Air-ground cross-view-angle pedestrian re-identification method based on multi-scale and view-angle perception double flow

Abstract

The invention belongs to the field of computer vision, particularly relates to an air-ground cross-view-angle pedestrian re-identification method based on multi-scale and view angle perception double flow, aims to match the same pedestrian individual under different view angles, and is applied to the industries of intelligent monitoring, public safety and the like. The system comprises two core branches, the multi-scale flow captures feature representations of different granularities, multi-scale semantic features are grouped and aggregated, and a feature representation which keeps consistency in a multi-scale space is learned, so that the model has robustness. According to the visual angle perception flow, a dynamic training strategy is adopted for different visual angle pairs, difficult samples are mined, the model can process visual angle changes more intelligently, and the cross-visual-angle matching capacity is remarkably improved. According to the method, feature representation with higher robustness and discrimination capability can be learned, and the Re-ID performance in a complex air-ground scene is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, and in particular relates to an air-ground cross-view pedestrian re-identification method based on multi-scale and viewpoint perception dual streams, which aims to match the same pedestrian individual under different viewpoints and can be applied to industries such as intelligent monitoring and public safety. Background Technology

[0002] Aerial-Ground Person Re-identification (AG-ReID) aims to match individuals captured by heterogeneous cameras from different perspectives. As a downstream task of image retrieval, it plays an irreplaceable role in fields such as intelligent surveillance and public safety. The basic process of an AG-ReID system is as follows: given a query image (such as an aerial view captured by a drone), it retrieves the most similar pedestrian image from a pre-built database of ground-view pedestrian images and uses the retrieval result as the pedestrian identity match for the query image. However, due to the significant differences between aerial and ground perspectives (such as overhead versus eye-level views, scale variations, and lighting variations), images of the same pedestrian exhibit significant appearance differences, posing a major challenge to the AG-ReID task. Therefore, extracting robust feature descriptions from pedestrian images that can cope with cross-viewpoint variations has become a key focus for improving the performance of AG-ReID algorithms. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a novel Multi-Scale and View-Specific Framework (MSVS), comprising two core branches. The Multi-Scale Stream (MS-Stream) captures feature representations at different granularities, groups and aggregates semantic features across multiple scales, and learns a feature representation that maintains consistency across the multi-scale space, thus enhancing the model's robustness. The View-Specific Stream (VS-Stream) employs a dynamic training strategy for different viewpoint pairs, mining challenging samples to enable the model to intelligently handle viewpoint changes and significantly improve cross-viewpoint matching capabilities. In this way, the MSVS framework learns more robust and discriminative feature representations, significantly improving Re-ID performance in complex air-to-ground scenarios.

[0004] The technical solution provided by this invention is as follows:

[0005] A method for cross-view pedestrian re-identification in air and ground based on multi-scale and viewpoint perception dual-stream includes the following steps:

[0006] Step 1: Obtain the image dataset, including open space scenes and pedestrian images;

[0007] Step 2: Input the image into the encoder and obtain image patch tokens, classification tokens clstoken, and view token;

[0008] Step 3: Use the multi-scale workflow module to extract global features and multi-scale local features, and construct contrastive learning tasks at each scale;

[0009] Step 4: Using the view perception workflow module, feature decoupling is performed on classification tokens and view tokens to extract view-independent identity features, and a similarity matrix is ​​constructed based on this. Positive and negative samples are defined by combining cls and view labels. The view perception contrast loss is calculated based on the weights and distances between anchor samples and positive and negative samples.

[0010] Step 5: Train the model using the total loss function and apply the trained model to the air-ground cross-view pedestrian re-identification task; the total loss function uses an adaptive weighting mechanism to dynamically adjust the weights of the comparison loss.

[0011] Step 2 is described in detail below:

[0012] The encoder is specifically a Transformer encoder. The input image is convolved to obtain image tokens patchtokens. After each Transformer layer of the Transformer encoder, the ID token is extracted from the viewpoint using inner subtraction, and finally the classification token cls token and the view token are output.

