Reloading pedestrian re-identification method based on decoupling network

Abstract

The invention provides a reloading pedestrian re-identification method based on a decoupling network. The method comprises the following steps: acquiring appearance information and semantic body type information of a pedestrian to be detected; inputting appearance information and body type information of a pedestrian to be detected into the constructed re-identification model to obtain final features; the re-recognition model comprises a multi-scale information fusion decoupling module, a median driving channel attention and human body shape enhancement module and a clothes mixing module. According to the method, the identity-related features and the clothes features are effectively decoupled, the characterization of the identity-related features is enhanced, and the negative influence of the clothes change on the model performance is reduced by mixing the clothes features.

Description

Technical Field

[0001] The present invention relates to the technical field of pedestrian re-identification, and specifically provides a method for re-identifying pedestrians after changing clothes based on a decoupling network. Background Art

[0002] Person re-identification (PIR) is a key task in computer vision, aiming to identify and track the same target individual across different camera viewpoints. This technology has broad applications in intelligent surveillance, public safety, and traffic management. While existing PIR methods have made significant progress in short-term scenarios, most assume clothing consistency and rely primarily on visual features such as color and texture. However, in real-world applications, individuals often change clothing, resulting in a significant performance degradation for traditional PIR systems. In short-term PIR tasks, clothing texture features can effectively identify the target individual. However, in long-term PIR scenarios, models that rely solely on clothing features can mistakenly identify some images as highly similar. When identity-related features are used for recognition, the model accurately identifies the images within the red box as highly similar, demonstrating that identity-related features are more robust to clothing changes.

[0003] Studies have shown that dressing scenes significantly increase intra-class differences and weaken inter-class discrimination because the identity cues provided by clothing become unreliable. In this case, re-identification relies more on intrinsic biometric features such as facial features, body shape, and hairstyle. To address these challenges, existing methods have begun to fuse multimodal inputs—including facial features, bone structure, gait patterns, silhouettes, and 3D body shapes—to enhance identity recognition. For example, representation learning is improved by fusing appearance features with body shape information. However, such models still have two major limitations: (1) it is difficult to capture long-range structural dependencies and has poor generalization ability for unseen clothing styles; (2) the problem of feature entanglement persists, and clothing-related features are coupled with identity features, resulting in limited generalization performance of the model under large-scale clothing changes.

[0004] Therefore, a new method for pedestrian re-identification after changing clothes based on a decoupled network is urgently needed to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned drawbacks, the present invention is proposed to provide a solution or partial solution to the above-mentioned problems.

[0006] The present invention provides a method for re-identifying pedestrians after changing clothes based on a decoupling network, comprising: obtaining appearance information and semantic body shape information of the pedestrian to be tested; inputting the appearance information and semantic body shape information of the pedestrian to be tested into a constructed re-identification model to obtain final features; the re-identification model includes a multi-scale information fusion decoupling module, a median-driven channel attention module, a human body shape enhancement module, and a clothing mixing module.

[0007] In a technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupling network, the appearance information and semantic body shape information of the pedestrian to be tested are input into the constructed re-identification model, and the process of obtaining the final features includes: introducing a median-driven channel attention mechanism into the multi-scale information fusion decoupling module, and optimizing human body features by fusing the semantic body shape information with the appearance information based on the multi-scale information fusion decoupling module to obtain identity-related body shape features and clothing features; inputting the clothing features into the clothing mixing module to obtain features that simulate clothing changes; inputting the identity-related body shape features into the human shape enhancement module, and perturbing the clothing features based on similarity; and performing feature splicing or cross-attention fusion on the mixed clothing features and the enhanced body shape features to form a joint feature.

[0008] In a technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupling network, the multi-scale information fusion and decoupling module includes: a cross-scale information fusion branch, a sensitive information aggregation branch, and an adaptive global enhancement branch; the multi-scale information fusion and decoupling module fuses the semantic body shape information with the appearance information to optimize human body features, and the process of decoupling identity-related features from clothing features includes: the cross-scale information fusion branch uses multi-scale convolution to construct a cross-scale receptive field to obtain identity clues of different spatial granularities, and at the same time introduces a median-driven channel attention mechanism to weighted optimize the fusion features; the sensitive information aggregation branch simulates the attention mechanism of the human visual system to salient areas through multi-scale pooling operations to obtain key features aggregated in different receptive fields; the adaptive global enhancement branch adaptively reweights and aggregates the features output by the cross-scale information fusion branch and the sensitive information aggregation branch.

