A dynamic confidence-based cross-domain person re-identification method and system

Abstract

The present invention discloses a cross-domain person re-identification method and system with dynamic confidence, which relates to the field of person re-identification. The method comprises pre-training a ResNet-50 network using a source domain training set; determining the mean average performance (mAP) of the current model on a source domain test set; and estimating the confidence of the current model on a target domain unlabeled dataset using mAP; extracting features of the target domain unlabeled dataset using the current model; clustering the extracted features using the DBSCAN algorithm, and using the clustered cluster IDs as pseudo-labels for corresponding pedestrian images in the target domain unlabeled dataset; initializing a memory bank using outlier features generated by clustering; training the current model using the target domain pseudo-labeled dataset, confidence, and memory bank, and then reconstructing the target domain pseudo-labeled dataset, confidence, and memory bank until the current model reaches an optimal state; and identifying the pedestrian images to be identified using the cross-domain person re-identification model. The present invention can improve the performance of person re-identification.

Description

Technical Field

[0001] The present invention relates to the field of pedestrian re-identification, and in particular to a cross-domain pedestrian re-identification method and system with dynamic confidence. Background Art

[0002] Person Re-identification (also known as pedestrian re-identification, abbreviated as ReID) is a technology that uses computer vision technology to determine whether a specific pedestrian exists in an image or video sequence; in other words, pedestrian re-identification refers to identifying the target pedestrian in a video sequence with existing possible sources and non-overlapping camera fields of view. It is widely considered to be a sub-problem of image retrieval. Given a surveillance pedestrian image, retrieve the pedestrian image across devices. In surveillance videos, due to the camera resolution and shooting angle, it is usually impossible to obtain very high-quality facial images. When face recognition fails, ReID becomes a very important alternative technology. A very important feature of ReID is that it is cross-camera, so when evaluating performance in academic papers, the goal is to retrieve the same pedestrian images under different cameras. ReID has been studied in academia for many years, but it was not until recent years with the development of deep learning that a huge breakthrough was achieved.

[0003] With the rapid development of deep learning, deep learning, which combines the advantages of deep neural networks in semantic feature extraction and end-to-end training, has gradually come into people's attention. Compared with traditional methods, deep networks can extract more discriminative features, thereby better completing the task of pedestrian re-identification. Supervised pedestrian re-identification technology has also made great breakthroughs in recent years. However, due to the domain differences between different datasets, the model that performs well on one dataset cannot be applied to other datasets. This has greatly hindered the practical application of pedestrian re-identification. Therefore, unsupervised domain adaptive person re-identification (UDA-ReID) has gradually attracted people's attention.

[0004] Shortcoming 1 of the existing technology: When clustering-based UDA-ReID assigns pseudo-labels, since the model does not perform well in the target domain at the beginning, there will inevitably be noisy samples. Previous methods generally further refine the pseudo-labels through some label refinement methods, but cannot completely remove the noisy samples. As a result, the noisy samples will greatly damage the model training.

[0005] Disadvantage 2 of existing technology: K-means clustering does not perform well for clustering complex data because the number of clusters must be set in advance. The DBSCAN clustering algorithm does not have this disadvantage. However, DBSCAN also has outliers during clustering. These outliers are not in the cluster and are therefore ignored during training. This leads to incomplete training data and the model may fall into a local optimum.

[0006] From the above, we can see that the existing methods all have their own instability problems. Therefore, there is an urgent need for a new pedestrian re-identification method or system to improve the performance of UDA-ReID technology. Summary of the Invention

[0007] The purpose of the present invention is to provide a cross-domain pedestrian re-identification method and system with dynamic confidence, which can improve the performance of pedestrian re-identification.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A dynamic confidence cross-domain pedestrian re-identification method, including:

[0010] The current model is obtained by pre-training a ResNet-50 network using a source domain training set; the source domain is a labeled person re-identification dataset; the source domain training set is the data used for model training in the person re-identification dataset;

[0011] Determine the mean average precision (mAP) of the current model on a source domain test set; and use mAP to estimate the confidence of the current model on a target domain unlabeled dataset; the source domain test set is divided into query sets and gallery sets;

[0012] The current model is used to extract features of the target domain unlabeled dataset. The extracted features are clustered using the DBSCAN algorithm. The clustered cluster IDs are used as pseudo labels for the corresponding pedestrian images in the target domain unlabeled dataset to obtain the target domain pseudo-labeled dataset.

