A cross-scenario wireless gesture recognition method based on unsupervised learning

Abstract

The present invention discloses a cross-scenario wireless gesture recognition method based on unsupervised learning, including: obtaining Doppler frequency shift data and preprocessing; inputting the preprocessed data into a shared feature extractor, and mapping the extracted features to the same latent space through an adversarial autoencoder; assigning pseudo-labels through classification prediction in the warm-up stage, and improving the assignment of pseudo-labels through target-oriented centroid points after the warm-up stage; constructing a target-oriented multivariate normal distribution through all pseudo-labels, and migrating the source domain to the target domain at the latent space level by adding a bias to obtain a trained cross-scenario wireless gesture recognition model; analyzing the newly input target domain samples through the trained cross-scenario wireless gesture recognition model, and predicting the gestures corresponding to the target domain samples in an unlabeled manner. The present invention can alleviate the problem of difficulty in perceptual data labeling, and can achieve stable and high-performance gesture recognition in different scenarios.

Description

Technical Field

[0001] The present invention relates to the field of information and communication technologies, and in particular to a cross-scenario wireless gesture recognition method based on unsupervised learning. Background Art

[0002] Gesture recognition provides users with a more intuitive and natural way of interaction, greatly enriching the interactive experience between people and the digital world. With the rapid development of sensing technology, gesture recognition technology is increasingly being used in fields such as smart home, healthcare, and virtual reality. Existing methods based on visual sensing require certain lighting conditions to achieve high gesture recognition accuracy, are easily interfered by low light conditions, smoke or opaque obstacles, and have privacy and security issues. Methods that rely on motion sensing usually require users to wear special equipment to track body movements to ensure effective and reliable sensing data, which not only increases the cost of additional equipment deployment, but is also inconvenient in life. In general, the above-mentioned gesture recognition technologies require additional expenses in hardware deployment, implementation, or maintenance, and privacy issues are difficult to ignore. In order to address the above issues, Wi-Fi gesture recognition technology relies on the existing widely deployed Wi-Fi infrastructure, making it a cost-effective, privacy-friendly, and non-invasive solution that does not require wearing additional equipment.

[0003] There are two main methods for gesture recognition based on Wi-Fi: model-based methods and learning-based methods. Model-based methods usually rely on establishing accurate and reliable mathematical models to predict the changes in Wi-Fi signals in space and time and map them into corresponding gestures. This method can build corresponding physical models based on expert knowledge and specific scenarios to identify different gestures. However, if the scenario or the gesture of a person is too complex, the differences between different gestures are difficult to describe with physical models. It is necessary to rely on more expert knowledge to build physical models, which is obviously not conducive to the specific implementation of the technology. The learning-based method can automatically extract features and explore patterns from a large amount of collected gesture data to accurately identify different gestures of users. Compared with the model-based method, the learning-based method is more versatile and stable. However, this type of method also has its disadvantages. The most intuitive disadvantage is that this type of method needs to rely on a large amount of gesture data for model training, which has a large demand for data collection and hardware resources.

[0004] In addition, in some scenarios, the collected gesture data not only contains gesture information, but also contains a large amount of information unrelated to gestures, such as the user, environment, user's position and orientation. If a model trained with data from one field is applied to another new field, it will often suffer a significant performance degradation. Therefore, the robustness of gesture recognition models in cross-scenario situations is a problem that must be faced in both model-based and learning-based methods. For the problem of Wi-Fi gesture recognition in cross-scenario situations, the existing small sample method collects a small number of samples in the target scene and marks them for model training to improve the gesture recognition performance of the model in the target scene. However, in practical applications, if there is no pre-planned data collection plan, the data labeling cost will be very high, and it is difficult to accurately label the perception data by self-judgment. In summary, the existing Wi-Fi-based gesture recognition methods have the following technical defects: the learning-based method requires the collection of a large amount of gesture data, and the cost of data labeling is high; the robustness of the gesture recognition model in cross-scenario situations is poor.

[0005] Therefore, the prior art needs to be improved. Summary of the invention

[0006] The technical problem to be solved by the present invention is that, in view of the defects of the prior art, the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning to solve the problems of high cost and poor robustness of the existing Wi-Fi-based gesture recognition method.

[0007] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0008] In a first aspect, the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, comprising:

[0009] Acquiring Doppler frequency shift data, and preprocessing the Doppler frequency shift data;

[0010] The preprocessed source domain data and target domain data are input into a shared feature extractor, and according to a given prior Gaussian distribution, the extracted source domain features and target domain features are mapped to the same latent space through an adversarial autoencoder to obtain source domain encoding features and target domain encoding features;

[0011] Different pseudo-labeling strategies are used to assign pseudo-labels. In the warm-up phase, pseudo-labels are assigned by classification prediction. After the warm-up phase, pseudo-label assignment is improved by target-oriented centroid points.

[0012] A target-oriented multivariate normal distribution is constructed through all pseudo-labels, and the source domain features are enhanced by adding biases. The source domain is migrated to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model.

[0013] The trained cross-scenario wireless gesture recognition model is used to analyze the newly input target domain samples, and the gestures corresponding to the target domain samples are predicted in an unlabeled manner.

[0014] In one implementation, the acquiring Doppler frequency shift data and preprocessing the Doppler frequency shift data includes:

[0015] Acquire the Doppler frequency shift data, and perform cubic spline interpolation processing on the time frame of the Doppler frequency shift data to obtain regularized Doppler frequency shift data;

[0016] The regularized Doppler frequency shift data is standardized to obtain the pre-processed source domain data and target domain data.

[0017] In one implementation, the acquiring the Doppler frequency shift data includes:

[0018] Conjugate multiply the channel state information of the transmitting and receiving antenna pairs on the same Wi-Fi network card, and filter out-of-band noise and static offset;

[0019] The principal component analysis method is used to perform dimensionality reduction processing on the acquired data, the dynamic multipath component of the Doppler frequency shift is retained, and the power distribution in the time domain and the Doppler frequency domain is described using short-time Fourier transform to obtain the Doppler frequency shift data.

