Universal target tracking enhancement method based on implicit language guidance
By generating implicit language representations from visual features and performing cross-modal feature fusion, the problems of dynamic adaptation and computational overhead in visual-language tracking methods under complex scenarios are solved, achieving efficient and real-time target tracking enhancement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-04-28
AI Technical Summary
Existing visual-language tracking methods rely on static explicit language descriptions, which cannot dynamically adapt to changes in the target's motion state and environment. Furthermore, their performance degrades when language input is lacking, and they incur high computational overhead, failing to meet real-time tracking requirements.
The text inversion module dynamically generates patch-level fine-grained implicit language representations from visual features, and combines a multi-layer semantic injection mechanism to perform cross-modal feature fusion, generate pseudo-descriptions and improve tracking performance. A two-stage training strategy is adopted to optimize the model.
It can dynamically adapt to the target state without explicit language input, improving tracking robustness and accuracy in complex scenarios, with low computational overhead, adapting to multiple tracking frameworks, and meeting real-time tracking requirements.
Smart Images
Figure CN121937705A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, specifically to a general target tracking enhancement method based on implicit language guidance. By dynamically generating implicit language representations and deeply fusing them with visual features, it improves target tracking performance in complex scenes and can be widely applied in multiple fields such as intelligent monitoring, autonomous driving, robot navigation, video analysis, and human-computer interaction. Background Technology
[0002] Visual object tracking is one of the core foundational tasks in the field of computer vision. Its core objective is to continuously and accurately predict the position information of a target in subsequent video frames, given the initial state of the target in the first frame of a video (such as bounding box, mask, etc.). As a key supporting technology for many downstream applications such as intelligent monitoring and autonomous driving, the performance of object tracking directly affects the reliability and practicality of the entire system.
[0003] Traditional visual target tracking methods rely solely on visual cues in images (such as the target's color, texture, shape, and other appearance features) for tracking. Early Siamese network-based two-stream methods achieved target localization through template-search region similarity matching, while more recent single-stream methods have simplified the tracking process and improved performance by jointly encoding templates and search regions. However, these purely visual tracking methods often face significant challenges in complex scenes: when visual information is insufficient or ambiguous, such as tracking a specific sheep in a flock of sheep with highly similar appearances, when the target is severely occluded, or when the target undergoes drastic deformation or motion blur, purely visual tracking methods are prone to tracking drift or even complete tracking failure.
[0004] To overcome the limitations of purely visual tracking methods, researchers have proposed Vision-Language Tracking (VLT). This type of method uses natural language descriptions as additional auxiliary input, leveraging the rich high-level semantic information contained in natural language (such as target attributes, categories, behaviors, and positional relationships) to provide more discriminative guiding cues for target tracking, thereby effectively alleviating the tracking challenges in complex scenes. Thanks to the rapid development of natural language processing technology, VLT methods have achieved significant performance improvements.
[0005] However, existing vision-language tracking methods still suffer from two major limitations, which severely restrict their practical application scope and performance: Static language descriptions are mismatched with dynamic scenes: Existing methods rely on predefined static language descriptions, which cannot dynamically adapt to changes in the target's motion state and context. For example, when tracking a bear, the initial description "a polar bear playing in a pool" loses its effective guiding meaning after the bear leaves the pool and moves to a rock. This leads to a semantic misalignment between the language information and the target's actual state, which not only fails to assist tracking but may also mislead model predictions and reduce tracking performance. Strong dependence on explicit language input: Existing visual-language tracking methods rely on externally provided explicit text descriptions. When text input is lacking, these methods typically degenerate into pure visual tracking models, resulting in a significant drop in tracking performance or even complete failure. Although some research has attempted to address the semantic inconsistency problem by generating frame-specific descriptions online, manually annotated text descriptions are difficult to scale. On the other hand, using large language models (LLMs) to generate descriptions online introduces enormous computational overhead, significantly reducing tracking inference speed and failing to meet the needs of real-time tracking scenarios.
