Model-independent single-domain generalization cross-domain gait recognition method, device and product

Abstract

The invention discloses a model-independent single-domain generalization cross-domain gait recognition method, device and product. The method comprises the following steps: firstly, collecting gait data from a plurality of source domains and extracting fine-grained features; then, extracting identity invariant domain features in the fine-grained features; calculating the centroid of each sub-field according to the extracted identity invariant field features; generating a sub-domain label by using the calculated centroid; and finally, extracting domain-invariant identity features to obtain a cross-domain gait recognition result. The method effectively solves the problem of domain offset, and has high application potential.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition technology, and relates to a single-domain generalized cross-domain gait recognition method, device and product, and specifically to a model-independent single-domain generalized cross-domain gait recognition method, device and product. Background Art

[0002] In today's society, with the rapid development of artificial intelligence and deep learning technology, gait recognition, as a biometric recognition technology, has attracted widespread attention due to its advantages in long-distance recognition and difficulty in disguise. Gait recognition identifies different people by analyzing their walking patterns, which has important application value in video surveillance, forensic identification, and crime prevention. In recent years, with the rise of deep learning technology, the field of gait recognition has made significant progress, and many methods have shown high accuracy on specific data sets. However, due to domain shift caused by scene changes, such as observation angle and clothing changes, the performance of these models on unseen data will drop significantly, limiting the widespread application of gait recognition technology.

[0003] In the existing gait recognition technology, researchers have been exploring how to improve the adaptability of the model under different conditions. As a biometric recognition technology, gait recognition relies on the unique movement pattern of an individual during walking. However, traditional gait recognition methods often face challenges brought by insufficient data sets and environmental changes. In order to enhance the robustness and accuracy of the system, researchers have adopted the following methods:

[0004] 1. Multimodal approach: In recent years, researchers have begun to combine different types of sensor data, such as video cameras, electromyographic signals, posture information, acceleration, gyroscopes, ground reaction force sensors, etc., to fully capture gait characteristics. This multimodal approach can improve the accuracy of feature extraction and enhance the model's adaptability to environmental changes (such as lighting, background, and individual status) by fusing data from different sensors. In addition, the use of multimodal data can also compensate for errors that may occur in a single sensor, thereby improving overall recognition performance.

[0005] For multimodal gait recognition technology, Chinese patent document number CN118349949A, published on July 16, 2024, records a multimodal gait recognition method and system. The invention uses an encoder to extract deep features of electromyographic signals and posture signals; uses a fusion module to fuse the deep features of electromyographic signals and posture signals to obtain deep fusion features; uses a decoder to decode the deep fusion features to obtain gait recognition results. This technology has the following problems: First, feature extraction has limitations. In practical applications, the collection of electromyographic signals and posture signals may be affected by environmental noise (such as other people's activities, background noise, etc.), resulting in unstable extracted features. In addition, differences in electromyographic activity and posture among different individuals will cause different signals to be generated for the same action, increasing the complexity of recognition. This difference between individuals may cause deviations in model training. Second, the fusion strategy is insufficient. Although deep feature fusion is adopted, if the fusion module fails to make full use of the complementary information between different modalities, it may lead to insufficient information after feature fusion, affecting the final recognition performance. If the fusion process fails to consider the nonlinear relationship between features, it may lead to the loss of some important information and affect the accuracy of the model.

[0006] 2. Multi-view recognition: Since gait may change significantly under different observation angles, researchers are also exploring cross-view recognition methods. By collecting gait data from multiple perspectives, researchers can integrate information from different angles to extract more stable and perspective-invariant gait features. Some researchers introduce perspective prediction mechanisms into the model to identify changes in gait features under different perspectives, thereby helping the system to accurately recognize under unseen perspectives. By using deep learning technologies such as deep convolutional neural networks (CNNs) and building a specific network structure, the model can learn the mapping relationship between different perspectives, thereby performing more effective feature extraction under different perspectives.

[0007] For gait recognition technology based on multi-perspective, Chinese patent document number CN114360058A, published on May 28, 2024, records a cross-perspective gait recognition method based on walking perspective prediction. Based on the use of deep convolutional network methods, the invention adds a perspective prediction network, uses sample perspective information as a supervisory signal to effectively model the perspective information in the gait energy map, and assists in network model training, and finally can extract gait features with perspective invariance. This technology has the following problems: First, it relies on diverse data sets. The performance of the perspective prediction network is highly dependent on the quality of the training data. If the training data set lacks diversity (such as different scenes, lighting conditions, and individual differences), it may lead to inaccurate perspective prediction, thereby affecting subsequent feature extraction. If the model fails to effectively cover various perspective changes during training, it may not be able to correctly identify new or unseen perspectives in practical applications, resulting in recognition failure. Second, model complexity: adding a perspective prediction network will increase the complexity of the model, which may lead to higher computational overhead, especially in scenarios that require real-time processing, which may affect the response speed of the system. Complex network structures may lead to increased training and inference times, especially when large-scale datasets need to be processed, which may affect the usability and practicality of the system.

