Shielding pedestrian re-identification method and system based on multi-stage feature refining

Abstract

The invention provides an occluded pedestrian re-identification method and system based on multi-stage feature refining, belongs to the technical field of computer vision, and aims to solve the problems that in the prior art, in a complex occluded environment, pedestrian identity identification is difficult, the feature representation capability is insufficient, cross-view alignment is inaccurate, and multi-view fusion is rigid. A three-stage progressive feature optimization framework is provided, and through a staged'purification-alignment-fusion 'strategy, the scheme can establish a fine-grained cross-view corresponding relation while suppressing single-view noise, and adaptively distribute fusion weights according to occlusion features of samples, thereby obtaining more robust identity feature representation. Compared with the prior art, the scheme has the advantages that the recognition accuracy is remarkably improved in a pedestrian shielding re-recognition scene, the low calculation complexity is kept, and the practical value and the popularization prospect are high.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to a method and system for re-identifying occluded pedestrians based on multi-stage feature refinement. Background Technology

[0002] With the rapid development of artificial intelligence and computer vision technologies, Person Re-identification (ReID) has become a key component of intelligent security and video surveillance systems. This technology aims to identify and match the same pedestrian from multiple non-overlapping camera views, providing support for public safety monitoring, trajectory analysis, and behavior prediction. However, in real-world complex scenarios, the accuracy of traditional ReID methods significantly decreases due to variations in camera angles, lighting differences, pose distortion, and occlusion. Among these, occlusion is one of the most challenging problems. When a pedestrian is partially occluded by other objects (such as vehicles, buildings, or other pedestrians), the discriminative features of key areas are easily lost, leading to inaccurate cross-camera matching. The core problems caused by occlusion are reflected in the following two aspects:

[0003] Incomplete feature information within the view leads to the loss of effective identity features in a single frame image;

[0004] Cross-view semantic alignment is difficult, as the visible parts of the same pedestrian differ significantly under different cameras, causing semantic correspondence confusion and feature misalignment.

[0005] Existing research mainly follows two technical routes:

[0006] Occlusion-aware methods, such as the PGFA (Pose-Guided Feature Alignment) model, which uses human pose estimation to generate part masks, and the HOReID method, which establishes part-level semantic correspondences through keypoint detection. These methods rely on external auxiliary models, and although they can locate visible areas, they are susceptible to the propagation of pose estimation errors and have high computational costs.

[0007] Occlusion-robust methods, such as the SORN network which utilizes a spatial reordering strategy, and the Pirt method which introduces a region confidence attention enhancement mechanism, perform well in mild occlusion scenarios, but still struggle to maintain feature consistency in cases of severe occlusion or cross-view misalignment.

[0008] The inventors have discovered that in recent years, multi-view fusion technology has been introduced into occluded pedestrian recognition tasks, becoming an effective approach to mitigate the impact of occlusion. For example, MVI²P (Multi-view Information Integration and Propagation) proposes a three-stage mechanism of "localization-quantization-distillation," while the FED (Feature Erasure and Diffusion) method utilizes a Transformer structure to encode cross-view dependency information. However, existing multi-view fusion methods generally suffer from the following problems:

[0009] (1) Noise within the view was not effectively suppressed, resulting in low-quality features interfering with the discrimination;

[0010] (2) The fusion granularity is too coarse and lacks a fine spatial local matching mechanism;

[0011] (3) Lack of dynamic fusion strategy, and failure to adaptively adjust fusion weight according to different occlusion modes. Summary of the Invention

[0012] This invention provides a method and system for re-identifying occluded pedestrians based on multi-stage feature refinement, which effectively solves the problems existing in the prior art.

[0013] According to a first aspect of the present invention, an occluded pedestrian re-identification method based on multi-stage feature refinement is provided, comprising:

[0014] Acquire pedestrian images in occluded scenes and perform corresponding preprocessing;

[0015] Using a preprocessed pedestrian image as input, a pre-constructed neural network model based on multi-stage feature refinement is used for feature extraction to obtain fused features refined through multi-stage feature refinement. The neural network model employs a dual-branch classification architecture, with both branches using the same network structure and connected to a first classifier and a second classifier, respectively. The neural network model specifically performs the following processing steps: For the input pedestrian image, original feature tensors are extracted using a pre-defined backbone network; based on the extracted original feature tensors, feature purification and channel-level dynamic adjustment are performed using a parameterized attention distillation mechanism to eliminate invalid feature responses caused by occlusion or background, obtaining refined features; based on the obtained refined features, cross-view spatial feature enhancement is performed using a cross-view information enhancement mechanism to obtain enhanced refined features; based on the obtained refined features, multi-view feature fusion is performed using an adaptive feature decoupling fusion mechanism to obtain fused features.

[0016] Based on the fused features refined through multiple stages, the occluded pedestrian re-identification result is obtained through a preset first classifier.