[0006] Furthermore, some existing methods attempt to mitigate the aforementioned problems by filtering inconsistent text or dynamically updating language references. However, these methods either lose valuable semantic information or further increase model complexity and inference latency, failing to fundamentally address the inherent defects of static explicit language descriptions. Therefore, there is an urgent need for a target tracking enhancement scheme that does not rely on explicit language input, can dynamically generate semantic guidance information adapted to the real-time state of the target, and has low computational overhead and high versatility, in order to overcome the bottlenecks of existing technologies. Summary of the Invention
[0007] To overcome the shortcomings of existing technologies, this invention provides a general target tracking enhancement method based on implicit language guidance. It dynamically generates patch-level fine-grained implicit language representations from visual features through a text inversion module, and achieves deep fusion of cross-modal features by combining a multi-layer semantic injection mechanism. Without the need for additional explicit text input, it significantly improves the robustness and accuracy of target tracking in complex scenarios, while ensuring low computational overhead and good versatility.
[0008] The technical solution adopted by this invention to solve its technical problem is: A general target tracking enhancement method based on implicit language guidance includes the following steps: Step 1) Input data preprocessing: Obtain template data and search images, and convert them into processable visual embedding vectors through a patch embedding layer; Step 2) Implicit Language Representation Generation: The template data and the visual embedding vector of the search image are concatenated into a unified feature representation, which is then input into the text inversion module to generate a patch-level fine-grained pseudo-description (implicit language representation). Step 3) High-level semantic language feature extraction: Input the pseudo-description into the frozen CLIP text encoder to extract high-level semantic language features; the Transformer layer of the CLIP text encoder is uniformly divided into N text blocks; Step 4) Multi-scale visual feature extraction: The visual features after patch embedding are extracted at multiple scales through the visual backbone network to obtain hierarchical visual features from shallow to deep layers. Step 5) Multi-layer semantic injection fusion: Through a multi-layer semantic injection mechanism, the semantic language features of each level are fused with the visual features of the corresponding level across modalities to generate enhanced visual features; Step 6) Target location prediction: Input the enhanced visual features into the prediction head and output the bounding box coordinates of the target in the search image; Step 7) Model training: A two-stage training strategy is adopted. The first stage is to pre-train the visual feature extraction network, and the second stage is to fine-tune the newly added modules to ensure model performance and training efficiency.
[0009] Furthermore, in step 1), the template data is visual data containing the historical state of the target, which can be selected as a continuous template video segment (containing N frames of images) or N discrete template images. The initial state of the target is determined by the manually labeled bounding box of the first frame. The search image is a single frame image to be used for target position prediction. The patch embedding layer consists of a convolutional layer and an MLP layer, which divides the template data and the search image into fixed-size image patches and converts them into high-dimensional dense visual embedding vectors.
[0010] Furthermore, in step 2), the visual embedding vector of the template data and the visual embedding vector of the search image are first concatenated along the spatial dimension to form a unified feature representation that includes the target's historical features and current context features. Then, the unified feature representation is input into the text inversion module, and the visual features are mapped to the CLIP text embedding space through projection, MLP, and alignment operations in sequence. The final generated pseudo-description is a patch-level fine-grained representation, with each image patch corresponding to a text token, which can faithfully capture local visual details and dynamically reflect the target's real-time state and environmental changes through global context interaction.
[0011] Furthermore, in step 3), a pre-trained clip-vit-large-patch14 model is used as the CLIP text encoder, and the parameters are kept frozen during training and inference. The stacked multi-layer Transformer of the CLIP text encoder is evenly divided into multiple text blocks (T-Blocks), and each text block contains multiple layers of Transformers. The pseudo-description is processed layer by layer through multiple text blocks to output high-level semantic language features at multiple levels. The language features at each level are consistent with the visual feature dimensions, laying the foundation for cross-modal fusion.
[0012] Furthermore, in step 4), the visual backbone network can adopt a Transformer-based network structure, such as Fast-iTPN or standard ViT; its Transformer layers are evenly divided into multiple visual blocks (V-Blocks), consistent with the number of text blocks; after the unified feature representation is input into the visual backbone network, it sequentially extracts features layer by layer through multiple visual blocks to obtain multi-level, multi-scale visual features. Among them, shallow features contain low-level information such as texture and edges, while deep features contain abstract information such as shape and category.
