A personnel re-identification method and system based on end-cloud cooperation

Abstract

The application discloses a personnel re-identification method and system based on end-cloud cooperation, and belongs to the field of computer vision. A personnel image to be queried and a personnel database are acquired; personnel features of the personnel image to be queried are extracted through an end-side personnel re-identification expert model, and the first K candidate results are obtained through preliminary screening; the personnel image to be queried is input into a cloud-side multi-modal large model, and a text description of the personnel image to be queried is generated through a specific prompt word; visual verification and text verification are performed on the candidate results in combination with the cloud-side multi-modal large model; and the final M re-identification results are output through a fusion decision mechanism by comprehensively considering the visual verification and the text verification results. Through the cooperative work of the end-side expert model and the cloud-side multi-modal large model, the small model efficiency and the large model semantic understanding ability are fully utilized, the problems of limited feature expression ability and poor adaptability to cross-camera view changes in the traditional personnel re-identification method are effectively solved, and the accuracy and robustness of personnel re-identification are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and in particular to a person re-identification method and system based on edge-cloud collaboration. Background Technology

[0002] People re-identification is an important research direction in the field of computer vision, aiming to identify the same pedestrian across different camera perspectives. With the rapid development of smart cities and intelligent security systems, people re-identification technology has significant application value in scenarios such as video surveillance and criminal suspect tracking. Current mainstream people re-identification methods are mainly divided into two categories: feature extraction-based methods and deep learning-based methods. Feature extraction-based methods, such as LBP and HOG, while computationally efficient, have limited feature representation capabilities and struggle to handle complex changes in lighting, pose, and viewpoint. Deep learning-based methods, such as Siamese networks and Triplet networks, offer high recognition accuracy, but their large number of model parameters and high computational complexity make them difficult to deploy on resource-constrained edge devices.

[0003] In recent years, with the development of cloud-based multimodal large models, some studies have attempted to apply these large models to person re-identification tasks. However, these methods typically use large models directly for end-to-end re-identification, resulting in huge computational overhead and problems such as slow inference speed and high resource consumption in practical deployments. Furthermore, existing methods lack sufficient semantic understanding of person appearance features, making it difficult to effectively describe and match detailed person features. The main problems with existing technologies are as follows: 1) A single model cannot simultaneously achieve both recognition accuracy and computational efficiency; 2) Lack of semantic understanding and descriptive ability regarding the physical characteristics of individuals; 3) Inconsistency in features caused by changes in viewpoint, lighting, and pose in cross-camera scenarios; 4) It cannot effectively integrate visual features and textual semantic information for comprehensive decision-making. Summary of the Invention

[0004] To overcome the above problems, this invention proposes a person re-identification method and system based on edge-cloud collaboration. By working collaboratively between the edge-side person re-identification expert model (small model) and the cloud-side multimodal large model, the efficiency of the small model and the semantic understanding ability of the large model are fully utilized, effectively improving the accuracy and robustness of person re-identification.

[0005] The specific technical solution adopted in this invention is as follows: Firstly, this invention proposes a person re-identification method based on edge-cloud collaboration, comprising the following steps: A person re-identification method based on edge-cloud collaboration includes the following steps: S1, Obtain the image of the person to be queried and a database of people including that person from across camera views; S2, extract the features of the image of the person to be queried through the end-side personnel re-identification expert model, and compare it with the features of all images in the personnel database to obtain the similarity ranking result; S3, Select the top K items from the similarity ranking results as the initial screening results, where K is a positive integer; S4. Input the image of the person to be queried into the cloud-side multimodal large model, and generate a text description of the image of the person to be queried through preset prompt words; S5. Input each candidate image in the preliminary screening results into the cloud-side multimodal large model, and perform visual verification and text verification through preset prompts to obtain visual verification score and text verification score; S6, by integrating visual verification score and text verification score through a fusion decision mechanism, obtains the final re-identification result; S7, select the first M entries from the final re-identification results as the final output, where M is a positive integer and M <K。

[0006] Furthermore, the on-device person re-identification expert model is a lightweight model based on the Transformer architecture, including an image segmentation processing module, a linear projection layer, a position encoding layer, a multi-head attention layer, and a feedforward network layer; the linear projection layer is used to map the features of the image segments to a unified dimension; the position encoding layer is used to add position information to each image segment; the multi-head attention layer and the feedforward network layer each contain... L Layers are used for feature extraction.

