Intelligent gesture recognition method fusing CSI amplitude phase information

Abstract

The invention discloses an intelligent gesture recognition method fusing CS I amplitude and phase information, and belongs to the technical field of intelligent recognition, S10, collecting CS I data including amplitude and phase information; s20, unwinding is carried out, and CS I phase data are calibrated; s30, removing a CS I data abnormal value by using a 3 sigma criterion; s40, carrying out wavelet decomposition denoising processing on the CS I data; s50, performing spline interpolation processing on the CS I data, and unifying the format; s60, processing the CS I data, and fusing amplitude and phase features; s70, a ResNet-18 model is trained, and gestures are classified; the method has the advantages that by effectively fusing the amplitude and phase information of the CS I and combining wavelet decomposition, spline interpolation and deep learning technologies, more comprehensive feature representation can be obtained, the accuracy of gesture recognition can be improved, and therefore the requirement for intelligent gesture recognition in a complex environment can be better met.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition technology, and more particularly, to an intelligent gesture recognition method integrating CSI amplitude and phase information. Background Art

[0002] With the rapid development of technology, human gesture recognition technology has demonstrated its unique application value in many fields, such as smart home control, human-computer interaction, and healthcare monitoring. However, existing human gesture recognition solutions still face numerous challenges in practical application, particularly in terms of privacy protection, ease of use, and recognition accuracy. Computer vision technology, a mainstream approach to gesture recognition, achieves relatively accurate gesture recognition by capturing human gesture images with a camera and performing processing and analysis. However, this approach has significant limitations in terms of privacy protection. In certain situations, such as private spaces or where personal privacy is a concern, the use of computer vision for gesture recognition may infringe on users' privacy rights, thus limiting its application.

[0003] To overcome the privacy concerns of computer vision technology, contact sensors have emerged. This approach uses a sensor device worn by the user to capture gesture information, thereby enabling gesture recognition. However, contact sensors require wearable devices, which can cause discomfort and restrict user freedom of movement. Furthermore, wearable devices may not be user-friendly or suitable for certain populations (such as the elderly and people with disabilities) and specific scenarios (such as medical surgery and emergency rescue).

[0004] In recent years, channel state information (CSI) has become a key concept in wireless communications, describing the channel properties of a communication link. Specifically, CSI technology has gained increasing attention in the field of gesture recognition due to its contactless and privacy-preserving features. CSI contains detailed information, such as the amplitude and phase of wireless signals during propagation, and can reflect the impact of gestures on the wireless signal propagation path, thus providing new insights for gesture recognition. However, most existing CSI-based gesture recognition methods utilize only CSI amplitude or phase information, failing to fully tap the potential of CSI. Some studies have utilized only CSI amplitude information and proposed algorithms based on wavelet transforms and short-term energy for gesture recognition. While these have achieved some success, they have neglected the use of phase information. Other studies have attempted to use CSI phase differences for gesture recognition. While these introduce novel features, they still fail to integrate both amplitude and phase information. Furthermore, some studies have employed deep learning methods, such as LSTM models, for training, but accuracy still needs to be improved. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, an intelligent gesture recognition method that integrates CSI amplitude and phase information is proposed. By effectively fusing the amplitude and phase information of CSI and combining wavelet decomposition, spline interpolation, and deep learning technology, a more comprehensive feature representation can be obtained, which helps to improve the accuracy of gesture recognition and better adapt to the needs of intelligent gesture recognition in complex environments.

[0006] The technical solution adopted by the invention to solve the technical problem is: an intelligent gesture recognition method integrating CSI amplitude and phase information, the improvement of which includes:

[0007] S10: Using a multi-antenna device access point and a multi-antenna mobile terminal, collect CSI data containing amplitude and phase information;

[0008] S20: performing deconvolution processing on the collected CSI phase data to eliminate phase offset and noise, and calibrating the phase information through linear transformation;

[0009] S30: Use the 3σ criterion to remove outliers in CSI data;

[0010] S40: De-noising the CSI data using a wavelet decomposition method;

[0011] S50: performing spline interpolation on the collected CSI data to unify the data format;

[0012] S60: Process the amplitude and phase data of CSI separately through the multi-scale channel attention module and splice them to form fused feature data;

[0013] S70: Use the ResNet-18 deep neural network model to train the fused feature data, identify gestures of different categories, and obtain gesture classification results.

