A text-based person re-identification method based on cross-modal semantic alignment

Abstract

This invention discloses a text-based person re-identification method based on cross-modal semantic alignment, comprising: 1. Inputting text statements and pedestrian images into a text encoder and a visual encoder respectively to obtain text features and image features; 2. Clustering the text features and image features using a token clustering learning module to obtain clustered text and visual prototype features; 3. Aligning the clustered text and visual prototype features by introducing an optimal transfer algorithm, and further calculating the loss function of the aligned prototype features to update the text encoder and visual encoder; 4. Calculating the similarity matrix between the text prototype features and the visual prototype features, and obtaining the retrieval results. This invention can effectively extract discriminative key information from pedestrian images and text descriptions, and aligns text and image modalities through an optimal transfer algorithm, thereby enabling more efficient and accurate matching of text modalities and pedestrian images.

Description

Technical Field

[0001] This invention belongs to the field of pedestrian re-identification, specifically a text-based pedestrian re-identification method based on cross-modal semantic alignment. Background Technology

[0002] Person re-identification (Re-ID) is the task of identifying a specific pedestrian across multiple non-overlapping cameras. In recent years, this task has attracted considerable attention due to its widespread applications in activity analysis, automated tracking, and smart retail. However, it remains challenging due to cluttered backgrounds, occlusion, varying lighting, camera angles, and the diversity of human poses. Most existing person re-identification methods focus primarily on image-based re-identification, formulating the task as an image-to-image matching problem. However, image-based re-identification requires at least one image as the query, which can be difficult to obtain in some scenarios. In contrast, text-based person re-identification methods employ free-form natural language descriptions, which are more user-friendly in practical applications. Given a textual description of the target person, text-based person re-identification aims to find the corresponding pedestrian image for the same person from a large-scale image database. Compared to traditional person re-identification methods, text-based person re-identification faces not only the challenges of person re-identification but also significant inter-modal discrepancies, making this task even more challenging.

[0003] Existing text-based person re-identification methods can be divided into two categories: global matching methods and local matching methods. Global matching methods extract global visual and text representations from both modalities and project them into a shared space. These methods reduce modal discrepancies through cleverly designed loss functions, but neglect noteworthy local details in both image and text modalities. Local matching methods typically consist of two steps: 1) extracting effective local representations using predefined rules or external models, and 2) then performing cross-modal matching using attention mechanisms. While these methods effectively mitigate modal discrepancies, the interaction between external models and expensive pairwise cross-modal similarity scores leads to significant computational costs. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing technologies by proposing a text-based pedestrian re-identification method based on cross-modal semantic alignment. This method effectively extracts discriminative key information from pedestrian images and text descriptions. Simultaneously, it optimizes the transmission alignment of text and image modalities to solve the problem of high computational overhead caused by traditional attention mechanisms, thereby enabling more efficient and accurate matching of text modalities and pedestrian images.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a text-based person re-identification method based on cross-modal semantic alignment, characterized by the following steps: Step 1: Collection and preprocessing of pedestrian data: Pedestrian visible light surveillance video was acquired using an optical camera. The video was then preprocessed frame-by-frame with pedestrian detection and size normalization. After data augmentation using random horizontal flipping, a training set of pedestrian images, X=, was obtained. ;in, This represents the i-th pedestrian image; Indicates the total number of images; The i-th pedestrian image Divide into N smaller blocks to obtain N sets of blocks. ,in, Represents the i-th pedestrian image The nth small block; right After vectorization, and The corresponding position codes are added together to obtain the nth block. Encoding vector ; pedestrian image set Describe and generate the corresponding text statement set. ; Represents the i-th pedestrian image The text description; right After word segmentation, the resulting L words are further embedded to obtain L feature vectors. These feature vectors are then added to the positional codes of the L words to obtain L encoded word vectors. ;in, express The Middle l Each encoded word vector; For the i-th pedestrian image and its corresponding text description Set the identity ID of the i-th pedestrian, denoted as . ,and The number of different pedestrian identity ID types; Step 2: Construct a text-based person re-identification network consisting of a feature extraction network, a token clustering learning module, a feature alignment learning module, and a prediction module; Step 2.1: The feature extraction network respectively... and After processing, the nth small block is obtained. pedestrian characteristics and the l Encoded word vectors Features ; Step 2.2, the token clustering learning module will respectively... and Clustered into the i-th pedestrian image The j-th visual part prototype feature and the i-th text description The j-th text portion prototype feature ; Step 2.3, the feature alignment learning module... and After processing, the i-th pedestrian image is obtained. Prototype features of the j-th aligned text portion and the i-th text description The j-th aligned visual part prototype feature ; Step 2.4, the prediction module... and After processing, the i-th pedestrian image is obtained. and text description The corresponding classification probability; Step 3: Optimize the text-based pedestrian re-identification network; Based on pedestrian training image set X and its text statement set The Adam optimization strategy was used to train the feature extraction network, and the total loss function of the network was calculated. To update network parameters until The optimal feature extraction network is obtained by continuing until convergence. Step 4: The search process; Step 4.1: Query the semantic prototype features of all text queries in the test set. and semantic prototype features of all image libraries After each part is concatenated, the corresponding text query features are obtained. Image features of the image library ;in, This represents the text prototype feature of the q-th text query. Represents the visual prototype features of the g-th pedestrian image; Step 4.2: Calculate text query features Image features of the image library The similarity matrix is ​​obtained, and the Sinkhorn algorithm in optimal transmission is used to update the similarity matrix. After sorting the updated similarity matrix row by row, the image retrieval results corresponding to the text description are obtained.

