Method, System and Storage Medium for Human Activity Sensing and Identification Based on Wi-Fi Signals

Abstract

The present invention relates to a method for human activity sensing and recognition based on Wi-Fi signals, comprising the following steps: determining whether high-precision requirements are met. If so, collecting channel state information data; if not, collecting orthogonal frequency division multiplexing symbol stream data; performing principal component analysis on the subcarrier domain of the channel state information data to obtain rough features, extracting the minimum value of cross-correlation data to obtain correlation artificial features, extracting amplitude data to obtain energy spectrum artificial features, and obtaining channel parameter features based on the spatially alternating generalized expectation maximization algorithm; processing the orthogonal frequency division multiplexing symbol stream data to obtain rough features; performing principal component analysis on the rough features to obtain fine features; inputting the sample types and feature sets of the training set samples into a support vector machine for classification and recognition training; and inputting the feature set of the sample to be tested into the trained support vector machine to obtain a classification and recognition result. Compared with the prior art, the present invention has the advantages of low complexity, high recognition accuracy, etc.

Description

Technical Field

[0001] The present invention relates to the fields of signal processing and intelligent perception, and in particular to a method, system and storage medium for human activity sensing and recognition based on Wi-Fi signals. Background Art

[0002] Indoor sensing technology has a wide range of applications. At present, its technical implementation paths are generally divided into two categories: The first category is active sensing, which often requires special devices to generate detection signals. For example, infrared technology: uses infrared rays to image the human body, is not restricted by light, but requires expensive additional devices and has great use limitations; dedicated sensor technology: uses intelligent sensors such as accelerometers and acoustic sensors to monitor the human body, with high sensing accuracy but high cost and inconvenient deployment; The second category is passive sensing, which does not rely on special devices. For example, computer vision technology: through image processing and then classification and recognition by means of machine learning, it is widely used but restricted by lighting conditions; environmental perception based on wireless signals, mainly Wi-Fi signals, has the advantages of wide deployment, device independence, non-line-of-sight perception, strong scalability, and good cost-effectiveness, and has received increasing attention.

[0003] In the current research on environmental perception technology based on Wi-Fi signals, the human activity sensing and recognition system based on channel state information and deep learning has become the main direction.

[0004] The literature "Understanding Wi-Fi Signal Frequency Features for Position-Independent Gesture Sensing" (K. Niu, F. Zhang, X. Wang, Q. Lv, H. Luo and D. Zhang. IEEE Transactions on Mobile Computing, 2021.) proposed an identification model based on channel state information, which extracts motion fragments and relative motion direction change information through phase calibration, smoothing filtering, and continuous wavelet transform, and the accuracy rate reaches 96% when identifying eight gestures.

[0005] The literature "Joint Activity Recognition and Indoor Localization with Wi-Fi Fingerprints" (F. Wang, J. Feng, Y. Zhao, X. Zhang, S. Zhang and J. Han. IEEE Access, 2019.) proposed a method based on channel state information amplitude position fingerprints and a one-dimensional dual-task convolutional neural network to recognize and classify six arm movements with an identification accuracy of 88.13%.

[0006] The literature "Human Behavior Recognition Based on Wi-Fi Channel State Information" (Z. Tang, A. Zhu, Z. Wang, K. Jiang, Y. Li and F. Hu. Chinese Automation Congress (CAC), 2020.) proposed an identification system that applies gated recurrent units and convolutional neural networks. Using the amplitude of the channel state information as a signal feature, it classifies seven human behaviors with an accuracy of 95.7%. It extracts images of the channel state information amplitude as training data, but the system complexity is high and it requires a long training time.

[0007] The literature "Human Behavior Recognition Method Based on Wi-Fi Channel Status Information" (Y. Zhou, Z. Cui, X. Lu, H. Wang, C. Sheng and Z. Zhang. 40th Chinese Control Conference (CCC), 2021.) proposed a method using support vector machines and discrete wavelet transform. By using the discrete wavelet transform information as the input of the support vector machine, it extracts features from the amplitude and phase of the channel state information and classifies seven human behaviors, achieving an accuracy of 93%, and fully utilizes the channel state information.