[0007] Furthermore, the training process of the end-user re-identification expert model includes: T1, construct a training dataset, including positive sample pairs and negative sample pairs; the positive sample pairs are different images of the same pedestrian, and the negative sample pairs are images of different pedestrians. T2 uses a joint loss function that includes identity loss and triplet loss to train the edge-side personnel re-identification expert model; T3 deploys the trained end-user re-identification expert model into practical applications.

[0008] Furthermore, the joint loss function of the end-user re-identification expert model is: ; ; ; in, For joint losses, For identity loss function, The triplet loss function, and These are the weighting coefficients. B For batch size, Q For the number of identity categories, For the first i The predicted score of each sample belonging to its true category. For the image features of the i-th sample, For the positive sample image features of the i-th sample, Let be the negative sample image features of the i-th sample, and margin be the boundary value.

[0009] Furthermore, in step S2, the feature extraction process of the end-user re-identification expert model includes: S2-1, The input image of the person to be queried is divided into blocks to obtain a set of image blocks; S2-2 maps the features of each image patch to a D-dimensional feature vector through a linear projection layer. And add learnable [CLS] labeled vectors to obtain features. ,in For learnable [CLS] tag vectors, Number of image patches; S2-3, Features Adding positional encoding preserves the spatial location information of image patches, resulting in numbered features. ; S2-4, via L Multi-head attention layers and feedforward network layers process the features after position encoding. The formula is as follows: ; ; in, This indicates a multi-head attention layer. Indicates a feedforward network layer. Representation layer normalization, It is the first Features of the output of the layer feedforward network It is the first Features output by the multi-head attention layer It is the first Features of the feedforward network layer output; hour, = This represents the feature vector of the person to be queried.

[0010] Furthermore, in step S4, the process of generating an image-text description of the person to be queried includes: S4-1, Construct prompts containing images of the person to be queried. The prompts should describe the person's physical features in detail and limit the number of characters in the text. S4-2, Input the prompt words and the image of the person to be queried into the cloud-based multimodal large model to obtain the text description results; S4-3 Post-processes the generated text description to remove irrelevant information and obtain the final text description.

[0011] Furthermore, in step S5, the calculation process for visual verification includes: S5-1, Construct prompt words containing the image of the person to be queried and candidate images. The prompt words require you to determine whether the two images are of the same person, and restrict the answer to "yes" or "no". S5-2, Input the prompt words and two images into the cloud-based multimodal large model to obtain the visual verification results; S5-3, convert "Yes" to a visual verification score of 1.0 and "No" to a visual verification score of 0.0.

[0012] Furthermore, in step S5, the text verification calculation process includes: S5-1, construct prompt words containing text descriptions of the person to be queried and text descriptions of candidate images. The prompt words are required to determine whether the two text descriptions refer to the same person, and the answer is limited to "yes" or "no". S5-2, input the prompt words containing two text descriptions into the cloud-side multimodal large model to obtain the text verification results; S5-3, convert "Yes" to a text verification score of 1.0, and "No" to a text verification score of 0.0.

[0013] Furthermore, in step S6, the calculation formula for the fusion decision-making mechanism is as follows: in, The final similarity score, For visual verification score, The text validation score. is the visual weight coefficient, with a value range of [0.4, 0.6].

[0014] Secondly, the present invention provides a person re-identification system based on edge-cloud collaboration, for implementing the aforementioned person re-identification method based on edge-cloud collaboration.

