A method, system and electronic device for visible-infrared pedestrian re-identification

Abstract

The application discloses a method and system for visible light-infrared pedestrian re-identification and electronic equipment, which decouples the training of the modal sharing branch and the modal specific branch: in the modal sharing branch, a structure prior driven adaptive dual-path enhancement module SPADE is introduced, and a strip region modeling and dynamic weight enhancement cross-modal consistent discrimination region are utilized; in the modal specific branch, a hierarchical attention collaborative network (HAC-DBN) based on multi-statistic quantity perception and asymmetric space enhancement is introduced, so as to realize dynamic fusion of shallow details and deep semantics; meanwhile, an identity level statistical distribution distillation loss (DU-ID-DistilLoss) based on dynamic uncertainty weighting is proposed, a learnable uncertainty parameter is introduced to adaptively adjust the alignment weight of mean, standard deviation and quartile range, so as to explicitly reduce the identity distribution difference between modes. The application can effectively alleviate the influence of cross-modal difference on the ranking performance, and maintain good robustness in a complex scene.

Description

Technical Field

[0001] This invention belongs to the technical field of machine learning, specifically relating to a method, system, and electronic device for pedestrian re-identification in visible light and infrared. Background Technology

[0002] With the rapid development of intelligent monitoring, public safety and other fields, pedestrian re-identification technology, as a core technology for achieving cross-scenario identity matching, has received widespread attention. Among them, visible light-infrared (VIS / IR) cross-modal pedestrian re-identification (VI-ReID) has irreplaceable application value in complex monitoring scenarios such as low light, nighttime or day-night switching, because it can overcome the limitations of lighting conditions. Examples include nighttime security patrols, cross-day-night park access control, and personnel trajectory tracking in tunnels. Its core objective is to achieve accurate matching and retrieval of the same pedestrian identity between visible light and infrared modalities, which have significantly different imaging principles.

[0003] Existing technologies for cross-modal person re-identification primarily address the problem by learning modality-invariant representations to reduce modal differences between visible light and infrared images, thereby improving cross-modal retrieval performance. However, because visible light images focus on capturing detailed information such as texture and color of pedestrians, while infrared images focus on reflecting the thermal radiation contour features of pedestrians, there are fundamental differences between the two in terms of data distribution and visual representation. Existing methods still face the following key technical bottlenecks:

[0004] 1. Insufficient Modeling Utilization of Fine-Grained Structure and Higher-Order Statistical Information: Most existing methods over-rely on standard symmetric convolutional kernels or single global average pooling for cross-modal alignment, failing to fully exploit discriminative cues in local human body regions (such as the head, torso, and limbs). In real-world scenarios, pedestrians often face issues such as occlusion, missing limbs, and pose variations. Furthermore, visual confusion easily arises between pedestrians with similar appearances, and relying solely on global features is insufficient to distinguish these fine-grained differences, leading to a high mismatch rate. In addition, traditional channel attention mechanisms often ignore second-order statistical information about feature distributions, resulting in the model's inability to extract sufficiently discriminative cues when faced with difficult-to-distinguish samples with similar means but significantly different texture distributions.

[0005] 2. Insufficient cross-layer information synergy: In the feature extraction process of deep neural networks, shallow features are rich in detailed information such as texture and edges, but have poor robustness; deep features have more stable semantic information, but lose a lot of key details. Existing methods lack an adaptive coordination mechanism when fusing shallow and deep features. As the network depth increases, shallow detailed information is easily attenuated or interfered with by redundant information, and deep semantic features lack discriminative power due to the lack of detailed support, making it difficult to achieve complementary enhancement between the two.

[0006] 3. Insufficient identity-level distribution alignment: Existing technologies mostly use sample-level alignment or classification loss for model optimization, which can only constrain the representational similarity of a single sample in different modalities, and cannot explicitly suppress the statistical distribution drift of the same identity in visible light and infrared modalities. This difference in distribution level will cause the feature embeddings of the same identity to be scattered and the feature embeddings of different identities to overlap when the model performs cross-modal retrieval, thereby reducing the stability and reliability of retrieval results, especially in complex scenarios with significant modal shifts, where the performance degradation is more obvious.

[0007] Therefore, how to construct a technical solution that can accurately capture fine-grained discrimination regions, achieve adaptive fusion of cross-level information, and reduce modal differences from the perspective of identity statistical distribution has become the key to improving the performance of visible light-infrared cross-modal pedestrian re-identification, and is also a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a visible light-infrared pedestrian re-identification method, system, and electronic device that combines structural prior guidance, hierarchical attention collaboration, and identity statistical distribution constraints for joint optimization, so as to improve the discriminativeness, robustness, and stability of cross-modal retrieval.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention provides a method for pedestrian re-identification based on visible light and infrared light, comprising the following steps:

[0011] In a first aspect, the present invention provides a method for pedestrian re-identification based on visible light-infrared light, comprising the following steps:

[0012] S1. Acquire visible light images and infrared images, and construct training batches after preprocessing the visible light images and infrared images. The preprocessing includes size normalization and data augmentation. The training batches adopt an identity equalization sampling strategy and include both visible light and infrared samples.

[0013] S2. Input the preprocessed visible light image and infrared image into the dual-branch backbone network respectively, and extract multi-layer features that contain at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-sharing branch and a modality-specific branch.

[0014] S3. Input the multi-layer features into the modality-sharing branch and the modality-specific branch respectively for representation learning, wherein the modality-sharing branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues;

[0015] S4. Perform structure prior-driven adaptive dual-path enhancement processing in the modality sharing branch to generate a spatial attention map and perform weighted residual enhancement on deep features to obtain modality sharing enhanced features.