[0008] The literature "Wi-Fi-Based Activity Recognition using Activity Filter and Enhanced Correlation with Deep Learning" (Z. Shi, J. A. Zhang, R. Y. Xu and Q. Cheng. IEEE International Conference on Communications Workshops, 2020.) proposed a long short-term memory (LSTM)-recurrent neural network (RNN) classifier based on the SoftMax method. It was trained according to the activity-related information extracted from the channel state information and used to classify six behavior types with an accuracy of 93.4%. This method includes channel state information compensation and enhancement, showing good overall performance.

[0009] The above literature mainly relies on channel state information data and deep learning perception and recognition models, which has the problem of high complexity in practical applications. Summary of the Invention

[0010] The purpose of the present invention is to provide a Wi-Fi signal-based human activity perception and recognition method, system and storage medium with low complexity.

[0011] The purpose of the present invention can be achieved through the following technical solutions:

[0012] A Wi-Fi signal-based human activity perception and recognition method includes the following steps:

[0013] Judge whether the accuracy requirement of the application scenario for the recognition result is high precision.

[0014] If so, collect the channel state information data of the Wi-Fi signal corresponding to each sample.

[0015] If not, collect the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to each sample.

[0016] Wherein, the samples include training set samples and samples to be tested.

[0017] Perform principal component analysis on the subcarrier domain of the channel state information data, retain the first n principal components, and perform one-dimensional processing on the first n principal component information of the subcarrier domain to extract the amplitude and phase information of the processed data to obtain the first rough features of the samples, where n is the pre-configured number of retained principal components.

[0018] Extract the minimum values of the time-domain cross-correlation data of the amplitudes of different subcarriers of the channel state information data to obtain correlation artificial features, and include them in the feature set of the sample; perform a short-time Fourier transform on the amplitude data of the channel state information data to obtain the energy spectral density, accumulate the energy in the low-frequency interval, obtain the energy impact curve, extract the peak value of the energy impact curve to obtain energy spectral artificial features, and include them in the feature set of the sample;

[0019] Divide the collected channel state information data into multiple segments in the time domain, apply the spatially alternating generalized expectation maximization algorithm to each segment to obtain the estimation of the multi-path signal parameters of the propagation channel in the time domain, extract the variance of the time-domain Doppler parameters to obtain channel parameter features, and include them in the feature set of the sample;

[0020] For the collected orthogonal frequency division multiplexing symbol stream data, extract the data of a pre-configured number of sampling points, and count the amplitude and phase information of the data to obtain the second rough feature;

[0021] Perform principal component analysis on the first rough feature or the second rough feature, select the principal components with the cumulative variance greater than the pre-configured cumulative feature ratio threshold to obtain the first fine feature or the second fine feature respectively, and include them in the feature set of the sample;

[0022] Input the sample type and feature set of the training set samples into a support vector machine for classification and recognition training;

[0023] Input the feature set of the sample to be tested into the trained support vector machine for classification and recognition to obtain the classification and recognition result.

[0024] The Wi-Fi signal is transmitted from the transmitter and received by the receiver after passing through the sample space, where the sample space is the human activity scene space.

[0025] The principal component analysis is as follows:

[0026]

[0027] Y = XW

[0028] where W=(w1,w2,…,wq) is composed of the eigenvectors (w1,w2,…,w q ) corresponding to the first q largest eigenvalues (λ1,λ2,…,λ q ) of the covariance matrix S. q is the number of principal components to be retained. Y is the data after reducing the original data from high dimension to low dimension, X is the original data, and N is the dimension of the original data.

[0029] The first rough feature includes the amplitude histogram feature, and the second rough feature includes the amplitude histogram feature, the phase histogram feature, and the fast Fourier transform amplitude feature.

[0030] The channel state information data is divided into multiple segments in the time domain based on the channel state information data, the packet sampling frequency, the signal center frequency, the bandwidth, and the number of frequency points.

