CN112069929A - Unsupervised pedestrian re-identification method and device, electronic equipment and storage medium - Google Patents

Abstract

The invention discloses an unsupervised pedestrian re-identification method, a device, electronic equipment and a storage medium, wherein the method comprises the following steps: pre-training a pedestrian re-recognition model in a labeled source domain data set; extracting training characteristics of a training set in the label-free target domain by using the model; according to the training characteristics, dividing a target domain training set into a plurality of clusters based on a self-adaptive clustering method, and distributing corresponding pseudo labels; setting each cluster as a prototype, selecting a sample with the distance from the center of the prototype to be smaller than a set threshold value from the prototype, and retraining the model by using the training characteristics and the pseudo labels of the sample to obtain a pedestrian re-identification model with updated parameters; and inputting the query set and the to-be-selected set of the target domain into the model, respectively obtaining the test characteristics of the query set and the to-be-selected set, and selecting pictures meeting the requirements of the query pictures from the to-be-selected set according to the similarity of the test characteristics. The method effectively relieves the problem of inter-domain separation and improves the accuracy of cross-domain pedestrian re-identification.

Description

An unsupervised pedestrian re-identification method, device, electronic device and storage medium

technical field

The invention belongs to the technical field of artificial intelligence and computer vision, and in particular relates to an unsupervised pedestrian re-identification method, device, electronic device and storage medium.

Background technique

With the acceleration of urbanization, public safety has become the focus and demand of people. Many important public health areas such as university campuses, theme parks, hospitals, and streets are widely covered by surveillance cameras, creating good objective conditions for automated surveillance using computer vision technology.

In recent years, person re-identification, as an important research direction in the field of video surveillance, has received increasing attention. Specifically, pedestrian re-identification refers to the technology of using computer vision technology to determine whether there is a specific pedestrian in an image or video sequence under cross-camera and cross-scene conditions. As an important supplement to face recognition technology, this technology can recognize pedestrians based on their clothing, posture, hairstyle and other information, and continuously track pedestrians whose faces cannot be clearly photographed in actual monitoring scenarios, enhancing the spatiotemporal data. Continuity helps save a lot of manpower and material resources, and has important research significance.

Thanks to the rapid development of deep neural networks, person re-identification technology based on supervised deep learning has been able to achieve very high recognition rates on mainstream public datasets. On the public Market-1501 dataset, the rank1 (first hit rate) has reached more than 95%, exceeding the recognition accuracy of the human eye. However, as an important vision task, person re-identification still has many challenges. In actual open application scenarios, the distribution of pedestrian data will vary greatly due to different seasons, clothing, lighting, and cameras. If the model trained in the labeled source domain data set is directly transferred to a new application scenario In the open environment, the generalization ability of the model is poor, the recognition performance is significantly reduced, and even the pedestrian recognition cannot be completed normally. Identify tasks.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present invention is to provide an unsupervised pedestrian re-identification method, device, electronic device and storage medium, so as to solve the problem of directly migrating the existing model trained in the labeled source domain data set to a new application scenario The problem of generating domain gaps.

In order to achieve the above object, the technical scheme adopted in the present invention is as follows:

In a first aspect, an embodiment of the present invention provides an unsupervised pedestrian re-identification method, including:

The pre-training step is used to pre-train the deep person re-identification model with supervised learning in the labeled source domain dataset;

The training feature extraction step is used to extract the training features of the training set samples in the unlabeled target domain by using the pre-trained deep person re-identification model;

The dividing step is used to divide the training set samples of the target domain into several clusters based on the adaptive clustering method according to the training features, and assign corresponding pseudo labels;

The retraining step is used to set each cluster as a prototype, and the samples in the cluster are the visible samples in the prototype, calculate the distance between the visible sample and the prototype center, and select the visible samples whose distance is less than the set threshold. sample, screen the training features according to the selected visual samples, obtain the training features after screening, and retrain the pre-trained deep pedestrian re-identification model by using the training features after screening and the assigned pseudo-labels, Obtain the deep pedestrian re-identification model after updating the parameters;

The identification step is used to input the query set of the target domain and the set to be selected into the deep pedestrian re-identification model after updating the parameters, respectively obtain the test features of the query set pictures and the test features of the pictures to be selected, and calculate the difference between the two in the metric space. Similarity, according to the similarity, the candidate pictures that meet the query picture requirements are identified from the candidate set.