[0013] Furthermore, in step 5), the multi-layer semantic injection fusion process is as follows: First, the dimensionality adjustment and distribution alignment of the language features at the i-th level are performed through the alignment module; then, the visual features and the aligned language features are subjected to layer normalization processing respectively; next, the normalized language is used as the query, and the normalized visual features are used as the key and value, and the semantic association weight cross attention is calculated through the multi-head cross attention unit; then, the cross attention is multiplied by the learnable scaling factor and residually connected with the visual features to obtain the enhanced visual features; finally, the enhanced visual features are used as the input of the next layer visual block, and the above process is repeated to complete the fusion of all levels.
[0014] Furthermore, in step 6), the final level of enhanced visual feature input prediction head decodes the features output by the backbone network, and the bounding box regression branch outputs the target bounding box coordinates to achieve target tracking of a single frame image.
[0015] Furthermore, in step 7), the two-stage training strategy is as follows: In the first stage, the visual backbone network and prediction head are trained on a large-scale visual tracking dataset. The training parameters, such as the number of training epochs, learning rate, and learning rate decay, are adjusted according to the different visual backbone network and prediction head. The training ends when the loss converges and stabilizes, enabling the visual backbone network to learn high-quality visual features. In the second stage, the visual backbone network and CLIP text encoder are frozen, and only the text inversion module, semantic injection module, and prediction head are trained. This stage is preparation for a general target tracking enhancement method. Training is performed on the visual tracking dataset with 10 training epochs and an initial learning rate of 0.001. The learning rate is decayed after the 6th training epoch. The loss function is the IoU loss and other commonly used losses for visual tracking tasks, consistent with the training in the first stage. It can be adjusted according to the visual backbone network and prediction head to achieve optimal results.
[0016] The beneficial effects of this invention are mainly reflected in: 1. Breaking the dependence on explicit language input: Implicit language representations are dynamically generated directly from visual features through the text inversion module, without relying on any external explicit text input. This fundamentally solves the performance degradation problem of existing visual-language tracking methods when there is a lack of language input, and expands the application scenarios. 2. Dynamically adapt to target and scene changes: The generated pseudo-description is a patch-level fine-grained representation that can accurately capture local visual details of the target. At the same time, through the global context interaction between the template and the search image, it reflects the target's motion state and environmental changes in real time, effectively avoiding semantic misalignment caused by static language description and improving tracking robustness in complex dynamic scenes. 3. Superior cross-modal fusion effect: The multi-layer semantic injection mechanism is adopted to realize the gradual fusion of linguistic features and visual features from shallow to deep layers. It not only preserves the low-level details of visual features, but also enhances the high-level semantic expression, significantly improving the integrity of cross-modal feature representation and the accuracy of semantic alignment. 4. Low computational overhead and good real-time performance: Only two lightweight modules are added, and the visual backbone network and CLIP text encoder are frozen in the second stage of training. Only the newly added modules and prediction head are trained, avoiding the computational burden caused by large-scale parameter updates. After integration, it only brings about 10fps frame rate loss, which meets the real-time tracking requirements. 5. High versatility and easy integration: Adopting a plug-in design, it does not change the input and output structure of the original tracking framework and can be flexibly integrated into various Transformer-based single-stream tracking frameworks without requiring large-scale modifications to the original tracker, thus exhibiting good versatility and scalability. 6. Significantly improved tracking performance: Experimental results show that after the method is integrated into advanced mainstream trackers, the core indicators on multiple authoritative benchmark datasets are significantly improved, especially in challenging scenarios such as similar appearance, occlusion, and fast movement. Attached Figure Description
[0017] Figure 1 is a schematic diagram of the method of the present invention.
[0018] Figure 2 is a schematic diagram of the model architecture for the first stage of training according to the present invention.
[0019] Figure 3 is a schematic diagram of the model architecture for the second stage of training according to the present invention.
[0020] Figure 4 is a flowchart of the reasoning stage of the present invention. Detailed Implementation
[0021] The present invention will now be further described with reference to the accompanying drawings.
[0022] Referring to Figures 1-4, a general target tracking enhancement method based on implicit language guidance includes the following steps: Step 1) Input data preprocessing: Obtain template data and search images, and convert them into processable visual embedding vectors through a patch embedding layer; In this embodiment, the template data is visual data containing the initial state of the target, which can be selected as a continuous template video segment (containing N frames of images) or N discrete template images. The initial state of the target is determined by the manually annotated bounding box of the first frame. The search image is a single-frame image to be used for target location prediction. The patch embedding layer consists of a convolutional layer and an MLP layer. A downsampling rate R is set, and the template data and search image are respectively divided into fixed-size image patches and converted into high-dimensional dense visual embedding vectors. The visual embedding vector of the template data is... The visual embedding vector of the search image is C is the number of feature channels. The height of the template image. The width of the template image. For the height of the search image, The width of the template image.