[0014] Furthermore, the specific steps of step S10 are:

[0015] S101: In a daily office environment, a multi-antenna access point is used as the transmitter and a multi-antenna mobile terminal is used as the receiver.

[0016] S102: Setting up an experimental scenario, having the test subject perform different gestures within the scenario, and collecting CSI data corresponding to each gesture. The CSI data includes the channel frequency response of multiple orthogonal frequency division multiplexing subcarriers, consisting of amplitude information and phase information.

[0017] S103: Collect the channel frequency response of each pair of receiving antennas and subcarriers and record it as a CSI data matrix.

[0018] Furthermore, the specific steps of step S20 are:

[0019] S201: Unwrap the subcarrier phase information of the receiving antenna to eliminate cross-cycle phase changes;

[0020] S202: Eliminate the time offset and unknown phase offset between the transmitter and the receiver using a linear fitting method;

[0021] S203: Subtract the linear fitting part from the unwrapped CSI phase to obtain calibrated CSI phase information for subsequent analysis.

[0022] Furthermore, in step S20, the formula for calibrating the CSI phase is:

[0023]

[0024] in, represents the true phase, the CSI phase measured in the yth subcarrier of the xth pair of transmit and receive antennas; kn represents the highest subcarrier index, k1 represents the lowest subcarrier index; k y represents the subcarrier index varying from -28 to 28 in IEEE802.11n, δ(t) is the time offset between the transmitter and the receiver, β(t) is the unknown phase offset, and Z(t) is the noise introduced by the measurement process; Represents the calibrated CSI phase of the xth transmit and receive antenna in the yth carrier at time t.

[0025] Furthermore, the specific steps of step S40 are:

[0026] S401: Decompose the CSI signal into three layers using the db4 wavelet function to obtain approximate coefficients and detail coefficients. The approximate coefficients describe the overall shape of the CSI waveform, while the detail coefficients capture noise and subtle changes.

[0027] S402: Calculate the standard deviation and data length of the CSI data sample and set a threshold; apply a hard threshold function to the approximation coefficient and detail coefficient to suppress the noise coefficient and retain the useful signal coefficient;

[0028] S403: Recombining the approximate coefficients and detail coefficients after hard threshold processing, and using a wavelet reconstruction method to obtain a CSI signal after noise reduction.

[0029] Furthermore, the specific steps of step S50 are:

[0030] S501: Fit each pair of adjacent data points in the CSI data set with a low-order polynomial;

[0031] S502: Ensure that the interpolation function is continuous at each data point, and that the first-order derivative and the second-order derivative at the point are continuous, thereby ensuring that the function values, derivatives, and second-order derivatives of the cubic polynomials in adjacent intervals are consistent at the nodes;

[0032] S503: setting natural boundary conditions, fixed boundary conditions, and periodic boundary conditions, and constructing a linear equation system based on the boundary conditions;

[0033] S504: Solve the linear equations to obtain the coefficients of the polynomials in each interval and complete the spline interpolation.

[0034] Furthermore, the step S60 includes the following steps:

[0035] S601: Perform global average pooling on the CSI amplitude data, adjust the channel scale through point convolution and batch normalization, use the ReLU activation function for nonlinear transformation, and perform batch normalization and point convolution again to obtain global features;

[0036] S602: Extract local features from the CSI amplitude data and obtain local feature representation through point convolution and activation function;

[0037] S603: The global features and local features are batch normalized and then fused through an addition operation. After the addition, the two are processed through a Sigmoid activation function. The processed features are multiplied with the original CSI amplitude data to obtain feature data composed of global and local features.

