Obstructed pedestrian re-identification method based on multi-branch cooperation

Abstract

The application belongs to the technical field of pedestrian re-identification, and discloses a kind of occluded pedestrian re-identification method based on multi-branch cooperation, specifically comprising the following steps: truncation based on OSNet main network;Input pedestrian image to truncated OSNet main network, extract initial feature map;Build multi-branch feature extraction module, and the multi-branch feature extraction module includes parallel global complementary branch, double-path local branch and multi-scale channel branch;Initial feature map is input to multi-branch feature extraction module;Extract explicit global feature and implicit global feature in initial feature map by global complementary branch;Local features in initial feature map are extracted by double-path local branch, and local features are reorganized and spliced, to generate fine-grained discriminant fine-grained features and high-dimensional discriminant features;Solve the problem of low recognition accuracy and insufficient robustness in the prior art under complex occlusion scene.

Description

Technical Field

[0001] This invention belongs to the field of pedestrian re-identification technology, specifically relating to a method for re-identifying occluded pedestrians based on multi-branch collaboration. Background Technology

[0002] With the continuous development of technology, pedestrian re-identification is a key technology in intelligent surveillance, public safety and smart city construction. Its goal is to match images or videos of the same pedestrians across camera views. The core challenge of this task is how to extract highly discriminative identity representations from appearance features and maintain good generalization ability under different views and complex scenarios. In recent years, the rapid development of deep convolutional neural networks (CNNs) and visual Transformer architectures, along with the continuous emergence of methods based on global features, local features, attention mechanisms, metric learning, and representation learning, has provided diverse and effective solutions for person re-identification. However, in real-world complex environments, occlusion problems severely impact the performance of person re-identification models. Occlusion leads to incomplete information about the human body region in pedestrian images, making it difficult for models to accurately capture key identity features, thus affecting the accuracy and robustness of recognition. Current mainstream methods still have shortcomings in handling occlusion problems: on the one hand, methods based on global features lack the ability to model fine-grained local regions and are easily affected by background interference and occlusion; on the other hand, while methods based on local features enhance the ability to perceive details, they are prone to feature alignment errors when dealing with occlusion misalignment or changes in viewpoint, affecting the overall recognition effect.

[0003] In recent years, researchers have proposed various improvement strategies, such as introducing attention mechanisms to enhance the feature representation ability of key regions, or improving the expressive ability of occluded images through generation mechanisms. However, these methods are limited in their performance and practical application efficiency in complex occlusion scenarios due to issues such as feature redundancy, unstable training, excessive reliance on prior information, and complex structures. In addition, some methods rely on external pose estimation or semantic segmentation modules, which increases the complexity and computational cost of the model, and performance degradation occurs in real-world scenarios. Therefore, existing technologies suffer from low recognition accuracy and insufficient robustness in complex occlusion scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a multi-branch collaborative method for re-identifying occluded pedestrians, which solves the problems of low recognition accuracy and insufficient robustness in complex occlusion scenarios in existing technologies.

[0005] The objective of this invention can be achieved through the following technical solutions: The occluded person re-identification method based on multi-branch collaboration includes the following steps: Truncation is performed based on the OSNet backbone network; Input pedestrian images into a truncated OSNet backbone network to extract initial feature maps; A multi-branch feature extraction module is constructed, which includes parallel global complementary branches, dual-path local branches, and multi-scale channel branches; The initial feature map is input into the multi-branch feature extraction module; Explicit and implicit global features are extracted from the initial feature map using global complementary branches; Local features are extracted from the initial feature map through dual-path local branching, and the local features are recombined and spliced ​​to generate fine-grained discriminative features and high-dimensional discriminative features. Multi-scale channel features are extracted from the initial feature map through multi-scale channel branching, and the multi-scale channel features are enhanced based on multi-scale deformable convolution and channel attention mechanism. Based on the output of the multi-branch feature extraction module, features of the unit space and rank space are defined and optimized collaboratively. Based on the optimized rank space features, the final pedestrian feature representation is generated, and the re-identification result is output.

[0006] The initial feature map is expressed in the form of C×H×W with dimensions of 512×24×8, where C represents the number of channels in the feature map; H represents the height of the feature map; and W represents the width of the feature map. The OSNet backbone network consists of five convolutional blocks: conv1, conv2, conv3, conv4, and conv5. The OSNet backbone is truncated at conv3_0, dividing it into two parts: conv1, conv2, conv3_0 and conv3_x, conv4, conv5. Here, conv3_0 represents the first convolutional layer in the third convolutional block, and conv3_x represents the remaining convolutional layers in the third convolutional block. The pedestrian image is input into conv1, conv2, and conv3_0, and the initial feature map is output from conv3_0 after processing. The input of conv3_x receives the initial feature map output by conv3_0; The three branches of the multi-branch feature extraction module all follow the structure of conv3_x, conv4, and conv5. The initial feature map is extracted and refined sequentially by conv3_x, conv4, and conv5, and the first feature map with a dimension of 512×24×8 is output by conv5.

[0007] Among them, the global complementary branches include parallel explicit feature paths and implicit feature paths; The global complementary branch extracts explicit and implicit global features from the initial feature map, specifically including the following steps: The initial feature map is sequentially processed by conv3_x, conv4, and conv5 to extract and refine features, and conv5 outputs the first feature map with dimensions of 512×24×8. The first feature map is input into the explicit feature path and the implicit feature path respectively; In the explicit feature path, the first feature map is explicitly extracted by generalized mean pooling (GeM) to obtain an explicit global feature G with a dimension of 512×1×1. In the implicit feature path, a Top DropBlock strategy is used to generate a mask. The mask is used to mask the high-activation regions of the first feature map, outputting a low-response feature region with dimensions of 512×24×8. Then, implicit features in the low-response feature region are extracted by global max pooling (GMP), resulting in an implicit global feature G with dimensions of 512×1×1. drop ; Explicit global feature G and implicit global feature G drop Both are subjected to batch normalization. The explicit global feature G and the implicit global feature G after batch normalization are... drop They are respectively represented as and .

