Air-ground pedestrian re-identification method, system, device and storage medium

Abstract

This invention discloses a ground-to-air pedestrian re-identification method, system, device, and storage medium, which correspond to a crime scene. The solution employs a learnable feature projection mechanism to explicitly model and bridge the significant geometric and appearance gap between aerial and ground-based perspectives, thereby more effectively extracting identity-sensitive and perspective-sensitive feature representations while suppressing interference information introduced by extreme perspective differences. Furthermore, by collaboratively achieving cross-perspective alignment and decoupling at the levels of feature transformation, supervisory signals, and optimization objectives, the model is guided to learn the optimal cross-perspective pedestrian representation, thus maintaining high robustness and strong generalization ability in complex scenarios and accurately achieving ground-to-air pedestrian re-identification. This solution has broad application prospects in scenarios requiring integrated ground-to-air perception, such as urban three-dimensional security, emergency search and rescue, and security for large-scale events.

Description

Technical Field

[0001] This invention relates to the field of pedestrian re-identification technology, and in particular to a ground-to-air pedestrian re-identification method, system, device, and storage medium. Background Technology

[0002] Pedestrian re-identification technology, as a core component of intelligent video surveillance systems, has played an increasingly important role in the field of public safety in recent years. With the deepening of smart city construction and the rapid development of the low-altitude economy, an integrated ground-air collaborative monitoring network, combining fixed ground surveillance cameras with mobile drone cameras, has emerged. This three-dimensional monitoring model can effectively expand the monitoring coverage, eliminate blind spots in traditional ground surveillance, and provide more comprehensive technical support for urban management, emergency command, and public safety. However, while this ground-air collaborative monitoring model brings new opportunities for pedestrian re-identification, it also introduces unprecedented technical challenges.

[0003] The primary challenge in ground-to-air pedestrian re-identification lies in the significant imaging differences between ground and aerial perspectives, such as... Figure 1 As shown, (a) is a ground-view image and (b) is an aerial-view image. It can be seen that the same pedestrian appears drastically different from the ground-view and aerial-view images. Ground cameras typically use eye-level or low-angle overhead views, while drones often use oblique or high-angle overhead views when performing missions. This drastic change in perspective causes the same pedestrian to appear completely different from different angles. Simultaneously, the dynamic changes in altitude during drone flight significantly alter the scale of pedestrian targets in the image, from richly detailed low-altitude images to tiny targets occupying only a few pixels in high-altitude images. This significant scale change places extremely high demands on feature extraction algorithms. Furthermore, the background environment from an aerial perspective is usually more complex and variable, with more severe occlusion problems caused by trees, buildings, etc. Combined with the influence of lighting conditions and weather factors, this further increases the difficulty of implementing ground-to-air pedestrian re-identification technology.

[0004] To address these challenges, researchers have proposed various technical solutions. Early studies primarily employed global feature-based methods, enhancing the feature extraction capabilities of deep neural networks to learn the overall representation of pedestrians. While these methods can capture the macroscopic features of pedestrians to some extent, they remain inadequate in handling drastic changes in ground and aerial perspectives, struggling to maintain feature discriminativeness under extreme viewpoint differences. Subsequent methods based on part alignment attempted to achieve cross-viewpoint local feature matching by locating pedestrian body parts or dividing horizontal stripes. This approach improved the efficiency of utilizing fine-grained features to some extent; however, when the ground and aerial perspectives are nearly perpendicular, establishing the correspondence between body parts becomes difficult, and predefined part division methods struggle to adapt to deformations caused by extreme viewpoint changes.

[0005] In recent years, feature decoupling-based methods have gradually become a research hotspot. These methods attempt to separate pedestrian features into viewpoint-dependent and viewpoint-invariant components, focusing on identity-discriminating features robust to viewpoint changes. While this approach has theoretical advantages, it often struggles to achieve perfect feature decoupling in practical applications. Viewpoint-invariant features may retain viewpoint information, or crucial identity-discriminating information may be lost during the decoupling process. Meanwhile, to alleviate the problem of scarce labeled data, unsupervised and self-supervised learning methods have also begun to be applied in the field of ground-to-air pedestrian re-identification. These methods utilize unlabeled data for training through cross-camera information interaction and spatiotemporal semantic awareness techniques; however, their performance is generally inferior to fully supervised methods, and they are easily affected by domain differences and noise interference in complex ground-to-air scenes.