Further, it also includes:

The iterative convergence step is used to repeat the training feature extraction step, the division step and the retraining step, and update the iterative weights until convergence.

Further, according to the described training features, the pedestrian images in the target domain are divided into several clusters by the adaptive clustering method, and corresponding pseudo-labels are assigned, including:

Calculate the distance between the pedestrian images in the target domain training set according to the training features to form a distance matrix;

Based on the distance matrix, use a density-based adaptive clustering algorithm to perform unsupervised clustering on pedestrian images in the target domain training set to generate several clusters;

After unsupervised clustering, each sample in the training set of the target domain is assigned a corresponding pseudo-label.

Further, each cluster is designated as a prototype, the samples in the cluster are the visible samples in the prototype, the distance between the visible samples in the prototype and the prototype center is calculated, and the visible samples whose distance is less than the set threshold are selected. , screen the training features according to the selected visual samples to obtain the filtered training features, and retrain the pre-trained deep pedestrian re-identification model by using the screened training features and the assigned pseudo-labels to obtain The deep person re-identification model after updating the parameters, including:

个簇设定为一个原型

，其中，

，

为簇的个数，簇

中每个对象设定为原型

的可视样本，计算对应的原型中心

；will

clusters are set as a prototype

,in,

,

is the number of clusters, the cluster

set each object as a prototype

The visual samples of , calculate the corresponding prototype center

;

中可视样本与原型中心

的距离，形成距离向量

，

中第

个元素

代表

与原型中心

的距离；Computational prototype

China Visual Sample and Prototype Center

distance, forming a distance vector

,

B

elements

represent

with Prototype Center

the distance;

中选出

小于阈值的可视样本，挑选方式如下：exist

selected

Visual samples smaller than the threshold are selected as follows:

代表原型

挑选出的可视样本，

为设置的距离阈值，

代表示性函数，若成立则为1，反之，则为0；in,

representative prototype

selected visual samples,

is the set distance threshold,

On behalf of the representative function, if it is established, it is 1, otherwise, it is 0;

对所述的训练特征进行相应的筛选，得到挑选后的训练特征，然后对预训练好的深度行人重识别模型进行训练，得到更新参数后的深度行人重识别模型。Use selected samples

The training features are screened accordingly to obtain the selected training features, and then the pre-trained deep pedestrian re-identification model is trained to obtain a deep pedestrian re-identification model with updated parameters.

的计算公式如下：Further, the prototype center

The calculation formula is as follows:

为原型

的可视样本个数，

为原型

的第

个样本，

。in,

as a prototype

the number of visible samples,

as a prototype

First

samples,

.

Further, input the query set of the target domain and the set to be selected into the deep pedestrian re-identification model after the update parameters, obtain the characteristics of the two respectively, calculate the similarity of the two in the metric space, and select the set from the set according to the similarity. Identify the candidate images that meet the query image requirements, including:

Input the query set of the target domain and the set to be selected into the deep pedestrian re-identification model to obtain the test features of the query set pictures and the test features of the pictures to be selected respectively;

Calculate the Euclidean distance of the test feature of the query set and the test feature of the set to be selected in the metric space, obtain the similarity matrix of the test feature of the query set and the test feature of the set to be selected, and identify the test feature from the set to be selected according to the similarity matrix Candidate images that meet the query image requirements.

