An air-ground pedestrian re-identification method based on frequency domain structure alignment and view angle identity decoupling

Abstract

The application discloses a kind of air-ground pedestrian re-identification methods based on frequency domain structure alignment and view angle identity decoupling, comprising: obtaining to be identified pedestrian image, extracting spatial domain feature map using visual backbone network;Frequency domain ready structure alignment is carried out to spatial domain feature map;Frequency domain transformation is carried out to the spatial domain feature map after structure alignment, and frequency domain feature map is obtained;Unified frequency semantic representation is constructed;Frequency adaptive routing is carried out based on frequency semantic representation, and frequency domain feature is decomposed into identity-biased frequency component and view angle-biased frequency component;Identity reliability fidelity processing is carried out to identity-biased frequency component;View angle-biased frequency component is carried out to view angle bias recovery processing;After fidelity, identity-biased frequency component and shape identity component are fused in frequency domain, and frequency domain identity feature is obtained, and spatial domain identity feature is extracted, and frequency domain identity feature and spatial domain identity feature are fused, and final identity is obtained;The application improves accuracy and feature robustness.

Description

Technical Field

[0001] This invention relates to the field of computer vision image technology, and in particular to a method for re-identifying pedestrians in the air and ground based on frequency domain structure alignment and viewpoint identity decoupling. Background Technology

[0002] Most existing pedestrian re-identification methods are designed for homogeneous camera networks, such as ground-to-ground or air-to-air matching tasks, and have achieved good results under relatively consistent imaging perspectives and shooting conditions. However, in practical applications, ground cameras and UAV platforms are often deployed collaboratively, forming heterogeneous air-ground camera networks. In this case, the query image and the image library image come from air and ground perspectives, respectively, and the appearance of pedestrians will change significantly with different viewing angles such as overhead, oblique, and eye-level views, thus forming a more complex air-ground pedestrian re-identification task.

[0003] Pedestrian re-identification in both air and ground not only needs to address the common problems in traditional pedestrian re-identification such as pose changes, occlusion, resolution differences, and background interference, but also must overcome challenges caused by significant differences in air and ground perspectives, including geometric distortion, semantic misalignment, inconsistent visible areas, and damage to identity-discriminating information. Existing methods can generally be categorized as follows: First, methods based on attributes or auxiliary information, which guide the model to learn cross-view shared representations by introducing identity attributes or additional supervision signals; second, methods based on shared space representation learning, which mitigate the interference of air-ground perspective differences on identity representations through mechanisms such as viewpoint decoupling, viewpoint-invariant constraints, cue learning, or explicit alignment. While the above methods alleviate the cross-view matching problem between air and ground to some extent, most of them still directly process mixed viewpoint-related and identity-related information in the spatial domain. When the air-ground perspective difference is large, if the viewpoint components are suppressed or decoupled in an integrated manner in the shared space, the key identity-discriminating information coupled with them is often weakened simultaneously, resulting in the destruction of some identity clues that should have been retained in the final representation. On the other hand, if spatially misaligned local features are directly transformed into the frequency domain, the resulting spectrum will simultaneously encode identity differences, geometric deformations, and local positional drifts, which is not conducive to subsequent frequency domain modeling and cross-view matching. Summary of the Invention

[0004] Purpose of the Invention: The purpose of this invention is to provide a method for re-identifying pedestrians in air and ground based on frequency domain structure alignment and viewpoint identity decoupling. By employing a frequency domain-ready structure alignment strategy, a more stable local spatial topology is restored before frequency domain modeling, solving the problems of spectral pollution, semantic misalignment, and insufficient cross-viewpoint comparability caused by directly transforming unaligned features in the frequency domain. Through frequency adaptive routing, identity reliability fidelity preservation, and viewpoint bias recovery mechanisms, the method performs structured diversion and fusion of identity-biased and viewpoint-biased components in the frequency domain, reducing the loss of key identity information caused by overall spatial domain decoupling, and improving the cross-viewpoint matching accuracy and feature robustness in pedestrian re-identification in air and ground.

