Cross-domain pedestrian re-identification method and system based on global-local feature fusion

Abstract

The invention discloses a cross-domain pedestrian re-identification method and system based on global-local feature fusion, and belongs to the technical field of pedestrian re-identification, and the method comprises the steps: carrying out the style alignment of a source domain data set, constructing a first model for learning, obtaining a pedestrian global feature vector, and carrying out the joint optimization through combining cross entropy loss and soft triple loss; a target domain fine tuning stage: constructing a second model including a teacher model and a student model based on the first model, fusing the global features with the local features to obtain corresponding upper body local fusion features and lower body local fusion features, clustering the local fusion features and the global features to obtain a corresponding pseudo-tag set, and performing target domain fine tuning on the pseudo-tag set; and evaluating and optimizing to obtain a cross-domain pedestrian re-identification model. According to the method, the cross-domain difference between the source domain and the target domain is effectively relieved, collaborative modeling of the global feature and the local discrimination feature of the pedestrian is realized, and the training stability of the target domain model and the pedestrian re-identification precision are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of pedestrian re-identification technology, and in particular to a cross-domain pedestrian re-identification method and system based on global-local feature fusion. Background Technology

[0002] With the rapid development of public safety, intelligent transportation, and smart cities, the deployment of surveillance cameras in urban public spaces is constantly expanding. Accurate pedestrian identification across different cameras has become a key technology in intelligent video analytics systems. Pedestrian re-identification technology aims to match and retrieve the identity of the same pedestrian from images or video sequences collected at different times, locations, and from different camera devices. It is of great significance in practical applications such as security monitoring, criminal investigation, and personnel trajectory analysis.

[0003] Traditional pedestrian re-identification methods primarily rely on deep convolutional neural networks (CNNs) to extract global features of pedestrians. However, in practical applications, different cameras exhibit significant differences due to factors such as shooting angle, lighting, background, and image style, leading to substantial inter-domain discrepancies between training data and the actual deployment environment. This severely impacts the accuracy and robustness of the identification. To reduce the cost of manual annotation and improve the model's adaptability in unknown scenarios, researchers have proposed unsupervised cross-domain pedestrian re-identification methods. These methods typically utilize source domain data with annotation information for pre-training and, without relying on manual annotation in the target domain, achieve model adaptation to the target domain through feature transfer, clustering to generate pseudo-labels, and other methods. However, unsupervised cross-domain pedestrian re-identification tasks still face many complex challenges. On the one hand, the significant differences in appearance styles between the source and target domains make direct feature transfer prone to introducing noise and affecting clustering quality. On the other hand, most existing methods rely mainly on global features for modeling, making it difficult to fully characterize the local structural information of pedestrians. In cases of occlusion, pose changes, or weak discriminative power in local areas, model performance tends to degrade. Summary of the Invention

[0004] This application aims to provide a cross-domain pedestrian re-identification method and system based on global-local feature fusion, which effectively alleviates the cross-domain differences between the source domain and the target domain, realizes collaborative modeling of global pedestrian features and local discriminative features, and significantly improves the training stability of the target domain model and the accuracy of pedestrian re-identification.

[0005] To achieve the above objectives, the technical solution of this application is as follows: A cross-domain person re-identification method based on global-local feature fusion includes, In the source domain pre-training stage, style alignment is performed on the source domain dataset, and the first model is constructed to learn and obtain the global feature vector of pedestrians; Construct a source domain pre-training loss function, train the first model, and obtain the trained first model; In the target domain fine-tuning stage, a second model including a teacher model and a student model is constructed based on the first model. The parameters of the second model are initialized using the parameters of the trained first model. The global features, upper body local features and lower body local features of the target domain image are extracted through the teacher model and the student model, respectively. The global features are fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features. Clustering of local fusion features and global features yields corresponding pseudo-label sets. Based on the pseudo-label sets, the teacher model parameters are updated using an exponential moving average method. The second model is then evaluated and optimized to obtain a cross-domain pedestrian re-identification model. Based on the cross-domain pedestrian re-identification model, global features, upper body local fusion features, and lower body local fusion features of the teacher model are extracted for similarity matching to obtain cross-domain pedestrian re-identification results.

[0006] Optionally, constructing the first model includes: constructing a backbone network and an identity branch network; constructing the backbone network uses attention units combined with a residual network, including: inputting a source domain dataset, adding attention operations between the convolution operations of the second and third residual blocks and between the convolution operations of the third and fourth residual blocks to expand the scope of global dependency modeling, and obtaining the residual processing result after the convolution operation of the fourth residual block as the input of the identity branch network; The identity branch network construction includes: inputting the residual processing result, performing a global average pooling operation to obtain the global average pooling residual processing result, performing a batch normalization operation on the global average pooling residual processing result to obtain the pedestrian global feature vector, and performing a fully connected operation on the pedestrian global feature vector to obtain the predicted pedestrian identity probability. ,in, Indicates the first n The predicted probability value of a pedestrian's identity. N Indicates the total number of pedestrian identities; Style alignment of the source domain dataset includes: treating the source domain dataset as a single style domain, using a cyclic consistency generative adversarial network with color consistency constraints to learn the overall appearance style features of the target domain dataset, uniformly converting the source domain dataset into an image style consistent with the target domain dataset, and aligning the source domain dataset with the target domain dataset.

[0007] Optionally, construct the source domain pre-training loss function, including: constructing the source domain pre-training loss function using the cross-entropy loss function and the soft triplet loss function, as shown below: in, This represents the identity classification loss during the source domain pre-training phase; Indicates the number of samples in the source domain. i Indicates the first i Each source domain sample refers to pedestrian image data with pedestrian identification labels from the source domain dataset; Represents the cross-entropy loss function; This represents the parameters of the first model in the current loop; This represents a source domain identity classifier that predicts the probability of pedestrian identities. get; Indicates in the parameter Below, the first model extracts the feature vector from the source domain sample input; This represents the soft triplet metric loss during the source domain pre-training phase. , , They represent the first i The anchor sample, the first i The positive sample, the first i One negative sample; Representing vectors Euclidean norm; This represents the source domain pre-training loss function.