[0013] Step 3 is as follows:

[0014] Using cls tokens as coarse-grained global features ,Will Each Patch Token is directly used as a fine-grained local feature. Where D is a real vector space; then these features are grouped and aggregated to construct feature representations at different scales, resulting in coarse-grained features at the global scale, regional-scale features at the medium scale, and local-scale features at the fine-grained scale, forming a multi-scale feature set; the multi-scale feature set Specifically, this includes coarse-grained features at the global scale. Medium-grained regional scale characteristics and fine-grained local scale features It can be obtained through the following formula:

[0015]

[0016]

[0017]

[0018] in, Indicates the first spatial regions ( The set of Patch indices.

[0019] For a feature in a multi-scale feature set, its representation at scale s is as follows: For anchor point samples Through identity tags and perspective tags Perform positive and negative sample classification; first, based on Select all samples with the same identity as potential positive samples, and then based on Subdivision yields a set of positive samples from the same perspective. Cross-perspective positive sample set Let the positive sample set Finally, samples from different identities were selected as the negative sample set. , For all sample sets, Represents a set of viewpoint labels;

[0020] Then, based on the positive sample set and all sample sets, we calculate the intra-scale contrast loss of features of the same identity at scale s and the cross-scale contrast loss of features of the same identity at different scales.

[0021] Intra-scale contrast loss The calculation formula is:

[0022]

[0023]

[0024] in, Here, p is the temperature parameter, and p is any element in the positive sample set. It is the cosine similarity between the current sample i and the positive sample p. It is the cosine similarity between the current sample i and sample k; Is the i-th sample at scale The following features For scale The positive sample set and all sample sets;

[0025] Cross-scale contrast loss The complete form is:

[0026]

[0027] It represents the feature of the i-th sample at scale x. Let y be the positive sample set and all sample sets at scale y;

[0028] Ultimately, multi-scale loss is obtained. :

[0029]

[0030] in This is a hyperparameter.

[0031] Step 4 is described in detail below:

[0032] Extract the cls token as a global feature Extract view tokens as view features Then, the global features are subtracted from the view features to obtain the view-invariant identity features. ;

[0033] Use identity features Construct a similarity matrix within the batch Each element Indicates sample and samples The cosine similarity between them is then dynamically determined by the batch size, which in turn dynamically determines the number of difficult negative samples K, and the anchor samples. From all negative sample sets Choose the one with the highest similarity Each sample is used as a difficult negative sample. ;

[0034] Each anchor point sample Computational perspective-perceptual contrast loss :

[0035]

[0036]

[0037] , , These are the viewpoint weights of anchor point i and positive sample p, negative sample n, and difficult negative sample h, respectively. , , Let be the cosine similarity between anchor point i and positive sample p, negative sample n, and difficult negative sample h, respectively, where n and p are any elements in the negative sample set and the positive sample set, respectively. , As an intermediate variable;

[0038] formula Using the same perspective positive sample set The calculation is the same as the same-view contrast loss. Using cross-perspective positive sample sets The calculation is the cross-view contrast loss. .

[0039] The viewpoint-aware contrast loss introduces a time-dependent weight scheduling function during training. The implemented incremental learning is as follows:

[0040]

[0041]

[0042] t is the iteration round number. This refers to the number of rounds in the preheating phase. This represents the total number of training rounds.

[0043] The adaptive weighting mechanism is as follows:

[0044]

[0045] in, For adaptive weights, Based on the weights, This is the weight decay factor. As a weighting enhancement factor, For the performance improvement;

[0046] Therefore, the total loss function is expressed as follows:

[0047]

[0048] in, and For standard supervised learning, the loss is for identity classification and the loss is for perspective classification. , For the corresponding loss weights.

[0049] The beneficial effects of this invention are as follows:

[0050] This invention designs a multi-scale flow to ensure that the feature representation of the same identity remains consistent in the feature space at different scales and is clearly distinguishable from the features of other identities. It effectively integrates coarse-grained contour information and fine-grained texture information, improving the model's robustness to scale changes.

[0051] This invention designs a visual perception flow to mine and dynamically weight difficult samples through viewpoint features, intelligently optimizing the feature space. This alleviates the information differences and matching difficulties caused by different viewpoints, and improves cross-viewpoint matching accuracy.