[0006] In summary, existing methods generally have poor adaptability to extreme perspective differences, making it difficult to effectively address the feature distribution shift caused by drastic perspective changes. Under multiple interferences such as complex backgrounds, severe occlusion, and dynamic scale variations, the discriminative power of features extracted by existing methods is significantly insufficient. Furthermore, the models have weak generalization ability; models trained on specific datasets show significant performance degradation when transferred to new ground-to-air surveillance scenarios. Some methods also suffer from high computational complexity and reliance on large amounts of labeled data, hindering their practical application in real-world ground-to-air collaborative surveillance systems. Therefore, there is an urgent need to develop a novel ground-to-air pedestrian re-identification scheme that can adaptively learn perspective-invariant features, intelligently mine discriminative regions, and maintain high robustness and strong generalization ability in complex scenarios.

[0007] In view of this, the present invention is hereby proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a ground-to-air pedestrian re-identification method, system, device, and storage medium that can maintain high robustness and strong generalization ability in complex scenarios, accurately realize ground-to-air pedestrian re-identification, and has broad application prospects in scenarios that require integrated ground-to-air collaborative perception, such as urban three-dimensional security, emergency search and rescue, and security for large-scale events.

[0009] The objective of this invention is achieved through the following technical solution: A method for re-identifying pedestrians in ground and air includes: A cross-view learning model is constructed, including: a backbone network, a deformable homography transformation module, a cross-view learning strategy module, and a view decoupling contrast loss module; The cross-view learning model is trained by: extracting features from aerial and ground-view images using a backbone network to obtain aerial and ground-view features; implementing a bidirectional and learnable geometric projection between the aerial and ground-view planes using a deformable homography transformation module to obtain projection features corresponding to each aerial and ground-view feature; calculating cross-view identity consistency loss and cross-view contrast loss using a cross-view decoupling contrast loss module, including a view decorrelation loss term calculated based on the distance between pedestrian identity information and the corresponding view prototype in the view features, and a contrast loss term calculated using different view features of the same pedestrian identity; and training the cross-view learning model by combining all calculated losses. After training, features of the aerial and ground-based visual images to be queried are extracted through the backbone network, and ground-to-air pedestrian re-identification is achieved based on the similarity between the features.

[0010] A ground-to-air pedestrian re-identification system, used to implement the aforementioned method, includes: The model building unit is used to build cross-view learning models, including: a backbone network, a deformable homography transformation module, a cross-view learning strategy module, and a view decoupling contrast loss module; The model training unit, used to train the cross-view learning model, includes: extracting features from aerial and ground-view images by the backbone network to obtain aerial and ground-view features; implementing a bidirectional and learnable geometric projection between the aerial and ground-view planes using a deformable homography transformation module to obtain projection features corresponding to each aerial and ground-view feature; calculating cross-view identity consistency loss and cross-view contrast loss using the cross-view learning strategy module, based on the aerial, ground-view features, and corresponding projection features; calculating the view decoupling contrast loss using the view decoupling contrast loss module, including a view decorrelation loss term calculated based on the distance between pedestrian identity information and the corresponding view prototype in the view features, and a contrast loss term calculated using different view features of the same pedestrian identity; and training the cross-view learning model by combining all calculated losses. The ground-to-air pedestrian re-identification unit is used to extract features from the aerial and ground-based visual images to be queried through the backbone network after training, and to achieve ground-to-air pedestrian re-identification based on the similarity between the features.

