A gait recognition method based on sample adaptive representation

Abstract

This invention discloses a gait recognition method based on sample-adaptive representation. By incorporating meta-knowledge into gait recognition, this method achieves sample adaptation, enabling the model to better perceive a wide range of complex scenarios, such as angles and conditions. The method specifically involves the following steps: acquiring gait data; defining an optimization objective; learning meta-knowledge using a meta-hypernetwork; applying attention to the spatial, temporal, and channel dimensions using meta-knowledge; integrating temporal information using meta-knowledge; and iterative training. This method is suitable for gait recognition in complex scenarios, leveraging meta-knowledge to achieve good results and generalize to diverse internal and external conditions.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a gait recognition method based on sample adaptive representation. Background Art

[0002] Gait recognition is defined as the problem of identifying pedestrians' specific walking patterns and identifying them from video sequences. Gait recognition has broad applications in fields such as smart cities, security systems, and video retrieval. In recent years, significant research has focused on improving gait recognition performance through human body modeling. However, gait recognition is often affected by numerous influencing factors, such as whether a person is wearing a coat or carrying a backpack. Accurately extracting effective features from gait under these numerous conditions is a key challenge in gait recognition.

[0003] Mainstream gait learning methods utilize deep learning technology. Existing deep learning-based methods input video sequences and output pedestrian identities. Direct end-to-end deep learning can address basic feature representation, but gait recognition faces challenges in adaptively capturing changes in gait motion patterns caused by changing conditions, including temporal, spatial, channel, and temporal integration. Summary of the Invention

[0004] To address the above-mentioned issues, the present invention aims to resolve the contradiction between the numerous conditional changes in gait recognition and robust gait representation, and to provide a gait recognition method based on sample adaptive representation. First, the present invention proposes a meta-supernetwork for learning meta-knowledge. Meta-knowledge can be used to learn condition-related knowledge, thereby parameterizing the backbone network to achieve sample adaptation. In addition, the present invention also introduces meta-knowledge into the calibration network of the attention mechanism, and simultaneously applies it in the three dimensions of space, time, and channels. Furthermore, the present invention introduces meta-knowledge into the attention mechanism and temporal information integration stage of gait recognition.

[0005] The technical solutions specifically adopted in the present invention are as follows:

[0006] A gait recognition method based on sample adaptive representation includes the following steps:

[0007] S1. Obtain a training dataset for training a gait recognition model;

[0008] S2. Build a gait recognition model framework and define the optimization objectives of the entire framework;

[0009] The gait recognition model framework includes a feature extractor, a meta-triple attention module, a meta-temporal pooling module, a fully connected layer and a classifier. Its input data is a set of silhouette image sequences collected by a person during walking. The input data is first extracted by the feature extractor to form a four-dimensional feature. Then, a parameterized hypernetwork is used to perform meta-knowledge learning, and then network parameters related to the input data are adaptively generated for the meta-triple attention module and the meta-temporal pooling module as learned meta-knowledge. In the meta-triple attention module, the learned meta-knowledge is used to apply an attention mechanism to the three dimensions of space, time and channel, and an adaptive scale feature is obtained for each dimension. The adaptive scale features of the three dimensions are then added and input into the meta-temporal pooling module. The learned meta-knowledge is used to fuse the time information. The fused features are passed through the fully connected layer to obtain the final features for reasoning. The final features are input into the classifier to output the classification results.

[0010] S3. Combining the triplet loss in the fully connected layer and the cross entropy loss in the classifier as the total loss function, and iteratively training the gait recognition model framework until convergence;

[0011] S4. When performing gait recognition, the silhouette image sequence of the pedestrian to be queried during walking is input into the trained gait recognition model framework, and similarity matching is performed in the library based on the final features output by the fully connected layer to determine the identity of the pedestrian.

[0012] Preferably, the training data set I train It contains silhouette image sequences collected and segmented from different people at different viewing angles and under different walking conditions.

[0013] Preferably, the feature extractor adopts a 6-layer 3D CNN model.

