Action recognition method and related equipment based on lightweight supervised diffusion model

Abstract

The present application discloses a method for motion recognition based on a lightweight supervised diffusion model and related equipment, and relates to the field of wireless communication technology. The method for motion recognition based on a lightweight supervised diffusion model comprises: in response to a start recognition signal, obtaining wireless channel state information data; according to a preset lightweight supervised diffusion model, performing diffusion restoration on the wireless channel state information data to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model comprises the steps of introducing sample label information and exponential sliding average; inputting the restored channel state information data into a preset human motion recognition model to obtain a human motion recognition result. The present application adopts a lightweight supervised diffusion model to perform diffusion restoration on the wireless channel state information data, uses fewer training rounds and sampling iteration rounds to obtain enhanced samples, and improves the accuracy of human motion recognition by extracting and classifying features of the enhanced samples.

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and in particular to an action recognition method and related equipment based on a lightweight supervised diffusion model. Background Art

[0002] Although human body recognition based on tools such as cameras, wearable sensors and radars can meet the needs of specific scenarios to a certain extent, these methods often have obvious limitations. First of all, cameras and wearable sensors are not only restricted in use, but also require users to carry or wear these devices in person, which brings many inconveniences to practical applications. Researchers have begun to explore the use of CSI (Channel State Information) data for human body recognition. CSI data has the characteristics of non-contact, good data protection, adaptability to more complex recognition scenarios, and low requirements for data collection. Therefore, it has received widespread attention in recent years.

[0003] At present, traditional action recognition methods based on lightweight supervised diffusion models often learn the semantic spatial features of human activities from CSI data through advanced neural network models. However, when processing data, there will be local missing or insufficient data size, resulting in low accuracy of wireless human action recognition.

[0004] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The main purpose of this application is to provide a motion recognition method and related equipment based on a lightweight supervised diffusion model, aiming to solve the technical problem of low accuracy of wireless human motion recognition.

[0006] To achieve the above objectives, the present application proposes an action recognition method based on a lightweight supervised diffusion model, the method comprising:

[0007] In response to the start identification signal, acquiring wireless channel state information data;

[0008] According to a preset lightweight supervised diffusion model, the wireless channel state information data is diffused and restored to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average;

[0009] The restored channel state information data is input into a preset human action recognition model to obtain a human action recognition result.

[0010] In one embodiment, before the step of performing diffusion restoration on the wireless channel state information data according to a preset lightweight supervised diffusion model to obtain restored channel state information data, the step includes:

[0011] Obtain channel state information data samples and corresponding human motion category labels;

[0012] Encoding the preset time step and the human action category label to obtain a target time step encoding;

[0013] The target time step encoding is used as a bias parameter and embedded into a preset residual network model;

[0014] Performing feature processing on the channel state information data sample to obtain target sample features;

[0015] Inputting the target sample features into the residual network model to obtain a residual output result;

[0016] Training a preset denoising function through a preset output layer network and the residual output result;

[0017] Updating the time step, inputting the residual output result into the residual network model for iteration, until the time step reaches a preset time step range, and obtaining a target denoising function;

[0018] According to the target denoising function, a lightweight supervised diffusion model is determined.

[0019] In one embodiment, the step of performing feature processing on the channel state information data sample to obtain target sample features includes:

[0020] Adding noise to the channel state information data sample according to preset Gaussian noise to obtain a diffuse noise sample;

[0021] Randomly sampling from the diffuse noise samples to obtain sampled noise samples;

[0022] The features of the sampled noise samples are extracted through a preset first convolutional layer and a nonlinear activation function to obtain target sample features.

[0023] In one embodiment, the step of encoding the preset time step and the human motion category label to obtain the target time step encoding comprises:

[0024] Perform step encoding and linear projection on each preset time step to obtain time step encoding information;

[0025] Encode the human action category label to obtain label encoding information;

[0026] Mixing the label encoding information with the time step encoding information to obtain label guided time step information;

[0027] The label guidance time step information is encoded through a preset linear layer to obtain a target time step encoding.

[0028] In one embodiment, the step of training a preset denoising function through a preset output layer network and a residual output result includes:

[0029] Restoring the residual output result through a deconvolution layer to obtain first restored data;

[0030] Extracting features of the first restored data through the nonlinear activation function and the second convolutional layer to obtain first restored data features;

[0031] The preset denoising function is trained through the first restored data feature.

[0032] In one embodiment, the step of inputting the residual output result into the residual network model for iteration until the time step reaches a preset time step range to obtain a target denoising function includes:

[0033] If the number of iterations reaches the preset range of iterations, the preset model parameters are subjected to exponential weighted averaging;

[0034] The residual output result is input into the residual network model of the model parameters to complete the exponential weighted average processing and iterate until the time step reaches the preset time step range to obtain the target denoising function.