[0017] Furthermore, the parameterized attention distillation mechanism is used for feature purification and channel-level dynamic adjustment to eliminate invalid feature responses in the image caused by occlusion or background, thereby obtaining refined features. Specifically:

[0018] For the input raw feature tensor, the class activation mapping is calculated using the classifier weights;

[0019] The original feature tensor is parameterized channel modulated based on computational class activation mapping and an introduced learnable channel scaling parameter.

[0020] Furthermore, the method of enhancing spatial features across views through a cross-view information enhancement mechanism to obtain enhanced refined features specifically involves:

[0021] The refined features obtained are processed by sliding window block division to obtain several blocks;

[0022] Features of each block are extracted using average pooling, and cross-view calculations are performed for each block based on these features. Figure 1 Consistency score;

[0023] Based on the obtained cross-view Figure 1 Consistency score, spatial feature enhancement is performed on the refined features to obtain enhanced refined features.

[0024] Furthermore, the obtained cross-view Figure 1 Consistency score, spatial feature enhancement is performed on the refined features to obtain enhanced refined features, specifically:

[0025] Cosine similarity is used to calculate cross-view block-level matching relationships, and the cross-view similarity of each block is calculated. Figure 1 Consistency score, specifically expressed as:

[0026]

[0027] Based on the obtained cross-view Figure 1 Consistency score, spatial feature enhancement of the refined features, specifically expressed as:

[0028]

[0029] in, The consistency score is given for the k-th block under the v-th view of the i-th target to be identified. For the local features of the k-th block under the v-th view of the i-th target to be identified, For the local features of the l-th block under the u-th view of the i-th target to be identified, For the enhanced and refined features corresponding to the v-th view of the i-th target to be identified, For the refined features corresponding to the v-th view of the i-th target to be identified, This represents the number of vertical blocks. H represents the number of horizontal blocks, and H and W represent the height and width of the original feature tensor.

[0030] Furthermore, the adaptive feature decoupling and fusion mechanism is used to perform feature fusion of multiple views to obtain fused features, specifically as follows:

[0031] Based on the obtained enhanced refined features, global average pooling is performed on the enhanced refined features of each view to obtain compact global features;

[0032] The compact global feature vectors of multiple views are concatenated and divided into several independent subspaces. Each subspace is decoupled and nonlinearly mapped through an independent MLP subnetwork to obtain the fusion weights of different views.

[0033] Based on the obtained fusion weights, the fusion of refined features for multi-view enhancement is performed to obtain fused features.

[0034] Furthermore, based on the obtained fusion weights, the fusion of refined features for multi-view enhancement is performed to obtain fused features, specifically as follows:

[0035]

[0036] in, Let the fusion weight be the v-th view corresponding to the i-th target to be identified. ( ) represents global average pooling. The enhanced and refined features corresponding to the v-th view of the first target to be identified. This is a feature of fusion.

[0037] Furthermore, the preset backbone network uses a ResNet-50 network to extract the original feature tensors.

[0038] According to a second aspect of the present invention, an occluded pedestrian re-identification system based on multi-stage feature refinement is provided, comprising:

[0039] The data acquisition unit is used to acquire pedestrian images in occluded scenes and perform corresponding preprocessing.

[0040] The feature fusion unit takes a preprocessed pedestrian image as input and extracts features through a pre-built neural network model based on multi-stage feature refinement to obtain fused features refined through multi-stage feature refinement. The neural network model based on multi-stage feature refinement adopts a dual-branch classification architecture, with both branches using the same network structure and connected to a first classifier and a second classifier, respectively. The neural network model specifically performs the following processing steps: For the input pedestrian image, the original feature tensor is extracted through a pre-defined backbone network; based on the extracted original feature tensor, feature purification and channel-level dynamic adjustment are performed through a parameterized attention distillation mechanism to eliminate invalid feature responses caused by occlusion or background in the image, obtaining refined features; based on the obtained refined features, cross-view spatial feature enhancement is performed through a cross-view information enhancement mechanism to obtain enhanced refined features; based on the obtained refined features, multi-view feature fusion is performed through an adaptive feature decoupling fusion mechanism to obtain fused features.

[0041] The recognition unit is used to obtain the occluded pedestrian re-recognition result based on the fused features refined through multi-stage features and a preset first classifier.

[0042] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the occluded pedestrian re-identification method based on multi-stage feature refinement.

[0043] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the occluded pedestrian re-identification method based on multi-stage feature refinement.

