A gait recognition method combining local correlation features and discriminative features

Abstract

The application belongs to the technical field of deep learning, and particularly relates to a gait recognition method combining local correlation features and discriminative features. First, in order to utilize the global and local information of the human body simultaneously and make the extracted local detail features retain the edge information of the local area of the human body, a local correlation feature extraction convolution network module LRConv is designed. Second, in order to solve the problem that the current method based on the local area of the human body does not focus on the local part of the human body which is less affected by clothing, a human body part focusing attention mechanism module HBF is used to focus on the local position of the human body which is less affected by the clothing, reduce the influence of the clothing of the human body on the gait recognition, and thus extract gait detail features which are more discriminative. The experimental results show that in the NM, BG and CL gait environments set in the CASIA-B dataset, GaitLRDF is improved by 0.48%, 0.74% and 2.12% respectively compared with the current most advanced algorithm, and the recognition accuracy on the OU-MVLP dataset reaches 91.35%.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to a gait recognition method that combines local correlation features and discriminative features. Background Technology

[0002] Gait recognition is a biometric technology that relies on the uniqueness of human walking posture, allowing gait information to be captured from a distance using ordinary camera equipment. Furthermore, the gait recognition process does not require the cooperation of the subject, making it widely applicable in various fields such as video surveillance and intelligent transportation. Most existing gait recognition methods employ convolutional neural networks (CNNs) to represent gait features for better recognition performance. Current gait feature representations can be divided into two categories: global feature-based representations and local feature-based representations. Global feature-based methods extract gait features from the entire gait image framework for gait recognition. Local feature-based methods extract and combine local gait features from specific gait regions for gait recognition. Global feature representations tend to ignore highly discriminative gait features, while local feature representations, although obtaining more accurate local details than global feature maps, neglect the correlation information between local gait regions, leading to a loss of overall coherence between gait areas. While GaitGL recognized the above problems and introduced 3D CNNs to extract gait spatiotemporal features—for example, patent application number "201811277328.2" describes "training global and local gait features through a three-dimensional convolutional neural network, and finally performing gait recognition by fusing these two gait types"—using 3D CNNs for gait image classification suffers from high computational complexity, consumes significant computational resources, and has low practicality. Furthermore, this method ignores edge details in local images when acquiring local features, resulting in insufficient feature extraction and the loss of some key features. Additionally, existing methods for acquiring local features often apply the same feature extraction algorithm to each segment of the gait image, failing to focus on specific gait regions less affected by occlusion. This leads to the loss of key features and insufficient extraction of gait discriminative features. Summary of the Invention

[0003] The purpose of this invention is to address the problem that existing technologies suffer from insufficient feature extraction and neglect of highly discriminative human body parts, resulting in the loss of some key features and inadequate extraction of gait discriminative features. This invention proposes a gait recognition method that combines local correlation features and discriminative features.

[0004] To achieve the above objectives, the technical solution of the present invention is: a gait recognition method combining local correlation features and discriminability features, comprising the following steps:

[0005] Step 1: Prepare the dataset using the popular gait datasets CASIA-B and OUMVLP in the current field of gait recognition.

[0006] Step 2: Perform data preprocessing on the dataset to ensure that the size of the human body outline is not affected by distance factors.

[0007] Step 3: Obtain the gait local correlation feature M that preserves edge details. LR :

[0008] 3.1 Input the B gait sample sequences containing N frames into the network model frame by frame, and use a 2-layer convolutional neural network and a 1-layer max pooling operation to extract the initial frame-level features;

[0009] 3.2 The LRConv convolutional network module for local correlation feature extraction is used to perform feature extraction operations on the frame-level feature sequence to obtain gait features Q with local correlation. GR Gait local detail features L that preserve edge detail information D , feature L D and Q GR The fusion process yields context-sensitive local human detail features M. LR ;

[0010] Step 4: Use the Human Body Part Focusing Module (HBF) to obtain discriminative features.

[0011] Step 5: Use the LRConv and HBF modules alternately.

[0012] Step 6: Assign local correlation features M to gait LR The input is fed into the horizontal pooling module, where global average pooling and global max pooling are used to process the features M. LR The horizontal plane is divided into n tensor strips.

[0013] Step 7: Train the network using a joint training strategy of cross-entropy loss function and triplet loss function.

[0014] Furthermore, in step 3.2 above, the convolutional network module LRConv consists of a global correlation feature extraction module GRFE and a detail feature extraction module LFE; the LFE module consists of a detail feature extraction module DFE and an edge feature extraction module EFE.

