An unsupervised domain adaptation pedestrian re-identification method and system

Abstract

The application provides a kind of unsupervised domain adaptation pedestrian re-identification method and system, related to the application field of domain adaptation.The unsupervised domain adaptation pedestrian re-identification method comprises: the sample features extracted are divided into global features and local features by feature mapping, and a base model of the global features and the local features of the sample is constructed by cross-entropy loss and triplet loss;The global features of the sample are subjected to clustering operation, pseudo-labels are generated by clustering operation, and the global and local features are trained by the generated pseudo-labels and the network is optimized in a supervised manner;The similarity of global features and local features is judged by cross-consistency score;The cross-consistency score and the local features of the sample are combined to purify the pseudo-labels, so as to complete pedestrian re-identification.The method improves the label noise problem in the existing unsupervised domain adaptation pedestrian re-identification method.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of application of domain adaptation, and in particular to an unsupervised domain adaptation pedestrian re-identification method and system. BACKGROUND

[0002] Unsupervised pedestrian re-identification method: According to whether additional source domain label information is used, existing unsupervised methods can be divided into two categories: one is an unsupervised domain adaptation method, and the other is an unsupervised learning method. Among them, the unsupervised domain adaptation method needs additional source domain label information, and the unsupervised learning method does not need additional source domain label information. Some unsupervised domain adaptation methods reduce the domain gap through feature distribution alignment and image style transfer. In recent years, unsupervised domain adaptation methods and unsupervised methods use clustering or near neighbor research to assign pseudo-labels for learning. Recently, some clustering-based methods apply clustering cluster proxy-based contrast learning strategies and show good advantages. However, the potential pseudo-label noise will affect the performance of the model, and recently some methods have begun to study this problem. However, these methods only consider the global context label information, unlike the above methods, the patent proposes to use fine-grained information from local features to purify the labels more effectively.

[0003] Label noise learning: Since it is difficult to obtain high-quality labels in many real-world scenarios, people have paid much attention to robust training with noisy labels. Loss adaptation methods use loss correction techniques through noise transformation matrices or sample reweighting schemes based on the reliability of given labels to reduce the influence of noisy labels. However, these methods need a certain number of clean labels to estimate the noise degree, and are not suitable for unsupervised person re-identification, in which pseudo-labels are very noisy at the beginning of training. Some people have also tried to design a robust loss function for label noise. However, these loss functions are designed for simple image classification tasks and are not suitable for open-set pedestrian re-identification tasks.

[0004] Label purification method: The generated pseudo-labels for unsupervised domain adaptation pedestrian re-identification methods themselves contain a part of label noise, which can significantly reduce the performance. To solve this problem, there are two strategies: one is to perform robust clustering on the dataset, and the other is a pseudo-label purification strategy. The modified clustering algorithm usually defines additional conditions to collect more reliable clusters, which are difficult to balance between the compactness and looseness of clustering. Although these methods have proven their effectiveness in some cases, they are always sensitive to parameters and highly dependent on the instance features themselves. Representative label purification algorithms include some nearest neighbor search methods and some additional information assisted pseudo-label purification methods. In contrast to the above methods, the patent proposes to use local features to reduce the label noise of global feature clustering by providing fine-grained information. SUMMARY

[0005] The technical problems solved by the present application

[0006] In view of the deficiencies of the prior art, the present application provides an unsupervised domain adaptation pedestrian re-identification method and system to improve the label noise problem in the prior art.

[0007] The technical solutions of the present application

[0008] To achieve the above object, the present application is realized by the following technical solutions:

[0009] In a first aspect, an unsupervised domain adaptation pedestrian re-identification method is provided, comprising:

[0010] The extracted sample features are divided into global features and local features through feature mapping, and a base model of the global features and the local features of the sample is constructed through cross-entropy loss and triplet loss;

[0011] The global features of the sample are clustered, pseudo-labels are generated through the clustering operation, and the global and local features are trained through the generated pseudo-labels and the network is optimized in a supervised manner;

[0012] The similarity of the global features and the local features is evaluated through cross-consistency score;

[0013] The cross-consistency score and the local features of the sample are combined to purify the pseudo-labels, so as to complete pedestrian re-identification.