[0038] Furthermore, the threshold value in step S402 is set as:

[0039]

[0040] Among them, m l is the length of the CSI data sample, and thr is the corresponding threshold.

[0041] Furthermore, the cubic polynomial fitting in step S502 satisfies the following conditions:

[0042] S i (x) = a i (xx i ) 3 +b i (xx i ) 2 +c i (xx i )+d;

[0043] Among them, a, b, c, and d are the polynomial coefficients of each interval.

[0044] The beneficial effect of the present invention is that by effectively fusing the amplitude and phase information of CSI, combined with wavelet decomposition, spline interpolation and deep learning technology, a more comprehensive feature representation can be obtained, which helps to improve the accuracy of gesture recognition, thereby better adapting to the needs of intelligent gesture recognition in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of an intelligent gesture recognition method that integrates CSI amplitude and phase information according to the present invention;

[0046] Figure 2 This is a CSI data set splicing flow chart of an intelligent gesture recognition method integrating CSI amplitude and phase information according to the present invention;

[0047] Figure 3 A flowchart of fine processing of CSI amplitude and phase information in an intelligent gesture recognition method integrating CSI amplitude and phase information of the present invention;

[0048] Figure 4 This is an experimental effect diagram of an intelligent gesture recognition method that integrates CSI amplitude and phase information in the present invention;

[0049] Figure 5 This is an experimental effect diagram of the present invention's intelligent gesture recognition method that integrates CSI amplitude and phase information using only amplitude;

[0050] Figure 6 This is an experimental effect diagram of the present invention's intelligent gesture recognition method that integrates CSI amplitude and phase information and uses only phase information;

[0051] Figure 7 This is a diagram showing the prediction accuracy of each gesture in an intelligent gesture recognition method that integrates CSI amplitude and phase information according to the present invention; DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to the accompanying drawings and examples.

[0053] The following will clearly and completely describe the concept, specific structure and technical effects of the present invention in combination with the embodiments and drawings, so as to fully understand the purpose, characteristics and effects of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention. In addition, all the connection / connection relationships involved in the patent do not refer to the direct connection of components, but refer to the fact that a better connection structure can be formed by adding or reducing connection accessories according to the specific implementation situation. The various technical features in the invention can be combined interactively without conflicting with each other.

[0054] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.

[0055] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. Secondly, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0056] See also Figure 1-Figure 2 As shown, the present invention provides an intelligent gesture recognition method integrating CSI amplitude and phase information, including:

[0057] S10: Collect and sample CSI data in a daily environment using multi-antenna device access points and multi-antenna mobile terminals to obtain a CSI dataset containing amplitude and phase information.

[0058] S20: performing deconvolution processing on the collected CSI phase data to eliminate phase offset and noise, and calibrating the phase information through linear transformation;

[0059] S30: Use the 3σ criterion to remove outliers from the CSI data and replace data that deviate from the mean by more than three standard deviations with the CSI data mean;

[0060] S40: De-noising the CSI data using a wavelet decomposition method, and processing the wavelet coefficients using a hard threshold function;

[0061] S50: performing spline interpolation processing on the collected CSI data, and performing unified format processing through spline interpolation to ensure consistency of the number of data packets;

[0062] S60: The amplitude and phase data of the CSI are processed separately through the multi-scale channel attention module to extract global features and local features, and the amplitude and phase information are spliced ​​to form fused feature data;

[0063] S70: Use the ResNet-18 deep neural network model to train the fused data, identify gestures of different categories, and obtain gesture classification results.