[0015] First, in the Global Relevance Feature Extraction (GRFE) module, a global convolution operation is performed on the initial frame-level feature sequence to obtain gait features Q with location correlation.GR ;

[0016] The DFE module in the LFE module performs a convolution operation on each of the divided blocks to obtain the detailed features L of the local region of the gait. D1 ;

[0017] The EFE module in the LFE module obtains the edge detail features missing by the DFE module. D2 .

[0018] Then L D1 and L D2 Pixel-level fusion is performed to obtain gait local detail features L that retain edge detail information. D .

[0019] Furthermore, in step 3.1 above, a two-layer convolutional neural network and a single-layer max-pooling operation are used to extract f for each frame. i Frame-level features g i Let i ∈ 1, 2, 3, ..., t, where i represents the index of the input gait sequence, expressed by the formula:

[0020] g i =Maxpool2d(Conv2d(Conv2d(f i ))) (1)

[0021] Furthermore, in section 3.2 above,

[0022] In the GRFE module, the initial frame-level feature sequence g = {g i |i∈1,2,3,...,t∈R B×N×C×H×W A single global convolution operation yields the global feature Q, which exhibits location-related characteristics. GR This can be expressed as a formula:

[0023] Q GR =f' 3×3 (g)∈R B×N×C×H×W (2)

[0024] In the detail feature extraction (DFE) module, feature g i Represented as w i w i Divide the data horizontally into a predefined number of m blocks. This represents the j-th feature map segmented from the i-th feature. For each piece Performing convolution operations yields the local detail features L of the corresponding blocks. D1

[0025] In the edge feature extraction module (EFE), feature g i Represented as q i In order to preserve and Edge detail features between, feature q i Divide into a predefined number of m+1 blocks, at this time Then for each piece Perform convolution operations to obtain the edge detail features L between each block. D2

[0026] Finally, L D1 and L D2 The fusion process yields local detail features L that retain edge detail characteristics. D Expressed by the formula:

[0027]

[0028] f' 3×3 =max[0,Conv2D(*)]+a*min[0,Conv2D(*)] (4)

[0029] The acquired local detail features L D Global features Q with location-related characteristics GR The fusion yields gait local correlation features M that retain edge detail information. LR .

[0030] Furthermore, in step 4 above:

[0031] Use the HBF module to process the local correlation features M output by the previous module. LR ∈R B×N×C×H×W Horizontally split into m blocks, at this time This represents the j-th part of the i-th feature map. The image blocks before the r block are the image blocks corresponding to the human head, and the image blocks after the mr block are the image blocks corresponding to the human legs below the knees.

[0032] In each feature map, the feature corresponding to the head is denoted as M1∈R. B×N×C×r×W The features corresponding to the area below the knee are denoted as M2∈R B×N×C×r×W The features corresponding to the remaining parts are denoted as M3∈R B×N×C×(H-2r)×W The channel attention mechanism network ECANet is introduced to perform feature enhancement operations on M1 and M2, resulting in features M'1∈R. B×N×C×r×W and M'2∈R B×N×C×r×W By fusing M'1 and feature M1 to preserve the global correlation of the image, the fused feature is M”1∈R. B×N×C×r×WSimilarly, fusing M'2 and M2 yields M”2∈R B×N×C×r×W Finally, M”1, M”2, and M3 are stitched together horizontally to obtain the local correlation feature N, which is less affected by occlusion and has stronger discriminative power. This is expressed by the formula:

[0033]

[0034]

[0035]

[0036] Where A represents the channel attention weight of feature M1, and B represents the channel attention weight of feature M2. Represents the feature map M * The i-th channel, C represents the number of channels, f(*) represents one-dimensional convolution, σ(*) represents the sigmoid activation function, conv represents one-dimensional convolution, and avg_pool represents AdaptiveAvgPool2d.

[0037] Furthermore, in step 6 above, the horizontal pooling module HP is used to pool feature M. LR The horizontal division into n tensor strips aims to extract gait feature sequences M. LR Extract more distinguishable human identity information from the feature map M, and use global average pooling and global max pooling to refine the feature map M. LR Perform a downsampling operation to obtain feature F i ,

[0038] F i =MaxPool2D(P i )+AvgPool2D(P i (8)

[0039] F = {F i |i=1,2,3,...,t},F∈R B×N×C×H .

[0040] Furthermore, in step 7 above, the final features extracted by the network are fed into the loss function, which is defined as: Loss = L t +L c L t L represents the triplet loss value. c This represents the cross-entropy loss value.