[0044] The above one or more technical solutions have the following beneficial effects:

[0045] This invention provides a method and system for re-identifying occluded pedestrians based on multi-stage feature refinement. Addressing the problems of difficulty in pedestrian identification under complex occlusion environments, insufficient feature representation capabilities, inaccurate cross-view alignment, and rigid multi-view fusion in existing technologies, the proposed scheme employs a three-stage progressive feature optimization framework. Through a phased "clean-align-fusion" strategy, the scheme can suppress single-view noise while establishing fine-grained cross-view correspondences and adaptively allocate fusion weights based on the occlusion features of the samples, thereby obtaining a more robust identity feature representation. Compared with existing technologies, this scheme not only significantly improves the recognition accuracy in occluded pedestrian re-identification scenarios but also maintains low computational complexity, demonstrating strong practical value and promising prospects for widespread application.

[0046] The proposed scheme employs a feature refinement module guided by class activation mapping to generate a spatial attention distribution based on classifier weights, identifying the most discriminative regions in the image. This module utilizes a parameterized channel modulation mechanism to transform dynamic CAM supervision into static learnable parameters, achieving adaptive weighting of feature channels, thereby effectively suppressing background noise and occlusion interference, and enhancing the feature expression of visible parts.

[0047] The proposed scheme, for multi-view features after feature purification, uses a cross-view information enhancement module to divide the feature map into overlapping local blocks through a sliding window mechanism, calculates a block-level similarity matrix to measure cross-view consistency, and utilizes the semantic consistency assumption to match unoccluded regions under multiple views. It also generates a spatial weight map through bilinear interpolation, thereby adaptively strengthening the semantically corresponding feature regions between multiple views and realizing feature reconstruction and compensation for different occlusion modes.

[0048] For the enhanced multi-view features, the proposed scheme uses an adaptive feature decoupling and fusion module. By employing a grouped multilayer perceptron structure, the high-dimensional feature vector is divided into several subspaces for parallel processing. Feature decoupling is achieved through independent learning between groups. This module is based on a sample-level adaptive weight generation strategy, which assigns dynamic fusion coefficients to each view and forms the final fusion feature through weighted averaging, thereby realizing semantic perception fusion of multi-view features.

[0049] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0050] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0051] Figure 1 This is a flowchart of the occluded pedestrian re-identification method based on multi-stage feature refinement described in this embodiment of the invention.

[0052] Figure 2 This is a schematic diagram of the overall framework structure of the neural network model based on multi-stage feature refinement as described in this embodiment of the invention.

[0053] Figure 3 This is a schematic diagram of the guiding feature refining module structure described in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the cross-view information enhancement module structure described in an embodiment of the present invention;

[0055] Figure 5This is a visual comparison chart of the features of different modules described in the embodiments of the present invention;

[0056] Figure 6 This is a comparison chart of multi-view and baseline visualizations as described in an embodiment of the present invention. Detailed Implementation

[0057] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0058] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0059] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0060] Example 1:

[0061] To address the above problems, in one or more implementations, such as Figure 1 As shown, this embodiment provides a method for re-identifying occluded pedestrians based on multi-stage feature refinement, including:

[0062] Acquire pedestrian images in occluded scenes and perform corresponding preprocessing;

[0063] Using a preprocessed pedestrian image as input, a pre-constructed neural network model based on multi-stage feature refinement is used for feature extraction to obtain fused features refined through multi-stage feature refinement. The neural network model employs a dual-branch classification architecture, with both branches using the same network structure and connected to a first classifier and a second classifier, respectively. The neural network model specifically performs the following processing steps: For the input pedestrian image, original feature tensors are extracted using a pre-defined backbone network; based on the extracted original feature tensors, feature purification and channel-level dynamic adjustment are performed using a parameterized attention distillation mechanism to eliminate invalid feature responses caused by occlusion or background, obtaining refined features; based on the obtained refined features, cross-view spatial feature enhancement is performed using a cross-view information enhancement mechanism to obtain enhanced refined features; based on the obtained refined features, multi-view feature fusion is performed using an adaptive feature decoupling fusion mechanism to obtain fused features.

[0064] Based on the fused features refined through multiple stages, the occluded pedestrian re-identification result is obtained through a preset first classifier.

[0065] In specific implementation, the neural network model based on multi-stage feature refinement adopts a dual-branch classification architecture. The two branches use the same network structure and are respectively connected to the first classifier and the second classifier. During the model training phase, the second classifier is used to learn the feature distribution of the second branch and guides the parameter update of the first classifier through inference loss. During the model field utilization phase, only the first classifier is used as the final recognition classifier.

[0066] In one or more embodiments, such as Figure 2 As shown, the neural network model based on multi-stage feature refinement described in this embodiment mainly includes a backbone network, a Class Activation Map Guided Feature Refinement (CGFR) module, a Cross-View Information Enhancement (CVIEM) module, an Adaptive Feature Decoupling and Fusion (AFDFM) module, and a classifier.

[0067] In one or more embodiments, the preset backbone network uses a ResNet-50 network to extract the original feature tensors.