[0006] The text-based person re-identification method based on cross-modal semantic alignment described in this invention is characterized in that the feature extraction network in step 2.2 is based on the CLIP-Transformer deep learning network and includes: a text encoder and a visual encoder, wherein the text encoder is composed of... The visual encoder consists of Transformer residual blocks and is composed of... It consists of Transformer residual blocks, each of which is composed of a self-attention layer, a layer normalization layer, and a feedforward network layer with the same structure but different parameters. Will The input is fed into the feature extraction network and processed by the visual encoder to obtain the nth small block. pedestrian characteristics Thus, the i-th pedestrian image is obtained. pedestrian characteristics Where D is the feature dimension; Will The input is fed into the feature extraction network and processed by the text encoder to obtain the first... l Encoded word vectors Features Thus obtain pedestrian text features .

[0007] Step 2.2 includes: Step 2.2.1: Define the visual semantic prototype and text semantic prototype ,in, This represents the j-th randomly generated visual semantic prototype. This represents the j-th randomly generated text semantic prototype; Step 2.2.2, Calculation and correlation matrix between , used to characterize the i-th pedestrian image The correlation between N image patch features and M visual semantic prototypes; calculate and correlation matrix between , used to characterize the i-th text description The correlation between L text features and M text semantic prototypes; Step 2.2.3, will transpose and After multiplication, the visual prototype features after clustering are obtained. ,in, Represents the i-th pedestrian image The j-th visual part prototype feature; Will transpose and After multiplication, we obtain the prototype features of the clustered text portion. ,in, Represents the i-th text description The prototype feature of the j-th text part.

[0008] The feature alignment learning module in step 2.3 uses the Sinkhorn algorithm in optimal transmission to iteratively calculate the prototype features of the text portion. and visual prototype features Optimal plan for alignment between ; The optimal plan and visual prototype features After multiplication, the prototype features of the aligned text portion are obtained. ; The optimal plan Transpose and textual prototype features After multiplication, the aligned visual prototype features are obtained. .

[0009] The prediction module in step 2.4 will align the visual prototype features. The final visual prototype features are obtained after splicing. The prototype features of the aligned text portion The final text prototype feature is obtained after concatenation. ,Will and After classification processing by a fully connected layer and then by the softmax function, the corresponding classification probabilities are obtained. ;in, Represents the i-th pedestrian image The output features are classified into the first category. The probability of a pedestrian identity ID. This indicates the corresponding text description. The output features are classified into the first category. The probability of a pedestrian identity ID.