[0014] Preferably, the extracted sample features are divided into global features and local features through feature mapping, specifically comprising:

[0015] Given the source domain and the target domain wherein represents the i-th sample of the source domain, y i represents the label of the i-th sample of the source domain, N S represents the number of samples in the source domain, represents the i-th sample of the target domain, N T represents the number of samples in the target domain;

[0016] Extract the shared feature representation F(x i )∈R C×H×W , wherein C, H, and W represent the channel number, height, and width of the feature mapping, respectively; global features f i g The feature mapping is input into a Bottleneck to obtain a new feature mapping, and the new feature mapping is evenly divided into N regions R And the local features are obtained by using a global average pooling operation on each region

[0017] Preferably, the base model for constructing the global feature and the local feature of the sample through the cross-entropy loss and the triplet loss specifically comprises:

[0018] The cross-entropy loss function of the global feature is:

[0019]

[0020]

[0021] Wherein, q i g is the predicted vector of the global feature, is a global feature classifier;

[0022] The cross-entropy loss function of the local feature is:

[0023]

[0024]

[0025] Wherein, represents the predicted vector of the nth local feature space p n , is the classifier of the local feature space p n ;

[0026] The triplet loss function is defined as:

[0027]

[0028] Wherein, ||·|| refers to the L2 norm, f i , f i,p and f i,n respectively represent the anchor sample feature, the hardest positive sample feature and the hardest negative sample feature obtained by using the hard batch triplet selection strategy in a mini-batch.

[0029] Preferably, the global feature of the sample is clustered, and the pseudo label is generated through the clustering operation, specifically comprising:

[0030] The DBSCAN clustering algorithm is used to cluster the global feature of the trained sample. The pseudo label generated by clustering adopts one-hot encoding form; according to the clustering result, a cluster-based storage library is initialized by clustering cluster centers, wherein m i is defined as:

[0031]

[0032] where f i denotes the feature of the i-th sample in the cluster C, |C| denotes the size of a cluster.

[0033] Preferably, the global and local features are trained by the generated pseudo labels and the network is optimized in a supervised manner, specifically including:

[0034] A contrastive learning loss is introduced:

[0035]

[0036] where m + denotes the feature f g the cluster center to which the feature belongs, m j denotes the j-th cluster center in the memory bank, denotes the cosine similarity between the vectors u and v, and τ is a temperature parameter;

[0037] The memory bank is updated in a momentum manner:

[0038] m i = μ·m i +(1-μ)·f g

[0039] where μ denotes an update factor for momentum update, f g denotes the feature of the sample belonging to the i-th cluster in the current batch;

[0040] The overall training objective is:

[0041] L = L gce + L pce + L tri + λL cl

[0042] where λ is a weighted hyperparameter that controls the importance of the contrastive learning loss.

[0043] Preferably, the similarity between the global feature and the local feature is evaluated by the cross-consistency score, specifically including:

[0044] The cross-consistency score is defined as the Jaccard similarity between the k-nearest neighbors of the global feature and the local feature, and the global feature space g of the picture and the n-th local feature space p n The calculation formula of the cross-consistency score is:

[0045]

[0046] wherein R i (g, k) and R i (p n , k) are k mutual nearest neighbors, i.e. R(f, k) = {g i |(g i ∈ N(f, k)) ∧ (f ∈ N(g i , k))}, N(f, k) is the k nearest neighbors of f, and |·| represents the number of candidates in the k mutual nearest neighbor set.

[0047] Preferably, the combination of the cross-consistency score and the local feature of the sample realizes the purification of the pseudo label, and specifically includes:

[0048] Constructing a label smoothing loss based on the cross-consistency score:

[0049] Given a sample , the pseudo label , the local feature , the label smoothing formula is:

[0050]

[0051] wherein u is a consistency vector, is a weight that determines the degree of label smoothing, and is dynamically updated by the cross-consistency score C i (g, p n ); i.e.