[0068] In specific implementation, the parametric attention distillation mechanism is used for feature purification and channel-level dynamic adjustment to eliminate invalid feature responses in the image caused by occlusion or background, thereby obtaining refined features. Specifically:

[0069] For the input raw feature tensor, the class activation mapping is calculated using the classifier weights;

[0070] The original feature tensor is parameterized channel modulated based on computational class activation mapping and an introduced learnable channel scaling parameter.

[0071] In one or more embodiments, such as Figure 3 As shown, the Class ActivationMap Guided Feature Refinement (CGFR) module is the first-stage module of the scheme described in this embodiment. Its purpose is to achieve feature purification and channel-level dynamic adjustment through a parameterized attention distillation mechanism, thereby eliminating invalid feature responses in the input image caused by occlusion or background. Its main concept is as follows:

[0072] An initial multi-channel feature map is extracted using a backbone network (such as ResNet-50). For this feature map, class activation maps (CAMs) are generated using classifier weights to reflect the importance distribution of the target identity in the feature space. Subsequently, a parameterized channel modulation mechanism is introduced to transform the dynamic CAM supervision into static learnable parameters, achieving channel-level weighting and residual feature fusion, thereby enhancing the response of identity-related regions and suppressing background noise. During the inference phase, the module degenerates into channel multiplication operations, effectively reducing computational complexity and maintaining the stability of the feature distribution.

[0073] Its specific implementation is as follows:

[0074] (1) For the input multi-channel feature map The classifier weights are Computational activation mapping The CAM (Camera Mapping) is used to measure the importance of different spatial locations in identity classification tasks. Based on knowledge distillation theory and parameterized function approximation theory, the spatial attention pattern reflected by CAM has statistical consistency across samples. Through this mechanism, the network can automatically identify and enhance salient regions related to identity in an image, while weakening the influence of background and occluded parts.

[0075] (2) Introduce learnable channel scaling parameters The importance of different channel features is dynamically adjusted. During the training phase, CAM mapping information is embedded into the channel weight update by jointly optimizing the loss function, enabling the network to automatically learn to distinguish salient features from redundant noise among different samples. During the inference phase, the module degenerates into lightweight channel multiplication operations, achieving a dimensionality reduction from O(CHW) to O(C). This degradation mechanism ensures that the model maintains efficient inference performance during deployment, making it suitable for real-time recognition needs in practical scenarios.

[0076] (3) A residual learning mechanism is introduced during the feature refinement process, so that the output features contain both the purified salient semantic features and retain the global consistency of the input features, thereby ensuring the stable propagation of gradients during backpropagation. Through this structural design, the model can effectively internalize the CAM supervision signal in a multi-layer network, forming an interpretable, learnable and stable feature purification mechanism, and finally outputting a high-quality identity feature representation with stronger discriminative ability and occlusion robustness.

[0077] The following example illustrates the specific working process of the CAM-guided feature refinement module:

[0078] Extracting the original feature tensor through a backbone network (such as ResNet-50) Its dimensions are denoted as C×H×W, where C is the number of channels, and H and W are the height and width, respectively.

[0079] Then, the CAM-guided feature refinement module is input, which performs the following processing steps:

[0080] (1) Calculate the class activation map. For the input feature map Through the weights of the first classifier Computational activation mapping By reflecting the response intensity of each spatial location to a specific category, the most discriminative regions are highlighted, and numerical stability is ensured through normalization operations, as shown below:

[0081]

[0082] in, For feature map The formula uses normalization to ensure that the values ​​are within the range of [0,1], highlighting the most important spatial regions.

[0083] (2) Parametric channel modulation. Learnable channel scaling parameters are introduced. Feature refinement is achieved during training through the following mechanisms:

[0084]

[0085] Where I is the unit tensor, used to control the fusion ratio of CAM information and original features. This design is based on residual learning theory, combining residual connection structure and layer normalization mechanism to maintain the continuity of feature representation and the stability of gradient propagation, prevent deep feature degradation, and ensure effective gradient propagation and the stability of feature representation.

[0086] By using a parameterized channel modulation mechanism to reweight the feature map channel by channel, the high-response region is enhanced while the low-response region is suppressed.

[0087] (3) Inference optimization. This module is designed based on attention distillation and parametric learning theory. During the training phase, dynamic CAM supervision is used to adaptively adjust the channel distribution, while during the inference phase, it gracefully degenerates into channel-level multiplication operations: This reduces the computational complexity from Down to It significantly improves feature recognition while maintaining extremely high computational efficiency.

[0088] In specific implementation, the method of enhancing spatial features across views through a cross-view information enhancement mechanism to obtain enhanced refined features specifically involves:

[0089] The refined features obtained are processed by sliding window block division to obtain several blocks;

[0090] Features of each block are extracted using average pooling, and cross-view calculations are performed for each block based on these features. Figure 1 Consistency score;

[0091] Based on the obtained cross-view Figure 1 Consistency score, spatial feature enhancement is performed on the refined features to obtain enhanced refined features.

