Cross-view pedestrian re-identification system and method based on implicit information compensation

Abstract

This invention relates to a cross-view pedestrian re-identification system and method based on implicit information compensation, belonging to the field of computer vision and intelligent video analysis technology. The invention includes a cross-view pedestrian image acquisition and preprocessing module, a basic feature extraction module, a multi-layer feature guidance module, a viewpoint conditional discrimination enhancement module, an implicit information compensation module, and a pedestrian identity discrimination and model training module, comprising the following steps: S1: cross-view pedestrian image acquisition, preprocessing, and basic feature extraction; S2: multi-layer feature-guided modeling; S3: viewpoint conditional discriminative feature enhancement; S4: implicit information compensation and feature reconstruction; and S5: pedestrian identity discrimination and end-to-end training with joint loss. This invention, while mitigating the cross-view feature distribution shift, fully preserves key identity discrimination information, significantly improving the discriminativeness, robustness, and generalization ability of air-to-ground cross-view pedestrian re-identification, and can be widely applied to the integrated air-to-ground monitoring system for smart city public safety and security.

Description

Technical Field

[0001] This invention relates to an aerial-ground cross-view pedestrian re-identification system and method based on implicit information compensation, which is applicable to the integrated air-ground monitoring system in smart city construction and public safety prevention and control, and belongs to the field of computer vision and intelligent video analysis technology. Background Technology

[0002] Person re-identification is one of the core technologies in computer vision and intelligent surveillance. Its goal is to retrieve and associate pedestrian targets with the same identity across a cross-view, non-overlapping network of cameras, making it a crucial link in achieving intelligent surveillance and personnel trajectory tracking. As surveillance systems evolve towards air-ground integration, unmanned aerial vehicles (UAVs) and high-altitude monitoring equipment have become important supplements to fixed ground-based surveillance, giving rise to the application scenario of Aerial-Ground ReID (AG-ReID). This scenario has significant practical application value in public safety, emergency response, and smart city management.

[0003] However, aerial-ground cross-view pedestrian re-identification faces technical challenges far exceeding those of traditional same-view / approximate-view re-identification: ground-view images are acquired at eye level, capturing complete structural and appearance information of the pedestrian's torso and limbs; while aerial views are often large overhead or even vertical top-view, resulting in severe perspective compression and morphological distortion in pedestrian images, preserving only limited structural features such as the head and shoulders. Due to the significant differences in shooting height, pitch angle, and imaging distance, the appearance features of the same pedestrian exhibit severe nonlinear distribution shifts under aerial-ground perspectives, leading to significant inter-domain drift in the feature space. When traditional re-identification models are directly applied to this scenario, recognition performance deteriorates drastically due to feature misalignment and missing semantic information.

[0004] To address the aforementioned issues, existing technologies primarily explore two directions: viewpoint-invariant feature learning and feature distribution alignment. 1. Viewpoint-invariant feature learning methods suppress viewpoint information in features through mechanisms such as adversarial learning and normalization, and extract common identity representations. However, while eliminating viewpoint differences, these methods are prone to losing key identity details attached to a specific viewpoint, resulting in insufficient identity discrimination ability of the features. 2. Feature distribution alignment methods reduce the feature distance between different viewpoints through geometric transformations and metric constraints. However, existing methods are mostly based on global statistics or coarse-grained structures for alignment, which makes it difficult to characterize the complex geometric deformation relationship between air and ground viewpoints. Furthermore, they are prone to introducing background noise interference, resulting in limited alignment effects.

[0005] In summary, existing technologies in AG-ReID tasks consistently struggle to strike a balance between effectively mitigating cross-view feature distribution shifts and fully preserving key identity-discriminating information. Furthermore, current methods fail to adequately mine implicit discriminative information within view-specific features, neglecting to fully utilize identity-related cues within view-specific information, thus limiting the model's cross-view robustness and generalization ability. A key technical challenge urgently needing to be addressed in the field of air-to-ground cross-view pedestrian re-identification is how to mitigate air-to-ground cross-view feature distribution shifts while simultaneously mining and compensating for the implicit discriminative information inherent in view-specific features, thereby constructing pedestrian feature representations that possess both view robustness and high discriminative power. Summary of the Invention

[0006] This invention provides a cross-view pedestrian re-identification system and method based on implicit information compensation, which solves the technical problems of large feature distribution shift, unstable identity discrimination, insufficient utilization of view information, and limited robustness and generalization ability in existing air-to-ground cross-view pedestrian re-identification methods. It achieves the dual goals of effectively reducing cross-view feature distribution shift and fully preserving key identity discrimination information, significantly improving the discriminativeness, robustness and generalization ability of the model in air-to-ground cross-view scenarios.

[0007] This invention constructs a multi-layer feature guidance mechanism to fully integrate feature information from different time sequences and semantic levels. On this basis, it introduces viewpoint conditional discrimination enhancement and confidence constraints to effectively mine discriminative clues in viewpoint-specific features. Furthermore, through an implicit information compensation strategy, it introduces structural information with weak viewpoint correlation to alleviate the feature distribution differences between aerial and ground views. Thus, while maintaining the ability to distinguish identities, it significantly improves the stability and generalization of cross-view pedestrian re-identification.

[0008] A cross-view pedestrian re-identification system based on implicit information compensation includes: Cross-view pedestrian image acquisition and preprocessing module: used to acquire pedestrian image data from aerial and ground perspectives, perform preprocessing operations such as pedestrian detection, cropping, size normalization and data augmentation, assign identity labels and viewpoint type labels to each image, and build an labeled training dataset. The basic feature extraction module constructs a backbone feature modeling network based on Vision Transformer, divides the preprocessed pedestrian image into image blocks and generates image block feature sequences, introduces classification feature vectors and view feature vectors to construct the initial input feature sequence, and extracts view-invariant identity semantic features and view-specific features as basic features after multi-layer feature modeling. The multi-layer feature guidance module divides the multi-layer feature modeling process of basic feature extraction into three stages: shallow, middle and deep. It integrates shallow spatial structure information and middle and high-level semantic information through two-level cross-level attention guidance, and then outputs a multi-layer feature guidance representation that integrates spatial structure, semantic and perspective condition information through cross-attention condition guidance of viewpoint features. The viewpoint conditional discrimination enhancement module integrates viewpoint-specific features with multi-layer feature-guided representations channel by channel to construct a joint representation. It generates a discriminative enhancement dynamic mask based on adaptive one-dimensional convolution and layer normalization, modulates the viewpoint-specific features through channels to obtain discriminative enhancement features, and constructs a confidence enhancement loss based on prediction uncertainty to constrain the prediction reliability of the enhanced features. The implicit information compensation module performs instance normalization on view-specific features to achieve feature distribution alignment. It generates a structure alignment dynamic mask through adaptive one-dimensional convolution and layer normalization, extracts view-weakly correlated implicit structural features, fuses implicit structural features with discriminative enhancement features and obtains compensated reconstructed features through lightweight reconstruction, and constructs a structural consistency loss based on identity centroid alignment to reduce the difference between cross-view features of the same identity. The pedestrian identity discrimination and model training module inputs the compensated and reconstructed features into the identity discrimination network composed of global pooling layers and fully connected layers, outputs the identity prediction probability, constructs a joint loss function that is a weighted fusion of identity discrimination loss, cross-view triplet constraint loss, confidence enhancement loss and structural consistency constraint loss, and uses gradient descent to train and update the overall model end-to-end. The cross-view pedestrian image acquisition and preprocessing module, basic feature extraction module, multi-layer feature guidance module, viewpoint condition discrimination enhancement module, implicit information compensation module, and pedestrian identity discrimination and model training module work together in sequence to realize the end-to-end pedestrian re-identification process from cross-view pedestrian image input to identity discrimination output.