[0005] Technical solution: The present invention provides a method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling, comprising the following steps:

[0006] (1) Obtain the image of the pedestrian to be identified and extract the spatial domain feature map using the visual backbone network;

[0007] (2) Perform frequency domain-ready structure alignment on the spatial domain feature map. Resample the local features by predicting continuous sampling offset and fusion weight to obtain the structure-aligned spatial domain feature map in order to restore the local semantic topology.

[0008] (3) Perform frequency domain transformation on the spatial domain feature map after structural alignment to obtain the frequency domain feature map, and construct a unified frequency semantic representation based on local frequency context, global structural conditions and frequency position prior;

[0009] (4) Frequency adaptive routing is performed based on frequency semantic representation, and the frequency domain features are decomposed into identity-biased frequency components and viewpoint-biased frequency components;

[0010] (5) Perform identity reliability fidelity processing on the biased identity frequency components to enhance the frequency response with stable identity discrimination capability and suppress unreliable frequency responses;

[0011] (6) Perform view offset recovery processing on the off-view frequency components to decompose them into explicit view offset components and recoverable shape identity components.

[0012] (7) The frequency component of the fidelity-preserved partial identity and the shape identity component are fused in the frequency domain to obtain the frequency domain identity features. The spatial domain identity features are extracted from the spatial domain feature map after structural alignment. The frequency domain identity features and the spatial domain identity features are fused to obtain the final identity representation for re-identification and retrieval of pedestrians in the air and ground.

[0013] Furthermore, in step (2), the frequency domain ready structure alignment is as follows: predict the structure state vector based on the global features extracted from the backbone network as global structure condition information; for each local feature region, predict multiple continuous sampling offsets and corresponding fusion weights in combination with the structure state vector; perform bilinear interpolation sampling and weighted fusion on the original spatial domain feature map according to the continuous sampling offsets and fusion weights to obtain the resampled local features, and reassemble them into a structure alignment feature map.

[0014] Furthermore, frequency-domain-ready structure alignment also includes: estimating spatial confidence of resampled local features and performing reliability gating based on spatial confidence to enhance the feature response of the aligned reliable region.

[0015] Further, step (3) is as follows: extract local frequency context from the frequency domain feature map; combine the global statistical information of the frequency domain features with the structural state vector to generate global conditional features and extend them to the entire frequency plane; introduce frequency position coding, fuse the local frequency context, global conditional features and frequency position coding to obtain frequency semantic representation.

[0016] Further, step (4) is as follows: predict the original scores of each frequency position belonging to the view branch and the identity branch according to the frequency semantic representation; calculate the route certainty and sharpen the route probability according to the route certainty, so that stronger unilateral allocation is achieved in high certainty positions and soft sharing is retained in low certainty positions; decompose the frequency domain features into identity-biased frequency components and view-biased frequency components according to the final route weight.

[0017] Further, step (5) is as follows: the identity reliability map is estimated by using the partial identity frequency component and the frequency semantic representation; the partial identity frequency component is weighted by the identity reliability map to enhance the contribution of the stable identity discrimination region and suppress the interference of the unreliable region.

[0018] Further, step (6) is as follows: the biased view frequency component is fused with the frequency semantic representation, and the view bias component and shape identity component are extracted through different mapping branches respectively; view classification supervision is introduced into the view bias component to enable it to distinguish view; view adversarial constraints are introduced into the shape identity component to suppress view-related information and retain identity structure semantics; orthogonal constraints are introduced between the view bias component and the shape identity component to reduce information redundancy.

