Cross-domain identity recognition method and system based on gait condition generative adversarial network
Through the generation of adversarial network models of gait conditions, learning domain mapping relationships and introducing Gaussian noises, the interference domain problem of gait recognition in WiFi signals is solved, improving the recognition accuracy, and ensuring accurate identification of different individuals.
Patent Information
- Application Number
- CN202411425202.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-10-12
AI Technical Summary
The prior art is difficult to effectively solve the interference domain problems introduced by walking path direction and equipment layout during gait recognition in WiFi signals, resulting in a decrease in recognition accuracy.
Adopt the adversarial network (CG-GAN) model based on gait conditions, learn the mapping relationship between different domains through feature extractor, generator, discriminator and classification discriminator, introduce Gaussian noise perturbation, and use KL divergence loss and loop Mechanisms such as consistency loss and identity mapping loss extract domain-independent information and retain gait characteristics to achieve domain adaptation and interference removal.
It improves the accuracy of gait recognition in WiFi signals, reduces the interference of walking path direction and device layout on the recognition, ensures accurate identification of different individuals, and alleviates the negative impact of multiple interference domains.
Smart Images

Figure CN118964957A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of identity recognition technology, and in particular to a cross-domain identity recognition method and system based on a gait conditional generative adversarial network. Background Art
[0002] WiFi devices (such as routers, mobile phones, etc.) transmit radio frequency signals in the form of WiFi spectrum during communication. These signals are reflected and diffracted by the human body in the area, and can be used to identify activities such as hands, body or breathing. In real life scenarios, these identification methods can be used in a variety of fields such as security monitoring systems, elderly motor function status monitoring, smart home devices, etc. Since each individual's natural walking posture (hereinafter referred to as "gait") theoretically has the specificity of being uniquely bound to biometrics, some existing technologies attempt to explore the possibility of using CSI (Channel State Information) corresponding to human gait collected by WiFi devices for HI (Human Identification). Compared with traditional sensors (such as cameras, microphones, etc.), non-invasive human identification systems based on WiFi signals involve fewer user privacy issues and are well adapted to personalized deployment applications in various sampling locations such as smart homes and offices. Therefore, their application in human identification work has important research value. In real environments, different walking trajectory directions and walking speeds will cause abnormal fluctuations in signal characteristics and introduce domain interference to the CSI signals collected by WiFi sensing, thereby negatively affecting the effectiveness of WiFi sensing. Existing technologies attempt to use transfer learning and data enhancement techniques to remove the influence of interference domains, but they still cannot effectively solve the problem of removing interference in the direction of complex human walking paths. Summary of the invention
[0003] The main technical problem solved by the present invention is to provide a cross-domain identity recognition method and system based on a gait conditional generative adversarial network, wherein the cross-domain identity recognition method can reduce the interference of gait determined by WiFi signals on the walking path direction, and improve the recognition accuracy of different individuals.
[0004] In order to solve the above problems, the present invention provides a cross-domain identity recognition method based on gait conditional generative adversarial network, which includes inputting WiFi CSI human gait data to be classified into a CG-GAN model, learning feature extraction with domain adaptation capability by the CG-GAN model, autonomously adapting complex source domain information of different receiver layout positions of WiFi signals to the target domain, and effectively extracting domain-independent information between interference domains, wherein:
[0005] The CG-GAN model includes a pair of feature extractors, a pair of generators, a pair of discriminators, and a classification discriminator that retains human gait information to learn the mapping relationship between different domains. The CG-GAN model uses the feature extractor to convert data samples in the source domain and data samples in the target domain into an embedding layer, and introduces Gaussian noise perturbation in the embedding layer, so that the feature extractor can cope with interference from multiple subdomains and obtain preliminary domain generalization ability.
[0006] In one embodiment, the CG-GAN model uses KL divergence loss to align the embedding layer distribution to a Gaussian distribution with zero mean and unit variance. In the above formula, KL divergence loss can be formally described as follows: , Among them, in the above formula Represents two different distributions and The KL divergence between encourages that the embedding representation of the latent space should be domain-independent. p and q refer to the probability distribution of z obtained by calculating the KL divergence.