[0008] Existing gait recognition methods show a high degree of dependence in data acquisition and testing under different scenarios. When tested with data from different scenarios, the performance of existing methods will be greatly reduced due to domain shift. Domain shift refers to the different distributions between training data (i.e., source domain) and test data (i.e., target domain). For example, changes in clothing, viewing angles, and even the segmentation method used to generate silhouettes can lead to distribution differences between datasets. Therefore, some researchers use unsupervised domain adaptation (UDA) techniques to improve the performance of cross-domain recognition. UDA methods use some labeled source domain data and some unlabeled target domain data to train the model, achieving knowledge transfer from the labeled source domain to the unlabeled target domain. However, when dealing with new data from an unseen target domain, UDA methods need to be updated with these new data, which is often difficult to achieve in practical applications due to privacy policies. Therefore, it is important that the model can generalize to data in new scenarios after being trained under old scenarios. Domain generalization (DG) is a promising solution to address the above challenges. The goal of domain generalization is to learn a model from one or more different but related source domains to perform well on unseen target domains without further updating on new data. It has received extensive attention in many vision domains such as face recognition and person re-identification. The usefulness of a gait recognition system depends on its ability to generalize to previously unencountered data. However, domain generalization in gait recognition has not been fully studied. Summary of the invention

[0009] In order to solve the technical problems of model dependence and domain shift existing in the current gait recognition technology, the present invention provides a model-independent single-domain generalized cross-domain gait recognition method, device and product, which learn a model from one or more different but related source domains so that it performs well in unseen domains without the need for further updating on new data.

[0010] The technical solution adopted by the method of the present invention is: a model-independent single-domain generalized cross-domain gait recognition method, comprising the following steps:

[0011] Step 1: Collect gait data from multiple source domains and extract fine-grained features;

[0012] Step 2: Extract identity-invariant domain features from fine-grained features;

[0013] Step 3: Calculate the centroid of each sub-domain based on the extracted identity-invariant domain features;

[0014] Step 4: Generate sub-domain labels using the calculated centroids;

[0015] Step 5: Extract domain-invariant identity features to obtain cross-domain gait recognition results.

[0016] Preferably, in step 1, the encoder f is first used to extract the gait features, and then the fine-grained features are used to extract the head h fine , convert the gait features obtained from the encoder f into a fine-grained representation;

[0017] The fine-grained feature extraction head h fine is a trained model; during training, we first collect gait data from multiple source domains and extract fine-grained features, and then use the identity classifier and domain classifiers

[0018] Assign fine-grained features to the corresponding identity labels y I and subdomain label y D ; The loss function L used in training fine for:

[0019]

[0020] Among them, f(x) represents the feature vector after gait data encoding, h fine represents the fine-grained feature extraction head, L ce is the cross entropy;

[0021] During the training process, back-propagation is used to update the model parameters, and the classification accuracy and F1 score are evaluated through the validation set to ensure the effectiveness of fine-grained feature extraction. Finally, the key gait features extracted by the analysis model are analyzed by visualizing the intermediate layer output to further optimize the model structure.

[0022] Preferably, in step 2, a domain classifier and an identity discriminator are used to extract identity-invariant domain features from fine-grained features;

[0023] The “domain classifier and identity discriminator” is a trained model; during training, the domain classifier g D and domain discriminator d I , change the domain header h D The output classification is the corresponding sub-domain label y D ; During the training process, d I and h D Participate in adversarial training, I Try h D The output of the classification is the corresponding identity label y I , and h D Try to minimize this classification error, thereby minimizing the differences between identities and obtaining identity-invariant domain features; the loss function used in training is:

[0024] L domain =L ce (g D (h D (f(x i ))),y D )+λL ce (d I (R λ (h D (f(x i )))),y I );

[0025] Among them, L ce is the cross entropy, g D is the domain classifier, h D is the domain head, f(x i ) is the feature vector after gait data encoding, y D is the subdomain label; λ is a hyperparameter, d I is the identity discriminator, R λ represents the gradient reversal layer, y I is the identity label; during training, the parameters of the encoder f are frozen and only the domain head h D and domain discriminator d I Make optimizations.

[0026] Preferably, the specific implementation of step 3 includes the following sub-steps:

[0027] Step 3.1: Data enhancement and style transfer to obtain enhanced data sets;

[0028] First, use different segmentation models S 1 ,S 2 ,…,S N To generate gait data x i The silhouette of the image is then migrated based on the segmentation model of StarGAN to convert the generated silhouette image into a unified style;

[0029] Step 3.2: After obtaining the enhanced dataset, perform feature extraction to obtain identity-invariant domain features;

[0030] Step 3.3: Initialize a centroid μ for each sub-domain k k ;

[0031] Step 3.4: Calculate the centroid by weighted average

[0032] Among them, represents the domain label of the i-th sample, k represents the index of the sub-domain; represents the indicator function, when sample i belongs to domain k, = 1, otherwise it is 0; represents the feature vector after gait data encoding; is the domain classifier, and is the domain head.