[0019] Furthermore, the method also includes: constructing a total loss function to jointly optimize the network, the total loss function including identity classification loss, triplet loss, spectral consistency loss, viewpoint bias recovery loss, and cross-branch consistency loss; wherein, the spectral consistency loss is used to constrain the frequency domain response of the air-ground view images of the same identity to tend to be consistent after structure alignment; the viewpoint bias recovery loss is used to constrain the viewpoint bias component and the shape identity component to each assume a clear function; and the cross-branch consistency loss is used to maintain the consistency of identity semantics extracted by the biased identity branch and the viewpoint recovery branch.

[0020] Furthermore, the cross-branch consistency loss is obtained by extracting auxiliary identity vectors from the fidelity-preserved partial identity frequency component and shape identity component, respectively, and constraining their similarity in the semantic space.

[0021] The present invention provides a computer-readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the steps of any of the methods described herein.

[0022] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: Before frequency domain transformation, this invention performs structural alignment. By predicting continuous sampling offsets and confidence-gated resampling, it recovers the local semantic topology, avoiding spectral confusion caused by direct transformation of unaligned features, and ensuring the comparability of spatial and ground frequency domain responses from the source. Frequency adaptive routing is used to split biased identity and biased perspective frequency components in a "high-determinism hard allocation, low-determinism soft sharing" manner, avoiding the destruction of key identity clues caused by traditional spatial domain holistic decoupling. Unreliable frequency responses are suppressed through identity reliability maps, while perspective offset recovery is performed on biased perspective components, extracting reusable shape identity components and effectively supplementing identity discrimination information. Finally, the faithfully reproduced frequency domain identity features and spatial domain identity features are fused, achieving a complementary advantage between fine-grained frequency domain discriminability and spatial domain structural semantics, resulting in a more robust final identity representation. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this embodiment of the invention provides a method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling, including the following steps:

[0026] Step 1: Divide the pedestrian re-identification dataset into training, validation, and test sets. Perform preprocessing on the training images, including resizing, random flipping, random cropping, random erasing, and normalization, to construct batch input data for backbone network training. Specifically: Read the images and identity label information from the pedestrian re-identification dataset, and divide the samples into training, validation, and test sets according to the official or preset partitioning method. The training set is used for network parameter updates, the validation set is used for performance monitoring and optimal model saving during training, and the test set is used for final retrieval performance evaluation. Input images are preferably scaled to 256×128 pixels and converted to tensor form. To improve model generalization ability and reduce overfitting, this invention employs enhancement methods such as random horizontal flipping, edge padding and random cropping, and random erasing during the training phase. Finally, the images are normalized using the channel mean and standard deviation corresponding to the ImageNet pre-trained model. To facilitate subsequent identity classification loss, triplet loss, and cross-view positive sample construction, this invention adopts a batch construction method based on identity-balanced sampling. Preferably, each batch size is set to 64, and each batch contains multiple different identities. Each identity is downsampled with several empty view and ground view images to ensure that cross-view samples of the same identity can participate in training in the same batch. This can simultaneously meet the needs of identity supervision, hard sample mining, and spectral consistency constraints.

[0027] Step 2: Input the preprocessed image into the Vision Transformer backbone network to extract global features and patch-level local features. Based on the global features, estimate the structural state vector of the current sample to provide global conditional information for subsequent structural alignment. This invention uses ViT-Base, pre-trained on ImageNet, as the feature extraction backbone network. For the input image... First, the image is divided into multiple non-overlapping image patches of size 16×16, and patch tokens are obtained through linear mapping. Then, the patches are encoded by the ViT backbone network to obtain a two-dimensional patch feature map and a global feature vector. This process is represented as follows:

[0028]

[0029] in, Indicates the ViT backbone network; This represents a local feature map reconstructed from the patch tokens. This represents the global feature vector. Used to represent local spatial structure information Used to summarize the overall structural state and global semantics of the current sample.