[0031] The spatial alternating generalized expectation maximization algorithm is as follows:

[0032] For the given observation h(m), output the maximum likelihood estimate of the multi-dimensional multipath signal parameter Θ, where the parameter to be estimated is Θ = {θ1, θ1,......, θ l} τ l , φ l and are the time of flight, the angle of arrival of the unit direction vector, and the Doppler frequency shift of the l-th path, and α l is the complex amplitude of the l-th path. The log-likelihood function of Θ is:

[0033]

[0034] where L is the total number of multipath components, and S l is the signal of the l-th path;

[0035] The maximum likelihood estimation problem to be solved is:

[0036]

[0037] In the expectation E step, for the l-th path, there is an expectation function:

[0038]

[0039] where is the parameter estimated in the previous iteration;

[0040] In the maximization M step, for the l-th path, there is a series of maximization functions:

[0041]

[0042]

[0043]

[0044]

[0045] where T, F, and A are the number of sampled data packets, subcarriers, and receiving sensors, respectively;

[0046]

[0047] Θ is initialized to 0. In one iteration, the expectation step and the maximization step are respectively executed for L paths to update the parameters. When the estimation of Θ converges, that is, when the difference between consecutive estimations is within a predefined threshold ∈, the iteration ends, and the parameter estimation values of the L paths are obtained.

[0048] If the collected data is channel state information data, the feature set of the sample includes first fine features, correlation artificial features, energy spectrum artificial features, and channel parameter features; if the collected data is orthogonal frequency division multiplexing symbol stream data, the feature set of the sample includes second fine features.

[0049] The parameter configuration of the support vector machine is as follows:

[0050] The kernel function is a radial basis kernel function:

[0051] (RBF)K(x, x i ) = exp{-γ‖x - x i ‖ 2}

[0052] where the scale factor

[0053] The objective function is:

[0054]

[0055] where C = 1;

[0056] The multi-classification strategy is "one-against-one" (OAO).

[0057] A human activity sensing and recognition device based on Wi-Fi signals includes a memory, a processor, and a program stored in the memory. When the processor executes the program, the method described above is implemented.

[0058] A storage medium stores a program, and when the program is executed, the method described above is implemented.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) According to the high or low requirements for the accuracy of the recognition result, the present invention selects different data and data processing processes to obtain the recognition result. The processing speed of orthogonal frequency division multiplexing symbol stream data is fast, but the accuracy is low. The result based on channel state information data has high accuracy but slow speed. The combination of the two can be applied to different application scenarios. Compared with the deep learning method, the overall complexity is low and the calculation speed is fast.

[0061] (2) Each step of the present invention is flexibly adjustable. The number of features included in the feature set can be increased or decreased according to actual needs, and the features from multiple angles complement each other. It has a wide application range and strong applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a flowchart of the method of the present invention;

[0063] Figure 2 is a flowchart of the method for obtaining the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal of the present invention;

[0064] Figure 3 is a layout diagram of the experimental scenario of Embodiment 1;

[0065] Figure 4 is a layout diagram of the experimental scenario of Embodiment 2. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and the detailed implementation manners and specific operation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0067] Embodiment 1

[0068] This embodiment provides a method for human activity sensing and recognition based on Wi-Fi signals, and the application scenario has a low requirement for recognition accuracy. Therefore, the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to each sample is collected as the data to be processed.

[0069] The transmitter in this embodiment is a TL-WDR7300 dual-band router; the receiver is a general software radio peripheral USRP X300, and an omnidirectional antenna with the model VERT2450 is used; the distance between the transmitter and the receiver is 1 meter; the transmission Wi-Fi signal frequency band is 5.745 GHz (149), and the bandwidth is 20 MHz. According to Figure 3 a Wi-Fi signal receiver system based on a general software radio peripheral shown, a signal transmission system is established. The signal transmitted from the router passes through the space of different human activity scenarios and is received by the signal receiver. According to Figure 2 the signal data selection method provided, the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal is obtained, and the obtained stream data is taken as a group of sample data with 3000 sampling points. The samples include training set samples and samples to be tested.

[0070] For each orthogonal frequency division multiplexing symbol stream data segment of each sample, normalize the amplitude and phase of the data respectively. Extract the amplitude statistical histogram features and phase statistical histogram features of each sample at intervals of 0.01. Extract the fast Fourier transform amplitude features through fast Fourier transform. There are 100 amplitude statistical histogram features and 100 phase statistical histogram features respectively, and 3000 fast Fourier transform amplitude features. Combine them to obtain a total of 3200-dimensional rough features.