[0011] A processing device includes: one or more processors; and a memory for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0012] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0013] As can be seen from the technical solution provided by the present invention, by using a learnable feature projection mechanism, the huge geometric and appearance gap between the aerial view (aerial photography view) and the ground-level view (ground photography view) is explicitly modeled and bridged, thereby more effectively extracting feature representations that are sensitive to identity and robust to viewpoint changes, while suppressing interference information introduced by extreme viewpoint differences. Furthermore, by using a deformable homography transformation module, a cross-viewpoint learning strategy module, and a viewpoint decoupling contrast loss module, cross-viewpoint alignment and decoupling are achieved collaboratively at the levels of feature transformation, supervision signals, and optimization objectives, jointly guiding the model to learn the optimal cross-viewpoint pedestrian representation, thereby accurately achieving ground-air pedestrian re-identification. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0015] Figure 1 This is a schematic diagram illustrating the differences in pedestrian images from a ground-to-air camera, provided as part of the background technology of this invention.

[0016] Figure 2 A flowchart of a ground-to-air pedestrian re-identification method provided in an embodiment of the present invention.

[0017] Figure 3 This diagram illustrates the comparison between the coarse-grained global projection and fine-grained local optimization methods provided in this embodiment of the invention and the conventional homography transformation.

[0018] Figure 4 This is a schematic diagram illustrating the visualization effect of the cross-view feature representation learned by the cross-view learning strategy provided in this embodiment of the invention.

[0019] Figure 5 This diagram illustrates the perspective decoupling contrast loss, standard identity classification cross-entropy loss, and bounded triplet loss provided for embodiments of the present invention.

[0020] Figure 6 This is a schematic diagram of a ground-to-air pedestrian re-identification system provided in an embodiment of the present invention.

[0021] Figure 7 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0022] 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 protection scope of the present invention.

[0023] First, the following explanations are provided for the terms that may be used in this article: The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0024] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0025] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0026] The following is a detailed description of the ground-to-air pedestrian re-identification method, system, device, and storage medium provided by the present invention. Contents not described in detail in the embodiments of the present invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of the present invention, conventional conditions in the art or conditions recommended by the manufacturer shall apply. Instruments used in the embodiments of the present invention whose manufacturers are not specified are all conventional products that can be purchased commercially.

[0027] Example 1 like Figure 2 The flowchart shown is a method for re-identifying pedestrians in the air and on the ground according to an embodiment of the present invention, which mainly includes the following steps: Step 1: Construct a cross-perspective learning model.

[0028] In this embodiment of the invention, the constructed cross-view learning model mainly includes: a backbone network, a deformable homography transformation module, a cross-view learning strategy module, and a view decoupling contrast loss module.

[0029] The backbone network is primarily responsible for extracting image features and can be implemented using existing network structures as needed; for example, VIT (Vision Transformer) can be used as the backbone network. The deformable homography transformation module is mainly responsible for bidirectional, learnable geometric projection between the aerial and ground-based planes. The cross-view learning strategy module utilizes extracted viewpoint features and corresponding projection features to enhance identity consistency and maintain viewpoint invariance. The viewpoint decoupling contrast loss module actively suppresses viewpoint-specific biases in the features through a viewpoint prototype mechanism.

[0030] Step 2: Train the cross-perspective learning model.

[0031] The entire training process includes: extracting features from aerial and ground-view images using a backbone network to obtain aerial and ground-view features; implementing a bidirectional and learnable geometric projection between the aerial and ground-view planes using a deformable homography transformation module to obtain the projection features corresponding to each aerial and ground-view feature; calculating cross-view identity consistency loss and cross-view contrast loss using a cross-view learning strategy module with aerial, ground-view features and corresponding projection features; calculating the view decoupling contrast loss using a view decoupling contrast loss module, which includes a view decorrelation loss term calculated based on the distance between pedestrian identity information and the corresponding view prototype in the view features, and a contrast loss term calculated using different view features of the same pedestrian identity. This process implicitly encourages the model to learn view-independent identity information by pushing away the view prototype and bringing the view features of the pedestrian identity information closer; and finally, training the cross-view learning model by combining all the calculated losses.

[0032] The preferred implementation methods for each module's processing are described below.

[0033] (1) Deformable homography transformation module.

[0034] This module is primarily responsible for enabling bidirectional and learnable geometric projection between the aerial view plane and the ground view plane.