[0007] In one embodiment, after obtaining the embedding layer perturbed by Gaussian noise, the CG-GAN model uses a generator to synthesize the embedding layer representation into a simulated data sample, uses a discriminator to determine whether the data is a source domain or a target domain, distinguishes the real data input, and guides the generator to generate realistic samples through the synthetic samples simulated by the generator, that is, to obtain the reconstructed source domain channel state information.
[0008] In one embodiment, for the mapping process from the source domain to the target domain, or the mapping process from the target domain to the source domain, the target adversarial loss involved is formally described as follows: .
[0009] In one embodiment, the CG-GAN model uses cycle consistency loss CCL to preserve the semantic information of the original CSI during the domain adaptation process, and remaps the synthetic data in the target domain back to the original source domain. The cycle consistency loss formed by the cycle transformation is formally described as follows: .
[0010] In one embodiment, the CG-GAN model also applies the identity mapping loss so that when the sample is transferred to the generator corresponding to its original domain, it can generate samples that maintain domain invariance. The identity mapping loss can be formally described as follows: .
[0011] In one embodiment, the CG-GAN model uses a classification discriminator to maintain the uniqueness of the gait features reflected by the human body in different states during the domain adaptation process, and quantifies the difference between the predicted label of the gait sample transformed by the generative adversarial network and its true label. The gait classification loss can be formally described as follows: , in, represents the set of gait categories, Represents the real source domain data label category, represents the softmax operation, Using cross entropy loss for supervised classification, the CG-GAN model uses the label information corresponding to the source domain data with known labels for supervised classification learning.
[0012] In one embodiment, the training loss function of the CG-GAN model is formally expressed as follows: , in, , , and They represent the weights of the identity mapping loss, cycle consistency loss, gait classification loss, and KL divergence loss respectively.
[0013] The present invention also provides a cross-domain identity recognition system based on a gait conditional generative adversarial network, comprising an IoT device with a WiFi function, the IoT device having a domain filter model trained by a server, the domain filter model comprising a CG-GAN model, the CG-GAN model having input WiFi CSI human gait data to be classified, learning feature extraction with domain adaptation capability, autonomously adapting complex source domain information of different receiver layout positions of WiFi signals to a target domain, and effectively extracting domain-independent information between interference domains, wherein,
[0014] The CG-GAN model includes a pair of feature extractors, a pair of generators, a pair of discriminators and a classification discriminator that retains human gait information to learn the mapping relationship between different domains. The CG-GAN model uses the feature extractor to convert data samples in the source domain and data samples in the target domain into an embedding layer, and introduces Gaussian noise perturbation in the embedding layer, so that the feature extractor can cope with interference from multiple subdomains and obtain preliminary domain generalization ability.
[0015] In one embodiment, the CG-GAN model uses KL divergence loss to align the embedding layer distribution to a Gaussian distribution with zero mean and unit variance. In the above formula, KL divergence loss is formally described as follows: ,
[0016] Among them, in the above formula Represents two different distributions and The KL divergence between encourages that the embedding representation of the latent space should be domain-independent. p and q refer to the probability distribution of z obtained by calculating the KL divergence.
[0017] In one embodiment, after obtaining the embedding layer perturbed by Gaussian noise, the CG-GAN model uses a generator to synthesize the embedding layer representation into a simulated data sample, uses a discriminator to determine whether the data is a source domain or a target domain, distinguishes the real data input, and guides the generator to generate realistic samples through the synthetic samples simulated by the generator, that is, to obtain the reconstructed source domain channel state information.
[0018] In one embodiment, for the mapping process from the source domain to the target domain, or the mapping process from the target domain to the source domain, the target adversarial loss involved is formally described as follows: .
[0019] In one embodiment, the CG-GAN model uses cycle consistency loss to preserve the semantic information of the original CSI during the domain adaptation process and remap the synthesized data in the target domain back to the original source domain. The cycle consistency loss formed by the cycle transformation is formally described as follows: .
[0020] In one embodiment, the CG-GAN model also applies the identity mapping loss so that when the sample is transferred to the generator corresponding to its original domain, it can generate samples that maintain domain invariance. The identity mapping loss can be formally described as follows: .
[0021] In one embodiment, the CG-GAN model uses a classification discriminator to maintain the uniqueness of the gait features reflected by the human body in different states during the domain adaptation process, and quantifies the difference between the predicted label of the gait sample transformed by the generative adversarial network and its true label. The gait classification loss can be formally described as follows: , in, represents the set of gait categories, Represents the real source domain data label category, represents the softmax operation, Using cross entropy loss for supervised classification, the CG-GAN model uses the label information corresponding to the source domain data with known labels for supervised classification learning.