In a second aspect, an embodiment of the present invention provides an unsupervised pedestrian re-identification device, including:

A pre-training unit for pre-training a deep person re-identification model with supervised learning in a labeled source-domain dataset;

The training feature extraction unit is used to extract the training features of the training set samples in the unlabeled target domain by using the pre-trained deep person re-identification model;

a dividing unit, which is used to divide the training set samples of the target domain into several clusters according to the training characteristics by using an adaptive clustering method, and assign corresponding pseudo labels;

The retraining unit is used to set each cluster as a prototype, calculate the distance between the visible samples in the prototype and the prototype center, select the visible samples whose distance is less than the set threshold, and analyze the selected visual samples according to the selected visual samples. The pre-trained deep pedestrian re-identification model is retrained by using the filtered training characteristics and assigned pseudo-labels to obtain the deep pedestrian re-identification model with updated parameters. ;

The identification unit is used to input the query set and the set to be selected in the target domain into the deep pedestrian re-identification model after updating the parameters, respectively obtain the test features of the query set pictures and the test features of the pictures to be selected, and calculate the difference between the two in the metric space. Similarity, according to the similarity, the candidate pictures that meet the query picture requirements are identified from the candidate set.

Further, it also includes:

The iterative convergence unit is used to repeatedly execute the training feature extraction unit, the division unit and the retraining unit, and update the iterative weights until convergence.

In a third aspect, an embodiment of the present invention provides an electronic device, including:

one or more processors;

memory for storing one or more programs;

When the one or more programs are executed by the one or more processors, the one or more processors implement an unsupervised pedestrian re-identification method as described in the first aspect.

In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, implements the method for unsupervised pedestrian re-identification described in the first aspect.

In the unsupervised pedestrian re-identification and device of the embodiment of the present invention, an adaptive clustering method is used to assign pseudo-labels to unlabeled target domains, and the adaptive clustering method does not need to set the cluster number in advance during clustering. It not only overcomes the difficulty of the uncertainty of the number of identity categories in the target domain, but also fully mines the visual information inside the target domain, providing a richer context for cross-domain transfer; the pseudo-labels assigned by adaptive clustering cannot It represents the real label, and the pseudo-label assigned by the target domain has a certain noise. In the unsupervised pedestrian re-identification and device of the embodiment of the present invention, the prototype optimization method is adopted, and trustworthy samples are targeted to participate in training, and the elimination of Noise samples that may degrade model performance are eliminated, improving recognition accuracy.

Description of drawings

The accompanying drawings described herein are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the attached image:

1 is a flowchart of an unsupervised pedestrian re-identification method provided by an embodiment of the present invention;

FIG. 2 is a block diagram of an unsupervised pedestrian re-identification device provided by an embodiment of the present invention.

Detailed ways

To make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions and specific operation processes in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention. The scope of protection is not limited to the following examples.

Example 1:

As shown in FIG. 1 , an embodiment of the present invention discloses an unsupervised pedestrian re-identification method, which includes the following steps:

The pre-training step S101 is used for pre-training a deep pedestrian re-identification model by using a supervised learning method in a labeled source domain data set;

采用深度残差神经网络，并在带标签的源域数据集上进行监督式训练，使

获得在源域上相对鲁棒的性能。In this step, the deep person re-identification model

Using deep residual neural networks and supervised training on the labeled source-domain datasets, the

Obtain relatively robust performance on the source domain.

；The training feature extraction step S103 is used to extract the training features of the training set samples in the unlabeled target domain by using the pre-trained deep pedestrian re-identification model

;

The dividing step S105 is used to divide the training set samples of the target domain into several clusters according to the training features by using an adaptive clustering method, and assign corresponding pseudo labels;

Specifically, it includes the following sub-steps:

，利用基于密度的自适应聚类算法DBSCAN对目标域训练集行人图像进行无监督聚类。具体步骤如下：Step S1052: Based on the existing distance matrix

, using the density-based adaptive clustering algorithm DBSCAN to perform unsupervised clustering of pedestrian images in the target domain training set. Specific steps are as follows:

