Pedestrian re-identification method with enhanced spatio-temporal kinematic characteristics of skeleton

Abstract

This invention proposes a pedestrian re-identification method enhanced with spatiotemporal kinematic features of the skeleton, comprising: acquiring a contour image sequence, a skeleton mask sequence, and skeleton angle data; performing skeleton-aware selective rotational position encoding on the contour image sequence to incorporate skeleton angle information; performing motion dynamics modeling on the skeleton angle data, calculating angular velocity and fusing it with the original angles, constructing a multi-hop time window through multi-scale convolution for accurate modeling of complex motion patterns, while relying on Mamba's selective state space to capture fine-grained details and global long-term temporal correlations; fusing two types of features through a cross-attention mechanism to guide contour feature extraction with motion information; extracting multi-scale spatiotemporal features through max pooling and horizontal pyramid pooling; and performing pedestrian identification based on the pooled features. This invention effectively solves the problems of insufficient extraction of skeleton angle frequency information and insufficient cross-modal interaction in existing methods, and improves the robustness of recognition under complex posture changes.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a pedestrian re-identification method that enhances skeletal spatiotemporal kinematic features. Background Technology

[0002] Pedestrian re-identification technology aims to achieve pedestrian matching and tracking across camera scenes and is one of the core technologies of intelligent video surveillance systems. With the development of deep learning technology, pedestrian re-identification methods based on deep neural networks have made significant progress. However, in practical applications, the diverse postures, clothing, and viewing angles of pedestrians still pose serious challenges to the performance of pedestrian re-identification systems. Existing pedestrian re-identification methods are mainly divided into two categories: appearance-based methods and gait-based methods. Appearance-based methods focus on extracting visual features such as color and texture, but are easily affected by changes in lighting and occlusion. Gait-based methods utilize pedestrian walking postures for identification, offering advantages such as physiological stability, non-contact operation, and difficulty in camouflage, making them a research hotspot in recent years. In gait-based pedestrian re-identification methods, skeletal data contains the location information of key joint points of the pedestrian, effectively representing the structural features of the human body. Traditional skeletal map construction methods typically involve simply connecting bone points to generate a skeletal map, which is then directly fused using a convolutional neural network. However, this method suffers from the following technical problems: Firstly, traditional convolutional neural networks primarily extract image features through spatial receptive fields, and their design is not specifically intended for processing the rotation angle and frequency information of skeletal data. The core information of skeletal data includes not only the spatial location of joints but also the angular changes between adjacent joints and the frequency characteristics of these angles evolving over time. Simply inputting a skeletal diagram as a regular image into a convolutional neural network cannot effectively capture the rotation angle information of the bones, making it difficult to fully express the dynamic changes during walking. Secondly, existing skeletal data processing methods are insufficient in cross-modal information fusion. The contour and angle information of the bones are usually extracted independently and then simply spliced ​​or weighted. This post-fusion approach often remains superficial, failing to deeply explore the semantic relationships between modalities, making it difficult to achieve deep interaction of temporal and spatial features and fine-grained semantic modeling, resulting in insufficient robustness of the model to complex posture changes. Furthermore, the traditional skeletal diagram construction process lacks a noise processing mechanism, easily incorporating noise information such as the rotation and length change frequency of the bones into the contour features, affecting recognition accuracy.

[0003] In summary, researching a pedestrian re-identification method that can effectively extract bone rotation angle frequency information and achieve deep fusion of multimodal information has important theoretical significance and application value. Summary of the Invention

[0004] The purpose of this invention is to provide a pedestrian re-identification method that enhances the spatiotemporal kinematic features of the skeleton. By selectively encoding the rotational position of the skeleton through skeleton perception, the method explicitly integrates the skeleton angle information into the contour feature representation. It utilizes a selective state space propagation mechanism based on Mamba and a multi-hop time window to capture fine-grained motion details and global long-term temporal correlations. Furthermore, it employs a cross-attention mechanism to achieve deep fusion of contour features and motion dynamics features. This addresses the technical problems in existing methods, such as insufficient extraction of skeleton rotation angle frequency information, cross-modal interaction remaining at the surface stitching level, and easy mixing of skeleton noise into contour features. As a result, the method significantly improves the robustness of the model to complex posture changes and the accuracy of pedestrian identification.