[0009] In one technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupled network, the process of inputting the clothing features into the clothing mixing module to obtain features that simulate clothing changes includes: randomly shuffling the input clothing features along the batch dimension to generate mixed objects; calculating the similarity weights of the clothing features; randomly sampling the mixing coefficients, and then using the original clothing feature structure as a guide to adjust the mixing coefficients through the similarity weights; and linearly mixing the original clothing features and the shuffled clothing features based on the adjusted weights to generate features that simulate clothing changes.

[0010] In one technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupled network, the identity-related body shape features are input into the human body shape enhancement module, and the clothing features are perturbed based on similarity, specifically including: aggregating features along the vertical and horizontal directions of the human body, converting pixel-level features into semantic features that conform to the geometric structure of the human body; performing lightweight convolution processing after grouping by anatomical regions; and focusing on key structural areas through an attention mechanism to generate enhanced body shape features.

[0011] In one technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupled network, the process of aggregating features along the vertical and horizontal directions of the human body and converting pixel-level features into semantic features that conform to the geometric structure of the human body includes: dividing the decoupled human body features into several groups, each group containing multiple channels; for each group of feature maps, performing global average pooling and global maximum pooling operations along the height and width directions respectively to obtain pooled features.

[0012] In a technical solution of the above-mentioned decoupled network-based pedestrian re-identification method, the median-driven channel attention includes parallel multi-pooling layers: global maximum pooling, global median pooling, global average pooling, shared convolutional layers, attention weighting and output modules.

[0013] In a technical solution of the above-mentioned decoupled network-based pedestrian re-identification method, the sensitive information aggregation branch simulates the attention mechanism of the human visual system to salient areas through multi-scale pooling operations to obtain key features of aggregation of different receptive fields, specifically including: performing maximum pooling and average pooling operations at different scales on the input features to obtain local and global pooling feature combinations, and then upsampling to a uniform size and summing them to achieve multi-scale feature fusion; introducing a median-driven channel attention mechanism, generating channel statistics through global average pooling, maximum pooling and median pooling, obtaining channel attention weights through shared network processing, weighted optimization of the fused features, and outputting the final fused features.

[0014] In one technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupled network, after the mixed clothing features and the enhanced body features are subjected to feature splicing or cross-attention fusion to form a joint feature, it also includes: training the re-identification model by minimizing the total loss function, wherein the minimized total loss function is the sum of the classification loss of the fused features, the clothing classification loss, the body feature classification loss, and the minimized mutual information loss.

[0015] In a technical solution of the above-mentioned method for re-identifying pedestrians after changing clothes based on a decoupled network, the process of obtaining semantic body shape information of the pedestrian to be tested includes: converting a human body image into a posture heat map; and processing the posture heat map to obtain semantic body shape information.

[0016] The beneficial effects of the decoupled network-based method for re-identifying people after changing clothes provided by the present invention are as follows: the method effectively decouples identity-related features from clothing features, enhances the representation of identity-related features, and reduces the negative impact of clothing changes on model performance by mixing clothing features. Specifically, the multi-scale information fusion decoupling module is combined with the median-driven channel attention mechanism to effectively suppress the noise generated during the fusion of the two modalities and decouple identity-related body shape features from identity-irrelevant clothing features. Thereafter, the human shape enhancement module further optimizes the representation of body shape features. At the same time, the clothing mixing module perturbs clothing features based on similarity, thereby reducing the impact of clothing changes on model recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The disclosure of the present invention will be more easily understood with reference to the accompanying drawings. Those skilled in the art will readily appreciate that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Furthermore, similar numbers in the drawings represent similar components, wherein:

[0018] Figure 1 1 is a flow chart of the main steps of a method for re-identifying pedestrians after changing clothes based on a decoupled network according to an embodiment of the present invention.