[0052] This invention designs an advanced optimization strategy that incorporates progressive learning and adaptive weight adjustment. This enables the model to learn from easy to difficult, ensuring the stability and efficiency of the training process, and ultimately improving the model's convergence and generalization capabilities. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall framework structure of the method of the present invention;

[0054] Figure 2 This is a schematic diagram of the MS module workflow of the present invention;

[0055] Figure 3 This is a schematic diagram of the VS module workflow of the present invention;

[0056] Figure 4 This is a flowchart comparing the methods of the present invention with those of the prior art. Detailed Implementation

[0057] To better understand the purpose, structure, and function of this invention, the following description, in conjunction with the accompanying drawings, provides a more detailed account of an air-ground cross-view pedestrian re-identification method based on multi-scale and viewpoint perception dual streams.

[0058] like Figure 4 As shown, past methods typically rely on single-scale feature matching, applying a uniform processing strategy to all sample pairs. The method provided in this invention constructs a multi-scale feature space, capturing features of different granularities and employing a dynamic training strategy for different viewpoint pairs.

[0059] like Figure 1 As shown, we propose a novel Multi-Scale and View-Specific Dual-Stream Framework (MSVS), consisting of three core modules: an encoder, a multi-scale workflow, and a view-specific workflow. The encoder extracts core features with discriminative capabilities. The robustness of these core features is further enhanced by combining the complementary multi-scale and view-specific workflows. The core idea lies in fusing multi-scale feature information and further mining valuable view-related information.

[0060] like Figure 1As shown in (a), the encoder is specifically a Transformer encoder, which decouples the core features. This method uses the ViT backbone network as its core and introduces a view token for encoding specific viewpoints. The view token is used to perform refined hierarchical decoupling of the clstoken, separating the viewpoint-independent, more identity-discriminating core features. We designed two complementary workflows to enhance the robustness of these core features:

[0061] Multi-Scale Stream: This workflow aims to learn a feature representation that maintains consistency across multiple scales. For example... Figure 1 As shown in (b), it utilizes global features and multi-scale local features extracted from the ViT backbone network to construct contrastive learning tasks at various scales. Its core mechanism is to bring features of the same identity closer together across different scales and push features of different identities further apart. By combining intra-scale and cross-scale contrastive losses, the model can effectively fuse multi-granularity appearance information, significantly improving robustness to drastic scale changes. The core idea of ​​the multi-scale workflow is to learn a feature representation that maintains consistency across multiple scales by constructing intra-scale and cross-scale contrastive learning tasks when pedestrian scales change drastically, thereby ensuring the discriminativeness and consistency of features at different granularities.

[0062] View-Specific Stream: This workflow aims to guide the model to learn discriminative features that remain robust under extreme viewpoint changes, mine difficult samples, and dynamically assign learning strategies based on samples from different viewpoints. For example... Figure 1 As shown in (c), it extracts view-independent identity features through feature decoupling and uses them to construct a similarity matrix, combining cls and view labels to define positive and negative samples. Its core mechanism includes two aspects: first, a hard negative sample mining strategy, which identifies and focuses on processing the negative samples with the highest similarity to the anchor point; second, a view-aware weighting mechanism, which assigns higher loss weights to more challenging cross-view sample pairs, thereby strengthening the model's generalization ability under view changes.

[0063] The present invention provides a method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams, specifically including the following steps:

[0064] Step 1: Use the CARGO dataset (empty ground scene, composed using Unity3D), with a total of 13 cameras: 5 aerial cameras (Cam1~Cam5) + 8 ground cameras (Cam6~Cam13), totaling 5000 IDs and 108,563 pedestrian images. The training set contains 51,451 images (2,500 IDs), and the test set contains 51,024 images (2,500 IDs).

[0065] Step 2: The encoder uses a single-branch ViT but explicitly maintains two special tokens, forming a dual-stream output: an ID stream (cls) and a view stream (view). One cls token is used for identity representation, and the other view token specifically carries view factors. These tokens, along with patch tokens, participate in multi-layer Transformer encoding. The core of the encoding process is "feature decoupling": after each Transformer layer, inner subtraction is used to separate the ID token from view interference, making the ID representation more stable and generalizable across views.