[0008] Dual-path local branches include parallel local feature recombination paths and local feature splicing paths; Local features are extracted from the initial feature map through dual-path local branching, and these local features are then recombined and spliced ​​to generate fine-grained discriminative features and high-dimensional discriminative features. The specific steps include: The initial feature map is sequentially processed by conv3_x, conv4, and conv5 to extract and refine features, and conv5 outputs the first feature map with dimensions of 512×24×8. The first feature map with dimensions of 512×24×8 is input into the local branch of the dual path; The first feature map is divided into four first local feature regions arranged in sequence along the horizontal direction. The dimensions of the four first local feature regions are all 512×6×8. Input all four first local feature regions simultaneously into the local feature recombination path and the local feature splicing path; The local feature reconstruction path specifically includes the following steps: By splicing any two adjacent first local feature regions in the four first local feature regions arranged in sequence, three second local feature regions with overlapping parts are generated. The dimensions of the second local feature regions are all 512×12×8. The three second local feature regions are processed by global average pooling (GAP) to obtain three fine-grained features P1, P2 and P3, each with a dimension of 512×1×1. Batch normalization is performed on fine-grained features P1, P2, and P3 respectively to obtain batch-normalized fine-grained features. , and ; The local feature stitching path specifically includes the following steps: Global average pooling is performed on the four first local feature regions of the input to obtain four local feature vectors with dimensions of 512×1×1. The four local feature vectors are concatenated along the channel dimension to obtain a discriminative feature P4 with a dimension of 2048×1×1; High-dimensional discriminative features are obtained after batch normalization of P4. .

[0009] The multi-scale channel branch includes parallel multi-pooling paths and multi-scale deformable convolutional paths. The first feature map is input into the multi-pooling path and the multi-scale deformable convolutional path for processing to generate multi-scale channel features. The multi-pooling path generates diverse channel feature representations through various pooling methods, and then performs fusion and dimensionality reduction processing. Multi-scale deformable convolution path: Extracting and enhancing multi-scale features through multi-scale deformable convolution and channel attention mechanism; The multi-scale deformable convolution path includes a multi-scale deformable convolution module and a channel attention mechanism module.

[0010] The multi-pooling path specifically includes the following steps: The first feature map is processed by Global Average Pooling (GAP), Generalized Mean Pooling (GeM), and Global Max Pooling (GMP) to generate three different channel feature representations. The dimensions of the three different channel feature representations are all 512×1×1. The three different channel feature representations are concatenated along the channel dimension to form a first fusion feature representation with a dimension of 1536×1×1; The first fused feature representation is sampled by 1×1 convolution to reduce its dimensionality and compress it to generate a compact feature vector with a dimension of 512×1×1; Channel segmentation is performed on the compact feature vector, and a 1×1 convolution operation with shared parameters is performed to decouple the compact feature vector into two independent and complementary sub-features C1 and C2, each of which has a dimension of 256×1×1. Batch normalization is performed on the two sub-features C1 and C2 respectively, and the output is obtained. and .

[0011] Multi-scale deformable convolutional paths specifically include the following steps: The first feature map enters the multi-scale deformable convolution module; The multi-scale deformable convolution module uses three deformable convolution kernels with different receptive fields (3×3, 5×5, and 7×7) to perform convolution processing on the first feature map in order to capture feature information at different scales. Each deformable convolutional kernel dynamically adjusts the sampling point position to obtain the second feature maps at three scales, namely F3, F5 and F7, with each second feature map having a dimension of 512×24×8. The three second feature maps are concatenated along the channel dimension to form a second fused feature representation with dimensions of 1536×24×8; After performing channel compression and batch normalization on the second fused feature representation using 1×1 convolution, a third feature map with dimensions of 512×24×8 is obtained. The third feature map is input from the multi-scale deformable convolution module to the channel attention mechanism module; In the channel attention mechanism module, the height H and width W of the third feature map are compressed to a single real number using global average pooling (GAP), and the global statistic Z for each channel is obtained. C， The third feature map is transformed from C×H×W to C×1×1; Z C The expression is as follows: In the formula, c This represents the first feature map. c One channel; ( k , l ) represents the spatial dimension, where k This represents the vertical coordinate index of the current pixel. l Indicates the horizontal coordinate index of the current pixel; Then, two fully connected layers implemented with 1×1 convolutions are used to learn the nonlinear relationship between channels, generating a weight vector S. c ; The weight vector S c The weights are applied to the third feature map, and the output is the fourth feature map. .

[0012] Based on the output of the multi-branch feature extraction module, features in the unit space and rank space are defined and jointly optimized, specifically including the following steps: The characteristic representations of the unit space I and the rank space R can be formally defined as two sets of features, expressed as follows: Construct a label-smooth identity loss function L ID Multiple similarity loss function L MS and triplet loss function L Triplet Furthermore, a multi-objective joint loss function is constructed by comprehensively considering this. The expression for the multi-objective joint loss function is as follows: This is the balance factor for the identity loss function; It serves as a balance factor for the multiple similarity loss function and the triplet loss function; The feature set of a unit space I is smoothed only by the identity loss function of the label. L ID Perform optimized training; The feature set of rank space R is obtained through a multi-objective joint loss function. L total Optimize the training.

[0013] A label-smooth identity loss function is constructed to supervise the classification of pedestrian feature representations. The formula for constructing the label-smooth identity loss function is as follows: in, u Indicates pedestrian category, u ∈﹛1,2,...,N﹜;N represents the set number of pedestrian images in the training set; y This is a real label; pu For the network's predicted logical value, pu The expression is as follows: in, x t Let t be the feature of the t-th sample; C The total number of label types; W v Features x t Does it belong to the category of tags? v The classifier consists of fully connected layers; Indicates the first tThe classifier weight vector corresponding to the true label of each sample; Indicates the transpose operation; The specific expression is as follows: in, For hyperparameters; A multi-similarity loss function is constructed to perform metric learning on pedestrian feature representations, thereby enhancing the discriminative ability of the features. The formula for the multi-similarity loss function is as follows: in, m It is the number of samples; S ij For anchor point samples i Compared with positive samples j Cosine similarity; S ik For anchor point samples i With negative samples k Cosine similarity; P i and N i Samples i The set of positive samples and the set of negative samples; α , β These are hyperparameters that control the weights of difficult samples; λ It is the similarity offset; A triplet loss function is constructed to perform metric learning on the pedestrian feature representation, further improving the discriminative ability of the features. The formula for the triplet loss function is as follows: in, a Anchor point sample; p To and a Positive samples of the same category; n To and a Negative samples of different categories; margin It is a preset boundary hyperparameter that controls the minimum interval between positive and negative sample pairs.