[0022] In one embodiment, the training loss function of the CG-GAN model is formally expressed as follows: , in, , , and They represent the weights of the identity mapping loss, cycle consistency loss, gait classification loss, and KL divergence loss respectively.
[0023] The present invention is based on a cross-domain identity recognition method and system of a gait conditional generative adversarial network. By utilizing the gait conditional generative adversarial network model to learn the mapping relationship between multiple interference domains, the system improves the ability to adapt to multiple types of interference domains, introduces Gaussian noise into the middle hidden layer of the conditional generative adversarial network, and learns feature extraction with domain adaptability. The domain adaptive feature extractor acts as a domain filter to remove noise from complex interference domains mixed with multiple WiFi device layouts, while retaining the unique characteristics of different individuals, ensuring that the WiFi sensing system can relatively accurately identify the CSI gait characteristics of the individual. It can eliminate the negative impacts of WiFi device layout, walking direction, and speed, and accurately identify different individuals. It can alleviate the negative impacts of multiple interference domains. The CG-GAN feature extractor trained on the server It is used as a domain filter model and deployed as the backbone network of the inference model on devices with WiFi functions to extract the unique gait features of the human body in the WiFi CSI signal to realize human gait recognition. It can reduce the interference of determining the gait through WiFi signals on the walking path direction and improve the recognition accuracy of different individuals. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the description only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0025] Figure 1 It is a schematic diagram of a conditional gait generative adversarial network in a cross-domain identity recognition method based on a gait conditional generative adversarial network of the present invention.
[0026] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0027] The claims of the present invention are further described in detail below in conjunction with specific embodiments and drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments proposed by ordinary technicians in this field before making creative work also fall within the scope of protection of the present invention.
[0028] It should be understood that, in the description of the embodiments of the present invention, all directional indication terms, such as "up", "down", "left", "right", "front", "back", etc., indicate the orientation or position relationship based on the orientation, position relationship shown in the drawings or the orientation or position relationship in which the invention product is usually placed when used, which is only for the convenience of simplifying the description of the present invention, and does not expressly or implicitly indicate that the device, element or component referred to must have a specific orientation and a specific orientation structure, and should not be understood as a limitation on the present invention. It is only used to explain the relative position relationship, movement, etc. between the components shown in the drawings. When the specific posture changes, the directional indication may also change accordingly. For processes and methods, unless it is clearly stated in the specification that each step must be performed in sequence, excluding the case where the order of each step is changed, otherwise the corresponding invention purpose cannot be achieved, it is considered that for the sake of simplicity of description, the specification only selects the best or better one from various implementation methods as an example for explanation.
[0029] In addition, ordinal numbers such as "first" and "second" in the present invention are only used for the purpose of distinction, and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. The features defined as "first" and "second" can expressly or implicitly indicate at least one of the technical features. In order to facilitate the description, the present invention may have quantifiers such as "one", "two", "multiple", "at least one" or "several" to define the technical (structural) features, where "multiple" means at least two, that is, two or more; "at least one" and "several" mean one or one and the corresponding technical features; quantifiers and other definitions are only for illustrating the specific embodiment or the best implementation method, and should be included as long as they are not contrary to the overall concept and purpose of the invention, unless it is clearly stated that the quantifiers must be used to define them in order to be implemented or to achieve the purpose of the invention.
[0030] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "screw-on" should be understood in a broad sense. For example, the positional relationship between components can be relatively fixed, or the components can be physically fixedly connected; they can be detachably connected or integrated; they can be mechanically connected or electrically connected; they can be directly connected or indirectly connected through an intermediate medium or component; they can be internally connected between two elements or interact with each other. Unless otherwise clearly defined in the specification, other interpretations may not achieve the corresponding functions or effects. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0031] The controller and control circuit involved in the present invention are conventional control technologies or units of those skilled in the art. For example, the control circuit of the controller can be realized by ordinary technicians in the field using existing technologies, such as simple programming. If the software or program involved in the control result is not described in detail, it belongs to the use of existing technologies or conventional technologies of ordinary technicians in the field. The power supply also uses the existing technologies in the field, and the main technical point of the present invention is to improve the mechanical device, so the present invention will not describe the specific circuit control relationship and circuit connection in detail.
