The target prediction box and response score map of the current frame are obtained through template-search feature interaction branch processing. The specific steps are as follows:

[0051] The current frame search region image and the current frame template features are input into the initial target tracking model, and the current frame search feature representation is obtained through the calculation of the template-search feature interaction branch;

[0052] The search feature representation of the current frame is input into the head network to obtain the target prediction box and response score map of the current frame.

[0053] For further details, please refer to Figure 4 The current frame search region image is input into the template-search feature interaction branch of the backbone feature extraction network for processing; specifically, the search region image is first feature encoded by the block segmentation module, and position information is introduced into the block-level features to obtain a search feature representation containing spatial position information.

[0054] Step 5: Based on the response scoring map, obtain the template features for the current frame alignment through target sequence temporal alignment processing;

[0055] In step 5, based on the response scoring map, the template features for the current frame alignment are obtained through temporal alignment of the target sequence. The specific steps are as follows:

[0056] Using the response scoring map, the position with the strongest response is determined as the center position of the target sequence in the current frame;

[0057] If a response score map of the previous frame exists, the relative displacement of the target sequence center of the adjacent frames is calculated, and the dynamic template features updated in the previous frame are spatially aligned based on the relative displacement of the target sequence center of the adjacent frames to obtain the template features aligned to the current frame; otherwise, the search features of the current frame are used as the template features aligned to the current frame.

[0058] If a response score map of the previous frame exists, the relative displacement of the target sequence center of adjacent frames is calculated. Based on the relative displacement of the target sequence center of adjacent frames, the dynamic template features updated in the previous frame are spatially aligned to obtain the template features aligned to the current frame. Conversely, if no response score map exists, the following relationship applies when using the search features of the current frame as the template features aligned to the current frame:

[0059] ;

[0060] in, This represents the relative displacement of the target sequence centers in adjacent frames. Indicates the center position of the target sequence in the current frame. Indicates the center position of the target sequence in the previous frame. The template features representing the alignment of the current frame. This indicates processing using a time-series alignment transformation operator based on relative displacement. This represents the template feature aligned to the previous frame. Indicates the search features for the current frame;

[0061] It should be noted that the aligned template features obtained in this step It also belongs to the dynamic template features of the previous frame after modulation.

[0062] Step 6: Update the template based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features for the current frame;

[0063] The template is updated based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features for the current frame. The specific steps are as follows:

[0064] A dynamic spatial mask is generated using the target prediction bounding box of the current frame, and then the dynamic spatial mask is converted into learnable mask features through an embedding mapping function;

[0065] Learnable mask features are used to fuse and modulate the template features aligned to the current frame to generate dynamic template features;

[0066] The fixed template features and the dynamic template features are concatenated along the sequence dimension to obtain the updated template features for the current frame.

[0067] In the process of converting a dynamic space mask into learnable mask features through an embedding mapping function, the following relationship exists:

[0068] ;

[0069] in, This represents the learnable mask features of the current frame. This indicates the embedding of mapping function operations. Indicates the dynamic space mask of the current frame;

[0070] In the process of fusing and modulating the template features aligned to the current frame using learnable mask features to generate dynamic template features, the following relationship exists:

[0071] ;

[0072] in, Indicates the dynamic template features of the current frame. This indicates processing via feature fusion operators;

[0073] In the process of concatenating fixed template features and dynamic template features along the sequence dimension to obtain the updated template features for the current frame, the following relationship exists:

[0074] ;

[0075] in, This represents the template features updated in the current frame. Indicates a fixed template feature. This indicates a splicing operation.

[0076] It should be noted that the overall expression for obtaining the dynamic template features of the current frame in this invention is:

[0077] ;

[0078] in, This indicates that the process is performed using a spatial segmentation modulation operator based on the target prediction box. This indicates processing using an alignment operator based on the temporal features of the target sequence;

[0079] In the overall expression, The specific processing procedure corresponds to the formula:

[0080] ;

[0081] In the overall expression, The specific processing procedure corresponds to the formula:

[0082] ;

[0083] .

[0084] Furthermore, through the template-search feature interaction process, the responsiveness of the search features to the target region is enhanced, and the distinction between the target and the background is improved. After the layer-by-layer feature interaction of the multi-layer attention module, a fused feature representation for target localization and template update is output, and the fused feature is passed to the template update module to generate dynamic template features for tracking in the next frame.

[0085] For further details, please refer to Figure 2The dynamic spatial mask guides the subsequent spatial segmentation and modulation process. It modulates features using aligned search features to highlight the feature response of the target region and suppress interference from the background region. The segmented and modulated features serve as dynamic template features, which are concatenated with the fixed template features in the sequence dimension to form updated template features.