[0030] Furthermore, in order to enable subsequent local alignment to perceive the view-induced deformation pattern of the current sample as a whole, this invention utilizes global features. Predicting the structural state vector using a multilayer perceptron It is represented as:

[0031]

[0032] in, This represents a structural state prediction network. This represents the structural state encoding of the current sample; The dimension of the structural state vector is represented, and in this embodiment, it is preferably set to 256. This structural state vector will be used as global conditional information in subsequent offset prediction, sampling weight prediction, and spatial confidence estimation, so that structural alignment no longer depends solely on the appearance of local patches, but can be modeled by combining the global structural information of the entire human body.

[0033] Step 3: Based on the structural state vector and local patch features, predict multiple consecutive sampling offsets and fusion weights for each patch, and perform frequency domain ready structure alignment; on this basis, further estimate the patch-level spatial confidence, perform reliability gating on the resampling features, and obtain an alignment feature map suitable for subsequent frequency domain analysis;

[0034] Considering that changes in the aerial and ground perspectives can cause significant shifts in the local semantics of the human body on the patch mesh—for example, the relative positions of semantic regions such as the head, shoulders, torso, and lower limbs change significantly between the aerial and ground views—directly performing a two-dimensional discrete cosine transform on the unaligned feature maps would simultaneously introduce identity differences, perspective-induced deformation, and spatial misalignment into the frequency domain response, hindering subsequent identity-perspective decoupling. Therefore, this invention designs a frequency domain-ready structure alignment module to perform structure-guided continuous resampling of local features.

[0035] For feature maps The p-th patch feature First, it is compared with the structural state vector obtained in Step 2. By splicing together, conditional features are formed. It integrates conditional features that combine local appearance information and global structural information. Through these conditional features, subsequent local alignment can simultaneously perceive "the semantic meaning of the current position" and "the structural state of the entire human body." Based on conditional features... This invention predicts continuous sampling offsets and corresponding fusion weight scores using multiple offset prediction heads and multiple weight prediction heads, respectively, as follows:

[0036]

[0037] in, Indicates the first One offset prediction head, Indicates the first Two-dimensional offset of each sampling point relative to the patch position; Indicates the first Each weighted prediction head; Indicates the first Unnormalized weighted scores for each sampling point; This represents the sampling fusion weights after softmax normalization; This indicates the number of sampling points. In this invention, it is preferably set to 4 to balance the accuracy of structural compensation and computational complexity.

[0038] After obtaining the offset and weights, bilinear interpolation is used to sample from the original feature map X at continuous locations to obtain the multi-point fused resampled features:

[0039]

[0040] in, Indicates the first Resampled features of a location after continuous sampling at multiple points; This indicates bilinear interpolation sampling of the feature map X on continuous coordinates; Indicates the first The sampling points are located at consecutive positions in the feature map. In this way, the feature representation of the current patch is no longer fixed to its original discrete position, but is composed of multiple neighborhood sampling points that are more suitable for that semantic position, thereby restoring a more comparable local structure topology.

[0041] Furthermore, considering that some local areas may be invisible, occluded, or contaminated by the background from the current viewpoint, or may remain unstable even after resampling, this invention introduces patch-level spatial confidence gating to screen the reliability of resampled features. Specifically, the resampled features... Original patch features and structural state vector A common input spatial confidence prediction head is used to obtain spatial confidence, and the resampling results are gated, which is expressed as follows:

[0042]

[0043] in, Indicates a spatial confidence prediction network; This represents the Sigmoid function; Indicates the first The reliability coefficient of each patch; This represents the aligned patch features after gating. The spatial confidence is not set manually, but is jointly estimated from the "aligned local information", "original local information" and "global structural state". Its role is to suppress unreliable regions and enhance the contribution of comparable regions before entering frequency domain modeling.

[0044] By recombining all the aligned patch features, we can obtain the structure alignment feature map. To achieve a more consistent spectral distribution in the feature map across different viewpoints, this invention further introduces a spectral consistency loss. Let the outputs of the empty view and the ground view with the same identity after structural alignment be respectively... and Then the spectral consistency loss is defined as:

[0045]

[0046] in, Indicates the number of channels; and These represent the alignment feature maps of the same identity's empty view and ground view at the [number]th [location]. Response on each channel; Represents the two-dimensional discrete cosine transform; express Norm; This represents a minimal constant to prevent the denominator from being zero.