[0010] The network total loss function in step three It is constructed according to the following steps: Step 3.1: Construct the identity loss function using equation (1) : (1) In equation (1), and The i-th pedestrian image and the i-th text description The output is classified into The probability of; Step 3.2: Construct the triplet loss function for the image modality and the text modality using equation (2). : (2) In equation (2), Indicates the first i Zhang Xingren Image Aligned visual prototype features and They represent Prototype features of the text portion aligned with the same pedestrian identity ID and different pedestrian identity IDs; Represents the i-th text description Aligned text portion prototype features. and They represent Visual prototype features of the same type of pedestrian identity ID and different types of pedestrian identity ID; It is the margin parameter. This represents the function that takes the maximum value. Represents the cosine similarity function; Step 3.3: Construct the alignment loss function for the visual and textual prototype features using equation (3). : (3) In equation (3), Let represent the alignment function from the text part prototype features to the visual part prototype features, and obtain it from equation (4). The alignment function from the visual prototype features to the text prototype features is given by equation (5). (4) (5) In equations (4) and (5), It is the temperature hyperparameter that controls the probability distribution; Step 3.4: Construct the total network loss function using equation (6). : (6).

[0011] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the text pedestrian re-identification method, and the processor is configured to execute the program stored in the memory.

[0012] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the text pedestrian re-identification method.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention proposes a novel cross-modal semantic alignment learning method, which learns distinctive image-text features through efficient and accurate cross-modal alignment, thereby significantly reducing intermodal gaps and improving the accuracy of retrieval and recognition in text pedestrian re-identification tasks.

[0014] 2. This invention uses a token clustering learning module, which maintains a set of visual and textual semantic prototypes. It aggregates the textual and visual token features extracted from the two modalities into fine-grained and compact textual and visual partial prototype features. These prototype features correspond to identity-related discriminative semantic information, thereby enhancing the feature representation capability of the textual pedestrian re-identification task and improving the accuracy of retrieval and recognition.

[0015] 3. This invention uses a feature alignment learning module, which introduces the Sinkhorn algorithm from the optimal transmission strategy to align the visual and textual prototype features to effectively solve semantic misalignment and greatly reduce computational costs, thereby accelerating the retrieval and matching speed of text pedestrian re-identification tasks. Attached Figure Description

[0016] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0017] In this embodiment, the flowchart of a text-based person re-identification method based on cross-modal semantic alignment is as follows: Figure 1 Specifically, it is done in the following steps: Step 1: Collection and preprocessing of pedestrian data: Pedestrian visible light surveillance video was acquired using an optical camera. The video was then preprocessed frame-by-frame with pedestrian detection and size normalization. After data augmentation using random horizontal flipping, a training set of pedestrian images, X=, was obtained. ;in, This represents the i-th pedestrian image; Indicates the total number of images; Take the i-th pedestrian image Divide into N smaller blocks to obtain N sets of blocks. ,in, Represents the i-th pedestrian image The nth small block; right After vectorization, and The corresponding position codes are added together to obtain the nth block. Encoding vector ; pedestrian image set Describe and generate the corresponding text statement set. ; Represents the i-th pedestrian image The text description; right After word segmentation, the resulting L words are further embedded to obtain L feature vectors. These feature vectors are then added to the positional codes of the L words to obtain L encoded word vectors. ;in, express The Middle l Each encoded word vector; For the i-th pedestrian image and its corresponding text description Set the identity ID of the i-th pedestrian, denoted as . ,and This represents the number of different pedestrian identity ID types; in this embodiment, B = 64. Each collection of training data includes 16 pedestrian IDs, with 4 text descriptions and 4 pedestrian images collected for each pedestrian ID.

[0018] Step 2: Construct a text-based person re-identification network consisting of a feature extraction network, a token clustering learning module, a feature alignment learning module, and a prediction module; Step 2.1: The feature extraction network is based on the CLIP-Transformer deep learning network and includes a text encoder and a visual encoder. The text encoder is composed of... The visual encoder consists of Transformer residual blocks and is composed of... It consists of several Transformer residual blocks, each of which is composed of a self-attention layer, a layer normalization layer, and a feedforward network layer with the same structure but different parameters; in this embodiment, .