[0035] In addition, to achieve the above purpose, the present application also proposes an action recognition device based on a lightweight supervised diffusion model, and the action recognition device based on a lightweight supervised diffusion model includes:

[0036] An acquisition module, configured to acquire wireless channel state information data in response to a start identification signal;

[0037] A diffusion module, used for performing diffusion restoration on the wireless channel state information data according to a preset lightweight supervised diffusion model to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average;

[0038] The recognition module is used to input the restored channel state information data into a preset human action recognition model to obtain a human action recognition result.

[0039] In addition, to achieve the above-mentioned purpose, the present application also proposes an action recognition device based on a lightweight supervised diffusion model, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the action recognition method based on a lightweight supervised diffusion model as described above.

[0040] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the action recognition method based on the lightweight supervised diffusion model as described above are implemented.

[0041] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the action recognition method based on the lightweight supervised diffusion model as described above.

[0042] One or more technical solutions proposed in this application have at least the following technical effects:

[0043] Compared with the related art, the traditional action recognition method based on the lightweight supervised diffusion model often learns the semantic spatial characteristics of human activities from CSI data through an advanced neural network model. However, there will be local missing or insufficient data size when processing data, resulting in low prediction accuracy of wireless human action recognition. In comparison, the present application obtains wireless channel state information data in response to the start recognition signal; according to the preset lightweight supervised diffusion model, the wireless channel state information data is diffusely restored to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes the steps of introducing sample label information and exponential sliding average; the restored channel state information data is input into the preset human action recognition model to obtain the human action recognition result. It can be understood that the present application adopts the lightweight supervised diffusion model trained by the steps of introducing sample label information and exponential sliding average to diffuse and restore the wireless channel state information data, helping the model to obtain more information-rich enhanced samples in the context of smaller training rounds and fewer sampling iteration rounds, and further extracts and classifies data features of the enhanced samples through the human action recognition model to improve the accuracy of human action recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0045] In order to more clearly illustrate the embodiments of the present application 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, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0046] Figure 1 A flowchart diagram of the first embodiment of the method for action recognition based on a lightweight supervised diffusion model of the present application;

[0047] Figure 2 A flowchart diagram of the second embodiment of the action recognition method based on the lightweight supervised diffusion model of the present application is provided;

[0048] Figure 3 A flowchart diagram of the third embodiment of the method for action recognition based on a lightweight supervised diffusion model of the present application;

[0049] Figure 4 This is a schematic diagram of the module structure of the action recognition device based on the lightweight supervised diffusion model in an embodiment of the present application;

[0050] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the action recognition method based on the lightweight supervised diffusion model in the embodiment of the present application.

[0051] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0053] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0054] The main solutions of the embodiments of this application are:

[0055] In response to the start identification signal, acquiring wireless channel state information data;

[0056] According to a preset lightweight supervised diffusion model, the wireless channel state information data is diffused and restored to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average;

[0057] The restored channel state information data is input into a preset human action recognition model to obtain a human action recognition result.

[0058] In this embodiment, the present application uses a motion recognition device based on a lightweight supervised diffusion model as the execution subject. For ease of description, it is specifically described below as “device”.

[0059] Due to existing technologies, traditional action recognition methods based on lightweight supervised diffusion models often learn the semantic spatial features of human activities from CSI data through advanced neural network models. However, when processing data, there will be local missing or insufficient data size, resulting in low prediction accuracy of wireless human action recognition.

[0060] The present application provides a solution, which adopts the steps of introducing sample label information and exponential sliding average to train a lightweight supervised diffusion model to perform diffusion restoration on wireless channel state information data, so as to help the model obtain more information-rich enhanced samples in the context of smaller training rounds and fewer sampling iteration rounds, and further extract and classify data features of the enhanced samples through the human motion recognition model, so as to improve the accuracy of human motion recognition.

[0061] Based on this, the embodiment of the present application provides an action recognition method based on a lightweight supervised diffusion model. Figure 1 , Figure 1 This is a flow chart of the first embodiment of the action recognition method based on the lightweight supervised diffusion model of the present application.

[0062] In this embodiment, the action recognition method based on the lightweight supervised diffusion model includes steps S10 to S30:

[0063] Step S10, in response to the start identification signal, obtaining wireless channel state information data;

[0064] It should be noted that the start recognition signal is a trigger signal that instructs the system to start executing the human motion recognition process. The wireless channel state information data (CSI data) is the information about the impact of the channel on the transmission signal in wireless communication, including parameters such as signal amplitude, phase and time delay, which can be used to recognize human motion.