[0015] Compared with the prior art, the beneficial effects of this invention are: This invention leverages the collaborative work of an expert model and a cloud-based multimodal large model to fully utilize the efficiency of the small model and the semantic understanding capability of the large model, effectively addressing the limitations of traditional person re-identification methods in terms of limited feature representation and poor adaptability to changes in camera perspective. A preliminary screening mechanism significantly reduces the computational load of the large model, improving overall system efficiency. A dual verification mechanism combining visual and textual verification effectively enhances the accuracy and robustness of re-identification. A fusion decision mechanism comprehensively considers visual features and textual semantic information, making the re-identification results more reliable. Experimental results demonstrate that this invention achieves superior performance compared to existing methods on multiple public datasets while maintaining high computational efficiency. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating a person re-identification method based on edge-cloud collaboration, as shown in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the end-person re-identification expert model shown in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0018] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0019] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0020] See Figure 1 The present invention proposes a person re-identification method based on edge-cloud collaboration, which mainly includes the following steps: Step 1: Obtain the image of the person to be queried and the personnel database.

[0021] Specifically, an image of the person to be queried is acquired through a camera or image acquisition device. This image can be a single frame or a video frame. The personnel database pre-stores a large number of personnel images and their features, along with textual descriptions. These images can come from different perspectives of multiple cameras. These image features and textual descriptions are obtained using the same processing method as the image of the person to be queried, which will be explained below using only the image of the person to be queried as an example.

[0022] Furthermore, the personnel database is specifically designed for cross-camera viewpoint re-identification tasks. This database must contain a complete set of images, including the person to be queried, and all images must be captured from multiple different camera views within the same scene. Specifically, the database pre-stores images, feature vectors, and text descriptions of all registered personnel within a specific monitoring area (such as an airport terminal or shopping mall entrance). These images cover variations of the same person under multiple viewpoints, lighting conditions, and poses, ensuring that the image of the person to be queried must exist in the database.

[0023] It should also be noted that in actual deployment, all images of individuals to be queried and personnel database data are collected from legally authorized public security monitoring systems or compliant channels with explicit user consent, strictly adhering to relevant laws and regulations such as the "Cybersecurity Law of the People's Republic of China" and the "Personal Information Protection Law." Data collection is limited to public safety or authorized scenarios (such as smart city security and enterprise park management) to avoid involving sensitive areas of personal privacy.

[0024] Step 2: The image features of the person to be queried are processed by the on-side personnel re-identification expert model and compared with the image features of all people in the personnel database to obtain the similarity ranking results.

[0025] The edge-side personnel re-identification expert model is a lightweight model based on the Transformer architecture, which includes a linear projection layer, a position encoding layer, a multi-head attention layer, and a feedforward network layer.

[0026] like Figure 2 As shown, the specific processing procedure for step 2 is as follows: S2-1, Input image of the person to be queried The image is divided into blocks, with each block being the size of a given block. To obtain a set of image patches .

[0027] In this step, the image of the person being queried is... Perform block processing to segment the image into N There are 16×16 image blocks, each image block being 16×16 pixels in size (i.e., ... P =16). Assume the input image size is... H×W×C=256×128×3, then the feature dimension of each image patch is Image segmentation processing is as follows: in, For the input image, The number of image patches, A set of image patches N Number of image patches , These are the image's height, width, and number of channels, respectively.

[0028] S2-2 maps the features of each image patch to a D-dimensional feature vector through a linear projection layer. ,in For linear transformation; then add a learnable [CLS] label vector to obtain ,in This is a learnable [CLS] tag vector.

[0029] S2-3, Add positional encoding to preserve the spatial location information of image patches and obtain the numbered features. : in, pos Indicates the position index. i This indicates a dimension index.

[0030] S2-4, via L Multi-head attention layer and feedforward network layer processing, in, This indicates a multi-head attention layer. Indicates a feedforward network layer. Representation layer normalization, It is the first Features of the output of the layer feedforward network It is the first Features output by the multi-head attention layer It is the first Features of the feedforward network layer output; hour, = This represents the feature vector of the person to be queried.

[0031] S2-5, Calculate the cosine similarity between the feature vector of the person to be queried and all feature vectors in the database. in, The feature vector of the person to be queried. For the first in the database i Feature vectors of an image Let be the cosine similarity between the two.

[0032] In one specific embodiment of the present invention, the training process of the end-user re-identification expert model includes: T1. Constructing the training dataset: Select images from public datasets (such as Market-1501, DukeMTMC-reID) to construct a training set containing positive and negative sample pairs. Specifically, randomly select k=4 images from each identity to form a batch, with batch size B=32, so each batch contains 32×4=128 images; T2, using the identity loss function and triplet loss function Joint training: in, For joint losses, and As the weighting coefficient, this embodiment takes... =1.0, =1.0.