[0020] In one implementation, the preprocessed source domain data and target domain data are input into a shared feature extractor, and according to a given prior Gaussian distribution, the extracted source domain features and target domain features are mapped to a same latent space through an adversarial autoencoder to obtain source domain encoding features and target domain encoding features, including:

[0021] Inputting the preprocessed source domain data and target domain data into a shared feature extractor formed by a convolutional neural network to extract the source domain features and the target domain features;

[0022] The source domain features and the target domain features are input into the adversarial autoencoder, and according to a given prior Gaussian distribution, the source domain features and the target domain features are mapped to a same latent space to obtain the source domain encoding features and the target domain encoding features.

[0023] In one implementation, the method of allocating pseudo labels using different pseudo labeling strategies, obtaining preheating stage pseudo labels by classification prediction allocation in the preheating stage, and improving the allocation of pseudo labels by target-oriented centroid points after the preheating stage includes:

[0024] In the warm-up phase, the source domain encoding features and the target domain encoding features are input into the main module classifier, and pseudo labels are assigned to the target domain samples through the predicted maximum value of the main module classifier;

[0025] After the warm-up phase, target-oriented centroids are constructed according to the target domain, and the set of centroids is initialized, and the assignment of pseudo-labels is improved based on the initialized set of centroids.

[0026] In one implementation, the target-oriented multivariate normal distribution is constructed through all pseudo-labels, and the source domain features are enhanced by adding bias, and the source domain is migrated to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model, including:

[0027] The target-oriented multivariate normal distribution is constructed through all pseudo-labels, and the source domain features are enhanced by adding bias to obtain an enhanced source domain feature set;

[0028] The source domain is migrated to the target domain at the latent space level according to the enhanced source domain feature set to obtain the trained cross-scenario wireless gesture recognition model.

[0029] In one implementation, the source domain feature set is an enhanced source domain feature set that achieves an implicit infinite enhancement effect after calculating the upper limit of the classification loss when the number of enhancements N tends to infinity using the law of large numbers;

[0030] The step of migrating the source domain to the target domain at the latent space level according to the enhanced feature set to obtain the trained cross-scenario wireless gesture recognition model includes:

[0031] Determining the overall loss of the cross-scenario wireless gesture recognition model according to the classification loss upper limit;

[0032] Training is performed based on the overall loss of the cross-scene wireless gesture recognition model to obtain the trained cross-scene wireless gesture recognition model.

[0033] In a second aspect, the present invention provides a cross-scenario wireless gesture recognition device based on unsupervised learning, comprising:

[0034] A preprocessing module, used for acquiring Doppler frequency shift data and preprocessing the Doppler frequency shift data;

[0035] The adversarial mapping module is used to input the preprocessed source domain data and target domain data into a shared feature extractor, and map the extracted source domain features and target domain features to the same latent space through the adversarial autoencoder according to the given prior Gaussian distribution, so as to obtain the source domain encoding features and the target domain encoding features;

[0036] The pseudo-labeling module is used to allocate pseudo-labels using different pseudo-labeling strategies. In the warm-up phase, the pseudo-labels are allocated by classification prediction, and after the warm-up phase, the allocation of pseudo-labels is improved by the target-oriented centroid points.

[0037] The transfer training module is used to construct a target-oriented multivariate normal distribution through all pseudo-labels, enhance the source domain features by adding bias, and transfer the source domain to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model;

[0038] The gesture prediction module is used to analyze the newly input target domain sample through the trained cross-scenario wireless gesture recognition model, and predict the gesture corresponding to the target domain sample in an unlabeled manner.

[0039] In a third aspect, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory stores a cross-scene wireless gesture recognition program based on unsupervised learning, and when the cross-scene wireless gesture recognition program based on unsupervised learning is executed by the processor, it is used to implement the operation of the cross-scene wireless gesture recognition method based on unsupervised learning as described in the first aspect.

[0040] In a fourth aspect, the present invention also provides a medium, which is a computer-readable storage medium, and which stores a cross-scene wireless gesture recognition program based on unsupervised learning. When the cross-scene wireless gesture recognition program based on unsupervised learning is executed by a processor, it is used to implement the operation of the cross-scene wireless gesture recognition method based on unsupervised learning as described in the first aspect.

[0041] The present invention adopts the above technical solution to achieve the following effects:

[0042] (1) The present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning. By utilizing the technology of unsupervised domain adaptation, only the data of the target domain needs to be collected without any labeling to achieve high-precision gesture recognition effect, which can alleviate the problem of difficulty in labeling perception data and realize stable and high-performance gesture recognition in different scenarios.

[0043] (2) The present invention divides the cross-scenario Wi-Fi gesture recognition problem into two interrelated sub-problems. By iteratively applying the adversarial feature alignment and pseudo-label guided feature enhancement methods, the adaptation of source domain data to target domain data is completed in the latent space, thereby improving the recognition accuracy of the gesture classifier in the new domain.

[0044] (3) In order to reduce the negative impact of erroneous pseudo-labels in the early stage of training, the present invention adopts a preheating pseudo-labeling technology to divide the entire training cycle into two parts, and adopts corresponding pseudo-labeling strategies in different cycles. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] 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 described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0046] Figure 1 It is a flow chart of the cross-scenario wireless gesture recognition method based on unsupervised learning in the present invention.

[0047] Figure 2 It is a schematic diagram of the detection scene of cross-scene wireless gesture recognition in the present invention.

[0048] Figure 3 It is a schematic diagram of the device deployment and experimental setting of cross-scenario wireless gesture recognition in the present invention.

[0049] Figure 4 It is a schematic diagram of the overall process of the cross-scenario wireless gesture recognition method based on unsupervised learning in the present invention.

[0050] Figure 5 It is a schematic diagram of an ablation experiment of cross-scenario wireless gesture recognition in the present invention.

[0051] Figure 6 It is a functional principle diagram of a terminal in one implementation of the present invention.

[0052] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0054] Exemplary Methods

[0055] To address the problem of Wi-Fi gesture recognition in cross-scenario scenarios, existing small sample methods collect a small number of samples in the target scenario and mark them for model training to improve the gesture recognition performance of the model in the target scenario. However, in practical applications, if there is no pre-planned data collection plan, the data labeling cost will be very high, and it is difficult to accurately label the perceived data through self-judgment. Existing Wi-Fi-based gesture recognition methods have the following technical defects: learning-based methods require the collection of a large amount of gesture data, and the cost of data labeling is high; the robustness of gesture recognition models in cross-scenario scenarios is poor.