[0065] Step S20, performing diffusion restoration on the wireless channel state information data according to a preset lightweight supervised diffusion model to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average;

[0066] It should be noted that the lightweight supervised diffusion model is an optimized supervised diffusion model that improves the efficiency and accuracy of the model by reducing computational complexity and introducing supervisory information. Diffusion recovery refers to the process of restoring noisy or incomplete data to its original state using a diffusion model. The sample label information is the label corresponding to the sample data in model training, which is used for supervised learning and guides the model to correctly learn data features. The exponential moving average (EMA) is a smoothing technique that reduces volatility during training and improves model stability by performing exponential weighted averaging on model parameters.

[0067] Step S30, inputting the restored channel state information data into a preset human action recognition model to obtain a human action recognition result.

[0068] It should be noted that the human motion recognition model is a machine learning model for identifying and classifying human motions, and is based on a deep learning method.

[0069] Exemplarily, the device is equipped with multiple wireless access points (APs) to capture the changes in wireless signals generated by the movement of the occupants. When the device receives an instruction to start recognition (for example, triggered by a voice command or a mobile phone APP), the device starts to collect CSI data from each APs. The collected CSI data may be incomplete or contain noise due to various factors (such as signal interference, hardware limitations). A pre-trained lightweight supervised diffusion model is used to process these data. During the training process of the model, the label information of the samples is introduced to guide the model to learn the correct data features, and the exponential sliding average technique is used to smooth the training process and reduce parameter fluctuations. Through this model, the noise and incomplete information in the CSI data are removed to restore more accurate and complete CSI data. The restored CSI data is then input into a preset human action recognition model. This human action recognition model is a deep neural network that is trained to extract features from CSI data and recognize human actions. For example, the model may recognize that the occupant is walking, sitting or jumping. Finally, the device outputs the recognition result of the human action.

[0070] This embodiment provides an action recognition method based on a lightweight supervised diffusion model, which uses a lightweight supervised diffusion model trained by introducing sample label information and exponential sliding average steps to diffuse and restore wireless channel state information data, helping the model to obtain enhanced samples with richer information in the context of smaller training rounds and fewer sampling iteration rounds. The enhanced samples are further extracted and classified by the human action recognition model, thereby improving the accuracy of human action recognition.

[0071] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2 Before step S20, the action recognition method based on the lightweight supervised diffusion model further includes:

[0072] Obtain channel state information data samples and corresponding human motion category labels;

[0073] Encoding the preset time step and the human action category label to obtain a target time step encoding;

[0074] The target time step encoding is used as a bias parameter and embedded into a preset residual network model;

[0075] Performing feature processing on the channel state information data sample to obtain target sample features;

[0076] Inputting the target sample features into the residual network model to obtain a residual output result;

[0077] Training a preset denoising function through a preset output layer network and the residual output result;

[0078] Updating the time step, inputting the residual output result into the residual network model for iteration, until the time step reaches a preset time step range, and obtaining a target denoising function;

[0079] According to the target denoising function, a lightweight supervised diffusion model is determined.

[0080] It should be noted that the channel state information data samples are data samples collected from the wireless channel, which reflect the state of the wireless signal during the propagation process, such as the amplitude and phase of the signal. The human motion category label is a label corresponding to the channel state information data sample, which is used to indicate the human motion category corresponding to the sample, such as "walking", "running", etc. The time step encoding is to encode the time step (iterative step in the diffusion model) and the human motion category label, and the target time step encoding obtained is used to guide the training process of the model. The residual network model is a deep learning model, which solves the degradation problem in deep network training by introducing residual connections (i.e., skipping one or more layers of connections). The target sample feature is a channel state information data sample after feature processing, which is used to input into the residual network model for further processing. The residual output result is the output result obtained by the residual network model after processing the input feature sample, which reflects the model's understanding and processing of the input data. The denoising function is a function used in the diffusion model to recover clean data from noisy data, and is one of the goals of model training. The lightweight supervised diffusion model is an optimized supervised diffusion model, which improves the efficiency and accuracy of the model by reducing the computational complexity and introducing supervisory information.

[0081] Exemplarily, the device deploys multiple wireless access points to collect CSI data generated by the activities of residents, and assigns a corresponding human motion category label to each CSI data sample.

[0082] The device encodes the time steps and human action category labels so that the model can process them.

[0083] The target time step encoding is used as an additional input and embedded into the residual network model. This encoding acts as a bias parameter to affect the model's weight update, allowing the model to take into account the information of the time step and action category. The CSI data samples are preprocessed, such as normalization and denoising, to extract features useful for action recognition. The processed features are input into the residual network model, and the model calculates the residual output results, which reflect the model's processing of the input features.

[0084] The residual output is used to train a denoising function, which aims to recover clean CSI data from noisy data.

[0085] As the time step is updated, the model is continuously iterated until the preset time step range is reached. In this process, the denoising function is continuously optimized until the data that meets the preset requirements can be accurately recovered from the noise.

[0086] Finally, we obtain a trained denoising function, which constitutes the core of our lightweight supervised diffusion model and can accurately recover data that meets the preset requirements from noisy CSI data.