，聚类簇

，未访问样本集合

，簇划分

；First, initialize the cluster core object sample set

, clusters

, the sample collection is not visited

, the cluster division

;

中每一个样本

，根据

∈

找到其

-领域子样本集

，如果

，则将

加入核心对象样本集

，其中，

表示聚类扫描半径，

表示每簇最小样本数。如果扫描完毕后，

，则结束。其中，设置

，扫描半径

的计算方式如下：将距离矩阵

的右上角展开，获得不重复的样本两两之间的距离，按照从小到大的顺序进行排列，取第t-top个距离设为

。Second, for the sample set

each sample in

,according to

∈

find its

- Domain subsample set

,if

, then the

Add to core object sample set

,in,

represents the cluster scan radius,

Indicates the minimum number of samples per cluster. If after scanning,

, it ends. Among them, set

, scan radius

is calculated as follows: the distance matrix

Expand the upper right corner of , to obtain the distance between pairs of non-repeated samples, arrange them in order from small to large, and take the t-top distance as

.

中，随机选择一个核心对象

，初始化当前簇核心对象队列

，类别序号

，当前簇样本集合

，更新未访问样本集合

；Next, in the core object sample collection

, randomly select a core object

, initialize the current cluster core object queue

, category number

, the current cluster sample set

, to update the collection of unvisited samples

;

取出一个核心对象

，根据

∈

找到其

，令

，更新

，

，

。如果当前簇核心对象队列

，则当前簇

生成完毕，更新

，核心对象集合

。如果

，则结束；Then, from

take out a core object

,according to

∈

find its

,make

,renew

,

,

. If the current cluster core object queue

, then the current cluster

Generated, updated

, the core object collection

. if

, then it ends;

，

更新为得到的簇总数。Finally, the output cluster partition

,

Update to the total number of clusters obtained.

The density-based DBSCAN clustering method does not need to determine the number of clusters in advance, and is more in line with the real scene of unsupervised pedestrian re-identification in an open environment.

Step S1053: After adaptive clustering, assign a corresponding pseudo-label to each sample of the target domain training set.

Retraining step S107 is used to define each cluster as a prototype, calculate the distance between the visible sample in the prototype and the center of the prototype, select the visible sample whose distance is less than the set threshold, and pair according to the selected visible sample. The training features are screened to obtain the screened training features, and the pre-trained deep pedestrian re-identification model is retrained by using the screened training features and the allocated pseudo-labels to obtain the deep pedestrian re-identification model after updating the parameters. Model;

Specifically, it includes the following sub-steps:

中每个簇

设定为一个原型

，簇

中每个对象设定为原型

的可视样本。计算对应的原型中心，计算方式如下：Step S1071: Divide clusters

each cluster in

set as a prototype

,cluster

set each object as a prototype

visual sample. Calculate the corresponding prototype center, the calculation method is as follows:

为原型

的可视样本个数。in,

as a prototype

the number of visible samples.

中可视样本与原型中心

的距离，形成距离向量

，

中每个元素

代表可视样本

与原型中心

的距离，

为原型

的可视样本个数。

计算方式如下：Step S1072: Calculate the prototype

China Visual Sample and Prototype Center

distance, forming a distance vector

,

each element in

represents a visual sample

with Prototype Center

the distance,

as a prototype

the number of visible samples.

It is calculated as follows:

Step S1073: According to the principle of self-paced learning, set a threshold, and automatically select the target domain training samples in the prototype that are close enough to the prototype center to participate in the training. The selection method is as follows:

为设置的距离阈值，

代表示性函数，若成立则为1，反之，则为0。in,

is the set distance threshold,

Represents a representative function, 1 if true, 0 otherwise.