[0009] A recognition method for a cross-view pedestrian re-identification system based on implicit information compensation includes the following steps: S1: Cross-view pedestrian image acquisition, preprocessing and basic feature extraction: Acquire pedestrian image data from aerial and ground perspectives, perform preprocessing operations and assign identity and perspective labels, input the labeled images into a backbone feature modeling network based on VisionTransformer, and extract perspective-invariant identity semantic features and perspective-specific features as basic features. S2: Multi-layer feature-guided modeling: The basic feature modeling process of S1 is divided into three stages: shallow, medium and deep. The feature fusion from shallow to deep layers is achieved through two-level cross-level attention guidance. Then, through the cross-attention condition guidance of view features, a multi-layer feature-guided representation is obtained. S3: Viewpoint-specific discriminative feature enhancement: The viewpoint-specific features of S1 are fused with the multi-layer feature guidance representation of S2 to construct a joint representation, generate a discriminative enhancement dynamic mask, and modulate the viewpoint-specific features to obtain discriminative enhancement features, and construct a confidence enhancement loss constraint model for training. S4: Implicit Information Compensation and Feature Reconstruction: Instance normalization is performed on the perspective-specific features of S1 to generate a structure-aligned dynamic mask and extract implicit structural features. The implicit structural features are fused with the discriminative enhancement features of S3, and the compensated reconstructed features are obtained through lightweight reconstruction. The structural consistency loss constraint model is then constructed for training. S5: Pedestrian Identity Determination and Joint Loss End-to-End Training: Input the compensated reconstruction features from S4 into the identity determination network to obtain the identity prediction probability. Construct a joint loss function that is a weighted fusion of identity determination loss, cross-view triplet constraint loss, confidence enhancement loss, and structural consistency constraint loss. Perform end-to-end training and updates on the overall model until the model converges to obtain the optimal cross-view pedestrian re-identification model.

[0010] Furthermore, S1 specifically includes: S11: Collect pedestrian image data from both aerial and ground perspectives. Aerial perspective images are collected by drones, aerial platforms, or high-altitude monitoring equipment, while ground perspective images are collected by fixed or mobile monitoring cameras. Perform pedestrian detection and cropping on the original images to obtain single pedestrian target image samples. S12: Preprocess the image samples: normalize the image size to 256×128, perform data augmentation operations such as random brightness adjustment, horizontal flipping, random cropping, or random erasing; assign an identity label to each sample. and perspective type tags , From a ground perspective, From an aerial perspective, To determine the total number of identity categories, construct a training dataset. S13: Input the images from the training dataset into the Vision Transformer backbone network, divide the images into several non-overlapping image patches, and linearly map them into image patch feature sequences. , The number of image patches, For feature dimensions; S14: Introduce classification feature vectors into the image patch feature sequence and view feature vector Construct the initial input feature sequence, and construct the initial input feature sequence of the basic network. ; S15: Process the initial input feature sequence Layer feature modeling operations Feature modeling operations include at least self-attention feature interactions and nonlinear mappings; S16: Extract classification feature vectors and viewpoint feature vectors from the final layer output feature sequence as basic features: , Perspective-invariant identity semantic features. Features specific to the viewpoint.

[0011] Furthermore, S2 specifically includes: S21: Multi-stage feature segmentation: This involves dividing S1... The layer feature modeling process is divided into three stages: shallow, medium, and deep, and the feature sequences of image patches at each stage are extracted. , For batch size, The number of image patches, For feature dimensions; S22: First-level feature guidance: using deep features For query features, shallow features Using the reference features, the first-level enhanced features are obtained through cross-attention and feedforward mapping: , For cross-attention calculation, For feedforward mapping processing; S23: Second-level feature guidance: using first-level enhanced features For query features, mid-level features Using the reference features, the second-level guided features are obtained through cross-attention and feedforward mapping. ; S24: Viewpoint Condition Guidance: Specific features from the perspective of S1 To query features, the second level of guiding features Using the reference features, a multi-layer feature-guided representation is obtained through cross-attention and feedforward mapping, combined with residual connections. .

[0012] Furthermore, S3 specifically includes: S31: Construction of Joint Feature Representation: This involves constructing a viewpoint-specific representation from S1. Multi-layer feature-guided representation with S3 Channel-by-channel fusion yields a joint representation: , It is the Sigmoid activation function. This is a channel-by-channel multiplication operation; S32: Discriminative Enhancement Feature Generation: For Joint Representation An adaptive one-dimensional convolution with kernel size k, layer normalization, and sigmoid activation are performed sequentially to generate a discriminative enhancement mask. , To achieve layer normalization, a mask is applied to view-specific features to obtain discriminative enhanced features: ; S33: Confidence Enhancement Loss Construction: Based on Viewpoint-Specific Features and discriminative enhancement features For identity classification prediction, the information entropy function is used. To measure prediction uncertainty, an interval parameter is introduced. Construct confidence-enhancing loss: , It is a monotonically increasing function.

[0013] Furthermore, S4 specifically includes: S41: Feature Distribution Alignment: Viewpoint-Specific Features of S1 Perform instance normalization to obtain normalized features: , For the average statistics, For variance statistics, For smoothing terms; S42: Implicit Structure Feature Extraction: Extraction of Normalized Features An adaptive one-dimensional convolution with kernel size k, layer normalization, and sigmoid activation are performed sequentially to generate a structure-aligned mask. Applying the mask to the normalized features yields implicit structural features: ; S43: Feature Fusion and Reconstruction: Enhancing the Discriminative Features of S3 With implicit structural features The compensation feature is obtained by adding each channel sequentially: Perform fully connected mapping, GELU activation, and layer normalization on the compensated features to obtain the compensated reconstructed features: , For learnable parameters of fully connected layers, Use the GELU activation function; S44: Construction of Structural Consistency Loss: Define the average Euclidean distance between the feature centroids of the same identity from both aerial and ground perspectives as... ,set up , The viewpoint features are normalized from the perspective of the air and ground. , To correspond to implicit structural features, an interval parameter is introduced. Construct structural consistency loss: .

[0014] Furthermore, S5 specifically includes: S51: Constructing an identity discrimination network: Consists of a global pooling layer and a fully connected layer. The compensated and reconstructed features from S4 are transformed into high-dimensional feature vectors through global pooling. These vectors are then input into the fully connected layer, mapped to the class space, and normalized using Softmax. The output is an identity prediction probability vector. , For the first The sample was classified as the first The probability of class identity; S52: Constructing the Identity Discrimination Loss: Based on the identity prediction probability and the real identity label, construct the loss: B represents the batch size. For the first The real identity label of each sample; S53: Constructing Cross-Perspective Triple Constraint Loss: Compressing intra-class distance and amplifying inter-class differences, constructing a loss: , For the first Compensated reconstruction features of each sample, For positive sample features, For negative sample features, For margin parameters, This is a function to find the maximum value. S54: Joint Loss Training: Constructing a joint loss function by weighted fusion of individual losses. , The loss weight hyperparameters are used; the gradient descent method is used to train the overall model end-to-end, and the gradient is backpropagated to update the network parameters until the preset number of training rounds is reached or the loss converges, thus obtaining the optimal model.