[0035] For aerial viewpoint features, a homography matrix is ​​predicted using a regression network. This homography matrix defines a global projective transformation from the aerial viewpoint plane to the ground viewpoint plane. The aerial viewpoint features are then globally projected using this homography matrix and normalized using perspective division. A deformable offset field is predicted and applied to the normalized global projection result. Finally, differentiable bilinear interpolation sampling is used to obtain the projection features of the aerial viewpoint features onto the ground viewpoint plane. Based on the same principle, for ground-view features, the projection features of ground-view features onto the aerial view plane are obtained by predicting the inverse matrix of the homography matrix and the corresponding learnable deformable offset field.

[0036] In this embodiment of the invention, the deformable offset field is generated through an attention-based aggregation mechanism: a set of candidate spatial offsets and corresponding attention weights are predicted, and the deformable offset field is obtained by calculating the weighted average of the candidate spatial offsets.

[0037] (2) Cross-perspective learning strategy module.

[0038] This module uses aerial viewpoint features, ground viewpoint features, and their respective projection features to calculate cross-view identity consistency loss and cross-view contrast loss.

[0039] For cross-view identity consistency loss, the difference between the aerial view features of the same pedestrian and the projection features of the corresponding ground view plane, as well as the difference between the ground view features of the same pedestrian and the projection features of the corresponding aerial view plane, are used to calculate the loss. For cross-view contrast loss, the Euclidean distance between positive sample pairs of the same pedestrian and the Euclidean distance between negative sample pairs that do not belong to the same pedestrian are used to calculate the loss using the relationship between the Euclidean distance between them and the preset boundary threshold.

[0040] Among them, positive sample pairs of the same pedestrian include: aerial view features and ground view features of the same pedestrian; negative sample pairs include: negative sample pairs are aerial view features of pedestrian A and ground view features of other pedestrians, as well as ground view features of pedestrian A and aerial view features of other pedestrians.

[0041] (3) Perspective decoupling contrast loss module.

[0042] This module is responsible for calculating the viewpoint decoupling contrast loss, which includes two loss terms: a viewpoint decorrelation loss term calculated based on the distance between the pedestrian identity information in the viewpoint features and the corresponding viewpoint prototype, and an identity alignment loss term calculated using different viewpoint features of the same pedestrian identity.

[0043] (3.1) For aerial and ground views, the average vector of all features under each view is used as the corresponding view prototype. Both aerial and ground view features contain pedestrian identity information. The view decorrelation loss term is calculated based on the distance between the aerial view features and the aerial view prototype, and the distance between the ground view features and the ground view prototype, in order to maximize the distance between each view feature and its respective view prototype.

[0044] (3.2) The identity alignment loss term is a contrastive loss based on the idea of ​​noise contrastive estimation. It is calculated by using the similarity between the features of different perspectives of the same pedestrian identity to bring the features of different perspectives of the same pedestrian identity closer together. The above two loss terms together constitute the perspective decoupling contrastive loss, which achieves perspective-independent identity feature learning by balancing perspective decorrelation and identity alignment.

[0045] Since the contrast loss based on the idea of ​​noise contrast estimation is an existing technology, it will not be elaborated here.

[0046] Based on this, in this embodiment of the invention, a total training loss is constructed by integrating cross-view identity consistency loss, cross-view contrast loss, view decoupling contrast loss, and additionally calculated standard identity classification cross-entropy loss and bounded triplet loss, and configuring corresponding weight coefficients. The cross-view learning model is then trained using this total training loss. Considering that the subsequent training methods can be implemented with reference to conventional techniques, this invention will not elaborate further.

[0047] Those skilled in the art will understand that both viewpoint features contain pedestrian identity information in the viewpoint image. After projection and a series of loss constraints, the viewpoint features output by the backbone network can simultaneously possess identity discriminativeness and cross-viewpoint consistency, and can also be referred to as pedestrian identity features.

[0048] Step 3: Achieve ground-to-air pedestrian re-identification based on the similarity between features extracted from the trained backbone network.

[0049] In this embodiment of the invention, after training is completed, three modules in the model can be discarded, and only the backbone network can be retained; the features of the aerial and ground-based visual images to be queried are extracted through the backbone network, and ground-to-air pedestrian re-identification is achieved based on the similarity between the features.