[0008] Optionally, a second model is constructed based on the first model, including a teacher model and a student model. Global features, upper body local features, and lower body local features are extracted from the target domain dataset using the teacher and student models respectively. The global features are then fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features. This includes: constructing teacher and student models, each including a backbone network, based on the backbone network and identity branch network of the first model, as well as corresponding global and local branch networks for the teacher and student models, respectively, thus completing the construction of the second model. Specifically, this includes: In the teacher model and student model, the backbone network of the first model is input to the target domain training set. Attention operations are added between the convolution operations of the second residual block and the third residual block, and between the convolution operations of the third residual block and the fourth residual block to expand the scope of global dependency modeling. After the convolution operation of the fourth residual block, the high-dimensional global feature map extracted from the pedestrian image in the target domain is obtained as the input of the corresponding global branch network. In the global branch networks corresponding to the teacher model and the student model, based on the identity branch network of the first model, high-dimensional global feature maps are input respectively, and global average pooling operation is used to obtain the residual processing result after global average pooling. Instance normalization and batch normalization operations are performed on the residual processing result after global average pooling to obtain the global feature vector of the student model and the global feature vector of the teacher model respectively. In the local branch networks corresponding to the teacher and student models, based on the region partitioning strategy, the channel variance is calculated in the height direction and averaged along the width direction according to the high-dimensional global feature maps obtained from the teacher and student models, respectively, to obtain the height response distribution. The height response distribution is normalized, and the segmentation point at the preset ratio is determined based on the cumulative probability of the height response distribution. The upper body region feature sub-map and the lower body region feature sub-map are automatically generated. The upper body local features, lower body local features of the student model, and upper body local features, lower body local features of the teacher model are obtained through local pooling mapping layers, respectively. This enables adaptive region segmentation of pedestrian images based on the statistical changes in the height dimension of the feature maps output by the backbone network. Adaptive cross-enhancement channel attention operations are performed on the local features of the upper body and lower body of the student model. Enhanced local features of the upper body and lower body are obtained through bidirectional interactive channel enhancement. The enhanced upper body local features and enhanced lower body local features of the student model are fused with the global feature vector of the teacher model to obtain the upper body local fusion features and the lower body local fusion features.

[0009] Optionally, the adaptive cross-enhanced channel attention operation includes: performing global average pooling on the input features to obtain global average pooling channel statistical weights and global average pooling channel features; performing global max pooling on the input features to obtain global max pooling channel statistical weights and global max pooling channel features; performing adaptive weighted fusion on the global average pooling channel statistical weights, global average pooling channel vectors, global max pooling channel statistical weights, and global max pooling channel features; using a shared multilayer perceptron to perform nonlinear transformations on the global average pooling channel features and global max pooling channel features respectively to obtain global average pooling channel attention mapping features and global max pooling channel attention mapping features; summing these features and applying an activation function to obtain the pedestrian discrimination features corresponding to the channels with weight values ​​close to 1; finally, multiplying these features element-wise along the channel dimension by the pedestrian discrimination features corresponding to the channels to output the enhanced features.

[0010] Optionally, the fusion operation includes: two adaptive cross-enhancement channel attention operations. After inputting the upper body local features and lower body local features of the student model, adaptive cross-enhancement channel attention operations are performed on the upper body local features and lower body local features of the student model to obtain enhanced upper body local features and enhanced lower body local features, respectively. The enhanced upper body local features and enhanced lower body local features of the student model are structurally aligned and vector-level integrated with the global feature vector of the teacher model. After global attention operation and batch normalization operation, the upper body local fusion features and lower body local fusion features are obtained. The correspondence between global semantics and local structure is maintained through feature concatenation and additive fusion, thereby forming a more discriminative fusion feature representation.

[0011] Optionally, clustering the local fusion features and global features yields the corresponding pseudo-label set, specifically including: The mean clustering algorithm is used to cluster and remove outliers from the local fusion features of the upper body and the lower body, resulting in pseudo-label sets for the local fusion features of the upper body and the lower body. In the global branch networks corresponding to the teacher model and the student model, the local features of the upper body of the teacher model, the local features of the lower body of the teacher model, the local features of the upper body of the student model, and the local features of the lower body of the teacher model are respectively encoded through feature encoding to obtain the local features of the upper body of the teacher model, the local features of the lower body of the teacher model, and the local features of the lower body of the student model. The mean clustering algorithm is used to cluster and remove outliers from the local features of the upper body and the lower body, respectively, to obtain the pseudo-label set of the local features of the upper body and the pseudo-label set of the local features of the lower body. The samples in the upper body local feature pseudo-label set and the upper body local feature pseudo-label set are matched. When the pseudo-labels are consistent, the sample pair is retained. When the pseudo-labels are inconsistent, the pseudo-label with higher confidence is retained, resulting in the final upper body local feature pseudo-label set. The samples in the lower body local feature pseudo-label set and the lower body local feature pseudo-label set are matched. When the pseudo-labels are consistent, the sample pair is retained. When the pseudo-labels are inconsistent, the pseudo-label with higher confidence is retained. Samples with mismatched sample indices are retained, resulting in the final lower body local feature pseudo-label set. In the global branch networks corresponding to the teacher model and the student model, the global features of the teacher model and the student model are obtained through the generalized mean pooling layer based on the global features of the teacher model and the global features of the student model. The encoded global features of the teacher model and the encoded global features of the student model are obtained through feature encoding. The mean clustering algorithm is used to cluster and remove outliers from the global features of the encoded teacher model, resulting in a global feature pseudo-label set.

[0012] Optionally, constructing a target domain fine-tuning loss function based on the pseudo-label set includes: constructing a target domain fine-tuning loss function based on the final upper body local feature pseudo-label set, the final lower body local feature pseudo-label set, and the global feature pseudo-label set as supervision. The target domain fine-tuning loss function includes identity classification loss, triplet metric loss, and local similarity loss, as shown below: in, The identity classification loss represents the global features. The loss is represented by a soft triplet metric for global features, using... The loss function representing global features, using the global feature pseudo-label set. As a training supervisor; The soft triplet loss represents the loss of the upper body local features, using the upper body local features of the student model. Using these as input parameters, the final set of pseudo-labels for local features of the upper body is obtained. As a training supervisor; The soft triplet loss represents the local features of the lower body, using the local features of the lower body in the student model. Using these as input parameters, the final set of pseudo-labels for local features of the lower body is obtained. As a training supervisor; Hyperparameters representing the loss of local similarity; The local similarity loss is represented by the cosine similarity between the upper and lower body local features of the student model and the teacher model, respectively. Minimizing the negative similarity is used as the training objective to encourage consistency in local features of the same person across different body regions, as shown below: in, The number of samples in the target domain dataset used for training; Indicates the first One target domain pedestrian sample; , These represent the parameters of the second model. The extracted first Local feature vectors of the upper and lower body of each sample.