[0052] The purified local feature pseudo label is brought into , and the cross-entropy loss is rewritten using the KL divergence to obtain a label smoothing loss based on the cross-consistency score, and the formula is as follows:

[0053]

[0054] wherein H(·,·) and D DL (·||·) represent the cross-entropy loss and the KL divergence, respectively, and are weighted and balanced by the value of ;

[0055] Constructing a pseudo label purification loss based on the local feature:

[0056] Generating a pseudo label purified based on the local feature as the pseudo label of the global feature, and the formula is as follows:

[0057]

[0058]

[0059] wherein denotes the set weight, is the prediction vector of the local feature , β∈[0,1] is the weighting parameter controlling the one-hot encoding and the integrated prediction ratio; the purified global feature pseudo label is brought into the formula to train the global feature:

[0060]

[0061] The final training optimization target is obtained, and the formula is as follows:

[0062] L=L cals +L pglr +L tri +λL cl .

[0063] First, the extracted sample features are mapped to form two branches: one is the global feature, and the other is the local feature. The method proposed in this patent reduces label noise by utilizing the complementary information of global features and local features. Specifically, a cross-consistency score is designed as the similarity of k-nearest neighbors between feature spaces to utilize reliable complementary relationships. Based on the cross-consistency score, the method purifies the pseudo label of the global feature by combining the prediction of the local feature, and reduces the noise in the global feature clustering together. By applying label smoothing according to the applicability of each local label given, the pseudo label of the local feature is further purified. Thanks to the reliable complementary information provided by the cross-consistency score, the invention effectively reduces the influence of label noise.

[0064] In the second aspect, an unsupervised domain adaptation pedestrian re-identification system is provided, which comprises:

[0065] A basic model construction module is used to divide the extracted sample features into global features and local features through feature mapping, and a basic model of the global features and the local features of the sample is constructed through cross-entropy loss and triplet loss;

[0066] A pseudo label generation module is used to perform clustering operation on the global features of the sample, generate pseudo labels through clustering operation, train the global and local features through the generated pseudo labels, and optimize the network in a supervised manner;

[0067] A similarity evaluation module is used to evaluate the similarity of the global features and the local features through the cross-consistency score;

[0068] A purification module is used to purify the pseudo label by combining the cross-consistency score and the local features of the sample, so as to complete the pedestrian re-identification.

[0069] In a third aspect, a computer-readable storage medium storing one or more programs is provided, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods described.

[0070] In a fourth aspect, a computing device is provided, comprising:

[0071] one or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described.

[0072] (III) Advantages

[0073] The unsupervised domain adaptation pedestrian re-identification method and system of the present application first divides the extracted sample features into global features and local features through feature mapping, and then constructs the most basic model through cross-entropy loss and triplet loss. Secondly, a storage repository for storing clustering centers is constructed, and contrastive learning is introduced to make the features have higher similarity with their clustering centers. Then, cross-consistency loss is designed and calculated to determine the similarity between samples. The higher the cross-consistency score is, the higher the similarity between samples is, and therefore more complementary information can be provided. Finally, pseudo-label purification is performed. By constructing label smoothing loss based on cross-consistency score and label purification loss based on local label, the purification of pseudo-label is realized, and the label noise is reduced. The less the label noise is, the stronger the performance of the model is, and therefore better classification and retrieval effect can be achieved. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 The figure is a flowchart of the unsupervised domain adaptation pedestrian re-identification method of the present application;

[0075] Figure 2 The figure is a model diagram of the unsupervised domain adaptation pedestrian re-identification method of the present application;

[0076] Figure 3 The figure is a common multi-branch convolutional neural network model diagram in the embodiment of the present application;

[0077] Figure 4 The figure is a network architecture diagram of Bottleneck in the unsupervised domain adaptation pedestrian re-identification method of the present application;

[0078] Figure 5 The figure is a calculation principle diagram of cross-consistency score in the unsupervised domain adaptation pedestrian re-identification method of the present application. DETAILED DESCRIPTION

[0079] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0080] Embodiments

[0081] As shown in Figures 1-2 , the embodiment of the present application provides an unsupervised domain adaptation pedestrian re-identification method, comprising:

[0082] The extracted sample features are divided into global features and local features through feature mapping, and a basic model of the global features and the local features of the sample is constructed through cross-entropy loss and triplet loss;

[0083] The global features of the sample are subjected to clustering operation, pseudo labels are generated through the clustering operation, and the global and local features are trained through the generated pseudo labels and the network is optimized in a supervised manner;

[0084] The similarity of the global features and the local features is judged through cross-consistency score;

[0085] The cross-consistency score and the local features of the sample are combined to purify the pseudo labels, so as to complete pedestrian re-identification.

