A completely unsupervised luggage re-identification method based on dynamic queue depth clustering

Abstract

The application provides a completely unsupervised luggage re-identification method based on dynamic queue depth clustering, and the main steps include: creating a luggage data set LReID-50tohsouand without an identity tag, performing target detection by using YOLOv5, generating 5000 luggage images of different identities, and totally generating 50000 images. A local video acquisition environment is built, and video data of luggage check-in images and sorting lines are acquired. A DynQueueViT luggage re-identification model is deployed, advanced dynamic queue management features are adopted, memory requirements are reduced, and time constraints are used to reduce pseudo-label noise. The model decomposes the luggage image into global features, local features and mixed features, and performs dynamic queue updating training. The re-identification features are calculated through a multi-layer visual conversion module and a local feature data enhancement module. Finally, whether the luggage belongs to the same identity is judged through a distance formula. The model introduces time constraints and positive and negative sample matching algorithms, improves the re-identification accuracy, and evaluates the identity similarity of the luggage image by calculating the cosine distance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, and particularly relates to a completely unsupervised luggage re-identification method based on dynamic queue depth clustering. BACKGROUND

[0002] With the accelerated development of the domestic economy and the continuous improvement of people's living standards, more and more people choose to take a plane as a means of travel. According to the data in the "2023 November China Civil Aviation Main Production Index Statistics" released by the Ministry of Transport of the People's Republic of China, the total passenger transport volume of domestic flights in China reached 569 million person-times in 2023. In the process of taking a plane, usually a passenger will carry multiple pieces of luggage, which leads to the fact that the number of luggage handled by domestic airports is much larger than the number of passengers. In the face of such a large number of luggage, it is difficult for domestic airports to rely solely on manual sorting of luggage to meet people's travel needs. In this case, for airports, the importance of an automated luggage sorting system in transporting passengers is increasingly prominent. At present, the mainstream airport luggage automated sorting system adopts the form of Radio Frequency Identification (RFID) to sort luggage.

[0003] Radio Frequency Identification technology uses unique electromagnetic tags and electromagnetic readers for luggage identification and data transmission. As can be seen from the figure, the luggage sorting process is divided into two main stages: check-in and conveyor tracking. In the check-in stage, airport staff will attach an electromagnetic tag containing a unique identification code to the luggage to facilitate subsequent tracking and identification. In the subsequent conveyor tracking stage, a special electromagnetic reading and writing device is responsible for identifying and tracking the luggage. Specifically, when the luggage passes through the tag and enters the conveyor belt after check-in, it will move quickly along the conveyor belt, passing through multiple electromagnetic reading and writing device nodes. In this process, each node can identify the electromagnetic tag attached to the passing luggage and integrate the node information traversed by the electromagnetic tag to form the motion trajectory of the luggage, such as passing through node #1, node #3, node #4, etc.

[0004] Although Radio Frequency Identification technology has been widely used in luggage tracking and identification, it still has some inherent limitations. For example, high equipment cost, luggage tracking effect limited by the number of readers, electromagnetic interference may cause identification failure, etc. These problems limit the further development and promotion of automated luggage sorting systems. At the same time, in recent years, computer vision technology has shown excellent robustness, accuracy and real-time performance in tracking and monitoring pedestrians and vehicles, showing the potential to handle complex scenes. Therefore, in order to meet people's strong demand for efficient travel, it is necessary to upgrade the luggage tracking system based on Radio Frequency Identification to a luggage tracking system using computer vision technology.

[0005] A baggage tracking system in computer vision realizes the whole cross-camera baggage tracking process by recognizing, tracking and re-identifying the baggage in the captured image. Specifically, this process first involves identifying which objects in the image captured by the camera are baggage. Next, the system will monitor the movement trajectory of these bags in real time to complete the single-camera tracking task. The key last step is for the system to confirm the exact identity of these bags to match the trajectories of the bags captured by different cameras to construct the complete path of the bags. In the process of cross-camera tracking, the baggage re-identification technology plays a similar role to the electromagnetic tag, which identifies the identity of the baggage by analyzing the features of the baggage in the camera image. This is the core link to complete the cross-camera baggage trajectory matching and construct the complete baggage path.

[0006] Thanks to the rapid development of neural network technology in recent years, baggage recognition and tracking within a single camera have shown excellent performance. However, in the aspect of re-identification technology for cross-camera baggage identity matching, current solutions still do not fully meet the needs of practical applications, so this paper focuses on the method of baggage re-identification.