[0014] The beneficial effects of this invention are: This invention proposes a multi-branch collaborative occluded pedestrian re-identification method. Through the collaborative design of global complementary branches, dual-path local branches, and multi-scale channel branches, it effectively improves the model's recognition performance in complex occluded scenarios. The global complementary branch utilizes an explicit-implicit dual-path mining mechanism to enhance the modeling ability of overall pedestrian features, and can compensate for information loss caused by occlusion by supplementing and inferring global contextual information. The dual-path local branches capture local details of pedestrians through fine-grained feature extraction, further improving the model's ability to discriminate local features. The multi-scale channel branches, combined with multi-scale feature fusion and deformable convolution mechanisms, significantly enhance the model's adaptability to occluded scenarios, enabling flexible handling of features at different scales and effectively avoiding interference from occluded regions. Simultaneously, the fusion design of the multi-objective joint loss function achieves synergistic optimization of pedestrian feature discriminability and robustness, improving the model's discriminative ability and generalization. These designs work together to effectively solve the problems of low recognition accuracy and insufficient robustness of existing technologies in complex occluded scenarios, providing strong support for the promotion of pedestrian re-identification technology in practical applications. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the overall process of the pedestrian re-identification method under occlusion according to the present invention; Figure 2 These are schematic diagrams of conventional convolution and deformable convolution of the present invention; Figure 3 This is a schematic diagram of the global complementary branch of the present invention; Figure 4 This is a schematic diagram of the dual-path partial branch of the present invention; Figure 5 This is a schematic diagram of the multi-scale channel branching of the present invention; Figure 6 This is a schematic diagram of the multi-scale deformable convolution module of the present invention; Figure 7 This is a schematic diagram of the channel attention mechanism module of the present invention; Figure 8 This is an example image of the Occluded-Duke dataset used in the experiments of this invention; Figure 9 This describes the impact of different λ² values ​​on model performance during the experimental process of this invention. Figure 10 This is a schematic diagram of the pedestrian re-identification results of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figures 1 to 10 As shown, the occluded pedestrian re-identification method based on multi-branch collaboration specifically includes the following steps: Truncation is performed based on the OSNet backbone network; Input pedestrian images into a truncated OSNet backbone network to extract initial feature maps; A multi-branch feature extraction module is constructed, which includes parallel global complementary branches, dual-path local branches, and multi-scale channel branches; The initial feature map is input into the multi-branch feature extraction module; Explicit and implicit global features are extracted from the initial feature map using global complementary branches; Local features are extracted from the initial feature map through dual-path local branching, and the local features are recombined and spliced ​​to generate fine-grained discriminative features and high-dimensional discriminative features. Multi-scale channel features are extracted from the initial feature map through multi-scale channel branching, and the multi-scale channel features are enhanced based on multi-scale deformable convolution and channel attention mechanism. Based on the output of the multi-branch feature extraction module, features of the unit space and rank space are defined and optimized collaboratively. Based on the optimized rank space features, the final pedestrian feature representation is generated, and the re-identification result is output.

[0019] The initial feature map is expressed in the form of C×H×W with dimensions of 512×24×8, where C represents the number of channels in the feature map; H represents the height of the feature map; and W represents the width of the feature map. The OSNet backbone network consists of five convolutional blocks: conv1, conv2, conv3, conv4, and conv5. The OSNet backbone is truncated at conv3_0, dividing it into two parts: conv1, conv2, conv3_0 and conv3_x, conv4, conv5. Here, conv3_0 represents the first convolutional layer in the third convolutional block, and conv3_x represents the remaining convolutional layers in the third convolutional block. The pedestrian image is input into conv1, conv2, and conv3_0, and the initial feature map is output from conv3_0 after processing. The input of conv3_x receives the initial feature map output by conv3_0; The three branches of the multi-branch feature extraction module all follow the structure of conv3_x, conv4, and conv5. The initial feature map is extracted and refined sequentially by conv3_x, conv4, and conv5, and the first feature map with a dimension of 512×24×8 is output by conv5.

[0020] Among them, the global complementary branches include parallel explicit feature paths and implicit feature paths; The global complementary branch extracts explicit and implicit global features from the initial feature map, specifically including the following steps: The initial feature map is sequentially processed by conv3_x, conv4, and conv5 to extract and refine features, and conv5 outputs the first feature map with dimensions of 512×24×8. The first feature map is input into the explicit feature path and the implicit feature path respectively; In the explicit feature path, the first feature map is explicitly extracted by generalized mean pooling (GeM) to obtain an explicit global feature G with a dimension of 512×1×1. In the implicit feature path, a Top DropBlock strategy is used to generate a mask. The mask is used to mask the high-activation regions of the first feature map, outputting a low-response feature region with dimensions of 512×24×8. Then, implicit features in the low-response feature region are extracted by global max pooling (GMP), resulting in an implicit global feature G with dimensions of 512×1×1. drop ; Explicit global feature G and implicit global feature G drop Both are subjected to batch normalization. The explicit global feature G and the implicit global feature G after batch normalization are... drop They are respectively represented as and .

[0021] Generalized Mean Pooling (GeM) is a flexible feature pooling method that achieves adaptive interpolation between average pooling and max pooling by introducing a learnable exponent p. Its mathematical expression is: (1) In the formula, c This represents the first feature map. c One channel; (k , l ) represents the spatial dimension, where k This represents the vertical coordinate index of the current pixel. l Indicates the horizontal coordinate index of the current pixel; when When, GeM degenerates into average pooling; when At that time, GeM approximates max pooling; GeM can effectively improve the robustness of the model to scale changes by dynamically balancing global statistics and local extrema by adjusting the p value, and has a strong feature representation ability.

[0022] When in formula (1) In this case, GeM is equivalent to Global Average Pooling (GAP). GAP calculates the mean of the spatial dimensions of each channel of the input feature map and outputs a C × 1 × 1 vector (where C is the number of channels), ultimately achieving feature dimensionality reduction and global information extraction; given an input feature map... Its mathematical expression is: (2) in y c The output feature after global average pooling c The global average value of the channel; GAP can effectively suppress spatial noise interference by preserving the statistical characteristics of the channel dimension, but may weaken local saliency features, and is suitable for scenarios that require global context modeling; When in formula (1) In this case, GeM is equivalent to Global Max Pooling (GMP). GMP enhances local salient responses by extracting the maximum spatial dimension of each channel of the input feature map, and outputs a C × 1 × 1 vector. The mathematical expression is: (3) GMP is sensitive to outliers in feature maps and can highlight the most discriminative local features, but it may lose spatial distribution information due to overemphasis on extreme values. It is suitable for tasks that require enhanced feature sparsity and salient region detection.

