Thinking chain cross-view multi-target tracking method based on large language model attribute extraction

Abstract

The invention relates to the field of target tracking, in particular to a thinking chain cross-view multi-target tracking method based on large language model attribute extraction, which comprises the following steps: (1) constructing a training model, and learning an implicit relationship between text features and visual features through comparative learning and matching loss; (2) constructing an inference model, performing weighted fusion on the vision-text matching score and the vision-attribute matching score, and calculating a comprehensive score; (3) performing multi-stage thinking chain training on the training model and the inference model, and outputting an inter-frame association result and a comprehensive total score; and (4) inputting the inter-frame association result and the comprehensive total score into a prediction module, and generating a multi-target tracking object trajectory meeting a text description condition. The method has the beneficial effects that thinking chain multi-stage operation is integrated, overall surface information and deep attribute information of semantics are learned step by step, targets conforming to language description are tracked with high precision, and generalization ability, precision and robustness of the model in cross-view and cross-modal scenes are improved.

Description

Technical Field

[0001] This invention relates to the field of target tracking, and more particularly to a cross-perspective multi-target tracking method based on the extraction of attributes from a large language model. Background Technology

[0002] Traditional multi-object tracking (MOT) methods largely rely on visual feature embedding for object matching, extracting appearance features and performing spatiotemporal correlation through convolutional neural networks (CNNs) or Transformers. To enhance the semantic understanding capabilities of these models, researchers have attempted to introduce natural language descriptions to assist multi-object tracking. This involves fusing visual features with natural language descriptions generated by manual annotation or templates, thereby enhancing the semantic discriminativeness of features and making it suitable for interactive tracking needs in complex scenarios. With the widespread adoption of multi-camera surveillance and intelligent video analysis systems, cross-view multi-object tracking (CRMOT) has become an important research direction in computer vision. This task aims to maintain the identity and associate the trajectory of the same target that meets textual requirements from multiple perspectives, achieving continuous tracking over a global scale.

[0003] Multi-target tracking (RMOT) based on language features is divided into two paradigms: the two-stage paradigm and the end-to-end paradigm. The two-stage paradigm decouples trajectory generation from language matching, first explicitly constructing a complete set of candidate target trajectories, and then performing cross-modal matching; the end-to-end paradigm adopts integrated modeling, and predicts the target trajectory corresponding to the query directly from visual and language input through joint reasoning. However, existing multi-target tracking methods that incorporate text still have some shortcomings: (1) The alignment between visual features and language features is unstable. In the case of cross-viewpoints, due to changes in lighting, angle and occlusion, the distribution of visual features and text semantics are often difficult to align in a unified space, which leads to unreliable matching scores. (2) The training and reasoning process of traditional methods is a single-step decision. In the association stage, target matching is performed through a one-time decision, which cannot gradually filter, verify or correct the association. (3) Only the complete text and visual features are used for rough screening. The visual embedding lacks a deep understanding of semantics and does not introduce high-level cues such as semantic attributes to achieve detailed understanding.

[0004] The Chain-of-Thought (CoT) mechanism has demonstrated significant multi-step logical reasoning and semantic understanding capabilities in the field of Large Language Models (LLM). This mechanism, by explicitly constructing a chain of reasoning steps within the model, enables the model to progressively derive intermediate semantic conclusions from input information, thereby obtaining more stable and interpretable reasoning association results. However, in Referring-MOT tasks involving referencing language features, existing research has not yet applied the Chain-of-Thought concept to cross-perspective identity association. Traditional visual models lack progressive reasoning mechanisms, making it difficult to model multi-layered semantic information in time and space; furthermore, existing multimodal methods typically only perform one-time early feature fusion during the training phase, lacking the ability for dynamic guidance and multi-step optimization during the reasoning phase, making them prone to semantic drift and feature mismatch issues.