[0066] Specifically, the input image is convolved to obtain image tokens (patch tokens). Each layer of the Transformer encoder includes LayerNorm, Multi-Head Self-Attention, and residual connections, and can enable DropPath for random depth regularization. Inner subtraction is performed after each layer: the ID token is output as x_id, the view token as x_view, and x_id is updated to x_id = x_id − x_view. This operation accumulates layer by layer, gradually subtracting viewpoint factors from the ID channel, achieving layer-by-layer constraints for "viewpoint decoupling." Because patch tokens and two special tokens interact bidirectionally in self-attention, the semantics of ID and viewpoint intersect and then separate at both spatial (patch) and global (token) levels: the viewpoint token aggregates viewpoint information, the ID token absorbs identity-related context, and the inner subtraction constrains the decoupling boundary between the two. Finally, a cls token for ID representation is obtained, carrying an identity-related global representation, and a view token for viewpoint-supervised learning, carrying viewpoint (Aerial / Ground) related information.

[0067] Step 3: Use the MS module to extract global features and multi-scale local features, and construct a contrastive learning task at each scale;

[0068] like Figure 2As shown, a coarse-grained global feature is obtained through cls tokens. ,Will Each Patch Token is directly used as a fine-grained local feature. Where D is a real vector space, defaulting to 768. In this way, diverse multi-scale information can be obtained from the output of a single backbone network without incurring additional computational burden. To capture pedestrians' appearance features more comprehensively, including This is a local feature of the nth PatchTokens.

[0069] These features are then grouped and aggregated to construct feature representations at different scales: coarse-grained features at the global scale. Medium-grained regional scale characteristics and fine-grained local scale features Global features provide overall semantic information, regional features balance global and local information, and local features capture detailed features. The grouping and aggregation process is specifically represented by the following formula:

[0070]

[0071]

[0072]

[0073] in, Indicates the first spatial regions ( The set of Patch indices, The spatial region is the number of patches within the region, and the region is a sub-region on the patch grid; therefore, the entire construction process can be formally represented as: , It is a multi-scale feature set.

[0074] Based on the aforementioned multi-scale feature construction, this method designs intra-scale contrastive loss and cross-scale contrastive loss to improve the model's robustness to scale changes. Intra-scale contrastive loss fully utilizes the unique information of each scale, such as the macroscopic semantics of global features and the fine details of local features. Representation of a specific scale: For the same identity, its feature representation at that scale. They should be close to each other in the feature space. Let represent the feature vectors of the same identity i at scale s, originating from viewpoints a and b respectively; while for different identities, their feature representations are... They should be far apart in the characteristic space. Let i represent the feature representations of identity i and identity j at scale s. Learnable weight parameters are introduced. Satisfying normalization constraints S represents the total number of multi-scale features, i.e., the number of scales the model divides during the feature extraction stage. For each scale... Features Independently compute adaptive weighted loss, dynamically balancing the contributions of features at different scales. It is any element in the multi-scale feature set F.

[0075] For anchor point samples Through identity tags and perspective tags Perform positive and negative sample classification; first, based on Select all samples with the same identity as potential positive samples, and then based on Subdivision yields a set of positive samples from the same perspective. Cross-perspective positive sample set For the sake of brevity, In fact and Used to calculate the same viewpoint loss respectively and cross-view loss Finally, samples from different identities were selected as the negative sample set. , For all sample sets. The size of the sample set in the current batch is [size missing]. This is used to calculate the average loss. Specifically, it is the intra-scale contrast loss. The complete form is:

[0076]

[0077]

[0078] This is a temperature parameter used to control the smoothness of the feature similarity distribution, where p is any element belonging to the positive sample set. It is the cosine similarity between the current sample i and the positive sample p. It is the cosine similarity between the current sample i and sample k; Is the i-th sample at scale The following features For scale The positive sample set and all sample sets.