[0056] In view of the above technical problems, an embodiment of the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning. The method obtains Doppler frequency shift data and preprocesses the Doppler frequency shift data; inputs the preprocessed source domain data and target domain data into a shared feature extractor, and maps the extracted source domain features and target domain features to the same latent space through an adversarial autoencoder according to a given prior Gaussian distribution, so as to obtain source domain encoding features and target domain encoding features; different pseudo-labeling strategies are used to allocate pseudo-labels, and in the warm-up stage, preheating stage pseudo-labels are allocated through classification prediction, and after the warm-up stage, the allocation of pseudo-labels is improved through target-oriented centroid points; a target-oriented multivariate normal distribution is constructed through all pseudo-labels, and the source domain features are enhanced by adding bias, and the source domain is migrated to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model; the newly input target domain samples are analyzed through the trained cross-scenario wireless gesture recognition model, and the gestures corresponding to the target domain samples are predicted in an unlabeled manner. The embodiments of the present invention can alleviate the problem of difficulty in labeling perception data, and can achieve stable and high-performance gesture recognition in different scenarios.

[0057] like Figure 1 As shown, an embodiment of the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, comprising the following steps:

[0058] Step S100: acquiring Doppler frequency shift data, and preprocessing the Doppler frequency shift data.

[0059] In this embodiment, a cross-scenario wireless gesture recognition method based on unsupervised learning is designed. The challenge of cross-scenario wireless gesture recognition is that when the wireless scene changes, the distribution of sensor data collected by the Wi-Fi device will also change. In order to maintain satisfactory gesture recognition accuracy, the Wi-Fi device is usually required to re-collect and label data samples in the new environment, and then use these labeled new samples to re-train a gesture recognition model and deploy it in the new environment.

[0060] In this embodiment, multiple gesture detection scenarios are set for cross-scenario wireless gesture recognition. Figure 2 As shown, Figure 2 Scene A is an office scene, scene B is an auditorium, and scene C is a classroom. In these gesture detection scenes, the deployment of devices and the settings of the experiment are as follows: Figure 3 As shown, the green circle ( Figure 3 The middle circle mark) is the location of the sensing personnel, and the red box ( Figure 3 The square mark in the middle indicates the location of the receiver, and the blue triangle ( Figure 3The triangle mark in the middle is the location of the transmitter, and there are 5 orientations for the subjects at each location. The device deployment in the sensing area of ​​the above three scenarios is the same as Figure 3 Stay consistent.

[0061] For cross-environment gesture recognition, in this embodiment, one of the three scenes is taken as the target domain in turn, and the rest are taken as the source domain. For cross-position and cross-direction gesture recognition, the position and direction are respectively taken as Figure 3 As shown in the figure, the image is divided into five parts, and one of the positions or directions is still used as the target domain in turn, and the remaining positions and directions are regarded as the source domain. In addition, the channel state information data of the six gestures including push, pull, sweep, clap, slide, circle and Z-shaped gestures collected in the above scenarios are converted into Doppler frequency shift data as the input of the model.

[0062] Based on the above gesture detection scenario, in this embodiment, Doppler frequency shift data of a Wi-Fi device is obtained and preprocessed to obtain gesture sample data for training a gesture recognition model.

[0063] Specifically, in an implementation of this embodiment, step S100 includes the following steps:

[0064] Step S101, acquiring the Doppler frequency shift data, and performing cubic spline interpolation processing on the time frame of the Doppler frequency shift data to obtain regularized Doppler frequency shift data.

[0065] Specifically, in an implementation of this embodiment, obtaining the Doppler frequency shift data includes: conjugate multiplying the channel state information of the transmitting and receiving antenna pairs on the same Wi-Fi network card, and filtering out-band noise and static offset; using a principal component analysis method to perform dimensionality reduction processing on the acquired data, retaining the dynamic multipath component of the Doppler frequency shift, and using a short-time Fourier transform to describe the power distribution in the time domain and the Doppler frequency domain to obtain the Doppler frequency shift data.

[0066] In this embodiment, when acquiring Doppler frequency shift data, the channel state information of the transmitting and receiving antenna pairs on the same Wi-Fi network card is conjugate multiplied, and the out-of-band noise and static offset are filtered. Then, principal component analysis is applied to reduce the dimension of the data to retain the dynamic multipath component of the Doppler frequency shift. Finally, short-time Fourier transform is used to describe the power distribution in the time domain and the Doppler frequency domain. Since different users may have different execution times for the same gesture due to different action modes when making the same gesture, it is difficult for the same user to maintain the same execution time when making the same gesture. In order to keep the dimension of the data consistent, in this embodiment, the time frame of the sample is interpolated by cubic splines, so that the size of each sample is 6*121*1000, where 6 is the number of transceiver links, 121 is the number of frequency domain sampling points, and 1000 is the number of time frames; after cubic spline interpolation, the regularized Doppler frequency shift data is finally obtained.

[0067] Specifically, in an implementation of this embodiment, step S100 further includes the following steps:

[0068] Step S102, performing standardization processing on the regularized Doppler frequency shift data to obtain the pre-processed source domain data and target domain data.

[0069] In view of the problem that there may be differences when the same user makes the same gesture, in order to ignore these small differences and improve the stability of gesture recognition, in this embodiment, the regularized Doppler frequency shift data is standardized:

[0070]

[0071] Wherein, μ and σ represent the mean and standard deviation of the normalized Doppler frequency shift data, respectively.

[0072] This embodiment focuses on the problem of unlabeled cross-scene gesture recognition. Therefore, the labeled and unlabeled gesture data are directly regarded as source domain data and target domain data respectively. The source domain data and the target domain data have the same category, but due to the differences in users, user sensor positions, orientations and environments, the two have different data distributions. In order to achieve knowledge transfer from the source domain to the target domain, the input gesture dataset can be divided into two parts: the source domain labeled dataset and the target domain labeled dataset. and target domain unlabeled dataset Among them, N s and N t denote the number of labeled samples and unlabeled samples respectively. Represents gesture data The mark, A set of tags representing gesture data.