[0016] S5. Perform hierarchical attention collaboration and dynamic bridging processing in the modality-specific branch, refine the extraction of modality-specific cues through multi-statistic residual channel calibration and asymmetric multi-scale spatial enhancement, and achieve cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features.

[0017] S6. By constructing a multi-task joint objective function and combining iterative training with implicit knowledge extraction strategies, the dual-branch backbone network is trained iteratively. The implicit knowledge extraction strategies include Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. These strategies effectively transfer rich information from modality-specific enhancement features to modality-shared branches, and a joint loss function guides network learning. The joint loss function includes:

[0018] Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved.

[0019] Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity.

[0020] Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is enhanced.

[0021] S7. In the inference stage, the cross-modal shared enhanced features are input into the trained dual-branch backbone network to calculate the similarity between the query sample and the candidate sample, and pedestrian retrieval and ranking output are completed. The similarity measure adopts cosine similarity or inner product similarity.

[0022] As a preferred technical solution, in step S1, the data augmentation includes one or more of random horizontal flipping, random erasure, random cropping, random rotation, and random channel perturbation; the identity equalization sampling strategy samples P identities for each modality and K images for each identity, where P and K are preset positive integers.

[0023] As a preferred technical solution, in step S2, the dual-branch backbone network is a convolutional neural network, a Transformer network, or a combination of both, and adopts a feature extraction structure based on ResNet-50.

[0024] As a preferred technical solution, step S4, the structure-prior-driven adaptive dual-path enhancement processing specifically includes:

[0025] S41. Divide the mid-level features into a preset number S strip regions along the height direction, and apply global average pooling to each strip region to capture strip-level priors.

[0026] S42. For each strip region, pooling is performed using saliency-driven pooling and statistically invariant pooling branches to obtain strip-level saliency features and strip-level statistical prior features, and then the strip-level saliency features and strip-level statistical prior features are concatenated and fused; the saliency-driven pooling branch is an adaptive max pooling branch, and the statistically invariant pooling branch is a global average pooling branch for strip sub-blocks;

[0027] S43. For each strip, a mapping network with non-shared parameters is used to generate the original weights of the strip. The mapping network includes a feature extraction head and a weight generation head. Softmax normalization is performed on the original weights of the strip so that the sum of the weights of each strip is 1.

[0028] S44. Construct a strip mask based on the normalized strip weights, and generate a spatial attention map through convolutional transformation and nonlinear activation.

[0029] S45. Perform element-wise multiplication between the spatial attention map and the deep features, and then add the residuals to the original deep features to output cross-modal shared enhanced features.

[0030] As a preferred technical solution, the hierarchical attention collaboration in step S5 involves embedding a bimodal adaptive attention module after the bottleneck layer of the bimodal backbone network. The processing of the bimodal adaptive attention module includes:

[0031] a) Channel calibration: Perform global average pooling and global standard deviation pooling on the input features simultaneously to obtain first-order and second-order statistical features; concatenate the first-order and second-order statistical features and input them into a residual multilayer perceptron containing a dimensionality reduction layer, a nonlinear activation layer and an dimensionality increase layer to generate channel attention weights.

[0032] b) Spatial Augmentation: The features after channel calibration are reduced in dimension and compressed. Then, spatial features are extracted through two parallel paths: the first path uses consecutive convolutional layers with kernels of 1×7 and 7×1 to perform horizontal-vertical feature decomposition, and the second path uses 3×3 dilated convolutions with an inflation rate of 3 to expand the receptive field. Finally, the outputs of the two paths are fused with element-wise maximum values ​​to generate spatial attention weights.

[0033] c) Dynamic fusion: The modality discrimination coefficient is calculated based on the global mean of channel weights and spatial weights, and the channel enhancement features and spatial enhancement features are weighted and fused using this coefficient.

[0034] As a preferred technical solution, the dynamic bridging in step S5 specifically includes:

[0035] a) By using a cross-level adaptive bridging module, deep features are projected onto shallow channel dimensions through 1×1 convolution;

[0036] b) Adjust the shallow features to the same spatial resolution as the deep features using bilinear interpolation;

[0037] c) The projected deep features and the interpolated shallow features are spliced ​​together, and a fusion mask is generated based on the spliced ​​features through deformable convolution. The fusion mask is a continuous value between 0 and 1, which is used to perform convex combination weighted fusion of shallow features and deep features.

[0038] d) Perform multi-head self-attention operation on the weighted fused features to obtain cross-layer bridging features, which are output as modality-specific enhancement features.

[0039] As a preferred technical solution, in step S6, the identity-level statistical distribution distillation loss based on dynamic uncertainty weighting specifically includes:

[0040] a) Calculate statistical descriptor vectors for visible light features and infrared features for the same identity, wherein the statistical descriptors include a mean vector, a standard deviation vector, and an interquartile range vector;

[0041] b) Calculate the cross-modal Pearson correlation coefficient for each statistical descriptor vector, and construct the basic consistency loss term accordingly;

[0042] c) Introduce an independent learnable uncertainty parameter for each type of statistical descriptor, and construct the following dynamically weighted loss function:

[0043]

[0044] in, Indicates the statistical descriptor type, Based on the consistency loss term, The constraint weights of different statistics are adjusted synchronously with network training.

[0045] As a preferred technical solution, step S6, which iteratively trains the dual-branch backbone network, further includes the following steps:

[0046] S601, Forward Propagation and Feature Calculation: Input the training batch data into the dual-branch backbone network, extract modality-shared enhancement features through the SPADE module, extract modality-specific enhancement features through the HAC-DBN module, and calculate the output tensors of each level and classification.