[0014] Preferably, the meta-knowledge learning is performed by using a hypernetwork F with the four-dimensional feature X as input meta Implementation, by the super network F meta Calibration Network F for Attention Mechanism in Parameterized Gait Recognition cali； The specific approach to meta-knowledge learning is as follows:

[0015] First, the four-dimensional feature X is globally averaged pooled in the three dimensions of space, time, and channel to obtain their respective statistical characteristics m1, m2, and m3:

[0016]

[0017]

[0018]

[0019] Where: GAP represents the global average pooling operation, H, W, T, and C represent the height, width, time series length, and number of channels of the four-dimensional feature X, respectively. The h, w, t, and c in X() represent the element indexes of the height, width, time series, and channel dimensions in X, respectively.

[0020] Then, the hypernetwork F is implemented by a two-layer multilayer perceptron. meta , respectively generate calibration networks F for the three dimensions of space, time and channel cali Parameter W cali :

[0021]

[0022] Where: and Represent the super network F meta The network parameters of the first and second multilayer perceptrons in , σ is the activation function of the multilayer perceptron; m represents the statistical characteristics of a dimension, and for the three dimensions of space, time and channel, m is m1, m2, and m3 respectively;

[0023] Finally, the parameters W generated for the three dimensions of space, time and channel are cali , which includes the network parameters related to the current input data generated for the meta-triplet attention module and the meta-temporal pooling module.

[0024] Preferably, in the meta-triple attention module, the attention mechanism is applied simultaneously in the three dimensions of space, time, and channel, and can adaptively implement arbitrary scale representation; for each dimension of space, time, and channel, the attention mechanism process is as follows:

[0025] First, based on the statistical characteristics m obtained by global average pooling of the input four-dimensional feature X, the global feature f of the four-dimensional feature X is extracted through the fully connected network in the global dependency module. global , local features are extracted through L branches in the local dependency module, where each branch uses a convolution to extract a local feature The fully connected network parameters in the global dependency module and the convolution parameters in the local dependency module are derived from the parameter W cali confirmed in

[0026] Then, the global features extracted by the fully connected layer and the local features extracted by the convolution are fused through the gating mechanism:

[0027]

[0028] Among them, f mtaIt is the adaptive scale feature obtained by fusion of the current dimension; Represents element-by-element multiplication; G(·) is a learnable gate, using a two-layer MLP network, and the gate parameters are derived from the parameters W cali G(m)[L+1] represents the soft weight of the global feature generated by gating, and G(m)[l] represents the soft weight of the local feature extracted by the l-th branch generated by gating, which is fused to form f mta Soft weights are used to control the weighted fusion of different scales, thereby achieving all-round scale representation.

[0029] As a preference, in the meta-time series pooling module, meta-knowledge is introduced into the time series information integration process in gait recognition, and the maximum pooling, average pooling and general pooling are fused using a sample adaptive method to obtain the final feature f omni :

[0030] f omni =β1Mean(F mta )+β2Max(F mta )+β3GeM(F mta ).

[0031] Among them, Mean(·), Max(·), and GeM(·) represent the average pooling operation, the maximum pooling operation, and the general pooling operation respectively. mta Represents the adaptive scale features f corresponding to the three dimensions of space, time and channel mta β1, β2, and β3 are soft weights generated by a gate, and the network parameters of the gate are based on the parameters W generated by the super network. cali Sure.

[0032] Preferably, the total loss function is the weighted sum of the triplet loss in the fully connected layer and the cross entropy loss in the classifier.

[0033] Compared with existing gait recognition methods, the gait recognition method based on sample adaptive representation of the present invention has the following advantages:

[0034] First, the gait recognition method based on sample adaptive representation of the present invention proposes to use a meta-hypernetwork to implicitly model the numerous changing conditions in gait, including angles, carried objects, etc., so as to achieve the effect of sample adaptation.

[0035] Secondly, this paper proposes to incorporate meta-knowledge into the attention mechanism and temporal information integration stages of gait recognition. The calibration network in the attention mechanism is parameterized by meta-knowledge and applied simultaneously in the spatial, temporal, and channel dimensions. In the temporal integration stage, sample-adaptive weighting of three commonly used temporal integration methods is employed to achieve a fine-grained, meta-knowledge-based temporal integration process.

[0036] Finally, our method achieves significant improvements on mainstream datasets, particularly in the most complex "wearing a coat" scenario. Specifically, average performance improves by at least 1.5%, and in the most complex scenario, it achieves at least a 2.7% improvement over previous methods. Furthermore, our method can be embedded in methods using attention mechanisms, achieving significant performance improvements, demonstrating the method's strong generalization capabilities.