[0005] To achieve the above objectives, the specific solution of the present invention is as follows:

[0006] A method for pedestrian re-identification enhanced by skeletal spatiotemporal kinematic features includes the following steps:

[0007] S1. Data acquisition step: acquire the outline image sequence, skeleton mask sequence and skeleton angle data of the pedestrian to be identified;

[0008] S2. Skeleton-aware selective rotation position encoding step: Perform skeleton-aware selective rotation position encoding processing on the contour image sequence, integrate the skeleton angle data into the contour feature representation, and generate skeleton-aware rotation position encoding features.

[0009] S3. Motion dynamics encoding step: Motion dynamics modeling is performed on the bone angle data, the angle change between adjacent frames is calculated to obtain angular velocity features, and the angular velocity features are fused with the original bone angle features. Then, temporal modeling is performed through multi-scale feature extraction and a selective state space model based on Mamba to generate motion dynamics features.

[0010] S4. Cross-modal fusion step: Based on the cross-attention mechanism, the rotational position encoding features of the skeleton perception are fused with the motion dynamics features, so that the motion information guides the extraction of contour features to obtain fused features;

[0011] S5. Pooling feature extraction step: Pooling feature extraction is performed on the fused features. Max pooling is used to retain key temporal information and horizontal pyramid pooling is used to extract multi-scale spatial features to enhance the ability to express different body parts.

[0012] S6. Identity recognition step: Pedestrian identity recognition and classification based on pooled features.

[0013] Furthermore, the skeleton-aware selective rotational position encoding step in step S2 includes:

[0014] The angle encoding sub-step uses a multilayer perceptron to perform non-linear encoding on the skeletal angle data to generate an angle hidden representation.

[0015] The rotation accumulation sub-step uses short convolution to perform one-dimensional convolution processing on the angle hidden representation, and enhances the expressive power of the features through a gating mechanism. Finally, the rotation angle is calculated by accumulating and summing.

[0016] The spatial mapping sub-step uses the skeleton mask sequence to map the rotation angle to a spatial dimension, generating a spatially aware rotation angle.

[0017] The feature rotation sub-step involves performing rotation encoding processing on the contour image sequence based on the spatially perceived rotation angle to generate the skeleton-aware rotation position encoding features.

[0018] Furthermore, the feature rotation sub-step includes:

[0019] The contour image sequence is extended to a higher dimension to obtain a high-dimensional contour tensor;

[0020] The pixel vector of the high-dimensional contour tensor is divided into multiple pairs of two-dimensional components;

[0021] A rotational position encoding is applied to each two-dimensional component. Based on the spatially perceived rotation angle, each pair of two-dimensional components is rotated to obtain the rotationally encoded contour features.

[0022] Furthermore, the motion dynamics encoding step in step S3 includes:

[0023] Motion feature construction sub-step: Calculate the angle change of bone angle data in adjacent frames to obtain angular velocity features, and then splice and fuse the angular velocity features with the original bone angle features to obtain fused motion features;

[0024] Multi-scale feature extraction sub-step: Multi-scale one-dimensional convolution is used to extract multi-scale temporal features from the fused motion features, and a multi-hop time window is constructed to capture temporal features and complex motion patterns under different time spans;

[0025] Mamba temporal modeling sub-steps: After concatenating the multi-scale temporal features, input them into the selective state-space model of Mamba for global temporal modeling, mine fine-grained details and global long-term temporal correlations, and output the motion dynamics features.

[0026] Furthermore, the multi-scale one-dimensional convolution includes using depthwise separable convolution with different kernel sizes to extract short-term, medium-term, and long-term temporal features respectively, wherein the short-term scale captures rapid changes between frames, the medium-term scale captures patterns within the gait cycle, and the long-term scale captures the complete walking cycle.

[0027] Furthermore, the cross-modal fusion step in step S4 includes:

[0028] The query projection sub-step projects the skeleton-aware rotational position encoding features into a query vector;

[0029] The key-value projection sub-step projects the kinematic features into key vectors and value vectors;

[0030] The attention calculation sub-step calculates the similarity between the query vector and the key vector using a multi-head attention mechanism, and uses the value vector to perform feature aggregation to obtain the fused feature.

[0031] Furthermore, the pooling feature extraction step in step S5 includes:

[0032] Max pooling is performed on the fused features along the temporal dimension to retain key temporal information;

[0033] Horizontal pyramid pooling is applied to the features after max pooling to extract multi-scale spatial features.