[0033] Identity loss function Calculation process: in, B =128 is the batch size. Q =751 represents the number of identity categories in the Market-1501 dataset. For the first i The predicted score of each sample belonging to its true category.

[0034] Triple loss function Calculation process, in, For the first i Image features of each sample For the first i Positive sample features of a sample (different images of the same identity). For the first i The negative sample features of each sample (images with different identities), with margin=0.2 as the boundary value.

[0035] During training, the stochastic gradient descent optimization algorithm was used, with the learning rate set to 0.0003, the batch size set to 64, and the number of training epochs set to 120. After training, the T3 expert model for end-user re-identification will be deployed in practical applications.

[0036] Step 3: Select the top K results from the similarity ranking results as the initial screening results, preferably K=20.

[0037] Based on the similarity ranking results obtained in step 2, the top 20 candidate images with the highest similarity are selected as the initial screening results. This step significantly reduces the amount of data processed by the subsequent large model, improving the overall efficiency of the system. Experiments have verified that the top 20 results usually contain correct matches, while avoiding the computational overhead of processing the entire database.

[0038] Step 4: Input the image of the person to be queried into the cloud-based multimodal large model, and generate a text description of the image of the person to be queried using preset prompt words.

[0039] The cloud-based multimodal large model is a pre-trained cloud-based multimodal large model (such as GPT-4V, Qwen-VL, etc.), which does not undergo additional training and completes the task solely through prompt word engineering. The specific processing procedure is as follows: S4-1, Construct prompts containing images of the person to be queried. The prompt format is "Please describe in detail the physical characteristics of the person in the picture, including but not limited to: clothing, hairstyle, accessories, posture, facial features, shoe type, items being carried, etc. The description should be accurate, specific, and comprehensive. Avoid using vague descriptions. For example, 'wearing dark clothes' should be specified as 'wearing a black jacket.' The description should be controlled between 150 and 200 words." S4-2, Input the prompt words and the image of the person to be queried into the cloud-based multimodal large model to obtain the text description results; S4-3 Post-processes the generated text description to remove irrelevant information and obtain the final text description.

[0040] For example, the generated text description might be: "An adult male wearing a black jacket and blue jeans, carrying a red backpack, is walking. He has short black hair, a well-defined facial structure, large eyes, and a high nose bridge. He is wearing black sneakers, his right hand is in his pocket, and his left hand is swinging naturally." Step 5: Input the features and text descriptions of each candidate image in the preliminary screening results, as well as the features and text descriptions of the image of the person to be queried, into the cloud-based multimodal large model. Perform visual verification and text verification using preset prompts to obtain visual verification scores and text verification scores.

[0041] In this embodiment, the visual verification process is as follows: S5-1, construct prompts containing images of the person to be queried and candidate images. The prompt format is: "Please determine whether the people in these two images are the same person. Please carefully observe the facial features, hairstyle, clothing, posture, accessories, and other details. If the people in the two images are highly consistent in the above features, they are considered to be the same person; if there are obvious differences, they are considered not to be the same person. Please only answer 'yes' or 'no' and do not add any other content." S5-2, Input the prompt words and two images into the cloud-based multimodal large model to obtain the visual verification results; S5-3, convert "Yes" to a visual verification score of 1.0 and "No" to a visual verification score of 0.0.

[0042] The text verification process is as follows: S5-1, construct prompts containing text descriptions of the person to be queried and text descriptions of candidate images. The prompt format is: "Please determine whether the following two descriptions describe the same person. Please carefully compare the details of the person's facial features, hairstyle, clothing, posture, accessories, shoe type, etc. in the two descriptions. If the two descriptions are highly consistent in the above characteristics, they are considered to be the same person; if there are obvious differences, they are considered to be different people. Please only answer 'yes' or 'no' and do not add any other content. Description 1: {Description of the person to be queried}; Description 2: {Description of the candidate person}"; S5-2, input the prompt words into the cloud-based multimodal large model to obtain the text verification results; S5-3, convert "Yes" to a text verification score of 1.0, and "No" to a text verification score of 0.0.