[0073] like Figure 1 As shown, an embodiment of the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, comprising the following steps:

[0074] In step S200, the preprocessed source domain data and target domain data are input into a shared feature extractor, and according to a given prior Gaussian distribution, the extracted source domain features and target domain features are mapped to a same latent space through an adversarial autoencoder to obtain source domain encoding features and target domain encoding features.

[0075] In this embodiment, different perception and recognition scenarios are divided into source domains and target domains, and the perception data samples collected in the new environment are regarded as target domain data. The difficulty of cross-scenario perception gesture recognition is that if there is no prior data collection plan, it is difficult to obtain labeled samples in the new environment, and the labeling of samples requires a certain cost. Without the support of sufficient target domain labeled samples, it is difficult to fully train the cross-scenario gesture recognition model. To address this problem, this embodiment adopts unsupervised domain adaptation technology to deal with it. Although labeled data is difficult to obtain, the acquisition of unlabeled data is low-cost. Through unsupervised domain adaptation technology, this embodiment can achieve accurate recognition of gestures in the target scene by relying only on unlabeled training data in the target domain. Unsupervised domain adaptation technology allows the data distribution of the target domain to be adapted in the absence of new domain labeled data, thereby improving the generalization ability of the model.

[0076] In this embodiment, the overall flow chart of the cross-scenario wireless gesture recognition method is as follows: Figure 4 As shown in the figure, in the training phase, this embodiment divides the cross-scenario gesture recognition problem into two interrelated sub-problems and adopts a two-stage optimization strategy to solve it. In the first stage, an adversarial autoencoder is used to align features to minimize the domain difference between the source domain and the target domain. With the help of pseudo-labeling of target domain samples, the adversarial alignment process retains the category information of the samples. In the second stage, a pseudo-labeling-guided feature enhancement method is applied in the latent space to assist the training of the gesture classifier.

[0077] Specifically, in an implementation of this embodiment, step S200 includes the following steps:

[0078] Step S201: input the preprocessed source domain data and target domain data into a shared feature extractor formed by a convolutional neural network to extract the source domain features and the target domain features.

[0079] In this embodiment, the gesture data x∈x s ∪x t Input into the shared feature extractor composed of convolutional neural networks, the characteristics of the input data can be expressed as:

[0080]

[0081] in, Represents the parameters of the feature extractor.

[0082] The data features of the gesture sample data are extracted by the above shared feature extractor, and then these data features are input into the adversarial autoencoder for encoding.

[0083] Specifically, in an implementation of this embodiment, step S200 further includes the following steps:

[0084] Step S202: input the source domain features and the target domain features into the adversarial autoencoder, and map the source domain features and the target domain features to a same latent space according to a given prior Gaussian distribution to obtain the source domain encoding features and the target domain encoding features.

[0085] In this embodiment, in order to narrow the data distribution difference between the source domain and the target domain, the features extracted by the feature extractor are input into the adversarial autoencoder, and the prior distribution is set to the standard Gaussian distribution, so that both the source domain features and the target domain features are mapped to the same latent space. The encoded features can be expressed as:

[0086] z=AAE(f;Θ A ) (3)

[0087] Among them, Θ A is the parameter of the adversarial autoencoder. Let q(z|f) and p(f) represent the encoding distribution and marginal distribution of feature f respectively, and the aggregated posterior distribution of the encoding feature z is expressed as:

[0088] q(z)=∫q(z|f)p(f)df (4)

[0089] Among them, p(z′) represents the introduced known prior distribution, and z′ represents the regular code derived from the prior distribution. Then, in order to map q(z) to p(z′), it is necessary to minimize the adversarial loss:

[0090] L adv =E z′～p(z′) [logDis(z′)]+E z～p(z) [log(1-Dis(z))] (5)

[0091] In order to make the encoding retain the information of the input features as much as possible, the reconstruction loss also needs to be minimized in this embodiment:

[0092]

[0093] in, is the mean square error, and D(·) represents the decoded output. At the same time, in order to ensure that the feature enhancement guided by the pseudo-label in the second stage is not affected, this embodiment adds an auxiliary classifier after the encoder to retain the category information of the gesture data, which is achieved by minimizing the following loss:

[0094]

[0095] Among them, s (j) and t (j) Represents the index set of source samples and target samples in the jth mini-batch. |s (j) | represents the size of the source domain mini-batch index set, |t (j) | represents the size of the target domain mini-batch index set. and are the predictions of the auxiliary classifier on the source domain and target domain samples, respectively. is the pseudo label assigned to the target domain samples.

[0096] Considering that pseudo-labeling is not accurate in the early stage of training, this embodiment sets a pseudo-labeling warm-up stage in the first few training cycles. In the warm-up stage, θ is set to 0, and after the warm-up is completed, θ is set to 1. At the same time, γ=jγ0 / J is set in formula (7), where γ0 is a hyperparameter, γ0 is set to 0.5, and j and J represent the current iteration and the maximum iteration, respectively. As training progresses, γ will gradually increase to γ0, thereby reducing the impact of incorrect pseudo-labeling during training again.

[0097] From formulas (5), (6), and (7), we can get the loss of the adversarial autoencoder as:

[0098] L aae =L adv +L re +L ac (8)

[0099] In this embodiment, an adversarial autoencoder is used to align gesture data features to minimize the domain difference between the source domain and the target domain. With the help of pseudo-labeling of target domain samples, the adversarial alignment process retains the category information of the samples, so that pseudo-labeling-guided feature enhancement methods can be applied in the latent space to assist the training of the gesture classifier.

[0100] like Figure 1 As shown, an embodiment of the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, comprising the following steps:

[0101] Step S300, using different pseudo-labeling strategies to allocate pseudo-labels, obtaining preheating stage pseudo-labels through classification prediction allocation in the preheating stage, and improving the allocation of pseudo-labels through target-oriented centroid points after the preheating stage.

[0102] In this embodiment, in the second stage of cross-scenario gesture recognition strategy optimization, this embodiment applies a pseudo-marker guided feature enhancement method in the latent space to assist in the training of the gesture classifier.