[0050] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0051] I. Overall Overview of the Plan

[0052] This invention provides a ground-to-air pedestrian re-identification method, which aims to learn a feature representation that is robust to changes in perspective and highly sensitive to identity by explicitly modeling and narrowing the huge visual gap between aerial and ground perspectives, thereby solving the problem of performance degradation caused by cross-perspective differences in aerial-to-ground pedestrian re-identification.

[0053] The core innovation of this invention lies in three tightly coupled key components: first, a designed Deformable Homography Transformation (DHT) module for bidirectional, learnable geometric projection between aerial and ground-based feature planes; second, a proposed Cross-View Learning (CVL) module that utilizes original features and DHT projected features to enhance identity consistency while maintaining viewpoint invariance; and third, a constructed View Decorrelation Contrastive Loss (VDCL) module that actively suppresses viewpoint-specific biases in features through a viewpoint prototyping mechanism. Through the collaborative work of these modules, the model is guided to learn the optimal cross-view pedestrian representation.

[0054] II. Detailed introduction of the plan.

[0055] Since the backbone network involved in this invention is a feature extraction network, existing network structures can be selected as needed for implementation, so they will not be elaborated upon. The following sections will provide a detailed introduction to the three models mentioned above.

[0056] 1. Deformable homography transformation module.

[0057] To address the severe geometric distortion and spatial layout misalignment issues between aerial and ground-based images, this invention designs a Deformable Homography Transform (DHT) module. This module adaptively projects features from one viewpoint onto the geometric plane of another viewpoint, providing an alignment foundation for subsequent cross-view learning. The features referred to here mainly consist of feature maps extracted by the backbone network. This module comprises two core sub-operations.

[0058] (1) Coarse-grained global projection. Given an aerial view feature map, the deformable homography transformation module predicts a 3x3 homography matrix through a lightweight regression network. This homography matrix defines a global projective transformation from the aerial view plane to the ground view plane. For any spatial location on the feature map, its projected coordinates on the ground view plane can be calculated using this homography matrix and normalized by perspective division.

[0059] (2) Fine-grained local optimization. Standard homography transformation can only handle rigid planar transformations and cannot cope with local non-rigid deformations caused by changes in human posture. To address this, a learnable deformable offset field is further introduced into the deformable homography transformation module. This offset field is generated through an attention-based aggregation mechanism: the network simultaneously predicts a set of candidate spatial offsets and their corresponding attention weights, and the weighted average of the candidate offsets is the deformable offset field. The coordinates obtained by global projection are added to the deformable offset field to obtain the final accurate sampling position. Then, through differentiable bilinear interpolation sampling, a precisely aligned ground-view feature map is obtained by sampling from the original feature map (i.e., the given aerial view feature map).

[0060] Similarly, for ground-based camera viewpoint features, by estimating the inverse of the homography matrix and the corresponding learnable deformable offset field, a symmetrical projection from the ground-based camera viewpoint to the aerial camera viewpoint can be achieved.

[0061] like Figure 3 As shown, this paper demonstrates the significant difference between the coarse-grained global projection and fine-grained local optimization methods proposed in this invention and conventional homography transformation (HT) in handling ground-to-air perspective misalignment; the original features refer to the aerial and ground perspective feature maps extracted by the backbone network.

[0062] 2. Cross-perspective learning strategy module.

[0063] To effectively utilize the projected features (also known as alignment features) generated by the deformable homography transformation module, this invention proposes a cross-view learning (CVL) strategy. This strategy forces the model to learn consistent identity information in the original feature space and the projected feature space through two complementary loss functions.

[0064] The first is Cross-View Identity Consistency Loss (CICL). This loss requires that the original aerial features of the same pedestrian should be as similar as possible to the features projected onto the ground plane, and vice versa. This constraint uses a smoothed L1 (mean absolute error) loss function to measure the difference between the two, ensuring that identity information remains stable during viewpoint changes.

[0065] The second is Cross-View Contrast Loss (CVCL). This loss directly narrows the feature distance between the same identity from different viewpoints (or their projections) in the feature embedding space, while widening the feature distance between different identities. Specifically, for positive sample pairs belonging to the same pedestrian, it minimizes the Euclidean distance between them; for negative sample pairs not belonging to the same pedestrian, it ensures that their distance is greater than a preset boundary threshold. This contrastive learning mechanism significantly enhances the model's discriminative ability.

