A human activity detection and early warning method based on Wi-Fi signals

Abstract

The application relates to a human activity detection and early warning method based on a Wi-Fi signal, comprising the following steps: acquiring system requirements, judging whether the system requirements are initialization, if yes, obtaining a first rough data representation; performing principal component analysis on the first rough data representation to obtain a first fine feature, and counting the first fine feature into a first feature set of a training set sample; if no, obtaining a second rough data representation; performing principal component analysis linear transformation matrix on the second rough data representation to obtain a second fine feature, and counting the second fine feature into a second feature set of a to-be-tested sample; inputting the second feature set into a trained feedforward neural network classifier to perform recognition, and obtaining a recognition detection result; and outputting early warning information based on the recognition detection result. Compared with the prior art, the application has the advantages of fast calculation speed, real-time performance and the like.

Description

Technical Field

[0001] This invention relates to the field of signal processing, and in particular to a method for detecting and warning of human activity based on Wi-Fi signals. Background Technology

[0002] Human activity detection and hazard prevention is a key research area. Systems applying indoor human activity detection can mitigate potential risks, reduce rescue time after dangerous situations occur, and help the elderly reduce their dependence on caregivers. Currently, human activity detection systems, based on different sensing technologies, range from wearable sensor-based to computer vision-based to dedicated sensor-based methods. However, each of these technologies has its own limitations. Wearable sensor-based technologies require constant wearing of the device and frequent charging; computer vision-based technologies face issues such as privacy violations and the difficulty of real-time processing due to high computational demands; and dedicated sensing technologies require pre-installed specialized equipment, resulting in high costs. Given these drawbacks, the widespread deployment of Wi-Fi devices and advancements in commercial Wi-Fi-based sensing technologies have spurred extensive research into Wi-Fi signal-based human activity detection and hazard warning systems.

[0003] Existing research includes: a method based on channel state information, using wavelet analysis-based preprocessing and feature extraction techniques combined with a support vector machine (SVM) recognition model; extracting frequency components as features using discrete wavelet transform and establishing an SVM recognition model based on transfer learning theory; a multi-level fall detection system utilizing channel state information from Wi-Fi devices, performing low-pass filtering and discrete wavelet denoising on the raw channel state information data, then training a classification model to identify different activities, using SVM for initial judgment and deep learning algorithms for final judgment; a method using SVM and a residual neural network, employing a motion segmentation method based on channel state information variance, extracting features from the channel state information amplitude using a Gabor filter, fully utilizing channel state information; and a phase difference-based activity segmentation method, performing signal preprocessing (e.g., interpolation, bandpass filtering) to segment and separate fall activities, extracting eight statistical features in real time, etc.

[0004] While the aforementioned literature has studied human activity detection and danger warning based on Wi-Fi signals, it mainly relies on channel state information data and complex data processing recognition models, which are highly complex and lack real-time performance in practical applications. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a human activity detection and early warning method based on Wi-Fi signals, which has low overall complexity, fast calculation speed, and real-time performance.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for detecting and warning of human activity based on Wi-Fi signals, comprising the following steps:

[0008] Receive Wi-Fi signal, obtain system requirements, and determine if the system requirements are initialized.

[0009] If so, the first orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to the training set samples is collected, the first sampling data of the pre-configured number of sampling points is extracted from the collected first orthogonal frequency division multiplexing symbol stream data, and the amplitude and phase information of the first sampling data are statistically analyzed to obtain the first coarse data representation;

[0010] Principal component analysis is performed on the first coarse data representation. Principal components with cumulative variance greater than the pre-configured cumulative feature proportion threshold are selected to obtain the first fine features, which are then included in the first feature set of the training set samples. The parameter-adjustable feedforward neural network classifier is trained based on the first feature set.

[0011] If not, the second orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to the sample to be tested is collected, the second sampling data of the pre-configured number of sampling points is extracted from the collected second orthogonal frequency division multiplexing symbol stream data, and the amplitude and phase information of the second sampling data are statistically analyzed to obtain the second coarse data characterization;

[0012] For the second coarse data representation, the principal component analysis linear transformation matrix of the training set samples is used to perform the same linear transformation to obtain the second fine features, which are then included in the second feature set of the test samples.

[0013] The second feature set is input into the trained feedforward neural network classifier for recognition, and the recognition and detection results are obtained.

[0014] Early warning information is output based on the identification and detection results.

[0015] Furthermore, in the principal component analysis process, the expression for the covariance matrix is:

[0016]

[0017] Where S is the covariance matrix, N is the number of samples, and X is the mean-reduced data for each dimension of the original d-dimensional data containing N samples, where d is the number of dimensions.

[0018] Furthermore, during principal component analysis, based on the covariance matrix, refined features are obtained. The expression for the refined features is:

[0019] Y = XV

[0020] Where Y is the fine feature, V is the feature vector group, V is composed of the feature vectors corresponding to the first q largest eigenvalues ​​of the covariance matrix S, and q is the number of principal components retained.