[0034] Furthermore, a triplet loss function is used for model training to improve the distinguishability of samples with the same identity by narrowing the feature distance and widening the feature distance between samples with different identities.

[0035] The method is trained and / or tested on the Gait3D public dataset and / or the GREW public dataset.

[0036] The method is applicable to deep learning-based pedestrian re-identification systems, with inputs including RGB images, binary contour images, and / or human pose images.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The pedestrian re-identification method with enhanced spatiotemporal kinematic features proposed in this invention is a gait recognition framework that integrates skeleton-aware selective rotational position encoding, motion dynamics encoding, and early cross-modal fusion, which can effectively enhance the model's ability to extract bone rotation angle frequency information.

[0039] (1) By using skeleton perception selective rotation position encoding, the skeleton angle information is explicitly integrated into the rotation encoding of the contour feature. The skeleton mask is used to achieve pixel-level selective spatial mapping. Without changing the contour pixel modulus and spatial position, the motion characteristics of the contour points are reflected by vector rotation. This effectively suppresses the pollution of contour features by non-corresponding skeleton segment noise and significantly enhances the ability of contour features to express the law of skeleton rotation change.

[0040] (2) By constructing a multi-hop time window, multi-scale one-dimensional convolution is used to capture motion patterns across multiple time spans, including rapid changes between short-term frames, patterns within the mid-term gait cycle, and long-term complete walking cycles. Combined with the global temporal modeling capability of the Mamba selective state space model, the computational complexity is reduced while fully exploring the fine-grained details of joint motion and the global long-term temporal correlation, effectively solving the problem of insufficient extraction of bone angle and frequency information by traditional methods.

[0041] (3) By using the cross-attention mechanism, the rotation position encoding features of skeleton perception are deeply fused with the motion dynamics features in the early stage, so that the motion information actively guides the extraction process of contour features. This achieves deep interaction of temporal and spatial features and fine-grained semantic alignment, overcoming the problem of superficial interaction caused by simple splicing or weighted fusion of cross-modal information in existing methods, and significantly improving the robustness of the model to complex pose changes.

[0042] (4) The present invention achieves a Rank-1 accuracy of 69.1% and an mAP of 60.7% on the Gait3D public dataset and a Rank-1 accuracy of 68.9% on the GREW public dataset. Both are significantly better than mainstream baseline methods such as GaitSet, GaitPart, GaitGL and GaitBase, which verify the effectiveness and generalization ability of the method in real complex scenarios. Attached Figure Description

[0043] Figure 1 This is a flowchart of the pedestrian re-identification method with enhanced skeletal spatiotemporal kinematic features according to the present invention.

[0044] Figure 2 This is the overall design diagram of the pedestrian re-identification method with enhanced skeletal spatiotemporal kinematic features according to the present invention. Detailed Implementation

[0045] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments are not intended to limit the scope of the present invention.

[0046] like Figure 1 and Figure 2 As shown in the figure, this specific embodiment provides a pedestrian re-identification method with enhanced skeletal spatiotemporal kinematic features, which includes the following steps:

[0047] S1. Data acquisition step: acquire the outline image sequence, skeleton mask sequence and skeleton angle data of the pedestrian to be identified.

[0048] Specifically, the contour image sequence is a binary contour image sequence captured by a regular camera during the pedestrian's walking process and then segmented; the skeletal mask sequence includes masked regions of skeletal segments such as the upper arm, forearm, lower leg, and thigh; the skeletal angle data includes the angle information between skeletal segments. The contour image sequence is represented as follows: The skeleton mask sequence is represented as Skeletal angle data is represented as Where T is the number of frames, H and W are the height and width of the image, respectively, and B is the number of bone segments;

[0049] S2. Skeleton-aware selective rotation position encoding step: Based on the contour image sequence, the skeleton mask sequence, and the skeleton angle data, the contour image sequence is subjected to skeleton-aware selective rotation position encoding processing, and the skeleton angle data is integrated into the contour feature representation to generate skeleton-aware rotation position encoding features.

[0050] The skeleton-sensing selective rotational position encoding step specifically includes the following sub-steps:

[0051] (2.1) Angle encoding sub-step: Use a multilayer perceptron to process the skeletal angle data. Perform non-linear encoding to generate an angle-hidden representation. The calculation formula is as follows:

[0052]

[0053] in The angular hiding vector represents the vector. It is a multilayer perceptron.