[0092] In one or more embodiments, such as Figure 4 As shown, a Cross-View Information Enhancement (CVIEM) module is designed. Based on the concept that even under complex occlusion conditions, semantically consistent unoccluded local regions still exist in multi-view images of the same identity, the module offers the following technical concept: Based on the semantic consistency assumption of multi-view samples, the feature map purified by the CGFR module is processed into blocks using a sliding window. Features of each local block are extracted through average pooling, and cross-view cosine similarity is calculated to construct a block-level matching matrix, measuring the semantic correspondence between local regions. A cross-view weight map is generated using consistency score mapping, and then back-projected into the original feature space through bilinear interpolation to achieve spatial-level feature enhancement. This module can adaptively enhance stable visible regions for different occlusion modes, thereby achieving higher accuracy in cross-view feature alignment.

[0093] It includes the following processing steps:

[0094] (1) Based on the observation of semantically consistent unoccluded local regions for the same identity under multiple camera views, a sliding window mechanism is used to decompose the feature map into overlapping local blocks, with a window size of [missing information]. With a step size of s, some overlap exists between blocks to ensure feature continuity and spatial smoothness. This local partitioning can capture fine-grained spatial semantic information while ensuring computational controllability, laying the foundation for subsequent cross-view matching;

[0095] (2) Perform average pooling on each local block to obtain its feature representation. This operation is based on the assumption of local stationarity, that is, the features of adjacent regions have similar semantic content, thereby preserving semantic consistency while reducing noise sensitivity. Through this feature compression method, the module effectively reduces local noise interference and enhances stability;

[0096] (3) The cosine similarity function is used to calculate the matching relationship between corresponding local blocks under different views. This similarity measure has good stability and orientation sensitivity under scale changes and illumination differences, and can accurately reflect the structural similarity between local regions in multi-view images. The cross-view matching matrix generated thereby can characterize the degree of correspondence between blocks;

[0097] (4) Calculate cross-view based on block-level matching results Figure 1Consistency scores are used to quantify the reliability of each local block in a multi-view context. A bilinear interpolation method is used to map the block-level weight map back to the original feature map size, generating a spatially enhanced weight map. This weight map is dynamically updated during training, thereby strengthening semantically consistent regions and suppressing occluded regions. Ultimately, the enhanced feature map significantly improves both global semantic representation and local alignment accuracy, providing more discriminative input features for subsequent fusion stages.

[0098] The following example illustrates in detail the working process of the cross-view information enhancement module:

[0099] (1) Sliding window decomposition. Given a refined feature map The feature map is decomposed into overlapping local blocks using a sliding window. In the embodiment described in this city, the window size is... The step size s = 8. The number of blocks is:

[0100]

[0101] in, The number of blocks.

[0102] (2) Local feature extraction. Features of each block are extracted using average pooling to reduce noise sensitivity and maintain local semantic consistency:

[0103]

[0104] in, For block indexes, Let be the coordinates of the top-left corner of the k-th block. This pooling strategy is based on the assumption of local stationarity, which assumes that features of neighboring spatial locations have similar semantic content.

[0105] (3) Cross-view matching. Cosine similarity is used to calculate cross-view block-level matching relationships. The choice of cosine similarity is based on its scale invariance and orientation sensitivity. Cross-view matching is defined for each block. Figure 1 Consistency score:

[0106]

[0107] For each local block, the most similar corresponding region is found in other views, based on the cross-view results of each block. Figure 1 Consistency scores establish robust cross-view matching relationships.

[0108] (4) Feature enhancement. Spatial-level feature enhancement is achieved by mapping block-level weights back to the original feature map size using bilinear interpolation.

[0109]

[0110] It should be noted here that cross-view generation is based on block-level similarity. Figure 1 The consistency score is used to map the weights back to the original feature map size using bilinear interpolation, resulting in a spatial weight distribution map. Finally, the enhancement weights are multiplied point-by-point with the feature map to strengthen highly consistent regions and suppress occluded regions, thereby constructing fine-grained, multi-scale cross-view enhancement features and achieving multi-angle feature complementarity and semantic alignment.

[0111] In specific implementation, the adaptive feature decoupling and fusion mechanism is used to perform feature fusion of multiple views to obtain fused features, specifically as follows:

[0112] Based on the obtained enhanced refined features, global average pooling is performed on the enhanced refined features of each view to obtain compact global features;

[0113] The compact global feature vectors of multiple views are concatenated and divided into several independent subspaces. Each subspace is decoupled and nonlinearly mapped through an independent MLP subnetwork to obtain the fusion weights of different views.

[0114] Based on the obtained fusion weights, the fusion of refined features for multi-view enhancement is performed to obtain fused features.