[0047] Step 4: Perform a two-dimensional discrete cosine transform on the aligned feature map obtained in Step 3 to construct a frequency semantic representation that integrates local frequency context, global conditional information and location prior; then perform frequency adaptive routing to decompose the frequency domain features into identity-biased frequencies and viewpoint-biased frequencies, and perform identity reliability fidelity preservation and viewpoint bias recovery respectively to obtain frequency domain identity representation.

[0048] Applying a two-dimensional discrete cosine transform to the structure alignment feature map obtained in Step 3 maps the spatial representation to the frequency domain, yielding a frequency domain feature map:

[0049]

[0050] in, This represents the aligned frequency domain feature map.

[0051] To enable decisions at each frequency location to simultaneously perceive the local frequency context, the global structural state, and the location prior, this invention further constructs a unified frequency semantic representation. Specifically, frequency domain features are first... enter Convolution yields frequency embeddings. ; and then through Depthwise separable convolution obtains local frequency context At the same time, Global average pooling result and structure state vector Perform concatenation, input condition mapping network The output is then broadcast across the entire frequency plane to obtain global conditional features. Combined with frequency position coding Finally, a fused frequency semantic representation is constructed:

[0052]

[0053] in, This means copying the global condition vector across the entire frequency plane; Indicates global pooling; express Depthwise separable convolution; This represents a frequency semantic representation that integrates local frequency context, global structural conditions, and positional priors. Each frequency position in the sample not only carries its own local context information, but also explicitly perceives the overall structural state and positional prior of the current sample.

[0054] Subsequently, this invention performs frequency-adaptive routing. Specifically, based on the fused representation Z, the original scores for each frequency location biased towards the viewpoint branch and the identity branch are predicted respectively, and the basic routing probabilities are obtained through softmax:

[0055]

[0056] in, and These represent the original route scores for perspective-biased branches and identity-biased branches, respectively. and They represent the first The first channel, the first , The basic routing probabilities of each frequency location belonging to the viewpoint branch and the identity branch.

[0057] To measure the degree of certainty that the current frequency position belongs to a certain branch, this invention further defines routing certainty and, based on this, derives a probabilistic sharpening mechanism to construct the final routing weights:

[0058]

[0059]

[0060] in, This indicates the routing certainty at the current frequency location. and These represent the final viewpoint routing weight and identity routing weight, respectively. When and When frequencies are close in frequency, it indicates that the semantics of that frequency are ambiguous; when one class clearly dominates, it indicates that the semantics of that frequency are more defined. Based on this, this invention constructs the final routing weights through a probability sharpening mechanism guided by certainty:

[0061]

[0062] This leads to the partial viewpoint frequency and partial identity frequency:

[0063]

[0064] Where ⊙ represents element-wise multiplication; Indicates the off-angle frequency; This indicates a frequency bias towards identity. Here, routing is not a simple hard partitioning, but rather a stronger unilateral allocation in high-determinism locations and a moderate amount of soft sharing in low-determinism locations, thereby avoiding overly aggressive frequency cutting.

[0065] Step 5: Perform identity reliability fidelity preservation and viewpoint offset recovery on the biased identity frequency and biased viewpoint frequency obtained in Step 4 to obtain the frequency domain identity representation. Then, extract the spatial domain identity representation from the aligned spatial domain feature map and fuse it with the frequency domain identity representation to obtain the final identity representation.