[0013] Initializing a memory bank using outlier features generated by clustering; the memory bank is used to store outlier features; and is used to calculate an outlier feature loss function;

[0014] The current model is trained using the target domain pseudo-label dataset, confidence, and memory library. In each subsequent round of training, the target domain pseudo-label dataset, confidence, and memory library are reconstructed until the current model reaches the optimal value, thus obtaining a cross-domain pedestrian re-identification model.

[0015] Use the cross-domain pedestrian re-identification model to identify the pedestrian images to be identified.

[0016] Optionally, using the source domain training set to pre-train a ResNet-50 network to obtain the current model specifically includes:

[0017] The ResNet-50 network is pre-trained using the cross entropy loss function and the triplet loss function.

[0018] Optionally, the joint pre-training process is:

[0019] The features after the pooling layer of the ResNet-50 network use the triplet loss function;

[0020] The cross entropy loss function is used after the fully connected layer of the ResNet-50 network.

[0021] Optionally, determining the mean average precision (mAP) of the current model on the source domain test set specifically includes:

[0022] Remove the last fully connected layer of the current model and retain the adaptive average pooling layer;

[0023] Use the processed current model to extract features of query sets and gallery sets respectively;

[0024] Calculate the similarity between each person image in the query set and each person image in the reference set gallery sets;

[0025] Calculate the mAP of the current model on the source domain test set based on the similarity.

[0026] Optionally, the method of training the current model using the target domain pseudo-label dataset, the current confidence level, and the memory library until the current model reaches an optimal state to obtain a cross-domain person re-identification model specifically includes:

[0027] Add a fully connected layer after the current model, and the output dimension is the number of clusters;

[0028] The features extracted from the target domain unlabeled dataset are used to calculate the triplet loss function; and the features extracted from the target domain unlabeled dataset are input into the fully connected layer to obtain the classification results and used to calculate the classification loss function;

[0029] Supervised learning is performed based on the target domain pseudo-label dataset and confidence level; an outlier feature loss function is determined based on the memory library, and the outlier feature is updated in real time using the outlier feature loss function.

[0030] Optionally, the outlier feature loss function is:

[0031]

[0032] Among them, L discret is the outlier feature loss function value, f i t is the target domain pedestrian image feature, f k The top k features in similarity ranking, f n are other features whose similarity ranking is not in the top k, m is a hyperparameter, N B This is the number of pedestrian images in the mini-batch at this time, and N2 is the number of outlier features.

[0033] Optionally, an outlier feature loss function is determined based on the memory bank, and the outlier feature is updated in real time using the outlier feature loss function, specifically including the following formula:

[0034]

[0035] in, is the outlier feature loss function relative to feature f j The partial derivative value is used to update the feature, K i are the top k features in similarity ranking.

[0036] A dynamic confidence cross-domain person re-identification system, comprising:

[0037] A pre-training module is used to pre-train the ResNet-50 network using a source domain training set to obtain the current model; the source domain is a labeled pedestrian re-identification dataset; the source domain training set is the data in the pedestrian re-identification dataset used for model training;

[0038] A confidence determination module is used to determine the mean average precision (mAP) of the current model on a source domain test set and use mAP to estimate the current confidence of the current model on a target domain unlabeled dataset. The source domain test set is divided into query sets and gallery sets.

[0039] The target domain pseudo-label dataset determination module is used to extract features of the target domain unlabeled dataset using the current model; the extracted features are clustered using the DBSCAN algorithm, and the cluster IDs after clustering are used as pseudo-labels for the corresponding pedestrian images in the target domain unlabeled dataset to obtain the target domain pseudo-label dataset;

[0040] A memory bank initialization module is used to initialize a memory bank using outlier features generated by clustering; the memory bank is used to store outlier features; and is used to calculate an outlier feature loss function;

[0041] The cross-domain person re-identification model determination module is used to train the current model using the target domain pseudo-label dataset, confidence level, and memory library. In each subsequent round of training, the target domain pseudo-label dataset, confidence level, and memory library are rebuilt until the current model reaches the optimal value, thus obtaining the cross-domain person re-identification model.