[0087] It is understandable that, based on the existing diffusion model, this application has built a set of lightweight supervised diffusion models applicable to CSI analysis and identification by introducing label information and exponential sliding average schemes, abandoning the time-consuming spectrum graph calculation in the original diffusion model, and under the experimental conditions of missing values ​​or predictiveness, only a small number of iterations and a smaller network can be used to obtain better data generation accuracy. The model can adapt to data enhancement or enhancement under missing conditions, and can also generate data at future times. It can effectively improve the anti-interference ability and recognition accuracy of WiFi human behavior recognition under interference conditions.

[0088] This application proposes a SLDM training framework, which effectively overcomes the limitations of existing diffusion models in the CSI enhancement process. SLDM introduces label information and an exponential sliding average strategy, and incorporates more explicit prior knowledge into the training process of the diffusion model, thereby greatly improving the efficiency of model training and enabling the model to be trained more smoothly. In short, SLDM compresses the model size proposed by existing work by introducing label information of data samples and an exponential sliding average strategy, helping the model to obtain more information-rich enhanced samples in the context of smaller training rounds and fewer sampling iteration rounds. Figure 3 As shown on the left, SLDM proposes a supervised diffusion model for data enhancement, and uses a convolutional neural network to further extract data features and classify them, thereby improving the accuracy of human motion recognition based on WiFi and realizing wireless human perception with certain anti-interference capabilities.

[0089] The basic idea of ​​the diffusion model is to slowly destroy the structure in the data distribution through an iterative forward diffusion process. The model learns the process of back-diffusion from images close to random noise to real data, thereby recovering the data structure, resulting in a highly flexible and easy-to-handle data generation model. In short, the diffusion model can learn a model distribution through forward and backward diffusion processes. To approximate the data distribution The forward process will ) is converted to a fixed Gaussian distribution of step Markov chain , and the reverse process is achieved through a learnable Step Markov chain, Gaussian distribution Convert to The learning formula of the forward process is shown in formula (1).

[0090] (1)

[0091] in, is a variable with variance The fixed Gaussian transfer equation of . By using the increased variance schedule By gradually and slowly adding noise to the CSI data, this formula can Convert to .

[0092] After the forward diffusion noise addition process is completed, the reverse recovery process can be formally expressed as shown in formula (2).

[0093] (2)

[0094] in, Is a parameter Learnable Gaussian transfer equation for the joint distribution Will Convert to The goal of the diffusion model is to minimize the negative log-likelihood , which is equivalent to minimizing and Since they are both Gaussian distributions, The mean and variance of The mean and variance of . The mean and variance of can be formulated as shown in formula (3) and formula (4) respectively.

[0095] (3)

[0096] (4)

[0097] in, , , .therefore, The mean and variance of can be expressed parameterized as shown in Equation (5) and Equation (6).

[0098] (5)

[0099] (6)

[0100] in, , Represents the denoising function, which is used to extract the current sample Estimated noise ,because and The variance of is constant, and the KL divergence is simplified to and distance, optimize the parameters The objective function can be formally expressed as shown in formula (7).

[0101] (7)

[0102] By minimizing The loss function To train the denoising function After that, you can iterate To use synthesis , as shown in formula (8).

[0103] (8)

[0104] in, .when hour, ;when hour, .

[0105] In a feasible implementation manner, the step of performing feature processing on the channel state information data sample to obtain target sample features includes:

[0106] Adding noise to the channel state information data sample according to preset Gaussian noise to obtain a diffuse noise sample;

[0107] Randomly sampling from the diffuse noise samples to obtain sampled noise samples;

[0108] The features of the sampled noise samples are extracted through a preset first convolutional layer and a nonlinear activation function to obtain target sample features.

[0109] It should be noted that the Gaussian noise is a random noise that statistically obeys a normal distribution and is usually used to simulate random errors or background noise of electronic devices. The channel state information data samples (CSI data samples) are data samples collected from wireless communication channels. These data reflect the state of the signal during propagation, such as the amplitude and phase of the signal.

[0110] The diffuse noise sample is a data sample with noise obtained by adding Gaussian noise to the CSI data sample to simulate the interference of the data during transmission. The random sampling is the process of randomly selecting a part of individuals from the population as samples, and these samples are used to represent the population for analysis. The first convolutional layer is a convolutional layer used to extract input data features in deep learning, especially in convolutional neural networks (CNNs).

[0111] The nonlinear activation function is a function that introduces nonlinear factors, such as ReLU (Rectified Linear Unit), which allows the model to learn and simulate complex nonlinear relationships. The target sample features are the features of the CSI data samples obtained after feature extraction, and these features are used for subsequent data processing and analysis.