[0005] Meanwhile, the performance of cross-view multi-object tracking, which incorporates text, is significantly lower in cross-domain tracking (where the training and test sets are different datasets) compared to in-domain tracking (where the training and test sets are the same dataset). This performance gap is primarily due to the model's lack of deep learning of nuanced linguistic attributes. The semantic space learned in the training set cannot generalize to new descriptive methods, making accurate visual-text alignment difficult for datasets with different visual styles, attribute distributions, and even linguistic descriptions. Summary of the Invention

[0006] To overcome the above-mentioned shortcomings, this invention aims to provide a cross-perspective multi-target tracking method for thought chains based on attribute extraction from a large language model. It introduces a multi-stage training mechanism for thought chains, dividing target detection, attribute extraction, semantic alignment, and identity association into multiple progressive and logically continuous ordered stages, thereby achieving the gradual fusion of visual and textual information and semantic enhancement.

[0007] This invention achieves the above objective through the following approach: a cross-perspective multi-target tracking method for thought chains based on attribute extraction from a large language model, comprising the following steps:

[0008] (1) Constructing a training model: Calculate the image-text cosine similarity, cross-entropy loss function and cross-view reference language feature multi-target tracking loss function between visual features and text features, and learn the implicit relationship between text features and visual features through contrastive learning and matching loss;

[0009] (2) Constructing a reasoning model: Calculate the visual-text matching score and the visual-attribute matching score; weight and fuse the visual-text matching score and the visual-attribute matching score to calculate the comprehensive score;

[0010] (3) Perform multi-stage thinking chain training on the training model and the inference model, and output the inter-frame correlation results and the overall score;

[0011] (4) Input the inter-frame correlation results and the overall score into the prediction module to generate the multi-target tracking object trajectory that meets the text description conditions.

[0012] Preferably, step (1) specifically includes the following steps:

[0013] (1.1) Obtain video frames and their corresponding text descriptions;

[0014] (1.2) Video frames are output via the backbone network as detection heads, single-view Re-ID heads, cross-view Re-ID heads, and full Re-ID heads;

[0015] (1.3) Video frames are processed by an image encoder to extract Transformer image features. Extracting local representations of CNN images from full Re-ID heads Visual features are obtained by simple linear fusion after normalization. ; Calculate the multi-target tracking loss function ,in It is a parameter. This indicates damage to the detection head. This represents the single-view Re-ID loss. Indicates cross-view Re-ID loss, and These are learnable parameters;

[0016] (1.4) Text descriptions are encoded to output text features;

[0017] (1.5) Calculate the image-text cosine similarity and cross-entropy loss function for visual and text features. Where N represents the number of objects, and K represents the number of all language descriptions in the training data. This represents the label for the j-th language description corresponding to the i-th object. This represents the probability that the i-th object is predicted to be the j-th label value;

[0018] (1.6) Obtaining the multi-target tracking loss function based on cross-perspective reference language features Through optimization By learning the implicit relationship between text features and visual features, visual features for image-text matching can be obtained.

[0019] Preferably, the text description is converted into text features aligned with visual features by a text encoder.

[0020] Preferably, in step (2), the dot product is used to calculate the cosine similarity between the visual features and text features of the image-text matching to obtain the visual-text matching score.

[0021] Preferably, in step (2), the large language model performs semantic parsing on the text description to extract explicit attribute information, and extracts attribute features through a text encoder.

[0022] Preferably, the explicit attribute information is preprocessed, including deleting unnecessary formats and merging attributes that have been incorrectly separated.

[0023] Preferably, the multi-stage thinking chain training in step (3) includes:

[0024] (3.1) Conduct two-stage mental chain training on the training model;

[0025] (3.2) Conduct three-stage thinking chain training on the reasoning model.