[0054] (2.2) Rotation Accumulation Sub-step: The angle hidden representation is processed by one-dimensional convolution using short convolution to extract temporal features, and the expressive power of the features is enhanced by a gating mechanism. Then, the rotation angle is calculated by accumulating and summing, and the calculation formula is as follows:

[0055]

[0056]

[0057]

[0058]

[0059] in This represents the angular feature vector after convolution. The gated signal vector, This represents the Sigmoid activation function. This indicates the angular characteristics after gating enhancement. This represents the cumulative rotation angle encoding vector. For temperature parameters, the symbol is... This is element-wise multiplication.

[0060] (2.3) Spatial mapping sub-step: using the skeleton mask sequence The rotation angle encoding is mapped to the spatial dimension, meaning that the angle encoding feature of each bone segment pixel only applies to the corresponding spatial position of that bone segment in the contour map, generating a spatially aware rotation angle. The calculation formula is as follows:

[0061]

[0062] in, Indicates the first Frame skeleton mask, For the first The rotation angle vector of the frame. For the first Spatial awareness of rotation angle of the frame.

[0063] (2.4) Feature rotation sub-step: Based on the spatially aware rotation angle, the contour image sequence is rotated. Perform rotation encoding. First, process the contour data... Extended to We obtain a high-dimensional contour tensor. Next, the data point is rotated and encoded based on the angle of the bone segment it belongs to, so that it reflects the motion characteristics of the contour point through vector rotation without changing the modulus or spatial position. Specifically, the contour pixel vector is divided into... Yes, each pair of dimension indices is :

[0064]

[0065] in It is the first Frame number Each component has two dimensions. Applying standard RoPE to each component yields the rotated-encoded component. :

[0066]

[0067] The rotation matrix is ​​defined as:

[0068]

[0069] in Indicates the first The frame contour feature tensor, For the first Frame number Two-dimensional components, For the first A rotation matrix with two-dimensional components.

[0070] The overall calculation process of the above skeleton-aware selective rotational position encoding step is expressed as follows:

[0071]

[0072] S3. Motion dynamics encoding step: Motion dynamics modeling is performed on the bone angle data, the angle change between adjacent frames is calculated to obtain angular velocity features, and the angular velocity features are fused with the original bone angle features. Then, temporal modeling is performed through multi-scale feature extraction and a selective state space model based on Mamba to generate motion dynamics features.

[0073] The motion dynamics coding step specifically includes the following sub-steps:

[0074] (3.1) Sub-step for constructing motion features: Calculate the angular change of bone angle data in adjacent frames to obtain angular velocity features. The angular velocity features are then spliced ​​and fused with the original skeletal angle features, and the calculation formula is as follows:

[0075]

[0076]

[0077]

[0078] in It is the first Angular velocity feature vectors of all skeletal segments in the frame. It is the total number of frames in the sequence. It is the fused motion input feature vector.

[0079] (3.2) Multi-scale feature extraction sub-step: Multi-scale one-dimensional convolution is used to extract multi-scale temporal features from the fused motion features, and a multi-hop time window is constructed to capture temporal features and complex motion patterns at different time spans. Depth-separable convolutions with different kernel sizes are used to extract short-term, medium-term, and long-term temporal features respectively. The short-term scale captures rapid changes between frames, the medium-term scale captures patterns within the gait cycle, and the long-term scale captures the complete walking cycle. The calculation formula is as follows:

[0080]

[0081] in This represents depthwise separable convolution. This represents the fused motion features of the input.

[0082] The Mamba temporal modeling sub-step involves concatenating multi-scale features and then performing global temporal modeling using the Mamba selective state-space model to capture long-distance dependencies between motion frames. The calculation formula is as follows:

[0083]

[0084]

[0085] in, This represents the temporal modeling process of the Mamba selective state-space model. To hide the dimension.

[0086] The overall calculation process of the above motion dynamics coding steps is expressed as follows:

[0087]

[0088] S4. Cross-modal fusion step: Based on the cross-attention mechanism, the rotational position encoding features of the skeleton perception are fused with the motion dynamics features, so that the motion information guides the extraction of contour features and obtains fused features.