[0115] In one or more implementations, the Adaptive Feature Decoupling Fusion Module (AFDFM) is based on personalized learning theory, which holds that effective multi-view fusion should be sample-adaptive.

[0116] Its specific implementation is as follows:

[0117] (1) Based on the theory of personalized learning, it is believed that effective multi-view feature fusion should have sample adaptability. Different samples have large differences in occlusion ratio, viewpoint difference and image quality, so it is necessary to design a targeted fusion strategy. To this end, this module introduces a dynamic weight learning mechanism in its design, so that the model can automatically adjust the fusion method according to the feature distribution of the input samples;

[0118] (2) To achieve independent modeling of features in multiple subspaces, a Grouped Multilayer Perceptron (MLP) architecture is introduced. Inspired by group regularization theory and modular neural network design, this structure divides high-dimensional feature vectors into several independent subspaces. Each subspace is processed through an independent MLP subnetwork, thereby achieving feature decoupling and nonlinear mapping. This grouping approach can reduce the risk of overfitting and improve the flexibility of feature representation;

[0119] (3) Divide the concatenated feature vectors from the four views into sequential categories. Each group of features undergoes feature transformation and activation function mapping via an independent MLP sub-network. Through parallel computation within groups and collaborative optimization between groups, the module can efficiently capture the differential and complementary features between views, thereby significantly reducing optimization complexity while maintaining high expressive power.

[0120] (4) The fusion layer concatenates the hidden representations of all subspaces and generates fusion weight vectors for each view through the Softmax output layer. Based on the generated weights, the model performs a weighted average of the features from different views to obtain the final fused feature representation. This fusion strategy not only effectively integrates visual information from different angles but also enhances the model's adaptability to occlusion changes and environmental disturbances through sample-level adaptive adjustment. Furthermore, the grouping strategy reduces the parameter complexity from O(D²) to ,in This significantly reduces the computational burden and improves training and inference efficiency, making the module both high-performance and highly deployable.

[0121] The following example illustrates the working process of the adaptive feature decoupling and fusion module in detail:

[0122] (1) Feature compaction. Global average pooling is performed on the enhanced features to obtain compact features:

[0123]

[0124] (2) Feature concatenation. For the four views of the i-th identity, construct a concatenated feature vector:

[0125]

[0126] (3) Grouped Feature Decoupling. To reduce parameter complexity, the cascaded feature vectors are divided into G=4 groups for independent processing, with each group having a dimension of... Each group performs nonlinear feature transformation and semantic decoupling through an independent MLP subnetwork, enabling independent modeling of different subspaces:

[0127]

[0128] in, This is the weight matrix. This is the bias vector.

[0129] (4) Weight generation and fusion. After processing within the group, the hidden representations of all groups are concatenated to generate the fusion weights of the four views:

[0130]

[0131]

[0132] in, To map the hidden layer features into a linear projection matrix of the final grouped output, This is the corresponding bias vector.

[0133] The final fusion features were obtained through a weighted average:

[0134]

[0135] The grouping strategy reduces the parameter complexity from O(D²) to ,in Where Hg is a vector Dimensions.

[0136] In practical implementation, the training of the neural network model based on multi-stage feature refinement adopts a two-branch collaborative end-to-end training strategy, wherein:

[0137] A composite loss function is introduced during training. This function includes identity classification loss and triplet distance loss for the first and second classifiers, as well as inference distillation loss connecting the two classifiers. The identity classification loss and triplet distance loss are used to ensure the discriminativeness of the features extracted by each branch. The inference distillation loss is used to measure the difference between the output distributions of the first and second classifiers. By minimizing this difference, the first classifier (main branch) is forced to learn and approximate the feature distribution of the second classifier (multi-view branch), thereby transferring the prior knowledge of the multi-view branch to the first classifier, balancing the model's feature discriminativeness and cross-view performance in real-world applications. Figure 1 To the point of being responsive.

[0138] In the inference stage, the trained model is used to encode the input image, output pedestrian feature vectors, and calculate the Euclidean distance between features to complete identity matching, thereby achieving accurate re-identification of pedestrians with complex occlusions.

[0139] Specifically, such as Figure 2 As shown, the training strategy for the model is as follows:

[0140] (1) Loss Function Design. A multi-task learning strategy is adopted, and the overall loss function consists of four key components:

[0141]

[0142] in, For identity classification loss, To fuse feature classification loss, For reasoning loss, The triplet loss is set with weights λ1=0.007 and λ2=0.5.

[0143] (2) Training configuration. The Adam optimizer was used, with a batch size of 64 (8 identities, 8 images per identity) and an initial learning rate of 3×10⁻. 4 The training time was reduced to 0.1 times the original value at the 40th and 70th epochs, for a total of 120 epochs.

[0144] (3) Data augmentation. Data augmentation techniques such as random flipping, padding, and pruning are used to improve the model’s generalization ability and resistance to overfitting.