[0013] Optionally, the teacher model parameters are updated using an exponential moving average method, and the second model is evaluated and optimized to obtain a cross-domain person re-identification model. This includes: performing backpropagation and parameter updates on the student model with the goal of minimizing the target domain fine-tuning loss function; using an exponential moving average strategy to synchronously update the teacher model parameters using the updated student model parameters, so that the teacher model maintains a smoother and more stable feature representation capability; and updating the teacher model parameters using an exponential moving average strategy.

[0014] A cross-domain pedestrian re-identification system based on global-local feature fusion includes: one or more processors; and a storage device for storing one or more programs that, when executed by the one or more processors, enable the one or more processors to implement the cross-domain pedestrian re-identification method based on global-local feature fusion as described above.

[0015] The cross-domain pedestrian re-identification method and system proposed in this application, based on global-local feature fusion, effectively suppresses the negative impact of noisy pseudo-labels on the training of the cross-domain pedestrian re-identification model by fusing global and local discriminative features and introducing a stable pseudo-label generation and update mechanism. This alleviates the cross-domain differences between the source and target domains, enables collaborative modeling of global and local discriminative features of pedestrians, and significantly improves the recognition accuracy and training stability of the cross-domain pedestrian re-identification model in the target domain.

[0016] To make the above-mentioned features and advantages of the application more apparent and understandable, specific embodiments are provided below, and detailed descriptions are given in conjunction with the accompanying drawings. Attached Figure Description

[0017] Figure 1 The flowchart shows the cross-domain pedestrian re-identification method based on global-local feature fusion provided in this application.

[0018] Figure 2 This is a schematic diagram of the backbone network and identity branch network of the first model proposed in this application.

[0019] Figure 3 This is a schematic diagram of the second model proposed in this application.

[0020] Figure 4 This is a schematic diagram of the adaptive cross-enhanced channel attention operation of the second model proposed in this application.

[0021] Figure 5 This is a schematic diagram of the fusion operation of the second model proposed in this application.

[0022] Figure 6 Visualization results of the cross-domain pedestrian re-identification method based on global-local feature fusion provided in this application. Detailed Implementation

[0023] To make the objectives and technical solutions of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0024] In the first embodiment of this application, please refer to Figure 1 , Figure 1 The flowchart of the cross-domain pedestrian re-identification method based on global-local feature fusion provided in this application includes the following steps: S1~S5.

[0025] Step S1: In the source domain pre-training stage, style alignment is performed on the source domain dataset, and the first model is constructed to learn and obtain the global feature vector of pedestrians; Step S2: Construct the source domain pre-training loss function, train the first model, and obtain the trained first model; Step S3: Target domain fine-tuning stage. Based on the first model, a second model including a teacher model and a student model is constructed. The parameters of the second model are initialized using the parameters of the trained first model. The global features, upper body local features and lower body local features of the target domain image are extracted through the teacher model and the student model, respectively. The global features are fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features. Step S4: Cluster the local fusion features and global features to obtain the corresponding pseudo-label set. Based on the pseudo-label set, update the teacher model parameters using the exponential moving average (EMA) method. Evaluate and optimize the second model to obtain the cross-domain pedestrian re-identification model. Step S5: Extract global features, upper body local fusion features, and lower body local fusion features of the teacher model based on the cross-domain pedestrian re-identification model for similarity matching to obtain the cross-domain pedestrian re-identification results.

[0026] The cross-domain pedestrian re-identification method based on global-local feature fusion provided in this application effectively suppresses the negative impact of noisy pseudo-labels on the training of the cross-domain pedestrian re-identification model by fusing global and local discriminative features and introducing a stable pseudo-label generation and update mechanism. It also alleviates the cross-domain differences between the source and target domains, realizes the collaborative modeling of global and local discriminative features of pedestrians, and significantly improves the recognition accuracy and training stability of the cross-domain pedestrian re-identification model in the target domain.

[0027] In step S1, please refer to Figure 1 In step S1, the source domain pre-training stage, style alignment is performed on the source domain dataset to construct the first model and learn to obtain the global feature vector of pedestrians.

[0028] As an example, during the source domain pre-training phase, a source domain dataset is constructed, including: a source domain training set, a source domain validation set, a source domain query set, and a source domain graph set.

[0029] As an example, style alignment of the source domain dataset includes: treating the source domain dataset as a single style domain, using a CycleGAN with color consistency constraints to learn the overall appearance style features of the target domain dataset, uniformly converting the source domain dataset into an image style consistent with the target domain dataset, and style aligning the source domain dataset and the target domain dataset to generate enhanced samples in the target domain style, thereby reducing the appearance difference between the source domain and the target domain and improving the diversity of training samples.

[0030] In step S2, please refer to Figure 1 In step S2, the source domain pre-training loss function is constructed, and the first model is trained to obtain the trained first model.

[0031] For example, please refer to Figure 2 , Figure 2 This diagram illustrates the backbone network and identity branch network of the first model proposed in this application. Constructing the first model includes building the backbone network and the identity branch network. The backbone network is constructed using attention units combined with a residual network (ResNet), including: inputting a source domain dataset; adding non-local attention operations between the convolutional operations of the second residual block (Layer 2) and the third residual block (Layer 3), and between the convolutional operations of the third residual block (Layer 3) and the fourth residual block (Layer 4), respectively, to expand the scope of global dependency modeling; and obtaining the residual processing result after the convolutional operation of the fourth residual block (Layer 4), which serves as the input to the identity branch network. The residual network includes: a first residual block (Layer 1), a second residual block (Layer 2), a third residual block (Layer 3), and a fourth residual block (Layer 4).

[0032] Furthermore, constructing the identity branch network includes: inputting the residual processing result, applying a global average pooling (GAP) operation to obtain the global average pooling residual processing result, performing a batch normalization operation (BatchNorm) on the global average pooling residual processing result to obtain the pedestrian global feature vector, and performing a fully connected (FC) operation on the pedestrian global feature vector to obtain the predicted pedestrian identity probability. , where represents the first n The predicted probability value of a pedestrian's identity. N This indicates the total number of pedestrian identities.

[0033] In one embodiment of this application, the residual network includes a ResNet50 residual network.

[0034] As an example, constructing a source domain pre-training loss function This includes: constructing a source domain pre-training loss function using the cross-entropy loss function and the soft triplet loss function. , means as follows: (1) (2) (3) in, This represents the identity classification loss during the source domain pre-training phase; Indicates the number of samples in the source domain. i Indicates the first i Each source domain sample refers to pedestrian image data with pedestrian identification labels from the source domain dataset; Represents the cross-entropy loss function; This represents the parameters of the first model in the current loop; This represents a source domain identity classifier that predicts the probability of pedestrian identities. get; Indicates in the parameter Below, the first model extracts the feature vector from the source domain sample input; This represents the soft triplet metric loss during the source domain pre-training phase. , , They represent the first i The anchor sample, the first i The positive sample, the first i One negative sample; Representing vectors Euclidean norm; This represents the source domain pre-training loss function.