[0086] The specific steps are as follows:

[0087] Step 1: constructing an initial model framework through cross-entropy loss and triplet loss

[0088] The present application first constructs an initial local unsupervised domain adaptation pedestrian re-identification framework through cross-entropy loss and triplet loss, which utilizes the fine-grained information of local features. Existing unsupervised domain adaptation pedestrian re-identification methods only utilize global features and ignore the fine-grained information of local features. Unlike existing methods, the present application adopts a multi-branch manner to purify sample pseudo labels using both global features and local features. Common multi-branch convolutional neural network models are as shown in Figure 3 .

[0089] For convenience of modeling, in the unsupervised domain adaptation scenario, given a source domain and a target domain wherein represents the i-th sample of the source domain, y i represents the label of the i-th sample of the source domain, N S represents the number of samples of the source domain, represents the i-th sample of the target domain, N TThis represents the number of samples in the target domain. The model first extracts the shared feature representation F(x). i )∈R C×H×W Where C, H, and W represent the number of channels, height, and width of the feature map, respectively. Then, a global average pooling operation is performed on the feature map to obtain the global feature f. i g The feature map is input into Bottleneck to obtain a new feature map, which is then divided into N regions on average. Then, a global average pooling operation is performed on each region to obtain local features. Bottleneck's network architecture is as follows: Figure 4 As shown. Since the target domain has no labels, the DBSCAN clustering algorithm is used based on global features f. i g Perform clustering and assign pseudo-labels using the clustering results. Then, assign target domain sample images... The pseudo-tag is marked as The generated pseudo-labels are used to train global and local features for pedestrian re-identification.

[0090] For global features, the form of the cross-entropy loss function is shown in Equation (1):

[0091]

[0092]

[0093] in, It is the prediction vector of global features. It is a global feature classifier, which consists of a fully connected layer and a softmax function. Similarly, the cross-entropy loss function for local features takes the form of Equation (3):

[0094]

[0095]

[0096] in, p represents the nth local feature space n The predicted vector, It is a local feature space p n The classifier. Next, we define a softmax-triplelet loss function, as shown in Equation (5):

[0097]

[0098] Where ||·|| refers to the L2 norm, f i f i,p and fi,n respectively represent the anchor sample feature, the hardest positive sample feature and the hardest negative sample feature obtained by using the hard batch triplet selection strategy in one mini-batch. The hardest positive sample refers to the positive sample farthest from the anchor sample, and the hardest negative sample refers to the negative sample closest to the anchor sample.

[0099] Step 2: Construct a repository storing clustering centers

[0100] This step is divided into two stages: a clustering stage and a preliminary network model training stage. In the clustering stage, global and local features are extracted, and pseudo labels are assigned according to the global feature clustering.

[0101] Stage One: Clustering

[0102] At the beginning of each training round, the DBSCAN clustering algorithm is used to cluster the global features of the training samples. The pseudo labels generated by clustering are in one-hot encoding form. At the same time, according to the clustering results, a clustering-based repository is initialized by clustering cluster centers, where m i is defined as formula (6):

[0103]

[0104] where f i represents the feature of the i-th sample in the clustering cluster C, and |C| represents the size of a clustering cluster.

[0105] Stage Two: Preliminary Network Model Training

[0106] After obtaining the pseudo labels, the network is optimized in a supervised manner to achieve the training goal. Here, we introduce a contrastive learning loss, as shown in formula (7):

[0107]

[0108] where m + represents the clustering cluster center to which the feature f g belongs, and m j represents the j-th clustering cluster center in the repository, is the cosine similarity between vectors u and v, and τ is the temperature parameter. The purpose of formula (7) is to encourage the feature f g and its related positive clustering cluster center m + to have a higher similarity. The repository is updated in a momentum manner, as shown in formula (8):

[0109] m i = μ·m i +(1-μ)·fg (8)

[0110] where μ denotes the update factor of momentum update, f g denotes the feature of the sample belonging to the i-th cluster in the current batch. Therefore, the overall preliminary training target of this stage is shown in formula (9):

[0111] L = L gce + L pce + L tri + λL cl (9)

[0112] where λ is a weighted hyperparameter, which controls the importance of the contrastive learning loss.

