Unsupervised cross-domain pedestrian re-identification method based on location attention and refined clustering

Abstract

The present application is an unsupervised cross-domain pedestrian re-identification method based on position attention and refined clustering, which overcomes the problems of pseudo label noise generated by clustering and insufficient feature discrimination in the prior art. The present application comprises the following steps: (1) based on the ResNet50 network, a position attention module is added before the layer1 layer and after the layer4 layer; before the start of the depth convolution of the model, the position information of the sample is focused by using the position attention block to perform preliminary feature extraction; after the model extracts the features, the position attention block is used again to mine the features and extract more fine-grained features; (2) a refined clustering algorithm module is used to remove noise interference samples to realize a clustering process from coarse to fine; (3) a memory module is introduced, and the class center obtained by each period of training is stored in the memory module.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and relates to an unsupervised cross-domain pedestrian re-identification method based on position attention and refined clustering. BACKGROUND

[0002] With the development of deep learning and the increasing demand for public safety, pedestrian re-identification has become an important research topic in the field of computer vision. Pedestrian re-identification, also known as pedestrian re-identification, uses deep learning methods to determine whether the person captured by the camera deployed at different locations has consistent identity information with the candidate person, and to re-identify the process. Pedestrian re-identification has great practical significance, and its technology can be combined with pedestrian detection algorithms and widely used in intelligent security and search for lost children and other social security issues.

[0003] Pedestrian re-identification technology mainly includes supervised and unsupervised directions. The supervised method is easily affected by environmental factors, resulting in insufficient generalization ability. The unsupervised method learns the information of labeled source domain data and unlabeled target domain data, and the samples collected by the two data sets are different. Therefore, in actual application, the unsupervised pedestrian re-identification method is becoming more and more important.

[0004] The clustering-based pedestrian re-identification strategy generates pseudo-labels by clustering and uses the pseudo-labels for subsequent training. Compared with other unsupervised methods, the clustering-based method has obvious advantages, without the need for a large amount of manpower and material resources to manually annotate data, and without pre-modeling. However, it has two problems, one is the pseudo-label noise generated by clustering, and the other is the insufficient feature discrimination. The pseudo-label prediction pedestrian re-identification method usually produces outliers, and most works choose to directly discard the outlier samples. Some scholars propose a bottom-up clustering (BUC) framework, which iteratively trains the network with pseudo-labels. Some scholars introduce a clustering validity criterion and propose a simple and practical density-based clustering method. Some scholars propose a self-similarity grouping (SSG) method, which uses the potential similarity of unlabeled samples to automatically construct multiple clusters from different views. Some scholars first assign a multi-label vector to each image, and after several cycles of multi-label training, they use clustering methods to assign a pseudo-label to each image. Some scholars use similarity calculation and cyclic consistency to ensure the quality of the pseudo-labels. Some scholars exclude outliers based on clustering results to ensure that the model selects samples as much as possible during training. Some scholars improve the clustering algorithm, assign pseudo-labels to noise samples, and propose a dynamic and symmetric cross-entropy loss to resist the noise labels generated by clustering. Some scholars increase the clustering points of the target domain samples and expand the data to improve the discriminability. Some scholars propose a new adaptive contrastive learning framework, and the proposed adaptive method gradually produces more reliable clusters to refine mixed memories and learning goals. Some scholars propose a new component-based pseudo-label refinement that reduces label noise by using the complementary relationship between global and partial features. Although the above methods use clustering to generate labels, there is still a noise problem, and some methods choose to directly discard the noise samples to ensure the reliability of the pseudo-labels. However, the discarded samples contain important pedestrian information, and simply and roughly removing them will only harm the performance of the model.

[0005] For the problem of insufficient feature discrimination, there are great differences in illumination, background and perspective under different environments. Even the same pedestrian under different cameras, the extracted features will be different. Some scholars propose a "squeeze and activation" attention block to connect channel attention and spatial attention. Some scholars propose a new method of improving the representation ability of CNN network, a convolution bottleneck attention module. Some scholars propose a new domain adaptive attention model that maps features to domain shared feature mapping and domain specific feature mapping. Some scholars propose a unified learning framework that combines attribute information and attention. Some scholars propose a network architecture that integrates attention mechanism and diversity regularization. Although the above methods have greatly improved the performance of the model, they ignore the position information, which is very important for generating spatial selective attention, and the position information is even more important in the re-identification framework. Ignoring it will only damage the accuracy of the model. SUMMARY