[0033] Preferably, the specific implementation of step 4 includes the following sub-steps:

[0034] Step 4.1: For each x i , calculate its centroid μ with all sub-domains k The distance between

[0035] Step 4.2: Get x i The subfield label y P =arg min dist(h D (f(x i )),μ k ), where f(x i ) represents x i The encoded feature vector, dist(), represents each x i and the subdomain centroid μ k The distance between D For the field head.

[0036] Preferably, in step 5, an identity classifier and a domain discriminator are used to extract domain-invariant identity features;

[0037] The “identity classifier and domain discriminator” is a trained model; during training, the identity classifier gI and domain discriminator d D , the identity header h I The output of the classification is the corresponding identity label y I ; During the training process, d D and h I Participate in adversarial training, D Try h I The output classification is the corresponding domain label y D , and h I Try to maximize this classification error, thereby minimizing the difference between domains and obtaining domain-invariant identity features. The loss function used in training is:

[0038] L i dentity =

[0039] L ce (g I (h I (f(x))),y I )+L trp (x,x n ,x p )+L ce (d D (R λ (h I (f(x)))),y D );

[0040] Among them, L ce is the cross entropy, g I is the identity classifier, h I is the identity head, f(x) represents the feature vector after gait data encoding, y I is the identity tag; L frp represents the triple loss, x n represents a sample with a different identity from x, x p represents the sample with the same identity as x; d D represents the domain discriminator, λ is a hyperparameter, R λ represents the gradient reversal layer, y D The sub-domain label.

[0041] Preferably, in step 5, after extracting the domain-invariant identity feature, the feature is compared with the existing features in the database. Specifically, for each gait feature in the database, the distance between the data to be identified and it is calculated. If the distance between the two is within a suitable range, the identification is successful and the gait identification result is returned.

[0042] Preferably, in step 4, the sub-domain label y P , is the trained sub-domain label yP Computational models;

[0043] The training process includes: (1) self-supervised learning: using the generated pseudo labels for self-supervised learning to improve label quality; (2) cross-validation: verifying the effectiveness of the generated labels through cross-validation technology to ensure the accuracy and reliability of the labels; (3) based on the verification results, P The calculation model parameters are adjusted to optimize the model performance and improve the accuracy of gait recognition.

[0044] The technical solution adopted by the device of the present invention is: a model-independent single-domain generalized cross-domain gait recognition device, including a user dynamic data uploading module, a sub-domain centroid calculation module, a sub-domain label generation module, a self-supervised learning and model optimization module, and a domain-invariant identity feature learning module;

[0045] The user dynamic data uploading module is used to receive dynamic data generated by users in different fields or environments;

[0046] The sub-domain centroid module is used to calculate the centroid of each sub-domain based on different data distributions;

[0047] The sub-domain label generation module is used to generate corresponding labels based on the output of the sub-domain centroid calculation module to identify different data points in each sub-domain; at the same time, the sub-domain label generation module optimizes the labels, reduces label conflicts between different domains by adjusting the boundaries and distribution of the labels, and ensures that the labels can accurately reflect the true distribution of the data;

[0048] The self-supervised learning and model optimization module is used to adopt a self-supervised learning method to improve the learning effect of the model by using the intrinsic characteristics of the data itself without relying on manually annotated data; the module adopts cross-validation and triple loss mechanism to ensure that the model can better distinguish different categories of data by calculating the distance relationship between samples, and then train and verify the model under different data divisions to further reduce the overfitting problem of the model; at the same time, the self-supervised learning and model optimization module also feeds back the optimized model information to the sub-domain label generation module to adjust the label generation strategy to form a closed-loop optimization;

[0049] The domain-invariant identity feature learning module is used to extract domain-invariant features from the optimized model.

[0050] The technical solution adopted by the product of the present invention is: a model-independent single-domain generalized cross-domain gait recognition product, including computer program instructions. When the computer program instructions are run on a computer, the computer executes the model-independent single-domain generalized cross-domain gait recognition method.

[0051] Compared with the prior art, the beneficial effects of the present invention include:

[0052] 1. Revealing the underlying data distribution: Through adversarial training, we deeply explore and reveal the underlying distribution characteristics in the data set, and successfully separate identity information and domain information. This process not only improves the recognition ability of the model, but also provides a more accurate basis for subsequent gait analysis.

[0053] 2. Enhanced adaptability and robustness: This method significantly improves the model's adaptability in different environments and conditions, enabling it to accurately recognize unseen gait data.

[0054] 3. Reduce data dependence and cost: Through model-independent design, the dependence on specific training data sets is significantly reduced, and the cost of data annotation and collection is reduced. This feature makes the model easier to promote in resource-constrained environments and is suitable for various practical applications, including smart homes and medical health monitoring.

[0055] 4. Optimize feature extraction and consistency: Optimize the feature extraction process in a cross-domain environment to ensure that the gait features remain consistent, thereby improving the accuracy of recognition. By introducing fine-grained feature learning and domain invariance, the model can extract more reliable gait features and improve overall performance.