[0113] Step 3: Determine the similarity between sample features by calculating cross-consistency score

[0114] Unlike the overall preliminary model framework in formula (9), the present application finally uses the purified pseudo-label to train the model, which considers the complementary relationship between global features and local features. However, due to the differences in feature similarity structure between global features and local features, there is unreliable complementary information. Specifically, local features contain information unrelated to people, which is not suitable for purifying the pseudo-label of global features. In addition, global features only consider global context, sometimes ignoring information related to local features. Therefore, determining whether the given complementary information is reliable is a basic task of the present application.

[0115] To solve this problem, the present application designs a cross-consistency score to capture the mutual similarity of the k-nearest neighbors of global features and local features, as shown in formula (10). Figure 5 The global feature space g of the picture and the cross-consistency score of the n-th local feature space p n are shown in formula (10):

[0116]

[0117] where R i (g, k) and R i (p n , k) are the k mutual nearest neighbors, i.e. R(f, k) = {g i | (g i ∈ N(f, k)) ∧ (f ∈ N(g i , k))}, N(f, k) is the k nearest neighbor of f, and |·| represents the number of candidates in the k mutual nearest neighbor set. We believe that the more repeated samples between the k mutual nearest neighbor sets of g and p n , i.e. C i (g, p nThe higher the value of the cross-consistency score C (y

[0118] Step 4: Pseudo-label purification is performed by cross-consistency score based label smoothing loss and local feature based label purification method, the specific steps are as follows:

[0119] Step 4-1) Construct cross-consistency score based label smoothing loss:

[0120] Learning all local features using the same global pseudo-label may be detrimental to model training, ignoring the local context of local features. For example, some parts contain information unrelated to people (e.g., occlusion), and it is best to exclude these unrelated information from training. To solve this problem, we use label smoothing to purify the pseudo-label of each local feature according to the corresponding cross-consistency score.

[0121] Given a sample Pseudo-label Local feature The label smoothing formula of the local feature

[0122]

[0123] Where u is a consistency vector, is a weight that determines the degree of label smoothing, which is dynamically updated by the cross-consistency score C i (g, p n ) That is Then the purified local feature pseudo-label is brought into formula (3) and the cross-entropy loss is rewritten with KL divergence to get the cross-consistency score based label smoothing loss, as shown in formula (12):

[0124]

[0125] Where H(·,·) and D DL (·||·) represent cross-entropy loss and KL divergence, respectively, which are weighted balanced by the value of .

[0126] In formula (12), the first term encourages the prediction to approach high confidence of y i , and the second term encourages the prediction to collapse into a consistency vector. By scaling the two opposite terms using the cross-consistency score, we calibrate the prediction of the local feature according to the reliability of the pseudo-label of each local feature.

[0127] Step 4-2) Construct local feature based pseudo-label purification loss:

[0128] We propose a label refinement method for local features, which uses the prediction results of local features to generate refined labels for global features. Local features with rich local context information can be used to handle label noise in global feature clustering, because global feature clustering often ignores fine-grained information. However, due to the poor discriminability of local features, misleading information will be provided. Therefore, according to each cross-consistency score, we use different weights to aggregate the predictions of partial features, so as to use more reliable information to improve the label.

[0129] Generating pseudo labels based on local feature refinement As the pseudo label of global feature, as shown in equation (13):

[0130]

[0131]

[0132] wherein, represents the set weight, which can be calculated by equation (14), is the prediction vector of local feature , and β ∈ [0, 1] is a weighting parameter that controls the one-hot encoding and the integrated prediction ratio. Then, the refined global feature pseudo label is brought into equation (1) to obtain equation (15) to train the global feature:

[0133]

[0134] With the refined label guided by local features, the global feature can learn from the local features with rich fine-grained information.