[0035] As an example, the first model is trained to obtain the trained first model, including: a loss function pre-trained in the source domain. With the goal of minimizing, the first model is trained using the source domain training set to obtain the trained first model.

[0036] Furthermore, after training the first model and obtaining the trained first model, the process also includes: evaluating the performance of the first model using a source domain validation set, and adjusting the model parameters and training strategy based on the evaluation results.

[0037] As an example, the performance of the first model is evaluated using a source domain validation set. Based on the evaluation results, the model parameters and training strategies are adjusted, including: using the source domain validation set as input to the first model, calculating the evaluation index of the first model on the source domain validation set; when the evaluation index continues to rise in several consecutive rounds of validation calculations, it is determined that the first model has not yet fully converged; the model problem is determined based on the evaluation index; and the corresponding model parameters or training strategies are selected for optimization and parameter updates; when the evaluation index tends to stabilize in several consecutive rounds of validation calculations and the improvement is no longer significant, it is determined that the first model has basically converged, the optimization of the first model ends, and a set of first model parameters with strong discriminative ability is obtained.

[0038] Specifically, the evaluation metrics include Rank-1 accuracy and mean average precision (mAP).

[0039] Specifically, select the corresponding model parameters or training strategies for optimization and parameter updates. For example, increase network capacity when underfitting and add regularization when overfitting.

[0040] In step S3, please refer to Figure 1 In step S3, the target domain fine-tuning stage, a second model including a teacher model and a student model is constructed based on the first model. The parameters of the second model are initialized using the parameters of the trained first model. Global features, upper body local features, and lower body local features of the target domain dataset are extracted through the teacher model and the student model, respectively. The global features are fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features.

[0041] As an example, during the target domain fine-tuning phase, a target domain dataset is constructed, including: a target domain training set, a target domain query set, and a target domain graph set. The target domain dataset is an unlabeled dataset.

[0042] For example, please refer to Figure 3 , Figure 3 This is a schematic diagram of the second model proposed in this application. Based on the first model, a second model including a teacher model and a student model is constructed. Global features, upper body local features, and lower body local features are extracted from the target domain dataset using the teacher and student models respectively. The global features are then fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features. This includes: constructing a teacher model and a student model, each including a backbone network, based on the backbone network and identity branch network of the first model, as well as corresponding global and local branch networks for the teacher and student models respectively, thus completing the construction of the second model. Specifically, this includes steps S31 to S35.

[0043] Step S31: In the teacher model and student model, based on the backbone network of the first model, the target domain training set is input respectively. Attention operations (Non-local) are added between the convolution operations of the second residual block Layer2 and the third residual block Layer3, and between the convolution operations of the third residual block Layer3 and the fourth residual block Layer4 to expand the scope of global dependency modeling. After the convolution operation of the fourth residual block Layer4, the high-dimensional global feature map extracted from the pedestrian image in the target domain is obtained as the input of the corresponding global branch network.

[0044] Step S32: In the global branch networks corresponding to the teacher model and the student model, based on the identity branch network of the first model, input the high-dimensional global feature map respectively, and use global average pooling to obtain the residual processing result after global average pooling. Perform instance normalization and batch normalization (IBN) operations on the residual processing result after global average pooling to obtain the global feature vector of the student model respectively. Global feature vector of the teacher model Further analysis of the global feature vector of the student model Perform a fully connected operation to obtain the probability of the student model predicting the pedestrian's identity.

[0045] In one embodiment of this application, the residual network includes a ResNet50 residual network.

[0046] Step S33: In the local branch networks corresponding to the teacher model and the student model, based on the region partitioning strategy, the channel variance is calculated in the height direction and averaged along the width direction according to the high-dimensional global feature maps obtained from the teacher model and the student model respectively to obtain the height response distribution. Normalization Softmax is performed on the height response distribution, and the segmentation point at the preset ratio is determined based on the cumulative probability of the height response distribution. The upper body region feature sub-map (upper) and the lower body region feature sub-map (lower) are automatically generated. The upper body local features of the student model are obtained through the local pooling mapping layer (GAM). Local features of the lower body of the student model and partial features of the upper body of the teacher model Local features of the lower body of the teacher model This enables adaptive region segmentation of pedestrian images based on statistical changes in the height dimension of the feature maps output by the backbone network.

[0047] Specifically, the local pooling mapping layer includes max pooling, average pooling, and fully connected mapping operations.

[0048] Specifically, determining the cutoff point at a preset proportion based on the cumulative probability of the high response distribution includes: the preset proportion is expressed as... ,in The number of regions to be divided. This is the current partition index; when divided into upper and lower regions, i.e. Preset ratio The value is fixed at 0.5, meaning that when the cumulative probability of the height response distribution exceeds 0.5, the corresponding image location is determined as the segmentation point.

[0049] Step S34: Local features of the upper body of the student model Local features of the lower body of the student model Adaptive Cross-Enhanced Channel Attention (ACECA) is performed to obtain enhanced upper body local features through bidirectional interactive channel enhancement. Enhance the local features of the lower body This allows the channel responses of the two to mutually constrain and compensate each other, thereby strengthening the discrimination features of key parts of the human body and suppressing local noise areas, thus improving the ability to distinguish local features.

[0050] Specifically, please refer to Figure 4 , Figure 4 This diagram illustrates the adaptive cross-enhancement channel attention operation of the second model proposed in this application. The adaptive cross-enhancement channel attention operation employs a cross-gate structure to achieve cross-region feature dependency modeling, establishing bidirectional channel modulation between upper body local features and lower body local features for feature enhancement. The adaptive cross-enhancement channel attention operation includes: performing global average pooling on the input features to obtain global average pooling channel statistical weights and global average pooling channel features; and simultaneously performing global max pooling (GMP) on the input features to obtain global max pooling channel statistical weights and global max pooling channel features. Furthermore, the statistical weights of the global average pooling channels are calculated along one path. Global average pooling channel vector Global max pooling channel statistical weights and global max pooling channel characteristics Adaptive weighted fusion is performed to obtain Another approach uses a shared multilayer perceptron (Shared MLP) to process the global average pooling channel features. Global max pooling channel characteristics A nonlinear transformation is performed to obtain the global average pooling channel attention mapping features. and global max pooling channel attention mapping features After summing, the sigmoid activation function is used to obtain pedestrian discrimination features corresponding to channels with weight values ​​close to 1. Finally, Pedestrian discrimination features corresponding to the passage Perform element-wise multiplication along the channel dimension to output enhanced features.