[0066] like Figure 4 As shown, the visualization effect of the cross-view feature representation learned through the cross-view learning strategy is intuitively demonstrated.

[0067] 3. Perspective decoupling and contrast loss module.

[0068] To further refine feature representations, this invention designs a Viewpoint Decoupling Contrast Loss (VDCL). This loss has two objectives: first, viewpoint decoupling. During training, a dynamically updated memory is maintained to calculate the viewpoint prototypes for both aerial and ground-based perspectives—that is, the average vector of the same viewpoint features for all pedestrians. Specifically, the average vector of aerial viewpoint features for all pedestrians is the aerial viewpoint prototype, and the average vector of ground-based viewpoint features for all pedestrians is the ground-based viewpoint prototype. Through a dedicated loss term (viewpoint decorrelation loss term), the viewpoint features of each pedestrian are actively pushed away from their respective viewpoint prototypes, thereby explicitly suppressing viewpoint-related information.

[0069] The second is the identity alignment term. It employs a contrast loss based on the idea of ​​noise contrast estimation to bring the features of the same pedestrian from different perspectives closer together, thereby strengthening the consistency of identity across perspectives.

[0070] Based on the above three modules, a multi-task joint optimization framework is finally constructed, combining all the above losses with the standard identity classification cross-entropy loss and the bounded triplet loss, such as... Figure 5 As shown, the viewpoint decoupling contrast loss is illustrated. The calculation principle is presented, along with the standard identity classification cross-entropy loss. and the loss of triples with boundaries The overall optimization objective consists of a weighted sum of multiple loss terms, where the weight coefficients are determined experimentally to balance the relationship between identity discriminability and perspective invariance. For example, , , The same weighting coefficients can be set. Crucially, during the inference phase, only the backbone network needs to be used to extract features from aerial and ground-view images for comparison; the other three modules do not need to participate, ensuring efficient and simple deployment.

[0071] Those skilled in the art will understand that both the standard identity classification cross-entropy loss and the bounded triple loss are general losses for pedestrian re-identification. The standard identity classification cross-entropy loss takes the model's predicted distribution and the true label distribution as inputs to ensure that the model can accurately classify different identities. The triple loss takes anchor samples, positive and negative samples as inputs to optimize the feature space, making similar features closer together and dissimilar features farther away. Considering that their calculation methods can refer to conventional techniques, they will not be elaborated further.

[0072] III. Effects Description

[0073] The solution provided in this invention, by introducing a deformable homography transformation module, achieves explicit and learnable geometric alignment between aerial and ground-based features for the first time in the AG-ReID (Aerial-to-Ground Person Re-identification) task. This module combines global homography transformation with local deformable optimization, effectively handling multi-level perspective differences from overall layout to local details, fundamentally bridging the cross-viewpoint gap. Secondly, the cross-viewpoint learning strategy cleverly utilizes the relationship between original and projected features, employing identity consistency loss and contrast loss to dually ensure that the model learns viewpoint-invariant representations while maintaining strong identity discrimination capabilities. Furthermore, the viewpoint decoupling contrast loss, through a viewpoint prototyping mechanism, actively removes residual viewpoint-specific information from features, further improving feature purity and generalization ability.

[0074] Of particular note is that this method does not rely on any auxiliary information (such as human attributes, key points, or additional cue words) and is trained entirely end-to-end. More importantly, all the complex modules designed for training (DHT, CVL, VDCL) can be completely discarded during the inference phase, and the final pedestrian matching can be completed using only a standard feature extraction backbone network. This feature greatly simplifies the deployment process and reduces computational overhead, making it highly suitable for large-scale application in practical urban security, emergency command, and other ground-air collaborative monitoring systems.

[0075] Furthermore, the method of this invention demonstrates excellent generalization ability and robustness. Its modular design is clear and easy to integrate with existing pedestrian re-identification frameworks, showing broad prospects for engineering applications and promotional value.