[0066] Regarding the identity bias frequency stream obtained in Step 4, this invention does not directly... Instead of using the final identity frequency as the primary identifier, the invention first determines which positions are "reliable identity frequencies" and which are merely low-reliability responses affected by noise, misalignment, or unstable local patterns. To achieve this, the invention utilizes both the partial identity frequency itself and the shared frequency semantic representation. Jointly estimate the identity reliability map. Specifically, this involves partial identity frequency estimation. It provides local evidence of the current frequency location itself, while sharing frequency semantic representation. This provides auxiliary judgment criteria that combine global structural state, local frequency context, and frequency position prior. The combined effect of these two factors allows for a more accurate determination of whether the current frequency response truly possesses stable identity discrimination value. Based on this, the present invention defines an identity reliability graph. And obtain the identity frequency after authentication. It is represented as:

[0067]

[0068] in, For identity bias frequency stream; For identity reliability graph; This is a frequency representation of the identity after ensuring fidelity; This represents a convolutional mapping that extracts reliable evidence from partial identity frequencies. This represents a convolutional mapping that extracts reliable priors from the shared frequency semantic representation. This process enhances the contribution of stable identity frequencies while suppressing the interference of unreliable frequency responses on the final identity representation.

[0069] Secondly, regarding the viewpoint offset frequency stream, this invention does not simply... Direct deletion or overall suppression. Because While the data contains significant viewpoint bias information, it may also contain retrievable identity clues related to local human structures. Therefore, this invention further incorporates viewpoint bias frequency... It is decomposed into two parts: one part is used to explicitly represent the viewpoint offset, denoted as The other part is used to represent the shape identity information recovered from the viewpoint branch, denoted as... To achieve this objective, the present invention will deflect the viewing angle frequency. With shared frequency semantics This indicates that the components are stitched together, and then the view offset component and shape identity component are extracted through two different mapping branches, as shown below:

[0070]

[0071] in, For the viewpoint bias frequency flow, Indicates the viewpoint offset component. This represents the identity shape component recovered from the dominant viewpoint frequency. In this way, the information originally mixed in the biased viewpoint frequency is further separated into "viewpoint bias information that should be retained" and "auxiliary structural information that can still be used for identity determination".

[0072] In order to ensure that the obtained and Each component performs a clearly defined and distinct function, and this invention further imposes targeted constraints on these two components. Specifically, firstly, regarding the viewpoint offset component... This invention aims to reflect the differences between different perspectives such as aerial and ground as accurately as possible, therefore... The input is a standard viewpoint classifier, and its viewpoint separability is enhanced through cross-entropy loss. Secondly, for the shape identity component, this invention aims to remove viewpoint information as much as possible, retaining only stable structural semantics related to identity. Therefore, it is input into a viewpoint classifier with a gradient inversion layer to suppress viewpoint cues through inverse gradient methods. Finally, to avoid... and Since both components learn too much redundant information, this invention introduces an orthogonal constraint between them to separate the viewpoint bias component and the shape identity component as much as possible in the representation space. Combining the above three constraints, the viewpoint bias recovery loss is defined as:

[0073]

[0074]

[0075] in, This indicates the viewpoint bias recovery loss; , , These represent the weighting coefficients of the three losses; Represents cross-entropy loss; This represents a standard viewpoint classifier; This represents a view classifier with a gradient inversion layer; Indicates perspective label; This represents a minimal constant to prevent the denominator from being zero. Preferably, , , Through this loss, it is possible to More focused on modeling perspective bias, while It focuses more on preserving the structural semantics related to identity.

[0076] Obtained in two frequency domain branches and Subsequently, the present invention performs layer-level fusion on these features to obtain the final frequency domain identity feature map. :

[0077]

[0078] Furthermore, to avoid conflicting identity semantics learned by the identity-biased branch and the perspective recovery branch, this invention respectively... and Auxiliary identity vectors are extracted from the vectors, and then a cross-branch consistency loss is defined using cosine similarity to ensure that the two remain similar in the semantic space. This loss is expressed as:

[0079]

[0080] in, and These represent the linear mapping parameters of the two auxiliary identity projection heads, respectively;