进行训练。其中，经由步骤S103，对得到训练特征

进行相应的删选，得到挑选后的特征

，最小化三元组损失

。计算方式如下：Step S1074: the selected sample

to train. Wherein, through step S103, the training features are obtained for

Carry out the corresponding deletion and get the selected features

, which minimizes the triplet loss

. It is calculated as follows:

为

其中一个元素，其伪标签为

，取离它最近的负样本

，其伪标签为

，取离它最远的正样本

，其伪标签为

，记：make

for

One of the elements whose pseudo tag is

, take the nearest negative sample

, whose pseudo-label is

, take the positive sample farthest from it

, whose pseudo-label is

,remember:

，

为边界值，则度量损失

的计算公式为：make

,

is the boundary value, then measure the loss

The calculation formula is:

为hinge函数。in,

is the hinge function.

The identification step S109 is used to input the query set and the set to be selected in the target domain into the deep pedestrian re-identification model after updating the parameters, respectively obtain the test feature of the query set picture and the test feature of the set of pictures to be selected, and calculate the two in the metric space. According to the similarity, the candidate pictures that meet the query picture requirements are identified from the candidate set according to the similarity.

Specifically, it includes the following sub-steps:

和待选集

，

为查询集元素个数，

为待选集元素个数；

和G中元素皆为RGB图片，大小被调整为256×128×3，分别输入经由步骤S107后的深度行人重识别模型，获得相应特征集合，其中：Step S1091: the query set

and selections

,

is the number of query set elements,

is the number of elements to be selected;

The elements in and G are all RGB images, and the size is adjusted to 256×128×3, respectively input the deep pedestrian re-identification model after step S107 to obtain the corresponding feature set, where:

和

之间的欧氏距离，构建距离矩阵

，针对每个查询图片将待选图片按照距离大小进行排序，设置查询个数s，取距离较小的前s个待选图片作为该查询图片的检索候选列表；并用mAP和Rank@1对结果的准确率进行评估。Step S1092: Calculate

and

Euclidean distance between, build a distance matrix

, for each query picture, sort the pictures to be selected according to the size of the distance, set the number of queries s , and take the first s pictures to be selected with a smaller distance as the retrieval candidate list of the query picture; and use mAP and Rank@1 to compare the results accuracy is evaluated.

To improve the robustness and accuracy of the person re-ID model, it also includes:

The iterative convergence step S108 is used to repeat the training feature extraction step S103, the division step S105 and the retraining step S107, and update the weight of the iterative model M until convergence.

The following Table 1 shows the recognition accuracy results obtained based on the methods provided by the above embodiments of the present invention. From the top to the bottom, other benchmark methods for comparison are listed in order compared with the results implemented in this embodiment. It can be seen that the recognition performance of the above-mentioned embodiment of the present invention is well improved.

Table 1: Recognition Accuracy Results

To sum up, the embodiment of the present invention discloses an unsupervised pedestrian re-identification method, which utilizes the advantages of existing deep learning and extracts features through a deep residual neural network; constructs a method based on adaptive clustering and prototype optimization. Unsupervised pedestrian re-identification model, in which the adaptive clustering method can automatically assign pseudo-labels to pedestrians in the target domain, which not only overcomes the difficulty of the uncertain number of identity categories in the target domain, but also fully exploits the visual information inside the target domain. Provides a richer context for cross-domain transfer. Since the pseudo-labels assigned by adaptive clustering cannot represent real labels, the training samples in the target domain have a certain amount of noise. The method based on prototype optimization automatically selects trustworthy for the training process of the target domain. , removing untrusted samples that may degrade model performance. To sum up, the present invention effectively alleviates the problem of domain interval, improves the accuracy of cross-domain transfer of pedestrian re-identification, and has good robustness and general applicability.