[0019] Figure 2 2. It is a schematic diagram of the structure of a multi-scale information fusion decoupling module according to an embodiment of the present invention;

[0020] Figure 3 2 is a schematic diagram of the structure of median-driven channel attention according to one embodiment of the present invention;

[0021] Figure 4 is a structural schematic diagram of a clothing mixing module according to an embodiment of the present invention;

[0022] Figure 5 FIG. 4 is a schematic structural diagram of a human body shape enhancement module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] Some embodiments of the present invention are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] like Figure 1 As shown, a method for re-identifying pedestrians after changing clothes based on a decoupled network in an embodiment of the present invention mainly includes the following steps S1 to S2.

[0025] Step S1: Obtain appearance information and semantic body shape information of the pedestrian to be measured;

[0026] In this embodiment, appearance information refers to information related to surface features and external forms that is directly extracted from the visual presentation of a target (such as a pedestrian), specifically including: basic visual features: pixel-level or local area features such as color (such as a red top and blue pants), texture (such as the plaid pattern of clothes, the texture of leather), shape (such as the loose / slim style of clothes), and gloss (such as the reflectivity of leather shoes); component-level semantic features: through semantic segmentation or target detection, the appearance is disassembled into identifiable components (such as tops, pants, shoes, accessories, etc.), and the attributes of each component are described (such as "white baseball cap" and "black backpack"); overall matching features: the combination relationship of each component (such as the color combination of "red top + blue pants", the matching style of "suit + tie"), and the overall outline (such as long / short, loose / tight clothing).

[0027] Semantic body shape information refers to semantically descriptive body shape features extracted from the body structure of a target (such as a pedestrian). Specifically, it includes: geometric physical features: limb proportions (such as leg length / height, shoulder width / hip width) calculated through the coordinates of human joints, the relative position relationship between the trunk and limbs (such as the ratio of upper and lower body), and the dimensions of various body dimensions (such as height and quantitative values ​​of body fatness); semantic abstract labels: mapping geometric features into human-understandable semantic descriptions, such as "tall and thin", "strong", "short and fat", "slender lower limbs", etc., or expressing them through numerical ranges (such as "height over 1.8m", "BMI>25"); structural invariant features: independent of external clothing, determined only by the inherent structure of the human skeleton, muscles, etc., and stable across appearance changes (such as body shape characteristics remain unchanged after changing clothes).

[0028] The process of obtaining the appearance and semantic body shape information of the pedestrian under test includes: inputting the pedestrian RGB image or video frame into the convolutional neural network to extract basic visual features, such as color (RGB three-channel values), texture (clothing pattern), edge (style outline), etc.

[0029] In this example, ResNet-50 and SSE modules are used as the backbone network to extract appearance information from pedestrian images and encode semantic body shape information from the human pose heat map generated by the off-the-shelf estimator.

[0030] Step S2: Input the appearance information and semantic body shape information of the pedestrian to be tested into the constructed re-identification model to obtain the final features; Figure 1 As shown in the figure, the re-identification model includes a multi-scale information fusion decoupling module, a median-driven channel attention module, a human shape enhancement module, and a clothing mixing module.

[0031] In one embodiment, Figure 1As shown, step S2, inputting the appearance information and semantic body shape information of the pedestrian to be detected into the constructed re-identification model, and obtaining the final features includes:

[0032] Step S21: The median-driven channel attention mechanism is introduced into the multi-scale information fusion and decoupling module. The semantic body shape information and the appearance information are fused based on the multi-scale information fusion and decoupling module to optimize human features and obtain identity-related body shape features and clothing features.

[0033] Step S22: inputting the clothing features into the clothing mixing module to obtain features simulating clothing changes;

[0034] Step S23: inputting the identity-related body shape features into the human body shape enhancement module, and perturbing clothing features based on similarity;

[0035] Step S24: perform feature splicing or cross-attention fusion on the mixed clothing features and the enhanced body features to form a joint feature.

[0036] In one embodiment, Figure 2 As shown, the multi-scale information fusion decoupling module includes: a cross-scale information fusion branch, a sensitive information aggregation branch, and an adaptive global enhancement branch; this module optimizes human body features by fusing semantic body shape information with appearance data, effectively realizing the decoupling of identity-related features and clothing features.