[0086] The updated template features are fed back to the backbone feature extraction network for feature interaction and target prediction in the search region image of the next frame. This process is repeated to achieve recursive updating of template features and stable tracking of the target in the video sequence.

[0087] Step 7: Use the updated template features of the current frame as the template features of the next frame, repeat steps 4 to 6, and iterate the search region image of each subsequent frame to obtain a continuous target prediction box, so as to complete the tracking of the target in the video sequence.

[0088] In step 7, the updated template features of the current frame are used as the template features of the next frame. Steps 4 to 6 are repeated, and the search region image of each subsequent frame is iterated to obtain a continuous target prediction box, thereby completing the tracking of the target in the video sequence. The corresponding process has the following relationship:

[0089] ;

[0090] in, Indicates the search features for the next frame. This indicates that the core feature extraction network is used for processing. This indicates the search area image for the next frame. This indicates a template update operation guided by the target prediction bounding box. This represents the target prediction bounding box in the current frame. This represents the template features updated in the previous frame.

[0091] Furthermore, in this step, To utilize the target prediction box in the current frame and the template features updated from the previous frame , the overall expression for updating template features.

[0092] In this invention, in order to achieve effective training of the target tracking model, three loss constraints are introduced for the task objectives such as target discrimination, location localization and bounding box regression, and a joint loss function is constructed to optimize the model parameters.

[0093] Based on the cross-entropy loss form, the first loss function is constructed using the centerness response results output by the classification branch in the prediction head network and the corresponding ground truth labeling information. Its relationship can be expressed as:

[0094] ;

[0095] in, This represents the degree of deviation between the predicted result and the true label, i.e., the first loss function; The actual marker value representing the target location. This represents the probability value that the model predicts the location belongs to the target. Represents logarithmic operations;

[0096] A second loss function is constructed based on the generalized intersection-union ratio (GIoU), and its relational expression is as follows:

[0097] ;

[0098] in, This represents the degree of spatial matching between the predicted bounding box and the ground truth bounding box, i.e., the second loss function; A reference window representing the actual bounding box. This represents the reference window for the bounding box predicted by the model. This represents the generalized cross-union ratio (CUP) function calculated based on the cross-union ratio between the predicted window and the true window and its minimum bounding region.

[0099] The third loss function based on the center location regression error is expressed as follows:

[0100] ;

[0101] in, This represents the average absolute deviation between the predicted target center position and the actual target center position, i.e., the third loss function; This represents the target center coordinates obtained through the offset regression branch in the prediction head network. This represents the coordinates of the center position of the corresponding actual bounding box. Represents the L1 norm;

[0102] Based on the first, second, and third loss functions, the total loss function of the target tracking model is constructed, and its relational expression is as follows:

[0103] ;

[0104] in, Represents the total loss function. This represents the weighting coefficients of the second loss function. This represents the weighting coefficients of the third loss function.

[0105] During model training, the network parameters are continuously updated by minimizing the total loss function to achieve joint optimization of the target position, scale, and center coordinates. The target tracking model is trained by minimizing the total loss through updating the weights and learning parameters.

[0106] Please see Figure 5 This invention also provides a target tracking system based on target sequence alignment and prediction box modulation, the system comprising:

[0107] The model building module is used for:

[0108] A target tracking model is constructed using a backbone feature extraction network and a head network that includes a template feature extraction branch and a template-search feature interaction branch;

[0109] The model initialization module is used for:

[0110] By loading pre-trained parameters into the target tracking model, the backbone feature extraction network of the target tracking model is initialized to obtain the initialized target tracking model.

[0111] The feature extraction and prediction module is used for:

[0112] The first frame template image is input into the initial target tracking model. The initial template features are obtained by feature encoding processing of the template feature extraction branch in the initial target tracking model. The initial template features are then used as fixed template features.

[0113] The current frame search region image and the current frame template features are input into the initial target tracking model. The target prediction box and response score map of the current frame are obtained through template-search feature interaction branch processing.

[0114] The sequence alignment module is used for:

[0115] Based on the response scoring map, the template features for the current frame alignment are obtained through temporal alignment of the target sequence.

[0116] The template update module is used for:

[0117] The template is updated based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features of the current frame;

[0118] The iteration and tracing module is used for;

[0119] The updated template features of the current frame are used as the template features of the next frame, and the search region image of each subsequent frame is iterated to obtain a continuous target prediction box, thus completing the tracking of the target in the video sequence.