[0051] Using partial features of the upper body of the student model For example, input the local features of the upper body of the student model. Local features of the upper body of the student model Global average pooling is performed to obtain the global average pooling channel statistical weights. and global average pooling channel characteristics At the same time, local features of the upper body of the student model were analyzed. Perform global max pooling to obtain global max pooling channel statistical weights. and global max pooling channel characteristics Furthermore, the statistical weights of the global average pooling channels are calculated along one path. Global average pooling channel vector Global max pooling channel statistical weights and global max pooling channel characteristics Adaptive weighted fusion is performed to obtain Another approach uses a shared multilayer perceptron to process the global average pooling channel features. Global max pooling channel characteristics A nonlinear transformation is performed to obtain the global average pooling channel attention mapping features. and global max pooling channel attention mapping features After summing, the sigmoid activation function is used to obtain pedestrian discrimination features corresponding to channels with weight values ​​close to 1. Finally, Pedestrian discrimination features corresponding to the passage Perform element-wise multiplication along the channel dimension to output enhanced upper body local features. .

[0052] Step S35: Enhance the upper body local features of the student model Enhance the local features of the lower body Respectively compared with the global feature vector of the teacher model By performing fusion, the partial fusion features of the upper body are obtained. Lower body partial fusion features .

[0053] Specifically, please refer to Figure 5 , Figure 5This is a schematic diagram of the fusion operation of the second model proposed in this application. The fusion operation includes: two adaptive cross-enhancement channel attention operations, inputting local features of the upper body of the student model. Local features of the lower body of the student model Then, the local features of the upper body of the student model were analyzed. Local features of the lower body of the student model Adaptive cross-enhancing channel attention operation is performed to obtain enhanced upper body local features. Enhance the local features of the lower body Enhance the upper body local features of the student model Enhance the local features of the lower body Respectively compared with the global feature vector of the teacher model Structural alignment and vector-level integration are performed, followed by global attention (GAM) and batch normalization (BatchNorm) operations to obtain the upper body local fusion features. Lower body partial fusion features By using feature concatenation and additive fusion, the correspondence between global semantics and local structure is maintained, thereby forming a more discriminative fusion feature representation.

[0054] As an example, the second model parameters are initialized using the first model parameters after training, including: when entering the target domain fine-tuning stage, the first model parameters are copied and used as the initial parameters of the teacher model and the student model respectively.

[0055] Specifically, the parameters of the first model include: learning rate, number of training epochs, and batch size. The parameters of the second model include: learning rate, number of training epochs, and momentum coefficient of the exponential moving average strategy. The number of samples input to the model at one time. Among these, the momentum coefficient of the exponential moving average strategy. .

[0056] In step S4, please refer to Figure 1 In step S4, clustering the local fusion features and global features yields the corresponding pseudo-label sets. Based on the pseudo-label sets, a target domain fine-tuning loss function is constructed. The teacher model parameters are updated using an exponential moving average method. The second model is then evaluated and optimized to obtain the cross-domain pedestrian re-identification model.

[0057] As an example, clustering the local fusion features and global features yields the corresponding pseudo-label set, specifically including steps S41 to S45.

[0058] Step S41: Use the mean clustering algorithm to fuse local features of the upper body. Lower body partial fusion features Clustering and outlier removal are performed to obtain a pseudo-label set of local fusion features for the upper body. Lower body partial fusion feature pseudo-label set .

[0059] Specifically, outlier removal involves: after clustering using the mean clustering algorithm to obtain multiple clusters, calculating the local density deviation value for each sample within each cluster. When the local density of a sample is significantly lower than that of most samples within the cluster, and the corresponding Local Outlier Factor (LOF) exceeds a set local outlier factor threshold, the sample is identified as a low-density outlier within the cluster and removed from the pseudo-label set. This filters out approximately 5% of unreliable samples in the clustering results, improving the purity of the pseudo-labels and the stability of the subsequent self-training stage. The local outlier factor threshold is determined through ablation experiments.

[0060] In one embodiment of this application, a comparative analysis of model performance under different outlier removal ratios shows that when the local outlier factor threshold, i.e., the removal ratio, is about 5%, a better balance can be achieved between effectively removing untrusted pseudo-labels and maintaining the integrity of the cluster structure.

[0061] Step S42: In the global partial branch networks corresponding to the teacher model and the student model, respectively analyze the local features of the upper body of the teacher model. Local features of the lower body of the teacher model Partial features of the upper body of the student model Local features of the lower body of the teacher model Local features of the upper body of the encoded teacher model were obtained through feature encoding. Local features of the lower body of the teacher model Partial features of the upper body of the student model Local features of the lower body of the student model The mean clustering algorithm was used to analyze the local features of the upper body of the teacher model. Local features of the lower body of the teacher model Clustering and outlier removal are performed to obtain a pseudo-label set of local features of the upper body. Lower body partial feature pseudo-label set .

[0062] Step S43: Pseudo-label set based on upper body local fusion features Upper body partial feature pseudo-label set The samples in the dataset are matched. If the pseudo-labels match, the sample pair is retained. If the pseudo-labels do not match, the pseudo-label with higher confidence is retained, resulting in the final pseudo-label set of upper body local features. ; Pseudo-label set based on lower body local fusion features Lower body partial feature pseudo-label set The samples in the dataset are matched. If the pseudo-labels match, the sample pair is retained. If the pseudo-labels do not match, the pseudo-label with higher confidence is retained. Samples with mismatched sample indices are also retained, resulting in the final pseudo-label set for lower body local features. .

[0063] Step S44: In the global branch networks corresponding to the teacher model and student model, based on the global features of the teacher model... Global features of the student model Global features of the teacher model are obtained through a generalized mean pooling layer (GeM). Global features of the student model The global features of the encoded teacher model are obtained through feature encoding. Global features of the encoded student model .

[0064] Step S45: Use the mean clustering algorithm to analyze the global features of the encoded teacher model. Clustering and outlier removal are performed to obtain a global feature pseudo-label set. .