[0007] In the specific scenario of baggage sorting pipeline, object re-identification technology faces new challenges, which mainly come from the fact that the accuracy of re-identification is highly dependent on the quality of images captured by the camera and whether the appearance features of the baggage remain consistent in different scenarios. The operating environment of the baggage re-identification task is more complex compared to common pedestrian or vehicle re-identification scenarios, and complex environmental conditions often cause changes in the appearance features of the same baggage.

[0008] In the baggage sorting pipeline environment, cameras are installed at multiple key locations on the conveyor belt, and there are often other auxiliary sorting equipment near these locations.

[0009] The following briefly explains how auxiliary sorting equipment causes changes in the appearance and posture of baggage:

[0010] (1) Signal indicator light

[0011] There are multiple signal indicator lights on the baggage sorting pipeline that flash at regular intervals. The rapid alternation of light and dark causes the appearance color of some metal-painted baggage to change rapidly.

[0012] (2) Conveyor belt

[0013] The conveyor belt of the baggage sorting pipeline has both ups and downs and many curves. The baggage often reverses during its forward movement, and there is often a large difference in the posture of the baggage before and after the reversal.

[0014] (3) Mechanical swing arm

[0015] The luggage sorting pipeline uses mechanical swing arms for sorting, which change the posture of the luggage when changing the luggage conveying channel.

[0016] (4) Fences

[0017] The fences on the luggage sorting pipeline will block the luggage in the camera field of view.

[0018] In summary, due to the same piece of luggage may present significantly different surface color and running posture under different camera views, and sometimes the camera shot picture may not completely cover the target luggage. Therefore, in this complex and variable environment, the luggage re-identification task faces the challenge of accurately identifying a large number of target identities.

[0019] In addition, in the airport scene with a daily passenger flow of about 100,000 people, the luggage videos taken from different angles by the sorting pipeline every day provide rich data resources for the training of the luggage re-identification model. However, the large scale and complexity of these data make it infeasible to rely on traditional manual annotation methods. Therefore, fully unsupervised training algorithms become a better choice for training luggage re-identification models using large-scale data sets. However, most existing fully unsupervised re-identification training algorithms are mainly trained and tested on small-scale data sets. In the environment with smaller data volume, these algorithms need to consume a large amount of video memory resources for data comparison and clustering. This approach makes model training more difficult when dealing with large-scale data sets. Therefore, another challenge of unsupervised training of luggage re-identification is how to handle the training of large-scale data sets with limited video memory resources.

[0020] In addition, this paper relies on the "Key Technology Development of Aviation Logistics Multi-source Data Fusion and Coordination" school-level project led by the Second Institute of Civil Aviation Administration of China. In summary, in-depth research and development of luggage re-identification models not only an important scientific research topic, but also has great economic significance. SUMMARY

[0021] The purpose of the present application is to solve the problem that the existing fully unsupervised re-identification model has too high memory requirement during training, which makes it difficult to train on large-scale unlabeled luggage data sets. A fully unsupervised luggage re-identification model DynQueueViT based on dynamic queue deep clustering is proposed. The model reduces the memory requirement for training while improving the re-identification performance of the model.

[0022] The purpose of the present application is achieved as follows:

[0023] A completely unsupervised luggage re-identification model based on dynamic queue depth clustering, DynQueueViT, is proposed. The model aims to make full use of the large amount of unlabeled data generated by the luggage sorting pipeline, and improve the accuracy of luggage re-identification under limited GPU resources.

[0024] The specific method is as follows:

[0025] 1. A completely unsupervised luggage re-identification method based on dynamic queue depth clustering, characterized in that the method comprises the following steps:

[0026] Step 1, build a luggage dataset, create a luggage dataset LReID-50tohsouand without identity label. These videos are recorded by five cameras covering various positions of the sorting pipeline, with a total length of 148 hours, containing more than 100,000 different identities of luggage. The dataset is made using YOLOv5 tools. YOLOv5 is used for target detection, and the detection results are cropped to build the dataset. LReID-50tohsouand contains 5000 different identities of luggage, a total of 50000 pictures, and an average of 10 different angle photos for each luggage.