[0037] The gait recognition method based on sample adaptive representation of the present invention can learn meta-knowledge for various complex scene conditions in gait recognition during the learning process, and integrate meta-knowledge into the attention mechanism and temporal information integration process. The present invention can effectively improve the accuracy and efficiency of gait recognition and has considerable application value. In the future, in the field of urban biometric recognition, the gait recognition method based on sample adaptive representation of the present invention can perform better in more diverse and complex scenes, and can provide strong guarantees for tasks such as identity verification, park management, and criminal investigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the process of the present invention;

[0039] Figure 2 This is a framework diagram of the gait recognition model of the present invention;

[0040] Figure 3 Schematic diagram of the attention mechanism using meta-knowledge of the present invention;

[0041] Figure 4 Schematic diagram of temporal information integration using meta-knowledge according to the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0044] refer to Figure 1 In a preferred embodiment of the present invention, a gait recognition method based on sample adaptive representation is provided, comprising the following steps:

[0045] S1. Obtain a training dataset for training a gait recognition model.

[0046] In this embodiment, the obtained training data set I train It contains silhouette image sequences of different people (with unique person IDs) collected and segmented under different viewing angles and different walking conditions. train It can be directly implemented using existing datasets such as CASIA-B and OU-MVLP.

[0047] S2. Construct a gait recognition model framework, which includes a feature extractor, a meta triple attention module, a meta temporal pooling module, a separate FC layer, and a classifier. The input data is a set of silhouette image sequences collected from a person walking. The input data is first extracted by the feature extractor to form a four-dimensional feature, and then a parameterized hypernetwork (meta hyper The meta-knowledge is learned by using a meta-triple attention network (MHN), and then the network parameters related to the input data are adaptively generated for the meta-triple attention module and the meta-temporal pooling module as the learned meta-knowledge. In the meta-triple attention module, the learned meta-knowledge is used to apply the attention mechanism to the three dimensions of space, time, and channel, and each dimension obtains an adaptive scale feature. The adaptive scale features of the three dimensions are then added together and input into the meta-temporal pooling module, and the learned meta-knowledge is used to fuse the time information. The fused features are passed through the fully connected layer to obtain the final features for reasoning, which are input into the classifier to output the classification results. The gait recognition model framework is as follows: Figure 2 shown.

[0048] In this embodiment, the feature extractor adopts a 6-layer 3D CNN model.

[0049] In this embodiment, the above meta-knowledge learning is achieved by using a hypernetwork F with the four-dimensional feature X as input. meta Implementation, by the super network F meta Calibration Network F for Attention Mechanism in Parameterized Gait Recognition cali , that is, W cali =F meta (X). The specific approach to meta-knowledge learning is as follows:

[0050] First, the four-dimensional feature X is globally averaged pooled in the three dimensions of space, time, and channel to obtain their respective statistical characteristics m1, m2, and m3:

[0051]

[0052]

[0053]

[0054] Where: GAP represents the global average pooling operation, H, W, T, and C represent the height, width, time series length, and number of channels of the four-dimensional feature X, respectively. The h, w, t, and c in X() represent the element indexes of the height, width, time series, and channel dimensions in X, respectively.

[0055] Then, the hypernetwork F is implemented by a two-layer multilayer perceptron. meta , respectively generate calibration networks F for the three dimensions of space, time and channel cali Parameter W cali :

[0056]

[0057] Where: and Represent the super network F meta The network parameters of the first and second multilayer perceptrons in , σ is the activation function of the multilayer perceptron; m represents the statistical characteristics of a dimension, and for the three dimensions of space, time and channel, m is m1, m2, and m3 respectively;

[0058] Finally, the parameters W generated for the three dimensions of space, time and channel are cali , which includes the network parameters related to the current input data generated for the meta-triple attention module and the meta-temporal pooling module. Specifically, the parameter W cali It includes the fully connected network parameters in the subsequent meta-triplet attention module, the convolution parameters in the branch, the learnable gated unit parameters, and the soft weight parameters in the meta-temporal pooling module.