[0064] In this invention, spline interpolation ensures the consistency and continuity of data packets, avoiding information loss or deviation caused by missing data or irregular intervals. This is particularly important for subsequent feature extraction and model training, ensuring that each data point has the same dimensions and time sequence. A multi-scale channel attention module is used to extract global and local features of amplitude and phase data, respectively, enabling a more comprehensive understanding of the different levels of signal information. This feature extraction method enables precise fusion of amplitude and phase information, thereby improving model performance. Training using the ResNet-18 deep neural network allows for efficient processing and recognition of gestures of various categories. Due to its residual structure, the ResNet network effectively avoids the vanishing gradient problem in deep networks, improving the accuracy and robustness of gesture classification. Through this series of data preprocessing and feature fusion methods, the final model is able to extract high-quality features from rich and processed data, significantly improving the accuracy of gesture recognition. Utilizing multi-antenna devices and deep neural networks, it is not only capable of processing complex CSI data but also accurately recognizing a variety of gestures in real-world environments, demonstrating strong adaptability and generalization capabilities.

[0065] Furthermore, the specific steps of step S10 are:

[0066] S101: In a daily office environment, a multi-antenna access point is used as the transmitter and a multi-antenna mobile terminal is used as the receiver.

[0067] S102: Setting up an experimental scenario, having the test subject perform different gestures within the scenario, and collecting CSI data corresponding to each gesture. The CSI data includes the channel frequency response of multiple orthogonal frequency division multiplexing subcarriers, consisting of amplitude information and phase information.

[0068] S103: Collect the channel frequency response of each pair of receiving antennas and subcarriers and record it as a CSI data matrix.

[0069] In this embodiment, the multi-antenna device used is a WiFi device. The experiment is carried out in a daily office environment. People make different gestures in this environment. The collected data is divided into six categories according to the different gestures, namely "pushing and pulling the hand forward", "clapping", "waving upward", "waving downward", "waving to the left" and "waving to the right".

[0070] Orthogonal Frequency Division Multiplexing (OFDM) is a multicarrier modulation technique with strong resistance to multipath fading, making it suitable for multipath environments. Due to the orthogonality of the subcarriers, OFDM eliminates the need for guard bands between subcarriers, thereby improving spectrum efficiency. The channel frequency response (CFR) describes how the frequency content of a signal changes from the transmitter to the receiver over a specific transmission medium.

[0071] This Wi-Fi uses Orthogonal Frequency Division Multiplexing (OFDM), which means the actual transmission channel consists of multiple orthogonal communication subchannels with different frequency bands (i.e., subcarriers). The CSI is the channel frequency response (CFR) between each pair of transmit and receive antennas, covering 30 OFDM subcarriers.

[0072] The receiving end uses Intel5300 as three receiving antennas, and the collected data matrix is ​​3x30, that is:

[0073]

[0074] H(t) is the channel state information at time t, H x,y (t) is the channel state information of the y-th subcarrier received by the x-th receiving antenna at time t, H x,y (t) can be rewritten as

[0075]

[0076] Where X x,y (t) and Y x,y (t) are the transmitted and received signals, A x,y (t) and H x,y (t) is the magnitude and phase of the

[0077] Furthermore, the specific steps of step S20 are:

[0078] S201: Unwrap the subcarrier phase information of the receiving antenna to eliminate cross-cycle phase changes;

[0079] S202: Eliminate the time offset and unknown phase offset between the transmitter and the receiver using a linear fitting method;

[0080] S203: Subtract the linear fitting part from the unwrapped CSI phase to obtain calibrated CSI phase information for subsequent analysis.

[0081] Since the phase data of a signal varies within a specific range (usually -π to π), when the phase of the signal crosses this range, a "phase jump" or "blurring" occurs, which makes it impossible to accurately reflect the actual changes in the signal. The unwrapping process can eliminate this cross-cycle phase change, thereby restoring the true phase information. By unwrapping, the problem of periodic phase jumps is eliminated, making the phase information more coherent and easier to process and analyze later. In actual experiments, there may be a slight time deviation between the transmitter and the receiver, which will cause the signal to shift in time at the receiver, affecting the synchronization of the data. The use of linear fitting methods can effectively identify and correct this time offset, ensuring the temporal consistency of the signal.