[0056] 5. Improve reliability and wide application: This method effectively copes with the differences between data from different sources and improves the reliability of gait recognition in complex situations. It can adapt to various practical application scenarios, such as monitoring, health monitoring, smart assistants, etc. It not only meets security requirements, but also provides support for personalized services, showing a wide range of application potential and market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The technical solution of the present invention is further described below using embodiments and specific implementation methods. In addition, in the process of describing the technical solution, some drawings are also used. For those skilled in the art, other drawings and the intention of the present invention can also be obtained based on these drawings without paying creative work.

[0058] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0060] The present invention is mainly based on the development needs of biometric recognition technology. Considering the problem that the existing gait recognition methods have performance degradation due to domain offset in different scenarios, a single-domain generalization framework called GaitDG is proposed. This framework fully reveals the potential distribution in the data set, uses adversarial training to effectively separate identity information and domain information, and combines data enhancement to diversify the data set distribution to improve the generalization ability of the model. As a model-agnostic method, GaitDG can directly enhance the cross-domain recognition performance of the gait recognition model without changing its structure. Experiments on public gait datasets show that GaitDG significantly improves the cross-domain performance of several latest gait recognition models. Without the need for target domain data, the present invention achieves high generalization and practicality of the gait recognition model, providing strong support for the application of gait recognition technology in real complex environments.

[0061] Please see Figure 1 , this embodiment provides a model-independent single-domain generalized cross-domain gait recognition method, comprising the following steps:

[0062] Step 1: Collect gait data from multiple source domains and extract fine-grained features;

[0063] In one embodiment, gait data is collected from multiple source domains and fine-grained features are extracted. In order to ensure the diversity and comprehensiveness of the data, this embodiment collects data from different angles and under different lighting conditions. Subsequently, the encoder f is used to conduct an in-depth analysis of these data, and fine-grained features that can represent the individual's gait, including key parameters such as gait cycle, stride length, and gait speed, are extracted through advanced signal processing and machine learning techniques. This embodiment is trained through two classifiers: one for identifying individual identities and the other for identifying the domain of the data source. The model is trained through a cross-entropy loss function to accurately classify these features, and the back-propagation algorithm is used to optimize the model parameters. During the training process, the validation set is continuously used to evaluate the classification accuracy and F1 score to ensure the effectiveness of feature extraction.

[0064] In one embodiment, the encoder f is first used to extract gait features, and then the head h is extracted using fine-grained features. fine , convert the gait features obtained from the encoder f into a fine-grained representation;

[0065] The fine-grained feature extraction head h fine is a trained model; during training, we first collect gait data from multiple source domains and extract fine-grained features, and then use the identity classifier and domain classifiers

[0066] Assign fine-grained features to the corresponding identity labels y I and subdomain label yD ; The loss function L used in training fine for:

[0067]

[0068] Among them, f(x), h fine (f(x)) represent the gait feature encoding result and the extracted fine-grained features, L ce is the cross entropy;

[0069] During the training process, back-propagation is used to update the model parameters, and the classification accuracy and F1 score are evaluated through the validation set to ensure the effectiveness of fine-grained feature extraction. Finally, the key gait features extracted by the analysis model are analyzed by visualizing the intermediate layer output to further optimize the model structure.

[0070] Step 2: Extract identity-invariant domain features from fine-grained features;

[0071] In one implementation, a domain classifier and a discriminator are used to extract identity-invariant domain features from fine-grained features;

[0072] The “domain classifier and discriminator” is a trained model; during training, the domain classifier g D and domain discriminator d I , change the domain header h D The output classification is the corresponding sub-domain label y D ; During the training process, d I and h D Participate in adversarial training, I Try h D The output of the classification is the corresponding identity label y I , and h D Try to minimize this classification error, thereby minimizing the difference between identities and obtaining identity-invariant domain features; this adversarial process is achieved through the gradient reversal layer, maximizing the goal of the discriminant module and minimizing the goal of the encoder. The loss function used in training is:

[0073] L domain =L ce (g D (h D (f(x i ))),y D )+λL ce (d I (R λ (h D (f(x i )))),y I );

[0074] Among them, f(xi ) represents the gait feature encoding result, L ce is the cross entropy; λ is a hyperparameter, R λ represents the gradient reversal layer; the hyperparameter λ determines the weight of the gradient during the back-propagation process. In addition, in order to ensure that the identity-invariant domain features are obtained without affecting the learning process of the encoder, in this stage, the parameters of the encoder are frozen and only the domain head h is used. D and domain discriminator d I In this way, training can focus on improving the learning of domain features, thereby enhancing the model's robustness to identity changes and effectively alleviating the problem of domain offset. Through this step, the model is able to extract domain features with high generalization ability, laying the foundation for subsequent cross-domain gait recognition.

[0075] Step 3: Calculate the centroid of each sub-domain based on the extracted identity-invariant domain features;

[0076] In one implementation, the specific implementation of step 3 includes the following sub-steps:

[0077] Step 3.1: Data enhancement and style transfer. In order to improve the generalization ability of the model to different perspectives, this embodiment uses random perspective transformation to transform gait data to different perspectives. This allows the model to be trained under a wider range of perspectives, effectively alleviating the domain shift problem caused by perspective changes.