[0103] Specifically, in an implementation of this embodiment, step S300 includes the following steps:

[0104] Step S301, in the warm-up phase, the source domain encoding feature and the target domain encoding feature are input into the main module classifier, and a pseudo label is assigned to the target domain sample through the predicted maximum value of the main module classifier;

[0105] Step S302, after the warm-up phase, construct target-oriented centroid points according to the target domain, initialize the centroid point set, and improve the allocation of pseudo-labels according to the initialized centroid point set.

[0106] Specifically, in an implementation of the present embodiment, the class centroid points are constructed according to the target domain, and the constructed centroid points are used to improve the pseudo-label assignment results of the target domain samples, including: in each round of iteration, pseudo-labels are assigned to the target domain samples through the predicted maximum value of the main module classifier, and the target coding feature memory library and the pseudo-label memory library are updated; or a set of centroid points is initialized, pseudo-labels are assigned according to the set of centroid points, and the target coding feature memory library and the pseudo-label memory library are updated.

[0107] In this embodiment, a preheating pseudo-labeling strategy is used to assign pseudo-labels to target samples; specifically, preheating pseudo-labeling is to use different pseudo-labeling methods in stages to act on model training. In the preheating stage, pseudo-labels are assigned to target samples based on the classifier's prediction of the target sample. After the preheating stage, class centroids are constructed based on the target domain, and these centroids are used to improve the pseudo-label assignment of target samples. This method helps to reduce the deviation of centroid estimation, thereby reducing the mislabeling of target domain samples. In addition, several memory library modules are introduced in this embodiment. To cache the sample's encoded features, pseudo-labels, and class centroid points respectively. Compared with the iterative update method, the memory library update method can replace outdated information in real time to reduce cumulative deviations and memory costs.

[0108] In this embodiment, the encoding feature z is input into the main module classifier to obtain the prediction of the sample:

[0109]

[0110] Among them, Θ C is the parameter of the main module classifier. According to different training cycles, any one of the following pseudo-labeling methods will be used in this embodiment:

[0111] Pseudo-labeling method 1: In each iteration, the target sample is assigned a pseudo-label by the maximum value of the classifier's prediction:

[0112]

[0113] Then, update the target encoding feature memory library and pseudo-labeled memory banks

[0114] Pseudo-labeling method 2: Initialize the centroid point set c = [c1, c2, ..., c K ], K is the size of the sample label set. The centroid is calculated as follows:

[0115]

[0116] in, is the indicator function. When x=k, but According to the centroid point set c=[c1,c2,…,c K ], the pseudo-label assignment is as follows:

[0117]

[0118] Then, update the target encoding feature memory library and pseudo-labeled memory banks At the same time, the centroid needs to be updated in each iteration:

[0119]

[0120] In this embodiment, through the above preheating pseudo-labeling strategy, pseudo-labels can be assigned to target samples according to the classifier's prediction of target samples in the preheating stage, the pseudo-label assignment of target samples can be improved, the deviation of centroid estimation can be reduced, and thus the incorrect labeling of target domain samples can be reduced.

[0121] like Figure 1 As shown, an embodiment of the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, comprising the following steps:

[0122] Step S400, constructing a target-oriented multivariate normal distribution through all pseudo-labels, and enhancing the source domain features by adding bias, migrating the source domain to the target domain at the latent space level, and obtaining a trained cross-scenario wireless gesture recognition model.

[0123] In this embodiment, a target-oriented multivariate normal distribution constructed by pseudo-labeling of target samples is used, and a bias is added to achieve domain migration from the source domain to the target domain at the latent space level. By alternately optimizing the parameters of the adversarial autoencoder and the parameters of the main module of the gesture recognition model, the cross-scenario gesture recognition problem is solved layer by layer until the model converges.

[0124] Specifically, in an implementation of this embodiment, step S400 includes the following steps:

[0125] Step S401, constructing the target-oriented multivariate normal distribution through all pseudo-labels, and enhancing the source domain features by adding bias to obtain an enhanced source domain feature set;

[0126] Step S402: Migrating the source domain to the target domain at the latent space level according to the enhanced source domain feature set to obtain the trained cross-scenario wireless gesture recognition model.

[0127] In this embodiment, during the training process, feature enhancement is used to transfer knowledge from the source domain to the target domain. First, a multivariate normal distribution needs to be constructed as:

[0128]

[0129] Among them, γ=jγ0 / J is consistent with the setting in formula (7), Δ is the difference between the target average encoding feature and the source average encoding feature. k Describes the relative distance between the target domain and source domain features at the latent space level. is the intra-class covariance of the target domain encoding features, calculated as follows:

[0130]

[0131] Where k = 1, 2, ..., K. The intra-class covariance of the target domain encoding feature can reflect the unique semantic changes of the target domain samples at the latent space level. In this embodiment, the source domain encoding feature z is enhanced by adding a bias s :

[0132]

[0133] in, is the label of the source domain sample. In this way, we get This is equivalent to simulating the labeled source domain features that conform to the target domain distribution at the level of latent space. Assume that the source domain features are enhanced N times:

[0134]

[0135] The enhanced source domain encoding feature set of the current mini batch can be obtained In order to improve the adaptability of source domain features to the target domain, N usually needs to be set to a relatively large value. Obviously, if explicit feature enhancement operations are used, that is, the actual generation of these enhanced features, a very large computational cost is required. Therefore, choosing to deal with this problem through implicit enhancement only requires deriving the upper limit of the loss after infinitely enhancing the source domain features through the law of large numbers, and still achieving the model training effect after explicitly enhancing the features infinitely.

[0136] During model training, the strong correlation between the target features and the predictions is conducive to more accurate assignment of pseudo-labels. To this end, it is necessary to minimize the mutual information between the target features and the predictions. The mutual information between the target features and the predictions is expressed as:

[0137]

[0138] in, represents the average target prediction of the kth category of the jth mini-batch. During the pseudo-label warm-up phase, we calculate for:

[0139]

[0140] After the pseudo-label warm-up is completed, calculate for:

[0141]

[0142] in, Represents the Euclidean distance vector between the target encoding feature and the centroid point.