[0076] To visually demonstrate the aforementioned effects of the present invention, comparative experiments were conducted. The results of the comparative experiments show that the method proposed in this invention, by explicitly modeling cross-view geometric relationships and designing a specialized learning strategy, has achieved state-of-the-art performance on multiple publicly available aerial-ground pedestrian re-identification benchmark datasets, fully verifying its technical superiority. Specific performance improvement data are shown in Table 1.

[0077] Table 1: Ablation study on the enhancement effectiveness of deformable homography transformation modules on the CARGO dataset

[0078] In Table 1, ALL represents the results across all datasets. During testing, there was no specific distinction between aerial and ground-based perspectives. Rank1 represents the first-place hit rate, mAP represents the mean accuracy, and A↔G represents aerial↔ground. Aerial refers to the aerial perspective, and ground refers to the ground-based perspective. As shown in Table 1, on the CARGO dataset, after introducing the deformable homography transformation module, the Rank1 metric for the A↔G task improved from 47.50% to 50.00%.

[0079] In this embodiment of the invention, Figure 4 Both the baseline method in the previous experiment and the baseline method in the comparative experiment can be directly used as the VIT-based basic pedestrian re-identification method.

[0080] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0081] Example 2 This invention also provides a ground-to-air pedestrian re-identification system, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 6 As shown, the system mainly includes: The model building unit is used to build cross-view learning models, including: a backbone network, a deformable homography transformation module, a cross-view learning strategy module, and a view decoupling contrast loss module; The model training unit, used to train the cross-view learning model, includes: extracting features from aerial and ground-view images by the backbone network to obtain aerial and ground-view features; implementing a bidirectional and learnable geometric projection between the aerial and ground-view planes using a deformable homography transformation module to obtain projection features corresponding to each aerial and ground-view feature; calculating cross-view identity consistency loss and cross-view contrast loss using the cross-view learning strategy module, based on the aerial, ground-view features, and corresponding projection features; calculating the view decoupling contrast loss using the view decoupling contrast loss module, including a view decorrelation loss term calculated based on the distance between pedestrian identity information and the corresponding view prototype in the view features, and a contrast loss term calculated using different view features of the same pedestrian identity; and training the cross-view learning model by combining all calculated losses. The ground-to-air pedestrian re-identification unit is used to extract features from the aerial and ground-based visual images to be queried through the backbone network after training, and to achieve ground-to-air pedestrian re-identification based on the similarity between the features.

[0082] Since the main technical details of this system have been described in detail in previous embodiments, they will not be repeated here.

[0083] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0084] Example 3 The present invention also provides a processing device, such as Figure 7 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0085] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0086] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example: Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc. The output device can be a display terminal; The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0087] Example 4 The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0088] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0089] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for aerial pedestrian re-identification, characterized in that, include: A cross-view learning model is constructed, including: a backbone network, a deformable homography transformation module, a cross-view learning strategy module, and a view decoupling contrast loss module; The cross-view learning model is trained by: extracting features from aerial and ground-view images using a backbone network to obtain aerial and ground-view features; implementing a bidirectional and learnable geometric projection between the aerial and ground-view planes using a deformable homography transformation module to obtain projection features corresponding to each aerial and ground-view feature; calculating cross-view identity consistency loss and cross-view contrast loss using a cross-view decoupling contrast loss module, including a view decorrelation loss term calculated based on the distance between pedestrian identity information and the corresponding view prototype in the view features, and a contrast loss term calculated using different view features of the same pedestrian identity; and training the cross-view learning model by combining all calculated losses. After training, features of the aerial and ground-based visual images to be queried are extracted through the backbone network, and ground-to-air pedestrian re-identification is achieved based on the similarity between the features.

2. The method for ground-to-air pedestrian re-identification according to claim 1, characterized in that, The bidirectional and learnable geometric projection between the aerial and ground-based viewpoints achieved by the deformable homography transformation module includes: For aerial viewpoint features, a homography matrix is ​​predicted using a regression network. This homography matrix defines a global projective transformation from the aerial viewpoint plane to the ground viewpoint plane. The aerial viewpoint features are then globally projected using this homography matrix and normalized using perspective division. A deformable offset field is predicted and applied to the normalized global projection result. Finally, differentiable bilinear interpolation sampling is used to obtain the projection features of the aerial viewpoint features onto the ground viewpoint plane. Based on the same principle, for ground-view features, the projection features of ground-view features onto the aerial view plane are obtained by predicting the inverse matrix of the homography matrix and the corresponding learnable deformable offset field.