[0037] Step S21, the multi-scale information fusion and decoupling module fuses the semantic body shape information with the appearance information to optimize human body features, and the process of decoupling identity-related features from clothing features includes:

[0038] Step S211: The cross-scale information fusion branch uses multi-scale convolution to construct a cross-scale receptive field to obtain identity clues of different spatial granularities, and at the same time introduces a median-driven channel attention mechanism to perform weighted optimization on the fusion features;

[0039] In this embodiment, the multi-scale convolution layer contains three types of convolution kernels: 3×3, 5×5, and 7×7. The multi-scale pooling layer integrates 2×2 and 4×4 maximum pooling and 2×2 average pooling operations. DCN stands for deformable convolution. The two feature maps extracted from ResNet-50 and SSE are spliced ​​along the channel dimension to form the input feature map F. c In order to capture both fine-grained details (such as texture and pattern) and coarse-grained structural features (such as body shape) under clothing variation conditions, the cross-scale information fusion (CSIF) branch uses multi-scale convolution to construct a cross-scale receptive field. This design enables the model to fuse identity cues of different spatial granularities. Specifically, for a given input feature F c, using three convolution kernels of 3×3, 5×5 and 7×7 for feature extraction, the calculation process is expressed as:

[0040] F a =Conv 3×3 (F c , W θ )+Conv 3×3 (F c , W ζ ),

[0041] F b =Conv5 ×5 (F c , W ζ )+Conv 7×7 (F c , W φ ),

[0042] Among them, F a represents the features obtained by 3×3 and 5×5 convolution operations, F b represents the features obtained by 5×5 and 7×7 convolution operations, W θ 、W ζ and W φ are the learnable convolution parameters.

[0043] At the same time, the median-driven channel attention mechanism is introduced to perform weighted optimization on the fusion features, which can enhance the expression of key information while suppressing redundant noise. The final output feature F of this branch is CSIF Expressed as:

[0044] F CSIF =MDCA(F a )·F a +MDCA(F b )·F b ,

[0045] where median-driven channel attention (·) represents the process of applying median-driven channel attention.

[0046] Step S212: The sensitive information aggregation branch simulates the attention mechanism of the human visual system to salient areas through a multi-scale pooling operation to obtain key features aggregated in different receptive fields;

[0047] In this example, a Sensitive Information Aggregation (SIA) branch is designed to enhance the model's focus on key areas and suppress irrelevant details. This branch simulates the human visual system's attention mechanism for salient areas through multi-scale pooling operations, aggregating key features from different receptive fields.

[0048] Specifically, for a given input feature F c, a multi-scale pooling operation is used to extract key sensitive features from different regions and generate smooth global features at the same time. The pooling results after upsampling are summed to achieve multi-scale feature fusion. The calculation process is expressed as:

[0049] F pool1 =Up(MaxPppl 2×2 (F c ))+Up(AvgPopl 2×2 (F c )),

[0050] F pool2 =U p (AvgPool 2×2 (F c ))+Up(MaxPool 4×4 (F c )),

[0051] Among them, F pool1 and F pool2 represents the combination of the maximum pooling and average pooling results after upsampling at different scales. Up(·) represents the upsampling function. A median-driven channel attention mechanism is also introduced to further optimize feature representation and suppress redundant information. The final fused feature FSIA is used as the output of this branch. The process can be expressed as:

[0052] F SIA =MDFCA(F pool1 )·F pool1 +MDCA(F pool2 )·F pool2 ,

[0053] where MDCA(·) denotes the process of applying median-driven channel attention.

[0054] Step S213: The adaptive global enhancement branch adaptively reweights and aggregates the features output by the cross-scale information fusion branch and the sensitive information aggregation branch.