[0015] The core innovation of this invention lies in: A multi-layer feature guidance mechanism is proposed to integrate the spatial structure information of the shallow layer, the semantic information of the middle layer and the identity discrimination information of the deep layer across layers. At the same time, adaptive modulation of viewpoint conditions is introduced so that the feature representation takes into account both the fine geometric structure of pedestrians and highly discriminative identity semantics. A viewpoint-conditional discriminative feature enhancement strategy is proposed, which combines adaptive reweighting of feature channels of dynamic masking with confidence constraints based on prediction uncertainty to adaptively enhance feature channels that are highly correlated with identity discrimination, suppress viewpoint noise interference, and improve the reliability and stability of cross-viewpoint identity prediction. An implicit information compensation mechanism is proposed. The distribution of view features is aligned through instance normalization. The implicit structural features with weak view correlation are mined by combining structural alignment dynamic mask. After being fused with discriminative enhancement features, lightweight reconstruction is performed. At the same time, the structural consistency constraint of identity centroid alignment is introduced. Without relying on explicit view alignment, the feature differences of the same identity under the air-ground view are effectively reduced. A multi-loss joint optimization strategy is proposed, which integrates identity discrimination loss, cross-view triplet constraint loss, confidence enhancement loss and structural consistency constraint loss in a weighted manner to achieve joint optimization of the model's identity discrimination capability and cross-view robustness, thus ensuring the model's comprehensive performance in complex air-to-ground cross-view scenarios.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Effectively solve the problem of feature distribution offset between air and ground: Through instance normalization, structure alignment mask and identity centroid alignment constraint of implicit information compensation mechanism, the feature difference between air and ground view is reduced from the two levels of sample distribution and identity centroid, which significantly alleviates the drift between feature spatial domains caused by shooting height, pitch angle and geometric distortion, and improves the cross-view consistency of features. 2. Enhance feature robustness while preserving key identity information: This invention does not simply suppress viewpoint information, but enhances the mining of identity-related clues in viewpoint-specific features through viewpoint condition discrimination, and then introduces viewpoint-weakly related structural features through implicit information compensation, thereby achieving effective utilization of viewpoint information and precise suppression of viewpoint noise, avoiding the loss of identity discrimination information and ensuring high feature discriminability. 3. Significantly improves the reliability and stability of cross-view identity discrimination: The multi-layer feature guidance mechanism integrates complementary feature information from different levels, making the feature representation more comprehensive; the confidence enhancement loss based on prediction uncertainty forces the model to output more certain cross-view identity prediction results, effectively reducing prediction errors under extreme view distortion and improving the inference reliability of the model. 4. The model has strong generalization ability and is suitable for complex air-ground integrated monitoring scenarios: The end-to-end training architecture of this invention does not require explicit geometric transformation or annotation of air-ground perspectives. It achieves adaptive learning and compensation of cross-view features through data-driven methods, which can adapt to air-ground monitoring scenarios with different shooting heights and pitch angles. It also has good robustness to changes in lighting, changes in pedestrian posture, and scale differences. 5. Modular architecture design, easy to deploy and expand: The system of this invention adopts a modular design, with each functional module decoupled and highly collaborative. The parameters of each module (such as feature dimension, convolution kernel size, loss weight, etc.) can be flexibly adjusted according to the needs of the actual monitoring scenario. It can also be seamlessly integrated with existing pedestrian detection and trajectory tracking systems, and has good engineering practicality and scalability.

[0017] The technical solution of this invention can be widely applied to scenarios such as public safety prevention and control in smart cities, integrated air-ground monitoring of airports / stations / parks, and personnel trajectory tracking in emergency response. It provides core cross-view pedestrian re-identification technology support for intelligent video analysis and has important practical application value and industrialization prospects. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the cross-view pedestrian re-identification system and method based on implicit information compensation according to the present invention.

[0020] Figure 2 This is a feature-diversified block structure diagram of the cross-view pedestrian re-identification system and method based on implicit information compensation according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: This embodiment provides an aerial-ground cross-view pedestrian re-identification system based on implicit information compensation, including a cross-view pedestrian image acquisition and preprocessing module, a basic feature extraction module, a multi-layer feature guidance module, a viewpoint condition discrimination enhancement module, an implicit information compensation module, and a pedestrian identity discrimination and model training module. The overall system structure can be referred to... Figure 1 As shown, the modules work together to realize an end-to-end pedestrian re-identification process from cross-view pedestrian image input to pedestrian identity judgment output.

[0023] The cross-view pedestrian image acquisition, preprocessing, and basic feature extraction module is used to acquire pedestrian image data under cross-view conditions and perform pedestrian detection, cropping, and preprocessing operations on the acquired images. Subsequently, based on VisionTransformer, feature modeling is performed on the preprocessed pedestrian images to extract basic feature representations that simultaneously contain pedestrian identity semantic information and shooting view condition information, providing input for subsequent feature guidance and compensation processing.

[0024] The multi-layer feature guidance module is used to perform multi-layer partitioning and progressive guided modeling of features at different network depth stages. Through a cross-layer feature guidance mechanism, this module uses the spatial structure information in shallow features to supplement and constrain mid-to-high-level semantic features, and gradually merges shallow, mid-level and deep features, so that the final feature representation can simultaneously take into account the pedestrian's fine spatial structure information and discriminative high-level semantic information.

[0025] The viewpoint-conditional discrimination enhancement module is used to introduce viewpoint conditions to discriminate and aggregate features based on multi-layer feature guidance. This module adaptively enhances feature channels highly correlated with pedestrian identification through viewpoint-conditional guidance and a dynamic masking mechanism, and improves the reliability and stability of the model's identity prediction results under cross-viewpoint conditions by combining confidence constraints based on prediction uncertainty.

[0026] An implicit information compensation module is used to further reduce the feature distribution differences between aerial and ground-based perspectives. This module aligns the distribution of view-specific features through instance normalization and constructs a structure-aligned dynamic mask to extract implicit structural features with weak view correlation. Subsequently, the implicit structural features are fused with discriminative enhancement features, and a lightweight feature reconstruction operation is performed to generate a more robust feature representation under extreme cross-view conditions. Simultaneously, a structural consistency constraint based on identity centroid alignment is introduced to enhance the feature consistency of the same identity across different perspectives.

[0027] The pedestrian identification and model training module is used to perform identity identification based on the compensated and reconstructed pedestrian features, and to perform end-to-end training and updating of the model by constructing a joint loss function. This module improves the cross-view robustness of the model while ensuring its discrimination ability through joint optimization of identity discrimination loss, cross-view triplet constraint loss, confidence enhancement loss, and structural consistency constraint loss, ultimately obtaining the best-performing cross-view pedestrian re-identification model.