[0032] The disclosure of the present invention provides many different embodiments or examples to realize different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described in the present invention. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplicity and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed. In addition, the present invention provides various specific examples of processes and materials, but those of ordinary skill in the art can be aware of the application of other processes and / or the use of other materials.
[0033] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0034] like Figure 1 As shown, an embodiment of a cross-domain identity recognition method based on a gait-conditional generative adversarial network is provided in the present invention.
[0035] The cross-domain identity recognition method based on gait conditional generative adversarial network includes inputting WiFiCSI human gait data to be classified into the CG-GAN (Conditional Gait Generative Adversarial Network) model, and the CG-GAN model learns feature extraction with domain adaptation ability, autonomously adapts the complex source domain information of different receiver layout positions of WiFi signals to the target domain, and effectively extracts domain-independent information between interference domains. The CG-GAN model includes a pair of feature extractors, a pair of generators, a pair of discriminators and a classification discriminator that retains human gait information. The mechanism of the CG-GAN model is used to learn the mapping relationship between different domains. The feature extractor represents the extraction of domain-independent features in CSI, and the discriminator used to generate data is the source domain or the target domain. The CG-GAN model uses the feature extractor to extract the source domain The data samples in the target domain and the data samples in the target domain are transformed into the embedding layer , Gaussian noise perturbation is introduced in the embedding layer.
[0036] Specifically, the complex source domain information of different receiver layout positions of WiFi signals is autonomously adapted to the target domain, that is, the source domain contains multiple device layout positions or walking paths, and the target domain is a specific position or path. The correspondence between the source domain and the target domain is learned, and the model is migrated and converted to a specific target domain. The domain-independent information between the interference domains is effectively extracted. The domain-independent information is common information. The CG-GAN model can be used as a domain filter to migrate the source domain to the target domain while retaining the human gait features contained in the WiFi CSI. The gait features are unique to the human body and have no correlation with the interference domain. Therefore, the domain-independent information refers to the human gait information. The model can migrate and adapt to the target domain without losing the human gait information. Among them, the CG-GAN model includes a pair of feature extractors, namely, feature extractors and feature extractor , a pair of generators, namely the generator and generator , a pair of discriminators, namely, the discriminator and the discriminator , and a classifier that retains human gait information , using the mechanism of the CG-GAN model to learn the mapping relationship between different domains, the feature extractor , Feature Extractor Indicates that domain-independent features in CSI are extracted, and the data used for generation is the source domain The discriminator or the discriminator of the target domain 𝑌 , the CG-GAN model uses a feature extractor and feature extractor Source Domain The data samples 𝑥 in the target domain 𝑌 and the data samples 𝑦 in the target domain 𝑌 are transformed into the embedding layer and embedding layer , in the embedding layer and embedding layer Gaussian noise perturbations are introduced respectively. That is, the real source domain information state information is extracted by the feature extractor Source Domain The data sample 𝑥 in is transformed into the embedding layer , in the embedding layer Gaussian noise perturbation is introduced and passed through the generator The simulated synthetic samples, such as the target domain channel state information. The synthesized target domain channel state information is passed through the feature extractor Transform the data sample 𝑦 in the original target domain 𝑌 into the embedding layer , in the embedding layer Gaussian noise perturbation is introduced, and then the generator Guided to produce realistic samples, that is, to generate reconstructed source domain channel state information.
[0037] The CG-GAN model inputs the WiFi CSI human gait data to be classified, and can first divide the source domain data and the target domain data. The source domain refers to the data pre-recorded by the equipment manufacturer, including multiple interference domains, and the target domain refers to the data collected by the user, including a specific interference domain. This patent can migrate and adapt to the specific target domain to extract human gait information.
[0038] CG-GAN model uses feature extractor and feature extractor Source domain The data samples 𝑥 in the target domain 𝑌 and the data samples 𝑦 in the target domain 𝑌 are transformed into embedded representations and Considering the existence of various heterogeneous interference domains, such as walking paths, directions, device locations, etc., this embodiment adopts and embedding layer Gaussian noise perturbations are introduced to help the corresponding feature extractors cope with interference from multiple subdomains and obtain preliminary domain generalization capabilities, such as the embedding layer When Gaussian noise is introduced, the feature extractor To cope with interference from multiple subdomains, embedding layer When Gaussian noise is introduced, the feature extractor To cope with interference from multiple subdomains. Subsequently, the CG-GAN model uses KL divergence loss to align the embedding layers separately and embedding layer to a Gaussian distribution with zero mean and unit variance The KL divergence loss can be formally expressed as formula (1).