[0023] Step 2) Implicit Language Representation Generation: The template data and the visual embedding vector of the search image are concatenated into a unified feature representation, which is then input into the text inversion module to generate a patch-level fine-grained pseudo-description (implicit language representation). In step 2), the visual embedding vector of the template data is first... and the visual embedding vector of the search image By splicing along the spatial dimension, a unified feature representation is formed that includes both the target's historical features and current context features. The unified feature representation is then input into the text inversion module, where it is sequentially processed through projection, multilayer perceptron, and alignment operations to precisely map the visual features to the CLIP text embedding space. The projection operation is used to adjust the matching between the visual feature dimension and the text embedding space dimension, the multilayer perceptron captures the nonlinear correlation between features, and the alignment operation ensures that the visual features and the text embedding space are distributed in a consistent manner. The final generated pseudo-description is a patch-level fine-grained representation, with each image patch corresponding to a text token, which can faithfully capture local visual details and dynamically reflect the real-time state of the target and environmental changes through global context interaction.
[0024] Step 3) High-level semantic language feature extraction: Input the pseudo-description into the frozen CLIP text encoder to extract high-level semantic language features; In this embodiment, a pre-trained clip-vit-large-patch14 model is used as the CLIP text encoder. Parameters are frozen during training and inference to fully utilize its pre-trained visual-language alignment prior knowledge. Its 12-layer Transformer is evenly divided into four text blocks (T-blocks), each containing three Transformer layers, corresponding to semantic feature extraction from shallow to deep layers. The pseudo-description is processed layer by layer through the four text blocks, outputting four levels of high-level semantic and linguistic features. ,in The language features at each level are consistent with the visual features, laying the foundation for cross-modal fusion.
[0025] Step 4) Multi-scale visual feature extraction: The visual features after patch embedding are extracted at multiple scales through the visual backbone network to obtain hierarchical visual features from shallow to deep layers. In this embodiment, the visual backbone network adopts a Transformer-based network structure, which can effectively capture global and local features of the image; its Transformer layer is evenly divided into 4 visual blocks (V-Blocks), consistent with the number of text blocks, ensuring the hierarchical correspondence between language features and visual features; unified feature representation. After being input into the visual backbone network, features are extracted layer by layer through four visual blocks to obtain multi-scale visual features at four levels. ,in The shallow visual block output contains basic visual features such as texture and edges, while the deep visual block output contains high-level visual features such as shape and category.
[0026] Step 5) Multi-layer semantic injection fusion: Through a multi-layer semantic injection mechanism, the semantic and linguistic features of each level are fused with the visual features of the corresponding level across modalities to generate enhanced visual features; In this embodiment, the specific process of multi-layer semantic injection fusion is as follows: 5.1 Feature alignment: The language features of the i-th level are aligned using the MLP alignment module. Perform dimensional adjustments and distribution alignment to ensure it aligns with visual features at the same level. By achieving perfect matching in both dimension and distribution, aligned language features are obtained. 5.2 Layer Normalization: This involves normalizing the visual features separately. and aligned language features Layer normalization is performed to eliminate distribution differences between different modal features and improve the fusion effect; 5.3 Multi-head cross-attention calculation: The normalized layer is then processed... As the query (Q), the normalized Simultaneously, these functions serve as both keys (K) and values (V), inputting into the multi-head cross-attention unit to calculate the semantic association weights between linguistic and visual features. This weight reflects the importance of linguistic semantic information to each part of the visual features; 5.4 Feature weighted fusion: incorporating attention weights With learnable scaling factor Multiply, then combine with the original visual features Perform residual connections to obtain enhanced visual features guided by implicit language. 5.5 Layer-by-layer iterative fusion: enhancing visual features As input to the next layer of visual blocks, repeat steps 5.1-5.4 to complete the semantic injection fusion of all four levels, and finally obtain the global visual features with multi-layer semantic enhancement.