[0078] Due to privacy policy restrictions, many original RGB images of the data cannot be obtained, which makes it impossible to use the segmentation model of the target domain to process the data and generate a consistent silhouette model style. Therefore, this embodiment migrates the segmentation model based on StarGAN and converts the generated silhouette images into a unified style to enhance the diversity and robustness of the dataset.

[0079] Specifically, this embodiment has a data set containing different segmentation models, and uses these segmentation models S 1 ,S 2 ,…,S N To generate silhouettes of these images. Then this embodiment trains a starGAN to achieve style transfer between the two segmentation models. During training, this embodiment randomly converts the silhouettes to the styles of different segmentation models to enhance the robustness of the model and reduce style optimization. This embodiment denotes these enhanced datasets as x aug , where χ aug The specific representation is as follows:

[0080] χ aug ={A p (x i )|x i ∈χ},

[0081] Among them, A p is an enhancement operation, including random perspective transformation and segmentation model transfer, and the hyperparameter p represents the probability of applying the enhancement.

[0082] Step 3.2: Feature extraction. After obtaining the enhanced dataset, feature extraction is performed to obtain identity-invariant domain features. The specific process includes: (1) Using encoder f: Input the enhanced gait data into encoder f to extract basic feature representation. (2) Generate domain features: Generate domain features through domain header h D Identity-invariant domain features are extracted from the features output by the encoder.

[0083] Step 3.3: Initialize a centroid μ for each sub-domain k k ;

[0084] Step 3.4: Calculate the centroid by weighted average

[0085] Among them, e i represents the domain label of the i-th sample, k represents the index of the subdomain; δ k Represents the indicator function. When sample i belongs to domain k, δ k =1, otherwise 0; f(x i ) represents the feature vector after gait data encoding; g D is the domain classifier, h D For the field head.

[0086] Step 4: Generate sub-domain labels using the calculated centroids;

[0087] In one implementation, the specific implementation of step 4 includes the following sub-steps:

[0088] Step 4.1: For each x i , calculate its centroid μ with all sub-domains k The distance between

[0089] Step 4.2: Get x i The subfield label y P =arg min dist(h D (f(x i )),μ k ), where f(x i ) represents x i The encoded feature vector, dist(), represents each x i and the subdomain centroid μ k The distance between D For the field head.

[0090] In one embodiment, the sub-domain label y P , is the trained sub-domain label y P Computational models;

[0091] The training process includes: (1) Self-supervised learning: Use the generated pseudo labels for self-supervised learning to improve label quality. (2) Cross-validation: Use cross-validation technology to verify the effectiveness of the generated labels to ensure the accuracy and reliability of the labels. (3) Model optimization: Based on the verification results, adjust the model parameters and optimize the model performance to improve the accuracy of gait recognition.

[0092] Step 5: Extract domain-invariant identity features to obtain cross-domain gait recognition results.

[0093] In one embodiment, similar to step 2, an identity classifier and a domain discriminator are used to extract domain-invariant identity features;

[0094] The “identity classifier and domain discriminator” is a trained model; during training, the identity classifier g I and domain discriminator d D , the identity header h I The output of the classification is the corresponding identity label y I ; During the training process, d D and h I Participate in adversarial training, D Try h I The output classification is the corresponding domain label y D , and h I Try to maximize this classification error, thereby minimizing the difference between domains and obtaining domain-invariant identity features. The loss function used in training is:

[0095] L identity =L ce (g I (h I (f(x))),y I )+L trp (x,x n ,x p )+L ce (d D (R λ (h I (f(x)))),y D );

[0096] Among them, L ce is the cross entropy, g I is the identity classifier, h I is the identity head, f(x) represents the feature vector after gait data encoding, y Iis the identity tag; L trp represents the triple loss, x n represents a sample with a different identity from x, x p represents the sample with the same identity as x; d D represents the domain discriminator, λ is a hyperparameter, R λ represents the gradient reversal layer, y D The sub-domain label.

[0097] The triple loss function improves the recognition ability of the model, enabling the model to accurately recognize gait data in different fields. This process not only improves the generalization ability of the model, but also reduces the dependence on specific training data sets, thereby reducing the cost of data annotation and collection, and enhancing the adaptability and robustness of the model.

[0098] After extracting the domain-invariant identity feature, the feature is compared with the existing features in the database. Specifically, for each gait feature in the database, the distance between the data to be identified and it is calculated. If the distance between the two is within a suitable range, the identification is successful and the gait recognition result is returned.

[0099] This embodiment also provides a model-independent single-domain generalized cross-domain gait recognition device, including: a user dynamic data uploading module, a sub-domain centroid calculation module, a sub-domain label generation module, a self-supervised learning and model optimization module, and a domain-invariant identity feature learning module. Each module works closely together to process the dynamic data uploaded by the user, and ultimately realizes identity feature extraction and model optimization in a cross-domain environment.

[0100] The user dynamic data upload module is mainly responsible for receiving dynamic data generated by users in different fields or environments. The uploaded data will serve as the basis for subsequent sub-field analysis and label generation, ensuring that the system can accurately process data in different fields.