[0039] (1), In the above formula Represents two different distributions and The KL divergence between them encourages that the embedding representation of the latent space should be domain-independent. p and q refer to the calculation method of KL divergence, that is, the probability distribution of z.
[0040] Obtained the embedded representation after Gaussian noise perturbation and After that, the CG-GAN model uses the generator and generator Synthesize the embedding representation into simulated data samples while leveraging the discriminator and the discriminator Distinguish between real data input and synthetic samples simulated by the generator, thereby guiding the generator to produce realistic samples. The mapping process from 𝑌 to the target domain 𝑌, the target adversarial loss can be formally described as shown in formula (2): (2).
[0041] Similarly, the target domain 𝑌 to the source domain The target adversarial loss involved in the mapping process can also be formally described as shown in formula (3): (3).
[0042] In order to enhance the model's ability to retain the semantic information of the original CSI to a certain extent during the domain adaptation process, the CG-GAN model adopts cycle consistency loss (CCL) to remap the synthetic data in the target domain back to the original source domain. The loss expression of this cycle transformation is as follows: (4).
[0043] At the same time, the CG-GAN model also applies the identity mapping loss (IML) so that when the sample is sent to the generator corresponding to its original domain, it can generate samples that maintain domain invariance. This loss can be formally described as formula (5): (5).
[0044] In addition, the CG-GAN model uses a classification discriminator , in order to maintain the uniqueness of the gait features reflected by the human body in different states during the domain adaptation process. It quantifies the difference between the predicted label of the gait sample after the generative adversarial network transformation and its true label. The gait classification loss can be formally described as shown in formula (6).
[0045] (6), in, represents the set of gait categories, Represents the real source domain data label category, represents the softmax operation, Cross entropy loss is used for supervised classification. The CG-GAN model uses label information corresponding to source domain data with known labels to perform supervised classification learning.
[0046] Considering the above-mentioned joint constraints, the training loss function of the CG-GAN model can be formally expressed as formula (7): (7), in, , , and They represent the weights of the identity mapping loss, cycle consistency loss, gait classification loss, and KL divergence loss respectively.
[0047] The CG-GAN model has domain adaptation capabilities and can alleviate the negative impact of multiple interference domains. It is used as a domain filter model, as the backbone network of the inference model, and deployed on devices with WiFi functions to extract the unique gait features of the human body in the WiFi CSI signal to realize human gait recognition.
[0048] Because the gait-conditional generative adversarial network model is used to learn the mapping relationship between multiple interference domains, adapt to various types of interference domains, introduce Gaussian noise in the hidden layer in the middle of the CG-GAN model, and learn feature extraction with domain adaptation capabilities, the domain adaptive feature extractor acts as a domain filter to remove the noise of the complex interference domain mixed with multiple WiFi device layouts, while retaining the unique characteristics of different individuals, ensuring that the WiFi perception system can relatively accurately identify the CSI gait characteristics of the person. It can eliminate the negative impact of WiFi device layout, walking direction, and speed and accurately identify different individuals. It can alleviate the negative impact of multiple interference domains. The CG-GAN model feature extractor trained on the server It is used as a domain filter model and deployed as the backbone network of the inference model on devices with WiFi functions, and human gait recognition is realized based on the unique gait features of the human body in the WiFi CSI signal. It can reduce the interference of determining the gait through WiFi signals on the walking path direction and improve the recognition accuracy of different individuals.
[0049] The present invention also provides a cross-domain identity recognition system based on gait conditional generative adversarial network, the cross-domain identity recognition system includes an IoT (Internet of Things) device with WiFi function, the IoT device has a domain filter model trained by a server, and the domain filter model includes a CG-GAN model. The CG-GAN model adopts the above-mentioned implementation structure or function, that is, the CG-GAN model includes a pair of feature extractors, namely, feature extractors and feature extractor , a pair of generators, namely the generator and generator , a pair of discriminators, namely, the discriminator and the discriminator , and a classification discriminator , to learn the mapping relationship between different domains, the feature extractor , Indicates the extraction of domain-independent features in CSI, the generator and generator , discriminator and the discriminator , the CG-GAN model uses a feature extractor and feature extractor Source Domain The data samples 𝑥 in the target domain 𝑌 and the data samples 𝑦 in the target domain 𝑌 are transformed into the embedding layer and embedding layer , in the embedding layer and embedding layer A Gaussian noise perturbation is introduced.