[0027] Step 6) Target location prediction: Input the enhanced visual features into the prediction head and output the bounding box coordinates of the target in the search image; In this embodiment, the final layer of enhanced visual features is input into the prediction head, which includes a multilayer perceptron and a convolutional layer. The prediction head decodes the visual features to generate the bounding box coordinates of the target, and finally completes the target tracking of a single frame image.
[0028] Step 7) Model training: A two-stage training strategy is adopted to ensure model performance and training efficiency; In this embodiment 7), the two-stage training strategy is specifically as follows: 7.1 First Stage: Pre-training of the Visual Feature Extraction Network. Following the training settings of the original tracking framework, the embedding layer, visual backbone network, and prediction head are trained on a large-scale visual tracking dataset, enabling the visual backbone network to learn high-quality visual feature representations; 7.2 Second Stage: Adding Modules and Training. The proposed modules are added to the tracker, and all parameters of the embedding layer, visual backbone network, and CLIP text encoder are frozen. Only the text inversion module, the alignment module and injection module in the multi-layer semantic injection mechanism, and the prediction head are trained. The focus is on learning the mapping from visual features to text space and optimizing cross-modal fusion.
[0029] Figure 1 In this invention, an implicit language-guided tracking process is presented. The core of this method is that after the tracker integrates TIMI, it can generate implicit semantic descriptions of corresponding visual information in real time, obtain semantic features, and fuse them into the backbone network of the tracker to enhance the target feature representation.
[0030] In Figure 2, the embedding layer, visual backbone network, and prediction head are components of the tracker; the input is the template image and the search image, and the output is the target bounding box; the embedding layer receives the template image and the search image and generates an embedding representation divided into patches, the visual backbone network extracts visual features, and the prediction head generates the final target bounding box output.
[0031] In Figure 3, the patch embedding layer, the visual backbone network (divided into 4 V-Blocks), and the prediction head are the core components of the tracker, while the text inversion module, the text backbone network (divided into 4 T-Blocks), and the semantic injection module are new components added in this invention. The input is a template image and a search image, and the output is the target bounding box. The text inversion module receives the visual features after patch embedding, generates a pseudo-description, and inputs it into the text backbone network to extract language features. The semantic injection module realizes the fusion of language features and visual features at each level. Figure 4 shows the complete inference process of this invention. The inputs are a template image and a search image. The embedding layer receives the template image and the search image and performs patch embedding. The patch embedding is input to the text inversion module and the visual backbone network respectively. The text inversion module generates a pseudo-description. The pseudo-description is extracted through multi-layer language features. The visual backbone network performs multi-layer visual feature extraction. The language features of the corresponding layer are injected into the visual features and enter the next layer of visual feature extraction. After multi-layer feature extraction and injection, the final features are decoded to generate coordinates and output the tracking result.
[0032] This embodiment redefines monocular target tracking enhancement as a cross-modal feature fusion task guided by implicit language. The core process is as follows: (a) First, the template data and search image are converted into a unified visual feature representation through a patch embedding layer, providing a foundation for subsequent implicit language generation and feature fusion; (b) The visual features are mapped to the language embedding space using a text inversion module to generate patch-level fine-grained pseudo-descriptions, which can obtain semantic guidance information that dynamically adapts to the target state without explicit language input; (c) High-level semantic language features are extracted through a frozen CLIP text encoder, and multi-scale visual features are extracted through a visual backbone network to construct a hierarchical cross-modal feature set; (d) A multi-layer semantic injection mechanism is adopted to achieve deep fusion of language features and visual features through cross-attention calculation, generating enhanced visual features and improving the recognition of target features; (e) The enhanced visual features are used to predict the target position, and a two-stage training strategy is used to ensure model performance and training efficiency, ultimately achieving robust target tracking in complex scenarios.
[0033] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms stated in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.