[0055] In this embodiment, in order to effectively integrate the cross-scale features of the first two branches and enhance the global representation, an adaptive global enhancement (AGE) branch is introduced. Given that the CSIF and SIA branches focus on different complementary aspects of the representation, AGE adaptively reweights and aggregates these features, highlighting key information while suppressing redundant interference. Specifically, the output features of the first two branches are added to the original input features Fc to obtain the feature f. Then, deformable convolution (DCN) is applied to introduce an additional offset to dynamically adjust the feature f to obtain the feature

[0056]

[0057] Where DCN(·, θ) represents the deformable convolution operation and θ is the process parameter. Finally, the fused features are optimized through the median-driven channel attention mechanism. The process can be expressed as:

[0058]

[0059] Finally, the fusion feature F AGE Divided into identity-related body shape features F along the channel dimension Shape and clothing characteristics F Cloth .

[0060] To ensure effective separation of features, mutual information is used to quantify their correlation. Specifically, the MI loss L MI The mutual information neural estimator (MINE) is combined with Monte Carlo integration, and its expression is as follows:

[0061]

[0062] where F′ Shape represents the identity-related body shape features after random rearrangement, T(·,·,θ) is a fully connected neural network parameterized by θ, which is responsible for extracting the paired features ( FShape , F Cloth ) as a joint distribution, and from the unpaired features (F′ Shape , F Cloth ) is sampled as the marginal distribution. Ultimately, the correlation between the two is reduced by minimizing the mutual information loss, achieving the goal of feature decoupling.

[0063] In one embodiment, step S22, inputting the clothing features into the clothing mixing module to obtain features for simulating clothing changes, includes:

[0064] Step S221: randomly shuffle the input clothing features along the batch dimension to generate mixed objects;

[0065] Step S222: Calculate the similarity weight of clothing features;

[0066] Step S223: randomly sample the mixing coefficients, and then use the original clothing feature structure as a guide to adjust the mixing coefficients through similarity weights;

[0067] Step S224: linearly mix the original clothing features and the disrupted clothing features based on the adjusted weights to generate features that simulate clothing changes.

[0068] In this embodiment, Figure 4As shown in Figure 2, a clothing blending module is introduced to mitigate the model's over-reliance on clothing appearance and encourage the learning of more identity-distinguishing feature representations in clothing variations. Its core idea is to simulate clothing variations by blending clothing features from samples with similar appearances during training, thereby breaking the correlation between identity and clothing and enhancing the model's robustness to clothing appearance variations.

[0069] Specifically, first randomly shuffle the input clothing features F along the batch dimension cloth , to generate the hybrid object F rand Then, the similarity between samples is measured by calculating the cosine similarity and linearly mapping it to the interval [0,1]. The calculation method can be expressed as:

[0070]

[0071] At the same time, following the original Mixup method, the mixing coefficient α is randomly sampled from the Beta distribution with a hyperparameter λ. In order to ensure that the mixed features remain closer to the original features, the structure of the original clothing features is used as a guide for generation, and the mixing coefficient λ is set * :

[0072] λ=Beta(α, α), λ * =Max(λ, 1-λ),

[0073] Finally, the weight of the mixed feature is adjusted by the similarity weight of the clothing feature to ensure that the final mixed result F mix . Not only randomness is taken into account, but also the similarity between samples. This hybrid strategy can be expressed as:

[0074] F mix =λ * ·F cloth +(1-λ * )F rand ·sim,

[0075] This strategy reduces the model's dependence on clothing changes and helps it gradually adapt to the transition between different clothing characteristics.

[0076] In one embodiment, step S23, inputting the identity-related body shape features into the human body shape enhancement module, and perturbing clothing features based on similarity specifically includes:

[0077] Step S231: Aggregate features along the vertical and horizontal directions of the human body to convert pixel-level features into semantic features that conform to the geometric structure of the human body;

[0078] Step S232: performing lightweight convolution processing after grouping by anatomical region;

[0079] Step S233: Focus on key structural areas through the attention mechanism to generate enhanced body shape features.

[0080] In this embodiment, Figure 5 As shown in Figure 2, considering that human body shape provides stable and clothing-independent clues, a human body shape enhancement module is introduced. This module aggregates features along the two spatial directions of the human body shape and groups the features using a body region-based partitioning strategy, enabling the model to better focus on subtle changes in shape structure while reducing the computational burden. Specifically, the decoupled human body feature F Shape Divide into G groups (in training, G is set to 8), each group contains C / G channels. Then, for each group of feature maps, perform global average pooling and global maximum pooling operations along the height and width directions respectively. Get the pooled feature It is then fed into a shared convolutional layer consisting of two 1×1 convolutional layers, a batch normalization layer, and a ReLU activation function:

[0081]

[0082] Here, Conv1(·,ζ) denotes a 1×1 convolution operation, where ζ is a learnable convolution parameter, and BN(·) denotes batch normalization.