[0065] Furthermore, constructing a target domain fine-tuning loss function based on the pseudo-label set includes: based on the final upper body local feature pseudo-label set. Final lower body local feature pseudo-label set Global feature pseudo-label set As a supervised construction of the target domain fine-tuning loss function fine-tuning the loss function in the target domain The loss includes identity classification loss, triplet metric loss, and local similarity loss, as shown below: (4) in, The identity classification loss represents the global features. The loss is represented by a soft triplet metric for global features, using... The loss function representing global features, its construction principle and the source domain pre-training loss function Similarly, based on the global features of the student model For input parameters, the global feature pseudo-label set As a training supervisor; The soft triplet loss represents the loss of the upper body local features, using the upper body local features of the student model. Using these as input parameters, the final set of pseudo-labels for local features of the upper body is obtained. As a training supervisor; The soft triplet loss represents the local features of the lower body, using the local features of the lower body in the student model. Using these as input parameters, the final set of pseudo-labels for local features of the lower body is obtained. As a training supervisor; Hyperparameters representing the loss of local similarity; The local similarity loss is represented by the cosine similarity between the upper and lower body local features of the student model and the teacher model, respectively. Minimizing the negative similarity is used as the training objective to encourage consistency in local features of the same person across different body regions, as shown below: (5) in, The number of samples in the target domain dataset used for training; Indicates the first One target domain pedestrian sample; , These represent the parameters of the second model. The extracted first Local feature vectors of the upper and lower body of each sample.

[0066] As an example, the teacher model parameters are updated using an exponential moving average method, and the second model is evaluated and optimized to obtain a cross-domain person re-identification model, including: fine-tuning the loss function with the target domain. To minimize the impact of the student model, backpropagation and parameter updates are performed. An exponential moving average strategy is employed to synchronously update the teacher model parameters using the updated student model parameters, thereby ensuring a smoother and more stable feature representation capability for the teacher model. The exponential moving average strategy for updating the teacher model parameters is expressed as follows: (6) in, This represents the time-averaged parameters of the network in the previous iteration, i.e., the teacher model parameters to be updated. T Indicates the current moment; The time-averaged parameters of the network in the current iteration; the initial time-averaged parameters. That is, the parameters obtained after the first model evaluation and optimization; This represents the momentum coefficient of the exponential moving average strategy. .

[0067] Furthermore, the second model is evaluated and optimized to obtain a cross-domain pedestrian re-identification model, including: using the target domain validation set as the input of the second model, calculating the evaluation index of the second model on the target domain validation set; when the evaluation index continues to rise in several consecutive rounds of validation calculations, it is determined that the first model has not yet fully converged; the model problem is determined based on the evaluation index; and the corresponding model parameters or training strategies are selected for optimization and parameter updates; when the evaluation index tends to stabilize in several consecutive rounds of validation calculations and the improvement is no longer significant, it is determined that the second model has basically converged, the optimization of the second model ends, and the optimized second model is used as the cross-domain pedestrian re-identification model.

[0068] In step S5, please refer to Figure 1 Step S5 in the process extracts global features of the teacher model based on the cross-domain person re-identification model. , Upper body partial fusion features and lower body partial fusion features Used for similarity matching to obtain cross-domain pedestrian re-identification results.

[0069] As an example, extracting global features of the teacher model based on a cross-domain person re-identification model. , Upper body partial fusion features and lower body partial fusion features For similarity matching, the cross-domain person re-identification results are obtained, including: inputting a query image from the target domain query set, and extracting the teacher model global features of the query image based on the cross-domain person re-identification model. , Upper body partial fusion features and lower body partial fusion features Used to perform similarity matching with images in the target domain image library, query pedestrians in the images, and obtain cross-domain pedestrian re-identification results.

[0070] In one embodiment of this application, please refer to Figure 6 , Figure 6 The image shows a visualization of the cross-domain person re-identification method based on global-local feature fusion provided in this application. Figure 6As shown, the visualization results of the unsupervised cross-domain person re-identification task from the Duke dataset to the Market dataset are presented. Red dashed boxes represent query images, and red solid boxes mark incorrect matching samples in the search results. The upper row within each red dashed box represents the search results of the baseline method, and the lower row represents the search results of the method proposed in this application. The corresponding Rank-10 ranking results are given on the right. It can be observed that, compared to the baseline method, the cross-domain person re-identification method based on global-local feature fusion provided in this application can return correctly matched pedestrians at higher ranking positions in most query samples, and the number of incorrect matching samples is significantly reduced. Secondly, taking the second pedestrian sample as an example for comparative analysis, it can be clearly observed that the baseline method begins to show multiple incorrect matches after Rank-8. Some search results incorrectly rank pedestrians with similar appearances but different identities (such as pedestrians wearing similar short sleeves and shorts) at the top, indicating that its ability to model fine-grained identity discrimination information is limited and easily affected by clothing similarity and posture changes. In contrast, the cross-domain person re-identification method based on global-local feature fusion provided in this application can stably maintain correctly matched pedestrians in a higher ranking range at the corresponding positions, while significantly reducing the number of incorrectly matched samples. This indicates that by introducing fusion modeling of global features and upper and lower body local features, the cross-domain person re-identification method based on global-local feature fusion provided in this application can more accurately focus on the key discrimination regions of pedestrians, effectively suppressing misjudgments caused by background interference and pedestrians with similar appearances, thus further verifying its effectiveness and advantages in the cross-domain person re-identification task from the Duke dataset to the Market dataset.

[0071] The effectiveness of this invention will now be verified using specific data.

[0072] In another embodiment of this application, two publicly available person re-identification datasets, DukeMTMC-reID and MSMT17, are selected. Specifically, the DukeMTMC-reID dataset is used as the source dataset, and MSMT17 is used as the target dataset. The DukeMTMC-reID dataset contains over 2700 pedestrians, captured by 8 cameras, totaling 36411 images. The training set contains 702 pedestrians, totaling 16522 images, and the test set contains 702 pedestrians, totaling 17661 images. The MSMT17 dataset, constructed by Microsoft Research, contains 126441 pedestrian images collected from 15 cameras, involving 4101 pedestrian identities. This dataset covers different time periods and diverse environments, with significant variations in pedestrian appearance, including complex situations such as full-body and partial occlusion, and more diverse background conditions, making it highly challenging for recognition and one of the most challenging benchmark datasets in the current field of person re-identification. The training set contains 1041 identities, and the test set contains 3060 identities. In the proposed model, the model parameters include: learning rate, number of training epochs, and momentum coefficient of the exponential moving average strategy. The number of samples input to the model at one time should be set with reference to commonly used values ​​in unsupervised cross-domain person re-identification and combined with experience. These values ​​are 0.00035, 60, 0.999, and 64 respectively to ensure the stability and reproducibility of model training.