[0023] Dual-path local branches include parallel local feature recombination paths and local feature splicing paths; Local features are extracted from the initial feature map through dual-path local branching, and these local features are then recombined and spliced ​​to generate fine-grained discriminative features and high-dimensional discriminative features. The specific steps include: The initial feature map is sequentially processed by conv3_x, conv4, and conv5 to extract and refine features, and conv5 outputs the first feature map with dimensions of 512×24×8. The first feature map with dimensions of 512×24×8 is input into the local branch of the dual path; The first feature map is divided into four first local feature regions arranged in sequence along the horizontal direction. The dimensions of the four first local feature regions are all 512×6×8. Input all four first local feature regions simultaneously into the local feature recombination path and the local feature splicing path; The local feature reconstruction path specifically includes the following steps: By splicing any two adjacent first local feature regions in the four first local feature regions arranged in sequence, three second local feature regions with overlapping parts are generated. The dimensions of the second local feature regions are all 512×12×8. The three second local feature regions are processed by global average pooling (GAP) to obtain three fine-grained features P1, P2 and P3, each with a dimension of 512×1×1. Batch normalization is performed on fine-grained features P1, P2, and P3 respectively to obtain batch-normalized fine-grained features. , and ; The local feature stitching path specifically includes the following steps: Global average pooling is performed on the four first local feature regions of the input to obtain four local feature vectors with dimensions of 512×1×1. The four local feature vectors are concatenated along the channel dimension to obtain a discriminative feature P4 with a dimension of 2048×1×1; High-dimensional discriminative features are obtained after batch normalization of P4. .

[0024] The multi-scale channel branch includes parallel multi-pooling paths and multi-scale deformable convolutional paths. The first feature map is input into the multi-pooling path and the multi-scale deformable convolutional path for processing to generate multi-scale channel features. The multi-pooling path generates diverse channel feature representations through various pooling methods, and then performs fusion and dimensionality reduction processing. Multi-scale deformable convolution path: Extracting and enhancing multi-scale features through multi-scale deformable convolution and channel attention mechanism; The multi-scale deformable convolution path includes a multi-scale deformable convolution module and a channel attention mechanism module.

[0025] The multi-pooling path specifically includes the following steps: The first feature map is processed by Global Average Pooling (GAP), Generalized Mean Pooling (GeM), and Global Max Pooling (GMP) to generate three different channel feature representations. The dimensions of the three different channel feature representations are all 512×1×1. The three different channel feature representations are concatenated along the channel dimension to form a first fusion feature representation with a dimension of 1536×1×1; The first fused feature representation is sampled by 1×1 convolution to reduce its dimensionality and compress it to generate a compact feature vector with a dimension of 512×1×1; Channel segmentation is performed on the compact feature vector, and a 1×1 convolution operation with shared parameters is performed to decouple the compact feature vector into two independent and complementary sub-features C1 and C2, each of which has a dimension of 256×1×1. Batch normalization is performed on the two sub-features C1 and C2 respectively, and the output is obtained. and .

[0026] Multi-scale deformable convolutional paths specifically include the following steps: The first feature map enters the multi-scale deformable convolution module; The multi-scale deformable convolution module uses three deformable convolution kernels with different receptive fields (3×3, 5×5, and 7×7) to perform convolution processing on the first feature map in order to capture feature information at different scales. Each deformable convolutional kernel dynamically adjusts the sampling point position to obtain the second feature maps at three scales, namely F3, F5 and F7, with each second feature map having a dimension of 512×24×8. The three second feature maps are concatenated along the channel dimension to form a second fused feature representation with dimensions of 1536×24×8; After performing channel compression and batch normalization on the second fused feature representation using 1×1 convolution, a third feature map with dimensions of 512×24×8 is obtained. The third feature map is input from the multi-scale deformable convolution module to the channel attention mechanism module; In the channel attention mechanism module, the height H and width W of the third feature map are compressed to a single real number using global average pooling (GAP), and the global statistic Z for each channel is obtained. C， The third feature map is transformed from C×H×W to C×1×1; Z C The expression is as follows: Then, two fully connected layers (the first and second fully connected layers) implemented with 1×1 convolutions are used to learn the non-linear relationships between channels and generate weight vectors. Sc , S c The specific expression is as follows: W1 represents the weight matrix of the first fully connected layer, used for dimensionality reduction; δ represents the ReLU activation function; W2 represents the weight matrix of the second fully connected layer, used for dimensionality increase; σ represents the Sigmoid activation function. The weight vector S c The weights are applied to the third feature map, and the output is the fourth feature map. ; The specific expression is as follows: Based on the output of the multi-branch feature extraction module, features in the unit space and rank space are defined and jointly optimized, specifically including the following steps: The characteristic representations of the unit space I and the rank space R can be formally defined as two sets of features, expressed as follows: Construct a label-smooth identity loss function L ID Multiple similarity loss function L MS and triplet loss function L Triplet Furthermore, a multi-objective joint loss function is constructed by comprehensively considering this. The expression for the multi-objective joint loss function is as follows: This is the balance factor for the identity loss function; It serves as a balance factor for the multiple similarity loss function and the triplet loss function; The feature set of a unit space I is smoothed only by the identity loss function of the label. L ID Perform optimized training; The feature set of rank space R is obtained through a multi-objective joint loss function. L total Optimize the training.

[0027] Construct a label-smooth identity loss function to supervise the classification of pedestrian feature representations; The label-smoothing identity loss function (ID Loss) smooths the true labels by introducing label smoothing regularization (LSR), preventing the model from overconfident in its predictions and thus improving the model's robustness and generalization ability. The formula for constructing the label-smoothing identity loss function is as follows: ,in in, u Indicates pedestrian category, u ∈﹛1,2,...,N﹜;N represents the set number of pedestrian images in the training set; y This is a real label; pu For the network's predicted logical value, pu The expression is as follows: in, x t Let t be the feature of the t-th sample; C The total number of label types; W v Features x t Does it belong to the category of tags? v The classifier consists of fully connected layers; Indicates the first t The classifier weight vector corresponding to the true label of each sample; Indicates the transpose operation; The specific expression is as follows: in, This is a hyperparameter, usually set to a small value to prevent the model from becoming overconfident in the training labels and reduce overfitting.