[0112] For example, assume that there is a wireless signal receiver that collects CSI data within a certain period of time. In order to simulate the interference that the wireless signal may encounter during transmission, the device adds noise to the CSI data sample according to a preset Gaussian noise distribution. This preset Gaussian noise distribution has a specific mean and standard deviation, which simulates the random interference that the signal may encounter during transmission.

[0113] After adding Gaussian noise, the device obtains a series of diffuse noise samples. Next, the device randomly selects a part of these samples as training samples. This process is called random sampling. For example, the device can randomly select 10 samples from 100 samples collected in each time period as the device's sampling noise samples.

[0114] The sampled noise samples are input into a deep learning model and passed through a convolutional layer. This convolutional layer contains multiple filters, each of which is responsible for extracting specific features from the sample. The output of the convolutional layer then passes through a nonlinear activation function, such as ReLU, which sets negative values ​​to zero, enhancing the model's nonlinear expression of features. After this step, the device obtains the target sample features, which can be used for subsequent human action recognition or other analysis tasks.

[0115] In a feasible implementation manner, the step of encoding the preset time step and the human motion category label to obtain the target time step encoding includes:

[0116] Perform step encoding and linear projection on each preset time step to obtain time step encoding information;

[0117] Encode the human action category label to obtain label encoding information;

[0118] Mixing the label encoding information with the time step encoding information to obtain label guided time step information;

[0119] The label guidance time step information is encoded through a preset linear layer to obtain a target time step encoding.

[0120] It should be noted that the time step in the sequence model or time series analysis refers to each discrete time point in the sequence, which can be regarded as a slice of the data on the time axis.

[0121] The step encoding is the process of converting the time steps into a numerical or vector representation that can be used in the model to capture the relative or absolute relationship between the time steps. The linear projection is a linear transformation that maps the input data to a new space, usually used for dimensionality reduction or feature extraction.

[0122] The time step encoding information refers to the encoding information of the time step obtained after step encoding and linear projection, which is used for subsequent model processing. The human action category label refers to a classification label related to human actions, which is used to indicate a specific action category, such as "walking", "running", etc.

[0123] The label encoding is the process of converting a category label into a numerical value or vector so that the model can process and learn it. The label guidance time step information is data that combines the time step encoding information and the label encoding information, and is used to provide the model with specific action category guidance at the time step.

[0124] The linear layer is a layer that performs linear transformation in a neural network and is usually used for further processing and learning of features. The target time step encoding refers to the label guidance time step information after processing by the linear layer, which is used as the input of the model for final action recognition or other tasks.

[0125] It is understandable that in order to integrate the time step information of the diffusion into In SLDM, for each time step Step encoding and linear projection were performed to obtain step embedding information Specifically, SLDM uses the classic sine and cosine functions to calculate the time step encoding , as shown in formula (9).

[0126]

[0127] in, .

[0128] This application further introduces label information into the training architecture of the SLDM framework to help the model converge more quickly. The specific approach is to input the human activity category label information of the sample while inputting the CSI data. After the same scale encoding operation as time step encoding, the encoded information of the label is then The time step encoding information obtained Mixing, obtaining time step information with label guidance (still recorded as ). SLDM then uses two linear projection layers to calculate as steps embedded in which and are the weights of the two linear layers, and is the bias of the two layers. The time step encoding process here can be expressed as , the final code obtained The embedded features at the time step with label guidance are input into the subsequent residual block to guide the model to perform diffusion learning.

[0129] In a feasible implementation manner, the step of training a preset denoising function through a preset output layer network and a residual output result includes:

[0130] Restoring the residual output result through a deconvolution layer to obtain first restored data;

[0131] Extracting features of the first restored data through the nonlinear activation function and the second convolutional layer to obtain first restored data features;

[0132] The preset denoising function is trained through the first restored data feature.

[0133] It should be noted that the deconvolution layer is also called a transposed convolution layer, which is a neural network layer used to upsample or restore the data dimension. It is usually used to recover a higher resolution representation from compressed or reduced dimensional data.

[0134] The residual output result is the output obtained after the residual network model processes the input data, and includes the feature representation after being processed by the network layer.

[0135] The first restored data is data obtained after the residual output result is processed by the deconvolution layer, and is an intermediate product of the denoising or data recovery process.

[0136] The nonlinear activation function is a function used to introduce nonlinearity, such as ReLU (Rectified Linear Unit), which allows the model to learn and simulate complex nonlinear relationships.

[0137] The second convolutional layer is a convolutional layer in a neural network, which is usually used to extract features of the input data.

[0138] The first restored data feature refers to the feature extracted from the first restored data after being processed by the nonlinear activation function and the second convolution layer.

[0139] Exemplarily, the device uses a residual network model to process CSI data. At a certain stage of the model, the device uses a deconvolution layer to upsample or restore the output of the residual network to obtain a higher resolution data representation. This operation can help us restore a representation that is closer to the original CSI data, which we call the first restored data.