[0143] Specifically, in an implementation of this embodiment, the source domain feature set is an enhanced source domain feature set that achieves an implicit infinite enhancement effect after using the law of large numbers to calculate the upper limit of the classification loss when the number of enhancements N tends to infinity; the source domain is migrated to the target domain at the latent space level according to the enhanced feature set to obtain the trained cross-scene wireless gesture recognition model, including: determining the overall loss of the cross-scene wireless gesture recognition model according to the classification loss upper limit; and training based on the overall loss of the cross-scene wireless gesture recognition model to obtain the trained cross-scene wireless gesture recognition model.

[0144] In this embodiment, the source domain feature set is enhanced by: not using an explicit enhancement method, but using the law of large numbers to calculate the upper limit of the classification loss when the number of enhancements N tends to infinity, so as to achieve the effect of implicitly enhancing the source domain feature set infinitely, and then obtaining the enhanced source domain feature set; then, determining the overall loss of the cross-scene wireless gesture recognition model according to the classification loss upper limit; and training based on the overall loss of the cross-scene wireless gesture recognition model to obtain the trained cross-scene wireless gesture recognition model.

[0145] In this embodiment, the training method is:

[0146] For the enhanced feature set, the law of large numbers is used in this embodiment to calculate the upper limit of the classification loss when N tends to infinity:

[0147]

[0148] For a random variable X, the concavity of the logarithmic function and Jensen's inequality show that E[log(X)]≤log(E[X]), which converts the above formula into:

[0149]

[0150] Where w=[w1,w2,...,w K ] T ∈R K×H and b=[b1,b2,…,b K ] T ∈R K are the weight vector and bias vector of the last fully connected layer of the classifier. K represents the number of categories, and H is The length of the characteristic dimension of and The loss limit is derived as:

[0151]

[0152] in,

[0153]

[0154] in, The source domain sample The predicted output for the kth category.

[0155] The overall model loss is defined as:

[0156] L=ηL aae -ξL m +ωL c (25)

[0157] In order to perform phased alternating optimization, formula (25) is divided into:

[0158] L=L AAE +L GC (26)

[0159] in,

[0160] L AAB =ηL aae (27)

[0161] denotes the adversarial autoencoder loss.

[0162] L GC =-ξL m +ωL c (28)

[0163] represents the gesture classification loss. Set η, ω to 1 and ξ to 0.1.

[0164] Define the model parameter set as where Θ E The encoder parameters are updated by minimizing formula (27) A , minimize formula (28) to update Θ. The specific algorithm flow is shown in Table 1 below:

[0165] Table 1 Model training algorithm process

[0166]

[0167] The framework of the adversarial autoencoder and the main module framework of the model are shown in Table 2 below:

[0168] Table 2 Framework of adversarial autoencoder and main module framework of the model

[0169]

[0170] The core goal of this embodiment is to address the problem of wireless gesture recognition across scenarios by designing an unsupervised domain adaptation technology, so that the gesture recognition model can achieve high-precision gesture recognition in the target domain by relying only on unlabeled training samples in the target domain. This method greatly reduces the dependence on labeled data in the new domain and improves the adaptability of the gesture recognition model in different environments.

[0171] like Figure 1 As shown, an embodiment of the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, comprising the following steps:

[0172] Step S500: Analyze the newly input target domain sample by using the trained cross-scenario wireless gesture recognition model, and predict the gesture corresponding to the target domain sample in an unlabeled manner.

[0173] In this embodiment, based on the above trained cross-scene wireless gesture recognition model, in different gesture detection scenarios, the newly acquired target domain samples are taken as input, and the newly input target domain samples are analyzed by the trained cross-scene wireless gesture recognition model, that is, the gesture corresponding to the target domain sample can be predicted in an unlabeled manner.

[0174] Experimental verification:

[0175] This embodiment evaluates the overall performance of this method on the Widar3.0 gesture data. The experimental subjects include: 14, 15, 16 users in the classroom environment, 1, 3, 6 users in the auditorium environment, and 7, 8, 9 users in the office environment. For the intra-domain performance evaluation, 80% of all the data in the three environments constitute the training set, and the remaining 20% ​​is used as the test set. For cross-domain performance evaluation, the data is divided according to different cross-position / orientation / environment scenarios, 80% of the target domain data is used as training data, and the remaining 20% ​​is used as test data. At the same time, for the preheating setting of pseudo-labeling, 20% of the entire training cycle is used as the preheating stage of the model pseudo-labeling. In this embodiment, one evaluation entity is selected as the target domain, and the rest are used as the source domain. The specific performance is shown in Table 3 below:

[0176] Table 3 Overall performance of the model

[0177]

[0178] As can be seen from Table 3, the recognition accuracy in the domain can reach 97.63%, and the average recognition accuracy across positions, directions, and environments reaches 97.48%, 90.89%, and 95.48%, respectively. The recognition accuracy of this method exceeds 90% in all cross-scene gesture recognition. And the recognition accuracy across scenes is very stable, with a standard deviation of only 1.54.

[0179] In this example, an ablation experiment is performed to verify the effectiveness of the proposed adversarial feature alignment and pseudo-labeling guidance modules, using a cross-user experiment with users 1, 2, and 3 in a classroom environment. In each evaluation, one user dataset is used as the target domain, and the datasets of the remaining users are used as the source domain. The specific experimental results are shown in Figure 2. Figure 5 As shown. After removing feature alignment and feature enhancement, the recognition accuracy of the model decreased by about 8% and 13% respectively. Without the domain adaptation method, the average recognition accuracy was reduced by 17.5% compared with the present invention. Removing feature alignment or feature enhancement increased the accuracy by about 5% to 9% compared with the method without domain adaptation, which shows the effectiveness of the two components proposed in this embodiment, which can enable the source domain knowledge to be transferred to the target domain.

[0180] In this embodiment, 20% of the total training cycle is used as the warm-up stage to evaluate the effectiveness of the warm-up pseudo-labeling in this embodiment. The cross-user experiment is also conducted using users 1, 2, and 3 in the classroom for verification. Also in each evaluation, one user's dataset is used as the target domain, and the datasets of the remaining users are used as the source domain. The warm-up pseudo-labeling ablation experiment settings are shown in Table 4 below.