[0082] Furthermore, the measured channel state information (CSI) phase in the yth subcarrier of the xth pair of transmit and receive antennas is given by the following formula

[0083]

[0084] is the true phase, δ(t) is the time offset between the transmitter and the receiver, β(t) is the unknown phase offset, Z(t) is the noise introduced by the measurement process, and k y is the subcarrier index that varies from -28 to 28 in IEEE 802.11n, and N is the number of (Fast Fourier Transform) window points. Basically, the two components δ(t) and β(t) can be eliminated by linear transformation. Therefore, we define two variables a x (t) and b x (t) are as follows:

[0085]

[0086] Where n is the total number of subcarriers. According to the IEEE 802.11n protocol, the frequencies of the subcarriers are completely symmetrical, that is, Therefore, b x (t) is rewritten as

[0087]

[0088] From the measured phase Subtract the linear expression a from x (t)k y +b x (t), to obtain the calibration phase information of the y-th subcarrier As shown below

[0089]

[0090] represents the true phase, the CSI phase measured in the yth subcarrier of the xth pair of transmit and receive antennas; k n Indicates the highest subcarrier index; k1 indicates the lowest subcarrier index; k y Indicates the subcarrier index ranging from -28 to 28 in IEEE 802.11n.

[0091] The 3σ criterion (also known as the "three sigma criterion" or the "3 standard deviation rule") is a method used in statistics to determine the range of data distribution. Its main application is to determine whether the data is within the normal range by performing standard deviation analysis on the data. Specifically, the 3σ criterion describes the probability that a data point is within 3 standard deviations from the mean in a normal distribution. In the collected CSI, there are always some abnormal data that are much larger or smaller than the mean and standard deviation of the CSI samples. In general, the 3σ criterion is an effective way to delete and replace abnormal data, which is defined as a value that deviates from the mean by more than three times the standard deviation. When abnormal data is found, it needs to be replaced with the mean of the CSI sample. If the time series window size is set to 9, the mean μ of the time series CSI sample of the yth subcarrier in the xth pair of transmit and receive antennas at time t under the current window can be calculated x,y (t) and standard deviation σ X,Y (t), as shown below

[0092]

[0093] A x,y (t) is the channel state information (CSI) amplitude or phase of the yth subcarrier in the xth pair of transmit and receive antennas at time t. W is the time series window size. Based on the 3σ criterion, the time series window size is set to 7. If A x,y (t)∈[μ x,y (t)-3σ x,y (t),μ x,y (t)+3σ x,y (t)], the current data is regarded as abnormal data, and then the abnormal data is marked and replaced by μ x,y (t).

[0094] The 3σ criterion is used to remove outliers from CSI data. Data that deviates from the mean by more than three standard deviations is replaced with the CSI mean. By calculating the mean and standard deviation of the data, outliers that deviate from the normal range can be identified and replaced with the mean, thereby improving data accuracy.

[0095] Furthermore, the specific steps of step S40 are:

[0096] S401: Decompose the CSI signal into three layers using the db4 wavelet function to obtain approximate coefficients and detail coefficients. The approximate coefficients describe the overall shape of the CSI waveform, while the detail coefficients capture noise and subtle changes.

[0097] S402: Calculate the standard deviation and data length of the CSI data sample and set a threshold; apply a hard threshold function to the approximation coefficient and detail coefficient to suppress the noise coefficient and retain the useful signal coefficient;

[0098] S403: Recombining the approximate coefficients and detail coefficients after hard threshold processing, and using a wavelet reconstruction method to obtain a CSI signal after noise reduction.

[0099] Through wavelet decomposition, the approximate coefficients reflect the primary patterns and morphology of the CSI signal, while the detail coefficients capture the signal's noise and high-frequency components. The db4 wavelet function is a commonly used wavelet basis that effectively processes signals with varying frequency components. It exhibits excellent time-frequency localization, accurately capturing the local characteristics and variations of the CSI signal. By calculating the standard deviation and setting a threshold, it is possible to distinguish between signal and noise. After wavelet transform, the detail coefficients typically reflect high-frequency noise, while the approximate coefficients reflect the primary structure of the signal. Applying a hard threshold function effectively removes the noise coefficients and preserves important signal information. This reduces noise interference and improves signal quality. Using wavelet reconstruction, the approximate and detail coefficients, after hard thresholding, are recombined to produce a smoother, denoised CSI signal. This process helps restore the signal's essential characteristics and eliminates the effects of noise, providing more accurate and reliable input data for subsequent analysis.