[0081] Finally, after completing the frequency domain fusion, this invention extracts the final frequency domain identity vector from the frequency domain identity feature map, while retaining the spatial domain identity vector extracted from the spatial domain alignment feature map, and then performs late fusion on the two to obtain the final identity representation used for retrieval. This is represented as:

[0082]

[0083] in, This represents the final frequency domain identity vector; Indicates the parameters of the frequency domain projection head; Represents a spatial domain identity vector;

[0084] Indicates the parameters of the spatial domain projection head; This indicates a normalization operation; This represents the final identity representation after fusion. This fusion method, without requiring a positional correspondence between the spatial and frequency domain feature maps, fully preserves the complementary advantages of both types of representations, providing more robust identity features for subsequent final retrieval and joint loss optimization.

[0085] Step 6: Construct the total loss function, jointly optimize the identity classification loss, triplet loss, spectrum consistency loss, viewpoint bias recovery loss, and cross-branch consistency loss, and select the optimal weights on the validation set as the final model parameters.

[0086] After completing identity fidelity, viewpoint offset recovery, and spatial-frequency domain fusion in Step 5, this invention has obtained the final identity representation used for retrieval. To enable the structural alignment module, frequency domain decoupling module, and final identity learning process to be collaboratively optimized within the same training framework, this invention constructs a unified total loss function in Step 6 and combines classification supervision, metric learning supervision, spectral consistency constraints, viewpoint bias recovery constraints, and cross-branch consistency constraints to perform end-to-end training on the entire network. The purpose of this step is to: on the one hand, ensure the final representation... It possesses sufficiently strong identity discrimination capabilities; on the other hand, it ensures that the intermediate representations of the preceding modules are not optimized in isolation, but rather converge together in a direction that is conducive to cross-view matching between air and ground. The total loss function constructed in this invention is as follows:

[0087]

[0088] in, For identity classification loss, For triple loss, For spectrum consistency constraints, Restoring constraints for viewpoint bias. For cross-branch consistency constraints, , and For the corresponding hyperparameters, the preferred setting in this invention is... , , To balance structural accuracy and computational complexity.

[0089] In terms of training implementation, this invention uses the PyTorch framework and completes training on a single NVIDIA RTX 4090 GPU; the optimizer is preferably AdamW, the weight decay is set to 0.05, and the initial learning rate is set to... The total number of training epochs is preferably 120, with the first 20 epochs using linear warm-up and subsequent cosine annealing to decay the learning rate. The weights of the spectral consistency loss, viewpoint bias recovery loss, and cross-branch consistency loss are preferably set to 0.5, 0.5, and 0.2, respectively. During training, the model is evaluated using a validation set after each epoch, and the optimal network parameters are saved, represented as follows:

[0090]

[0091] in, Indicates the first The model parameters obtained from the first round of training, This represents the model parameters that achieve the best performance on the validation set. This represents the average accuracy evaluation metric.

[0092] During the testing phase, load the optimal model parameters. The final identity features of the query sample and the image library sample are extracted separately, and their similarity scores are calculated, with cosine similarity being the preferred method.

[0093]

[0094] in, Indicates the query sample. Represents a sample from a library. and These represent the final identity features of the corresponding samples. This represents the similarity score between the two. Sort the results by similarity score and output the final search results for pedestrians in empty areas.