[0028] Construct a multi-similarity loss function to perform metric learning on pedestrian feature representations and enhance the discriminative ability of features; The Multiple Similarity Loss (MS Loss) function is a loss function used in Deep Metric Learning (DML). MS Loss dynamically adjusts the weights of hard samples while considering three similarity relationships between sample pairs (self-similarity, positive sample relative similarity, and negative sample relative similarity), efficiently learning discriminative features. The formula for the MS Loss function is as follows: in, m It is the number of samples; Sij For anchor point samples i Compared with positive samples j Cosine similarity; S ik For anchor point samples i With negative samples k cosine similarity; P i and N i Let be the positive sample set and negative sample set of sample i, respectively; α and β are hyperparameters that control the weight of hard samples; λ is the similarity offset; the multi-similarity loss considers both the similarity between the anchor sample and the positive sample and the similarity between the anchor sample and the negative sample; the more similar the anchor sample is to the positive sample, the smaller the loss, and vice versa.

[0029] We construct a triplet loss function to perform metric learning on pedestrian feature representations, thereby further improving the discriminative ability of the features; The goal of the triplet loss function is to ensure that the distance between the anchor point and the positive sample is smaller than the distance between the anchor point and the negative sample. This is usually controlled by a hyperparameter. The formula for the triplet loss function is as follows: in, a Anchor point sample; p To and a Positive samples of the same category; n To and a Negative samples of different categories; margin These are preset boundary hyperparameters that control the minimum interval between positive and negative sample pairs. The goal of the triplet loss function is to make... a and p distance d(a,p) As small as possible, a and n distance d (a,n) As large as possible; Based on the identity loss function, the multiple similarity loss function, and the triplet loss function, a multi-objective joint loss function is constructed. The expression of the multi-objective joint loss function is as follows: This is the balance factor for the identity loss function; It serves as a balance factor for the multiple similarity loss function and the triplet loss function; The pedestrian feature representation is optimized using a multi-objective joint loss function; The pedestrian re-identification result is output based on the optimized pedestrian feature representation.

[0030] To verify the effectiveness of the occluded pedestrian re-identification method based on multi-branch collaboration proposed in this invention, the occluded pedestrian re-identification method based on multi-branch collaboration of this invention is applied to a pedestrian re-identification network model; the acquired pedestrian image dataset is input into the pedestrian re-identification network model, and the pedestrian re-identification result is output. Systematic experiments were conducted on three benchmark datasets: Market-1501, DukeMTMC-ReID, and Occluded-Duke. Cumulative Matching Characteristic (CMC) curves and mean Average Precision (mAP) were used as evaluation metrics. The specific definitions are as follows: The CMC curve measures the probability that a pedestrian re-identification network model successfully matches the target pedestrian in the first k retrieval results, reflecting the recognition ability at different Rank-k positions. Among them, Rank-1 accuracy is particularly critical, representing the proportion of correct matches appearing for the first time at the top. mAP is the mean of the average precision (AP) for each query sample. It measures the overall retrieval performance of the model across all query samples by comprehensively considering retrieval precision and recall.

[0031] The specific steps for model training configuration in the experimental phase of the method of this invention are as follows: I. Model Initialization Baseline network: OSNet (an efficient person re-identification model) pre-trained on the ImageNet dataset is used as the infrastructure; II. Input Preprocessing and Data Augmentation Standardization: Each channel of the pedestrian image is processed to have zero mean and unit variance (to make the data distribution more stable); Size adjustment: Scaling to a fixed resolution of 384×128 (to meet model input requirements); Data augmentation: Data augmentation includes random scaling and horizontal flipping to improve the generalization ability of the pedestrian re-identification network model; Random scaling: First, enlarge the image width and height to 105% of the original size, then randomly crop it back to 384×128 to simulate shooting angles at different distances; Horizontal Flip: Mirrors the image with a 50% probability, increasing the variety of viewpoints; III. Training Strategies The training cycle is set to 160 epochs, with each epoch iterating through all training data. Batch settings: Each batch contains 48 pedestrian images, consisting of 8 different pedestrian identities (6 pedestrian images are randomly selected for each identity). Optimizer: The Adam optimizer is used with momentum factors of β1=0.9 and β2=0.999 (to control the momentum decay rate). Learning rate scheduling: Set the initial learning rate to 6×10. -4 ; First 10 epochs: learning rate from 6×10 -5 Linear growth to 6×10 -4 To avoid instability in the initial training; The next 150 epochs will gradually decay according to the cosine function. The training of the pedestrian re-identification network model is optimized and supervised based on a multi-objective joint loss function to improve feature discriminativeness and robustness. During the testing phase, the final representation is obtained by concatenating multi-branch features, and cosine distance is used for matching and discrimination. The experimental environment parameters are shown in Table 1: Table 1 Experimental Environment Parameters Analysis of the effectiveness of the loss function during the experiment A multi-objective joint loss function is used for training pedestrian re-identification network models. The multi-objective joint loss function integrates the label smoothing identity loss function, the multi-similarity loss function, and the triplet loss function to achieve synergistic optimization of pedestrian feature discriminability and robustness. To balance the contributions of different loss terms, hyperparameters λ1 and λ2 are introduced. λ1 controls the weight of classification and measurement loss, and is empirically set to 0.5. λ2 is used to adjust the influence weight between MS Loss and Triplet Loss, and is optimized through experiments. Table 2 shows the performance changes with different λ2 values ​​on the Market-150 and DukeMTMC-ReID datasets; When λ2 = 0, only ID Loss and Triplet Loss are introduced; When λ2 = 1, it is a combination of ID Loss and MS Loss; Figure 9 Line graphs showing the impact of different λ2 values ​​on model performance on the Market-1501 and DukeMTMC-ReID datasets are presented. Experimental results show that appropriately adjusting the weights of MS Loss and Triplet Loss significantly improves the model's discriminative ability, verifying the positive effect of the multi-loss synergy mechanism on model performance, and further confirming the complementarity of different loss functions in optimization objectives and learning preferences; the impact of λ2 on model performance is relatively mild, and when λ2≥0.9 or λ2≤0.3, the Rank-1 and mAP indices show slight fluctuations. Specifically, when λ² is set to 0.7, the model achieves the best overall performance on both datasets, as follows: Market-1501's Rank-1 and mAP were 96.6% and 91.7%, respectively. DukeMTMC-ReID achieved Rank-1 and mAP of 91.9% and 83.9%, respectively. Therefore, in all experiments, the value of λ2 was 0.7 in this invention; Table 2 shows the differences between the Market-1501 and DukeMTMC-ReID datasets. λ Experimental results of 2 Effectiveness analysis of the multi-branch feature extraction module during the experiment To fully verify the effectiveness of the multi-branch feature extraction module proposed in this invention, this study conducted systematic ablation experiments on the Market-1501 and DukeMTMC-ReID benchmark datasets. As shown in Table 3, the experimental setup included the following seven sets of comparison configurations: Baseline represents the base model, Global(G), Part(P), and Channel(C) represent independent configurations using only global complementary branches, dual-path local branches, and multi-scale channel branches, respectively; G+P, G+C, and P+C represent combinations of two branches, respectively; and G+P+C represents the complete multi-branch feature extraction module. Table 3 Experimental results of different network branches and their combinations on the Market-1501 and DukeMTMC-ReID datasets. Experimental results show that the proposed multi-branch feature extraction module has a significant performance improvement advantage in person re-identification tasks. Each branch module demonstrates effectiveness and complementarity in both individual and joint applications. Detailed analysis follows: (1) Single-branch performance evaluation On the Market-1501 dataset: After introducing the global complementary branch, the model's Rank-1 accuracy improved by 0.5%, and mAP improved by 4.3%. After introducing dual-path local branching, the model's Rank-1 accuracy improved by 1.1%, and mAP improved by 4.6%. After introducing the multi-scale channel branch, the model's Rank-1 accuracy improved by 0.9%, and its mAP improved by 4.8%. On the DukeMTMC-ReID dataset: After introducing the global complementary branch, the model's Rank-1 accuracy improved by 0.9%, and its mAP improved by 5.4%. After introducing dual-path local branches, the model's Rank-1 accuracy improved by 1.8%, and mAP improved by 6.5%. After introducing the multi-scale channel branch, the model's Rank-1 accuracy improved by 2.3%, and mAP improved by 7.2%. These improvements indicate that each branching structure independently enhances the model's discriminative power to varying degrees; specifically: Global complementary branches enhance the modeling ability of global context information, helping the model better understand the overall characteristics of pedestrians; Dual-path local branching helps the model better capture the local details of pedestrians through fine-grained local feature extraction; Multi-scale channel branch enhances the model's ability to perceive features at different scales through multi-scale feature fusion, thereby improving the overall performance of the model. (2) Improved performance of multi-branch collaboration When the three branches work together, the model achieves optimal performance on multiple benchmark datasets; On the Market-1501 dataset: The model achieved a Rank-1 accuracy of 96.6% and a mAP of 91.7%. Compared to the baseline model, Rank-1 improved by 1.8%, and mAP improved by 6.8%. On the DukeMTMC-ReID dataset: The model achieved a Rank-1 accuracy of 91.9% and a mAP of 83.9%. Compared to the baseline model, Rank-1 improved by 3.3%, and mAP improved by 10.4%. Experimental results show that the multi-branch feature learning framework has a significant feature complementarity effect, specifically reflected in the following aspects: Global complementary branch: By modeling spatial context, the model's ability to understand the overall semantics of pedestrians is enhanced; this means that the model can better grasp the overall characteristics of pedestrians, and can supplement and infer them through contextual information even when some features are occluded.