[0181] Table 4 Preheating type pseudo-marking ablation experiment

[0182] method No preheat setting Preheating settings User 2, 3 to User 1 81.67％ 90.33％ User 1, 3 to User 2 87.83％ 92.00％ User 1, 2 to User 3 85.33％ 93.67％ average 84.94％ 92.00％

[0183] As shown in Table 4, after using the warm-up strategy, the average recognition accuracy of the model is improved by 7.06%. This shows that the use of the warm-up strategy can better align the initial centroid points with the actual category distribution in the target domain, so that the model can assign more accurate pseudo labels to target samples in the subsequent training stage.

[0184] This embodiment achieves the following technical effects through the above technical solution:

[0185] (1) This embodiment provides a cross-scenario wireless gesture recognition method based on unsupervised learning. By using the unsupervised domain adaptation technology, only the data of the target domain needs to be collected without any labeling to achieve high-precision gesture recognition effect, which can alleviate the problem of difficulty in labeling perception data and realize stable and high-performance gesture recognition in different scenarios.

[0186] (2) This embodiment divides the cross-scenario Wi-Fi gesture recognition problem into two interrelated sub-problems. By iteratively applying the adversarial feature alignment and pseudo-label guided feature enhancement methods, the adaptation of source domain data to target domain data is completed in the latent space, thereby improving the recognition accuracy of the gesture classifier in the new domain.

[0187] (3) In order to reduce the negative impact of erroneous pseudo-labels in the early stage of training, this embodiment uses a warm-up pseudo-labeling technique to divide the entire training cycle into two parts, and uses corresponding pseudo-labeling strategies in different cycles.

[0188] Exemplary Devices

[0189] Based on the above embodiments, the present invention further provides a cross-scenario wireless gesture recognition device based on unsupervised learning, comprising:

[0190] A preprocessing module, used for acquiring Doppler frequency shift data and preprocessing the Doppler frequency shift data;

[0191] The adversarial mapping module is used to input the preprocessed source domain data and target domain data into a shared feature extractor, and map the extracted source domain features and target domain features to the same latent space through the adversarial autoencoder according to the given prior Gaussian distribution, so as to obtain the source domain encoding features and the target domain encoding features;

[0192] The pseudo-labeling module is used to allocate pseudo-labels using different pseudo-labeling strategies. In the warm-up phase, the pseudo-labels are allocated by classification prediction, and after the warm-up phase, the allocation of pseudo-labels is improved by the target-oriented centroid points.

[0193] The transfer training module is used to construct a target-oriented multivariate normal distribution through all pseudo-labels, enhance the source domain features by adding bias, and transfer the source domain to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model;

[0194] The gesture prediction module is used to analyze the newly input target domain sample through the trained cross-scenario wireless gesture recognition model, and predict the gesture corresponding to the target domain sample in an unlabeled manner.

[0195] This embodiment achieves the following technical effects through the above technical solution:

[0196] (1) This embodiment provides a cross-scenario wireless gesture recognition method based on unsupervised learning. By using the unsupervised domain adaptation technology, only the data of the target domain needs to be collected without any labeling to achieve high-precision gesture recognition effect, which can alleviate the problem of difficulty in labeling perception data and realize stable and high-performance gesture recognition in different scenarios.

[0197] (2) This embodiment divides the cross-scenario Wi-Fi gesture recognition problem into two interrelated sub-problems. By iteratively applying the adversarial feature alignment and pseudo-label guided feature enhancement methods, the adaptation of source domain data to target domain data is completed in the latent space, thereby improving the recognition accuracy of the gesture classifier in the new domain.

[0198] (3) In order to reduce the negative impact of erroneous pseudo-labels in the early stage of training, this embodiment uses a warm-up pseudo-labeling technique to divide the entire training cycle into two parts, and uses corresponding pseudo-labeling strategies in different cycles.

[0199] Based on the above embodiment, the present invention further provides a terminal, whose principle block diagram can be as follows: Figure 6 shown.

[0200] The terminal includes: a processor, a memory, an interface, a display screen and a communication module connected through a system bus; wherein the processor of the terminal is used to provide computing and control capabilities; the memory of the terminal includes a storage medium and an internal memory; the storage medium stores an operating system and a computer program; the internal memory provides an environment for the operation of the operating system and the computer program in the storage medium; the interface is used to connect to external devices; the display screen is used to display corresponding information; and the communication module is used to communicate with a cloud server or other devices.

[0201] When the computer program is executed by a processor, it is used to implement the operation of a cross-scenario wireless gesture recognition method based on unsupervised learning.

[0202] It can be understood by those skilled in the art that Figure 6 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal to which the scheme of the present invention is applied. The specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0203] In one embodiment, a terminal is provided, comprising: a processor and a memory, wherein the memory stores a cross-scene wireless gesture recognition program based on unsupervised learning, and the cross-scene wireless gesture recognition program based on unsupervised learning is used to implement the operation of the above cross-scene wireless gesture recognition method based on unsupervised learning when executed by the processor.

[0204] In one embodiment, a storage medium is provided, wherein the storage medium stores a cross-scene wireless gesture recognition program based on unsupervised learning, and when the cross-scene wireless gesture recognition program based on unsupervised learning is executed by a processor, it is used to implement the operation of the above-mentioned cross-scene wireless gesture recognition method based on unsupervised learning.

[0205] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and volatile memory.

[0206] In summary, the present invention provides a cross-scenario wireless gesture recognition method based on unsupervised learning, including: obtaining Doppler frequency shift data and performing preprocessing; inputting the preprocessed data into a shared feature extractor, and mapping the extracted features to the same latent space through an adversarial autoencoder; assigning pseudo-labels through classification prediction in the warm-up stage, and improving the assignment of pseudo-labels through target-oriented centroid points after the warm-up stage; constructing a target-oriented multivariate normal distribution through all pseudo-labels, and migrating the source domain to the target domain at the latent space level by adding a bias to obtain a trained cross-scenario wireless gesture recognition model; analyzing the newly input target domain samples through the trained cross-scenario wireless gesture recognition model, and predicting the gestures corresponding to the target domain samples in an unlabeled manner. The present invention can alleviate the problem of difficult perceptual data labeling, and can achieve stable and high-performance gesture recognition in different scenarios.