[0019] Will The input is fed into the feature extraction network and processed by the visual encoder to obtain the nth small block. pedestrian characteristics Thus, the i-th pedestrian image is obtained. pedestrian characteristics Where D is the feature dimension; in this embodiment, .

[0020] Will The input is fed into the feature extraction network and processed by the text encoder to obtain the first... l Encoded word vectors Features Thus obtain pedestrian text features In this embodiment, .

[0021] Step 2.2, the token clustering learning module will respectively... and Clustering into fine-grained visual partial prototype features and text part prototype features and ;in, Represents the i-th pedestrian image The j-th visual part prototype feature, Represents the i-th text description The j-th text portion prototype feature; in this embodiment, Each prototype feature corresponds to a specific discriminative semantic information related to identity information, which is a key clue to distinguish different pedestrians, thus enhancing the feature expression capability of the text-based pedestrian re-identification method.

[0022] Step 2.2.1: Define the visual semantic prototype and text semantic prototype ,in, This represents the j-th randomly generated visual semantic prototype. This represents the j-th randomly generated text semantic prototype; Step 2.2.2, Calculation and correlation matrix between , used to characterize the i-th pedestrian image The correlation between N image patch features and M visual semantic prototypes; calculate and correlation matrix between , used to characterize the i-th text description The correlation between L text features and M text semantic prototypes; Step 2.2.3, will transpose and After multiplication, the visual prototype features after clustering are obtained. ,in, Represents the i-th pedestrian image The j-th visual part prototype feature; Will transpose and After multiplication, we obtain the prototype features of the clustered text portion. ,in, Represents the i-th text description The prototype feature of the j-th text part.

[0023] Step 2.3: The feature alignment learning module uses the Sinkhorn algorithm in optimal transmission to iteratively calculate the prototype features of the text portion. and visual prototype features Optimal plan for alignment between ; The optimal plan and visual prototype features After multiplication, the prototype features of the aligned text portion are obtained. ,in, Represents the i-th pedestrian image The prototype feature of the j-th aligned text portion; The optimal plan Transpose and textual prototype features After multiplication, the aligned visual prototype features are obtained. ,in, Represents the i-th text description The j-th aligned visual prototype feature; In this embodiment, since most traditional optimization transmission methods are complex and non-differentiable, the Sinkhorn algorithm is adopted for fast optimization. This algorithm uses entropy regularization constraints and iterates 100 times to obtain the optimal plan. This enables fine-grained semantic alignment of text and image prototype features while significantly reducing computational overhead. The Sinkhorn algorithm greatly reduces the complexity of the solution through entropy regularization and iterative computation, making it more feasible in practical applications and thus providing a more efficient and accurate feature matching method for the field of person re-identification.

[0024] The prediction module will align the visual prototype features. The final visual prototype features are obtained after splicing. The prototype features of the aligned text portion The final text prototype feature is obtained after concatenation. ,Will and After classification processing by a fully connected layer and then by the softmax function, the corresponding classification probabilities are obtained. ;in, Represents the i-th pedestrian image The output features are classified into the first category. The probability of a pedestrian identity ID. This indicates the corresponding text description. The output features are classified into the first category. The probability of a pedestrian identity ID.

[0025] Step 3: Optimize the text-based pedestrian re-identification network; Step 3.1: Construct the identity loss function using equation (1) : (1) In equation (1), and The i-th pedestrian image and the i-th text description The output is classified into The probability of; Step 3.2: Construct the triplet loss function for the image modality and the text modality using equation (2). : (2) In equation (2), Indicates the first i Zhang Xingren Image Aligned visual prototype features and They represent Prototype features of the text portion aligned with the same pedestrian identity ID and different pedestrian identity IDs; Represents the i-th text description Aligned text portion prototype features. and They represent Visual prototype features of the same type of pedestrian identity ID and different types of pedestrian identity ID; It is the margin parameter. This represents the function that takes the maximum value. This represents the cosine similarity function; in this embodiment, .

[0026] Step 3.3: Construct the alignment loss function for the visual and textual prototype features using equation (3). The alignment loss function further enhances the semantic alignment between the prototype features of the text and visual parts; (3) In equation (3), Let represent the alignment function from the text part prototype features to the visual part prototype features, and obtain it from equation (4). The alignment function from the visual prototype features to the text prototype features is given by equation (5). (4) (5) In equations (4) and (5), It is the temperature hyperparameter that controls the probability distribution; in this embodiment, .