[0042] The recognition module is used to identify the pedestrian images to be identified using the cross-domain pedestrian re-identification model.

[0043] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0044] The present invention provides a cross-domain person re-identification method and system with dynamic confidence. The method first pre-trains a model on a source domain. After pre-training, the model is adjusted using an unlabeled target domain to improve the model's performance on the target domain. Initially, the model's parameters are trained based on samples from the source domain. Therefore, at the beginning of the model adjustment phase, the model's performance on the source domain remains high. However, as the model is continuously adjusted on the target domain, its performance on the source domain gradually declines, while its performance on the target domain gradually improves. During the adjustment phase, the confidence of the model is expressed and calculated using mean average predictive value (MAP). This confidence is then combined with a loss function to calculate the loss. Since the confidence is relatively low at the beginning of model adjustment, it mitigates the damage caused by noisy samples to model updates and reduces their impact. As the model is updated and adjusted, the confidence increases, allowing for more efficient use of pseudo-labels. After the samples are assigned pseudo-labels, the outlier loss is combined to further encourage the model to learn discriminative features, enabling more efficient use of all data and improving model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A schematic flow chart of a cross-domain person re-identification method with dynamic confidence provided by the present invention;

[0047] Figure 2 Schematic diagram of the process determined for the cross-domain pedestrian re-identification model. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] The purpose of the present invention is to provide a cross-domain pedestrian re-identification method and system with dynamic confidence, which can improve the performance of pedestrian re-identification.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1 and Figure 2 As shown, the present invention provides a dynamic confidence cross-domain pedestrian re-identification method, including:

[0052] S101, using the source domain training set X S A ResNet-50 network is pre-trained to obtain the current model; the source domain is a labeled person re-identification dataset; and the source domain training set is data in the person re-identification dataset used for model training.

[0053] ResNet-50 is a residual learning network, including layers 1-4, pooling layers, and a fully connected layer with 1000-dimensional output. Only the network before layer 4 in the model is saved. This network is used to extract high-dimensional features of the image and add an adaptive average pooling layer to reduce the dimension of the high-dimensional features to f. i s , and used as the calculation of triplet loss later, a fully connected (FC) layer is added after the adaptive average pooling layer for classification, and the output dimension of the FC layer is S id Wei, S id is the number of IDs in the source domain training set.

[0054] The model is trained jointly using the cross entropy loss function and the triplet loss function. Specifically, the triplet loss function is used for the features after the pooling layer, which is regarded as a verification problem. The cross entropy loss is used after the FC layer, which is regarded as a classification problem. The two loss functions are used to train the model until the model converges to generate the initial Re-ID model. The loss calculation is shown as follows:

[0055]

[0056] Where, L cls is the classification loss value, L ce is the cross entropy loss value, NB L is the number of pedestrian images in the mini-batch at this time, ce is the calculated single-person cross entropy loss function value, and the average of all pedestrian cross entropy loss values is the classification loss value, f i s is the source domain pedestrian image feature, φ(f i s ) is the classification result after the fully connected layer, y i The true label of the current pedestrian image, L tri is the triplet loss value, Δ is a hyperparameter, which is a manually set constant, f i + is the most difficult positive sample of the current pedestrian in the mini-batch, that is, the positive sample with the farthest feature distance of the current pedestrian, f i - is the closest negative sample of the current pedestrian in the mini-batch, that is, the negative sample closest to the current pedestrian.

[0057] S102, determining the mean average precision (mAP) of the current model on a source domain test set; and using mAP to estimate the confidence of the current model on a target domain unlabeled dataset, wherein the source domain test set is divided into query sets and gallery sets.

[0058] S102 specifically includes:

[0059] Remove the last fully connected layer of the current model and retain the adaptive average pooling layer;

[0060] Use the processed current model to extract features of query sets and gallery sets respectively;

[0061] Calculate the similarity between each person image in the query set and each person image in the reference set gallery sets;

[0062] Calculate the mAP of the current model on the source domain test set based on the similarity. The specific calculation method is as follows:

[0063]

[0064] Where ω is the calculated confidence value, mAP is meanAveragePrecision, AP iis the average accuracy of the true positive samples of the current image x ranked in the i-th position in the gallery sets, where i is {1, 2, 3, ..., k}, k is the number of pedestrians with the same label as the current image x in the gallery sets, the true positive samples are the pedestrian pictures with the same label as the current image x in the gallery sets, the number of true positive samples is the number of all true positive samples before the true positive samples ranked in the i-th position (the same pedestrian picture), the predicted positive samples are all pedestrian pictures before the true positive samples ranked in the i-th position (the same pedestrian picture), and the number of predicted positive samples is the number of all pedestrian pictures before the true positive samples ranked in the i-th position (the same pedestrian picture).