[0100] For background noise, in theory, a low-pass filter (i.e., Butterworth filter) can eliminate high-frequency noise. However, traditional low-pass filters are not effective for removing burst and impulse noise, and strict low-pass filters will lead to the loss of useful signals. Instead, we use a threshold denoising method based on wavelet decomposition. This method can effectively protect the peaks of the expected signal and the burst signal, remove burst and impulse noise, and suppress the interference of high-frequency noise. Wavelet decomposition can decompose the channel state information (CSI) into two terms: approximate coefficients and detail coefficients. The former describes the shape of the CSI waveform, and the latter captures the noise and subtle details of the CSI. In the wavelet domain, the effective CSI corresponds to large coefficients, while the noise corresponds to small coefficients and satisfies the Gaussian distribution. Therefore, a threshold is determined in advance so that the coefficients in a certain interval in the wavelet domain are set to zero, thereby suppressing high-frequency noise. We use the WAVEDEC function provided by the MATLAB wavelet toolbox to perform signal decomposition. In this paper, we use the db4 wavelet function to decompose the CSI into three layers to obtain three detail coefficient vectors and one approximate coefficient vector. At the same time, the threshold is set to

[0101]

[0102] where m l is the length of the CSI data sample, and thr is the corresponding threshold. The detail coefficient and the approximation coefficient are processed by the hard threshold function, which is expressed as

[0103]

[0104] Where w is the original wavelet coefficient, w thr is the wavelet coefficient after hard thresholding. The wavelet coefficients after hard thresholding are recombined and reconstructed to obtain the denoised CSI.

[0105] Furthermore, the specific steps of step S50 are:

[0106] S501: For each pair of adjacent data points [x i ,x i+1 ], a low-order polynomial is used to fit between them;

[0107] S502: Ensure that the interpolation function is continuous at each data point, and that the first-order derivative and the second-order derivative at the point are continuous, thereby ensuring that the function values, derivatives, and second-order derivatives of the cubic polynomials in adjacent intervals are consistent at the nodes;

[0108] S503: setting natural boundary conditions, fixed boundary conditions, and periodic boundary conditions, and constructing a linear equation system based on the boundary conditions;

[0109] S504: Solve the linear equations to obtain the coefficients of the polynomials in each interval and complete the spline interpolation.

[0110] The cubic polynomial fitting in step S502 satisfies the following conditions:

[0111] S i (x) = a i (xx i ) 3 +b i (xx i ) 2 +c i (xx i )+d;

[0112] Among them, a, b, c, and d are the polynomial coefficients of each interval.

[0113] The polynomial coefficients a, b, c, d satisfy the relationship

[0114] S i (xi )=y i S I (x i+1 )=y i+1 S′ i (x i )=S′ i-1 (xi) S″ i (x) = S″ i-1 (x)

[0115] By fitting adjacent data points with a low-order polynomial, the transition between data points can be ensured to be smooth, avoiding abrupt changes or discontinuities. This is particularly important for CSI signals, as such signals typically contain some continuous changes and relatively smooth waveforms. Next, the interpolation function is ensured to be continuous not only in terms of function value at the data points, but also in terms of first-order derivative (i.e., slope) and second-order derivative (i.e., curvature). This means that the transition at the data points is more natural and smooth. This is particularly important for CSI signals, as signal changes are often continuous and require smooth transitions between each data point. The various boundary conditions mentioned above (such as natural boundary conditions, fixed boundary conditions, and periodic boundary conditions) can be flexibly selected according to different application scenarios to ensure that the interpolation process remains reasonable and physically consistent at both ends of the signal (i.e., at the boundaries). Natural boundary conditions: Generally, the second-order derivative of the signal at the boundary is zero, which is applicable to situations without obvious external constraints.