[0071] Perform principal component analysis on the rough features for feature dimensionality reduction. Select the principal components with the cumulative variance greater than the pre-configured cumulative feature proportion threshold to obtain fine features. The pre-configured cumulative feature proportion threshold is 0.99.

[0072] The principal component analysis is as follows:

[0073]

[0074] Y = XW

[0075] where W = (w1, w2,..., wq) is composed of the eigenvectors (w1, w2,..., w q ) corresponding to the first q largest eigenvalues (λ1, λ2,..., λ q ) of the covariance matrix S. q is the number of principal components retained. Y is the data after reducing the original data from high dimension to low dimension, X is the original data, and N is the dimension of the original data.

[0076] Through multiple tests and dataset partitioning, 100 training set samples and 200 test samples are obtained for both the one-person scenario and the no-person scenario. Input the sample types and fine features of the training set samples into the support vector machine for classification and recognition training.

[0077] The parameter configuration of the support vector machine is as follows:

[0078] The kernel function is the radial basis kernel function:

[0079] (RBF)K(x, x i ) = exp{-γ‖x - x i ‖ 2}

[0080] where the scale factor

[0081] The objective function is:

[0082]

[0083] where C = 1;

[0084] The multi-classification strategy is "one-against-one" (OAO).

[0085] Input the fine features of the sample to be tested into the trained support vector machine for classification and recognition to obtain the classification and recognition results.

[0086] After verification in the experimental scenario, when distinguishing between the scenarios of no person and one person, the recognition accuracy of this method can reach 99.3%; when distinguishing between the scenarios of no person, one person standing, and one person squatting, the overall recognition accuracy is 70%.

[0087] Embodiment 2

[0088] This embodiment provides a human activity sensing and recognition method based on Wi-Fi signals. Since the application scenario has high requirements for recognition accuracy, the channel state information data of the Wi-Fi signal corresponding to each sample is collected as the data to be processed.

[0089] In this embodiment, the transmitter is a Nokia Beacon 1 dual-band router; the receiver is a Nokia Beacon 1 dual-band router; the distance between the transmitter and the receiver is 2 meters; the Wi-Fi signal transmission frequency band is 5.775 GHz (155), and the bandwidth is 20 MHz.

[0090] According to Figure 4 As shown in a Wi-Fi signal receiver system based on a Nokia Beacon 1 dual-band router, a signal transmission system is established. The signal transmitted from the router passes through the space of different human activity scenarios and is received by the receiver. According to the channel state information tool, Wi-Fi data packets are obtained at a sampling frequency of 100 hz. One data packet lasts for 5 seconds, and the channel state information data of the Wi-Fi signal of one sample is obtained through one data packet.

[0091] Perform principal component analysis on the subcarrier domain of the collected channel state information data, retain the first 4 principal components, perform one-dimensional processing on the information of the first 4 principal components in the subcarrier domain, and then normalize the amplitude of the data. With an interval of 0.001, 1000 amplitude statistical histogram features of each sample data are extracted and merged to obtain a total of 1000-dimensional rough features.

[0092] Perform feature dimensionality reduction on the rough features through principal component analysis, select the principal components whose cumulative variance is greater than the pre-configured cumulative feature ratio threshold, and obtain fine features. The pre-configured cumulative feature ratio threshold is 0.99.

[0093] The principal component analysis is:

[0094]

[0095] Y = XW

[0096] where \(W=(w_1, w_2, \ldots, w_q)\) consists of the first \(q\) largest eigenvectors \((\lambda_1, \lambda_2, \ldots, \lambda\) q ) corresponding to the eigenvalues of the covariance matrix \(S\), \((w_1, w_2, \ldots, w\) q ), \(q\) is the number of principal components retained, \(Y\) is the data after reducing the original data from high dimension to low dimension, \(X\) is the original data, and \(N\) is the dimension of the original data.

[0097] Extract the 1st to 3rd minima of the time-domain cross-correlation data of the amplitudes of different subcarriers of the channel state information data to obtain correlation artificial features; perform a short-time Fourier transform on the amplitude data of the channel state information data to obtain the energy spectral density, accumulate the energy from 2 Hz to 30 Hz to obtain an energy impact curve, and extract the peak of the energy impact curve to obtain energy spectral artificial features.