[0021] Furthermore, the recognition and detection results are categorized into three scenarios: standing posture, sitting / lying posture, and falling posture. The warning information is also categorized into three types corresponding to the recognition and detection results: standing warning information, falling warning information, and sitting / lying warning information.

[0022] After identifying the standing posture scene, if a preset number or more of the test samples are continuously determined to be standing posture scenes, a standing warning message will be output.

[0023] After identifying a fall posture scenario, if a preset number or more of the test samples are continuously identified as fall posture scenarios, a fall warning message will be output.

[0024] After identifying a sitting or lying posture scene, determine whether it is during the daytime. If not, no warning information is output. If so, a sitting or lying warning information is output when a preset number or more of the test samples are continuously identified as sitting or lying posture scenes.

[0025] Furthermore, the coarse data characterization includes amplitude histogram features, phase histogram features, and fast Fourier transform amplitude features.

[0026] Furthermore, the Wi-Fi signal is transmitted from the transmitting end, passes through the sample space, and is received by the receiving end. The sample space is the human activity scene space.

[0027] Furthermore, the Wi-Fi signal is transmitted from the transmitting end, passes through the sample space, and is received by the receiving end. The sample space is the human activity scene space.

[0028] Furthermore, the algorithm used in the training process of the feedforward neural network classifier is the quantized conjugate gradient method.

[0029] Furthermore, the activation function of the hidden layer is the hyperbolic tangent function.

[0030] Furthermore, the acquisition process of orthogonal frequency division multiplexing symbol stream data is as follows:

[0031] The receiver receives the Wi-Fi signal sent by the Wi-Fi signal source and performs frame detection, frequency offset correction, symbol alignment, phase offset correction and channel estimation on the Wi-Fi signal in sequence to obtain orthogonal frequency division multiplexing symbol stream data.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) The present invention obtains orthogonal frequency division multiplexing symbol stream data segments from the received Wi-Fi signal, obtains coarse data representation from the orthogonal frequency division multiplexing symbol stream data, and further combines principal component analysis to obtain fine features. Compared with the method based on channel state information, it does not rely on special equipment, only requires single signal link data, has low overall complexity, fast calculation speed, and real-time performance.

[0034] (2) The feedforward neural network classifier used in this invention has adjustable parameters. By changing the pre-configured cumulative feature ratio threshold, the number of features in the feature set can be changed according to actual needs, thereby adjusting the sensitivity. It has a wide range of applications and strong applicability.

[0035] (3) This invention does not require additional sensing equipment and protects the privacy of the monitored object. Attached Figure Description

[0036] Figure 1 This is a flowchart of the present invention;

[0037] Figure 2 This is a flowchart illustrating the acquisition process of orthogonal frequency division multiplexing symbol stream data according to the present invention.

[0038] Figure 3 This is a diagram showing the experimental setup of an embodiment of the present invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0040] This invention proposes a method for detecting and warning of human activity based on Wi-Fi signals. The method utilizes a device to collect Wi-Fi signals from environmental routing equipment, obtaining orthogonal frequency division multiplexing (OFDM) symbol stream data of the signals in the environment to be detected. Based on a recognition and detection model trained during the initialization process, it can identify and detect human activities, providing warnings for scenarios involving falls and alerts for prolonged standing, sitting, or lying down. When different human activities are present in the detected environment, alert or warning information is output to the security system for comprehensive risk assessment.

[0041] The flowchart of the human activity detection and early warning method based on Wi-Fi signals proposed in this invention is as follows: Figure 1 As shown, the method includes the following steps:

[0042] Determine if the system requires initialization.

[0043] If so, collect the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signals corresponding to the training set samples.

[0044] If not, collect the orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to the sample to be tested;

[0045] For the training set samples and the samples to be tested, the data of the pre-configured number of sampling points are extracted from the collected orthogonal frequency division multiplexing symbol stream data, and the amplitude and phase information of the data are statistically analyzed to obtain a rough data representation.

[0046] Principal component analysis is performed on the coarse data representation of the training set samples. Principal components with cumulative variance greater than the pre-configured cumulative feature proportion threshold are selected to obtain fine features, which are then included in the feature set of the training set samples.

[0047] The sample types and feature sets of the training set samples are input into the feedforward neural network classifier for recognition and detection training.

[0048] For the coarse data representation of the test sample, the principal component analysis linear transformation matrix of the training set sample is used to perform the same linear transformation to obtain fine features, which are then included in the feature set of the test sample.

[0049] The feature set of the sample to be tested is input into the trained feedforward neural network classifier for recognition, and the recognition and detection results are obtained.

[0050] Therefore, the feature set of the training set samples is the refined feature obtained by principal component analysis feature extraction; the feature set of the test samples is the refined feature obtained by the same linear transformation of principal component analysis based on the training set samples.