[0083] Subsequently, the outputs of the shared convolutional layers along the horizontal and vertical directions are added together and the attention weights are generated through the Sigmoid activation function. This calculation can be expressed as:

[0084]

[0085] Here, σ(·) represents the Sigmoid activation function.

[0086] Finally, we get the weight W h and W w Used to perform weighted operations on the pedestrian shape feature map to generate an enhanced shape feature map F enhance The above process can be expressed as:

[0087] F enhance =F shape W h W w .

[0088] In one embodiment, Figure 3 As shown in

[15] , median-driven channel attention combines median pooling with multiple pooling operations, where A represents the attention map generated by summing multiple outputs of the shared convolutional layer.

[0089] In order to suppress the fusion noise from heterogeneous modalities (such as RGB images and pose heatmaps), a median-driven channel attention mechanism is proposed. Unlike traditional methods (such as relying only on average pooling or maximum pooling), median-driven channel attention introduces median pooling as a more statistically robust supplement. By integrating these three pooling strategies in a shared attention network, median-driven channel attention achieves a more balanced and noise-resistant channel importance estimate, thereby improving the quality of feature representation. Figure 4 As shown. For the input feature F input , global average pooling, global maximum pooling and global median pooling are first applied to all spatial locations of each channel, which can be expressed as:

[0090] F GAP =GAP(F input ), F CMP =GMP(F input ),

[0091] F GMedP =GMedP(F input ),

[0092] Among them, F GAP 、F GMP and F GMedP Represent the features extracted by global average pooling, maximum pooling, and median pooling, respectively. Next, a shared convolutional layer is used to process the pooling results. This convolutional layer consists of two 1×1 convolutions, a ReLU activation function, and a Sigmoid activation function. The processed results are added element-wise to obtain the final channel attention map. The above process can be expressed as:

[0093]

[0094] Among them, i∈{GAP, GMP, GMedP}, σ(·) represents the Sigmoid activation function, ReLU(·) represents the ReLU activation function, Conv1(·,ζ) represents the 1×1 convolution operation, and ζ is a learnable convolution parameter. Finally, the channel attention map A is combined with the original input feature map F input Multiply to get the weighted feature map F output :

[0095] F output =A·F input .

[0096] In one embodiment, after the mixed clothing features and the enhanced body features are subjected to feature splicing or cross-attention fusion to form a joint feature, the method further includes: training a re-identification model by minimizing a total loss function, wherein the minimized total loss function is the sum of the classification loss of the fused features, the clothing classification loss, the body feature classification loss, and the minimized mutual information loss.

[0097] In this embodiment, in order to ensure the clear decoupling of identity-related features and clothing features while maximizing their discrimination, the mutual information loss L is minimized. MI To achieve this goal. For the decoupled clothing feature F Cloth and human body shape characteristics F Shape , respectively apply classification loss L Cloth and L Shape To promote the learning of each feature. Then, the two features are spliced ​​together through interaction to obtain a more robust and discriminative feature F, and the classification loss L is applied ID

[0098] The total loss of the model can be expressed as:

[0099] L all =L ID +L Cloth +L Shape +λL MI

[0100] Among them, L Shape , L Cloth and L ID Denotes the cross entropy loss for classification, LMI denotes minimizing mutual information loss, and λ is the weight coefficient of the loss term. Considering that deep features are more relevant to the task, λ is empirically set to 0.1 in all experiments.

[0101] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the original technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for pedestrian re-identification after changing clothes based on a decoupled network, characterized by: include: Obtain the appearance information and semantic body shape information of the pedestrian to be tested; Input the appearance information and semantic body shape information of the pedestrian to be tested into the constructed re-identification model to obtain the final features; The re-identification model includes a multi-scale information fusion decoupling module, a median-driven channel attention module, a human shape enhancement module, and a clothing mixing module.