[0028] Example 2: This embodiment provides a cross-view pedestrian re-identification system and method based on implicit information compensation, such as... Figure 2 As shown, it includes the following steps: S1: Cross-view pedestrian image acquisition, preprocessing and basic feature extraction: Acquire pedestrian image data from aerial and ground perspectives, perform preprocessing operations and assign identity and perspective labels, input the labeled images into a backbone feature modeling network based on VisionTransformer, and extract perspective-invariant identity semantic features and perspective-specific features as basic features. First, pedestrian image data under cross-view conditions is acquired, including at least aerial and ground views. Aerial pedestrian images are acquired by drones, aerial platforms, or high-altitude monitoring equipment, while ground pedestrian images are acquired by fixed or mobile monitoring cameras. The acquired raw pedestrian images or video data are parsed and processed, and a pedestrian detection algorithm is executed to locate pedestrian regions in the images. The detected pedestrian regions are then cropped to obtain pedestrian image samples containing only a single pedestrian target. Subsequently, preprocessing operations are performed on the pedestrian image samples. These preprocessing operations include image size normalization, pixel normalization, and data augmentation operations such as brightness perturbation, random cropping, or horizontal flipping to reduce the impact of illumination changes, scale differences, and pose changes on the feature modeling process. In this example, the preset resolution is [resolution value missing]. After preprocessing, a corresponding identity label is assigned to each pedestrian image sample. and perspective tags The preprocessed pedestrian images are then input into the feature modeling network. Extracting basic feature representations from pedestrian images and The basic features include both semantic information related to pedestrian identity and perspective condition information introduced by the shooting perspective, providing input for subsequent feature guidance and compensation processing.

[0029] Preferably, in step S1, the acquisition, preprocessing, and feature extraction of pedestrian images from multiple viewpoints are achieved through a feature modeling network based on VisionTransformer. The specific steps are as follows: First, pedestrian image data under cross-view conditions is collected using diverse camera equipment. The cross-view conditions include at least aerial and ground views with significant geometric distortion and perspective differences. Aerial pedestrian images are acquired from a high-angle view using drones, aerial platforms, or high-altitude monitoring equipment, while ground pedestrian images are acquired using fixed or mobile monitoring cameras. Let the set of aerial pedestrian images be denoted as . The set of pedestrian images from a ground perspective is denoted as ; in, Indicates the first Zhang's aerial view of pedestrians. Indicates the first Zhang's ground-view pedestrian images and These represent the number of pedestrian images from the corresponding viewpoints. After parsing and processing the acquired raw video or image data, a pedestrian detection algorithm is executed to locate pedestrian regions. The detection results are then cropped to obtain pedestrian image samples containing only a single human target. Subsequently, preprocessing operations are performed on these image samples, including normalizing the image size to a preset resolution and performing data augmentation operations such as random brightness adjustment, horizontal flipping, random cropping, or random erasing as needed to reduce the impact of lighting changes, scale differences, and pose changes on the feature modeling process. After preprocessing, a corresponding identity label is assigned to each pedestrian image sample. And viewpoint type tags; in, This indicates the total number of pedestrian identity categories. Indicates the first The image is from a ground-level perspective. Indicates the first The images are taken from an aerial perspective, thus constructing a training dataset with identity and viewpoint annotations.

[0030] After preprocessing and annotation, pedestrian image samples are input into a backbone feature modeling network built on the Vision Transformer to uniformly model local structural cues and high-level semantic representations of pedestrian images across different viewpoints. Specifically, the base network first processes the input pedestrian images through image-blocking and embedding substructures. The image is divided into several non-overlapping blocks, and a linear mapping is performed on each block to obtain the corresponding image block feature vector, forming an image block feature sequence: in, Indicates the number of image patches. This represents the feature dimension. Based on the aforementioned feature sequence, to explicitly introduce global semantic modeling capabilities and viewpoint condition modeling capabilities, two types of learnable features are introduced into the feature sequence: classification features used to aggregate the overall semantic representation of pedestrians. And the view feature vector used to characterize the shooting view conditions. Therefore, the initial input feature sequence of the basic network is constructed as follows: Subsequently, the initial feature sequence is input into the backbone feature modeling network, and the feature sequence is updated layer by layer along the network depth direction. Then the... The feature update process of a layer is represented as follows: in, Indicates the first Layer feature modeling operations, which include at least feature interaction modeling based on a self-attention mechanism and nonlinear mapping processing. Indicates the number of feature modeling layers.

[0031] In the feature modeling process, the classification features Through global interaction with image patch features, high-level semantic cues related to pedestrian identity but not dependent on specific viewpoints are aggregated layer by layer; the viewpoint feature vector By participating in the feature interaction process, it explicitly characterizes the viewpoint-specific feature changes introduced by the difference between aerial and ground viewpoints, providing a basic representation for subsequent viewpoint condition guidance and information compensation.

[0032] go through After layer feature modeling, classification feature vectors and viewpoint feature vectors are extracted from the feature sequence output by the final layer, respectively, as basic feature representations: in, This represents view-invariant feature representations that contain semantic clues related to pedestrian identification. This represents the perspective features that are closely related to the shooting perspective conditions.

[0033] Through the above-mentioned basic network construction and feature modeling process, a structured basic feature representation is obtained for subsequent multi-layer feature guidance, viewpoint conditional feature aggregation, and viewpoint-specific feature implicit compensation processing.

[0034] S2: Multi-layer feature-guided modeling: The basic feature modeling process of S1 is divided into three stages: shallow, medium and deep. The feature fusion from shallow to deep layers is achieved through two-level cross-level attention guidance. Then, through the cross-attention condition guidance of view features, a multi-layer feature-guided representation is obtained. After completing the basic feature extraction described in S1, the feature modeling network is partitioned in the depth direction. Each stage. In this example, =3. Specifically, the features in the network are divided into shallow features, mid-level features, and deep features according to semantic hierarchy. Shallow features mainly represent the local spatial structure and geometric details of pedestrian images, mid-level features take into account both structural and semantic expression, and deep features focus on depicting high-level semantic information highly related to pedestrian identity discrimination. Based on this, a multi-layer feature guidance mechanism is introduced. Through cross-level feature interaction, the spatial structure information contained in the shallow features is progressively introduced into the mid- and high-level features, supplementing and constraining the deep semantic features. Through this progressive feature guidance process, effective association and fusion between features at different depth stages are achieved, resulting in a multi-layer feature-guided representation that integrates spatial structure information and identity semantic information. This provides a more sufficient feature foundation for subsequent modeling of viewpoint condition features.

[0035] Preferably, in step S2, the multi-layer feature-guided representation is achieved through multi-stage feature partitioning and a progressive feature-guided mechanism, with the specific steps as follows: S21. Multi-stage feature segmentation and hierarchical modeling. After completing the feature extraction described in S1, in order to fully explore the complementary information contained in the features of different network depth stages, the features described in S1 are segmented and hierarchically modeled. The layer feature modeling process is divided into layers according to the network depth direction. The three stages correspond to the shallow feature modeling stage, the medium feature modeling stage, and the deep feature modeling stage, respectively. Let the image patch feature sequences extracted in each stage be denoted as: in, Indicates batch size, Indicates the number of image patches. This represents the feature dimension. Because features exhibit a progressively abstract semantic evolution characteristic along the network depth direction, shallow features are mainly used to characterize the local spatial structure and geometric details of pedestrian images, mid-level features take into account both local structure and mid-level semantic expression, while deep features focus on depicting high-level abstract semantic information highly relevant to pedestrian identification. This multi-stage feature division lays the foundation for subsequent cross-level feature guidance and fusion.