[0135] Finally, the final training optimization target is obtained, as shown in equation (16):

[0136] L = L cals + L pglr + L tri + λL cl (16)

[0137] The present application effectively reduces the influence of noise labels in two ways. By formula (12), the local features with low cross-consistency scores are trained with pseudo labels close to uniform distribution, while the global features trained by the purified labels guided by local features obtain reliable fine-grained information from local features by formula (13). In addition, when the predictions of all local features have lower cross-consistency scores, the union prediction of the purified labels guided by local features will eventually collapse into a uniform vector due to the strong label smoothing effect of all local features, thus providing meaningless training information. This enables the present method to weaken the influence of noise pseudo labels and obtain better representation learning.

[0138] The present application provides an unsupervised domain adaptation pedestrian re-identification method based on multi-branch sample label purification. There are many specific implementation methods and approaches for this technical solution, and the above description is only the optimal implementation method of the present application. It should be noted that, without departing from the principles of the present application, the present application can be improved and enhanced for those skilled in the art, and these improvements and enhancements should be considered within the scope of the present application. The components not explicitly described in the present application can be realized by existing technology.

[0139] Another embodiment of the present application provides an unsupervised domain adaptation pedestrian re-identification system, which comprises:

[0140] A basic model construction module is configured to divide the extracted sample features into global features and local features through feature mapping, and construct a basic model of the global features and the local features of the sample through cross-entropy loss and triplet loss;

[0141] A pseudo label generation module is configured to perform clustering operation on the global features of the sample, generate pseudo labels through the clustering operation, train the global and local features through the generated pseudo labels, and optimize the network in a supervised manner;

[0142] A similarity evaluation module is configured to evaluate the similarity of the global features and the local features through cross-consistency scores;

[0143] A purification module is configured to purify the pseudo labels in combination with the cross-consistency scores and the local features of the sample, so as to complete pedestrian re-identification.

[0144] Embodiments of the present application can be provided as a method or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer-readable program code. The program code may

[0145] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in conjunction with the flowcharts described above. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks or in conjunction with the flowcharts described above.

[0146] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in conjunction with the flowcharts described above. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks or in conjunction with the flowcharts described above.

[0147] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks or in conjunction with the flowcharts described above. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks or in conjunction with the flowcharts described above.

[0148] It is to be noted that, as used in this specification and the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to "a component" can include a combination of two or more components. Additionally, the terms "comprise," "comprises," and "comprising," or any variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Furthermore, unless otherwise indicated herein, the terms "first," "second," "third," etc., are used herein merely as labels, and are not intended to impose ordinal import.

Claims

1. An unsupervised domain adaptation pedestrian re-identification method, characterized in that, The extracted sample features are divided into global features and local features through feature mapping, and a base model of the global features and the local features of the sample is constructed through cross-entropy loss and triplet loss; The global features of the sample are clustered, pseudo labels are generated through the clustering operation, and the global and local features are trained through the generated pseudo labels and the network is optimized in a supervised manner; The similarity of the global features and the local features is evaluated through cross-consistency score; The pseudo labels are purified by combining the cross-consistency score and the local features of the sample, so as to complete pedestrian re-identification; The pseudo labels are purified by combining the cross-consistency score and the local features of the sample, and the pseudo label purification loss based on the local features is constructed. The label smoothing loss based on the cross-consistency score is constructed. The pseudo label purification loss based on the local features is constructed. Given sample Pseudo labels Local features The label smoothing formula is: where u is a uniformity vector, is a weight that determines the degree of label smoothing, and is updated by the cross-consistency score C i (g, p n ) is dynamically updated i.e. The purified local feature pseudo label is brought into And the cross-entropy loss is rewritten with KL divergence to get the label smoothing loss based on cross-consistency score, as follows: where H(·, ·) and D DL (·||·) denote cross-entropy loss and KL divergence, respectively, and are weighted by the values of ; The final training optimization target is obtained, and the formula is as follows: Generating pseudo labels based on local feature purification As a global feature, the pseudo label is as follows: wherein, denotes the collective weight, is the predicted vector of local features , β ∈ [0, 1] is a weighting parameter controlling the one-hot encoding and the integrated prediction ratio; the purified global feature pseudo-label is brought into the formula to train the global feature: The extracted sample features are divided into global features and local features through feature mapping, and a base model of the global features and the local features of the sample is constructed through cross-entropy loss and triplet loss. where L trl represents the triplet loss function, L cl represents the contrastive learning loss function.