[0026] Preferably, the two-stage mental chain training of the training model specifically includes the following steps:

[0027] (3.1.1) First stage: Visual pre-training stage: Freeze the text encoding module, train only the visual branch, generate target boxes through the detection head and combine them with the corresponding identity labels, optimize the multi-object tracking loss function, and thus obtain stable visual features; the multi-object tracking loss function ,in Indicates detection loss, This represents the single-view Re-ID loss. Indicates cross-view Re-ID loss, and These are learnable parameters;

[0028] (3.1.2) Second stage: Image-text fusion stage: Unfreeze the text encoding module and load the visual features obtained from the first stage of training; fuse the natural language text with the visual features corresponding to the target box and calculate the cross-entropy loss function. Where N represents the number of objects, and K represents the number of all language descriptions in the training data. This represents the label for the j-th language description corresponding to the i-th object. This represents the probability that the i-th object is predicted to be the j-th label value;

[0029] (3.1.3) Obtain the multi-target tracking loss function based on cross-perspective reference language features Through optimization Obtain visual features for image-text matching with multimodal understanding capabilities.

[0030] Preferably, the three-stage thought chain training of the reasoning model specifically includes the following steps:

[0031] (3.2.1) First stage coarse screening stage: The input text description and image-text matching visual features are matched and calculated to obtain a preliminary visual-text matching score. The preliminary visual-text matching score is compared with a preset threshold, and high-confidence candidates that meet the requirements are retained and the corresponding visual-text matching scores are recorded. ;

[0032] (3.2.2) Second stage attribute selection stage: The large language model performs semantic parsing to extract explicit attributes, aligns attribute features with image-text matching visual features, and calculates visual-attribute matching scores. ;

[0033] (3.2.3) Third stage: Fusion decision stage: Integrating visual-text matching scores Visual-attribute matching score The weighted fusion is used to obtain the overall total score. Output the inter-frame correlation results and the overall score, where It is a parameter.

[0034] Preferably, step (4) specifically includes the following steps:

[0035] (4.1) Input the inter-frame correlation results and the overall score into the prediction module, summarize the fusion scores of each trajectory under all views and calculate the average score. If the average score exceeds the threshold, the trajectory is directly adopted.

[0036] (4.2) If the average score of the trajectory does not reach the threshold, then fine screening is carried out: check the score of each single-view Re-ID head, give weighted rewards to views that exceed the threshold, and penalize views that do not meet the threshold. Finally, only the trajectories with cumulative scores exceeding the threshold are adopted.

[0037] The beneficial effects of this invention are as follows: This invention incorporates multi-stage operations of the thought chain into both the training and inference model stages, gradually learning the overall surface information and deep attribute information of semantics, enabling high-precision tracking of targets in videos that conform to linguistic descriptions. Firstly, the two-stage design of the training model enhances the model's image-text alignment capability. In the first stage, the model can fully learn the spatial localization and ID discrimination features of the target box, ensuring the stability of visual features. In the second stage, after unfreezing the text module and introducing image-text alignment loss, the model further optimizes the image-text matching representation while maintaining the robustness of visual features, thereby obtaining a better image-text joint embedding and improving the model's generalization ability in cross-viewpoint and cross-modal scenarios. Secondly, the three-stage inference strategy of the inference model effectively improves the accuracy and robustness of cross-perspective matching. In the first stage, preliminary filtering of candidate targets is carried out through coarse screening, which can quickly eliminate detection boxes that are not related to the semantics of the text and reduce the computational burden. In the second stage, the explicit attribute descriptions in the text are automatically parsed using the Large Language Model (LLM) to achieve fine-grained filtering at the attribute level, enabling the model to focus on fine-grained semantic differences. In the third stage, the text score and attribute score are integrated to comprehensively consider the consistency of global semantics and local attributes, thereby obtaining a more stable and interpretable final matching result. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the steps of the method of the present invention;

[0039] Figure 2 This is a schematic diagram of the network framework process of the method of the present invention;

[0040] Figure 3 This is a schematic diagram of the network framework for two-stage thought chain training using the training model of the method of this invention.