[0006] The purpose of the present application is to provide an unsupervised cross-domain pedestrian re-identification method based on position attention and refined clustering, which overcomes the problems of pseudo label noise generated by clustering and insufficient feature discrimination in the prior art. The present application uses position attention to focus on position information to solve the problem of insufficient feature discrimination; a refined clustering algorithm is proposed to improve the reliability of clustering to solve the problem of pseudo label noise generated by clustering.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is:

[0008] An unsupervised cross-domain pedestrian re-identification system based on position attention and refined clustering, characterized by: comprising a backbone network module, a position attention module, a clustering algorithm module and a memory module; wherein:

[0009] The backbone network module: uses ResNet50 to extract the initial features of the image;

[0010] The position attention module: extracts more fine-grained features, and the position is to add position attention before layer1 and after layer4 of ResNet50;

[0011] The clustering algorithm module: generates pseudo labels by clustering, and uses a refined clustering method to overcome the influence of noise on the model;

[0012] The memory module: stores the class centers obtained by each cycle of training in the memory module to facilitate subsequent training.

[0013] An identification method for an unsupervised cross-domain pedestrian re-identification system based on position attention and refined clustering, characterized by: comprising the following steps

[0014] (1) Based on ResNet50 network, the position attention module is added before layer1 layer and after layer4 layer; before the beginning of the depth convolution of the model, the position information of the sample is focused by using the position attention block to perform preliminary feature extraction; after the model extracts the features, the position attention block is used again to mine the features and extract more fine-grained features;

[0015] (2) The fine clustering algorithm module is used to remove noise interference samples to realize the clustering process from coarse to fine;

[0016] (3) The memory module is introduced, and the class center obtained by each cycle training is stored in the memory module.

[0017] In step (1), the position attention module is added to the backbone network to extract more fine-grained features, including coordinate information embedding and coordinate information generation. Specifically, the following steps are included:

[0018] Given the input X, first use the pooling kernel with size (H, 1) or (1, W) to encode each channel along the horizontal and vertical coordinates respectively; the output of the c-th channel with height h is represented as:

[0019]

[0020] wherein represents the output of the c-th channel with height h; similarly, the output of the c-th channel with width w and height h is represented as:

[0021]

[0022] wherein represents the output of the c-th channel with width w;

[0023] Then the feature maps along the horizontal and vertical directions are spliced together, and then they are sent into a 1*1 convolution to generate an intermediate feature map m, which is represented as:

[0024] m = δ (F ([Z h ,z w ])) (3)

[0025] wherein [, ] represents concatenation operation, δ is a nonlinear activation function, and F is a 1*1 convolution change function; then the spatial dimension is divided into two independent tensors, and another two 1*1 convolutions are used to change the two independent tensors to obtain two tensors with the same number of channels as the input X;

[0026] g h = sigmoid (F h (m h )) (4)

[0027] g w = sigmoid(F w (m w )) (5)

[0028] where sigmoid() denotes the sigmoid activation function.

[0029] Finally, note that the output of the block Y is represented as:

[0030] Y(i,j) = x(i,j) x g h x g w (6)

[0031] In step (2), the fine-grained clustering algorithm specifically includes the following steps:

[0032] First, the Jaccard is calculated using the position attention feature Y(i,j) to perform preliminary clustering and generate pseudo labels; in the clustering process, by setting different clustering radii, three clustering centers with the same radius and different clustering radii are generated, and by cross-merging the three different clusters, an independence score is obtained, and the reliability is measured by using the score to determine the reliability of the clustering instance;

[0033] The method of measuring the independence of the cluster is defined by the cross-merging (IoU) score:

[0034]

[0035] where I(f i t ) represents the samples within the same f i t cluster, I loose (f i t ) is a cluster set containing f i t when the clustering condition becomes loose, I tight (f i t ) is a cluster set containing f i t when the clustering condition becomes tight; R indep (f i t ) is larger, indicating that the cluster f i t is more independent, where the value of R indep (f i t ) is between 0 and 1;

[0036] Before each cycle, clustering is performed, and screening of the clustering is performed according to a clustering reliability measure, and the clustering with an independence score between 0 and 1 is regarded as reliable clustering, and the rest is regarded as unreliable clustering; only reliable clustering is reserved, and the rest of the samples are regarded as clustering outliers; the outliers continue to be clustered in the subsequent process until reliable clustering is generated; with the increase of the number of iterations, unreliable clustering is screened according to the reliability determination criterion, and a coarse-to-fine clustering process is realized; and the center point of the reliable clustering is stored in a memory module for subsequent training.