[0073] By comparing with mainstream methods in the field of unsupervised cross-domain person re-identification, including: Mutual Averaging Teaching: a pseudo-label refinement method for unsupervised domain adaptation in person re-identification (MMT), Group-Aware Label Transfer Method: for Domain Adaptive Person Re-identification (GLT), Global Feature Data Augmentation and Pseudo-label Optimization Based on Generative Adversarial Networks: an Unsupervised Domain Adaptive Method for Person Re-identification (DAPRH), Domain Adaptive Person Re-identification Method Based on Domain Alignment and Mutual Pseudo-label Optimization (DAMPR), and Source Domain Knowledge Preservation Method for Cross-Domain Person Re-identification (PKSD), the experimental results are shown in Table 1. The results show that the cross-domain person re-identification method based on global-local feature fusion provided in this application achieves superior results in key performance indicators such as Rank-1 accuracy and mean average precision.

[0074] Table 1. Person Re-identification Results for the Duke→MSMT17 Dataset This application also provides a cross-domain pedestrian re-identification system, comprising: one or more processors; and a storage device for storing one or more programs that, when executed by the one or more processors, enable the one or more processors to implement the cross-domain pedestrian re-identification method based on global-local feature fusion as described above.

[0075] The cross-domain pedestrian re-identification method and system based on global-local feature fusion provided in this application effectively suppresses the negative impact of noisy pseudo-labels on the training of cross-domain pedestrian re-identification models by fusing global and local discriminative features and introducing a stable pseudo-label generation and update mechanism. This alleviates the cross-domain differences between the source and target domains, realizes collaborative modeling of global and local discriminative features of pedestrians, and significantly improves the recognition accuracy and training stability of the cross-domain pedestrian re-identification model in the target domain.

[0076] Although this application has been disclosed above with reference to embodiments, it is not intended to limit this application. Anyone skilled in the art may make some modifications and refinements without departing from the spirit and scope of this application. Therefore, the scope of protection of this application shall be determined by the appended claims.

Claims

1. A cross-domain person re-identification method based on global-local feature fusion, characterized in that, include, In the source domain pre-training stage, style alignment is performed on the source domain dataset, and the first model is constructed to learn and obtain the global feature vector of pedestrians; Construct a source domain pre-training loss function, train the first model, and obtain the trained first model; In the target domain fine-tuning stage, a second model including a teacher model and a student model is constructed based on the first model. The parameters of the second model are initialized using the parameters of the trained first model. The global features, upper body local features and lower body local features of the target domain image are extracted through the teacher model and the student model, respectively. The global features are fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features. Clustering of local fusion features and global features yields corresponding pseudo-label sets. Based on the pseudo-label sets, the teacher model parameters are updated using an exponential moving average method. The second model is then evaluated and optimized to obtain a cross-domain pedestrian re-identification model. Based on the cross-domain pedestrian re-identification model, global features, upper body local fusion features, and lower body local fusion features of the teacher model are extracted for similarity matching to obtain cross-domain pedestrian re-identification results.

2. The cross-domain person re-identification method based on global-local feature fusion as described in claim 1, characterized in that, The construction of the first model includes: constructing a backbone network and an identity branch network; the backbone network is constructed by combining attention units with a residual network, including: inputting the source domain dataset, adding attention operations between the convolution operations of the second and third residual blocks and between the convolution operations of the third and fourth residual blocks to expand the scope of global dependency modeling, and obtaining the residual processing result after the convolution operation of the fourth residual block as the input of the identity branch network; The identity branch network construction includes: inputting the residual processing result, performing a global average pooling operation to obtain the global average pooling residual processing result, performing a batch normalization operation on the global average pooling residual processing result to obtain the pedestrian global feature vector, and performing a fully connected operation on the pedestrian global feature vector to obtain the predicted pedestrian identity probability. ,in, Indicates the first n The predicted probability value of a pedestrian's identity. N Indicates the total number of pedestrian identities; Style alignment of the source domain dataset includes: treating the source domain dataset as a single style domain, using a cyclic consistency generative adversarial network with color consistency constraints to learn the overall appearance style features of the target domain dataset, uniformly converting the source domain dataset into an image style consistent with the target domain dataset, and aligning the source domain dataset with the target domain dataset.

3. The cross-domain person re-identification method based on global-local feature fusion as described in claim 2, characterized in that, Constructing the source domain pre-training loss function includes: using the cross-entropy loss function and the soft triplet loss function to construct the source domain pre-training loss function, as shown below: in, This represents the identity classification loss during the source domain pre-training phase; Indicates the number of samples in the source domain. i Indicates the first i Each source domain sample refers to pedestrian image data with pedestrian identification labels from the source domain dataset; Represents the cross-entropy loss function; This represents the parameters of the first model in the current loop; This represents a source domain identity classifier that predicts the probability of pedestrian identities. get; Indicates in the parameter Below, the first model extracts the feature vector from the source domain sample input; This represents the soft triplet metric loss during the source domain pre-training phase. , , They represent the first i The anchor sample, the first i The positive sample, the first i One negative sample; Representing vectors Euclidean norm; This represents the source domain pre-training loss function.

4. The cross-domain person re-identification method based on global-local feature fusion as described in claim 1, characterized in that, A second model, comprising a teacher model and a student model, is constructed based on the first model. Global features, upper body local features, and lower body local features are extracted from the target domain dataset using the teacher and student models, respectively. The global features are then fused with the local features to obtain the corresponding upper body local fusion features and lower body local fusion features. This includes: constructing teacher and student models, each with a backbone network, based on the backbone network and identity branch network of the first model, as well as corresponding global and local branch networks for the teacher and student models, thus completing the construction of the second model. Specifically, this includes: In the teacher model and student model, the backbone network of the first model is input to the target domain training set. Attention operations are added between the convolution operations of the second residual block and the third residual block, and between the convolution operations of the third residual block and the fourth residual block to expand the scope of global dependency modeling. After the convolution operation of the fourth residual block, the high-dimensional global feature map extracted from the pedestrian image in the target domain is obtained as the input of the corresponding global branch network. In the global branch networks corresponding to the teacher model and the student model, based on the identity branch network of the first model, high-dimensional global feature maps are input respectively, and global average pooling operation is used to obtain the residual processing result after global average pooling. Instance normalization and batch normalization operations are performed on the residual processing result after global average pooling to obtain the global feature vector of the student model and the global feature vector of the teacher model respectively. In the local branch networks corresponding to the teacher and student models, based on the region partitioning strategy, the channel variance is calculated in the height direction and averaged along the width direction according to the high-dimensional global feature maps obtained from the teacher and student models, respectively, to obtain the height response distribution. The height response distribution is normalized, and the segmentation point at the preset ratio is determined based on the cumulative probability of the height response distribution. The upper body region feature sub-map and the lower body region feature sub-map are automatically generated. The upper body local features, lower body local features of the student model, and upper body local features, lower body local features of the teacher model are obtained through local pooling mapping layers, respectively. This enables adaptive region segmentation of pedestrian images based on the statistical changes in the height dimension of the feature maps output by the backbone network. Adaptive cross-enhancement channel attention operations are performed on the local features of the upper body and lower body of the student model. Enhanced local features of the upper body and lower body are obtained through bidirectional interactive channel enhancement. The enhanced upper body local features and enhanced lower body local features of the student model are fused with the global feature vector of the teacher model to obtain the upper body local fusion features and the lower body local fusion features.