[0047] S602. Joint Loss Collaborative Calculation: Based on the forward propagation results, in addition to the basic pedestrian re-identification loss, relevant losses are calculated simultaneously. and conventional auxiliary losses , And assign preset weights to each loss term to construct a joint objective function;

[0048] S603, Parameter Adaptive Update: Update the backbone network weights using the backpropagation algorithm and update them synchronously. Learnable uncertainty parameters in As training iterates, the uncertainty parameters automatically converge based on the alignment difficulty, thereby adaptively adjusting the distillation intensity until the network converges and outputs the trained model.

[0049] Secondly, the present invention provides a system for visible light-infrared pedestrian re-identification, which is applied to the aforementioned method for visible light-infrared pedestrian re-identification, including an image acquisition module, an image preprocessing module, a representation learning module, a first processing module, a second processing module, a training module, and an inference module;

[0050] The image acquisition module is used to acquire visible light images and infrared images, and to construct training batches after preprocessing the visible light images and infrared images. The preprocessing includes size normalization and data augmentation. The training batches adopt an identity equalization sampling strategy and include both visible light and infrared samples.

[0051] The image preprocessing module is used to input the preprocessed visible light image and infrared image into a dual-branch backbone network respectively, and extract multi-layer features that contain at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-sharing branch and a modality-specific branch.

[0052] The representation learning module is used to input the multi-layer features into the modality-shared branch and the modality-specific branch respectively for representation learning, wherein the modality-shared branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues;

[0053] The first processing module is used to perform structure prior-driven adaptive dual-path enhancement processing in the modality sharing branch, generate a spatial attention map and perform weighted residual enhancement on deep features to obtain modality sharing enhanced features;

[0054] The second processing module is used to perform hierarchical attention collaboration and dynamic bridging processing in the modality-specific branch, refine the extraction of modality-specific cues through multi-statistic residual channel calibration and asymmetric multi-scale spatial enhancement, and achieve cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features.

[0055] The training module is used to iteratively train the dual-branch backbone network by constructing a multi-task joint objective function and combining it with an implicit knowledge extraction strategy. The implicit knowledge extraction strategy includes Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. Through these strategies, rich information from modality-specific enhancement features is effectively transferred to the modality-shared branch, and a joint loss function guides the network learning. The joint loss function includes:

[0056] Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved.

[0057] Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity.

[0058] Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is enhanced.

[0059] The reasoning module is used to input the cross-modal shared enhanced features into the trained dual-branch backbone network during the reasoning stage, calculate the similarity between the query sample and the candidate sample, and complete the pedestrian retrieval and ranking output. The similarity measure adopts cosine similarity or inner product similarity.

[0060] Thirdly, the present invention provides an electronic device, the electronic device comprising:

[0061] At least one processor; and,

[0062] A memory communicatively connected to the at least one processor; wherein,

[0063] The memory stores computer program instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for visible-infrared pedestrian re-identification.

[0064] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0065] Existing technologies rely excessively on global features for cross-modal alignment, making it difficult to handle issues such as pedestrian occlusion, missing limbs, pose changes, and confusion between pedestrians with similar appearances, resulting in a high mismatch rate. This patent innovatively introduces a structure prior-driven adaptive dual-path enhancement module (SPADE) into the modal sharing branch. By modeling striped regions, it divides the human body into local regions such as the head, torso, and limbs, accurately capturing fine-grained discrimination cues. Simultaneously, it combines dynamic weight modulation to generate a spatial attention map, strengthening the consistent discrimination regions across modalities. This effectively reduces the risk of mismatches caused by factors such as partial occlusion and missing limbs, significantly improving the model's robustness to local changes.

[0066] This patent designs a Hierarchical Attention Collaboration and Dynamic Bridging Network (HAC-DBN) in a modality-specific branch: Addressing the shortcomings of existing technologies that only use a single statistic (such as the mean) for channel weighting, ignoring the discreteness information of feature distribution, this invention introduces Global Standard Deviation Pooling (GSP) to supplement second-order statistical information in the modality-specific branch, and preserves the original distribution characteristics through residual structures, significantly enhancing the model's ability to judge identity stability under infrared modalities and avoiding misjudgments caused by changes in thermal radiation. Furthermore, it designs an asymmetric spatial convolution path, creatively utilizing different convolution kernel sizes to capture the longitudinal structural features of pedestrians (such as human posture and limb proportions), and combines dilated convolution to capture large-scale contour information. This asymmetric design effectively solves the problem of limited receptive field for a single convolution kernel in variable monitoring scenarios, improving robustness to occlusion and posture changes.

[0067] This invention innovatively proposes a dynamically uncertain weighted identity-level statistical distribution distillation loss (DU-ID-DistilLoss). This method selects the mean, standard deviation, and interquartile range as key statistical descriptors for the feature distribution of the same identity across different modalities, and constructs cross-modal consistency constraints using the Pearson correlation coefficient. Based on this, a learnable uncertainty parameter is introduced to adaptively and dynamically weight the distillation loss for different statistics. This mechanism effectively suppresses the interference of statistical distortion (especially second-order statistics) caused by thermal noise on optimization, explicitly eliminating cross-modal differences at the identity statistical distribution level. The final results show that the features of the same identity are more compactly embedded in the feature space, the feature boundaries of different identities are clearer, and the stability and robustness of cross-modal retrieval in complex scenarios are significantly improved.