[0065] S103, use the current model to extract the target domain unlabeled dataset features; and use the DBSCAN algorithm to cluster the extracted features, and use the cluster id after clustering as the pseudo label of the corresponding pedestrian image in the target domain unlabeled dataset to obtain the target domain pseudo label dataset.

[0066] Each data in the target domain Input into the network to extract features Cluster the features of all images using the DBSCAN algorithm and use the cluster ID as a pseudo-label. The dataset then becomes Because DBSCAN clustering will produce outliers, that is, data that is not in the cluster, collecting all the data features that are not in the cluster can obtain Where N1+N2=N T , where N1 is the number of pedestrian images assigned with pseudo labels, N2 is the number of outlier features, that is, the number of pedestrians without pseudo labels, and N1 is the number of all pedestrian images in the target domain.

[0067] S104, initializing a memory bank using the outlier features generated by clustering; the memory bank is used to store the outlier features; and is used to calculate the outlier feature loss function.

[0068] S105, use the target domain pseudo-label dataset, current confidence and memory library to train the current model, repeat S102-S105 until the current model reaches the optimal level, and obtain a cross-domain pedestrian re-identification model.

[0069] S105 specifically includes:

[0070] Add a fully connected layer after the current model, and the output dimension is the number of clusters;

[0071] Features f extracted from the target domain unlabeled dataset i tIt is used to calculate the triplet loss function; and the features extracted from the target domain unlabeled dataset are input into the fully connected layer to obtain the classification results and used to calculate the classification loss function;

[0072] The loss function is:

[0073]

[0074] Where, L cls is the classification loss value, L ce is the cross entropy loss value, N B L is the number of pedestrian images in the mini-batch at this time, ce is the calculated single-person cross entropy loss function value, and the average of all pedestrian cross entropy loss values is the classification loss value, f i s is the source domain pedestrian image feature, φ(f i s ) is the classification result after the fully connected layer, y i The true label of the current pedestrian image, L tri is the triplet loss value, Δ is a hyperparameter used to balance the distance between the current feature and the positive and negative sample features, f i + is the most difficult positive sample of the current pedestrian in the mini-batch, that is, the positive sample with the farthest feature distance of the current pedestrian, f i - is the nearest negative sample of the current pedestrian in the mini-batch, that is, the negative sample closest to the current pedestrian, and ω is the calculated confidence value.

[0075] Supervised learning is performed based on the target domain pseudo-label dataset and confidence level. The outlier loss function is determined based on the memory bank, and the outlier feature loss function is used to update the outlier features in real time.

[0076] Based on the memory bank initialized by outliers, an outlier feature loss function is proposed, which is combined with other loss functions to optimize the model. When a data is input into the model to obtain the feature f i t Afterwards, calculate this feature and F discret The similarity of each feature in the , assuming that the top k similarity rankings are positive samples and the rest are negative samples. The loss function is mainly used to increase the similarity between the positive samples and reduce the similarity between the negative samples. The top k features are recorded as K i , the outlier feature loss function is:

[0077]

[0078] Among them, Ldiscret is the outlier feature loss function value, f i t is the target domain pedestrian image feature, f k The top k features in similarity ranking, f n are other features whose similarity ranking is not in the top k, m is a hyperparameter, N B This is the number of pedestrian images in the mini-batch at this time, and N2 is the number of outlier features.

[0079] Calculate the loss function and use the gradient descent method to optimize the parameters of the pedestrian re-identification model with the goal of minimizing the loss function. The formula of the loss function is as follows:

[0080] L=L cls +L tri +L discret .

[0081] Where, L is the total loss value, L cls is the classification loss value, L ce is the cross entropy loss value, L discret is the outlier feature loss function value.