Claims
1. A general target tracking enhancement method based on implicit language guidance, characterized in that, The method includes the following steps: Step 1) Input data preprocessing: Obtain template data and search images, and convert them into processable visual embedding vectors through a patch embedding layer; Step 2) Implicit Language Representation Generation: The template data and the visual embedding vector of the search image are concatenated into a unified feature representation, which is then input into the text inversion module to generate a patch-level fine-grained pseudo-description, i.e., implicit language representation. Step 3) High-level semantic language feature extraction: Input the pseudo-description into the frozen CLIP text encoder to extract high-level semantic language features; the Transformer layer of the CLIP text encoder is uniformly divided into N text blocks; Step 4) Multi-scale visual feature extraction: The visual features after patch embedding are extracted at multiple scales through the visual backbone network to obtain hierarchical visual features from shallow to deep layers. Step 5) Multi-layer semantic injection fusion: Through a multi-layer semantic injection mechanism, the semantic and linguistic features of each level are fused with the visual features of the corresponding level across modalities to generate enhanced visual features; Step 6) Target location prediction: Input the enhanced visual features into the prediction head and output the bounding box coordinates of the target in the search image; Step 7) Model training: A two-stage training strategy is adopted. The first stage is to pre-train the visual feature extraction network, and the second stage is to fine-tune the newly added modules to ensure model performance and training efficiency.
2. The general target tracking enhancement method based on implicit language guidance as described in claim 1, characterized in that, In step 1), the template data is visual data containing the historical state of the target, which is a template video segment of N consecutive frames or N discrete template images. The initial state of the target is determined by the manually labeled bounding box of the first frame. The search image is a single frame image to be predicted for the target position. The patch embedding layer consists of a convolutional layer and an MLP layer, which divides the template data and the search image into fixed-size image patches and converts them into high-dimensional dense visual embedding vectors.
3. A general target tracking enhancement method based on implicit language guidance as described in claim 1 or 2, characterized in that, In step 2), the visual embedding vector of the template data and the visual embedding vector of the search image are first concatenated along the spatial dimension to form a unified feature representation that includes the target's historical features and current context features. Then, the unified feature representation is input into the text inversion module, and the visual features are mapped to the CLIP text embedding space through projection, MLP and alignment operations in sequence. The final generated pseudo-description is a patch-level fine-grained representation, with each image patch corresponding to a text token, which can faithfully capture local visual details and dynamically reflect the target's real-time state and environmental changes through global context interaction.
4. A general target tracking enhancement method based on implicit language guidance as described in claim 1 or 2, characterized in that, In step 3), a pre-trained clip-vit-large-patch14 model is used as the CLIP text encoder, and the parameters are kept frozen during training and inference. The stacked multi-layer Transformer is evenly divided into multiple text blocks, each text block containing multiple layers of Transformer. The pseudo-description is processed layer by layer through multiple text blocks to output high-level semantic language features at multiple levels, and the language features at each level are consistent with the dimensions of the visual features.
5. A general target tracking enhancement method based on implicit language guidance as described in claim 1 or 2, characterized in that, In step 5), the multi-layer semantic injection fusion process is as follows: First, the dimensionality of the language features at the i-th level is adjusted and the distribution is aligned through the alignment module; then, the visual features and the aligned language features are subjected to layer normalization; next, the normalized language is used as the query and the normalized visual features are used as the key and value, and the semantic association weight cross attention is calculated through the multi-head cross attention unit; then, the cross attention is multiplied by the learnable scaling factor and residually connected with the visual features to obtain the enhanced visual features; finally, the enhanced visual features are used as the input of the next layer visual block, and the above process is repeated to complete the fusion of all levels.
6. A general target tracking enhancement method based on implicit language guidance as described in claim 1 or 2, characterized in that, In step 6), the final level of enhanced visual features is input to the prediction head. By decoding the features output by the backbone network, the bounding box regression branch outputs the target bounding box coordinates, thereby achieving target tracking of a single frame image.
7. A general target tracking enhancement method based on implicit language guidance as described in claim 1 or 2, characterized in that, In step 7), the two-stage training strategy is as follows: In the first stage, the visual backbone network and the prediction head are trained on a large-scale visual tracking dataset. The training rounds, learning rate and learning rate decay parameters are adjusted according to the different visual backbone network and prediction head. The training ends when the loss converges and becomes stable, so that the visual backbone network learns high-quality visual features. The second stage freezes the visual backbone network and CLIP text encoder, training only the text inversion module, semantic injection module, and prediction head. This stage prepares for a general object tracking enhancement method. Training is performed on a visual tracking dataset, with training epochs and an initial learning rate set. After the nth training epoch, the learning rate is adjusted by decay. The loss function is set to be consistent with the training in the first stage and adjusted according to the visual backbone network and prediction head to achieve optimal results.