[0027] Step 2, build a local luggage sorting pipeline video collection environment, including multiple non-intersecting cameras in the luggage sorting pipeline, video stream transmission server;

[0028] Step 3, build a local luggage check-in video collection environment, including luggage check-in table shooting cameras, video stream transmission server;

[0029] Step 4, get luggage check-in images and luggage videos taken from different perspectives on the sorting line;

[0030] Step 5, deploy the local luggage re-identification feature calculation DynQueueViT model. The DynQueueViT luggage re-identification model does not store all the features of the training pictures in the video memory at once, but uses a dynamic queue to manage these features. As the training progresses, the features in the queue will be continuously updated to ensure that there are enough samples participating in the training while significantly reducing the demand for video memory. At the same time, the DynQueueViT luggage re-identification model uses video sequence time constraints on training data entering the model during training, which reduces the noise of pseudo-labels to a certain extent.

[0031] Step 6, use the local luggage re-identification feature calculation model to calculate the re-identification features between different luggage using multi-layer visual transformation modules and local feature data enhancement modules;

[0032] Step 7: Substitute the re-identification features between different bags into the distance formula; determine whether the two target bags belong to the same person;

[0033] 2. The creation process of the LReID-50tohsouand unlabeled baggage dataset in step 1. The creation steps include: creating a mini-batch detection dataset using LabelMe and video, initializing and fine-tuning the YOLOv5 network, generating videos labeled with baggage bounding boxes, extracting one image every 4 frames, cropping the bounding boxes to generate a backup dataset, and finally manually filtering out duplicate images and completing the manual annotation of the unlabeled dataset;

[0034] 3. In step 5, the DynQueueViT model uses a feature extraction network to decompose the luggage image input to the model into global features f. g Local features f l Comparison of features f with weighted reference data ;

[0035] 4. In step 5, the DynQueueViT model represents the global features f of the input image. g and local features f l Each model establishes a first-in, first-out (FIFO) dynamic queue. Simultaneously, to capture image information more comprehensively, a dynamic queue for storing mixed features is also introduced. When evaluating whether a sample pair is a positive or negative pair, the DynQueueViT baggage re-identification model primarily relies on global and local feature queues; however, when defining the contrastive loss to train the feature extraction network, the DynQueueViT baggage re-identification model chooses to use mixed features.

[0036] 5. In step 5, the DynQueueViT model introduces time constraints on video sequences when updating the dynamic queue. Specifically, in a real baggage sorting assembly line environment, multiple cameras simultaneously capture images of baggage, and the same piece of baggage may be captured by different cameras within the same time period. Therefore, using the capture timestamp attached to each image in the LRe-ID50thsouand dataset, images captured by different cameras within the same time period are grouped into the same training batch, and then this training batch of images is fed into the DynQueueViT baggage re-identification model for training.

[0037] 6. In step 5, the DynQueueViT model uses a positive-negative sample matching algorithm to find all luggage images belonging to the same identity in the dynamic queue and identifies them as the positive sample set. Correspondingly, the hybrid features obtained by fusing all image features in the positive sample set will be used as the feature representation of the anchor sample.

[0038] For all images {X} in the dynamic queue i, i = 1, 2, ..., M × batch_size}, calculate their global features f respectively. g and local features f l With all images {X} in the dynamic queue j The Euclidean distance between the corresponding features of the image in the dynamic queue (j = 1, 2, ..., M × batch_size) (including the image itself). During the process of assigning image features in the dynamic queue to different similarity sets, the positive and negative sample matching algorithm also assigns images in the queue to dissimilar sets and boundary sets. i The distance is relatively far, located at the total distance S total (X i X j The images that are furthest from the end of the ascending sort order constitute image X. i Dissimilar sets Image X i boundary set Then it is composed of those similar candidate sets that were excluded. Image X j Compared to those at a total distance S total (X i ,X j In ascending order, it is ranked earlier but does not belong to the candidate set. The image composition.

[0039] Regarding the selection of the positive sample set, from image X i Similar sets Select all samples whose distance is less than a threshold to form a positive sample set. In selecting the negative sample set, image X was chosen. i boundary set and dissimilar sets The negative sample set is composed of a portion of the images.

[0040] 7. The LocEnViT model in step 5, the model fuses the positive sample set. The global and local features of all images in the dataset are used to create a comprehensive global feature f. i mg and local features f i ml

[0041]

[0042] 8. In step 5, the DynQueueViT model uses a mixture of features contrastive loss L. m Compared with data augmentation, the loss L data The model is trained. The overall loss function L... total It can be represented as:

[0043] L total = βL m + (1 - β)L data

[0044] Mixed feature contrastive loss L m The feature extraction network is trained by using the positive sample set of real data. This loss is further divided into two parts: global mixed feature L mg contrastive loss and local mixed feature contrastive loss L ml :

[0045]

[0046] The contrastive loss can be expressed as:

[0047]

[0048] 9、Step 5 DynQueueViT model, when judging whether two pieces of luggage images belong to the same identity, LocEnViT luggage re-identification model performs cosine distance calculation:

[0049] BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a schematic diagram of the unlabeled luggage dataset LReID-50thsouand.