[0145] (4) Evaluation strategy. Standard pedestrian re-identification evaluation metrics were used, including Rank-1 accuracy, Rank-5 accuracy, Rank-10 accuracy, and mean precision (mAP).

[0146] In one or more implementations, dataset preparation and preprocessing are required before model training, mainly employing the following concepts:

[0147] A re-identification dataset containing pedestrian images from multiple viewpoints and varying degrees of occlusion was obtained. The dataset was divided into training and testing sets, with each identity including image samples from multiple camera viewpoints. For the input image samples, size normalization, color space normalization, and data augmentation operations were performed, including random horizontal flipping, random padding, random cropping, and normalization, to ensure the diversity of training samples in terms of pose, occlusion ratio, and lighting conditions, thereby improving the robustness and generalization ability of the model.

[0148] Specifically, the processing steps include the following:

[0149] (1) Collect relevant occluded pedestrian re-identification datasets. (1) Occluded-DukeMTMC dataset: the most challenging occluded dataset, containing 15,618 images for 702 training identities and 2,210 images for 519 query identities. All samples contain varying degrees of occlusion. (2) P-DukeMTMC-reID dataset: contains 1,299 identities, of which 665 are used for training. The query images in the test set are all subject to occlusion. (3) Market-1501 dataset: contains 1,501 identities and 32,668 images, used to verify the generalization ability of the method.

[0150] (2) Split the dataset according to the standard partitioning ratio. For the Occluded-DukeMTMC dataset, the training set contains 15,618 images, and the test set contains 2,210 query images and 17,661 candidate images. Ensure that the same identity does not overlap in the training and test sets, and each identity contains images from multiple perspectives.

[0151] (3) All input images were uniformly adjusted to 256×128 pixels and normalized. To improve the model's generalization ability, enhancement strategies such as random horizontal flipping, padding, random cropping, and color perturbation were adopted to increase sample diversity and viewpoint robustness. This step can effectively alleviate the overfitting problem and improve the model's performance stability under different scenes, lighting, and occlusion conditions.

[0152] like Figure 5 As shown, a feature visualization comparison chart of the different modules described in the embodiment is presented, such as... Figure 6 As shown, a multi-view and baseline visualization comparison chart of the solution described in the embodiment is presented.

[0153] Example 2:

[0154] In one or more embodiments, corresponding to the above method, this embodiment provides an occluded pedestrian re-identification system based on multi-stage feature refinement, including:

[0155] The data acquisition unit is used to acquire pedestrian images in occluded scenes and perform corresponding preprocessing.

[0156] The feature fusion unit takes a preprocessed pedestrian image as input and extracts features through a pre-built neural network model based on multi-stage feature refinement to obtain fused features refined through multi-stage feature refinement. The neural network model based on multi-stage feature refinement adopts a dual-branch classification architecture, with both branches using the same network structure and connected to a first classifier and a second classifier, respectively. The neural network model specifically performs the following processing steps: For the input pedestrian image, the original feature tensor is extracted through a pre-defined backbone network; based on the extracted original feature tensor, feature purification and channel-level dynamic adjustment are performed through a parameterized attention distillation mechanism to eliminate invalid feature responses caused by occlusion or background in the image, obtaining refined features; based on the obtained refined features, cross-view spatial feature enhancement is performed through a cross-view information enhancement mechanism to obtain enhanced refined features; based on the obtained refined features, multi-view feature fusion is performed through an adaptive feature decoupling fusion mechanism to obtain fused features.

[0157] The recognition unit is used to obtain the occluded pedestrian re-recognition result based on the fused features refined through multi-stage features and a preset first classifier.

[0158] In further embodiments, the following is also provided:

[0159] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the methods described in the above embodiments. For brevity, further details are omitted here.

[0160] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0161] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0162] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the methods described in the above embodiments.

[0163] The methods described in the above embodiments can be directly executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0164] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0165] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for re-identifying occluded pedestrians based on multi-stage feature refinement, characterized in that, include: Acquire pedestrian images in occluded scenes and perform corresponding preprocessing; Using a preprocessed pedestrian image as input, a pre-constructed neural network model based on multi-stage feature refinement is used for feature extraction to obtain fused features refined through multi-stage feature refinement. The neural network model employs a dual-branch classification architecture, with both branches using the same network structure and connected to a first classifier and a second classifier, respectively. The neural network model specifically performs the following processing steps: For the input pedestrian image, original feature tensors are extracted using a pre-defined backbone network; based on the extracted original feature tensors, feature purification and channel-level dynamic adjustment are performed using a parameterized attention distillation mechanism to eliminate invalid feature responses caused by occlusion or background, obtaining refined features; based on the obtained refined features, cross-view spatial feature enhancement is performed using a cross-view information enhancement mechanism to obtain enhanced refined features; based on the obtained refined features, multi-view feature fusion is performed using an adaptive feature decoupling fusion mechanism to obtain fused features. Based on the fused features refined through multiple stages, the occluded pedestrian re-identification result is obtained through a preset first classifier.