[0068] This patent adopts an architecture that decouples training between modality-shared branches and modality-specific branches. The modality-shared branch focuses on learning a universal discriminative representation that is consistent across modalities, providing basic support for identity matching between different modalities. The modality-specific branch retains modality-sensitive supplementary discriminative cues and makes full use of the unique feature information of the two modalities. The two work together to achieve comprehensive representation learning of "general features + specific features", avoiding the loss of feature information caused by training with a single branch, and enabling the model to maintain good generalization ability in complex scenarios. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a schematic diagram of the overall architecture of HA-SPGDL in this embodiment.

[0071] Figure 2 This embodiment shows a schematic diagram of the Dual Path Heterogeneous Graph Pooling (DHGP) module.

[0072] Figure 3 This embodiment shows a schematic diagram of the Cross-Level Adaptive Bridging (CLAB) module.

[0073] Figure 4 Schematic diagram of the dual-modal spatial channel attention module in this implementation.

[0074] Figure 5 This embodiment presents a schematic diagram of the principle of identity-level statistical distribution distillation loss (ID-DistilLoss).

[0075] Figure 6This embodiment presents a visual representation of the search results.

[0076] Figure 7 This is a schematic diagram of the visible light-infrared pedestrian re-identification system in this embodiment;

[0077] Figure 8 This is a schematic diagram of the electronic device in this embodiment. Detailed Implementation

[0078] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0079] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0080] like Figure 1 As shown, this embodiment provides a method for pedestrian re-identification based on visible light and infrared light, which includes the following steps:

[0081] S1. Acquire visible light and infrared images. After preprocessing the visible light and infrared images, construct a training set. Let the visible light and infrared training sets be respectively... and .

[0082] Furthermore, the input image is sized and normalized (e.g., to 288×144), and data augmentation is performed; the data augmentation may include one or more of random horizontal flipping, random erasing, random cropping, random rotation, and random channel perturbation.

[0083] When constructing training batches, it is preferable to use identity balanced sampling: each modality samples P identities, and each identity samples K images, where P and K are preset positive integers, so that a training batch contains both visible light and infrared samples.

[0084] S2. Input the preprocessed visible light image and infrared image into Block 1-2 (shared convolutional layer) of the dual-branch backbone network Res-Net50 respectively, and extract multi-layer features that contain at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-shared branch and a modality-specific branch.

[0085] Furthermore, the dual-branch backbone network can be a convolutional neural network, a Transformer network, or a combination of both; preferably, a feature extraction structure based on ResNet-50 is adopted, and the output is a multi-layer feature map containing at least shallow detail features and deep semantic features, as shown in a specific example.

[0086] S3. Input the multi-layer features into the modality-sharing branch and the modality-specific branch respectively for representation learning: the modality-sharing branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues.

[0087] S4. Perform structure prior-driven adaptive dual-path enhancement processing (SPADE) in the modality sharing branch to generate a spatial attention map and perform weighted residual enhancement on deep features to obtain cross-modality shared enhanced features.

[0088] Furthermore, such as Figure 2 As shown, step S4 specifically includes:

[0089] (1) The statistical invariant pooling branch will output the mid-level features of the mid-level features. The region is divided into stripes with a preset number S along the height direction, and then global average pooling is applied to each stripe to obtain... This allows for the capture of strip-level statistical priors, enhancing robustness to local variations.

[0090] (2) The saliency-driven pooling branch will first output the mid-level features. Max pooling and compression are performed to obtain a feature. To preserve the strongest responses of each channel, the signals are then rearranged into [B,S,C] to obtain the strip-level saliency features. And by combining the two, we obtain .

[0091] (3) Subsequently, dynamic weight modeling is performed on each strip through a mapping network with non-shared parameters. An independent dual-head MLP network is used to first pass the fused features through the feature head. After mapping to discriminative local features, a head is generated through weighting. Obtain weights And normalize, preferably using the softmax function to make the sum of the weights of each strip equal to 1;

[0092] (4) Based on strip weights Construct a weighted mask Spatial attention maps are generated by combining convolution operations and BN normalization. .

[0093] (5) Utilize the spatial attention map to analyze deep features Perform element-wise multiplication and weighting, and fuse it with the original deep feature residuals to output cross-modal shared enhanced features. .

[0094] S5. In the modality-specific branch, perform hierarchical attention collaboration and dynamic bridging processing (HAC-DBN) to calibrate the attention of modality-specific cues and achieve cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features.

[0095] Furthermore, such as Figure 3 and Figure 4 As shown, step S5 specifically involves:

[0096] (1) Hierarchical attention coordination:

[0097] A dual-modal adaptive channel-spatial attention module (DM-CSAM) is embedded after each residual block (Bottleneck) of Block 3 and Block 4 in the dual-branch backbone network to collaboratively optimize spatial resolution and semantic depth: given input features .

[0098] Multi-statistic residual channel attention: First- and second-order statistical information of features is extracted simultaneously through global average pooling (GAP) and global standard deviation pooling (GSP). The two sets of statistical vectors are concatenated and input into a multilayer perceptron consisting of a residual bottleneck structure. This structure includes a dimensionality reduction layer, a feature recalibration nonlinear activation layer, and a dimensionality increase layer, and retains the original statistics through skip connections.