[0032] Dual-path local branching: Focusing on fine-grained regions enhances the ability to model local structural information; this allows the model to capture pedestrian local features (such as clothing texture, facial features, etc.) in greater detail, thereby enhancing the ability to discriminate at the local level.

[0033] Multi-scale channel branching: Through multi-scale feature fusion and deformable convolution mechanism, the robustness of the model to occluded scenes is significantly enhanced; this means that the model can better adapt to features of different scales and can flexibly handle occluded parts, avoiding interference from occluded areas on the recognition results. In summary, the proposed multi-branch collaborative structure not only performs excellently when used alone, but also further enhances model performance through feature-level complementarity when working collaboratively. Especially in complex occlusion scenarios, this multi-branch collaborative structure achieves superior recognition results, fully validating its effectiveness and advancement in pedestrian re-identification tasks.

[0034] Effectiveness analysis of occlusion problem during experiment Table 4 shows the performance of each branch structure and its combination on the Occluded-Duke dataset, further verifying the effectiveness and robustness of the proposed multi-branch structure in the occluded pedestrian re-identification scenario. Table 4 Experimental results of different network branches and their combinations on the Occluded-Duke dataset. Based on the single-branch experimental results, the multi-scale channel branch achieved 72.9% and 60.0% in Rank-1 and mAP metrics, respectively, outperforming the global complementary branch and the dual-path local branch. This result demonstrates that the channel branch designed to address the occlusion problem has a significant advantage in enhancing the model's discriminative ability.

[0035] Further multi-branch collaborative experiments show that the three branches exhibit good complementarity during joint training, significantly improving the overall performance of the model in complex occlusion scenarios. When the global complementary branch, the dual-path local branch, and the multi-scale channel branch are used simultaneously, the model achieves state-of-the-art performance on the Occluded-Duke dataset, with Rank-1 and mAP reaching 76.5% and 66.1%, respectively, representing improvements of 3.6% and 6.1% compared to the single channel branch. The experimental results fully demonstrate the significant advantages of the proposed multi-branch architecture under occlusion conditions, reflecting the effective synergy of local fine-grained modeling, global context supplementation, and multi-scale channel fusion, which is of certain significance for improving the robustness of occluded pedestrian recognition.

[0036] Comparative analysis with advanced methods during the experiment The performance of the proposed method in this invention is compared with that of various methods addressing occlusion problems in the re-identification task, and the results are shown in Table 5. It can be seen that the proposed method outperforms existing methods in both Rank-1 accuracy and mAP. Compared to the HCGA method, Rank-1 accuracy is improved by 6.4 percentage points and mAP by 8.6 percentage points; compared to the DIP method, improvements are made by 5.4 percentage points and 3.0 percentage points, respectively; and compared to the BPB method, improvements are made by 1.4 percentage points and 3.6 percentage points, respectively. These results demonstrate that the multi-branch network proposed in this invention exhibits stronger recognition ability and robustness in complex occlusion scenarios.