[0059] In this embodiment, in the above-mentioned meta-triple attention module, the attention mechanism is applied simultaneously in the three dimensions of space, time and channel, and can adaptively realize arbitrary scale representation; for each dimension of space, time and channel, the attention mechanism process can be found in Figure 3 As shown, the specific process is as follows:

[0060] First, based on the statistical characteristics m obtained by global average pooling of the input four-dimensional feature X, the global feature f of the four-dimensional feature X is extracted through the fully connected network in the Global Dependency module. global , local features are extracted through L branches in the local dependency module, where each branch uses a convolution to extract a local feature In this embodiment, L=3, that is, there are 3 branches in the local dependency module, which are implemented by 1×1 convolution, 3×1 convolution and 5×1 convolution respectively. In addition, in this embodiment, the fully connected network in the global dependency module is composed of a cascade of FC layer, Leaky Relu layer and FC layer. The fully connected network parameters in the global dependency module and the convolution parameters in the local dependency module are all derived from the parameter W cali It is determined in the above, which is the result obtained by meta-knowledge learning in advance.

[0061] Then, the global features extracted by the fully connected layer and the local features extracted by the convolution are fused through the gating mechanism:

[0062]

[0063] Among them, f mta It is the adaptive scale feature obtained by fusion of the current dimension; represents the element-wise multiplication operation; G(·) is a learnable gating (weighting network) that uses a two-layer MLP network. The gating parameters are derived from the parameter W cali It is determined in the source code, which is the result obtained by meta-knowledge learning in advance; G(m)[L+1] represents the soft weight of the global feature generated by gating, and G(m)[l] represents the soft weight of the local feature extracted by the l-th branch generated by gating, which is fused to form f mta When using soft weights to control weighted fusion of different scales, Figure 3 Omni-Scale Representation shown in the middle right image.

[0064] In this embodiment, if Figure 4As shown in the above meta-time series pooling module, meta-knowledge is introduced into the temporal information integration process of gait recognition, and the maximum pooling, average pooling and general pooling are fused using the sample adaptive method to obtain the final feature f omni :

[0065] f ommni =β1Mean(F mta )+β2Max(F mta )+β3 GeM(F mta ).

[0066] Among them, Mean(·), Max(·), and GeM(·) represent the average pooling operation, the maximum pooling operation, and the general pooling operation respectively. mta Represents the adaptive scale features f corresponding to the three dimensions of space, time and channel mta The sum of β1, β2, and β3 are soft weights generated by another gating (weighting network), and the network parameters of the gating are based on the parameters W generated by the super network. cali Determined, it is the result obtained by meta-knowledge learning in advance.

[0067] S3. Define the optimization goal of the entire framework, and combine the triple loss (Triple Loss) in the fully connected layer and the cross entropy loss (CE L oss ) are combined as the total loss function, and the gait recognition model framework is iteratively trained until convergence.

[0068] It should be noted that the classifier in the above framework is only used during training. During testing or actual application, the classification head using cross entropy will be discarded, and the features output by the fully connected layer can be directly used for inference.

[0069] In this embodiment, the total loss function is the weighted sum of the triplet loss in the fully connected layer and the cross entropy loss in the classifier. The specific calculation methods of the triplet loss and the cross entropy loss belong to the prior art and will not be described in detail.

[0070] In this embodiment, when calculating triplet loss, the triplet consists of an anchor, a positive sample, and a negative sample.

[0071] In the present invention, the sample difficulty in the modeling training data set refers to defining the difficulty of each sample in the training data set for the gait recognition learning task.

[0072] S4. When performing gait recognition, the silhouette image sequence of the pedestrian to be queried during walking is input into the trained gait recognition model framework, and similarity matching is performed in the library based on the final features output by the fully connected layer to determine the identity of the pedestrian.

[0073] Specifically, when performing actual gait recognition, it is only necessary to input the video sequence of pedestrians into the trained gait recognition model, which will then output features. By comparing the features of the pedestrians most similar to the input probe to be queried in the database, the similarity is sorted and the most likely pedestrian identity is output.

[0074] The method provided by this invention can effectively improve performance under various conditions. The above method is applied to specific examples below so that those skilled in the art can better understand the effects of the invention. Subsequent examples, on the mainstream datasets CASIA-B and OU-MVLP, fully verify the effectiveness of meta-learning in the attention mechanism and temporal integration with experimental results.

[0075] Example

[0076] The implementation method of this embodiment is as described above (i.e., S1 to S3). The specific steps will not be elaborated in detail. The following only demonstrates its effect based on case data. The present invention is implemented on two datasets with true value annotations according to certain settings, namely:

[0077] CASIA-B dataset: contains 124 person IDs, 11 viewpoints, and 13,640 gait sequences. This dataset will be mainly used as a test dataset.