[0037] In step (3), the memory module is introduced and specifically includes the following steps:

[0038] First, the initialization of the memory module is performed, each cluster is regarded as a class, and the average feature of each cluster is used to initialize the class-level storage body K[i], and specifically represented as:

[0039]

[0040] Where I k represents a reliable cluster containing feature instances x i , |. | represents the number of features of the cluster, v i represents a target domain feature instance;

[0041] Then, the module is updated, all instances in the kth cluster are updated in each cycle, and a memory of the memory module is constructed to store the center point of the cluster; when each clustering is completed, the memory module is updated, and the center point of the reliable cluster is fed into the module;

[0042] K[i]←=K[i]+(1-β)v i (9)

[0043] Where β∈[0,1] is an update rate.

[0044] The memory module stores features, so that the target domain samples can be continuously updated even in the case that clustering and non-clustering outliers are constantly changing.

[0045] Compared with the prior art, the application has the following advantages and effects:

[0046] 1. The present application proposes an unsupervised cross-domain pedestrian re-identification method based on position attention and refined clustering. The present application embeds a position attention block into the backbone network based on the ResNet50 network to extract more discriminative features. A refined clustering strategy is used to remove noise interference samples and achieve a coarse-to-fine clustering process. The memory module dynamically stores the class center for invariance learning. The problems of insufficient feature discrimination and clustering pseudo-label noise are solved. The present application is fully experimentally verified on the Market1501 and DukeMTMC-ReID public data sets, and compared with related algorithms. The results show that the method proposed by the present application has the advantages of high accuracy and strong robustness in the unsupervised cross-domain scene.

[0047] 2. The present application adopts the following methods to solve the two problems of pseudo-label noise generated by clustering and insufficient feature discrimination: position attention is used to focus on position information to solve the problem of insufficient feature discrimination; a refined clustering algorithm is proposed to improve the reliability of clustering to solve the problem of pseudo-label noise generated by clustering; in addition, the present application uses a memory module to store the clustering center point for invariance learning to adapt to the change of clustering.

[0048] 3. Compared with the BUC, DBC and HCT methods based on pure clustering, the method proposed by the present application has better performance than them through pre-training of the source domain. Compared with the MMCL, SSG, MLC and D-MMD methods based on unsupervised cross-domain, the method proposed by the present application uses an unsupervised cross-domain method based on clustering, which uses a refined clustering method to train and pay attention to the information of noise samples without discarding them. Compared with the PTGAN and TJ-AIDL methods based on GAN, the method of the present application has good performance in mAP precision and Rank, which shows that the method proposed by the present application is much better than the method based on GAN. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The principle block diagram of the method of the present application is shown in the figure;

[0050] Figure 2 The detail diagram of the backbone network of the present application is shown in the figure;

[0051] Figure 3 The position attention structure diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0053] In view of the two problems of insufficient feature discrimination and pseudo label noise generated by clustering, the present application proposes an unsupervised cross-domain pedestrian re-identification method based on position attention and refined clustering, and a framework diagram of the method is as shown in Figure 1

[0054] From Figure 1 It can be seen that the present framework mainly includes four parts, which are backbone network, position attention, clustering algorithm and memory module. The backbone network uses ResNet50 to extract the initial features of the image; the position attention is introduced to extract more fine-grained features, and the position is before layer1 and after layer4 of ResNet50; a refined clustering method is proposed to overcome the influence of noise on the model; the memory module is introduced to store the class centers obtained in each cycle of training to facilitate subsequent training. The implementation details of each module are as follows:

[0055] Position attention:

[0056] ResNet50 is used as the backbone network to extract image features. ResNet50 includes 1 convolutional layer and 4 residual modules, each of which includes multiple convolutional layers, batch normalization layers and linear activation functions. Because of the influence of lighting conditions, background and viewing angle, the features extracted by ResNet50 are insufficient in discrimination, and the model precision is poor, so the present application introduces position attention to extract more fine-grained features.