[0089] The cross-modal fusion step specifically includes the following sub-steps:

[0090] (4.1) Query projection sub-step: Project the skeleton-aware rotational position encoding features into a query vector. ;

[0091] (4.2) Key value projection sub-step: Project the kinematic features into key vectors. Sum value vector ;

[0092] (4.3) Attention Calculation Sub-step: The similarity between the query and the key is calculated using a multi-head attention mechanism, and feature aggregation is performed using the value vector. The calculation formula is as follows:

[0093]

[0094] In the formula, Represents the query matrix. Represents the key matrix, Represents a value matrix, This represents the dimension of the key vector.

[0095] The overall calculation process of the above cross-modal fusion steps is expressed as follows:

[0096]

[0097] S5. Pooling feature extraction step: Pooling feature extraction is performed on the fused features. Max pooling is used to retain key temporal information and horizontal pyramid pooling is used to extract multi-scale spatial features to enhance the ability to express different body parts.

[0098] The pooling feature extraction step specifically includes:

[0099] Max pooling is performed on the fused features along the temporal dimension to retain key temporal information. The calculation formula is as follows:

[0100]

[0101] Horizontal pyramid pooling is then applied to the features after max pooling to extract multi-scale spatial features:

[0102]

[0103] in, This represents the feature tensor after temporal pooling. This represents the max pooling operation along the time dimension. This indicates a horizontal pyramid pooling operation.

[0104] S6. Identity recognition step: Pedestrian identity recognition and classification based on pooled features.

[0105] The identity recognition step employs a fully connected layer and a batch normalized neck structure for identity classification, and uses a triplet loss function for model training. The distinguishability of features is enhanced by narrowing the feature distance between samples of the same identity and widening the feature distance between samples of different identities. The triplet loss function value... The calculation formula is as follows:

[0106]

[0107]

[0108] in, , and These are anchor sample features. : Positive sample features (belonging to the same identity as the anchor sample) and negative sample features (belonging to a different identity than the anchor sample). It is the Euclidean distance function. The safety margin parameter represents the loss of the triplet.

[0109] The overall calculation process of the above identity recognition steps can be represented as follows:

[0110]

[0111] The present embodiment's skeletal spatiotemporal kinematic feature-enhanced pedestrian re-identification method is evaluated on the Gait3D and GREW public datasets. The Gait3D public dataset covers a large number of pedestrian gait sequences in real-world scenarios, including pedestrian contour images, skeletal pose data, and skeletal angle data. The GREW public dataset also contains rich pedestrian gait data, which is used to verify the generalization ability of the present method in different scenarios.

[0112] This embodiment sets up three main evaluation tasks, as follows: 1) Rank-1 accuracy evaluation; 2) Rank-K accuracy evaluation; 3) Mean Average Precision (mAP) evaluation. For these three tasks, a total of four open-source gait recognition baseline methods are evaluated, including the set-based method GaitSet, the part-based method GaitPart, the graph learning-based method GaitGL, and the transformer-based method GaitBase. For each method, its independent performance is evaluated first, followed by its performance when the method of this embodiment is applied.

[0113] In the evaluation phase, the method of this embodiment and the methods of the prior art were compared and evaluated in the following three parts:

[0114] (1) Rank-1 accuracy (%): refers to the proportion of the target pedestrian correctly matched in the first prediction result returned. It is used to reflect the top recognition ability of the model. The higher the value of this indicator, the better.

[0115] (2) Rank-5 / Rank-10 accuracy (%): refers to the proportion of the target pedestrian correctly matched in the first 5 or first 10 prediction results returned. It is used to reflect the overall ranking ability of the model. The higher the value of this indicator, the better.

[0116] (3) Average precision mean (mAP (%)): refers to the average precision of all queries, used to measure the overall performance of the model across all queries. The higher the value of this indicator, the better.

[0117] The accuracy of Rank-1 and Rank-5 is evaluated using the accuracy metric; the mean precision is evaluated using the mAP metric.