[0037] Table 5. Comparison results with state-of-the-art methods on the Occluded-Duke dataset. Furthermore, to comprehensively evaluate the performance of the multi-branch network proposed in this invention, it is divided into four categories: global feature methods, local feature methods, attention mechanism methods, and other mainstream methods. Comparative experiments are conducted on two mainstream person re-identification benchmark datasets, Market-1501 and DukeMTMC-ReID. Table 6 shows the performance comparison results of the proposed method and various advanced methods on the above datasets. Through multi-dimensional analysis, the adaptability and advantages of the proposed model under different feature modeling strategies are further verified. Table 6. Comparison results with state-of-the-art methods on the Market1501 and DukeMTMC-ReID datasets. As can be seen from Table 6, the model proposed in this invention exhibits excellent performance on both the Market-1501 and DukeMTMC-ReID public datasets. On the Market-1501 dataset, the proposed method achieves a Rank-1 accuracy of 96.6% and an mAP of 91.7%. On the DukeMTMC-ReID dataset, the Rank-1 and mAP are 91.9% and 83.9% respectively, outperforming most current mainstream methods in terms of overall performance; Compared with PLR-OSNet, which also has a multi-branch structure, the method of this invention enhances the expressive power of local features and introduces multi-scale channel branches to improve the richness of feature representation. On the Market-1501 dataset, it improves the Rank-1 by 1.0% and the mAP by 2.8%, respectively, and on the DukeMTMC-ReID dataset, it improves the mAP by 0.3% and 2.7%, respectively. The above results further verify the effectiveness and robustness of the method of this invention in cross-scene pedestrian re-identification tasks.

[0038] Visual analysis during the experiment To further demonstrate the robustness of the proposed multi-branch collaborative pedestrian re-identification network in occluded scenarios, this invention provides a visual analysis of the identification results under occlusion conditions. like Figure 10 As shown, the black rectangle represents the query image, and Top-10 represents the top ten candidate images with the highest similarity. The green box represents the correctly identified result, and the red box represents the incorrectly identified result. Figure 10 The results are visualized and compared under typical occlusion scenarios. The figures show that the baseline network, due to its single-scale feature extraction method, has significant limitations when facing occlusion interference. In the first set of test samples involving pedestrians occluding each other, the baseline method cannot effectively distinguish the fine-grained features between the target pedestrian and the background interfering pedestrians, leading to feature confusion and no correctly matching samples being returned in the Top-10 search results. In the second set of object occlusion cases, due to the lack of modeling ability for local occlusion, the baseline network failed to effectively capture key semantic information such as clothing texture, thus affecting the overall recognition accuracy.

[0039] In contrast, the multi-branch collaborative feature extraction architecture proposed in this invention, by introducing multi-scale semantic modeling and local structure perception modules, can improve the representation ability of occluded regions and alleviate recognition interference caused by occlusion; from Figure 10 The visualization results shown in the Ours section demonstrate that the method maintains high robustness and recognition accuracy even in complex occlusion scenarios, further validating its practicality and advancement in real-world applications.

[0040] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for re-identifying occluded pedestrians based on multi-branch collaboration, characterized in that, Specifically, the following steps are included: Truncation is performed based on the OSNet backbone network; Input pedestrian images into a truncated OSNet backbone network to extract initial feature maps; A multi-branch feature extraction module is constructed, which includes parallel global complementary branches, dual-path local branches, and multi-scale channel branches; The initial feature map is input into the multi-branch feature extraction module; Explicit and implicit global features are extracted from the initial feature map using global complementary branches; Local features are extracted from the initial feature map through dual-path local branching, and the local features are recombined and spliced ​​to generate fine-grained discriminative features and high-dimensional discriminative features. Multi-scale channel features are extracted from the initial feature map through multi-scale channel branching, and the multi-scale channel features are enhanced based on multi-scale deformable convolution and channel attention mechanism. Based on the output of the multi-branch feature extraction module, features of the unit space and rank space are defined and optimized collaboratively. Based on the optimized rank space features, the final pedestrian feature representation is generated, and the re-identification result is output. The initial feature map is expressed in the form of C×H×W with dimensions of 512×24×8, where C represents the number of channels in the feature map; H represents the height of the feature map; and W represents the width of the feature map. The OSNet backbone network consists of five convolutional blocks: conv1, conv2, conv3, conv4, and conv5. The OSNet backbone is truncated at conv3_0, dividing it into two parts: conv1, conv2, conv3_0 and conv3_x, conv4, conv5. Here, conv3_0 represents the first convolutional layer in the third convolutional block, and conv3_x represents the remaining convolutional layers in the third convolutional block. The pedestrian image is input into conv1, conv2, and conv3_0, and the initial feature map is output from conv3_0 after processing. The input of conv3_x receives the initial feature map output by conv3_0; The three branches of the multi-branch feature extraction module all continue the structure of conv3_x, conv4, and conv5; Among them, the global complementary branches include parallel explicit feature paths and implicit feature paths; The global complementary branch extracts explicit and implicit global features from the initial feature map, specifically including the following steps: The initial feature map is sequentially processed by conv3_x, conv4, and conv5 to extract and refine features, and conv5 outputs the first feature map with dimensions of 512×24×8. The first feature map is input into the explicit feature path and the implicit feature path respectively; In the explicit feature path, the first feature map is explicitly extracted by generalized mean pooling (GeM) to obtain an explicit global feature G with a dimension of 512×1×1. In the implicit feature path, a Top DropBlock strategy is used to generate a mask. The mask is used to mask the high-activation regions of the first feature map, outputting a low-response feature region with dimensions of 512×24×8. Then, implicit features in the low-response feature region are extracted by global max pooling (GMP), resulting in an implicit global feature G with dimensions of 512×1×1. drop ; Explicit global feature G and implicit global feature G drop Both are subjected to batch normalization. The explicit global feature G and the implicit global feature G after batch normalization are... drop They are respectively represented as and ; Based on the output of the multi-branch feature extraction module, features in the unit space and rank space are defined and jointly optimized, specifically including the following steps: The characteristic representations of unit space I and rank space R are formally defined as two sets of features; Construct a label-smooth identity loss function L ID Multiple similarity loss function L MS and triplet loss function L Triplet Furthermore, a multi-objective joint loss function is constructed by comprehensively considering this. The expression for the multi-objective joint loss function is as follows: This is the balance factor for the identity loss function; It serves as a balance factor for the multiple similarity loss function and the triplet loss function; The feature set of a unit space I is smoothed only by the identity loss function of the label. L ID Perform optimized training; The feature set of rank space R is obtained through a multi-objective joint loss function. L total Optimize the training.