[0027] Step 3.4: Construct the total network loss function using equation (6). : (6) Step 3.5: Based on the pedestrian image set X and its text statement set The Adam optimization strategy was used to train the feature extraction network, and the total loss function of the network was calculated. To update network parameters until The optimal feature extraction network is obtained by training until convergence. In this embodiment, a total of 60 epochs were trained, with an initial learning rate of 1e-5, and the learning rate was reduced by 0.1 in the 20th, 40th and 50th epochs.

[0028] Step 4: The search process; Step 4.1: Query the semantic prototype features of all text queries in the test set. and semantic prototype features of all image libraries After each part is concatenated, the corresponding text query features are obtained. Image features of the image library ;in, This represents the text prototype feature of the q-th text query. This represents the visual prototype features of the g-th pedestrian image; in this embodiment, .

[0029] Step 4.2: Calculate text query features Image features of the image library The similarity matrix is ​​obtained, and the Sinkhorn algorithm in optimal transmission is used to update the similarity matrix. After sorting the updated similarity matrix row by row, the image retrieval results corresponding to the text description are obtained. In this embodiment, the Sinkhorn algorithm is used to calculate the optimal plan, and the corresponding elements are multiplied with the original similarity matrix to obtain the updated similarity matrix.

[0030] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0031] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A text-based person re-identification method based on cross-modal semantic alignment, characterized in that, Includes the following steps: Step 1: Collection and preprocessing of pedestrian data: Pedestrian visible light surveillance video was acquired using an optical camera. The video was then preprocessed frame-by-frame with pedestrian detection and size normalization. After data augmentation using random horizontal flipping, a training set of pedestrian images, X=, was obtained. ;in, This represents the i-th pedestrian image; Indicates the total number of images; The i-th pedestrian image Divide into N smaller blocks to obtain N sets of blocks. ,in, Represents the i-th pedestrian image The nth small block; right After vectorization, and The corresponding position codes are added together to obtain the nth block. Encoding vector ; pedestrian image set Describe and generate the corresponding text statement set. ; Represents the i-th pedestrian image The text description; right After word segmentation, the resulting L words are further embedded to obtain L feature vectors. These feature vectors are then added to the positional codes of the L words to obtain L encoded word vectors. ;in, express The Middle l Each encoded word vector; For the i-th pedestrian image and its corresponding text description Set the identity ID of the i-th pedestrian, denoted as . ,and The number of different pedestrian identity ID types; Step 2: Construct a text-based person re-identification network consisting of a feature extraction network, a token clustering learning module, a feature alignment learning module, and a prediction module; Step 2.1: The feature extraction network respectively... and After processing, the nth small block is obtained. pedestrian characteristics and the l Encoded word vectors Features ; Step 2.2, the token clustering learning module will respectively... and Clustered into the i-th pedestrian image The j-th visual part prototype feature and the i-th text description The j-th text portion prototype feature ; Step 2.2.1: Define the visual semantic prototype and text semantic prototype ,in, This represents the j-th randomly generated visual semantic prototype. This represents the j-th randomly generated text semantic prototype; Step 2.2.2, Calculation and correlation matrix between , used to characterize the i-th pedestrian image The correlation between N image patch features and M visual semantic prototypes; calculate and correlation matrix between , used to characterize the i-th text description The correlation between L text features and M text semantic prototypes; Step 2.2.3, will transpose and After multiplication, the visual prototype features after clustering are obtained. ,in, Represents the i-th pedestrian image The j-th visual part prototype feature; Will transpose and After multiplication, we obtain the prototype features of the clustered text portion. ,in, Represents the i-th text description The prototype feature of the j-th text part; Step 2.3, the feature alignment learning module... and After processing, the i-th pedestrian image is obtained. Prototype features of the j-th aligned text portion and the i-th text description The j-th aligned visual part prototype feature ; Step 2.4, the prediction module... and After processing, the i-th pedestrian image is obtained. and text description The corresponding classification probability; Step 3: Optimize the text-based pedestrian re-identification network; Based on pedestrian training image set X and its text statement set The Adam optimization strategy was used to train the feature extraction network, and the total loss function of the network was calculated. To update network parameters until The optimal feature extraction network is obtained by continuing until convergence. Step 4: The search process; Step 4.1: Query the semantic prototype features of all text queries in the test set. and semantic prototype features of all image libraries After each part is concatenated, the corresponding text query features are obtained. Image features of the image library ;in, This represents the text prototype feature of the q-th text query. Represents the visual prototype features of the g-th pedestrian image; Step 4.2: Calculate text query features Image features of the image library The similarity matrix is ​​obtained, and the Sinkhorn algorithm in optimal transmission is used to update the similarity matrix. After sorting the updated similarity matrix row by row, the image retrieval results corresponding to the text description are obtained.