[0098] Divide the collected time-domain channel state information data into segments of 0.1 s based on the channel state information data, packet sampling frequency, signal center frequency, bandwidth, and number of frequency points, where the signal parameters are assumed to be static. Apply the space-alternating generalized expectation maximization algorithm to each segment to obtain the estimation of the signal parameters of 10 paths of the propagation channel in the time domain, and extract the variance of the time-domain Doppler parameters to obtain channel parameter features.

[0099] The space-alternating generalized expectation maximization algorithm is as follows:

[0100] For a given observation \(h(m)\), output the maximum likelihood estimation of the multi-dimensional multipath signal parameter \(\Theta\), where the parameter to be estimated is \(\Theta = \{\theta_1, \theta_1,......, \theta_1\}\), \(\tau_1, \varphi\) l and are the time of flight, unit direction vector angle of arrival, and Doppler frequency shift of the \(l\)th path, \(\alpha\) l is the complex amplitude of the \(l\)th path, and the log-likelihood function of \(\Theta\) is:

[0101]

[0102] where \(L\) is the total number of multipath components, \(S\) l is the signal of the \(l\)th path;

[0103] The maximum likelihood estimation problem to be solved is:

[0104]

[0105] In the expectation E step, for the \(l\)th path, there is an expectation function:

[0106]

[0107] where, is the parameter estimated in the previous iteration;

[0108] In the maximization M step, for the l-th path, there is a series of maximization functions:

[0109]

[0110]

[0111]

[0112]

[0113] where T, F, and A are the numbers of sampled data packets, subcarriers, and receiving sensors respectively;

[0114]

[0115] Θ is initialized to 0. In one iteration, the expectation step and the maximization step are respectively executed for the L paths to update the parameters. When the estimation of Θ converges, that is, when the difference between consecutive estimations is within a predefined threshold ∈, the iteration ends, and the parameter estimation values of the L paths are obtained.

[0116] Through multiple tests and dataset partitioning, 100 training set samples and 190 samples to be tested are obtained for each of the no-person scenario, the one-person standing and waving scenario, and the one-person moving and waving scenario. The sample types and feature sets of the training set samples are input into the support vector machine for classification and recognition training. The feature set includes fine features, correlation artificial features, energy spectrum artificial features, and channel parameter features.

[0117] The parameter configuration of the support vector machine is as follows:

[0118] The kernel function is the radial basis kernel function:

[0119] (RBF)K(x,x i )=exp{-γ‖x - x i ‖ 2}

[0120] where the scale factor

[0121] The objective function is:

[0122]

[0123] where C = 1;

[0124] The multi-classification strategy is "one-against-one" (OAO).

[0125] The feature set of the samples to be tested is input into the trained support vector machine for classification and recognition to obtain the classification and recognition results.

[0126] After being verified by the experimental scenario, it has an overall recognition accuracy of 85% when distinguishing scenarios of no one, one person standing and waving, and one person moving and waving.