[0057] Referring to Figure 2 , the structure of the backbone network is fine-tuned, and the position attention is added before layer1 and after layer4. Before the start of deep convolution of the model, the position attention block is used to focus on the position information of the sample to perform preliminary feature extraction to facilitate subsequent convolution training. In addition, after the model extracts the features, the position attention block is used again to mine the features to extract more fine-grained features and improve the discrimination of the model

[0058] Referring to Figure 3 , in view of the problem of insufficient feature discrimination, the position attention block is added to the backbone network to extract more fine-grained features to alleviate the loss of position information caused by global pooling, which includes coordinate information embedding and coordinate information generation.

[0059] Given the input X, first use the pooling kernel with size (H, 1) or (1, W) to encode each channel along the horizontal and vertical coordinates respectively. Therefore, the output of the c-th channel with height h can be expressed as:

[0060] ​

[0061] wherein represents the output of the c-th channel with height h. Similarly, the output of the c-th channel with width w and height h can be represented as:

[0062]

[0063] wherein represents the output of the c-th channel with width w.

[0064] The feature maps along the horizontal and vertical directions are then concatenated together, and then they are fed into a 1x1 convolution to generate an intermediate feature map m, which is represented as:

[0065] m = δ (F ([Z h ,Z w ])) (3)

[0066] wherein [, ] represents a concatenation operation, δ is a nonlinear activation function, and F is a 1*1 convolution transformation function. Then the spatial dimension is divided into two independent tensors, and two independent tensors are transformed by using two other 1x1 convolutions to obtain two tensors with the same number of channels as the input X.

[0067] g h = sigmoid (F h (m h )) (4)

[0068] g w = sigmoid (F w (m w )) (5)

[0069] wherein sigmoid() represents a sigmoid activation function.

[0070] Finally, the output of the attention block Y is represented as:

[0071] Y(i,j) = x(i,j) x g h x g w (6)

[0072] Refinement clustering algorithm:

[0073] The clustering-based pedestrian re-identification method inevitably produces outliers, and simply removing them will damage the model accuracy. To solve this problem, the present application proposes a fine clustering strategy, which determines the reliability of the clustering instance by designing a reliability measurement criterion, removes the interference of noise samples, and realizes the clustering process from coarse to fine.

[0074] The specific fine clustering strategy is to first calculate the Jaccard using the position attention feature Y(i,j) to perform preliminary clustering and generate pseudo labels. In the clustering process, different clustering radii are set to generate three clustering centers with the same radius and different clustering radii. The independence score is obtained by performing intersection over union (IoU) on the three different clusters. The reliability is measured using the obtained score to determine the reliability of the clustering instances.

[0075] The method for measuring the independence of the cluster is defined by the intersection over union (IoU) score.

[0076]

[0077] where I(f i t ) represents the same f i t within the cluster, I loose (f i t ) is a cluster set containing f i t when the clustering condition becomes loose, and I tight (f i t ) is a cluster set containing f i t when the clustering condition becomes tight. R indep (f i t ) is larger, indicating that the cluster f i t is more independent, where the value of R indep (f i t ) is between 0 and 1.

[0078] Before each cycle, clustering is performed, and the reliable clustering is selected according to the clustering reliability measurement. Clusters with an independence score between 0 and 1 are considered reliable clusters, and the rest are considered unreliable clusters. Only reliable clusters are retained, and the rest of the samples are considered as clustering outliers. Outliers will continue to be clustered in subsequent processes until reliable clusters are generated. With the increase of the number of iterations, unreliable clusters are screened according to the reliability determination criterion, to realize the clustering process from coarse to fine. In addition, the center point of the reliable cluster is stored in the memory module for subsequent training.

[0079] Memory module:

[0080] In order to improve the generalization ability of the network on the target test set, invariance learning of the network is proposed by estimating the similarity between target images.

[0081] First, the initialization of the memory module, each cluster is regarded as a class, using the average feature of each cluster to initialize the class-level memory K[i]. Specifically represented as:

[0082]

[0083] Where I k represents the reliable cluster containing feature instance x i , |. | represents the number of features of the cluster, v i represents the target domain feature instance.

[0084] Then the module is updated, according to all instances in the kth cluster, constantly updated in each cycle, a memory of the memory module is constructed to store the center point of the cluster. At each clustering, it is updated, and the center point of the reliable cluster is fed into the module.

[0085] K[i]←=K[i]+(1-β)v i (9)

[0086] Where β∈[0,1] is the update rate.