2. The occluded pedestrian re-identification method based on multi-branch collaboration according to claim 1, characterized in that, Dual-path local branches include parallel local feature recombination paths and local feature splicing paths; Local features are extracted from the initial feature map through dual-path local branching, and these local features are then recombined and spliced ​​to generate fine-grained discriminative features and high-dimensional discriminative features. The specific steps include: The initial feature map is sequentially processed by conv3_x, conv4, and conv5 to extract and refine features, and conv5 outputs the first feature map with dimensions of 512×24×8. The first feature map with dimensions of 512×24×8 is input into the local branch of the dual path; The first feature map is divided into four first local feature regions arranged in sequence along the horizontal direction. The dimensions of the four first local feature regions are all 512×6×8. Input all four first local feature regions simultaneously into the local feature recombination path and the local feature splicing path; The local feature reconstruction path specifically includes the following steps: By splicing any two adjacent first local feature regions in the four first local feature regions arranged in sequence, three second local feature regions with overlapping parts are generated. The dimensions of the second local feature regions are all 512×12×8. The three second local feature regions are processed by global average pooling (GAP) to obtain three fine-grained features P1, P2 and P3, each with a dimension of 512×1×1. Batch normalization is performed on fine-grained features P1, P2, and P3 respectively to obtain batch-normalized fine-grained features. , and ; The local feature stitching path specifically includes the following steps: Global average pooling is performed on the four first local feature regions of the input to obtain four local feature vectors with dimensions of 512×1×1. The four local feature vectors are concatenated along the channel dimension to obtain a discriminative feature P4 with a dimension of 2048×1×1; High-dimensional discriminative features are obtained after batch normalization of P4. .

3. The occluded pedestrian re-identification method based on multi-branch collaboration according to claim 2, characterized in that, The multi-scale channel branch includes parallel multi-pooling paths and multi-scale deformable convolutional paths. The first feature map is input into the multi-pooling path and the multi-scale deformable convolutional path for processing to generate multi-scale channel features. The multi-pooling path generates diverse channel feature representations through various pooling methods, and then performs fusion and dimensionality reduction processing. Multi-scale deformable convolution path: Extracting and enhancing multi-scale features through multi-scale deformable convolution and channel attention mechanism; The multi-scale deformable convolution path includes a multi-scale deformable convolution module and a channel attention mechanism module.

4. The occluded pedestrian re-identification method based on multi-branch collaboration according to claim 3, characterized in that, The multi-pooling path specifically includes the following steps: The first feature map is processed by Global Average Pooling (GAP), Generalized Mean Pooling (GeM), and Global Max Pooling (GMP) to generate three different channel feature representations. The dimensions of the three different channel feature representations are all 512×1×1. The three different channel feature representations are concatenated along the channel dimension to form a first fusion feature representation with a dimension of 1536×1×1; The first fused feature representation is sampled by 1×1 convolution to reduce its dimensionality and compress it to generate a compact feature vector with a dimension of 512×1×1; Channel segmentation is performed on the compact feature vector, and a 1×1 convolution operation with shared parameters is performed to decouple the compact feature vector into two independent and complementary sub-features C1 and C2, each of which has a dimension of 256×1×1. Batch normalization is performed on the two sub-features C1 and C2 respectively, and the output is obtained. and .

5. The occluded pedestrian re-identification method based on multi-branch collaboration according to claim 4, characterized in that, Multi-scale deformable convolutional paths specifically include the following steps: The first feature map enters the multi-scale deformable convolution module; The multi-scale deformable convolution module uses three deformable convolution kernels with different receptive fields (3×3, 5×5, and 7×7) to perform convolution processing on the first feature map in order to capture feature information at different scales. Each deformable convolutional kernel dynamically adjusts the sampling point position to obtain the second feature maps at three scales, namely F3, F5 and F7, with each second feature map having a dimension of 512×24×8. The three second feature maps are concatenated along the channel dimension to form a second fused feature representation with dimensions of 1536×24×8; After performing channel compression and batch normalization on the second fused feature representation using 1×1 convolution, a third feature map with dimensions of 512×24×8 is obtained. The third feature map is input from the multi-scale deformable convolution module to the channel attention mechanism module; In the channel attention mechanism module, the height H and width W of the third feature map are compressed to a single real number using global average pooling (GAP), and the global statistic Z for each channel is obtained. C The third feature map is transformed from C×H×W to C×1×1; Z c The expression is as follows: In the formula, c This represents the first feature map. c One channel; ( k , l ) represents the spatial dimension, where k This represents the vertical coordinate index of the current pixel. l Indicates the horizontal coordinate index of the current pixel; Then, two fully connected layers implemented with 1×1 convolutions are used to learn the nonlinear relationship between channels, generating a weight vector S. c ; The weight vector S c The weights are applied to the third feature map, and the output is the fourth feature map. .

6. The occluded pedestrian re-identification method based on multi-branch collaboration according to claim 5, characterized in that, The unit space I and the rank space R are respectively represented as: 。 7. The occluded pedestrian re-identification method based on multi-branch collaboration according to claim 6, characterized in that, A label-smooth identity loss function is constructed to supervise the classification of pedestrian feature representations. The formula for constructing the label-smooth identity loss function is as follows: ,in in, u Indicates pedestrian category, u ∈﹛1,2,...,N﹜;N represents the set number of pedestrian images in the training set; y This is a real label; For hyperparameters, For the network's predicted logical value, The expression is as follows: in, x t Let t be the feature of the t-th sample; C The total number of label types; W v Features x t Does it belong to the category of tags? v The classifier consists of fully connected layers; Indicates the first t The classifier weight vector corresponding to the true label of each sample; Indicates the transpose operation; A multi-similarity loss function is constructed to perform metric learning on pedestrian feature representations, thereby enhancing the discriminative ability of the features. The formula for the multi-similarity loss function is as follows: in, m It is the number of samples; S ij For anchor point samples i Compared with positive samples j Cosine similarity; S ik For anchor point samples i With negative samples k Cosine similarity; P i and N i These represent the positive sample group and the negative sample group, respectively. α , β These are hyperparameters that control the weights of difficult samples; λ It is the similarity offset; A triplet loss function is constructed to perform metric learning on the pedestrian feature representation, further improving the discriminative ability of the features. The formula for the triplet loss function is as follows: in, This is the anchor point sample in the current triplet; p To and Positive samples of the same category; n To and Negative samples of different categories; margin It is a preset boundary hyperparameter that controls the minimum interval between positive and negative sample pairs.

Images