5. The cross-domain person re-identification method based on global-local feature fusion as described in claim 4, characterized in that, The adaptive cross-enhanced channel attention operation includes: performing global average pooling on the input features to obtain global average pooling channel statistical weights and global average pooling channel features; performing global max pooling on the input features to obtain global max pooling channel statistical weights and global max pooling channel features; performing adaptive weighted fusion on the global average pooling channel statistical weights, global average pooling channel vectors, global max pooling channel statistical weights, and global max pooling channel features; and using a shared multilayer perceptron to perform nonlinear transformations on the global average pooling channel features and global max pooling channel features to obtain global average pooling channel attention mapping features and global max pooling channel attention mapping features, respectively. After summing, an activation function is used to obtain the pedestrian discrimination features corresponding to the channels with weight values ​​close to 1. Finally, the channel-dimensional pedestrian discrimination features are multiplied element-wise to output the enhanced features.

6. The cross-domain person re-identification method based on global-local feature fusion as described in claim 4, characterized in that, The fusion operation includes two adaptive cross-enhancement channel attention operations. After inputting the upper body local features and lower body local features of the student model, adaptive cross-enhancement channel attention operations are performed on the upper body local features and lower body local features respectively to obtain enhanced upper body local features and enhanced lower body local features. The enhanced upper body local features and enhanced lower body local features of the student model are structurally aligned and vector-level integrated with the global feature vector of the teacher model. After global attention operation and batch normalization operation, the upper body local fusion features and lower body local fusion features are obtained. The correspondence between global semantics and local structure is maintained through feature concatenation and additive fusion, thereby forming a more discriminative fusion feature representation.

7. The cross-domain person re-identification method based on global-local feature fusion as described in claim 4, characterized in that, Clustering of local fusion features and global features yields corresponding pseudo-label sets, specifically including: The mean clustering algorithm is used to cluster and remove outliers from the local fusion features of the upper body and the lower body, resulting in pseudo-label sets for the local fusion features of the upper body and the lower body. In the global branch networks corresponding to the teacher model and the student model, the local features of the upper body of the teacher model, the local features of the lower body of the teacher model, the local features of the upper body of the student model, and the local features of the lower body of the teacher model are respectively encoded through feature encoding to obtain the local features of the upper body of the teacher model, the local features of the lower body of the teacher model, and the local features of the lower body of the student model. The mean clustering algorithm is used to cluster and remove outliers from the local features of the upper body and the lower body, respectively, to obtain the pseudo-label set of the local features of the upper body and the pseudo-label set of the local features of the lower body. The samples in the upper body local feature pseudo-label set and the upper body local feature pseudo-label set are matched. When the pseudo-labels are consistent, the sample pair is retained. When the pseudo-labels are inconsistent, the pseudo-label with higher confidence is retained, resulting in the final upper body local feature pseudo-label set. The samples in the lower body local feature pseudo-label set and the lower body local feature pseudo-label set are matched. When the pseudo-labels are consistent, the sample pair is retained. When the pseudo-labels are inconsistent, the pseudo-label with higher confidence is retained. Samples with mismatched sample indices are retained, resulting in the final lower body local feature pseudo-label set. In the global branch networks corresponding to the teacher model and the student model, the global features of the teacher model and the student model are obtained through the generalized mean pooling layer based on the global features of the teacher model and the global features of the student model. The encoded global features of the teacher model and the encoded global features of the student model are obtained through feature encoding. The mean clustering algorithm is used to cluster and remove outliers from the global features of the encoded teacher model, resulting in a global feature pseudo-label set.

8. The cross-domain person re-identification method based on global-local feature fusion as described in claim 7, characterized in that, The target domain fine-tuning loss function is constructed based on pseudo-label sets. This involves using the final upper body local feature pseudo-label set, the final lower body local feature pseudo-label set, and the global feature pseudo-label set as supervision to construct the target domain fine-tuning loss function. The target domain fine-tuning loss function includes identity classification loss, triplet metric loss, and local similarity loss, as shown below: in, The identity classification loss represents the global features. The loss is represented by a soft triplet metric for global features, using... The loss function representing global features, using the global feature pseudo-label set. As a training supervisor; The soft triplet loss represents the loss of the upper body local features, using the upper body local features of the student model. Using these as input parameters, the final set of pseudo-labels for local features of the upper body is obtained. As a training supervisor; The soft triplet loss represents the local features of the lower body, using the local features of the lower body in the student model. Using these as input parameters, the final set of pseudo-labels for local features of the lower body is obtained. As a training supervisor; Hyperparameters representing the loss of local similarity; The local similarity loss is represented by the cosine similarity between the upper and lower body local features of the student model and the teacher model, respectively. Minimizing the negative similarity is used as the training objective to encourage consistency in local features of the same person across different body regions, as shown below: in, The number of samples in the target domain dataset used for training; Indicates the first One target domain pedestrian sample; , These represent the parameters of the second model. The extracted first Local feature vectors of the upper and lower body of each sample.

9. The cross-domain person re-identification method based on global-local feature fusion as described in claim 1, characterized in that, The teacher model parameters are updated using an exponential moving average method, and the second model is evaluated and optimized to obtain a cross-domain person re-identification model. This includes: performing backpropagation and parameter updates on the student model with the goal of minimizing the target domain fine-tuning loss function; using an exponential moving average strategy to synchronously update the teacher model parameters with the updated student model parameters, so that the teacher model maintains a smoother and more stable feature representation ability; and updating the teacher model parameters using an exponential moving average strategy.

10. A cross-domain person re-identification system based on global-local feature fusion, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the cross-domain pedestrian re-identification method based on global-local feature fusion as described in any one of claims 1-9.

Images