[0041] Figure 4 This is an example of a cross-view multi-target tracking visualization trajectory map implemented based on a text description;

[0042] Figure 5 This is a comparison chart of the evaluation results of this invention on a cross-domain dataset. Detailed Implementation

[0043] The present invention will be further described below with reference to specific implementation examples, but the scope of protection of the present invention is not limited thereto:

[0044] Example: The training set in this example uses the DIVOTrack dataset, which contains 7 different scenes (Floor, Gate1, Ground, Moving, Park, Shop, Square). Each scene contains video frames captured by cameras from three different perspectives (View1, View2, View3). A total of 84 text descriptions are manually annotated for the 7 scenes. The test set in this embodiment uses both in-domain and cross-domain formats. The in-domain test set uses the DIVOTrack dataset, which contains three different scenes (Circle, Gate2, Side). Each scene contains video frames captured by cameras from three different perspectives (View1, View2, View3). The three scenes are annotated with a total of 40 text descriptions using manual annotation. The cross-domain test set uses the Campus dataset, which contains three different scenes (Garden1, Garden2, ParkingLot). Each scene contains video frames captured by cameras from three or four different perspectives (View1, View2, View3 / View1, View2, View3, View4). The three scenes are annotated with a total of 48 text descriptions using manual annotation.

[0045] like Figure 1 , Figure 2 As shown, the cross-perspective multi-target tracking method for thought chains based on attribute extraction from a large language model includes the following steps:

[0046] (1) Constructing a training model: Calculate the image-text cosine similarity, cross-entropy loss function and cross-view reference language feature multi-target tracking loss function between visual features and text features, and learn the implicit relationship between text features and visual features through contrastive learning and matching loss.

[0047] like Figure 3 As shown, specifically:

[0048] (1.1) Randomly select 12 video frames from different scenes and different perspectives, and the text descriptions corresponding to the video frames by index;

[0049] (1.2) The video frames are output via the backbone network as a detection header, a single-view Re-ID header, a cross-view Re-ID header, and a full Re-ID header, wherein:

[0050] Detection head: Performs target detection on the extracted features and identifies key regions;

[0051] Single-view Re-ID head: Generates a unique feature representation for each independent view, treating the same object in different views as different objects in single-view tracking;

[0052] Cross-view Re-ID header: Establishes feature associations and correspondences between different viewpoints, treating the same object in different views as the same object;

[0053] Full Re-ID Head: Integrates information from all perspectives to form a comprehensive feature representation for language description computation;

[0054] (1.3) Video frames are processed by an image encoder to extract Transformer image features. Extracting local representations of CNN images from full Re-ID heads That is, the features within the cropped target bounding box are normalized by L2 and then simply linearly fused with the features of the transformer image to obtain the visual features. ; Calculate the multi-target tracking loss function ,in This indicates damage to the detection head. This represents the single-view Re-ID loss. Indicates cross-view Re-ID loss, and These are learnable parameters;

[0055] (1.4) After text encoding (the first six layers of BERT), the text description converts the text information into text features aligned with the visual features;

[0056] (1.5) Calculate the image-text cosine similarity and cross-entropy loss function for visual and text features using the dot product. The text reference loss is represented by cross-entropy loss, where N represents the number of objects and K represents the number of language descriptions in the training data. This represents the label for the j-th language description corresponding to the i-th object. This represents the probability that the i-th object is predicted to be the j-th label value;

[0057] (1.6) Obtaining the multi-target tracking loss function based on cross-perspective reference language features And the similarity between all text descriptions and all targets, through optimization By learning the implicit relationship between text features and visual features, visual features for image-text matching can be obtained.

[0058] In the process of learning the implicit relationship between text features and visual features, ITC (Image_TextConstructive loss) is calculated to bring correct image-text matching pairs closer together and incorrect image-text matching pairs further apart; ITM (Image_Text Matching loss) is calculated to determine whether the image and text match.