[0140] The first restored data then passes through a nonlinear activation function, such as ReLU, to increase the nonlinear expression ability of the model. Next, the data enters the second convolutional layer, which is used to further extract features from the first restored data. This series of operations helps the device extract useful features from the restored data, which will be used to train the denoising function.

[0141] Finally, the device uses the first restored data features obtained from the second convolutional layer to train a denoising function. The goal of this denoising function is to reduce the noise in the CSI data and restore a clearer signal. In this way, the model can learn how to restore qualified data from noisy CSI data, thereby improving the accuracy of human action recognition.

[0142] In a feasible implementation manner, the step of inputting the residual output result into the residual network model for iteration until the time step reaches a preset time step range to obtain a target denoising function includes:

[0143] If the number of iterations reaches the preset range of iterations, the preset model parameters are subjected to exponential weighted averaging;

[0144] The residual output result is input into the residual network model of the model parameters to complete the exponential weighted average processing and iterate until the time step reaches the preset time step range to obtain the target denoising function.

[0145] It should be noted that the number of iterations is the number of times a specific step is repeated during the algorithm or model training process. In deep learning, the number of iterations usually refers to the number of repetitions during the forward propagation and backpropagation processes.

[0146] The preset iteration range is an interval or target value of the number of iterations set before training the model, which is used to control the duration of the training process or to achieve a certain performance standard.

[0147] Exponential Weighted Average (EWA) is a method of calculating an average that gives more weight to recent data and less weight to older data. This averaging method can smooth the data and reduce the impact of short-term fluctuations.

[0148] Model parameters are variables that are adjusted during the learning process in machine learning models to optimize the model to better fit the training data. Parameters can include weights, biases, etc. Exponential weighted average processing is a method of applying exponential weighted average to model parameters to update parameter values, making the model more stable during training, reducing overfitting, and improving the generalization ability of the model.

[0149] The residual output result is the output obtained after the residual network model processes the input data, which contains the feature representation after the network layer processing. The time step is each discrete time point in the sequence.

[0150] The preset time step range is an interval or target value of a time step set before model training or inference, which is used to control the length or range of time series data processed by the model. The target denoising function is a denoising function in signal processing. The goal of the denoising function is to reduce the noise in the data and restore a clearer signal. The target denoising function refers to a function that can achieve this goal after training and optimization.

[0151] It is understandable that in order to avoid possible weight fluctuations during the training process and make the training phase relatively smooth, SLDM also introduces an exponential moving average strategy (EMA). EMA aims to reduce volatility during the training process and make the model more robust by performing exponential weighted averaging on model parameters. A total of 1,000 iterations were used in the SLDM module training, and different parameter update strategies were used in different training phases. Specifically, in the first 300 iterations, SLDM used the traditional parameter update method, while in the subsequent iterations, SLDM used the EMA strategy to update the model parameters. The core of the EMA strategy lies in the update of parameter values, and its mathematical expression is shown in formula (10).

[0152] (10)

[0153] In the formula, Indicates at time The exponential moving average of Indicates at time The exponential moving average of Indicates at time The original parameter value of is the decay coefficient of the exponential moving average, which controls the weight of historical data to the average value. In the actual implementation, 0.995 was empirically selected. This strategy effectively smooths the parameter update process, improves the stability and generalization ability of the SLDM model, and can restore CSI data with authenticity and enhancement effects.

[0154] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the action recognition method based on the lightweight supervised diffusion model of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0155] This application also provides an action recognition device based on a lightweight supervised diffusion model, please refer to Figure 4 , the action recognition device based on the lightweight supervised diffusion model includes:

[0156] An acquisition module 10, configured to acquire wireless channel state information data in response to a start identification signal;

[0157] A diffusion module 20 is used to perform diffusion restoration on the wireless channel state information data according to a preset lightweight supervised diffusion model to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average;

[0158] The recognition module 30 is used to input the restored channel state information data into a preset human action recognition model to obtain a human action recognition result.

[0159] And / or, the action recognition device based on the lightweight supervised diffusion model includes:

[0160] A first acquisition module is used to acquire channel state information data samples and corresponding human motion category labels;

[0161] A first encoding module, used for encoding a preset time step and the human motion category label to obtain a target time step code;

[0162] A first embedding module, used for embedding the target time step encoding as a bias parameter into a preset residual network model;

[0163] A first processing module, configured to perform feature processing on the channel state information data sample to obtain target sample features;

[0164] A first input module, used to input the target sample features into the residual network model to obtain a residual output result;

[0165] A first training module, used for training a preset denoising function through a preset output layer network and the residual output result;

[0166] A first iteration module is used to update the time step, input the residual output result into the residual network model for iteration until the time step reaches a preset time step range, and obtain a target denoising function;

[0167] The first determination module is used to determine a lightweight supervised diffusion model according to the target denoising function.