[0082] Since the model update method is batch processing, the outlier features may be inconsistent with the model update during the batch processing process. That is, after a batch update of the model, the outlier features are not updated accordingly. To address this situation, the outlier features are updated in real time in combination with the outlier feature loss as follows:

[0083] in, For the outlier feature loss function, consider the feature f j The partial derivative value is used to update the feature, K i are the top k features in similarity ranking.

[0084] S106: Using a cross-domain person re-identification model, the pedestrian images to be identified are identified. The pedestrian images to be identified are the pedestrian images to be searched and the pedestrian images to be identified.

[0085] According to the pedestrian features extracted by the feature extraction unit, the distance between the two pedestrian features is calculated. The distance calculation formula is as follows:

[0086]

[0087] Where, d(x i ,x j ) is pedestrian x i With pedestrians x j The characteristic distance value between i is the pedestrian image to be found, xj is the pedestrian picture to be identified, f(x i ) is the pedestrian image feature to be found, f(x j ) is the pedestrian feature to be identified.

[0088] According to the characteristic distance between the two pedestrians obtained by the calculation unit, it is determined whether it is less than a pre-set threshold. If it is less than, they are the same pedestrian, otherwise they are not.

[0089] The principle of implementing the above scheme of the present invention is as follows: in unsupervised domain adaptive pedestrian re-identification, due to the domain difference between the source domain and the target domain, the performance of the model pre-trained on the source domain on the target domain will be greatly reduced. However, due to the background of the sample, the pixels of the camera, etc. in the source domain and the target domain, the target domain features extracted by the clustering-based method do not have particularly good discrimination ability. Therefore, in the initial clustering process, there may be more than one person in the same cluster, that is, the pseudo-labels of different classes are consistent, resulting in a large amount of noise in the pseudo-labels.

[0090] Different clustering methods will also produce different effects. The DBSCAN clustering method used in this invention has the advantages of not needing to set the number of clusters in advance and the variable cluster shapes. However, after clustering is completed, there will be a large number of outliers, that is, there are many samples that are not divided into clusters. In previous methods, these samples may be directly ignored, but these samples also have certain value.

[0091] As another specific embodiment, the present invention provides a cross-domain person re-identification system with dynamic confidence, including:

[0092] A pre-training module is used to pre-train a ResNet-50 network using a source domain training set to obtain the current model; the source domain is a labeled pedestrian re-identification dataset; the source domain training set is the data in the pedestrian re-identification dataset used for model training.

[0093] The confidence determination module is used to determine the mean average precision (mAP) of the current model on the source domain test set; and use mAP to estimate the current confidence of the current model on the target domain unlabeled dataset. The source domain test set is divided into query sets and reference sets (gallery sets).

[0094] The target domain pseudo-label dataset determination module is used to use the current model to extract the features of the target domain unlabeled dataset; and cluster the extracted features using the DBSCAN algorithm, and use the clustered cluster ID as the pseudo label of the corresponding pedestrian image in the target domain unlabeled dataset to obtain the target domain pseudo-label dataset.

[0095] A memory bank initialization module is used to initialize a memory bank using outlier features generated by clustering; the memory bank is used to store outlier features; and is used to calculate an outlier feature loss function.

[0096] The cross-domain pedestrian re-identification model determination module is used to train the current model using the target domain pseudo-label dataset, the current confidence level, and the memory library until the current model reaches the optimal level, thereby obtaining a cross-domain pedestrian re-identification model.

[0097] The recognition module is used to identify the pedestrian images to be identified using the cross-domain pedestrian re-identification model.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0099] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A cross-domain person re-identification method with dynamic confidence, characterized by: include: Use the source domain training set to pre-train the ResNet-50 network to obtain the current model; The source domain is a labeled person re-identification dataset; The source domain training set is data in the person re-identification dataset used for model training; Determine the mean average precision (mAP) of the current model on the source domain test set; The mAP is used to estimate the confidence of the current model on the target domain unlabeled dataset. The source domain test set is the data in the pedestrian re-identification dataset used to test the model performance, which is divided into query sets and reference sets. The current model is used to extract features of the target domain unlabeled dataset. The extracted features are clustered using the DBSCAN algorithm. The clustered cluster IDs are used as pseudo labels for the corresponding pedestrian images in the target domain unlabeled dataset to obtain the target domain pseudo-labeled dataset. Initialize the memory bank using the outlier features generated by clustering; The memory bank is used to store outlier features; And used to calculate the outlier feature loss function; The current model is trained using the target domain pseudo-label dataset, confidence, and memory library. In each round of training, the target domain pseudo-label dataset, confidence, and memory library are rebuilt until the current model reaches the optimal value, thus obtaining a cross-domain pedestrian re-identification model. Use the cross-domain pedestrian re-identification model to identify the pedestrian images to be identified.