[0099] Asymmetric Parallel Spatial Attention: 1×1 channel compression is performed on the channel-calibrated features to generate spatial thumbnails. Two sets of parallel spatial convolutional paths are introduced. One is a horizontal-vertical decomposition path using consecutive convolutional layers with kernel sizes of 1×7 and 7×1. The other uses 3×3 dilated convolutions with a dilation rate of d=3 to expand the spatial receptive field and capture pedestrian contours. Element-wise max fusion is performed on the results of the two paths to generate the final spatial attention map. The final attention features are fused using a dynamic gating mechanism, utilizing channel weights. and A modality discrimination coefficient α is calculated from the global mean. The channel weights and spatial weights are then multiplied by the original feature elements and then multiplied by the α coefficient to obtain... .

[0100] (2) Dynamic bridging (cross-level adaptive bridging): deep features are obtained by channel projection. The shallow features are adjusted by bilinear interpolation to obtain The two are fused, and a fusion mask is generated through deformable convolution to perform weighted fusion of shallow details and deep semantics; in terms of fusion features Multi-head self-attention is performed, where each attention head is represented as follows: To obtain cross-layer bridging characterization and output mode-specific enhancement features. .

[0101] S6. By constructing a multi-task joint objective function and combining iterative training with implicit knowledge extraction strategies, the dual-branch backbone network is trained iteratively. The implicit knowledge extraction strategies include Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. These strategies effectively transfer rich information from modality-specific enhancement features to modality-shared branches, and a joint loss function guides network learning. The joint loss function includes:

[0102] Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved.

[0103] Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity.

[0104] Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is improved.

[0105] Furthermore, such as Figure 5 As shown, step S6 specifically involves:

[0106] Calculate the statistical descriptor vector for the same identity in both modalities. and The statistical descriptor includes at least a mean vector. Standard deviation vector With the interquartile range vector The Pearson correlation coefficient was used to calculate the identity relevance scores of the three types of statistics across modalities. Subsequently, the basic distillation loss term for each type of statistic is defined. (in To address the statistical distortion caused by infrared modal thermal noise, this embodiment introduces a learnable uncertainty parameter for each type of statistical descriptor. The aforementioned The parameters are updated synchronously with the backbone network during network training, and their physical meaning represents the model's "degree of distrust" of the consistency constraints of this type of statistics. This leads to the construction of a dynamically weighted identity-level distribution distillation loss. , which is represented as When a certain statistic (such as standard deviation) exhibits significant instability in cross-modal alignment (i.e., When the value is large, the model will increase This reduces the penalty weight of the term. In some implementations, cross-modal collaborative constraint loss and hierarchical consistency loss can also be superimposed to align key semantic local regions and enhance the semantic coherence between shallow and deep features.

[0107] Furthermore, step S6, which iteratively trains the dual-branch backbone network, also includes the following steps:

[0108] S601, Forward Propagation and Feature Calculation: Input the training batch data into the dual-branch backbone network, extract modality-shared enhancement features through the SPADE module, extract modality-specific enhancement features through the HAC-DBN module, and calculate the output tensors of each level and classification.

[0109] S602. Joint Loss Collaborative Calculation: Based on the forward propagation results, in addition to the basic pedestrian re-identification loss, relevant losses are calculated simultaneously. and conventional auxiliary losses , And assign preset weights to each loss term to construct a joint objective function;

[0110] S603, Parameter Adaptive Update: Update the backbone network weights using the backpropagation algorithm and update them synchronously. Learnable uncertainty parameters in As training iterates, the uncertainty parameters automatically converge based on the alignment difficulty, thereby adaptively adjusting the distillation intensity until the network converges and outputs the trained model.

[0111] S7. In the reasoning stage, the similarity between the query sample and the candidate sample is calculated based on the cross-modal shared enhanced features to complete the pedestrian retrieval and sorting output. The similarity measure adopts cosine similarity or inner product similarity.

[0112] like Figure 6 As shown, under different retrieval modes and cross-modal directions in the public dataset, this method can output more stable ranking results. However, in complex situations such as occlusion, low-quality imaging, or incomplete target parts, a small number of mismatches may still occur. For example, when the candidate image contains multi-target interference, it will affect the prior knowledge of the local structure.

[0113] In some implementations, the method of the present invention can be implemented based on a deep learning training framework and trained and inferred on a computing platform equipped with a graphics processing unit (GPU). Preferably, the feature extraction network is initialized with network parameters pre-trained on a large-scale dataset, for example, using a residual network pre-trained on the ImageNet dataset as the backbone network, and on this basis, the modality sharing enhancement and modality-specific enhancement structures of the present invention are introduced.

[0114] During the data preprocessing stage, the input image can be uniformly adjusted to a preset resolution, such as 288×144, and one or more data augmentation strategies can be introduced to improve generalization ability. The data augmentation strategies include random horizontal flipping, random erasing, random cropping, random rotation, and random channel perturbation.

[0115] During training, the learning rate scheduling can employ a warm-up mechanism combined with a segmented decay strategy. Preferably, the learning rate is linearly increased from 0 to the initial learning rate within the first few preset training cycles to achieve stable convergence; then, the initial learning rate is kept unchanged before a preset first adjustment point, the learning rate is scaled to 0.1 times the initial learning rate between the first and second adjustment points, and further scaled to 0.01 times the initial learning rate after the second adjustment point.

[0116] In terms of batch construction, a balanced sampling strategy is preferred: multiple identities are randomly sampled for each training batch, and a preset number of visible light images and infrared images are sampled for each identity, so that a batch contains samples of both modalities at the same time; for example, 12 identities can be randomly sampled, and each identity contains 5 visible light images and 5 infrared images.

[0117] Regarding the setting of loss weights, the weight coefficients of each loss term in the joint loss function can be preset according to the task requirements; in one example, the weight coefficients can be a set of preset constants (for example, γ1, γ2 and γ3 are 0.4, 0.3 and 5 respectively), but are not limited to these values.