[0059] (2) Constructing the inference model: Calculate the visual-text matching score and the visual-attribute matching score; weight and fuse the visual-text matching score and the visual-attribute matching score to calculate the comprehensive score. Specifically:

[0060] (2.1) After the text description is processed by the text encoder (the first six layers of BERT), the text information is converted into text feature vectors. The dot product is used to calculate the cosine similarity between the image and the text to obtain the visual-text matching score. ;

[0061] (2.2) Short word attributes are extracted from text descriptions using a Large Language Model (LLM, Qwen3.5 is used in this invention), and attribute features are extracted through a text encoder. Because unnecessary markdown formatting may occur during extraction, leading to attribute segmentation errors, this invention verifies and removes unnecessary formatting from attribute words, and re-merges attributes that were incorrectly separated due to markdown formatting removal. The cross-encoder in this invention is implemented using the last six layers of BERT. The cross-encoder embeds and aligns attribute features with visual features to learn the similarity between the same video frame image and each attribute. The original binary logits matching values ​​are converted into sigmoid probabilities and their average values ​​are calculated. The average matching probability of the attributes is then used to obtain the comprehensive visual-attribute matching score for each image. ;

[0062] (2.3) Visual-text matching score Visual-attribute matching score Perform weighted fusion to obtain the fusion score. In this invention .

[0063] (3) Perform multi-stage thinking chain training on the training model and the inference model, and output the inter-frame correlation results and the overall score.

[0064] (3.1) Conduct two-stage mental chain training on the training model.

[0065] (3.1.1) First stage: Visual pre-training stage: Freeze the text encoding module, train only the visual branch, generate target boxes through the detection head and combine them with the corresponding identity labels, optimize the multi-object tracking loss function, and thus obtain stable visual features; the multi-object tracking loss function , among which This indicates damage to the detection head. This represents the single-view Re-ID loss. Indicates cross-view Re-ID loss, and These are learnable parameters;

[0066] (3.1.2) Second stage: Image-text fusion stage: Unfreeze the text encoding module and load the visual features obtained from the first stage of training; fuse the natural language text with the visual features corresponding to the target box and calculate the cross-entropy loss function. Where N represents the number of objects, and K represents the number of all language descriptions in the training data. This represents the label for the j-th language description corresponding to the i-th object. This represents the probability that the i-th object is predicted to be the j-th label value;

[0067] (3.1.3) Obtain the multi-target tracking loss function based on cross-perspective reference language features Through optimization Obtain visual features for image-text matching with multimodal understanding capabilities.

[0068] (3.2) Conduct three-stage thinking chain training on the reasoning model.

[0069] (3.2.1) First stage: Coarse screening stage: The complete text description input is matched with the visual features of the target bounding box to obtain a preliminary text matching score. This score is compared with a preset threshold of 0.3, and candidate bounding boxes with low scores are filtered out. High-confidence candidates that meet the requirements are retained, and the corresponding visual-text matching scores are recorded. ;

[0070] (3.2.2) Second stage: Attribute fine-tuning stage: The large language model performs semantic parsing on the complete text description to extract explicit attribute information. In one embodiment, the explicit attribute information extracted from a complete text description "A man in a black coat and blue trousers, carrying a blue bag and holding a book" includes "man", "black coat", "blue trousers", "blue bag", "book", etc. Because unnecessary markdown formatting may be generated during extraction, leading to attribute segmentation errors, this invention preprocesses the explicit attribute information, verifies whether there are unnecessary formats in the attribute words and deletes them, and re-merges attributes that were incorrectly separated due to the deletion of markdown formatting, such as "black" and "coat". The attribute features are aligned with the visual features of the target box. This invention uses the last six layers of BERT as a visual-text cross-encoder to calculate the visual-attribute matching score. This allows for more granular semantic filtering.

[0071] (3.2.3) Third stage: Fusion decision stage: Integrating visual-text matching scores Visual-attribute matching score The weighted fusion is used to obtain the overall total score. Output the inter-frame correlation results and the overall score.

[0072] (4) Input the inter-frame correlation results and the overall score into the prediction module to generate the multi-target tracking object trajectory that meets the text description conditions.