[0036] S22. First-level feature guidance processing based on shallow features. To achieve cross-level association and collaborative modeling between features at different stages, this step adopts a progressive feature guidance mechanism, utilizing the spatial structure information contained in lower-level features to supplement and constrain higher-level features. Specifically, the first-level feature guidance processing is executed first. This is based on deeper features... As query features, shallow features As a reference feature, a long-range correlation between the deep semantic representation and the shallow geometric structure is established through cross-attention computation. After cross-attention computation, feedforward mapping is combined to perform non-linear updates on the features, thereby obtaining the first-level enhanced feature representation: in, This represents the cross-attention calculation process, used to incorporate geometric cues from shallow features. This indicates feedforward mapping, used to enhance the nonlinear expressive power of features.

[0037] S23. Second-level feature guidance processing based on mid-level features. After completing the first-level feature guidance processing described in S22, a second-level feature guidance processing is performed to further integrate mid-level semantic information and refine feature representation. Specifically, the first-level enhanced features are... As the query features at the current stage, using mid-level features As reference features, the features are further integrated and refined through cross-attention calculation and feedforward mapping to obtain the second-level guided feature representation: Through the aforementioned step-by-step feature guidance and alignment process, features at different depth stages are progressively fused in the feature space, enabling the final feature representation to simultaneously take into account the pedestrian's detailed spatial structural information and discriminative identity semantic information.

[0038] S24. Conditional guidance based on viewpoint features. After completing the multi-stage feature guidance processing described in S23, a multi-layered feature guidance representation integrating shallow spatial structure information and mid-to-high-level semantic information has been obtained. Building upon this, to further adapt to the differences in feature distribution under different shooting angles, viewpoint-specific features are introduced to conditionally guide and adaptively refine the multi-layer feature representation.

[0039] Specifically, the view-specific feature vectors obtained during the feature extraction process in S1 are denoted as... This is used to characterize the shooting perspective information corresponding to the pedestrian image. Based on the specific features of the stated perspective... As query features, the multi-layered features obtained in S23 guide the representation. As a reference feature, a correlation between viewpoint features and multi-layered guided features is established through cross-attention computation, thereby adaptively aggregating feature information related to pedestrian identification under different shooting viewpoint conditions. After completing the cross-attention computation, feedforward mapping processing is further performed on the updated features, and the viewpoint features are updated using a residual connection method to obtain the feature representation guided by viewpoint conditions: Through the above-mentioned conditional guidance processing based on viewpoint features, information highly related to pedestrian identity determination in the multi-layer feature guidance representation is adaptively selected and enhanced under different viewpoint conditions, thereby effectively alleviating the feature inconsistency problem caused by the difference between aerial and ground viewpoints.

[0040] Finally, the feature representation guided by the viewpoint conditions is... As the output feature of S2, it is used for subsequent steps such as feature aggregation based on viewpoint conditions, implicit information compensation, and pedestrian identity discrimination.

[0041] S3: Viewpoint-specific discriminative feature enhancement: The viewpoint-specific features of S1 are fused with the multi-layer feature guidance representation of S2 to construct a joint representation, generate a discriminative enhancement dynamic mask, and modulate the viewpoint-specific features to obtain discriminative enhancement features, and construct a confidence enhancement loss constraint model for training. After completing the multi-layer feature-guided modeling described in S2, viewpoint conditions are introduced to enhance the discriminative power of viewpoint-specific features, in order to further uncover discriminative clues in viewpoint-specific features that are highly correlated with pedestrian identity determination. Specifically, let the viewpoint-specific features extracted in S1 be represented as... The multi-layer feature guidance obtained in S2 is represented as First, the multi-layer feature-guided representation. Perform sigmoid activation and apply it as a guiding weight to view-specific features. Joint feature representation is constructed through a channel-by-channel fusion method. This allows the viewpoint features to undergo adaptive reweighting at the semantic level before enhancement.

[0042] Based on this, and using the joint feature representation Construct a discriminative enhancement dynamic mask. Specifically, through one-dimensional convolution and normalization operations, learn the contribution of each feature channel to the identity discrimination task from the channel dimension, and generate the corresponding discriminative enhancement mask. Then, the mask is applied to specific features of the original viewpoint. The system performs targeted enhancement on feature channels that are highly correlated with pedestrian identification, while suppressing noisy features that are significantly affected by viewpoint changes or have weak discriminative ability, thereby obtaining a discriminatively enhanced viewpoint feature representation. Furthermore, to further improve the predictive reliability of enhanced features under cross-view conditions, a confidence constraint based on prediction uncertainty is introduced into the discriminative enhancement process. By comparing the uncertainties of the original viewpoint features and the enhanced features in the identity prediction task, the model is guided to prioritize optimizing discriminative enhancement features during training. This enables the model to output more certain and stable identity discrimination results in extreme viewpoint change scenarios such as open fields and ground. Through this process, while maintaining the integrity of identity discrimination information, the discrimination stability of viewpoint-specific features under cross-viewpoint conditions is significantly improved, providing a more reliable feature foundation for subsequent implicit information compensation processing.

[0043] Preferably, in step S3, feature guidance and aggregation based on viewpoint conditions are jointly achieved through discriminative feature enhancement processing and confidence enhancement constraints. The specific steps are as follows: S31. Discriminative Feature Enhancement Processing Based on Viewpoint Conditions and Multi-layer Guidance. To dynamically mine discriminative cues highly correlated with pedestrian identity determination from viewpoint-specific features, this method jointly utilizes viewpoint features and multi-layer feature guidance representations obtained in S2 to perform discriminative feature enhancement processing on viewpoint features.

[0044] Specifically, let the view-specific features extracted in S1 be denoted as... The multi-layer feature-guided representation output by S2 is denoted as First, regarding the multi-layered guidance representation... An activation function is applied to obtain normalized guiding weights, which are then fused with viewpoint features channel by channel to construct a joint representation. in, This represents the Sigmoid activation function. This represents a channel-wise multiplication operation, and through this joint representation, the viewpoint features are guided by semantic reweighting before feature aggregation.

[0045] S32. Construct a discriminatively enhanced dynamic mask. Subsequently, based on the joint representation... This paper utilizes adaptive one-dimensional convolution and layer normalization techniques to construct a dynamic mask, aiming to capture discriminative identity features from the channel dimension. Specifically, it addresses the joint representation... By sequentially performing one-dimensional convolution, normalization, and non-linear activation mapping, a discriminative enhancement mask is obtained: in, Indicates the kernel size as Adaptive one-dimensional convolution operation, The representation layer undergoes normalization. After obtaining the discriminative enhancement mask, it is applied to the original viewpoint features. Channel-by-channel modulation of the viewpoint features yields a discriminatively enhanced feature representation: Through the above discriminative enhancement processing, the feature channels related to pedestrian identity discrimination in the viewpoint features are enhanced in a targeted manner, while background or geometric noise information that is severely affected by viewpoint changes and has weak discrimination ability is suppressed, thereby significantly improving the expressive ability of features in extreme cross-viewpoint environments.