2. The text-based person re-identification method based on cross-modal semantic alignment according to claim 1, characterized in that, The feature extraction network in step 2.2 is based on the CLIP-Transformer deep learning network and includes a text encoder and a visual encoder, wherein the text encoder is composed of... The visual encoder consists of Transformer residual blocks and is composed of... It consists of Transformer residual blocks, each of which is composed of a self-attention layer, a layer normalization layer, and a feedforward network layer with the same structure but different parameters. Will The input is fed into the feature extraction network and processed by the visual encoder to obtain the nth small block. pedestrian characteristics Thus, the i-th pedestrian image is obtained. pedestrian characteristics Where D is the feature dimension; Will The input is fed into the feature extraction network and processed by the text encoder to obtain the first... l Encoded word vectors Features Thus obtain pedestrian text features .

3. The text-based person re-identification method based on cross-modal semantic alignment according to claim 2, characterized in that, The feature alignment learning module in step 2.3 uses the Sinkhorn algorithm in optimal transmission to iteratively calculate the prototype features of the text portion. and visual prototype features Optimal plan for alignment between ; The optimal plan and visual prototype features After multiplication, the prototype features of the aligned text portion are obtained. ; The optimal plan Transpose and textual prototype features After multiplication, the aligned visual prototype features are obtained. .

4. The text-based person re-identification method based on cross-modal semantic alignment according to claim 3, characterized in that, The prediction module in step 2.4 will align the visual prototype features. The final visual prototype features are obtained after splicing. The prototype features of the aligned text portion The final text prototype feature is obtained after concatenation. ,Will and After classification processing by a fully connected layer and then by the softmax function, the corresponding classification probabilities are obtained. ;in, Represents the i-th pedestrian image The output features are classified into the first category. The probability of a pedestrian identity ID. This indicates the corresponding text description. The output features are classified into the first category. The probability of a pedestrian identity ID.

5. A text-based person re-identification method based on cross-modal semantic alignment according to claim 4, characterized in that, The network total loss function in step three It is constructed according to the following steps: Step 3.1: Construct the identity loss function using equation (1) : (1) In equation (1), and The i-th pedestrian image and the i-th text description The output is classified into The probability of; Step 3.2: Construct the triplet loss function for the image modality and the text modality using equation (2). : (2) In equation (2), Indicates the first i Zhang Xingren Image Aligned visual prototype features and They represent Prototype features of the text portion aligned with the same pedestrian identity ID and different pedestrian identity IDs; Represents the i-th text description Aligned text portion prototype features. and They represent Visual prototype features of the same type of pedestrian identity ID and different types of pedestrian identity ID; It is the margin parameter. This represents the function that takes the maximum value. Represents the cosine similarity function; Step 3.3: Construct the alignment loss function for the visual and textual prototype features using equation (3). : (3) In equation (3), Let represent the alignment function from the text part prototype features to the visual part prototype features, and obtain it from equation (4). The alignment function from the visual prototype features to the text prototype features is given by equation (5). (4) (5) In equations (4) and (5), It is the temperature hyperparameter that controls the probability distribution; Step 3.4: Construct the total network loss function using equation (6). : (5)。 6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing any of the text pedestrian re-identification methods of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of any of the text pedestrian re-identification methods described in claims 1-5.

Images