3. The method for ground-to-air pedestrian re-identification according to claim 2, characterized in that, The deformable offset field is generated through an attention-based aggregation mechanism, which includes: predicting a set of candidate spatial offsets and corresponding attention weights, and obtaining the deformable offset field by calculating the weighted average of the candidate spatial offsets.

4. The method for ground-to-air pedestrian re-identification according to claim 1, characterized in that, The calculation of cross-view identity consistency loss and cross-view comparison loss using aerial viewpoint features, ground viewpoint features, and corresponding projection features includes: For cross-view identity consistency loss, the difference between the aerial view features of the same pedestrian and the projection features of the corresponding ground view plane, as well as the difference between the ground view features of the same pedestrian and the projection features of the corresponding aerial view plane, are used to calculate the loss. For cross-view contrast loss, the Euclidean distance between positive sample pairs of the same pedestrian and the Euclidean distance between negative sample pairs that do not belong to the same pedestrian are used to calculate the loss by means of a preset boundary threshold. Among them, positive sample pairs of the same pedestrian include: aerial view features and ground view features of the same pedestrian; negative sample pairs include: negative sample pairs are aerial view features of pedestrian A and ground view features of other pedestrians, as well as ground view features of pedestrian A and aerial view features of other pedestrians.

5. The method for ground-to-air pedestrian re-identification according to claim 1, characterized in that, The decorrelation loss term is calculated in the following manner: For both aerial and ground-based perspectives, the average vector of all perspective features under each perspective is used as the corresponding perspective prototype. Both aerial and ground-based perspective features contain pedestrian identity information. The perspective decorrelation loss term is calculated based on the distance between the aerial perspective features and the aerial perspective prototype, as well as the distance between the ground-based perspective features and the ground-based perspective prototype, in order to maximize the distance between each perspective feature and its respective perspective prototype.

6. The method for ground-to-air pedestrian re-identification according to claim 1, characterized in that, The identity alignment loss term is calculated in the following manner: Identity alignment loss is a contrastive loss based on the idea of ​​noise contrast estimation. It is calculated by using the similarity between different perspective features of the same pedestrian identity to bring the different perspective features of the same pedestrian identity closer together.

7. The method for ground-to-air pedestrian re-identification according to claim 1, characterized in that, The process of training the cross-view learning model using all the calculated losses includes: The cross-perspective identity consistency loss, cross-perspective contrast loss, perspective decoupling contrast loss, and additionally calculated standard identity classification cross-entropy loss and bounded triplet loss are combined and configured with corresponding weight coefficients to construct the total training loss. The cross-perspective learning model is then trained using the total training loss.

8. A ground-to-air pedestrian re-identification system, characterized in that, To implement the method according to any one of claims 1 to 7, comprising: The model building unit is used to build cross-view learning models, including: a backbone network, a deformable homography transformation module, a cross-view learning strategy module, and a view decoupling contrast loss module; The model training unit, used to train the cross-view learning model, includes: extracting features from aerial and ground-view images by the backbone network to obtain aerial and ground-view features; implementing a bidirectional and learnable geometric projection between the aerial and ground-view planes using a deformable homography transformation module to obtain projection features corresponding to each aerial and ground-view feature; calculating cross-view identity consistency loss and cross-view contrast loss using the cross-view learning strategy module, based on the aerial, ground-view features, and corresponding projection features; calculating the view decoupling contrast loss using the view decoupling contrast loss module, including a view decorrelation loss term calculated based on the distance between pedestrian identity information and the corresponding view prototype in the view features, and a contrast loss term calculated using different view features of the same pedestrian identity; and training the cross-view learning model by combining all calculated losses. The ground-to-air pedestrian re-identification unit is used to extract features from the aerial and ground-based visual images to be queried through the backbone network after training, and to achieve ground-to-air pedestrian re-identification based on the similarity between the features.

9. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images