[0073] The inter-frame correlation results and fusion scores are input into the prediction module. The prediction module acts as an intelligent filter. Its core step is to perform a two-stage screening of the input candidate trajectories based on the multi-view fusion scores, generating multi-target tracking object trajectories that conform to natural language description conditions, thus achieving identity consistency determination and cross-view trajectory updates. Specific steps include:

[0074] (4.1) The prediction module summarizes the fusion scores of each trajectory under all views and calculates the average score. If the average score exceeds the threshold, the trajectory is directly adopted.

[0075] (4.2) If the average score does not reach the threshold, the second level of fine screening is entered: the score of each single view is checked one by one, the view that exceeds the threshold is weighted and rewarded, the view that does not meet the threshold is penalized, and finally only the trajectory with the cumulative hit score exceeding the threshold is adopted, so as to output a high-quality and reliable final trajectory set.

[0076] like Figure 4 The image shown is a final visualization example of an embodiment of the present invention. For the text description "A man in a black coat and blue trousers, carrying a blue bag and holding a book," the present invention processes the data to obtain trajectory tracking videos from three different perspectives that match the text description. Furthermore, all targets in the videos that match the text description are labeled with IDs and bounding boxes. Figure 4 As can be seen, the tracking results from the three perspectives are consistent and the object is continuously tracked. Furthermore, compared to perspectives 1 and 2, perspective 3 detects and tracks one more target that matches the text description, which does not appear in perspectives 1 and 2, further verifying the robustness of the invention.

[0077] Cross-view multi-target tracking is challenging on cross-domain test sets, where method performance tends to drop significantly compared to in-domain test sets. However, the method described in this invention demonstrates good MOTA and IDF1 performance across various scenarios on the Cross-domain Campus dataset, exhibiting strong generalization ability. Figure 5As shown, the present invention achieves 12.41% CVRIDF1 and 2.6% CVRMA in all scenarios, and has good generalization effect in specific scenarios such as Garden1, Garden2 and ParkingLot. The bolded part represents the optimal effect.

[0078] This invention introduces a multi-stage reasoning mechanism based on a thought chain, dividing target detection, attribute extraction, semantic alignment, and identity association into multiple progressive and logically continuous ordered stages, achieving gradual fusion and semantic enhancement of visual and textual information. Specifically, in the reasoning stage, this invention uses a large language model to parse the natural language descriptions of candidate targets, automatically extracting semantic attribute information (such as clothing, carried items, and action features), and performing multimodal cross-fusion and alignment with visual encoded features. This design effectively improves the model's feature consistency, semantic interpretability, and matching robustness across cross-viewpoint and cross-modal scenarios, thereby achieving high-precision identity association and global tracking.

[0079] The above description describes specific embodiments of the present invention and the technical principles employed. Any changes made in accordance with the concept of the present invention that do not exceed the spirit of the specification and drawings should still fall within the protection scope of the present invention.

Claims

1. A thought chain cross-view multi-target tracking method based on large language model attribute extraction, characterized by The method comprises the following steps: (1) constructing a training model: calculating the image-text cosine similarity of visual features and text features, cross-entropy loss function and cross-view reference language feature multi-object tracking loss function, learning the implicit relationship between text features and visual features through contrastive learning and matching loss; (2) constructing an inference model: calculating visual-text matching score and visual-attribute matching score; Weighted fusion of visual-text matching score and visual-attribute matching score to calculate the comprehensive score; (3) multi-stage thinking chain training of the training model and the inference model, outputting the inter-frame association result and the comprehensive total score; (4) inputting the inter-frame association result and the comprehensive total score into a prediction module to generate the multi-object tracking object trajectory meeting the text description condition.

2. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 1, characterized in that, The step (1) specifically comprises the following steps: (1.1) obtaining video frames and corresponding text descriptions; (1.2) the video frames are output by the backbone network to the detection head, single-view Re-ID head, cross-view Re-ID head and full Re-ID head; (1.3) Video frame extracts Transformer image features through image encoder CNN image local representation is extracted from full Re-ID head Visual features are obtained by simple linear fusion after normalization A multi-target tracking loss function is calculated Wherein is a parameter, represents the detection head loss, represents the single-view Re-ID loss, represents the cross-view Re-ID loss, and are learnable parameters; (1.4) the text description is output as text features by the text encoder; (1.5) calculating the image-text cosine similarity and cross-entropy loss function of the visual features and the text features where N represents the number of objects, K represents the number of all language descriptions in the training data, represents the label of the jth language description corresponding to the ith object, represents the probability that the ith object is predicted as the jth label value; (1.6) obtain a cross-view reference language feature multi-object tracking loss function by optimizing learning the implicit relationship between text features and visual features, obtaining a picture-text matching visual feature.

3. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 2, characterized in that, The text description is converted into text features aligned with the visual features by the text encoder.

4. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 1, characterized in that, In the step (2), the dot product calculates the image-text cosine similarity of the image-text matching visual features and the text features to obtain the visual-text matching score.

5. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 1, characterized in that, In the step (2), the large language model performs semantic analysis on the text description to extract explicit attribute information, and extracts attribute features through the text encoder.

6. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 5, characterized in that, The explicit attribute information is preprocessed, including deleting unnecessary formats that do not exist and merging attributes that are incorrectly separated.

7. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 1, characterized in that, The multi-stage thinking chain training in the step (3) comprises: (3.1) two-stage thinking chain training of the training model; (3.2) three-stage thinking chain training of the inference model.

8. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 7, characterized in that, The two-stage thinking chain training of the training model specifically comprises the following steps: (3.1.1) the first stage visual pre-training stage: freezing the text encoding module, only training the visual branch, generating the target box through the detection head and combining the corresponding identity label, optimizing the multi-target tracking loss function, so as to obtain stable visual features; the multi-target tracking loss function wherein represents the detection head loss, represents the single-view Re-ID loss, represents the cross-view Re-ID loss, and are learnable parameters; (3.1.2) the second stage of text and image fusion stage: thaw text encoding module, load the visual features obtained by the first stage training; the natural language text and the visual features corresponding to the target box are fused, and the cross entropy loss function is calculated where N represents the number of objects, K represents the number of all language descriptions in the training data, represents the label of the jth language description corresponding to the ith object, represents the probability that the ith object is predicted as the jth label value; (3.1.3) obtain a cross-view reference language feature multi-object tracking loss function , by optimizing obtain a graph-text matching visual feature with multi-modal understanding ability.

9. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 7, characterized in that, The three-stage thinking chain training of the inference model specifically comprises the following steps: (3.2.1) The first stage is a rough screening stage: the input text description and the image-text matching visual features are matched and calculated to obtain a preliminary visual-text matching score. The preliminary visual-text matching score is compared with a preset threshold, and high-confidence candidates that meet the requirements are retained and the corresponding visual-text matching scores are recorded ; (3.2.2) The second stage attribute screening stage: semantic analysis is performed by the large language model to extract explicit attributes, the attribute features are aligned with the visual features of the text and pictures, and the visual-attribute matching score is calculated ; (3.2.3) the third stage fusion decision stage: the visual-text matching score and the visual-attribute matching score are weighted to obtain a comprehensive total score , and the inter-frame association result and the comprehensive total score are output, wherein is a parameter.

10. The thought chain cross-view multi-target tracking method based on large language model attribute extraction according to claim 1, characterized in that, The step (4) specifically comprises the following steps: (4.1) inputting the inter-frame association result and the comprehensive total score into the prediction module, summarizing the fusion scores of each trajectory under all views and calculating the average score, and directly adopting the trajectory if the average score exceeds the threshold; (4.2) if the average score of the trajectory does not reach the threshold, fine screening is entered: checking the score of each single-view Re-ID head, giving weighted rewards to the views exceeding the threshold and punishing the views not meeting the standard, and finally only adopting the trajectory with the cumulative score exceeding the threshold.

Images