[0087] The memory module stores the features, so that the target domain samples can still be updated and store the cluster center to adapt to the changes of the cluster even in the case of constantly changing clustering and non-clustering outliers.

[0088] Embodiment:

[0089] The entire network is built using the Pytorch framework, and the input picture size is 256*128 pixels. Random data augmentation is applied to each image before the image is sent into the network, including random flipping, cropping and erasing. The source domain is pre-trained for 50 epochs, the batch size 1 is set to 128, the initial learning rate α1 is set to 0.0003, the target domain is trained for 70 epochs, the initial learning rate α2 is set to 0.00035, the batch size 2 is set to 64, the original cluster radius eps is set to 0.6, the sparse cluster radius eps_loose is set to 0.62, the compact cluster radius eps_tight is set to 0.58, denoted as Δd = 0.02, the hyperparameters k1 of the jaccard distance are set to 30 and k2 are set to 6, and the memory module update rate β is set to 0.2.

[0090] To verify the rationality of the method, ablation experiments are carried out on two benchmark datasets Market-1501 and DukeMTMC-reID, and four model methods of baseline network plus DBSCAN clustering (Baseline+Cluster), baseline network combined with clustering and position attention (Baseline+CA+Cluster), baseline network plus refined clustering (Baseline+Refine-Cluster), and baseline network combined with refined clustering and position attention (Baseline+CA+Refine-Cluster) are used for experimental analysis, and the performance evaluation results are shown in Table 1 and Table 2.

[0091] Table 1 Performance evaluation results of different models on Market1501 dataset

[0092]

[0093] Table 2 Performance evaluation results of different models on DukeMTMC-reID dataset

[0094]

[0095] From Table 1 and Table 2, it can be concluded that the new method has an identification accuracy Rank-1 of 13.8% and mAP of 22.5% higher than the benchmark model method when the source domain is DukeMTMC-ReID and the target domain is Market-1501, and an identification accuracy Rank-1 of 15.1% and mAP of 11.2% higher when the source domain is Market-1501 and the target domain is DukeMTMC-ReID. The results show that the position attention module overcomes the problem of insufficient feature discrimination, and the refined clustering method overcomes the problem of pseudo-label noise caused by clustering. If the position attention and refined clustering methods are added to the model at the same time, the model precision reaches the best effect.

[0096] To objectively evaluate the performance of the method, the algorithm proposed in the application is compared with typical unsupervised cross-domain pedestrian re-identification methods on the Market-1501 and DukeMTMC-reID benchmark datasets, and the performance is compared on the Market-1501 public dataset. Among the 11 comparison methods, BUC, DBC, SSG, MLC, MMCL, AD-Cluster, HCT, SSL, D-MMD, TJ-AIDL, ADTC, and HHL are analyzed according to the evaluation indicators mAP and Rank to verify the effectiveness of the proposed pedestrian re-identification algorithm. The experimental results are shown in Table 3 and Table 4, and the upper right superscript in bold font in the table represents the optimal, better and less optimal results in the comparison methods.

[0097] Table 3 Comparison of algorithm recognition performance on Market1501 dataset

[0098]

[0099] Table 4 Comparison of algorithm recognition performance on DukeMTMC-reID dataset

[0100]

[0101] From the results in Table 3 and Table 4, compared with other methods on Market1501 and DukeMTMC-ReID public datasets, the method has been improved. Specifically, compared with BUC, DBC and HCT based on pure clustering methods, the method proposed in the application is obviously better than them through the pre-training of the source domain. Compared with MMCL, SSG, MLC and D-MMD based on unsupervised cross-domain methods, the method proposed in the application adopts a clustering-based unsupervised cross-domain method, which uses a fine clustering method for training, focuses on the information of noise samples and does not discard them. Compared with PTGAN and TJ-AIDL based on GAN, the method of the application achieves good performance in mAP precision and Rank, which shows that the method proposed in the application is much better than the method based on GAN.

[0102] The above description is only a preferred embodiment of the present application, and is not intended to limit the protection scope of the present application. Any equivalent structural changes made according to the content of the specification and drawings of the present application shall be included in the patent protection scope of the present application.