[0116] Fixed boundary conditions: Applicable to situations where the boundary values ​​are known, which can ensure that the signal values ​​at the boundaries are not modified.

[0117] Periodic boundary conditions: Applicable to scenarios where the signal needs to change periodically, ensuring that the start and end of the signal can be smoothly connected.

[0118] Reasonable setting of boundary conditions helps to more accurately control the range of data fitting, avoid the influence of boundary effects on the results, and make the interpolation results more consistent with the expectations of the actual signal.

[0119] Furthermore, the step S60 includes the following steps:

[0120] S601: Perform global average pooling on the CSI amplitude data, adjust the channel scale through point convolution and batch normalization, use the ReLU activation function for nonlinear transformation, and perform batch normalization and point convolution again to obtain global features;

[0121] S602: Extract local features from the CSI amplitude data and obtain local feature representation through point convolution and activation function;

[0122] S603: The global features and local features are batch normalized and then fused through an addition operation. After the addition, the two are processed through a Sigmoid activation function. The processed features are multiplied with the original CSI amplitude data to obtain feature data composed of global and local features.

[0123] like Figure 3 As shown, we use the MS-CAM channel attention module for amplitude and phase respectively. Taking amplitude as an example, there are three main routes. The first route is to perform global average pooling on the CSI amplitude data 3x30x300. At this time, the output data format is 3x1x1. In order to pay attention to the scale of the channel, point convolution is used to reduce the parameter size and computational complexity. After batch normalization, it is activated by the ReLu function and then batch normalization, and then a point convolution is performed to obtain global features. The second route is to perform point convolution on the CSI amplitude data, and then pass it through the activation function and a point convolution to obtain local features. The third route is to batch normalize the global features of the first output and the local features of the second output, add them together, pass the Sigmund function, and then multiply the CSI amplitude data with the output of the Sigmund activation function to obtain the output features, which are composed of global + local features.

[0124] The dataset consists of six categories: "Hand Pushing and Pulling Forward," "Applauding," "Waving Upward," "Waving Downward," "Waving Leftward," and "Waving Rightward." Each category contains 500 data points, for a total of 3,000 data points, divided into a training set and a test set with a 70 / 30 split. The CSI amplitude and phase are first passed through the MS-CAM channel attention module. The CSI amplitude and phase are then concatenated into 6x30x300 fused data, which is then trained using ResNet-18. ResNet-18 has strong feature extraction capabilities. It can extract rich global and local features from the input data. This is crucial for gesture recognition, as different gesture types often contain subtle local variations as well as macroscopic global patterns, which ResNet-18 can effectively capture.

[0125] like Figure 4-Figure 7 As shown in the experimental results of this solution, we can see that in this daily office environment, people make 6 different gestures in the environment, and then collect CSI data. These data are trained using the method of this patent. Figure 7 The prediction accuracy is very high, close to 100% at the highest, combined with the following two separate CSI amplitudes ( Figure 5 ) or the experimental effect diagram of phase training ( Figure 6), it is not difficult to find that the method proposed in this application is more effective than the traditional method of using the amplitude and phase of CSI.

[0126] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. An intelligent gesture recognition method integrating CSI amplitude and phase information, characterized in that: include: S10: Using a multi-antenna device access point and a multi-antenna mobile terminal, collect CSI data containing amplitude and phase information; S20: performing deconvolution processing on the collected CSI phase data to eliminate phase offset and noise, and calibrating the phase information through linear transformation; S30: Use the 3σ criterion to remove outliers in CSI data; S40: De-noising the CSI data using a wavelet decomposition method; S50: performing spline interpolation on the collected CSI data to unify the data format; S60: Process the amplitude and phase data of CSI separately through the multi-scale channel attention module and splice them to form fused feature data; S70: Use the ResNet-18 deep neural network model to train the fused feature data, identify gestures of different categories, and obtain gesture classification results.

2. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 1 is characterized in that: The specific steps of step S10 are: S101: In a daily office environment, a multi-antenna access point is used as the transmitter and a multi-antenna mobile terminal is used as the receiver. S102: Setting up an experimental scenario, having the test subject perform different gestures within the scenario, and collecting CSI data corresponding to each gesture. The CSI data includes the channel frequency response of multiple orthogonal frequency division multiplexing subcarriers, consisting of amplitude information and phase information. S103: Collect the channel frequency response of each pair of receiving antennas and subcarriers and record it as a CSI data matrix.

3. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 1, characterized in that: The specific steps of step S20 are: S201: Unwrap the subcarrier phase information of the receiving antenna to eliminate cross-cycle phase changes; S202: Eliminate the time offset and unknown phase offset between the transmitter and the receiver using a linear fitting method; S203: Subtract the linear fitting part from the unwrapped CSI phase to obtain calibrated CSI phase information for subsequent analysis.

4. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 1, characterized in that: In step S20, the formula for calibrating the CSI phase is: in, represents the true phase, the CSI phase measured in the yth subcarrier of the xth pair of transmit and receive antennas; kn represents the highest subcarrier index, k1 represents the lowest subcarrier index; k y represents the subcarrier index varying from -28 to 28 in IEEE802.11n, δ(t) is the time offset between the transmitter and the receiver, β(t) is the unknown phase offset, and Z(t) is the noise introduced by the measurement process; Represents the calibrated CSI phase of the xth transmit and receive antenna in the yth carrier at time t.

5. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 1 is characterized in that: The specific steps of step S40 are: S401: Decompose the CSI signal into three layers using the db4 wavelet function to obtain approximate coefficients and detail coefficients. The approximate coefficients describe the overall shape of the CSI waveform, while the detail coefficients capture noise and subtle changes. S402: Calculate the standard deviation and data length of the CSI data sample and set a threshold; Apply hard threshold function to the approximation coefficient and detail coefficient to suppress the noise coefficient and retain the useful signal coefficient; S403: Recombining the approximate coefficients and detail coefficients after hard threshold processing, and using a wavelet reconstruction method to obtain a CSI signal after noise reduction.

6. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 5 is characterized in that: The specific steps of step S50 are: S501: Fit each pair of adjacent data points in the CSI data set with a low-order polynomial; S502: Ensure that the interpolation function is continuous at each data point, and that the first-order derivative and the second-order derivative at the point are continuous, thereby ensuring that the function values, derivatives, and second-order derivatives of the cubic polynomials in adjacent intervals are consistent at the nodes; S503: setting natural boundary conditions, fixed boundary conditions, and periodic boundary conditions, and constructing a linear equation system based on the boundary conditions; S504: Solve the linear equations to obtain the coefficients of the polynomials in each interval and complete the spline interpolation.

7. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 6, characterized in that: The step S60 includes the following steps: S601: Perform global average pooling on the CSI amplitude data, adjust the channel scale through point convolution and batch normalization, use the ReLU activation function for nonlinear transformation, and perform batch normalization and point convolution again to obtain global features; S602: Extract local features from the CSI amplitude data and obtain local feature representation through point convolution and activation function; S603: The global features and local features are batch normalized and then fused through an addition operation. After the addition, the two are processed through a Sigmoid activation function. The processed features are multiplied with the original CSI amplitude data to obtain feature data composed of global and local features.

8. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 5, characterized in that: The threshold value in step S402 is set as: Among them, m l is the length of the CSI data sample, and thr is the corresponding threshold.

9. The intelligent gesture recognition method integrating CSI amplitude and phase information according to claim 5, characterized in that: The cubic polynomial fitting in step S502 satisfies the following conditions: S i (x)=a i (x-x i ) 3 +b i (x-x i ) 2 +c i (x-x i )+d; Among them, a, b, c, and d are the polynomial coefficients of each interval.

Images