2. The cross-domain person re-identification method with dynamic confidence according to claim 1, characterized in that: The method of using the source domain training set to pre-train the ResNet-50 network to obtain the current model specifically includes: The ResNet-50 network is pre-trained using the cross entropy loss function and the triplet loss function.

3. The cross-domain person re-identification method with dynamic confidence according to claim 2 is characterized in that: The process of joint pre-training is: The features after the pooling layer of the ResNet-50 network use the triplet loss function; The cross entropy loss function is used after the fully connected layer of the ResNet-50 network.

4. The cross-domain person re-identification method with dynamic confidence according to claim 1, characterized in that: Determining the mean average precision (mAP) of the current model on the source domain test set specifically includes: Remove the last fully connected layer of the current model and retain the adaptive average pooling layer; Use the processed current model to extract features of query sets and gallery sets respectively; Calculate the similarity between each person image in the query set and each person image in the reference set gallery sets; Calculate the mAP of the current model on the source domain test set based on the similarity.

5. The cross-domain person re-identification method with dynamic confidence according to claim 1, characterized in that: The method uses the target domain pseudo-label dataset, confidence level, and memory library to train the current model until the current model reaches the optimal state, thereby obtaining a cross-domain person re-identification model. Specifically, the method includes: Add a fully connected layer after the current model, and the output dimension is the number of clusters; The features extracted from the target domain unlabeled dataset are used to calculate the triplet loss function; and the features extracted from the target domain unlabeled dataset are input into the fully connected layer to obtain the classification results and used to calculate the classification loss function; Supervised learning is performed based on the target domain pseudo-label dataset and confidence level; the outlier feature loss function is calculated based on the memory library, and the outlier feature is updated in real time using the outlier feature loss function.

6. The cross-domain person re-identification method with dynamic confidence according to claim 5, characterized in that: The outlier feature loss function is: Among them, L discret is the outlier feature loss function value, f i t is the target domain pedestrian image feature, f k The top k features in similarity ranking, f n are other features whose similarity ranking is not in the top k, m is a hyperparameter, N B This is the number of pedestrian images in the mini-batch at this time, and N2 is the number of outlier features.

7. The cross-domain person re-identification method with dynamic confidence according to claim 6, characterized in that: The outlier feature loss function is determined based on the memory library, and the outlier feature loss function is used to update the outlier feature in real time. Specifically, the formula includes the following: in, is the outlier feature loss function relative to feature f j The partial derivative value is used to update the feature, K i are the top k features in similarity ranking.

8. A dynamic confidence cross-domain pedestrian re-identification system, characterized by: include: The pre-training module is used to pre-train the ResNet-50 network using the source domain training set to obtain the current model; The source domain is a labeled person re-identification dataset; The source domain training set is data in the person re-identification dataset used for model training; Confidence determination module, used to determine the mean average precision (mAP) of the current model on the source domain test set; The mAP is used to estimate the confidence of the current model on the target domain unlabeled dataset; the source domain test set is divided into query sets and reference set gallery sets; The target domain pseudo-label dataset determination module is used to extract features of the target domain unlabeled dataset using the current model; the extracted features are clustered using the DBSCAN algorithm, and the cluster IDs after clustering are used as pseudo-labels for the corresponding pedestrian images in the target domain unlabeled dataset to obtain the target domain pseudo-label dataset; A memory bank initialization module is used to initialize the memory bank using outlier features generated by clustering; The memory bank is used to store outlier features; And used to calculate the outlier feature loss function; The cross-domain person re-identification model determination module is used to train the current model using the target domain pseudo-label dataset, confidence level, and memory library. In each round of training, the target domain pseudo-label dataset, confidence level, and memory library are rebuilt until the current model reaches the optimal value, thus obtaining the cross-domain person re-identification model. The recognition module is used to identify the pedestrian images to be identified using the cross-domain pedestrian re-identification model.

Images