Example 2:

As shown in FIG. 2 , this embodiment provides an unsupervised pedestrian re-identification device, which is a virtual device corresponding to the unsupervised pedestrian re-identification method provided in Embodiment 1, and the device has function modules corresponding to executing the method. and beneficial effects, the device includes:

A pre-training unit 901, used for pre-training a deep person re-identification model by using a supervised learning method in a labeled source domain data set;

The training feature extraction unit 903 is used to extract the training features of the training film set samples in the unlabeled target domain by using the pre-trained deep pedestrian re-identification model;

A dividing unit 905 is used to divide the training set samples of the target domain into several clusters by means of adaptive clustering according to the training features, and assign corresponding pseudo labels;

The retraining unit 907 is used to designate each cluster as a prototype, calculate the distance between the visible sample in the prototype and the prototype center, select the visible sample whose distance is less than the set threshold, and select the visible sample according to the selected visual sample. The training features are screened to obtain the screened training features, and the pre-trained deep pedestrian re-identification model is retrained by using the screened training features and the allocated pseudo-labels to obtain the deep pedestrian re-identification model after updating the parameters. Model;

The identification unit 909 is used to input the query set of the target domain and the set to be selected into the deep pedestrian re-identification model after updating the parameters, respectively obtain the test feature of the query set picture and the test feature of the set of pictures to be selected, and calculate the two in the metric space. According to the similarity, the candidate pictures that meet the query picture requirements are identified from the candidate set according to the similarity.

Further, it also includes:

The iterative convergence unit 908 is configured to repeatedly execute the training feature extraction unit, the division unit and the retraining unit, and update the iterative weights until convergence.

Example 3:

This embodiment provides an electronic device, including:

one or more processors;

memory for storing one or more programs;

When the one or more programs are executed by the one or more processors, the one or more processors implement an unsupervised pedestrian re-identification method as described in Embodiment 1.

Example 4:

This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the unsupervised pedestrian re-identification method described in Embodiment 1 is implemented.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

In the above-mentioned embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are only illustrative, for example, the division of the units may be a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components may be combined or Integration into another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of units or modules, and may be in electrical or other forms.

The units described as separate components may or may not be physically separated, and components shown as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes .

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the present invention. within the scope of protection.

Claims (10)

1. an unsupervised pedestrian re-identification method, is characterized in that, comprises: The pre-training step is used to pre-train the deep person re-identification model with supervised learning in the labeled source domain dataset; The training feature extraction step is used to extract the training features of the training set samples in the unlabeled target domain by using the pre-trained deep person re-identification model; The dividing step is used to divide the training set samples of the target domain into several clusters based on the adaptive clustering method according to the training features, and assign corresponding pseudo labels; The retraining step is used to set each cluster as a prototype, and the samples in the cluster are the visible samples in the prototype, calculate the distance between the visible sample and the prototype center, and select the visible samples whose distance is less than the set threshold. sample, filter the training features according to the selected visual samples, obtain the training features after screening, and retrain the pre-trained deep pedestrian re-identification model by using the training features after screening and the assigned pseudo-labels, Obtain the deep pedestrian re-identification model after updating the parameters; The identification step is used to input the query set of the target domain and the set to be selected into the deep pedestrian re-identification model after updating the parameters, respectively obtain the test features of the query set pictures and the test features of the pictures to be selected, and calculate the difference between the two in the metric space. Similarity, according to the similarity, the candidate pictures that meet the query picture requirements are identified from the candidate set. 2. a kind of unsupervised pedestrian re-identification method according to claim 1, is characterized in that, also comprises: The iterative convergence step is used to repeat the training feature extraction step, the division step and the retraining step, and update the iterative weights until convergence. 3. a kind of unsupervised pedestrian re-identification method according to claim 1 is characterized in that, according to described training characteristic, utilizes the method of adaptive clustering to divide the pedestrian image of target domain into several clusters, and assign corresponding Pseudo tags, including: Calculate the distance between the pedestrian images in the target domain training set according to the training features to form a distance matrix; Based on the distance matrix, use a density-based adaptive clustering algorithm to perform unsupervised clustering on pedestrian images in the target domain training set to generate several clusters; After unsupervised clustering, each sample in the training set of the target domain is assigned a corresponding pseudo-label. 4. A kind of unsupervised pedestrian re-identification method according to claim 1, is characterized in that, each cluster is designated as a prototype respectively, and the samples in the cluster are the visible samples in the prototype, and the visible samples in the prototype are calculated. The distance from the prototype center, select the visual samples whose distance is less than the set threshold, filter the training features according to the selected visual samples, obtain the training features after screening, and use the training features and distribution after screening. The good pseudo-label retrains the pre-trained deep person re-identification model, and obtains the deep person re-identification model with updated parameters, including: will