[0118] The final performance of the method in this embodiment on the Gait3D and GREW datasets is shown in Table 1 below:

[0119] Table 1

[0120]

Claims

1. A pedestrian re-identification method with enhanced skeletal space-time kinematic features, characterized in that, Includes the following steps: S1. Data acquisition step: acquire the outline image sequence, skeleton mask sequence and skeleton angle data of the pedestrian to be identified; S2. Skeleton-aware selective rotation position encoding step: Perform skeleton-aware selective rotation position encoding processing on the contour image sequence, integrate the skeleton angle data into the contour feature representation, and generate skeleton-aware rotation position encoding features. S3. Motion dynamics encoding step: Motion dynamics modeling is performed on the bone angle data, the angle change between adjacent frames is calculated to obtain angular velocity features, and the angular velocity features are fused with the original bone angle features. Then, temporal modeling is performed through multi-scale feature extraction and a selective state space model based on Mamba to generate motion dynamics features. S4. Cross-modal fusion step: Based on the cross-attention mechanism, the rotational position encoding features of the skeleton perception are fused with the motion dynamics features, so that the motion information guides the extraction of contour features to obtain fused features; S5. Pooling feature extraction step: Pooling feature extraction is performed on the fused features. Max pooling is used to retain key temporal information and horizontal pyramid pooling is used to extract multi-scale spatial features to enhance the ability to express different body parts. S6. Identity recognition step: Pedestrian identity recognition and classification based on pooled features.

2. The method of claim 1, wherein, The skeleton-aware selective rotational position encoding step in step S2 includes: The angle encoding sub-step uses a multilayer perceptron to perform non-linear encoding on the skeletal angle data to generate an angle hidden representation. The rotation accumulation sub-step uses short convolution to perform one-dimensional convolution processing on the angle hidden representation, and enhances the expressive power of the features through a gating mechanism. Finally, the rotation angle is calculated by accumulating and summing. The spatial mapping sub-step uses the skeleton mask sequence to map the rotation angle to a spatial dimension, generating a spatially aware rotation angle. The feature rotation sub-step involves performing rotation encoding processing on the contour image sequence based on the spatially perceived rotation angle to generate the skeleton-aware rotation position encoding features.

3. The method of claim 2, wherein, The feature rotation sub-step includes: The contour image sequence is extended to a higher dimension to obtain a high-dimensional contour tensor; The pixel vector of the high-dimensional contour tensor is divided into multiple pairs of two-dimensional components; A rotational position encoding is applied to each two-dimensional component. Based on the spatially perceived rotation angle, each pair of two-dimensional components is rotated to obtain the rotationally encoded contour features.

4. The method of claim 1, wherein, The motion dynamics encoding step in step S3 includes: Motion feature construction sub-step: Calculate the angle change of bone angle data in adjacent frames to obtain angular velocity features, and then splice and fuse the angular velocity features with the original bone angle features to obtain fused motion features; Multi-scale feature extraction sub-step: Multi-scale one-dimensional convolution is used to extract multi-scale temporal features from the fused motion features, and a multi-hop time window is constructed to capture temporal features and complex motion patterns under different time spans; Mamba temporal modeling sub-steps: After concatenating the multi-scale temporal features, input them into the selective state-space model of Mamba for global temporal modeling, mine fine-grained details and global long-term temporal correlations, and output the motion dynamics features.

5. The method of claim 4, wherein, The multi-scale one-dimensional convolution includes depthwise separable convolution with different kernel sizes to extract short-term, medium-term and long-term temporal features respectively. The short-term scale captures rapid changes between frames, the medium-term scale captures patterns within the gait cycle, and the long-term scale captures the complete walking cycle.

6. The method of claim 1, wherein, The cross-modal fusion step in step S4 includes: The query projection sub-step projects the skeleton-aware rotational position encoding features into a query vector; The key-value projection sub-step projects the kinematic features into key vectors and value vectors; The attention calculation sub-step calculates the similarity between the query vector and the key vector using a multi-head attention mechanism, and uses the value vector to perform feature aggregation to obtain the fused feature.

7. The method of claim 1, wherein, The pooling feature extraction step in step S5 includes: Max pooling is performed on the fused features along the temporal dimension to retain key temporal information; Horizontal pyramid pooling is applied to the features after max pooling to extract multi-scale spatial features.

8. The method of claim 1, wherein, The model is trained using a triplet loss function, which improves the distinguishability by narrowing the feature distance between samples with the same identity and widening the feature distance between samples with different identities.

9. The method according to any one of claims 1 to 8, characterized in that, The method is trained and / or tested on the Gait3D public dataset and / or the GREW public dataset.

10. The method according to any one of claims 1 to 8, characterized in that, The method is applicable to deep learning-based pedestrian re-identification systems, with inputs including RGB images, binary contour images, and / or human pose images.

Images