[0046] S33. Construction of Confidence Enhancement Loss Based on Prediction Uncertainty. To further constrain the reliability and stability of enhanced features in identity discrimination tasks, this method introduces a confidence enhancement loss based on prediction uncertainty to guide the model training process.

[0047] Specifically, based on the original viewpoint features Features after discriminative enhancement Identity classification prediction is performed to obtain the corresponding classification prediction results. The uncertainty of the prediction results is measured using the information entropy function, denoted as prediction entropy. To ensure that discriminative enhancement features can produce more deterministic and reliable prediction results under cross-view conditions, an interval parameter is introduced. And construct a confidence enhancement loss function: in, It is a monotonically increasing function, which effectively avoids the numerical overflow problem in the optimization process and improves training stability.

[0048] By constraining the confidence enhancement loss as described above, the model continuously optimizes the discriminative enhancement branch during training, enabling it to achieve significantly reduced prediction uncertainty compared to the original features when facing extreme viewpoint distortion. This fundamentally improves the system's discriminative confidence and robustness in air-to-ground retrieval tasks.

[0049] S4: Implicit Information Compensation and Feature Reconstruction: Instance normalization is performed on the perspective-specific features of S1 to generate a structure-aligned dynamic mask and extract implicit structural features. The implicit structural features are fused with the discriminative enhancement features of S3, and the compensated reconstructed features are obtained through lightweight reconstruction. The structural consistency loss constraint model is then constructed for training. First, specific characteristics of the viewpoint Instance normalization is performed to suppress global statistical offsets introduced by differences in shooting perspective while preserving the relative structural relationships between feature channels. Based on the normalized features, a structure-aligned dynamic mask is constructed. The implicit structural information that is robust to changes in viewpoint is selectively extracted. Then, the extracted implicit structural features are fused with discriminatively enhanced features, and a compensated pedestrian feature representation is generated through lightweight feature reconstruction. Through the aforementioned implicit information compensation processing, while preserving pedestrian identification information, the difference in feature distribution between aerial and ground perspectives is effectively reduced, improving the stability and consistency of features in complex cross-view scenarios.

[0050] Preferably, in step S4, the implicit information compensation for viewpoint-specific features is achieved jointly through feature distribution alignment, structural alignment modeling, feature fusion reconstruction, and structural consistency constraints. The specific steps are as follows: S41. Alignment of view-specific feature distribution based on instance normalization. To address the feature distribution shift caused by viewpoint changes, this method first performs instance-level normalization on the viewpoint-specific features in S3 to suppress statistical information strongly correlated with viewpoint, while maintaining the structural relationship between channels.

[0051] Specifically, let the view-specific features obtained in S3 be represented as Instance normalization is then performed on the feature to obtain the normalized feature representation: in, and These represent the mean and variance statistics performed on the features, respectively. This is a smoothing term.

[0052] Through the above example normalization process, global statistical information strongly correlated with viewpoint is effectively suppressed, while maintaining the relative structural relationship between channels, providing a stable distribution basis for subsequent implicit feature extraction.

[0053] S42. Construct a structure-aligned dynamic mask. This is done after obtaining the normalized feature representation. Subsequently, in order to selectively retain implicit information related to the pedestrian identity structure, this method further constructs a structure alignment mask and performs weighted filtering on the normalized features.

[0054] Specifically, for normalized features Adaptive one-dimensional convolution processing, layer normalization processing, and activation mapping are performed sequentially to generate a structure alignment mask: in, Indicates the kernel size as Adaptive one-dimensional convolution operation, Presentation layer normalization processing, This represents the Sigmoid activation function. Subsequently, this mask is applied to the normalized features. By weighting the features channel by channel, we obtain a viewpoint-weakly correlated implicit structural feature representation: The above processing can effectively capture implicit geometric cues that are robust to changes in viewpoint, and significantly reduce the impact of viewpoint interference on identity determination.

[0055] S43: Implicit Information Compensation Fusion and Feature Reconstruction. This involves obtaining discriminatively enhanced features. With implicit structure invariant features Subsequently, this method further fuses the two types of features to achieve implicit information compensation for viewpoint-specific features.

[0056] Specifically, the discriminative enhancement features obtained in S3 are represented The implicit structural feature representation obtained in S42 By performing channel-by-channel addition and fusion, the compensated feature representation is obtained: This fusion method, while retaining identity-related information in the discriminative enhancement features, adaptively introduces implicit structural information with weak viewpoint correlation, thereby constructing a more robust feature representation under extreme aerial and ground viewpoint changes. To ensure the compensated feature representation is compatible with the subsequent Transformer-based feature modeling process, this method further performs lightweight feature reconstruction processing on the compensated features. Specifically, the compensated features are transformed through fully connected mapping and nonlinear activation functions, combined with layer normalization, to obtain the reconstructed feature representation: in, This represents the learnable parameters of the fully connected layer. This represents the GELU activation function. The reconstructed features not only retain the essential discriminative power of the original viewpoint features, but also possess better structural consistency, making them suitable as high-quality input for subsequent re-identification decisions.

[0057] S44. Structural Consistency Constraints Based on Identity Centroid Alignment. In addition to sample-level discriminative and confidence constraints, to further enhance the consistency of the same identity feature structure under cross-perspective conditions, this method introduces structural consistency constraints based on identity centroid alignment in the feature space.

[0058] Specifically, in the feature space In this context, the average Euclidean distance between the feature centroids of the same identity from both aerial and ground-based perspectives is defined as: .set up and These represent the viewpoint features after instance normalization from aerial and ground perspectives, respectively. and Let each represent a corresponding implicit structural feature, then construct the centroid alignment loss function: This loss function explicitly reduces structural mismatch caused by differences in shooting height and angle by forcibly reducing the distance between the centroids of the viewpoints in the compensated feature space. Through the multi-level compensation processing from sample distribution alignment to identity centroid alignment, this method can effectively extract and utilize implicit discriminative signals in viewpoint-specific features, thereby significantly improving the matching accuracy and generalization ability of the pedestrian re-identification model in complex open-air surveillance scenarios.

[0059] S5: Pedestrian Identity Determination and Joint Loss End-to-End Training: Input the compensated reconstruction features from S4 into the identity determination network to obtain the identity prediction probability. Construct a joint loss function that is a weighted fusion of identity determination loss, cross-view triplet constraint loss, confidence enhancement loss, and structural consistency constraint loss. Perform end-to-end training and updates on the overall model until the model converges to obtain the optimal cross-view pedestrian re-identification model.

[0060] After implicit information compensation, the compensated pedestrian features are input into the pedestrian identity discrimination network to classify and predict pedestrian identities. During the model training phase, a joint loss function is constructed. The entire network is trained and updated end-to-end, and the joint loss function comprehensively considers identity discrimination constraints. Cross-perspective feature differentiation constraints Confidence enhancement constraints and structural consistency constraints The model is guided to improve its robustness across different viewpoints while maintaining its ability to identify individuals. The network parameters are iteratively updated using the backpropagation algorithm until the model converges, ultimately yielding a pedestrian re-identification model suitable for complex cross-viewpoint surveillance scenarios.