[0168] And / or, the first processing module includes:

[0169] A first diffusion module, configured to add noise to the channel state information data sample according to a preset Gaussian noise to obtain a diffusion noise sample;

[0170] A first sampling module, used for randomly sampling from the diffuse noise samples to obtain sampled noise samples;

[0171] The first extraction module is used to extract features of the sampled noise sample through a preset first convolution layer and a nonlinear activation function to obtain target sample features.

[0172] And / or, the first encoding module includes:

[0173] A first projection module is used to perform step encoding and linear projection on each preset time step to obtain time step encoding information;

[0174] The second encoding module is used to encode the human body action category label to obtain label encoding information;

[0175] A first mixing module, used for mixing the label encoding information with the time step encoding information to obtain label guidance time step information;

[0176] The third encoding module is used to encode the label guidance time step information through a preset linear layer to obtain a target time step encoding.

[0177] And / or, the first training module includes:

[0178] A first recovery module, used for recovering the residual output result through a deconvolution layer to obtain first recovered data;

[0179] A first extraction module, used for performing feature extraction on the first restored data through the nonlinear activation function and the second convolution layer to obtain first restored data features;

[0180] The second training module is used to train a preset denoising function through the first restored data feature.

[0181] And / or, the first iteration module includes:

[0182] A first processing module, configured to perform exponential weighted averaging processing on preset model parameters if the number of iterations reaches a preset range of iterations;

[0183] The first iteration module is used to input the residual output result into the residual network model that completes the exponential weighted average processing of the model parameters for iteration until the time step reaches a preset time step range to obtain a target denoising function.

[0184] The motion recognition device based on a lightweight supervised diffusion model provided by the present application adopts the motion recognition method based on a lightweight supervised diffusion model in the above-mentioned embodiment, which can solve the technical problem of low accuracy of wireless human motion recognition. Compared with the prior art, the beneficial effects of the motion recognition device based on a lightweight supervised diffusion model provided by the present application are the same as the beneficial effects of the motion recognition method based on a lightweight supervised diffusion model provided by the above-mentioned embodiment, and the other technical features of the motion recognition device based on a lightweight supervised diffusion model are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0185] The present application provides an action recognition device based on a lightweight supervised diffusion model, and the action recognition device based on the lightweight supervised diffusion model includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the action recognition method based on the lightweight supervised diffusion model in the above-mentioned embodiment one.

[0186] Reference below Figure 5 , which shows a schematic diagram of the structure of a motion recognition device based on a lightweight supervised diffusion model suitable for implementing the embodiment of the present application. The motion recognition device based on a lightweight supervised diffusion model in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, tablet computers, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The action recognition device based on the lightweight supervised diffusion model shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0187] like Figure 5As shown, the action recognition device based on the lightweight supervised diffusion model may include a processing device 1001 (such as a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM: Read Only Memory) 1002 or the program loaded from the storage device 1003 to the random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the action recognition device based on the lightweight supervised diffusion model are also stored. The processing device 1001, ROM1002 and RAM1004 are connected to each other through a bus 1005. The input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the action recognition device based on the lightweight supervised diffusion model to communicate wirelessly or wired with other devices to exchange data. Although the figure shows an action recognition device based on a lightweight supervised diffusion model with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have alternatively.

[0188] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0189] The motion recognition device based on a lightweight supervised diffusion model provided by the present application adopts the motion recognition method based on a lightweight supervised diffusion model in the above-mentioned embodiment, which can solve the technical problem of low accuracy of wireless human motion recognition. Compared with the prior art, the beneficial effects of the motion recognition device based on a lightweight supervised diffusion model provided by the present application are the same as the beneficial effects of the motion recognition method based on a lightweight supervised diffusion model provided by the above-mentioned embodiment, and the other technical features in the motion recognition device based on a lightweight supervised diffusion model are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0190] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0191] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0192] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the action recognition method based on the lightweight supervised diffusion model in the above-mentioned embodiment.

[0193] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM: Random Access Memory), a read-only memory (ROM: Read Only Memory), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency: Radio Frequency), etc., or any suitable combination of the above.

[0194] The computer-readable storage medium may be included in the action recognition device based on the lightweight supervised diffusion model; or may exist independently without being assembled into the action recognition device based on the lightweight supervised diffusion model.

[0195] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by the action recognition device based on the lightweight supervised diffusion model, the action recognition device based on the lightweight supervised diffusion model:

[0196] In response to the start identification signal, acquiring wireless channel state information data;

[0197] According to a preset lightweight supervised diffusion model, the wireless channel state information data is diffused and restored to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average;

[0198] The restored channel state information data is input into a preset human action recognition model to obtain a human action recognition result.