[0101] The sub-domain centroid module is mainly responsible for calculating the centroid of each sub-domain based on different data distributions. The centroid represents the central position of the data in each sub-domain. The centroid calculation can not only reflect the differences in data in each domain, but also provide key support for further label generation and optimization.

[0102] The sub-domain label generation module is mainly responsible for generating corresponding labels based on the output of the sub-domain centroid calculation module to identify different data points in each sub-domain. In order to ensure the stability and accuracy of the generated labels under different data distributions, the sub-domain label generation module optimizes the labels by adjusting the boundaries and distribution of the labels to reduce label conflicts between different domains, ensuring that the labels can accurately reflect the real distribution of the data, thereby improving the generalization ability of the system.

[0103] The self-supervised learning and model optimization module adopts a self-supervised learning method, which does not rely on manually labeled data and uses the intrinsic characteristics of the data itself to improve the learning effect of the model. This module adopts a cross-validation and triple loss mechanism to ensure that the model can better distinguish different categories of data by calculating the distance relationship between samples, and then train and verify the model under different data partitions to further reduce the overfitting problem of the model. At the same time, the self-supervised learning and model optimization module will also feed back the optimized model information to the sub-domain label generation module to adjust the label generation strategy and form a closed-loop optimization. In this closed-loop optimization process, label generation and model optimization interact with each other to ensure that the system can dynamically adjust data in different fields and improve the overall recognition accuracy.

[0104] The domain-invariant identity feature learning module is mainly responsible for extracting domain-invariant features from the optimized model. Domain-invariant features refer to features that can maintain consistency in different domains or distributions, ensuring that the system can accurately identify the user's identity features in a cross-domain data environment. By learning domain-invariant features, the system can still accurately identify identity features when facing data from different domains or environments, with strong robustness and consistency.

[0105] Through the collaborative work of the above modules, the present invention can effectively realize cross-domain identity feature learning and model optimization, ensuring accurate recognition and classification in a multi-domain environment. The system forms a closed-loop optimization process through sub-domain centroid calculation, label generation, self-supervised learning and feature extraction, extracting domain-invariant identity features from data in different fields, greatly improving the system's generalization ability and recognition accuracy.

[0106] This embodiment also provides a model-independent single-domain generalized cross-domain gait recognition product, including computer program instructions. When the computer program instructions are executed on a computer, the computer executes the model-independent single-domain generalized cross-domain gait recognition method.

[0107] The present invention is further described below through specific experiments.

[0108] The proposed method is evaluated on three gait recognition datasets, including two widely adopted datasets CASIA-B and OUMVLP and a large-scale real-world dataset grow.

[0109] CASIA-B includes 10 gait sequences: 6 normal gaits (NM), 2 backpack gaits (BG), and 2 coat gaits (CL). Each group contains 11 viewing angles (from 0° to 180°, with an interval of 18°). Each subject has 3 dressing changes. In this experiment, 50 subjects were selected.

[0110] The OUMVLP dataset contains 14 viewing angles (0°-90° and 180°-270°, one viewing angle every 15°), and each viewing angle has 2 sets of normal gaits (NM). 5153 subjects were used in the experiment.

[0111] The GREW dataset is a large-scale, real-world gait dataset containing 102,887 gait sequences with rich gait transformations and diverse real-world conditions. In this experiment, 20,000 subjects were selected from the GREW dataset for training and evaluating the performance of the model in a cross-domain gait recognition task.

[0112] In the experimental setting, the source domain uses the GREW dataset for training, and the target domains are the CASIA-B and OUMVLP datasets for testing. All gait images are resized to 64×44 pixels, and each gait sequence contains 30 frames. For comparative experiments, we selected three benchmark models: GaitSet, GaitPart, and GaitGL, and improved GaitDG (the method of the present invention) based on these benchmark models, and compared the improved models with PSIM (another domain adaptation framework). In terms of training and optimization settings, all models use the Adam optimizer with a learning rate set to 0.0001. The training process of GaitDG includes 60,000 iterations of the baseline model, followed by 400 iterations of GaitDG, with 600 parameter updates per iteration. In the experiment, λ=1.0, p=0.2, and K=8 were selected, where K is a hyperparameter that controls the number of subdomains. In terms of data enhancement, two methods, random perspective transformation (RPT) and segmentation model transfer (SMT), were adopted. Four segmentation models, JPPNet, SAM, YOLOv8 and DeeplabV3+, were used to segment the CASIA-B images and generate four training data sets for training.

[0113] The experiment was conducted on an Ubuntu 22.04 system equipped with an NVIDIA GeForce RTX 4090 GPU and an Intel i9-12900K processor. Three benchmark models, GaitSet, GaitPart, and GaitGL, were used to evaluate the effect of the GaitDG method in improving the generalization ability of the model.

[0114] On the CASIA-B dataset, the GaitDG method significantly improves the performance of three baseline models, namely GaitSet (71.2%), GaitPart (70.2%), and GaitGL (73.4%), which are 4.9%, 5.1%, and 2.9% higher than the original models, respectively. Under BG and CL conditions, GaitDG surpasses other baseline models in most viewpoints, especially under the GaitGL model, where it performs well.