[0061] Preferably, in step S5, pedestrian identification and model training updates are achieved through the construction of an identity identification network, identity identification loss, cross-view constraint loss, and joint optimization objective function. The specific steps are as follows: S51. Construct a pedestrian identification network. After completing the cross-view pedestrian feature extraction, multi-layer feature guidance, view condition discrimination enhancement, and implicit information compensation processing described in S1 to S4, the compensated reconstructed features output from S4 are used as the input to the pedestrian identification network.

[0062] The network mainly consists of a global pooling layer and a fully connected layer.

[0063] In the specific implementation process, firstly, global pooling is performed on the compensated and reconstructed features to transform the feature map into a high-dimensional feature vector, which serves as the feature representation for each pedestrian image. Then, this feature vector is input into a fully connected layer to map to the class space, and probability normalization is performed using the Softmax function to output the pedestrian identity prediction probability. Let the identity prediction probability vector corresponding to a pedestrian sample be: in, Indicates the first The pedestrian sample was identified as the first... The probability of each identity category, where C represents the total number of pedestrian identity categories in the training dataset.

[0064] S52. Construct the identity discrimination loss. Based on the identity classification probability results obtained in S51, construct the pedestrian identity discrimination loss to constrain the model's ability to discriminate pedestrian identities. in, Indicates the first Each pedestrian sample was correctly classified as its true identity label. The probability is given by B, where B represents the batch size.

[0065] S53. Constructing a cross-view triplet constraint loss. To address the distribution differences between aerial and ground views, and to further compress intra-class distances and amplify inter-class differences, a cross-view triplet constraint loss is introduced into the feature space: in, Indicates the first Compensated reconstruction features of individual pedestrian samples This represents the features of positive samples that share the same identity. This represents the features of negative samples that are different from their identity. For margin parameters, This represents the function that takes the maximum value.

[0066] S54. Construct the overall loss and update the network parameters. Weight and fuse the above loss functions to construct the overall objective function for end-to-end model training: in, To balance the hyperparameters of the various loss weights, gradient descent is used during the training phase to jointly optimize the overall architecture, which includes a feature extraction network, a multi-layer feature guidance module, a viewpoint discrimination enhancement module, an implicit information compensation module, and a pedestrian identification network. The overall loss is calculated according to the above formula, and the gradient is backpropagated to update the network parameters until a preset number of training epochs is reached or the loss function converges, ultimately obtaining the optimal cross-viewpoint pedestrian re-identification model.

[0067] This invention addresses the problems of large differences in shooting height, severe geometric distortion, and inconsistent feature distribution between aerial and ground perspectives. By introducing an implicit information compensation mechanism in the feature modeling stage, it effectively alleviates the interference caused by perspective differences while preserving key information for pedestrian identification, significantly improving the matching accuracy and system robustness in cross-perspective pedestrian re-identification tasks.

[0068] This invention proposes a multi-layer feature-guided feature modeling strategy. By dividing features at different network depth stages into multiple layers and progressively guiding modeling, it fully integrates shallow spatial structure information and mid-to-high-level semantic information. This enables the final feature representation to simultaneously take into account the pedestrian's fine geometric structure features and high-level identity semantic information with discriminative capabilities, thereby effectively improving the feature expression capability in complex cross-view scenarios.

[0069] This invention proposes a joint constraint mechanism of discriminative feature enhancement and confidence based on viewpoint conditions. By introducing viewpoint-specific features to conditionally modulate multi-layer guiding features and combining dynamic masking strategy to adaptively reweight feature channels, the invention guides the model to output more certain and reliable identity prediction results under cross-viewpoint conditions, thereby significantly improving the discriminative stability and inference reliability of the model under extreme viewpoint change conditions.

[0070] This invention proposes an implicit information compensation and structural consistency enhancement strategy for cross-view scenarios. By combining instance normalization and dynamic structural alignment masking, it extracts view-weakly correlated implicit structural information from view-specific features and fuses it with discriminative enhancement features. At the same time, it introduces a structural consistency constraint based on identity centroid alignment, which effectively reduces the feature distribution differences between different views without relying on explicit view alignment, and significantly enhances the consistency of features and the generalization ability of the model under cross-view conditions.

[0071] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A cross-view pedestrian re-identification system based on implicit information compensation, characterized in that: include: Cross-view pedestrian image acquisition and preprocessing module: used to acquire pedestrian image data from aerial and ground perspectives, perform preprocessing operations such as pedestrian detection, cropping, size normalization and data augmentation, assign identity labels and viewpoint type labels to each image, and build an labeled training dataset. The basic feature extraction module constructs a backbone feature modeling network based on Vision Transformer, divides the preprocessed pedestrian image into image blocks and generates image block feature sequences, introduces classification feature vectors and view feature vectors to construct the initial input feature sequence, and extracts view-invariant identity semantic features and view-specific features as basic features after multi-layer feature modeling. The multi-layer feature guidance module divides the multi-layer feature modeling process of basic feature extraction into three stages: shallow, middle and deep. It integrates shallow spatial structure information and middle and high-level semantic information through two-level cross-level attention guidance, and then outputs a multi-layer feature guidance representation that integrates spatial structure, semantic and perspective condition information through cross-attention condition guidance of viewpoint features. The viewpoint conditional discrimination enhancement module integrates viewpoint-specific features with multi-layer feature-guided representations channel by channel to construct a joint representation. It generates a discriminative enhancement dynamic mask based on adaptive one-dimensional convolution and layer normalization, modulates the viewpoint-specific features through channels to obtain discriminative enhancement features, and constructs a confidence enhancement loss based on prediction uncertainty to constrain the prediction reliability of the enhanced features. The implicit information compensation module performs instance normalization on view-specific features to achieve feature distribution alignment. It generates a structure alignment dynamic mask through adaptive one-dimensional convolution and layer normalization, extracts view-weakly correlated implicit structural features, fuses implicit structural features with discriminative enhancement features and obtains compensated reconstructed features through lightweight reconstruction, and constructs a structural consistency loss based on identity centroid alignment to reduce the difference between cross-view features of the same identity. The pedestrian identity discrimination and model training module inputs the compensated and reconstructed features into the identity discrimination network composed of global pooling layers and fully connected layers, outputs the identity prediction probability, constructs a joint loss function that is a weighted fusion of identity discrimination loss, cross-view triplet constraint loss, confidence enhancement loss and structural consistency constraint loss, and uses gradient descent to train and update the overall model end-to-end. The cross-view pedestrian image acquisition and preprocessing module, basic feature extraction module, multi-layer feature guidance module, viewpoint condition discrimination enhancement module, implicit information compensation module, and pedestrian identity discrimination and model training module work together in sequence to realize the end-to-end pedestrian re-identification process from cross-view pedestrian image input to identity discrimination output.