[0207] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A cross-scenario wireless gesture recognition method based on unsupervised learning, characterized in that: include: Acquiring Doppler frequency shift data, and preprocessing the Doppler frequency shift data; The preprocessed source domain data and target domain data are input into a shared feature extractor, and according to a given prior Gaussian distribution, the extracted source domain features and target domain features are mapped to the same latent space through an adversarial autoencoder to obtain source domain encoding features and target domain encoding features; Different pseudo-labeling strategies are used to assign pseudo-labels. In the warm-up phase, pseudo-labels are assigned by classification prediction. After the warm-up phase, pseudo-label assignment is improved by target-oriented centroid points. A target-oriented multivariate normal distribution is constructed through all pseudo-labels, and the source domain features are enhanced by adding biases. The source domain is migrated to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model. Analyzing the newly input target domain sample by using the trained cross-scenario wireless gesture recognition model, and predicting the gesture corresponding to the target domain sample in an unlabeled manner; The method of allocating pseudo labels by using different pseudo labeling strategies, obtaining pseudo labels in the warm-up phase by classification prediction allocation in the warm-up phase, and improving the allocation of pseudo labels by target-oriented centroid points after the warm-up phase, includes: In the warm-up phase, the source domain encoding features and the target domain encoding features are input into the main module classifier, and pseudo labels are assigned to the target domain samples through the predicted maximum value of the main module classifier; After the warm-up phase, target-oriented centroids are constructed according to the target domain, and the set of centroids is initialized, and the assignment of pseudo-labels is improved based on the initialized set of centroids.

2. The cross-scenario wireless gesture recognition method based on unsupervised learning according to claim 1 is characterized in that: The acquiring of Doppler frequency shift data and preprocessing of the Doppler frequency shift data comprises: Acquire the Doppler frequency shift data, and perform cubic spline interpolation processing on the time frame of the Doppler frequency shift data to obtain regularized Doppler frequency shift data; The regularized Doppler frequency shift data is standardized to obtain the pre-processed source domain data and target domain data.

3. The cross-scenario wireless gesture recognition method based on unsupervised learning according to claim 2 is characterized in that: The obtaining of the Doppler frequency shift data comprises: Conjugate multiply the channel state information of the transmitting and receiving antenna pairs on the same Wi-Fi network card, and filter out-of-band noise and static offset; The principal component analysis method is used to perform dimensionality reduction processing on the acquired data, the dynamic multipath component of the Doppler frequency shift is retained, and the power distribution in the time domain and the Doppler frequency domain is described using short-time Fourier transform to obtain the Doppler frequency shift data.

4. The cross-scenario wireless gesture recognition method based on unsupervised learning according to claim 1 is characterized in that: The preprocessed source domain data and target domain data are input into a shared feature extractor, and according to a given prior Gaussian distribution, the extracted source domain features and target domain features are mapped to a same latent space through an adversarial autoencoder to obtain source domain encoding features and target domain encoding features, including: Inputting the preprocessed source domain data and target domain data into a shared feature extractor formed by a convolutional neural network to extract the source domain features and the target domain features; The source domain features and the target domain features are input into the adversarial autoencoder, and according to a given prior Gaussian distribution, the source domain features and the target domain features are mapped to a same latent space to obtain the source domain encoding features and the target domain encoding features.

5. The cross-scenario wireless gesture recognition method based on unsupervised learning according to claim 1 is characterized in that: The method constructs a target-oriented multivariate normal distribution through all pseudo-labels, enhances source domain features by adding bias, migrates the source domain to the target domain at the latent space level, and obtains a trained cross-scenario wireless gesture recognition model, including: The target-oriented multivariate normal distribution is constructed through all pseudo-labels, and the source domain features are enhanced by adding bias to obtain an enhanced source domain feature set; The source domain is migrated to the target domain at the latent space level according to the enhanced source domain feature set to obtain the trained cross-scenario wireless gesture recognition model.

6. The cross-scenario wireless gesture recognition method based on unsupervised learning according to claim 5 is characterized in that: The source domain feature set is an enhanced source domain feature set that achieves an implicit infinite enhancement effect after calculating the upper limit of the classification loss when the number of enhancements N tends to infinity using the law of large numbers; The step of migrating the source domain to the target domain at the latent space level according to the enhanced source domain feature set to obtain the trained cross-scenario wireless gesture recognition model includes: Determining the overall loss of the cross-scenario wireless gesture recognition model according to the classification loss upper limit; Training is performed based on the overall loss of the cross-scene wireless gesture recognition model to obtain the trained cross-scene wireless gesture recognition model.

7. A cross-scenario wireless gesture recognition device based on unsupervised learning, used to implement the cross-scenario wireless gesture recognition method based on unsupervised learning as described in any one of claims 1-6, characterized in that: include: A preprocessing module, used for acquiring Doppler frequency shift data and preprocessing the Doppler frequency shift data; The adversarial mapping module is used to input the preprocessed source domain data and target domain data into a shared feature extractor, and map the extracted source domain features and target domain features to the same latent space through the adversarial autoencoder according to the given prior Gaussian distribution, so as to obtain the source domain encoding features and the target domain encoding features; The pseudo-labeling module is used to allocate pseudo-labels using different pseudo-labeling strategies. In the warm-up phase, the pseudo-labels are allocated by classification prediction, and after the warm-up phase, the allocation of pseudo-labels is improved by the target-oriented centroid points. The transfer training module is used to construct a target-oriented multivariate normal distribution through all pseudo-labels, enhance the source domain features by adding bias, and transfer the source domain to the target domain at the latent space level to obtain a trained cross-scenario wireless gesture recognition model; The gesture prediction module is used to analyze the newly input target domain sample through the trained cross-scenario wireless gesture recognition model, and predict the gesture corresponding to the target domain sample in an unlabeled manner.

8. A terminal, characterized in that: include: A processor and a memory, wherein the memory stores a cross-scene wireless gesture recognition program based on unsupervised learning, and when the cross-scene wireless gesture recognition program based on unsupervised learning is executed by the processor, it is used to implement the operation of the cross-scene wireless gesture recognition method based on unsupervised learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a cross-scene wireless gesture recognition program based on unsupervised learning, and when the cross-scene wireless gesture recognition program based on unsupervised learning is executed by a processor, it is used to implement the operation of the cross-scene wireless gesture recognition method based on unsupervised learning as described in any one of claims 1-6.

Images