[0078] OU-MVLP dataset: contains 10,307 person IDs and 14 perspectives, making it the largest academic dataset.

[0079] This example compares the MetaGait method trained using this method with other mainstream methods. Training and testing were performed on the CASIA-B and OU-MVLP datasets to verify the effectiveness of this method. The final test results and recognition accuracy are shown in Tables 1 and 2. This method shows significant improvements in Rank-1 performance on the CASIA-B and OU-MVLP datasets, particularly in the most complex clothing-changing scenario. It also shows significant improvements under average conditions, making it more suitable for cross-condition retrieval in real-world conditions. This lays a solid foundation for real-world applications and allows for good generalization to more complex location scenarios.

[0080] Table 1 Test results of this example on the CASIA-B dataset

[0081]

[0082] Table 2 Test results of this example on the OU-MVLP dataset

[0083]

[0084] It should be noted that in Tables 1 and 2 above, NM represents the normal scene, BG represents the backpack scene, and CL represents the dressing scene. GaitSet, GaitPart, GLN, MT3D, CSTL, GaitGL, and 3DLocal are all cutting-edge gait recognition models in the existing technology. For details, please refer to the following prior art documents:

[0085] GaitSet:Chao,H.,He,Y.,Zhang,J.,&Feng,J.(2019,July).Gaitset:Regardinggait as a set for cross-view gait recognition.In Proceedings of the AAAIconference on artificial intelligence (Vol.33,No.01,pp.8126-8133).

[0086] GaitPart: Fan, C., Peng, Y., Cao, C., Liu,

[0087] GLN: Hou, S., Cao, C., Liu, X., & Huang, Y. (2020, August). Gait lateral network: Learning discriminative and compact representations for gait recognition. InEuropean Conference on Computer Vision (pp.382-398). Springer, Cham.

[0088] MT3D:Lin,B.,Zhang,S.,&Bao,F.(2020,October).Gait Recognition withMultiple- Temporal-Scale 3D Convolutional Neural Network.In Proceedings ofthe 28th ACM International Conference on Multimedia(pp.3054-3062).

[0089] X.Huang et al.,"Context-Sensitive Temporal Feature Learning for GaitRecognition,"2021 IEEE / CVF International Conference on Computer Vision(ICCV),2021,pp.12889-12898,doi: 10.1109 / ICCV48922.2021.01267.

[0090] Z.Huang et al.,"3D Local Convolutional Neural Networks for GaitRecognition,"2021 IEEE / CVF International Conference on Computer Vision(ICCV),2021,pp.14900-14909,doi: 10.1109 / ICCV48922.2021.01465.

[0091] Lin,B.,Zhang,S.,&Yu,X.(2021).Gait Recognition via Effective Global-Local Feature Representation and Local Temporal Aggregation.2021IEEE / CVFInternational Conference on Computer Vision(ICCV),14628-14636.

[0092] Through the above technical solutions, the present invention implements a sample-adaptive gait recognition method. This method leverages learned meta-knowledge to empower the attention mechanism and temporal information integration in gait recognition, enabling the model to adapt to diverse conditions. This model demonstrates impressive performance on mainstream datasets and is easy to implement, providing promising application prospects for deployment in real-world scenarios.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A gait recognition method based on sample adaptive representation, characterized in that: The following steps are involved: S1. Obtain a training dataset for training a gait recognition model; S2. Construct a gait recognition model framework, which includes a feature extractor, a meta-triple attention module, a meta-temporal pooling module, a fully connected layer, and a classifier. The input data is a sequence of silhouette images collected from a person walking. The input data is first extracted by the feature extractor to form a four-dimensional feature. A parameterized hypernetwork is then used for meta-knowledge learning. Network parameters related to the input data are then adaptively generated for the meta-triple attention module and the meta-temporal pooling module as learned meta-knowledge. In the meta-triple attention module, the learned meta-knowledge is used to apply an attention mechanism to the three dimensions of space, time, and channel, respectively, and an adaptive scale feature is obtained for each dimension. The adaptive scale features of the three dimensions are then added together and input into the meta-time series pooling module. The learned meta-knowledge is used to fuse the temporal information. The fused features are passed through a fully connected layer to obtain the final features used for inference. The final features are input into the classifier to output the classification results. S3. Combining the triplet loss in the fully connected layer and the cross entropy loss in the classifier as the total loss function, and iteratively training the gait recognition model framework until convergence; S4. When performing gait recognition, the silhouette image sequence of the pedestrian to be queried during walking is input into the trained gait recognition model framework, and similarity matching is performed in the library based on the final features output by the fully connected layer to determine the identity of the pedestrian.