2. A recognition method for a cross-view pedestrian re-identification system based on implicit information compensation according to claim 1, characterized in that: Includes the following steps: S1: Cross-view pedestrian image acquisition, preprocessing and basic feature extraction: Acquire pedestrian image data from aerial and ground perspectives, perform preprocessing operations and assign identity and perspective labels, input the labeled images into a backbone feature modeling network based on VisionTransformer, and extract perspective-invariant identity semantic features and perspective-specific features as basic features. S2: Multi-layer feature-guided modeling: The basic feature modeling process of S1 is divided into three stages: shallow, medium and deep. The feature fusion from shallow to deep layers is achieved through two-level cross-level attention guidance. Then, through the cross-attention condition guidance of view features, a multi-layer feature-guided representation is obtained. S3: Viewpoint-specific discriminative feature enhancement: The viewpoint-specific features of S1 are fused with the multi-layer feature guidance representation of S2 to construct a joint representation, generate a discriminative enhancement dynamic mask, and modulate the viewpoint-specific features to obtain discriminative enhancement features, and construct a confidence enhancement loss constraint model for training. S4: Implicit Information Compensation and Feature Reconstruction: Instance normalization is performed on the perspective-specific features of S1 to generate a structure-aligned dynamic mask and extract implicit structural features. The implicit structural features are fused with the discriminative enhancement features of S3, and the compensated reconstructed features are obtained through lightweight reconstruction. The structural consistency loss constraint model is then constructed for training. S5: Pedestrian Identity Determination and Joint Loss End-to-End Training: Input the compensated reconstruction features from S4 into the identity determination network to obtain the identity prediction probability. Construct a joint loss function that is a weighted fusion of identity determination loss, cross-view triplet constraint loss, confidence enhancement loss, and structural consistency constraint loss. Perform end-to-end training and updates on the overall model until the model converges to obtain the optimal cross-view pedestrian re-identification model.

3. The cross-view pedestrian re-identification method based on implicit information compensation according to claim 2, characterized in that: S1 specifically includes: S11: Collect pedestrian image data from both aerial and ground perspectives. Aerial perspective images are collected by drones, aerial platforms, or high-altitude monitoring equipment, while ground perspective images are collected by fixed or mobile monitoring cameras. Perform pedestrian detection and cropping on the original images to obtain single pedestrian target image samples. S12: Preprocess the image samples: normalize the image size to 256×128, perform data augmentation operations such as random brightness adjustment, horizontal flipping, random cropping, or random erasing; assign an identity label to each sample. and perspective type tags , From a ground perspective, From an aerial perspective, To determine the total number of identity categories, construct a training dataset. S13: Input the images from the training dataset into the Vision Transformer backbone network, divide the images into several non-overlapping image patches, and linearly map them into image patch feature sequences. , The number of image patches, For feature dimensions; S14: Introduce classification feature vectors into the image patch feature sequence and view feature vector Construct the initial input feature sequence, and construct the initial input feature sequence of the basic network. ; S15: Process the initial input feature sequence Layer feature modeling operations Feature modeling operations include at least self-attention feature interactions and nonlinear mappings; S16: Extract classification feature vectors and viewpoint feature vectors from the final layer output feature sequence as basic features: , Perspective-invariant identity semantic features. Features specific to the viewpoint.

4. The cross-view pedestrian re-identification method based on implicit information compensation according to claim 2, characterized in that: S2 specifically includes: S21: Multi-stage feature segmentation: This involves dividing S1... The layer feature modeling process is divided into three stages: shallow, medium, and deep, and the feature sequences of image patches at each stage are extracted. , For batch size, The number of image patches, For feature dimensions; S22: First-level feature guidance: using deep features For query features, shallow features Using the reference features, the first-level enhanced features are obtained through cross-attention and feedforward mapping: , For cross-attention calculation, For feedforward mapping processing; S23: Second-level feature guidance: using first-level enhanced features For query features, mid-level features Using the reference features, the second-level guided features are obtained through cross-attention and feedforward mapping. ; S24: Viewpoint Condition Guidance: Specific features from the perspective of S1 To query features, the second level of guiding features Using the reference features, a multi-layer feature-guided representation is obtained through cross-attention and feedforward mapping, combined with residual connections. .

5. The cross-view pedestrian re-identification method based on implicit information compensation according to claim 2, characterized in that: S3 specifically includes: S31: Construction of Joint Feature Representation: This involves constructing a viewpoint-specific representation from S1. Multi-layer feature-guided representation with S3 Channel-by-channel fusion yields a joint representation: , It is the Sigmoid activation function. This is a channel-by-channel multiplication operation; S32: Discriminative Enhancement Feature Generation: For Joint Representation An adaptive one-dimensional convolution with kernel size k, layer normalization, and sigmoid activation are performed sequentially to generate a discriminative enhancement mask. , To achieve layer normalization, a mask is applied to view-specific features to obtain discriminative enhanced features: ; S33: Confidence Enhancement Loss Construction: Based on Viewpoint-Specific Features and discriminative enhancement features For identity classification prediction, the information entropy function is used. To measure prediction uncertainty, an interval parameter is introduced. Construct confidence-enhancing loss: , It is a monotonically increasing function.

6. The cross-view pedestrian re-identification method based on implicit information compensation according to claim 2, characterized in that: S4 specifically includes: S41: Feature Distribution Alignment: Viewpoint-Specific Features of S1 Perform instance normalization to obtain normalized features: , For the average statistics, For variance statistics, For smoothing terms; S42: Implicit Structure Feature Extraction: Extraction of Normalized Features An adaptive one-dimensional convolution with kernel size k, layer normalization, and sigmoid activation are performed sequentially to generate a structure-aligned mask. Applying the mask to the normalized features yields implicit structural features: ; S43: Feature Fusion and Reconstruction: Enhancing the Discriminative Features of S3 With implicit structural features The compensation feature is obtained by adding each channel sequentially: Perform fully connected mapping, GELU activation, and layer normalization on the compensated features to obtain the compensated reconstructed features: , For learnable parameters of fully connected layers, Use the GELU activation function; S44: Construction of Structural Consistency Loss: Define the average Euclidean distance between the feature centroids of the same identity from both aerial and ground perspectives as... ,set up , The viewpoint features are normalized from the perspective of the air and ground. , To correspond to implicit structural features, an interval parameter is introduced. Construct structural consistency loss: .

7. The cross-view pedestrian re-identification method based on implicit information compensation according to claim 2, characterized in that: S5 specifically includes: S51: Constructing an identity discrimination network: Consists of a global pooling layer and a fully connected layer. The compensated and reconstructed features from S4 are transformed into high-dimensional feature vectors through global pooling. These vectors are then input into the fully connected layer, mapped to the class space, and normalized using Softmax. The output is an identity prediction probability vector. , For the first The sample was classified as the first The probability of class identity; S52: Constructing the Identity Discrimination Loss: Based on the identity prediction probability and the real identity label, construct the loss: B represents the batch size. For the first The real identity label of each sample; S53: Constructing Cross-Perspective Triple Constraint Loss: Compressing intra-class distance and amplifying inter-class differences, constructing a loss: , For the first Compensated reconstruction features of each sample, For positive sample features, For negative sample features, For margin parameters, This is a function to find the maximum value. S54: Joint Loss Training: Constructing a joint loss function by weighted fusion of individual losses. , The loss weight hyperparameters are used; the gradient descent method is used to train the overall model end-to-end, and the gradient is backpropagated to update the network parameters until the preset number of training rounds is reached or the loss converges, thus obtaining the optimal model.

Images