[0079] Cross-scale contrastive loss performs comparative learning between features at different scales. Let x and y represent different scales: for the same identity, its feature representation at different scales. They should be close to each other in the feature space. The feature representation of identity i at different scales x and y; and the feature representation of different identities. They should be far apart in the feature space, regardless of their scale. Specifically, cross-scale contrastive loss. The complete form is:

[0080]

[0081] Ultimately, multi-scale loss It is a weighted sum of these two losses, expressed through hyperparameters. To balance the contributions of both:

[0082]

[0083] Step 4: The perspective-aware workflow is based on a core idea: the matching difficulty between different perspective pairs varies significantly, and different training strategies should be adopted. Specifically, samples with high similarity are more challenging than ordinary samples and require further exploration and attention. At the same time, cross-perspective samples are more difficult to match with same-perspective sample pairs and should receive more attention.

[0084] like Figure 3 As shown, the cls token is extracted from the encoder output as a global identity feature. Extract view tokens as view-related features. This dual-feature representation enables the model to simultaneously capture identity discrimination information and view perception information. Then, a feature decoupling strategy is employed, subtracting the view features from the global features to obtain view-invariant identity features. .

[0085] This step innovatively introduces a hard negative sample mining strategy based on similarity ranking. It utilizes decoupled identity features. Construct a similarity matrix within the batch Each element Indicates sample and samples The cosine similarity between them. The number of difficult negative samples K is then dynamically determined by the batch size, and the anchor samples... From all negative sample sets Choose the one with the highest similarity Each sample is used as a difficult negative sample. .

[0086] Considering the significant differences in matching difficulty between different viewpoint pairs, a viewpoint-aware weight matrix was also designed. This allows for dynamic adjustment of the contrast learning intensity of different perspective pairs, ensuring that cross-perspective sample pairs receive higher learning weights.

[0087] For batch size Each anchor point sample in Define the set of positive samples negative sample set and difficult negative sample set In the calculation, for Assign higher weighting factor Finally, for each anchor point sample... calculate Weighted perspective perception contrast loss:

[0088]

[0089]

[0090] , , These are the viewpoint weights of anchor point i and positive sample p, negative sample n, and difficult negative sample h, respectively. , , Let be the cosine similarity between anchor point i and positive sample p, negative sample n, and difficult negative sample h, respectively, where n and p are arbitrary elements in the negative sample set and the positive sample set, respectively; further, the formula... Using the same perspective positive sample set The calculation is the same as the same-view contrast loss. Using cross-perspective positive sample sets The calculation is the cross-view contrast loss. ; , It is an intermediate variable.

[0091] To achieve stable and efficient training, an optimization strategy incorporating progressive learning and adaptive weights was designed. View-Specific Stream loss. A time-dependent weighted scheduling function was introduced. This achieves incremental learning. We define the number of rounds in the warm-up phase as... In the early stages of training, the model primarily learns same-viewpoint matching, gradually shifting to more challenging cross-viewpoint matching as training progresses.

[0092]

[0093]

[0094] , These are the same-view contrast loss and cross-view contrast loss, respectively. For the time scheduling function of progressive learning;

[0095] We also introduced an adaptive weighting mechanism based on performance monitoring to dynamically adjust the weights of the comparison loss. :

[0096]

[0097] Based on the weights, This is the weight decay factor when performance stagnates. As a weighting factor for performance improvement, This refers to the performance improvement.

[0098] Step 5: Train the model using the total loss function, and apply the trained model to the air-to-ground cross-view pedestrian re-identification task. The final total loss function consists of two parts. The first part is the standard supervised learning loss, which includes the identity classification loss. and perspective classification loss The second part is our proposed contrastive learning loss, uniformly determined by adaptive weights. Adjustments were made.

[0099]

[0100] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams, characterized in that, Includes the following steps: Step 1: Obtain the image dataset, including open space scenes and pedestrian images; Step 2: Input the image into the encoder to obtain image patch tokens, classification tokens (cls token), and view token; Step 3: Use the multi-scale workflow module to extract global features and multi-scale local features, and construct contrastive learning tasks at each scale; Step 4: Using the view perception workflow module, feature decoupling is performed on classification tokens and view tokens to extract view-independent identity features, and a similarity matrix is ​​constructed based on this. Positive and negative samples are defined by combining cls and view labels. The view perception contrast loss is calculated based on the weights and distances between anchor samples and positive and negative samples. Step 5: Train the model using the total loss function and apply the trained model to the air-ground cross-view pedestrian re-identification task; the total loss function uses an adaptive weighting mechanism to dynamically adjust the weights of the comparison loss.

2. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 1, characterized in that, Step 2 is described in detail below: The encoder is specifically a Transformer encoder. The input image is convolved to obtain image tokens patchtokens. After each Transformer layer of the Transformer encoder, the ID token is extracted from the viewpoint using inner subtraction, and finally the classification token cls token and the view token are output.

3. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 2, characterized in that, Step 3 is as follows: Using cls tokens as coarse-grained global features ,Will Each Patch Token is directly used as a fine-grained local feature. , where D is the real vector space; then these features are grouped and aggregated to construct feature representations at different scales, resulting in coarse-grained features at the global scale, regional scale features at the medium scale, and local scale features at the fine scale, with each feature forming a multi-scale feature set; For a feature in a multi-scale feature set, its representation at scale s is as follows: For anchor point samples Through identity tags and perspective tags Perform positive and negative sample classification; first, based on Select all samples with the same identity as potential positive samples, and then based on Subdivision yields a set of positive samples from the same perspective. Cross-perspective positive sample set Let the positive sample set Finally, samples from different identities were selected as the negative sample set. , For all sample sets, Represents a set of viewpoint labels; Then, based on the positive sample set and all sample sets, we calculate the intra-scale contrast loss of features of the same identity at scale s and the cross-scale contrast loss of features of the same identity at different scales.

4. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 3, characterized in that, Step 4 is as follows: Extract the cls token as a global feature Extract view tokens as view features Then, the global features are subtracted from the view features to obtain the view-invariant identity features. ; Use identity features Construct a similarity matrix within the batch Each element Indicates sample and samples The cosine similarity between them is then dynamically determined by the batch size, which in turn dynamically determines the number of difficult negative samples K, and the anchor samples. From all negative sample sets Choose the one with the highest similarity Each sample is used as a difficult negative sample. ; Each anchor point sample Computational perspective-perceptual contrast loss : ； ； ； , , These are the viewpoint weights of anchor point i and positive sample p, negative sample n, and difficult negative sample h, respectively. , , Let be the cosine similarity between anchor point i and positive sample p, negative sample n, and difficult negative sample h, respectively, where n and p are any elements in the negative sample set and the positive sample set, respectively. , As an intermediate variable; formula Using the same perspective positive sample set The calculation is the same as the same-view contrast loss. Using cross-perspective positive sample sets The calculation is the cross-view contrast loss. .

5. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 4, characterized in that, The multi-scale feature set Specifically, this includes coarse-grained features at the global scale. Medium-grained regional scale characteristics and fine-grained local scale features It can be obtained through the following formula: ； ； ； in, Indicates the first spatial regions ( The set of Patch indices.

6. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 5, characterized in that, Intra-scale contrast loss The calculation formula is: ； ； in, Here, p is the temperature parameter, and p is any element in the positive sample set. It is the cosine similarity between the current sample i and the positive sample p. It is the cosine similarity between the current sample i and sample k; Is the i-th sample at scale The following features For scale The positive sample set and all sample sets; Cross-scale contrast loss The complete form is: ； It represents the feature of the i-th sample at scale x. Let y be the positive sample set and all sample sets at scale y; Ultimately, multi-scale loss is obtained. : ； in This is a hyperparameter.

7. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 6, characterized in that, The viewpoint-aware contrast loss introduces a time-dependent weight scheduling function during training. The implemented incremental learning is as follows: ； ； t is the iteration round number. This refers to the number of rounds in the preheating phase. This represents the total number of training rounds.

8. The method for cross-view pedestrian re-identification based on multi-scale and viewpoint perception dual streams according to claim 7, characterized in that, The adaptive weighting mechanism is as follows: ； in, For adaptive weights, Based on weights, This is the weight decay factor. As a weighting enhancement factor, For the performance improvement; Therefore, the total loss function is expressed as follows: ； in, and For standard supervised learning, the loss is for identity classification and the loss is for perspective classification. , For the corresponding loss weights.

Images