2. The method according to claim 1, characterized in that The appearance information and semantic body shape information of the pedestrian to be tested are input into the constructed re-ID model. The process of obtaining the final features includes: The multi-scale information fusion and decoupling module introduces a median-driven channel attention mechanism, which optimizes human features by fusing the semantic body shape information with the appearance information to obtain identity-related body shape features and clothing features. Inputting the clothing features into the clothing mixing module to obtain features that simulate clothing changes; Inputting the identity-related body shape features into the human body shape enhancement module, and perturbing clothing features based on similarity; The mixed clothing features and the enhanced body features are subjected to feature splicing or cross-attention fusion to form joint features.

3. The method according to claim 2, characterized in that The multi-scale information fusion decoupling module includes: a cross-scale information fusion branch, a sensitive information aggregation branch, and an adaptive global enhancement branch; The multi-scale information fusion and decoupling module fuses the semantic body shape information with the appearance information to optimize human features and achieves decoupling of identity-related features from clothing features. The process includes: The cross-scale information fusion branch uses multi-scale convolution to construct a cross-scale receptive field to obtain identity clues of different spatial granularities, and introduces a median-driven channel attention mechanism to perform weighted optimization on the fusion features. The sensitive information aggregation branch simulates the attention mechanism of the human visual system to salient areas through multi-scale pooling operations to obtain key features aggregated in different receptive fields; The adaptive global enhancement branch adaptively reweights and aggregates the features output by the cross-scale information fusion branch and the sensitive information aggregation branch.

4. The method according to claim 2, characterized in that The process of inputting the clothing features into the clothing mixing module to obtain features simulating clothing changes includes: Randomly shuffle the input clothing features along the batch dimension to generate mixed objects; Calculate the similarity weights of clothing features; Randomly sample the mixing coefficients, then use the original clothing feature structure as a guide to adjust the mixing coefficients through similarity weights; The original clothing features and the shuffled clothing features are linearly mixed based on the adjusted weights to generate features that simulate clothing changes.

5. The method according to claim 2, characterized in that Inputting the identity-related body shape features into the human body shape enhancement module, and perturbing the clothing features based on similarity specifically includes: Aggregate features along the vertical and horizontal directions of the human body to convert pixel-level features into semantic features that conform to the geometric structure of the human body; Lightweight convolution processing is performed after grouping by anatomical region; By focusing on key structural areas through the attention mechanism, enhanced body shape features are generated.

6. The method according to claim 5, characterized in that The process of aggregating features along the vertical and horizontal directions of the human body and converting pixel-level features into semantic features that conform to the geometric structure of the human body includes: Divide the decoupled human features into several groups, each group contains multiple channels; For each set of feature maps, global average pooling and global maximum pooling operations are performed along the height and width directions respectively to obtain pooled features.

7. The method according to claim 2, characterized in that The median-driven channel attention includes parallel multi-pooling layers: global maximum pooling, global median pooling, global average pooling, shared convolutional layers, attention weighting and output modules.

8. The method according to claim 3, characterized in that The sensitive information aggregation branch simulates the attention mechanism of the human visual system on salient areas through multi-scale pooling operations, and obtains the key features of different receptive fields, including: Perform maximum pooling and average pooling operations on the input features at different scales to obtain local and global pooling feature combinations, which are then upsampled to a uniform size and summed to achieve multi-scale feature fusion; A median-driven channel attention mechanism is introduced to generate channel statistics through global average pooling, maximum pooling and median pooling. The channel attention weights are obtained through shared network processing, and the fusion features are weighted and optimized to output the final fusion features.

9. The method according to claim 8, characterized in that After the mixed clothing features and the enhanced body features are subjected to feature splicing or cross-attention fusion to form a joint feature, the method further includes: training a re-identification model by minimizing a total loss function, wherein the minimized total loss function is the sum of the classification loss of the fused features, the clothing classification loss, the body feature classification loss, and the minimized mutual information loss.

10. The method according to claim 1, characterized in that The process of obtaining the semantic body shape information of the pedestrian to be measured includes: Convert human body images into posture heatmaps; The posture heat map is processed to obtain semantic body shape information.

Images