[0050] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some technical features in one or more embodiments may be replaced or combined by equivalent means, and these modifications or replacements or combinations do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, unless the embodiments explicitly exclude the possibility of replacing or combining technical features in different embodiments, or this replacement or combination is contrary to the overall concept of the invention.
Claims
1. A cross-domain identity recognition method based on gait-conditional generative adversarial network, comprising: The WiFi CSI human gait data to be classified is input into the CG-GAN model, and the CG-GAN model learns feature extraction with domain adaptation ability, autonomously adapts the complex source domain information of different receiver layout positions of WiFi signals to the target domain, and effectively extracts domain-independent information between interference domains. The CG-GAN model includes a pair of feature extractors, a pair of generators, a pair of discriminators, and a classification discriminator that retains human gait information to learn the mapping relationship between different domains. The CG-GAN model uses the feature extractor to convert data samples in the source domain and data samples in the target domain into an embedding layer, and introduces Gaussian noise perturbation in the embedding layer, so that the feature extractor can cope with interference from multiple subdomains and obtain preliminary domain generalization ability.
2. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 1 is characterized in that: The CG-GAN model uses KL divergence loss to align the distribution of the embedding layer to a Gaussian distribution with zero mean and unit variance. In the above formula, KL divergence loss can be formally described as follows: , Among them, in the above formula Represents two different distributions and The KL divergence between encourages that the embedding representation of the latent space should be domain-independent. p and q refer to the probability distribution of z obtained by calculating the KL divergence.
3. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 2 is characterized in that: After obtaining the embedding layer perturbed by Gaussian noise, the CG-GAN model uses the generator to synthesize the embedding layer representation into simulated data samples, uses the discriminator to determine whether the data is the source domain or the target domain, distinguishes the real data input, and guides the generator to produce realistic samples through the synthetic samples simulated by the generator.
4. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 3 is characterized in that: For the mapping process from the source domain to the target domain, or the mapping process from the target domain to the source domain, the target adversarial loss involved is formally described as follows: 。 5. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 3 is characterized in that: The CG-GAN model adopts cycle consistency loss, retains the semantic information of the original CSI in the domain adaptation process, and remaps the synthetic data in the target domain back to the original source domain. The cycle consistency loss formed by the cycle transformation is formally described as follows: 。 6. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 5 is characterized in that: The CG-GAN model also applies the identity mapping loss so that when the sample is sent to the generator corresponding to its original domain, it can generate samples that maintain domain invariance. The identity mapping loss can be formally described as follows: 。 7. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 6 is characterized in that: The CG-GAN model uses a classification discriminator to maintain the uniqueness of the gait characteristics reflected by the human body in different states during the domain adaptation process, and quantifies the difference between the predicted label and the true label of the gait sample after the generative adversarial network transformation. The gait classification loss can be formally described as follows: , in, represents the set of gait categories, Represents the real source domain data label category, represents the softmax operation, Using cross entropy loss for supervised classification, the CG-GAN model uses the label information corresponding to the source domain data with known labels for supervised classification learning.
8. The cross-domain identity recognition method based on gait conditional generative adversarial network according to claim 7 is characterized in that: The training loss function of the CG-GAN model is formally expressed as follows: , in, , , and They represent the weights of the identity mapping loss, cycle consistency loss, gait classification loss, and KL divergence loss respectively.
9. A cross-domain identity recognition system based on gait conditional generative adversarial network, comprising an IoT device with WiFi function, the IoT device having a domain filter model trained by a server, the domain filter model comprising a CG-GAN model, characterized in that: The CG-GAN model has the features described in any one of claims 1-8.
Citation Information
Patent Citations
Multi-gesture human body target tracking method based on generative adversarial network positive sample enhancement
CN108960086A
Action recognition method based on generative adversarial network in WiFi environment
CN113378718A
Gesture recognition method and device and storage medium
CN114499712A
Athlete style recognition system and method
US20200394413A1