[0043] Step 6: By integrating the visual verification score and the text verification score through a fusion decision mechanism, the final re-identification result is obtained.

[0044] The fusion decision-making mechanism adopts a weighted fusion approach, and the calculation formula is as follows: in, The final similarity score, For visual verification score, The text validation score. is the visual weight coefficient, with a value range of [0.4, 0.6]. Extensive experiments have determined that when... At this time, the overall system performance is optimal.

[0045] Step 7: Select the first M results from the final re-identification results as the final output results, preferably M=10.

[0046] In this step, based on the final similarity score obtained in step 6, the preliminary screening results are reordered, and the top 10 with the highest similarity are selected as the final output results.

[0047] Taking airport security checkpoints as an example, this invention is applied to tracking the movement trajectories of staff across multiple cameras. Assuming an airport has deployed multiple cameras at check-in counters, security checkpoints, and boarding gates, when the security management center needs to query the movement path of a staff member (such as a uniformed ground staff member) within a specific time period, the system can quickly retrieve images of that staff member from other camera perspectives, such as those at the security checkpoint and boarding gates, by inputting a screenshot of that staff member at the check-in counter. For example, given a frontal image of a ground staff member, this invention first uses an edge-side expert model to quickly filter the top 20 candidate results from a database (containing all staff member images across multiple cameras), then performs visual and textual dual verification using a cloud-side large-scale model, finally outputting the top 10 matching images of that staff member at multiple locations. This not only improves the efficiency and accuracy of security monitoring but also achieves seamless tracking of personnel movement, effectively assisting airport traffic management and abnormal behavior early warning.

[0048] The above method will be applied to the following embodiments to demonstrate the technical effects of the present invention. The specific steps in the embodiments will not be repeated.

[0049] This invention was tested on the mainstream person re-identification test set Market-1501, where 750 pedestrian identities were used for evaluation. To objectively evaluate the performance of this invention, it was not trained on the Market-1501 test set distribution; all training data came from other publicly available datasets.

[0050] This comparative test employs mainstream quantitative metrics for people re-identification tasks, encompassing multiple dimensions such as recognition accuracy, feature discrimination capability, and computational efficiency. Specifically, it includes Rank-1, Rank-5, Rank-10, mAP, and inference time. Rank-k accuracy represents the probability of containing a correct match within the top k search results, used to evaluate the quality of the search results. Mean accuracy (mAP) comprehensively considers the ranking quality of the search results and is a core metric for evaluating the performance of the people re-identification system. Inference time is used to evaluate the actual deployment efficiency of the algorithm, measured in milliseconds (ms).

[0051] The benchmark models used for comparison include ResNet50, PCB, MGN, ABD-Net, OSNet, TransReID, and HCT. Similar to the method in this invention, ABD-Net and TransReID utilize attention mechanisms to enhance feature discrimination capabilities. Official open-source implementations and pre-trained models of these baselines were used during inference. Some benchmark models, such as ResNet50, use only a simple convolutional network architecture. For these methods, preprocessed images are treated as input.

[0052] The experimental results obtained by following the steps described in the specific implementation method are shown in Table 1.

[0053] Table 1: Comparative test results of person re-identification obtained from the Market-1501 dataset in this invention. As shown in Table 1, overall, the present invention, trained on publicly accessible datasets, achieves better performance in Rank-1, Rank-5, Rank-10, and mAP evaluation results, as illustrated in Table 1. HCT uses a Transformer-based feature extraction architecture, thus exhibiting better computational efficiency. Due to the design and fusion decision-making strategy of the collaborative mechanism between the expert model and the cloud-based multimodal large model, the mAP score of the present invention is significantly better than previous methods. For example, the mAP of the present invention is approximately 77.6, while the mAP of the benchmark models is below 74.5, indicating that the present invention has stronger robustness and accuracy in handling complex scenarios. Furthermore, the Rank-1 accuracy of the present invention reaches 89.8%, while the best existing model, HCT, only achieves 87.5%, demonstrating the significant advantage of the present invention in first-detection accuracy.