[0115] On the OU-MVLP dataset, GaitDG also showed significant improvements compared to the original model and PSIM. Specifically, GaitSet improved by 1.7%, GaitPart improved by 0.8%, and GaitGL improved by 1.7%. Although its performance on this dataset is slightly inferior to CASIA-B, GaitDG still shows strong cross-domain recognition capabilities.

[0116] In addition, this embodiment also conducts an ablation experiment to further explore the impact of each component on the performance of GaitDG.

[0117] First, fine-grained feature learning and data augmentation play a crucial role in improving the model. Experimental results show that after removing fine-grained feature learning, the average rank-1 accuracy drops by 1.4%, while after removing random viewpoint transformation, the accuracy drops by 0.6%, and after removing segmentation model transfer, the accuracy drops by 1.2%. These results show that FGFL and data augmentation play an important role in helping the model extract finer features and enhance the generalization ability of the model, significantly improving the performance of GaitDG in gait recognition.

[0118] Secondly, the setting of the number of sub-domains K also has a significant impact on model performance. Experiments have found that when K = 2 or K = 10, the performance of the model is poor, indicating that too few or too many sub-domains will affect the performance of the model in cross-domain migration. Too few sub-domains may not be able to fully capture the intrinsic distribution of the data, while too many sub-domains may introduce redundant features and increase the difficulty of model learning. Therefore, a reasonable setting of the K value is crucial for extracting domain-invariant features and improving cross-domain recognition capabilities.

[0119] Finally, the impact of the scale of unseen data also has a significant impact on model performance. Specifically, when the scale of the OU-MVLP dataset is small (for example, 1% of the dataset scale), GaitDG's performance improvement is most obvious, with the rank-1 accuracy improved by 4.2%, while when the dataset scale reaches 100%, the improvement is only 1.7%. This result shows that GaitDG can better adapt to the diversity of data and improve generalization ability when processing small-scale datasets, especially showing stronger adaptability and recognition ability when facing unseen data.

[0120] The present invention significantly improves the accuracy and robustness of gait recognition through experimental verification on multiple gait data sets. Compared with the prior art, the present invention can effectively recognize more unseen gait data under the same conditions and can adapt to changes in different environments and conditions. At the same time, the present invention reduces the cost of data annotation and collection by reducing the dependence on specific training data sets, which facilitates the widespread promotion of gait recognition technology.

[0121] It should be understood that the embodiments described above are part of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be combined with each other arbitrarily to form a feasible technical solution. Such combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0122] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A model-independent single-domain generalized cross-domain gait recognition method, characterized in that: The following steps are involved: Step 1: Collect gait data from multiple source domains and extract fine-grained features; Step 2: Extract identity-invariant domain features from fine-grained features; Step 3: Calculate the centroid of each sub-domain based on the extracted identity-invariant domain features; Step 4: Generate sub-domain labels using the calculated centroids; Step 5: Extract domain-invariant identity features to obtain cross-domain gait recognition results.

2. The model-independent single-domain generalized cross-domain gait recognition method according to claim 1, characterized in that: In step 1, the encoder f is first used to extract gait features, and then the head h is extracted using fine-grained features fine , convert the gait features obtained from the encoder f into a fine-grained representation; The fine-grained feature extraction head h fine is a trained model; during training, we first collect gait data from multiple source domains and extract fine-grained features, and then use the identity classifier and domain classifiers Assign fine-grained features to the corresponding identity labels y I and subdomain label y D ; The loss function L used in training fine for: Among them, f(x) represents the feature vector after gait data encoding, h fine represents the fine-grained feature extraction head, L ce is the cross entropy; During the training process, back-propagation is used to update the model parameters, and the classification accuracy and F1 score are evaluated through the validation set to ensure the effectiveness of fine-grained feature extraction. Finally, the key gait features extracted by the analysis model are analyzed by visualizing the intermediate layer output to further optimize the model structure.

3. The model-independent single-domain generalized cross-domain gait recognition method according to claim 1, characterized in that: In step 2, a domain classifier and an identity discriminator are used to extract identity-invariant domain features from fine-grained features; The “domain classifier and identity discriminator” is a trained model; during training, the domain classifier g D and identity discriminator d I , change the domain header h D The output classification is the corresponding sub-domain label y D ; During the training process, d I and h D Participate in confrontation training, I Try h D The output of the classification is the corresponding identity label y I , and h D Try to maximize this classification error, thereby minimizing the differences between identities and obtaining identity-invariant domain features; The loss function used in training is: L domain =L ce (g D (h D (f(x i ))),y D )+λL ce (d I (R λ (h D (f(x i )))),y I ); Among them, L ce is the cross entropy, g D is the domain classifier, h D is the domain head, f(x i ) is the gait data x i The encoded feature vector, y D is the subdomain label; λ is a hyperparameter, d I is the identity discriminator, R λ represents the gradient reversal layer, y I is the identity label; during training, the parameters of the encoder f are frozen and only the domain head h D and domain discriminator d I Make optimizations.