[0041]

[0042]

[0043]

[0044] C i C p and C n These represent the class, positive samples, and negative samples of the current input data, respectively. Positive samples represent other samples of the same class as the current input data, and negative samples represent samples of a different class than the current input data. N represents the number of samples, and C represents the number of classes. i This indicates the relationship with the current input data x. i The number of samples in the same category, N k This indicates the relationship with the current input data x. i The number of samples in different categories. This represents the Euclidean distance between the current input data and the j-th positive sample in the same category. Let [t] represent the Euclidean distance between the current input data and the k-th negative sample from other categories. + This indicates taking the maximum value between 0 and t.

[0045]

[0046] M is the total number of samples, N is the number of training objects, and y ic The sign function is 0 or 1, representing the predicted identity information of sample i. It takes the value 1 if the true class of sample i is equal to c, and 0 otherwise; P ic This is the predicted probability value for observed sample i to belong to category c.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. This invention improves the accuracy and robustness of gait recognition by using the Local Correlation Feature Extraction (LRConv) convolutional network module and the Body Part Focusing (HBF) module. The LRConv module can simultaneously utilize both global and local human body information, and the local detail features extracted by this module can preserve human body edge information. The HBF module focuses on gait regions less affected by occlusion, obtaining more discriminative gait detail features.

[0049] 2. To simultaneously utilize both global and local information about the human body and ensure that the extracted local detail features retain edge information of local human body regions, this invention first designs a local correlation feature extraction convolutional network module, LRConv. Secondly, to address the issue that current methods based on local human body regions do not focus on areas of the human body less affected by clothing occlusion, a human body part focusing attention mechanism module, HBF, is used to focus on local locations of the human body less affected by occlusion, reducing the impact of clothing on gait recognition and thus extracting more discriminative gait detail features.

[0050] 3. Experiments were conducted on two popular public datasets, CASIA-B and OU-MVLP. The results show that, in the NM, BG, and CL gait environments set in the CASIA-B dataset, GaitLRDF improves the performance of state-of-the-art algorithms by 0.48%, 0.74%, and 2.12%, respectively, and achieves a recognition accuracy of 91.35% on the OU-MVLP dataset, demonstrating the superiority of the GaitLRDF algorithm. Attached Figure Description

[0051] Figure 1 Basic information about the dataset used;

[0052] Figure 2 This is a flowchart illustrating the overall workflow of the GaitLRDF algorithm of this invention.

[0053] Figure 3 LRConv workflow diagram;

[0054] Figure 4 Comparison of local feature representations extracted by traditional methods (a) and LRConv (b);

[0055] Figure 5 HBF workflow diagram; Detailed Implementation

[0056] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, the following will describe the invention in detail with reference to some embodiments. It should be noted that the described embodiments are only a part of the embodiments of this invention, and not all of them. Embodiments obtained by those skilled in the art based on the embodiments of this invention without creative work are all within the protection scope of this invention.

[0057] In this network framework, we designed the Local Relationship Feature Extraction Convolutional Network LRConv module to obtain richer gait features with contextual relationships and solve the problem of losing edge detail features in existing methods based on local feature representation. Then, through our designed Human Body Part Focusing Module HBF, we focus on a gait part of the human body that is less affected by occlusion and extract more discriminative gait features, thereby improving the accuracy of gait recognition.

[0058] This invention provides a gait recognition method (GaitLRDF) that combines local correlation features and discriminative features, comprising the following steps:

[0059] Step 1: Prepare the dataset, see [link / reference] Figure 1The dataset mentioned is either CASIA-B or OUMVLP, which are currently popular gait datasets in the field of gait recognition. This embodiment uses the CASIA-B dataset as an example, specifically:

[0060] The proposed network model was trained and evaluated on the CASIA-B dataset. The CASIA-B dataset is a public gait dataset produced by the Chinese Academy of Sciences. This dataset contains the identity information of 124 individuals. Each individual's identity information consists of 110 gait sequences, which were captured by cameras located at 11 different viewpoints (e.g., 0, 18, 36, ..., 90, ..., 180). Under each viewpoint, each individual completed walking under three different conditions: normal walking (NM) (6 sequences), walking with a bag (BG) (2 sequences), and walking with a coat (CL) (2 sequences).