[0054] In terms of computational efficiency, the inference time of this invention is 152.7 milliseconds. Although this is higher than that of pure small model methods (such as ResNet50's 35.2 milliseconds), this computational cost is reasonable considering the significant performance improvement. More importantly, the preliminary screening mechanism of the expert model significantly reduces the number of candidates that need to be processed by the cloud-side multimodal large model. This allows the invention to maintain high accuracy while avoiding the huge computational overhead caused by directly using a large model for full-database retrieval. Experiments show that when the preliminary screening number is set to 20, the optimal balance between computational efficiency and recognition accuracy can be achieved. In summary, this invention, through the collaborative work of the expert model and the cloud-side multimodal large model, fully leverages the efficiency of the small model and the semantic understanding capability of the large model, achieving superior performance to existing methods on several mainstream person re-identification datasets while maintaining acceptable computational efficiency. This edge-cloud collaborative architecture provides a new solution for person re-identification tasks, with significant theoretical value and application prospects.

[0055] Based on the same inventive concept, this embodiment also provides a person re-identification system based on edge-cloud collaboration, which is used to implement the above embodiments. The terms "module," "unit," etc., used below can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible.

[0056] In this embodiment, a person re-identification system based on edge-cloud collaboration includes: The image input module is used to acquire images of the person to be queried. This module supports acquiring input data from various video surveillance sources and image files, and performs basic preprocessing to ensure that the image quality meets the recognition requirements.

[0057] The edge-side personnel re-identification expert module is used to extract the features of the person to be queried and compare them with a database containing the person across camera perspectives to obtain preliminary screening results. This module uses a lightweight Transformer model to efficiently extract features, quickly retrieve and return the top K candidate results with the highest similarity in a large personnel database.

[0058] The cloud-based multimodal large model module is used to generate image and text descriptions of the people to be queried based on preset prompts, and to perform visual and textual verification on the preliminary screening results. This module calls the pre-trained cloud-based multimodal large model to perform dual-path verification, which evaluates the visual similarity of the images and the semantic consistency of the text respectively, and generates complementary verification scores.

[0059] The fusion decision module integrates visual and textual verification results to obtain the final re-recognition result. This module adopts an adaptive weighting strategy to fuse the two types of verification scores to form the final ranking criteria, ensuring that the decision result takes into account both visual features and semantic understanding.

[0060] The results output module is used to output the final top M re-identification results. This module provides a sorted list and a visual comparison interface to intuitively display the matching results and confidence levels, facilitating rapid decision-making in practical application scenarios such as security monitoring.

[0061] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0062] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0063] The above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for re-identifying personnel based on edge-cloud collaboration, characterized in that, Includes the following steps: S1, Obtain the image of the person to be queried and a database of people including that person from across camera views; S2, extract the features of the image of the person to be queried through the end-side personnel re-identification expert model, and compare it with the features of all images in the personnel database to obtain the similarity ranking result; S3, Select the top K items from the similarity ranking results as the initial screening results, where K is a positive integer; S4. Input the image of the person to be queried into the cloud-side multimodal large model, and generate a text description of the image of the person to be queried through preset prompt words; S5. Input each candidate image in the preliminary screening results into the cloud-side multimodal large model, and perform visual verification and text verification through preset prompts to obtain visual verification score and text verification score; S6, by integrating visual verification score and text verification score through a fusion decision mechanism, obtains the final re-identification result; S7, select the first M entries from the final re-identification results as the final output, where M is a positive integer and M <K。 2. The person re-identification method based on edge-cloud collaboration according to claim 1, characterized in that, The aforementioned on-device person re-identification expert model is a lightweight model based on the Transformer architecture, comprising an image segmentation processing module, a linear projection layer, a position encoding layer, a multi-head attention layer, and a feedforward network layer. The linear projection layer maps the features of the segmented image to a unified dimension. The position encoding layer adds positional information to each image block. The multi-head attention layer and the feedforward network layer each contain... L Layers are used for feature extraction.