[0120] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0121] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0122] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A target tracking method based on target sequence alignment and prediction box modulation, characterized in that, The method includes the following steps: Step 1: Construct a target tracking model using a backbone feature extraction network and a head network that includes a template feature extraction branch and a template-search feature interaction branch; Step 2: Load the pre-trained parameters into the target tracking model and initialize the backbone feature extraction network of the target tracking model to obtain the initialized target tracking model; Step 3: Input the first frame template image into the initial target tracking model. Through the feature encoding process of the template feature extraction branch in the initial target tracking model, the initial template features are obtained and used as the fixed template features. Step 4: Input the current frame search region image and the current frame template features into the initialized target tracking model. The target prediction box and response score map for the current frame are obtained through template-search feature interaction branch processing. The specific steps are as follows: The current frame search region image and the current frame template features are input into the initial target tracking model, and the current frame search feature representation is obtained through the calculation of the template-search feature interaction branch; The search feature representation of the current frame is input into the head network to obtain the target prediction box and response score map of the current frame; Step 5: Based on the response scoring map, obtain the template features for the current frame alignment through temporal alignment of the target sequence. The specific steps are as follows: Using the response scoring map, the position with the strongest response is determined as the center position of the target sequence in the current frame; If a response score map of the previous frame exists, the relative displacement of the target sequence center of the adjacent frame is calculated, and the dynamic template features updated in the previous frame are spatially aligned based on the relative displacement of the target sequence center of the adjacent frame to obtain the template features aligned in the current frame. Conversely, the search features of the current frame are used as template features for alignment of the current frame, and the corresponding process has the following relationship: ； in, This represents the relative displacement of the target sequence centers in adjacent frames. Indicates the center position of the target sequence in the current frame. Indicates the center position of the target sequence in the previous frame. The template features representing the alignment of the current frame. This indicates processing using a time-series alignment transformation operator based on relative displacement. This represents the template feature aligned to the previous frame. Indicates the search features for the current frame; Step 6: Update the template based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features for the current frame. The specific steps are as follows: A dynamic spatial mask is generated using the target prediction bounding box of the current frame, and then the dynamic spatial mask is converted into learnable mask features through an embedding mapping function. The corresponding process has the following relationship: ； in, This represents the learnable mask features of the current frame. This indicates the embedding of mapping function operations. Indicates the dynamic space mask of the current frame; Learnable mask features are used to fuse and modulate the template features aligned to the current frame to generate dynamic template features. The corresponding process has the following relationship: ； in, Indicates the dynamic template features of the current frame. This indicates processing via feature fusion operators; The fixed template features and the dynamic template features are concatenated along the sequence dimension to obtain the updated template features for the current frame. Step 7: Use the updated template features of the current frame as the template features of the next frame, repeat steps 4 to 6, and iterate the search region image of each subsequent frame to obtain a continuous target prediction box, so as to complete the tracking of the target in the video sequence.

2. The target tracking method based on target sequence alignment and prediction box modulation according to claim 1, characterized in that, In the process of concatenating fixed template features and dynamic template features along the sequence dimension to obtain the updated template features for the current frame, the following relationship exists: ； in, This represents the updated template features of the current frame. Indicates a fixed template feature. This indicates a splicing operation.

3. The target tracking method based on target sequence alignment and prediction box modulation according to claim 2, characterized in that, In step 7, the updated template features of the current frame are used as the template features of the next frame. Steps 4 to 6 are repeated, and the search region image of each subsequent frame is iterated to obtain a continuous target prediction box, thereby completing the tracking of the target in the video sequence. The calculation expression for the search features of the next frame is: ； in, Indicates the search features for the next frame. This indicates that the process is performed through a backbone feature extraction network. This indicates the search area image for the next frame.

4. A target tracking system based on target sequence alignment and prediction box modulation, characterized in that, The system employs the target tracking method based on target sequence alignment and prediction box modulation as described in any one of claims 1 to 3, and the system comprises: The model building module is used for: A target tracking model is constructed using a backbone feature extraction network and a head network that includes a template feature extraction branch and a template-search feature interaction branch; The model initialization module is used for: By loading pre-trained parameters into the target tracking model, the backbone feature extraction network of the target tracking model is initialized to obtain the initialized target tracking model. The feature extraction and prediction module is used for: The first frame template image is input into the initial target tracking model. The initial template features are obtained by feature encoding processing of the template feature extraction branch in the initial target tracking model. The initial template features are then used as fixed template features. The current frame search region image and the current frame template features are input into the initial target tracking model. The target prediction box and response score map of the current frame are obtained through template-search feature interaction branch processing. The sequence alignment module is used for: Based on the response scoring map, the template features for the current frame alignment are obtained through temporal alignment of the target sequence. The template update module is used for: The template is updated based on the template features aligned to the current frame, the target prediction box of the current frame, and the fixed template features to obtain the updated template features of the current frame; The iteration and tracing module is used for; The updated template features of the current frame are used as the template features of the next frame, and the search region image of each subsequent frame is iterated to obtain a continuous target prediction box, thus completing the tracking of the target in the video sequence.

Images