[0061] Step 2: Perform data preprocessing operations on the dataset:

[0062] Since the pedestrian silhouettes in the official dataset typically occupy a small proportion of the overall gait image, the image preprocessing workflow is as follows to reduce background interference and ensure that the size of the human silhouette is unaffected by distance: First, the top, bottom, and horizontal center of the human silhouette region in each frame are obtained. Then, a moving average filter is applied to these locations to normalize the size of the human silhouette region, ensuring its height is only 128 pixels based on the average location, while maintaining the aspect ratio of each region. Finally, we produce an 88×128 pixel image, where the average median corresponds to the horizontal center of the image.

[0063] Step 3: Since traditional gait recognition methods lose edge detail features of local human body regions when acquiring local correlation features, this invention uses the Local Correlation Feature Extraction Convolutional Network LRConv module to obtain gait local correlation features M that retain edge detail information. LR This reduces the loss of key gait features, resulting in extracted gait features containing more crucial information, specifically including:

[0064] Step 301: Input the B gait sample sequences containing N frames into the network model frame by frame. First, use a 2-layer convolutional neural network and a single-layer max pooling operation to extract f for each frame. i Frame-level features g i Let i ∈ 1, 2, 3, ..., t, where i represents the index of the input gait sequence, expressed by the formula:

[0065] g i =Maxpool2d(Conv2d(Conv2d(f i ))) (1)

[0066] Step 302: Use the Local Correlation Feature Extraction Convolutional Network module LRConv to perform feature extraction on the frame-level feature sequence. See [link to relevant documentation]. Figure 3 In the diagram, ⊕ represents pixel-level addition. The LRConv module consists of the Global Relational Feature Extraction (GRFE) module and the Minute Feature Extraction (LFE) module.

[0067] In the GRFE module, the initial frame-level feature sequence g = {g i |i∈1,2,3,...,t∈R B×N×C×H×W A single global convolution operation yields the global feature Q, which exhibits location-related characteristics. GR This can be expressed as a formula:

[0068] Q GR =f' 3×3 (g)∈R B×N×C×H×W (2)

[0069] The LFE module consists of a detail feature extraction module (DFE) and an edge feature extraction module (EFE). The DFE module performs a convolution operation on each of the divided blocks to obtain the detail features (L) of the local region of the gait. D1 The EFE module retrieves the edge detail features L that are missing from the DFE module. D2 The specific details are as follows:

[0070] In the detail feature extraction (DFE) module, feature g i Represented as w i w i Divide the data horizontally into a predefined number of m blocks. This represents the j-th feature map segmented from the i-th feature. For each piece Performing convolution operations yields the local detail features L of the corresponding blocks. D1 ;

[0071] In the edge feature extraction module (EFE), feature g i Represented as q i In order to preserve and Edge detail features between, feature q i Divide into a predefined number of m+1 blocks, at this time Then for each piece Perform convolution operations to obtain the edge detail features L between each block. D2 .

[0072] Finally, L D1 and L D2The fusion process yields local detail features L that retain edge detail characteristics. D Expressed by the formula:

[0073]

[0074] f' 3×3 =max[0,Conv2D(*)]+a*min[0,Conv2D(*)] (4)

[0075] The parameter α is initialized to 0.01. (See the attached diagram in the manual.) Figure 4 This demonstrates a comparison between the local feature representation of traditional methods and LRConv local feature representation, from... Figure 4 As can be seen in (b), the local feature extraction module of LRConv can obtain richer gait detail features and preserve gait edge information.

[0076] Step 303: Transfer feature Q GR and L D Feature fusion is performed to obtain context-sensitive local detail features M of human gait that retain edge details. LR ∈R B×N×C×H×W This reduces the loss of key features, thereby enriching the extracted gait features.

[0077] Step 4: Since methods based on local human features perform the same feature extraction operations on the segmented gait local images, they do not consider focusing on local human regions less affected by occlusion after segmentation. Therefore, the Human Part Focusing Module (HBF) is used to obtain discriminative features. Based on the attention mechanism network ECANet, the HBF module is designed. This module focuses on gait local regions less affected by occlusion, obtaining more discriminative gait detail features. Specifically, it includes:

[0078] Step 401: Use the HBF module to process the local correlation features M output by the previous module. LR ∈R B×N×C×H×W The system is horizontally divided into m blocks. The specific workflow is shown in the attached diagram of the instruction manual. Figure 5 As shown, Figure 5 In the diagram, Func represents ECANet, ⊕ represents pixel-level addition, and ⊙ represents dot product operation. This indicates feature level splicing. This represents the j-th part of the i-th feature map. The image blocks before block r represent the human head, and the image blocks after block mr represent the legs below the knees. In each feature map, the feature corresponding to the head is denoted as M1∈R. B×N×C×r×W The features corresponding to the area below the knee are denoted as M2∈R B×N×C×r×W The features corresponding to the remaining parts are denoted as M3∈R B ×N×C×(H-2r)×W The channel attention mechanism network ECANet is introduced to perform feature enhancement operations on M1 and M2, resulting in feature M'. M ∈R B×N×C×r×W and M'2∈R B×N×C×r×W By fusing M'1 and feature M1 to preserve the global correlation of the image, the fused feature is M”1∈R. B×N×C×r×W Similarly, fusing M'2 and M2 yields M”2∈R B×N×C×r×W Finally, M”1, M”2, and M3 are stitched together horizontally to obtain the local correlation feature N, which is less affected by occlusion and has stronger discriminative power. This is expressed by the formula:

[0079]

[0080]

[0081]

[0082]

[0083] Where A represents the channel attention weight of feature M1, and B represents the channel attention weight of feature M2. Represents the feature map M * The i-th channel, C represents the number of channels, f(*) represents one-dimensional convolution, σ(*) represents the sigmoid activation function, conv represents one-dimensional convolution, and avg_pool represents AdaptiveAvgPool2d.

[0084] Step 5: Interleave the use of the LRConv and HBF modules, such as... Figure 2 As shown, a total of 3 layers of LRConv modules and 2 HBF modules were used.

[0085] Step 6: Use the horizontal pooling module HP to extract the gait local correlation features M LR The input is fed into a horizontal pooling module, which horizontally divides it into n tensor strips, aiming to extract gait feature sequences M. LR To extract more distinguishable human identity information from the feature map M, making the extracted features more discriminative, global average pooling and global max pooling are used. LR Perform a downsampling operation to obtain feature F i ,

[0086] F i =MaxPool2D(P i)+AvgPool2D(P i (8)

[0087] F = {F i |i=1,2,3,…,t},F∈R B×N×C×H .

[0088] Step 7: Train the network using a joint training strategy of cross-entropy loss function and triplet loss function. The cross-entropy loss function increases the inter-class distance of different identity features, while the triplet loss function reduces the intra-class distance of the same identity feature, thus obtaining the optimal model parameters to further improve gait recognition accuracy. Specifically, during training, the final features extracted by the network are fed into the loss function, which is defined as: Loss = L t +L c L t L represents the triplet loss value. c This represents the cross-entropy loss value.

[0089]

[0090]

[0091]

[0092] C i C p and C n These represent the class, positive samples, and negative samples of the current input data, respectively. Positive samples represent other samples of the same class as the current input data, and negative samples represent samples of a different class than the current input data. N represents the number of samples, and C represents the number of classes. i This indicates the relationship with the current input data x. i The number of samples in the same category, N k This indicates the relationship with the current input data x. i Number of samples in different categories. This represents the Euclidean distance between the current input data and the j-th positive sample in the same category. Let [t] represent the Euclidean distance between the current input data and the k-th negative sample from other categories. + This indicates taking the maximum value between 0 and t.

[0093]

[0094] M is the total number of samples, N is the number of training objects, and y ic The sign function is 0 or 1, representing the predicted identity information of sample i. It takes the value 1 if the true class of sample i is equal to c, and 0 otherwise; P icThis is the predicted probability value for observed sample i to belong to category c.