[0051] Figure 2 is a schematic diagram of a completely unsupervised re-identification model training framework.

[0052] Figure 3 is a schematic diagram of a deep clustering framework based on dynamic queue.

[0053] Figure 4 is a schematic diagram of the dynamic queue updating process. DETAILED DESCRIPTION

[0054] The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0055] As Figure 2As shown, the present application first constructs a luggage dataset: creates a luggage dataset LReID-50tohsouand without identity tags, uses YOLOv5 for target detection, and crops the detection results to construct the dataset. The dataset contains 5000 different identities of luggage, a total of 50000 pictures, and an average of 10 different angle photos for each piece of luggage. The local video acquisition environment is built: the luggage sorting pipeline and the luggage check-in video acquisition environment are built, including multiple cameras and video streaming servers. Luggage image and video acquisition: acquire luggage check-in images and videos taken from different views on the flow sorting line. DynQueueViT model deployment and training: deploy the DynQueueViT luggage re-identification model, use the first-in, first-out dynamic queue management feature to reduce the memory requirement, and use the time constraint of the video sequence to reduce the pseudo-label noise. The model decomposes the input luggage image into global features, local features, and hybrid features, and uses dynamic queue updating and time constraints for training. Re-identification feature calculation: use multi-layer visual conversion modules and local feature data enhancement modules to calculate the re-identification features between different luggage. Identity judgment: put the re-identification features between different luggage into the distance formula to judge whether two target luggage belong to the same identity.

[0056] The DynQueueViT model improves the accuracy of re-identification by introducing time constraints and positive and negative sample matching algorithms. When defining the contrast loss, the global and local feature queues select hybrid features, and the model uses hybrid feature contrast loss and data enhancement contrast loss for training. Finally, the cosine distance is calculated to determine whether two luggage images belong to the same identity, which includes the following steps:

[0057] 1. A fully unsupervised luggage re-identification method based on dynamic queue deep clustering, characterized in that the method comprises the following steps:

[0058] Step 1, construct a luggage dataset, create a luggage dataset LReID-50tohsouand without identity tags. These videos are recorded by five cameras covering various positions on the sorting pipeline, with a total duration of 148 hours, containing more than 100,000 different identities of luggage. The dataset is made using the YOLOv5 tool. YOLOv5 is used for target detection, and the detection results are cropped to construct the dataset. LReID-50tohsouand contains 5000 different identities of luggage, a total of 50000 pictures, and an average of 10 different angle photos for each piece of luggage.

[0059] Step 2, build a local luggage sorting pipeline video acquisition environment, including multiple non-intersecting cameras in the luggage sorting pipeline and a video streaming server.

[0060] Step 3, build a local baggage check-in video collection environment, including a baggage check-in table camera and a video streaming server;

[0061] Step 4, obtain baggage check-in images and baggage videos taken at different views on the flow sorting line;

[0062] Step 5, deploy a local baggage re-identification feature calculation DynQueueViT model. The DynQueueViT baggage re-identification model does not store all training picture features in the video memory at once, but uses a first-in, first-out dynamic queue to manage these features. As the training progresses, the features in the queue are continuously updated to ensure that there are enough samples participating in the training while significantly reducing the training requirements for video memory. At the same time, the DynQueueViT baggage re-identification model uses a time constraint on the training data entering the model during the training process, which reduces the noise of pseudo-labels to some extent.