2. The gait recognition method based on sample adaptive representation according to claim 1, characterized in that: The training data set I train It contains silhouette image sequences collected and segmented from different people at different viewing angles and under different walking conditions.

3. The gait recognition method based on sample adaptive representation according to claim 1, characterized in that: The feature extractor adopts a 6-layer 3D CNN model.

4. The gait recognition method based on sample adaptive representation according to claim 1, characterized in that: The meta-knowledge learning is achieved by using a hypernetwork F that takes the four-dimensional feature X as input. meta Implementation, by the super network F meta Calibration Network F for Attention Mechanism in Parameterized Gait Recognition cali The specific approach to meta-knowledge learning is as follows: First, the four-dimensional feature X is globally averaged pooled in the three dimensions of space, time, and channel to obtain their respective statistical characteristics m1, m2, and m3: Where: GAP represents the global average pooling operation, H, W, T, and C represent the height, width, time series length, and number of channels of the four-dimensional feature X, respectively. The h, w, t, and c in X() represent the element indexes of the height, width, time series, and channel dimensions in X, respectively. Then, the hypernetwork F is implemented by a two-layer multilayer perceptron. meta , respectively generate calibration networks F for the three dimensions of space, time and channel cali Parameter W cali : Where: and Represent the super network F meta The network parameters of the first and second multilayer perceptrons in , σ is the activation function of the multilayer perceptron; m represents the statistical characteristics of a dimension, and for the three dimensions of space, time and channel, m is m1, m2, and m3 respectively; Finally, the parameters W generated for the three dimensions of space, time and channel are cali , which includes the network parameters related to the current input data generated for the meta-triplet attention module and the meta-temporal pooling module.

5. The gait recognition method based on sample adaptive representation according to claim 4, characterized in that: In the meta-triple attention module, the attention mechanism is applied simultaneously in the three dimensions of space, time, and channel, and can adaptively achieve arbitrary scale representation. For each dimension of space, time, and channel, the attention mechanism process is as follows: First, based on the statistical characteristics m obtained by global average pooling of the input four-dimensional feature X, the global feature f of the four-dimensional feature X is extracted through the fully connected network in the global dependency module. global , local features are extracted through L branches in the local dependency module, where each branch uses a convolution to extract a local feature The fully connected network parameters in the global dependency module and the convolution parameters in the local dependency module are derived from the parameter W cali confirmed in Then, the global features extracted by the fully connected layer and the local features extracted by the convolution are fused through the gating mechanism: Among them, f mta It is the adaptive scale feature obtained by fusion of the current dimension; Represents element-by-element multiplication; G(·) is a learnable gate, using a two-layer MLP network, and the gate parameters are derived from the parameters W cali G(m)[L+1] represents the soft weight of the global feature generated by gating, and G(m)[l] represents the soft weight of the local feature extracted by the l-th branch generated by gating, which is fused to form f mta Soft weights are used to control the weighted fusion of different scales, thereby achieving all-round scale representation.

6. The gait recognition method based on sample adaptive representation according to claim 5, characterized in that: In the meta-time series pooling module, meta-knowledge is introduced into the time series information integration process in gait recognition, and the maximum pooling, average pooling and general pooling are fused using a sample adaptive method to obtain the final feature f omni : f omni =β1Maen(F mta )+β2Max(F mta )+β3GeM(F mta ). Among them, Mean(·), Max(·), and GeM(·) represent the average pooling operation, the maximum pooling operation, and the general pooling operation respectively. mta Represents the adaptive scale features f corresponding to the three dimensions of space, time and channel mta β1, β2, and β3 are soft weights generated by a gate, and the network parameters of the gate are based on the parameters W generated by the super network. cali Sure.

7. The gait recognition method based on sample adaptive representation according to claim 1, characterized in that: The total loss function is the weighted sum of the triple loss in the fully connected layer and the cross entropy loss in the classifier.

Images