[0095] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A gait recognition method combining local correlation features and discriminability features, comprising the following steps: Step 1: Prepare the dataset, using gait datasets that are currently popular in the field of gait recognition; Step 2: Perform data preprocessing on the dataset to ensure that the size of the human body outline is not affected by distance factors; Step 3: Obtain gait local correlation features M that preserve edge details LR : Step 3.1 Input the B gait sample sequences containing N frames into the network model frame by frame, and use a 2-layer convolutional neural network and a 1-layer max pooling operation to extract the initial frame-level features; Step 3.2 uses the Local Correlation Feature Extraction Convolutional Network module LRConv to perform feature extraction on the frame-level feature sequence, obtaining gait features Q with location correlation. GR Gait local detail features L that preserve edge detail information D , feature L D and Q GR The fusion process yields context-sensitive local human detail features M. LR ; Step 4: Use the Human Body Part Focusing Module (HBF) to obtain discriminative features; In step 4 Use the HBF module to extract the local correlation features output by the previous module. Horizontally split into m blocks, at this time , This represents the j-th part of the i-th feature map. The image blocks before the r block are the image blocks corresponding to the human head, and the image blocks after the mr block are the image blocks corresponding to the human legs below the knees. In each feature map, the feature corresponding to the head is denoted as... The features corresponding to the area below the knee are denoted as The features corresponding to the remaining parts are denoted as The channel attention mechanism ECANet was introduced to address this issue. and Perform feature enhancement operations to obtain the following features: and ,Will and characteristics To maintain the global correlation of images through fusion, the features after fusion are: Similarly, and By fusion Finally, for , and Horizontal stitching yields local correlation features N that are less affected by occlusion and have stronger discriminative power, expressed by the formula: Where A represents the feature The channel attention weights, where B represents the feature. Channel attention weights, Represents feature mapping The i-th channel, C represents the number of channels, and f(*) represents a one-dimensional convolution. (*) represents the sigmoid activation function, conv represents 1D convolution, and avg_pool represents AdaptiveAvgPool2d; Step 5: Interleave the use of the LRConv and HBF modules; Step 6: Assign local correlation features M to gait LR The input is fed into the horizontal pooling module, where global average pooling and global max pooling are used to process the features M. LR Divide the horizontal plane into n tensor strips; Step 7: Train the network using a joint training strategy of cross-entropy loss function and triplet loss function.

2. The gait recognition method combining local correlation features and discriminability features according to claim 1, characterized in that: In step 3.2, the convolutional network module LRConv consists of a global correlation feature extraction module GRFE and a detail feature extraction module LFE; the LFE module consists of a detail feature extraction module DFE and an edge feature extraction module EFE. First, in the Global Relevance Feature Extraction (GRFE) module, a global convolution operation is performed on the initial frame-level feature sequence to obtain gait features Q with location correlation. GR ; The DFE module in the LFE module performs a convolution operation on each of the divided blocks to obtain the detailed features of the gait local region. D1 ; The EFE module in the LFE module obtains the edge detail features missing by the DFE module. D2 ; Then L D1 and L D2 Pixel-level fusion is performed to obtain gait local detail features L that retain edge detail information. D .

3. The gait recognition method combining local correlation features and discriminability features according to claim 2, characterized in that: In step 3.1, a two-layer convolutional neural network and a single-layer max pooling operation are used to extract each frame. Frame-level features , , where i represents the index of the input gait sequence, is expressed by the formula: （1）。 4. The gait recognition method combining local correlation features and discriminability features according to claim 3, characterized in that: In section 3.2, In the GRFE module, the initial frame-level feature sequence A single global convolution operation yields global features with local correlations. This can be expressed as a formula: (2) In the detail feature extraction (DFE) module, features Represented as , Divide the data horizontally into a predefined number of m blocks. , Indicates the first The first feature is divided into Block feature map, For each piece Performing convolution operations yields the local detail features of the corresponding blocks. ; In the edge feature extraction module (EFE), features Represented as In order to preserve and Edge detail features between features Divide into a predefined number of m+1 blocks, at this time , , Then for each piece Perform convolution operations to obtain edge detail features between each block. ; Finally and The fusion process yields local detail features that retain edge detail characteristics. Expressed by the formula: （3） （4）。 5. The gait recognition method combining local correlation features and discriminability features according to claim 4, characterized in that: In step 6, the horizontal pooling module HP is used to pool the features. The horizontal division into n tensor strips aims to extract gait feature sequences. Extract more distinguishable human identity information from the feature map, and use global average pooling and global max pooling to refine the feature map. Features are obtained by performing downsampling operations. , （8）， ,F∈R B×N×C×H 。 6. The gait recognition method combining local correlation features and discriminability features according to claim 5, characterized in that: In step 7, the final features extracted by the network are fed into the loss function, which is defined as follows: , This represents the triplet loss value. This represents the cross-entropy loss value. , and These represent the category, positive samples, and negative samples of the current input data, respectively; positive samples represent other samples of the same category as the current input data, and negative samples represent samples of a different category than the current input data; N represents the number of samples, and C represents the number of categories. Indicates the current input data Number of samples in the same category Indicates the current input data The number of samples in different categories; This represents the Euclidean distance between the current input data and the j-th positive sample in the same category. This represents the Euclidean distance between the current input data and the k-th negative sample in other categories. This indicates taking the maximum value between 0 and t; M is the total number of samples, and N is the number of training objects. The sign function is 0 or 1, representing the predicted identity information of sample i. It is 1 if the true category of sample i is equal to c, and 0 otherwise. This is the predicted probability value for observed sample i to belong to category c.

Images