clusters

set as a prototype

,in,

,

is the number of clusters, the cluster

set each object as a prototype

The visual samples of , calculate the corresponding prototype center

; Computational prototype

China Visual Sample and Prototype Center

distance, forming a distance vector

,

B

elements

represent

with Prototype Center

the distance; exist

selected

Visual samples smaller than the threshold are selected as follows:

in,

representative prototype

selected visual samples,

is the set distance threshold,

On behalf of the representative function, if it is established, it is 1, otherwise, it is 0; Use selected samples

The training features are screened accordingly to obtain the selected training features, and then the pre-trained deep pedestrian re-identification model is trained to obtain a deep pedestrian re-identification model with updated parameters. 5. An unsupervised pedestrian re-identification method according to claim 4, wherein the prototype center

The calculation formula is as follows:

in,

as a prototype

the number of visible samples,

as a prototype

First

samples,

. 6. A kind of unsupervised pedestrian re-identification method according to claim 1, is characterized in that, inputting the query set picture and the to-be-selected set picture of the target domain into the deep pedestrian re-identification model after the updated parameters, respectively obtaining The characteristics of the two, calculate the similarity between the two in the metric space, and select the pictures that meet the requirements from the candidate set according to the similarity, including: Input the query set of the target domain and the set to be selected into the deep pedestrian re-identification model to obtain the test features of the query set pictures and the test features of the pictures to be selected respectively; Calculate the Euclidean distance of the test feature of the query set and the test feature of the set to be selected in the metric space, obtain the similarity matrix of the test feature of the query set and the test feature of the set to be selected, and identify the test feature from the set to be selected according to the similarity matrix Candidate images that meet the query image requirements. 7. An unsupervised pedestrian re-identification device, characterized in that, comprising: A pre-training unit for pre-training a deep person re-identification model with supervised learning in a labeled source-domain dataset; The training feature extraction unit is used to extract the training features of the training set samples in the unlabeled target domain by using the pre-trained deep person re-identification model; a dividing unit, which is used to divide the training set samples of the target domain into several clusters according to the training characteristics by using an adaptive clustering method, and assign corresponding pseudo labels; The retraining unit is used to set each cluster as a prototype, calculate the distance between the visible samples in the prototype and the prototype center, select the visible samples whose distance is less than the set threshold, and analyze the selected visual samples according to the selected visual samples. The pre-trained deep pedestrian re-identification model is retrained by using the filtered training characteristics and assigned pseudo-labels to obtain the deep pedestrian re-identification model with updated parameters. ; The identification unit is used to input the query set and the set to be selected in the target domain into the deep pedestrian re-identification model after updating the parameters, respectively obtain the test features of the query set pictures and the test features of the pictures to be selected, and calculate the difference between the two in the metric space. Similarity, according to the similarity, the candidate pictures that meet the query picture requirements are identified from the candidate set. 8. An unsupervised pedestrian re-identification device according to claim 7, characterized in that, further comprising: The iterative convergence unit is used to repeatedly execute the training feature extraction unit, the division unit and the retraining unit, and update the iterative weights until convergence. 9. An electronic device, characterized in that, comprising: one or more processors; memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement an unsupervised pedestrian re-identification method according to any one of claims 1-6. 10. A computer-readable storage medium on which a computer program is stored, characterized in that, when the program is executed by a processor, an unsupervised pedestrian re-identification as described in any one of claims 1-6 is realized method.

Images