2. The unsupervised domain adaptation pedestrian re-identification method according to claim 1, characterized in that: The base model of the global features and the local features of the sample is constructed through cross-entropy loss and triplet loss, and the cross-entropy loss function of the global features is as follows: Given source domain and target domain wherein denotes the i-th sample of the source domain, y i denotes the label of the i-th sample of the source domain, N S denotes the number of source domain samples, denotes the i-th sample of the target domain, N T denotes the number of target domain samples; Extracting shared feature representation where C, H, W represent the number of channels, height and width of the feature map respectively; global average pooling operation is performed on the feature map to obtain global feature f i g The feature map is input into the bottleneck to obtain a new feature map, and the new feature map is evenly divided into N regions and a global average pooling operation is performed on each region to obtain local features 3. The unsupervised domain adaptation pedestrian re-identification method according to claim 2, wherein: The cross-entropy loss function of the local features is as follows: The triplet loss function is defined as follows: wherein, is a predicted vector of global features, is a global feature classifier; The global features of the sample are clustered, and pseudo labels are generated through the clustering operation. wherein, represents the prediction vector of the n-th local feature space p n , is the classifier of the local feature space p n . The global and local features are trained through the generated pseudo labels, and the network is optimized in a supervised manner. where ||·|| denotes the L2 norm, f i , f i,p and f i,n represent the anchor sample feature, the hardest positive sample feature and the hardest negative sample feature obtained by using the hard batch triplet selection strategy in one mini-batch, respectively.

4. The unsupervised domain adaptation pedestrian re-identification method according to claim 3, wherein: The contrast learning loss is introduced. The DBSCAN clustering algorithm is used for clustering operation on the global features of the training samples, and a pseudo label is generated by clustering Adopt one-hot encoding form; according to the clustering result, a clustering-based storage library Through cluster center initialization, wherein m i Defined as: where f i represents the features of the i-th sample in the cluster C, and |C| represents the size of a cluster.

5. The unsupervised domain adaptation pedestrian re-identification method according to claim 4, characterized in that: The repository is updated in a momentum manner. The overall training target is as follows: where m + denotes the feature f g the cluster center m j denotes the jth cluster center in the repository, denotes the cosine similarity between vectors u and v, and τ is a temperature parameter. Wherein, λ is a weighted hyperparameter, which controls the importance of the contrast learning loss. m i = μ · m i + (1 - μ) · f g wherein μ denotes an update factor for the momentum update, f g denotes the feature of the sample belonging to the i-th cluster in the current batch; The similarity of the global features and the local features is evaluated through cross-consistency score. The system is used to implement the unsupervised domain adaptation pedestrian re-identification method of claim 1, and the system comprises:

6. The unsupervised domain adaptation pedestrian re-identification method according to claim 5, wherein: A base model construction module is configured to divide the extracted sample features into global features and local features through feature mapping, and construct a base model of the global features and the local features of the sample through cross-entropy loss and triplet loss. The cross-consistency score is defined as the Jaccard similarity between the k-nearest neighbors of the global and local features of the picture The global feature space g and the n-th local feature space p n The cross-consistency score of the picture is calculated as follows: wherein and are k nearest neighbors of each other, i.e. are k nearest neighbors of f, |·| denotes the number of candidates in the set of k mutual nearest neighbors.

7. An unsupervised domain adaptation pedestrian re-identification system, characterized in that, A pseudo label generation module is configured to cluster the global features of the sample, generate pseudo labels through the clustering operation, train the global and local features through the generated pseudo labels, and optimize the network in a supervised manner. A similarity evaluation module is configured to evaluate the similarity of the global features and the local features through cross-consistency score. A purification module is configured to purify the pseudo labels by combining the cross-consistency score and the local features of the sample, so as to complete pedestrian re-identification. The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods of claims 1-6. One or more processors, memories, and one or more programs, wherein the one or more programs are stored in the memories and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods of claims 1-6.

8. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions that when executed by a computer cause the computer to perform a method of any of claims 1-7. ​ 9. A computing device, comprising: ​ ​

Images