[0063] Step 6, calculate the re-identification features between different bags using a local baggage re-identification feature calculation model, using a multi-layer visual conversion module and a local feature data enhancement module;

[0064] Step 7, substitute the re-identification features between different bags into the distance formula; determine whether the two target bags belong to the same identity;

[0065] 2, the baggage data set LReID-50tohsouand creation process in step 1. The production steps include: using LabelMe and video to produce a small batch of detection data sets, initializing and fine-tuning the YOLOv5 network, generating videos labeled with baggage bounding boxes, extracting one picture every 4 frames, cropping the bounding box to generate a backup data set, and finally manually screening out duplicate pictures and completing the manual annotation of the unlabeled data set;

[0066] 3, the DynQueueViT model in step 5 uses a feature extraction network to decompose the baggage image input into global features f g , local features f l and weight reference comparison features f data ;

[0067] 4, the DynQueueViT model in step 5 decomposes the global features f g and local features f lEach of them builds a dynamic queue of first-in-first-out. Meanwhile, in order to capture more comprehensive image information, a dynamic queue of mixed features is also introduced. When evaluating whether a sample pair belongs to positive sample pair or negative sample pair, DynQueueViT luggage re-identification model mainly relies on global and local feature queues, and when defining contrastive loss to train feature extraction network, DynQueueViT luggage re-identification model chooses to use mixed features.

[0068] 5、In step 5, DynQueueViT model introduces time constraints of video sequences when updating dynamic queue. Specifically, in the actual luggage sorting pipeline environment, multiple cameras will simultaneously shoot luggage, and the same piece of luggage is likely to be captured by different cameras in the same time period. Therefore, by using the shooting timestamps attached to each image in the LRe-ID50thsouand dataset, images taken by different cameras in the same time period are grouped into the same training batch, and the training batch images are sent to the DynQueueViT luggage re-identification model for training.

[0069] 6、In step 5, DynQueueViT model uses positive and negative sample matching algorithm to find all luggage images belonging to the same identity in the dynamic queue and marks them as positive sample set. Correspondingly, the mixed features obtained by fusing all image features in the positive sample set will be used as the feature representation of the anchor sample.

[0070] For all images {X i ,i=1,2…,M×batch_size} in the dynamic queue, their global features f g and local features f l are calculated respectively. The Euclidean distance between the corresponding features of all images {X j ,j=1,2…,M×batch_size} in the dynamic queue (including the image itself) is calculated. The image features in the dynamic queue are assigned to different similarity sets, and the positive and negative sample matching algorithm will also assign the images in the queue to dissimilar sets and boundary sets. Those images far away from image X i , which are located at the back of the sorted image, constitute the dissimilar set of image X total i j The boundary set of image X i is composed of those images X i that are excluded from the similar candidate set j total i j ​​​​​​​​In ascending order, it is ranked earlier but does not belong to the candidate set. The image composition.

[0071] Regarding the selection of the positive sample set, from image X i Similar sets Select all samples whose distance is less than a threshold to form a positive sample set. In selecting the negative sample set, image X was chosen. i boundary set and dissimilar sets The negative sample set is composed of a portion of the images.

[0072] 7. The LocEnViT model in step 5, the model fuses the positive sample set. The global and local features of all images in the dataset are used to create a comprehensive global feature f. i mg and local features f i ml :

[0073] 8. In step 5, the DynQueueViT model uses a mixture of features contrastive loss L. m Compared with data augmentation, the loss L data The model is trained. The overall loss function L... total It can be represented as: L total =βL m +(1-β)L data .

[0074] Hybrid Feature Contrast Loss L m The feature extraction network is trained using a positive sample set of real data. This loss is further subdivided into two parts: global mixture features L... mg Contrast loss and local blended features contrast loss L ml :

[0075]

[0076] The contrast loss can be expressed as:

[0077] 9. In step 5, the DynQueueViT model calculates the cosine distance when determining whether two baggage images belong to the same person.

Claims

1. A fully unsupervised luggage re-identification method based on dynamic queue depth clustering, characterized in that, The method comprises the following steps: Step 1, constructing a luggage data set, creating a luggage data set LReID-50tohsouand without an identity label; these videos are recorded by five cameras covering various positions of the sorting pipeline, with a total duration of 148 hours, containing more than 100,000 different identities of luggage; the data set is made using the YOLOv5 tool; YOLOv5 is used for target detection, and the detection results are cropped to construct the data set; LReID-50tohsouand contains 5000 different identities of luggage, a total of 50000 pictures, and an average of 10 different angle photos for each piece of luggage; Step 2, constructing a local luggage sorting pipeline video acquisition environment, including multiple non-intersecting cameras of the luggage sorting pipeline, a video stream transmission server; Step 3, constructing a local luggage check-in video acquisition environment, including a luggage check-in table shooting camera, a video stream transmission server; Step 4, obtaining luggage check-in images and luggage shooting videos at different views on the sorting line; Step 5, deploy the local luggage weight recognition feature calculation DynQueueViT model; the DynQueueViT luggage weight recognition model does not store the features of all training pictures in the video memory at one time, but uses a first-in, first-out dynamic queue to manage these features; as the training progresses, the features in the queue are continuously updated to ensure that there are enough samples participating in the training while significantly reducing the demand for video memory for training; at the same time, the DynQueueViT luggage weight recognition model uses the time constraint of the video sequence for the training data entering the model during the training process, which reduces the noise of the pseudo label to a certain extent; wherein the DynQueueViT model uses a feature extraction network to decompose the luggage image input into global features f g , local features f l and weight reference comparison features f data ; Step 6, calculating the re-identification features between different pieces of luggage through a local luggage re-identification feature calculation model, using a multi-layer visual conversion module and a local feature data enhancement module; Step 7, substituting the re-identification features between different pieces of luggage into the distance formula; determine whether the two target pieces of luggage belong to the same identity.

2. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, The process of creating the luggage data set LReID-50tohsouand without an identity label in step 1; the production steps include: using LabelMe and video to make a small batch of detection data set, initializing and fine-tuning the YOLOv5 network, generating videos labeled with luggage bounding boxes, extracting one picture every 4 frames, cropping the bounding box to generate a backup data set, and finally manually screening out duplicate pictures and completing the manual deduplication of the unlabeled data set.

3. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, The DynQueueViT model in step 5 is a global feature f g and a local feature f l Each establishes a first-in, first-out dynamic queue; at the same time, in order to more comprehensively capture image information, a dynamic queue for storing mixed features is also introduced; in evaluating whether a sample pair belongs to a positive sample pair or a negative sample pair, the DynQueueViT baggage re-identification model mainly relies on the global and local feature queues, and in defining a contrast loss for training the feature extraction network, the DynQueueViT baggage re-identification model selects to use mixed features.

4. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, In step 5, the DynQueueViT model introduces a time constraint for video sequences when updating the dynamic queue; specifically, in the actual luggage sorting pipeline environment, multiple cameras will simultaneously shoot luggage, and the same piece of luggage is likely to be captured by different cameras within the same time period; therefore, using the shooting time stamp attached to each image in the LRe-ID50thsouand data set, images taken by different cameras within the same time period are grouped into the same training batch, and the training batch images are then fed into the DynQueueViT luggage re-identification model for training.

5. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, In step 5, the DynQueueViT model uses a positive and negative sample matching algorithm to find all luggage images belonging to the same identity in the dynamic queue and marks them as a positive sample set; correspondingly, the mixed feature obtained by fusing the features of all images in the positive sample set will be used as the feature representation of the anchor sample; For all images {X} in the dynamic queue i For each of the following groups, i = 1, 2, ..., M × batch_size, calculate its global feature f. g and local features f l With all images {X} in the dynamic queue j The Euclidean distance between the corresponding features of the images in the dynamic queue, j = 1, 2, ..., M × batch_size, including the images themselves; during the process of assigning image features in the dynamic queue to different similarity sets, the positive and negative sample matching algorithm will also assign images in the queue to dissimilar sets and boundary sets; those features that are similar to image X i The distance is relatively far, located at the total distance S total (X,X j The images that are furthest from the end of the ascending sort order constitute image X. i Dissimilar sets Image X i boundary set Then it is composed of those similar candidate sets that were excluded. Image X j Compared to those at a total distance S total (X i ,X j In ascending order, it is ranked earlier but does not belong to the candidate set. Image composition; Regarding the selection of the positive sample set, from image X i Similar sets Select all samples whose distance is less than a threshold to form a positive sample set. In selecting the negative sample set, image X was chosen. i boundary set and dissimilar sets The negative sample set is composed of a portion of the images.

6. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, The step 5 has a LocEnViT model which creates comprehensive global and local features by fusing global and local features of all images in the positive sample set ​​​ 7. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, The DynQueueViT model in step 5 uses a hybrid feature contrastive loss L m and data augmentation contrastive loss L data Train the model; overall loss function L total is represented as: L total = βL m +(1-β)L data Mixed feature contrastive loss L m By utilizing the positive sample set of real data to train the feature extraction network; this loss is further subdivided into two parts: global mixed feature L mg Contrastive loss and local mixed feature contrastive loss L ml : The contrast loss is expressed as:

8. The fully unsupervised luggage re-identification method based on dynamic queue depth clustering according to claim 1, wherein, In step 5, when judging whether two pieces of luggage belong to the same identity, the LocEnViT luggage re-identification model calculates the cosine distance:

Images