4. The model-independent single-domain generalized cross-domain gait recognition method according to claim 1, characterized in that: The specific implementation of step 3 includes the following sub-steps: Step 3.1: Data enhancement and style transfer to obtain enhanced data sets; First, we use different segmentation models S1, S2, ..., S N To generate gait data x i The silhouette of the image is then migrated based on the segmentation model of StarGAN to convert the generated silhouette image into a unified style; Step 3.2: After obtaining the enhanced dataset, perform feature extraction to obtain identity-invariant domain features; Step 3.3: Initialize a centroid μ for each sub-domain k k ; Step 3.4: Calculate the centroid by weighted average Among them, e i represents the domain label of the i-th sample, k represents the index of the subdomain; δ k Represents the indicator function. When sample i belongs to domain k, δ k =1, otherwise 0; f(x i ) represents the gait data x i Encoded feature vector; g D is the domain classifier, h D For the field head.

5. The model-independent single-domain generalized cross-domain gait recognition method according to claim 1, characterized in that: The specific implementation of step 4 includes the following sub-steps: Step 4.1: For each gait data x i , calculate its centroid μ with all sub-domains k The distance between Step 4.2: Get x i The subfield label y P =argmindist(h D (f(x i )),μ k );in, f(x i ) represents x i The encoded feature vector, dist(), represents each x i and the subdomain centroid μ k The distance between D For the field head.

6. The model-independent single-domain generalized cross-domain gait recognition method according to claim 1, characterized in that: In step 5, the identity classifier and domain discriminator are used to extract domain-invariant identity features; The “identity classifier and domain discriminator” is a trained model; during training, the identity classifier g I and domain discriminator d D , the identity header h I The output of the classification is the corresponding identity label y I ; During the training process, d D and h I Participate in confrontation training, D Try j I The output classification is the corresponding domain label y D , and h I Try to maximize this classification error, thereby minimizing the difference between domains and obtaining domain-invariant identity features. The loss function used in training is: L ident i ty = L ce (g I (h I (f(x))),y I )+L trp (x,x n ,x p )+L ce (d D (R λ (h I (f(x)))),y D ); Among them, L ce is the cross entropy, g I is the identity classifier, h I is the identity head, f(x) represents the feature vector after gait data encoding, y I is the identity tag; L frp represents the triplet loss, x n represents a sample with a different identity from x, x p represents the sample with the same identity as x; d D represents the domain discriminator, λ is a hyperparameter, R λ represents the gradient reversal layer, y D The sub-domain label.

7. The model-independent single-domain generalized cross-domain gait recognition method according to claim 1, characterized in that: In step 5, after extracting the domain-invariant identity feature, the feature is compared with the existing features in the database; for each gait feature in the database, the distance between the data to be identified and it is calculated. If the distance between the two is within a suitable range, the identification is successful and the gait recognition result is returned.

8. The model-independent single-domain generalized cross-domain gait recognition method according to any one of claims 1 to 7, characterized in that: In step 4, the sub-domain label y P , is the trained sub-domain label y P Computational models; The training process includes: (1) self-supervised learning: using the generated pseudo labels for self-supervised learning to improve label quality; (2) cross-validation: verifying the effectiveness of the generated labels through cross-validation technology to ensure the accuracy and reliability of the labels; (3) based on the verification results, P The calculation model parameters are adjusted to optimize the model performance and improve the accuracy of gait recognition.

9. A model-independent single-domain generalized cross-domain gait recognition device, characterized by: It includes user dynamic data uploading module, sub-domain centroid calculation module, sub-domain label generation module, self-supervised learning and model optimization module, and domain-invariant identity feature learning module; The user dynamic data uploading module is used to receive dynamic data generated by users in different fields or environments; The sub-domain centroid module is used to calculate the centroid of each sub-domain based on different data distributions; The sub-domain label generation module is used to generate corresponding labels based on the output of the sub-domain centroid calculation module to identify different data points in each sub-domain; at the same time, the sub-domain label generation module optimizes the labels, reduces label conflicts between different domains by adjusting the boundaries and distribution of the labels, and ensures that the labels can accurately reflect the true distribution of the data; The self-supervised learning and model optimization module is used to adopt a self-supervised learning method to improve the learning effect of the model by using the intrinsic characteristics of the data itself without relying on manually annotated data; the module adopts cross-validation and triple loss mechanism to ensure that the model can better distinguish different categories of data by calculating the distance relationship between samples, and then train and verify the model under different data divisions to further reduce the overfitting problem of the model; at the same time, the self-supervised learning and model optimization module also feeds back the optimized model information to the sub-domain label generation module to adjust the label generation strategy to form a closed-loop optimization; The domain-invariant identity feature learning module is used to extract domain-invariant features from the optimized model.

10. A model-independent single-domain generalized cross-domain gait recognition product, comprising computer program instructions, characterized in that: When the computer program instructions are executed on a computer, the computer is enabled to execute the model-independent single-domain generalized cross-domain gait recognition method according to any one of claims 1 to 8.

Images