[0118] During the testing phase, it is preferable to use only modality-shared enhanced features (or their pooling vectors) to calculate the similarity between query samples and candidate samples and complete the retrieval ranking in order to evaluate cross-modal retrieval performance.

[0119] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously.

[0120] Based on the same idea as the visible-infrared pedestrian re-identification method in the above embodiments, the present invention also provides a system for visible-infrared pedestrian re-identification, which can be used to execute the above-described visible-infrared pedestrian re-identification method. For ease of explanation, the structural schematic diagram of the system embodiment for visible-infrared pedestrian re-identification only shows the parts relevant to the embodiments of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0121] like Figure 7 As shown, in another embodiment of this application, a system 100 for visible light-infrared pedestrian re-identification is provided. The system includes an image acquisition module 101, an image preprocessing module 102, a representation learning module 103, a first processing module 104, a second processing module 105, a training module 106, and an inference module 107.

[0122] The image acquisition module 101 is used to acquire visible light images and infrared images, and to construct training batches after preprocessing the visible light images and infrared images. The preprocessing includes size normalization and data augmentation. The training batches adopt an identity equalization sampling strategy and include both visible light and infrared samples.

[0123] The image preprocessing module 102 is used to input the preprocessed visible light image and infrared image into a dual-branch backbone network respectively, and extract multi-layer features that include at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-sharing branch and a modality-specific branch.

[0124] The representation learning module 103 is used to input the multi-layer features into the modality-sharing branch and the modality-specific branch respectively for representation learning, wherein the modality-sharing branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues;

[0125] The first processing module 104 is used to perform structure prior-driven adaptive dual-path enhancement processing in the modality sharing branch, generate a spatial attention map and perform weighted residual enhancement on deep features to obtain modality sharing enhanced features;

[0126] The second processing module 105 is used to perform hierarchical attention collaboration and dynamic bridging processing in the modality-specific branch, refine the extraction of modality-specific cues through multi-statistic residual channel calibration and asymmetric multi-scale spatial enhancement, and realize cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features.

[0127] The training module 106 is used to iteratively train the dual-branch backbone network by constructing a multi-task joint objective function and combining it with an implicit knowledge extraction strategy. The implicit knowledge extraction strategy includes Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. Through these strategies, rich information from modality-specific enhancement features is effectively transferred to the modality-shared branch, and a joint loss function guides the network learning. The joint loss function includes:

[0128] Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved.

[0129] Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity.

[0130] Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is enhanced.

[0131] The reasoning module 107 is used to input the cross-modal shared enhanced features into the trained dual-branch backbone network during the reasoning stage, calculate the similarity between the query sample and the candidate sample, and complete the pedestrian retrieval and ranking output. The similarity measure adopts cosine similarity or inner product similarity.

[0132] It should be noted that the visible light-infrared pedestrian re-identification system of the present invention corresponds one-to-one with the visible light-infrared pedestrian re-identification method of the present invention. The technical features and beneficial effects described in the embodiments of the visible light-infrared pedestrian re-identification method described above are applicable to the embodiments of visible light-infrared pedestrian re-identification. For details, please refer to the description in the embodiments of the method of the present invention, which will not be repeated here.

[0133] Furthermore, in the implementation of the visible light-infrared pedestrian re-identification system in the above embodiments, the logical division of each program module is only an example. In actual applications, the above functions can be assigned to different program modules as needed, for example, for the sake of corresponding hardware configuration requirements or software implementation convenience. That is, the internal structure of the visible light-infrared pedestrian re-identification system is divided into different program modules to complete all or part of the functions described above.

[0134] Please see Figure 8 In one embodiment, an electronic device is provided for implementing a method for visible-infrared pedestrian re-identification. The electronic device 200 may include a first processor 201, a first memory 202, and a bus, and may also include a computer program stored in the first memory 202 and executable on the first processor 201, such as a program 203 for visible-infrared pedestrian re-identification.

[0135] The first memory 202 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the first memory 202 can be an internal storage unit of the electronic device 200, such as the portable hard drive of the electronic device 200. In other embodiments, the first memory 202 can also be an external storage device of the electronic device 200, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 200. Furthermore, the first memory 202 can include both internal and external storage units of the electronic device 200. The first memory 202 can be used not only to store application software and various types of data installed on the electronic device 200, such as the code of the visible light-infrared pedestrian re-identification program 203, but also to temporarily store data that has been output or will be output.

[0136] In some embodiments, the first processor 201 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The first processor 201 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the first memory 202 and calls data stored in the first memory 202 to perform various functions of the electronic device 200 and process data.

[0137] Figure 8 Only electronic devices with components are shown; those skilled in the art will understand that... Figure 8 The structure shown does not constitute a limitation on the electronic device 200, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0138] The program 203 for visible light-infrared pedestrian re-identification stored in the first memory 202 of the electronic device 200 is a combination of multiple instructions, which, when run in the first processor 201, can achieve the following:

[0139] S1. Acquire visible light images and infrared images, and construct training batches after preprocessing the visible light images and infrared images. The preprocessing includes size normalization and data augmentation. The training batches adopt an identity equalization sampling strategy and include both visible light and infrared samples.

[0140] S2. Input the preprocessed visible light image and infrared image into the dual-branch backbone network respectively, and extract multi-layer features that contain at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-sharing branch and a modality-specific branch.

[0141] S3. Input the multi-layer features into the modality-sharing branch and the modality-specific branch respectively for representation learning, wherein the modality-sharing branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues;

[0142] S4. Perform structure prior-driven adaptive dual-path enhancement processing in the modality sharing branch to generate a spatial attention map and perform weighted residual enhancement on deep features to obtain modality sharing enhanced features.