[0127] If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

Claims

1. A method for human activity sensing and recognition based on Wi-Fi signals, characterized in that, Including the following steps: Judge whether the accuracy requirement of the application scenario for the recognition result is high accuracy. If so, collect the channel state information data of the Wi-Fi signal corresponding to each sample. If not, collect the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to each sample. Wherein, the samples include training set samples and samples to be tested. Perform principal component analysis on the subcarrier domain of the channel state information data, retain the first n principal components, perform one-dimensional processing on the first n principal component information in the subcarrier domain, and extract the amplitude and phase information of the processed data to obtain the first rough features of the sample. Wherein, n is the pre-configured number of retained principal components. Extract the minimum value of the time-domain cross-correlation data of the amplitudes of different subcarriers of the channel state information data to obtain correlation artificial features, and include them in the feature set of the sample; perform short-time Fourier transform on the amplitude data of the channel state information data to obtain the energy spectral density, accumulate the energy in the low-frequency interval, obtain the energy impact curve, and extract the peak value of the energy impact curve to obtain energy spectral artificial features, and include them in the feature set of the sample. Divide the collected channel state information data into multiple segments in the time domain, apply the space-alternating generalized expectation maximization algorithm to each segment, obtain the estimation of the signal parameters of multiple paths in the time domain of the propagation channel, extract the variance of the time-domain Doppler parameters, obtain channel parameter features, and include them in the feature set of the sample. For the collected orthogonal frequency division multiplexing symbol stream data, extract the data of a pre-configured number of sampling points, and count the amplitude and phase information of the data to obtain the second rough features. Perform principal component analysis on the first rough features or the second rough features, select the principal components with the cumulative variance greater than the pre-configured cumulative feature ratio threshold, and obtain the first fine features or the second fine features respectively, and include them in the feature set of the sample. Input the sample types and feature sets of the training set samples into a support vector machine for classification and recognition training. Input the feature set of the samples to be tested into the trained support vector machine for classification and recognition to obtain the classification and recognition results. The space-alternating generalized expectation maximization algorithm is as follows: For a given observation \(h(m)\), output the maximum likelihood estimate of the multi-dimensional multipath signal parameter \(\Theta\), where the parameters to be estimated are \(\tau\) l , \(\varphi\) l and are the time of flight, the arrival angle of the unit direction vector, and the Doppler frequency shift of the \(l\)-th path, \(\alpha\) l is the complex amplitude of the \(l\)-th path, and the log-likelihood function of \(\Theta\) is: where L is the total number of multipath components, and S l is the signal of the l-th path; The maximum likelihood estimation problem to be solved is: In the expectation E step, for the l-th path, there is an expectation function: Among them, is the parameter estimated in the previous iteration; In the maximization M step, for the l-th path, there is a series of maximization functions: Wherein, T, F, and A are the number of sampled data packets, subcarriers, and receiving sensors respectively. Θ is initialized to 0. In one iteration, the expectation step and the maximization step are respectively executed for L paths to update the parameters. When the estimation of Θ converges, that is, when the difference between consecutive estimations is within the predefined threshold ∈, the iteration ends, and the parameter estimation values of L paths are obtained.

2. The method for human activity sensing and recognition based on Wi-Fi signals according to claim 1, wherein The Wi-Fi signal is transmitted from the transmitting end and received by the receiving end after passing through the sample space. Wherein, the sample space is the human activity scene space.

3. A method for human activity sensing and recognition based on Wi-Fi signals according to claim 1, characterized in that, The principal component analysis is as follows: Y = XW where \(W=(w_1, w_2, \ldots, w_q)\) consists of the eigenvectors \((w_1, w_2, \ldots, w\) q ) corresponding to the first \(q\) largest eigenvalues \((\lambda_1, \lambda_2, \ldots, \lambda\) q ) of the covariance matrix \(S\), \(q\) is the number of principal components to be retained, \(Y\) is the data after reducing the original data from high dimension to low dimension, \(X\) is the original data, and \(N\) is the dimension of the original data.

4. A method for human activity sensing and recognition based on Wi-Fi signals according to claim 1, characterized in that, The first rough features include amplitude histogram features, and the second rough features include amplitude histogram features, phase histogram features, and fast Fourier transform amplitude features.

5. A method for human activity sensing and recognition based on Wi-Fi signals according to claim 1, characterized in that, The channel state information data is divided into multiple segments in the time domain based on the channel state information data, the data packet sampling frequency, the signal center frequency, the bandwidth, and the number of frequency points.

6. A method for human activity sensing and recognition based on Wi-Fi signals according to claim 1, characterized in that If the collected data is channel state information data, the feature set of the sample includes first fine features, correlation artificial features, energy spectrum artificial features, and channel parameter features; if the collected data is orthogonal frequency division multiplexing symbol stream data, the feature set of the sample includes second fine features.

7. A method for human activity sensing and recognition based on Wi-Fi signals according to claim 1, characterized in that The parameter configuration of the support vector machine is as follows: The kernel function is a radial basis kernel function: (RBF)K(x,x i ) = exp{-γ‖x - x i ‖ 2} where the scale factor The objective function is: where C = 1; The multi-classification strategy is "one-against-one" (OAO).

8. A human activity sensing and identifying device based on Wi-Fi signals, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1-7.

9. A storage medium, on which a program is stored, characterized in that, When the program is executed, it implements the method described in any one of claims 1-7.

Images