3. The person re-identification method based on edge-cloud collaboration according to claim 2, characterized in that, The training process of the end-person re-identification expert model includes: T1, construct a training dataset, including positive sample pairs and negative sample pairs; the positive sample pairs are different images of the same pedestrian, and the negative sample pairs are images of different pedestrians. T2 uses a joint loss function that includes identity loss and triplet loss to train the edge-side personnel re-identification expert model; T3 deploys the trained end-user re-identification expert model into practical applications.

4. The person re-identification method based on edge-cloud collaboration according to claim 3, characterized in that, The joint loss function of the end-user re-identification expert model is: ； ； ； in, For joint losses, For identity loss function, The triplet loss function, and These are the weighting coefficients. B For batch size, Q For the number of identity categories, For the first i The predicted score of each sample belonging to its true category. For the image features of the i-th sample, For the positive sample image features of the i-th sample, Let be the negative sample image features of the i-th sample, and margin be the boundary value.

5. The person re-identification method based on edge-cloud collaboration according to claim 2, characterized in that, In step S2, the feature extraction process of the end-user re-identification expert model includes: S2-1, The input image of the person to be queried is divided into blocks to obtain a set of image blocks; S2-2 maps the features of each image patch to a D-dimensional feature vector through a linear projection layer. And add learnable [CLS] labeled vectors to obtain features. ,in For learnable [CLS] tag vectors, Number of image patches; S2-3, Features Adding positional encoding preserves the spatial location information of image patches, resulting in numbered features. ; S2-4, via L Multi-head attention layers and feedforward network layers process the features after position encoding. The formula is as follows: ； ； in, This indicates a multi-head attention layer. Indicates a feedforward network layer. Representation layer normalization, It is the first Features of the output of the layer feedforward network. It is the first Features output by the multi-head attention layer It is the first Features of the feedforward network layer output; hour, = This represents the feature vector of the person to be queried.

6. The person re-identification method based on edge-cloud collaboration according to claim 1, characterized in that, Step S4, the process of generating the image text description of the person to be queried includes: S4-1, Construct prompts containing images of the person to be queried. The prompts should describe the person's physical features in detail and limit the number of characters in the text. S4-2, Input the prompt words and the image of the person to be queried into the cloud-based multimodal large model to obtain the text description results; S4-3 Post-processes the generated text description to remove irrelevant information and obtain the final text description.

7. The person re-identification method based on edge-cloud collaboration according to claim 1, characterized in that, In step S5, the calculation process for visual verification includes: S5-1, Construct prompt words containing the image of the person to be queried and candidate images. The prompt words require determining whether the two images are of the same person, and restrict the answer to "yes" or "no". S5-2, Input the prompt words and two images into the cloud-based multimodal large model to obtain the visual verification results; S5-3, convert "Yes" to a visual verification score of 1.0 and "No" to a visual verification score of 0.

0.

8. The person re-identification method based on edge-cloud collaboration according to claim 1, characterized in that, In step S5, the text verification calculation process includes: S5-1, construct prompt words containing text descriptions of the person to be queried and text descriptions of candidate images. The prompt words are required to determine whether the two text descriptions refer to the same person, and the answer is limited to "yes" or "no". S5-2, input the prompt words containing two text descriptions into the cloud-side multimodal large model to obtain the text verification results; S5-3, convert "Yes" to a text verification score of 1.0, and "No" to a text verification score of 0.

0.

9. The personnel re-identification method based on edge-cloud collaboration according to claim 1, characterized in that, In step S6, the calculation formula for the fusion decision-making mechanism is as follows: ； in, The final similarity score, For visual verification score, The text validation score. is the visual weight coefficient, with a value range of [0.4, 0.6].

10. A person re-identification system based on edge-cloud collaboration, used to implement the method of claim 1; characterized in that, The system includes: The image input module is used to obtain the image of the person to be queried; The edge-side personnel re-identification expert module is used to extract features from the image of the person to be queried and compare it with a database containing the person across camera perspectives to obtain preliminary screening results. The cloud-based multimodal large model module is used to generate text descriptions of images of people to be queried based on preset prompts, and to perform visual and textual verification of the preliminary screening results. The fusion decision module is used to integrate the visual verification and text verification results to obtain the final re-identification result; The results output module is used to output the final first M re-identification results.

Images