2. The occluded pedestrian re-identification method based on multi-stage feature refinement as described in claim 1, characterized in that, The process of using a parameterized attention distillation mechanism to perform feature purification and channel-level dynamic adjustment eliminates invalid feature responses in the image caused by occlusion or background, thereby obtaining refined features. Specifically: For the input raw feature tensor, the class activation mapping is calculated using the classifier weights; The original feature tensor is parameterized channel modulated based on computational class activation mapping and an introduced learnable channel scaling parameter.

3. The occluded pedestrian re-identification method based on multi-stage feature refinement as described in claim 1, characterized in that, The method of enhancing spatial features across views through a cross-view information enhancement mechanism to obtain enhanced refined features specifically involves: The refined features obtained are processed by sliding window block division to obtain several blocks; Features of each block are extracted by average pooling, and cross-view consistency score of each block is calculated based on the features of each block. Based on the obtained cross-view consistency score, spatial feature enhancement is performed on the refined features to obtain enhanced refined features.

4. The occluded pedestrian re-identification method based on multi-stage feature refinement as described in claim 3, characterized in that, Based on the obtained cross-view consistency score, spatial feature enhancement is performed on the refined features to obtain enhanced refined features, specifically as follows: Cosine similarity is used to calculate cross-view block-level matching relationships, and the cross-view consistency score for each block is calculated, specifically as follows: ； Based on the obtained cross-view consistency score, spatial feature enhancement is performed on the refined features, specifically as follows: ； in, The consistency score is given for the k-th block under the v-th view of the i-th target to be identified. For the local features of the k-th block under the v-th view of the i-th target to be identified, For the local features of the l-th block under the u-th view of the i-th target to be identified, For the enhanced and refined features corresponding to the v-th view of the i-th target to be identified, For the refined features corresponding to the v-th view of the i-th target to be identified, This represents the number of vertical blocks. H represents the number of horizontal blocks, and H and W represent the height and width of the original feature tensor.

5. The occluded pedestrian re-identification method based on multi-stage feature refinement as described in claim 1, characterized in that, The adaptive feature decoupling and fusion mechanism is used to perform feature fusion of multiple views to obtain fused features, specifically as follows: Based on the obtained enhanced refined features, global average pooling is performed on the enhanced refined features of each view to obtain compact global features; The compact global feature vectors of multiple views are concatenated and divided into several independent subspaces. Each subspace is decoupled and nonlinearly mapped through an independent MLP subnetwork to obtain the fusion weights of different views. Based on the obtained fusion weights, the fusion of refined features for multi-view enhancement is performed to obtain fused features.

6. The occluded pedestrian re-identification method based on multi-stage feature refinement as described in claim 5, characterized in that, Based on the obtained fusion weights, the refined features of multi-view enhancement are fused to obtain fused features, specifically as follows: ； in, Let the fusion weight be the v-th view corresponding to the i-th target to be identified. ( ) represents global average pooling. The enhanced and refined features corresponding to the v-th view of the first target to be identified. This is a feature of fusion.

7. The occluded pedestrian re-identification method based on multi-stage feature refinement as described in claim 1, characterized in that, The preset backbone network uses a ResNet-50 network to extract the original feature tensors.

8. An occluded pedestrian re-identification system based on multi-stage feature refinement, characterized in that, include: The data acquisition unit is used to acquire pedestrian images in occluded scenes and perform corresponding preprocessing. The feature fusion unit takes a preprocessed pedestrian image as input and extracts features through a pre-built neural network model based on multi-stage feature refinement to obtain fused features refined through multi-stage feature refinement. The neural network model based on multi-stage feature refinement adopts a dual-branch classification architecture, with both branches using the same network structure and connected to a first classifier and a second classifier, respectively. The neural network model specifically performs the following processing steps: For the input pedestrian image, the original feature tensor is extracted through a pre-defined backbone network; based on the extracted original feature tensor, feature purification and channel-level dynamic adjustment are performed through a parameterized attention distillation mechanism to eliminate invalid feature responses caused by occlusion or background in the image, obtaining refined features; based on the obtained refined features, cross-view spatial feature enhancement is performed through a cross-view information enhancement mechanism to obtain enhanced refined features; based on the obtained refined features, multi-view feature fusion is performed through an adaptive feature decoupling fusion mechanism to obtain fused features. The recognition unit is used to obtain the occluded pedestrian re-recognition result based on the fused features refined through multi-stage features and a preset first classifier.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the occluded pedestrian re-identification method based on multi-stage feature refinement as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the occluded pedestrian re-identification method based on multi-stage feature refinement as described in any one of claims 1-7.

Images