[0143] S5. Perform hierarchical attention collaboration and dynamic bridging processing in the modality-specific branch, refine the extraction of modality-specific cues through multi-statistic residual channel calibration and asymmetric multi-scale spatial enhancement, and achieve cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features.

[0144] S6. By constructing a multi-task joint objective function and combining iterative training with implicit knowledge extraction strategies, the dual-branch backbone network is trained iteratively. The implicit knowledge extraction strategies include Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. These strategies effectively transfer rich information from modality-specific enhancement features to modality-shared branches, and a joint loss function guides network learning. The joint loss function includes:

[0145] Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved.

[0146] Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity.

[0147] Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is enhanced.

[0148] S7. In the inference stage, the cross-modal shared enhanced features are input into the trained dual-branch backbone network to calculate the similarity between the query sample and the candidate sample, and pedestrian retrieval and ranking output are completed. The similarity measure adopts cosine similarity or inner product similarity.

[0149] Furthermore, if the modules / units integrated in the electronic device 200 are implemented as software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0150] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0151] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0152] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for pedestrian re-identification based on visible light-infrared light, characterized in that, Includes the following steps: S1. Acquire visible light images and infrared images, and construct training batches after preprocessing the visible light images and infrared images. The preprocessing includes size normalization and data augmentation. The training batches adopt an identity equalization sampling strategy and include both visible light and infrared samples. S2. Input the preprocessed visible light image and infrared image into the dual-branch backbone network respectively, and extract multi-layer features that contain at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-sharing branch and a modality-specific branch. S3. Input the multi-layer features into the modality-sharing branch and the modality-specific branch respectively for representation learning, wherein the modality-sharing branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues; S4. Perform structure prior-driven adaptive dual-path enhancement processing in the modality sharing branch to generate a spatial attention map and perform weighted residual enhancement on deep features to obtain modality sharing enhanced features. S5. Perform hierarchical attention collaboration and dynamic bridging processing in the modality-specific branch, refine the extraction of modality-specific cues through multi-statistic residual channel calibration and asymmetric multi-scale spatial enhancement, and achieve cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features. S6. By constructing a multi-task joint objective function and combining iterative training strategies for implicit knowledge extraction, the dual-branch backbone network is trained iteratively. The implicit knowledge extraction strategies include Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. Through these implicit knowledge extraction strategies, the rich information in modality-specific enhancement features is effectively transferred to the modality-shared branch, and the network learning is guided by the joint loss function. The joint loss function includes: Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved. Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity. Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is enhanced. S7. In the inference stage, the cross-modal shared enhanced features are input into the trained dual-branch backbone network to calculate the similarity between the query sample and the candidate sample, and pedestrian retrieval and ranking output are completed. The similarity measure adopts cosine similarity or inner product similarity.

2. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, In step S1, the data augmentation includes one or more of random horizontal flipping, random erasing, random cropping, random rotation, and random channel perturbation; the identity equalization sampling strategy samples P identities for each modality and K images for each identity, where P and K are preset positive integers.

3. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, In step S2, the dual-branch backbone network is a convolutional neural network, a Transformer network, or a combination of both, and adopts a feature extraction structure based on ResNet-50.

4. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, In step S4, the structure-prior-driven adaptive dual-path enhancement processing specifically includes: S41. Divide the mid-level features into a preset number S strip regions along the height direction, and apply global average pooling to each strip region to capture strip-level priors. S42. For each strip region, pooling is performed using saliency-driven pooling and statistically invariant pooling branches to obtain strip-level saliency features and strip-level statistical prior features, and then the strip-level saliency features and strip-level statistical prior features are concatenated and fused; the saliency-driven pooling branch is an adaptive max pooling branch, and the statistically invariant pooling branch is a global average pooling branch for strip sub-blocks; S43. For each strip, a mapping network with non-shared parameters is used to generate the original weights of the strip. The mapping network includes a feature extraction head and a weight generation head. Softmax normalization is performed on the original weights of the strip so that the sum of the weights of each strip is 1. S44. Construct a strip mask based on the normalized strip weights, and generate a spatial attention map through convolutional transformation and nonlinear activation. S45. Perform element-wise multiplication between the spatial attention map and the deep features, and then add the residuals to the original deep features to output cross-modal shared enhanced features.

5. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, The hierarchical attention collaboration described in step S5 involves embedding a bimodal adaptive attention module after the bottleneck layer of the bimodal backbone network. The processing procedure of the bimodal adaptive attention module includes: a) Channel calibration: Perform global average pooling and global standard deviation pooling on the input features simultaneously to obtain first-order and second-order statistical features; concatenate the first-order and second-order statistical features and input them into a residual multilayer perceptron containing a dimensionality reduction layer, a nonlinear activation layer and an dimensionality increase layer to generate channel attention weights. b) Spatial Augmentation: The features after channel calibration are reduced in dimension and compressed. Then, spatial features are extracted through two parallel paths: the first path uses consecutive convolutional layers with kernels of 1×7 and 7×1 to perform horizontal-vertical feature decomposition, and the second path uses 3×3 dilated convolutions with an inflation rate of 3 to expand the receptive field. Finally, the outputs of the two paths are fused with element-wise maximum values ​​to generate spatial attention weights. c) Dynamic fusion: The modality discrimination coefficient is calculated based on the global mean of channel weights and spatial weights, and the channel enhancement features and spatial enhancement features are weighted and fused using this coefficient.

6. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, The dynamic bridging described in step S5 specifically includes: a) By using a cross-level adaptive bridging module, deep features are projected onto shallow channel dimensions through 1×1 convolution; b) Adjust the shallow features to the same spatial resolution as the deep features using bilinear interpolation; c) The projected deep features and the interpolated shallow features are spliced ​​together, and a fusion mask is generated based on the spliced ​​features through deformable convolution. The fusion mask is a continuous value between 0 and 1, which is used to perform convex combination weighted fusion of shallow features and deep features. d) Perform multi-head self-attention operation on the weighted fused features to obtain cross-layer bridging features, which are output as modality-specific enhancement features.

7. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, In step S6, the identity-level statistical distribution distillation loss based on dynamic uncertainty weighting specifically includes: a) Calculate statistical descriptor vectors for visible light features and infrared features for the same identity, wherein the statistical descriptors include a mean vector, a standard deviation vector, and an interquartile range vector; b) Calculate the cross-modal Pearson correlation coefficient for each statistical descriptor vector, and construct the basic consistency loss term accordingly; c) Introduce an independent learnable uncertainty parameter for each type of statistical descriptor, and construct the following dynamically weighted loss function: in, Indicates the statistical descriptor type, Based on the consistency loss term, The constraint weights of different statistics are adjusted synchronously with network training.

8. The method for pedestrian re-identification based on visible light-infrared as described in claim 1, characterized in that, Step S6, iteratively training the dual-branch backbone network, further includes the following steps: S601, Forward Propagation and Feature Calculation: Input the training batch data into the dual-branch backbone network, extract modality-shared enhancement features through the SPADE module, extract modality-specific enhancement features through the HAC-DBN module, and calculate the output tensors of each level and classification. S602. Joint Loss Collaborative Calculation: Based on the forward propagation results, in addition to the basic pedestrian re-identification loss, relevant losses are calculated simultaneously. and conventional auxiliary losses , And assign preset weights to each loss term to construct a joint objective function; S603, Parameter Adaptive Update: Update the backbone network weights using the backpropagation algorithm and update them synchronously. Learnable uncertainty parameters in As training iterates, the uncertainty parameters automatically converge based on the alignment difficulty, thereby adaptively adjusting the distillation intensity until the network converges and outputs the trained model.

9. A system for pedestrian re-identification based on visible light-infrared light, characterized in that, A method for pedestrian re-identification based on visible light-infrared light, applicable to any one of claims 1-8, includes an image acquisition module, an image preprocessing module, a representation learning module, a first processing module, a second processing module, a training module, and an inference module; The image acquisition module is used to acquire visible light images and infrared images, and to construct training batches after preprocessing the visible light images and infrared images. The preprocessing includes size normalization and data augmentation. The training batches adopt an identity equalization sampling strategy and include both visible light and infrared samples. The image preprocessing module is used to input the preprocessed visible light image and infrared image into a dual-branch backbone network respectively, and extract multi-layer features that contain at least shallow detail features and deep semantic features. The dual-branch backbone network includes a modality-sharing branch and a modality-specific branch. The representation learning module is used to input the multi-layer features into the modality-shared branch and the modality-specific branch respectively for representation learning, wherein the modality-shared branch is used to learn cross-modality consistent discriminative representations, and the modality-specific branch is used to retain modality-sensitive supplementary discriminative cues; The first processing module is used to perform structure prior-driven adaptive dual-path enhancement processing in the modality sharing branch, generate a spatial attention map and perform weighted residual enhancement on deep features to obtain modality sharing enhanced features; The second processing module is used to perform hierarchical attention collaboration and dynamic bridging processing in the modality-specific branch, refine the extraction of modality-specific cues through multi-statistic residual channel calibration and asymmetric multi-scale spatial enhancement, and achieve cross-layer fusion of shallow detail features and deep semantic features to obtain modality-specific enhanced features. The training module is used to iteratively train the dual-branch backbone network by constructing a multi-task joint objective function and combining it with an implicit knowledge extraction strategy. The implicit knowledge extraction strategy includes Triple Graph Alignment (TGSA) and Class Semantic Alignment (CSA) strategies. Through the implicit knowledge extraction strategy, the rich information in the modality-specific enhancement features is effectively transferred to the modality-shared branch, and the network learning is guided by the joint loss function. The joint loss function includes: Distillation loss based on identity-level statistical distribution weighted by dynamic uncertainty Utilizing learnable parameters Adaptive constraints are applied to ensure the consistency of the statistical distribution of the same identity in visible light and infrared modes. By dynamically identifying and suppressing the statistical distortion caused by infrared mode thermal noise, robust identity-level global distribution alignment is achieved. Cross-modal cooperative constraint loss for dynamic region sensing To address the local semantic shift in dual-path heterogeneous graph pooling, learnable weights are used to focus key regions, and a local semantic focusing mechanism is used to accurately align key discriminant units across modalities, thereby eliminating fine-grained semantic ambiguity caused by modal heterogeneity. Hierarchical adaptive bridging loss By minimizing the distance between deep feature projections and shallow features, cross-level consistency is enhanced, semantic coherence between feature levels is ensured, shallow details and deep semantics are effectively integrated, and the collaborative expressive ability of multi-scale features is enhanced. The reasoning module is used to input the cross-modal shared enhanced features into the trained dual-branch backbone network during the reasoning stage, calculate the similarity between the query sample and the candidate sample, and complete the pedestrian retrieval and ranking output. The similarity measure adopts cosine similarity or inner product similarity.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores computer program instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a method for visible-infrared pedestrian re-identification as described in any one of claims 1-8.

Images