Claims

1. A recognition method for an unsupervised cross-domain person re-identification system based on positional attention and refined clustering, characterized by: It includes backbone network module, position attention module, clustering algorithm module and memory module; among them: Backbone network module: Use ResNet50 to extract the initial features of the image; Position attention module: extracts more fine-grained features. Its position attention is added before layer 1 and after layer 4 of ResNet50. Clustering algorithm module: Clustering generates pseudo labels and uses refined clustering methods to overcome the impact of noise on the model; Memory module: The class centers obtained from each training cycle are stored in the memory module to facilitate subsequent training; The recognition method of the unsupervised cross-domain pedestrian re-identification system based on position attention and refined clustering includes the following steps: (1) Based on the ResNet50 network, the position attention module is added before layer 1 and after layer 4. Before the model depth convolution begins, the position attention block is used to focus on the position information of the sample and perform preliminary feature extraction. After the model extracts the features, the position attention block is used to perform feature mining to extract more fine-grained features. (2) Using the refined clustering algorithm module, we can eliminate noise interference samples and realize the clustering process from coarse to fine; (3) Introducing a memory module and storing the class centers obtained from each cycle of training into the memory module; In step (2), the refined clustering algorithm specifically includes the following steps: First, we use the feature Y(i,j) proposed by position attention to calculate Jaccard for preliminary clustering and generate pseudo labels. During the clustering process, we set different cluster radii to generate three clusters with the same cluster center but different radii. By cross-merging the three different clusters, we obtain their independence scores. The obtained scores are used to measure the reliability and determine the reliability of the clustering instance. The method to measure cluster independence is defined by the intersection-over-union (IoU) score: in Indicates the same The samples within the cluster, When the clustering conditions become loose, The cluster set of When the clustering condition becomes compact, it contains The cluster set of The larger the The more independent, The value of is between 0 and 1; Clustering is performed before each cycle, and clusters are screened based on the cluster reliability measure. Clusters with independence scores between 0 and 1 are considered reliable clusters, and the rest are considered unreliable clusters. Only reliable clusters are retained, and the remaining samples are considered cluster outliers. Outliers continue to be clustered in the subsequent process until reliable clusters are generated. As the number of iterations increases, unreliable clusters are gradually screened based on the reliability judgment criteria to achieve a clustering process from coarse to fine. The center points of reliable clusters are stored in the memory module for subsequent training.

2. The recognition method of the unsupervised cross-domain person re-identification system based on position attention and refined clustering according to claim 1 is characterized by: In step (1), the position attention module is added to the backbone network to extract more fine-grained features, including two parts: coordinate information embedding and coordinate information generation. Specifically, the following steps are included: Given an input X, we first encode each channel along the horizontal and vertical coordinates using a pooling kernel of size (H, 1) or (1, W), respectively; the output of the c-th channel with height h is represented as: in Represents the output of the cth channel with a height of h; similarly, the output of the cth channel with a width of w and a height of h is expressed as: in Represents the output of the cth channel with width w; Then the feature maps along the horizontal and vertical directions are concatenated together and then fed into a 1×1 convolution to generate an intermediate feature map m, which is represented as: m=δ(F([Z h ,Z w ])) (3) Where [,] represents a concatenation operation, δ is a nonlinear activation function, and F is a 1*1 convolution transformation function. The spatial dimension is then divided into two independent tensors, and two more 1×1 convolutions are used to transform the two independent tensors to obtain two tensors with the same number of channels as the input X. g h =sigmoid(F h (m h )) (4) g w =sigmoid(F w (m w )) (5) Where sigmoid() represents the sigmoid activation function; Finally, the output of the attention block Y is expressed as: Y(i,j)=x(i,j)×g h ×g w (6)。 3. The recognition method of the unsupervised cross-domain person re-identification system based on position attention and refined clustering according to claim 1 is characterized by: In step (3), introducing the memory module specifically includes the following steps: The first step is to initialize the memory module. Each cluster is regarded as a class, and the average feature of each cluster is used to initialize the class-level memory K[i]. Specifically, it is expressed as: Among them I k Indicates that it contains feature instance x i Reliable clustering, |.| represents the number of features of the cluster, v i Represents the target domain feature instance; Then, the module is updated. Based on all instances in the k-th cluster, it is continuously updated in each cycle to construct a memory module to store the center points of the clusters. After each clustering is completed, it is updated and the center points of the reliable clusters are fed into the module. K[i]←βK[i]+(1-β)v i (9) Where β∈[0,1] is the update rate; The memory module performs feature storage, allowing the target domain samples to be continuously updated even when the clustered and non-clustered outliers are constantly changing.

Images