Claims

1. A method for re-identifying pedestrians in open spaces based on frequency domain structure alignment and viewpoint identity decoupling, characterized in that, Includes the following steps: (1) Obtain the image of the pedestrian to be identified and extract the spatial domain feature map using the visual backbone network; (2) Perform frequency domain-ready structure alignment on the spatial domain feature map. Resample the local features by predicting continuous sampling offset and fusion weight to obtain the structure-aligned spatial domain feature map in order to restore the local semantic topology. (3) Perform frequency domain transformation on the spatial domain feature map after structural alignment to obtain the frequency domain feature map, and construct a unified frequency semantic representation based on local frequency context, global structural conditions and frequency position prior; (4) Frequency adaptive routing is performed based on frequency semantic representation, and the frequency domain features are decomposed into identity-biased frequency components and viewpoint-biased frequency components; (5) Perform identity reliability fidelity processing on the biased identity frequency components to enhance the frequency response with stable identity discrimination capability and suppress unreliable frequency responses; (6) Perform view offset recovery processing on the off-view frequency components to decompose them into explicit view offset components and recoverable shape identity components. (7) The frequency component of the fidelity-preserved partial identity and the shape identity component are fused in the frequency domain to obtain the frequency domain identity features. The spatial domain identity features are extracted from the spatial domain feature map after structural alignment. The frequency domain identity features and the spatial domain identity features are fused to obtain the final identity representation for re-identification and retrieval of pedestrians in the air and ground.

2. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, In step (2), the frequency domain ready structure alignment is as follows: predict the structure state vector based on the global features extracted from the backbone network as global structure condition information; for each local feature region, predict multiple continuous sampling offsets and corresponding fusion weights in combination with the structure state vector; perform bilinear interpolation sampling and weighted fusion on the original spatial domain feature map according to the continuous sampling offsets and fusion weights to obtain the resampled local features and reassemble them into a structure alignment feature map.

3. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 2, characterized in that, Frequency-domain-ready structure alignment also includes estimating spatial confidence of resampled local features and performing reliability gating based on spatial confidence to enhance the feature response of the aligned reliable region.

4. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, Step (3) is as follows: extract local frequency context from the frequency domain feature map; combine the global statistical information of the frequency domain features with the structural state vector to generate global conditional features and extend them to the entire frequency plane; introduce frequency position coding, fuse the local frequency context, global conditional features and frequency position coding to obtain frequency semantic representation.

5. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, Step (4) is as follows: predict the original scores of each frequency position belonging to the view branch and identity branch according to the frequency semantic representation; calculate the route certainty and sharpen the route probability according to the route certainty, so that stronger unilateral allocation is achieved in high certainty positions and soft sharing is retained in low certainty positions. The frequency domain features are weighted and decomposed into identity-biased frequency components and viewpoint-biased frequency components based on the final routing weight.

6. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, Step (5) is as follows: use the partial identity frequency components and frequency semantic representation to jointly estimate the identity reliability map; use the identity reliability map to weight the partial identity frequency components to enhance the contribution of the stable identity discrimination region and suppress the interference of the unreliable region.

7. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, Step (6) is as follows: the biased view frequency component is fused with the frequency semantic representation, and the view bias component and shape identity component are extracted through different mapping branches respectively; view classification supervision is introduced to the view bias component to enable it to distinguish view; view adversarial constraints are introduced to the shape identity component to suppress view related information and retain identity structure semantics; orthogonal constraints are introduced between the view bias component and the shape identity component to reduce information redundancy.

8. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, The method further includes: constructing a total loss function to jointly optimize the network, the total loss function including identity classification loss, triplet loss, spectral consistency loss, viewpoint bias recovery loss and cross-branch consistency loss; wherein, the spectral consistency loss is used to constrain the frequency domain response of the air-ground view images of the same identity to tend to be consistent after structure alignment; the viewpoint bias recovery loss is used to constrain the viewpoint bias component and the shape identity component to each assume a clear function; the cross-branch consistency loss is used to maintain the consistency of identity semantics extracted by the biased identity branch and the viewpoint recovery branch.

9. The method for re-identifying pedestrians in the air based on frequency domain structure alignment and viewpoint identity decoupling according to claim 1, characterized in that, The cross-branch consistency loss is obtained by extracting auxiliary identity vectors from the fidelity-preserved partial identity frequency component and shape identity component, respectively, and constraining their similarity in the semantic space.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When executed by a processor, the program instructions implement the steps of the method according to any one of claims 1-9.

Images