[0199] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0200] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0201] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0202] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned motion recognition method based on a lightweight supervised diffusion model, and can solve the technical problem of low accuracy of wireless human motion recognition. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the motion recognition method based on a lightweight supervised diffusion model provided by the above-mentioned embodiment, and will not be repeated here.

[0203] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned action recognition method based on the lightweight supervised diffusion model.

[0204] The computer program product provided by the present application can solve the technical problem of low accuracy of wireless human motion recognition. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the motion recognition method based on the lightweight supervised diffusion model provided by the above embodiment, which will not be repeated here.

[0205] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. An action recognition method based on a lightweight supervised diffusion model, characterized in that: The method includes: In response to the start identification signal, acquiring wireless channel state information data; Obtain channel state information data samples and corresponding human motion category labels; Perform step encoding and linear projection on each preset time step to obtain time step encoding information; Encode the human action category label to obtain label encoding information; Mixing the label encoding information with the time step encoding information to obtain label guided time step information; Encoding the label guidance time step information through a preset linear layer to obtain a target time step encoding; The target time step encoding is used as a bias parameter and embedded into a preset residual network model, wherein the target time step encoding serves as an embedded feature of a label-guided time step to guide the residual network model to perform diffusion learning; Performing feature processing on the channel state information data sample to obtain target sample features; Inputting the target sample features into the residual network model to obtain a residual output result; Training a preset denoising function through a preset output layer network and the residual output result; Updating the time step, inputting the residual output result into the residual network model for iteration, until the time step reaches a preset time step range, and obtaining a target denoising function; Determining a lightweight supervised diffusion model according to the target denoising function; According to a preset lightweight supervised diffusion model, the wireless channel state information data is diffused and restored to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average; The restored channel state information data is input into a preset human action recognition model to obtain a human action recognition result.

2. The method according to claim 1, characterized in that The step of performing feature processing on the channel state information data sample to obtain target sample features comprises: Adding noise to the channel state information data sample according to preset Gaussian noise to obtain a diffuse noise sample; Randomly sampling from the diffuse noise samples to obtain sampled noise samples; The features of the sampled noise samples are extracted through a preset first convolutional layer and a nonlinear activation function to obtain target sample features.

3. The method according to claim 2, characterized in that The step of training the preset denoising function through the preset output layer network and the residual output result comprises: Restoring the residual output result through a deconvolution layer to obtain first restored data; Performing feature extraction on the first restored data through the nonlinear activation function and the second convolutional layer to obtain first restored data features; The preset denoising function is trained through the first restored data feature.

4. The method according to claim 1, characterized in that The step of inputting the residual output result into the residual network model for iteration until the time step reaches a preset time step range to obtain a target denoising function comprises: If the number of iterations reaches the preset range of iterations, the preset model parameters are subjected to exponential weighted averaging; The residual output result is input into the residual network model of the model parameters to complete the exponential weighted average processing and iterate until the time step reaches the preset time step range to obtain the target denoising function.

5. An action recognition device based on a lightweight supervised diffusion model, characterized in that: The device comprises: An acquisition module, configured to acquire wireless channel state information data in response to a start identification signal; A first acquisition module is used to acquire channel state information data samples and corresponding human motion category labels; A first projection module is used to perform step encoding and linear projection on each preset time step to obtain time step encoding information; The second encoding module is used to encode the human body action category label to obtain label encoding information; A first mixing module, used for mixing the label encoding information with the time step encoding information to obtain label guidance time step information; A third encoding module, used for encoding the label guidance time step information through a preset linear layer to obtain a target time step encoding; A first embedding module is used to embed the target time step encoding as a bias parameter into a preset residual network model, wherein the target time step encoding serves as an embedding feature of a label-guided time step to guide the residual network model to perform diffusion learning; A first processing module, configured to perform feature processing on the channel state information data sample to obtain target sample features; A first input module, used to input the target sample features into the residual network model to obtain a residual output result; A first training module, used for training a preset denoising function through a preset output layer network and the residual output result; A first iteration module is used to update the time step, input the residual output result into the residual network model for iteration until the time step reaches a preset time step range, and obtain a target denoising function; A first determination module is used to determine a lightweight supervised diffusion model according to the target denoising function; A diffusion module, used for performing diffusion restoration on the wireless channel state information data according to a preset lightweight supervised diffusion model to obtain restored channel state information data, wherein the training process of the lightweight supervised diffusion model includes a step of introducing sample label information and a step of exponential sliding average; The recognition module is used to input the restored channel state information data into a preset human action recognition model to obtain a human action recognition result.

6. An action recognition device based on a lightweight supervised diffusion model, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the action recognition method based on a lightweight supervised diffusion model as described in any one of claims 1 to 4.

7. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the action recognition method based on a lightweight supervised diffusion model as described in any one of claims 1 to 4 are implemented.

8. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the action recognition method based on a lightweight supervised diffusion model as claimed in any one of claims 1 to 4 are implemented.

Images