[0051] In this invention, the Wi-Fi signal is transmitted from the transmitter, passes through the sample space, and is received by the receiver, wherein the sample space is the human activity scene space.

[0052] The specific steps of principal component analysis are as follows:

[0053] The covariance matrix is ​​calculated based on data representation. Then, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. The eigenvalues ​​are arranged in order, with different eigenvalues ​​corresponding to different eigenvectors. These eigenvectors are the projection vectors from the high-dimensional data space to the low-dimensional data space. Several projection vectors can form a linear transformation matrix from the high-dimensional data space to the low-dimensional data space. The new low-dimensional data is obtained by linearly transforming the original high-dimensional data along several dimensional projection directions in the low-dimensional data space. The magnitude of the eigenvalues ​​reflects the variance of the data values ​​after projection along a certain direction. The variance indicates the amount of information retained in the original data along that projection direction. The k eigenvectors with the largest eigenvalues ​​selected from largest to smallest are used as linear transformation matrices. The sample points representing the data are projected onto these k eigenvectors to reduce dimensionality. The resulting low-dimensional projection is called the principal component. Cumulative variance refers to the cumulative variance of the variances of the k-dimensional data retained in the low-dimensional space. The dimensionality reduction process is essentially removing the principal components with smaller variances after the linear transformation to reduce the data dimensionality and extract effective features. Cumulative variance reflects the degree of dimensionality reduction. By pre-configuring a cumulative variance percentage threshold, we can ensure that the data dimensionality reduction retains the effective information in the original data. The value of k can be controlled. The larger the cumulative variance percentage threshold, the more high-dimensional data information is retained, and the larger the value of k, the worse the dimensionality reduction effect.

[0054] In principal component analysis, the covariance matrix is ​​expressed as follows:

[0055]

[0056] Where S is the covariance matrix, N is the number of samples, and X is the mean-reduced data for each feature dimension based on the d-dimensional original data containing N samples, where d is the number of dimensions. X is the coarse data representation.

[0057] Based on the covariance matrix, the first refined feature is obtained, and the expression for the refined feature is:

[0058] Y = XV

[0059] Where Y is the fine feature, V is the feature vector group, V is composed of the eigenvectors corresponding to the first q largest eigenvalues ​​of the covariance matrix S, and q is the number of principal components retained. When X is the first coarse data representation, Y is the first fine feature; when X is the second coarse data representation, Y is the second fine feature.

[0060] The coarse data characterization of this invention includes amplitude histogram features, phase histogram features, and fast Fourier transform amplitude features.

[0061] The feedforward neural network classifier used in this invention is a type of neural network application, comprising an input layer, one or more hidden layers containing a number of neurons, and an output layer containing a number of neurons. Adjacent layers are fully connected. Each neuron in the neural network includes a non-linear activation function. A learning algorithm is used to execute the learning process and achieve the design goal. Its configuration is as follows:

[0062] The network learning and training algorithm is the quantized conjugate gradient method;

[0063] The activation function of the hidden layer is set to the hyperbolic tangent function;

[0064] The activation function of the output layer is set to a linear function.

[0065] The recognition and detection results of this invention are divided into three scenarios: standing posture, sitting / lying posture, and falling posture. The warning information is divided into three types corresponding to the recognition and detection results: standing warning information, falling warning information, and sitting / lying warning information.

[0066] After identifying the standing posture scene, if a preset number or more of the test samples are continuously determined to be standing posture scenes, a standing warning message will be output.

[0067] After identifying a fall posture scenario, if a preset number or more of the test samples are continuously identified as fall posture scenarios, a fall warning message will be output.

[0068] After identifying a sitting or lying posture scene, determine whether it is during the daytime. If not, no warning information is output. If so, a sitting or lying warning information is output when a preset number or more of the test samples are continuously identified as sitting or lying posture scenes.

[0069] The acquisition process of orthogonal frequency division multiplexing symbol stream data in this invention is as follows: Figure 2 As shown, the specific steps are as follows:

[0070] The receiver receives the Wi-Fi signal sent by the Wi-Fi signal source and performs frame detection, frequency offset correction, symbol alignment, phase offset correction and channel estimation on the Wi-Fi signal in sequence to obtain orthogonal frequency division multiplexing symbol stream data.

[0071] The following is a specific example for verification. 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. In this embodiment, the signal transmitter is a TL-WDR7300 dual-band router; the signal receiver is a general-purpose software radio peripheral USRP X300 using a single antenna, with the antenna model being a VERT2450 omnidirectional antenna; the signal receiver is connected to the host computer device executing the program via a network cable; the distance between the router antenna and the receiver antenna is 1.5 meters; the router device and the receiver device are placed 1 meter above the ground; the Wi-Fi signal transmission frequency band is 5.745 GHz (149), and the bandwidth is 20 MHz.

[0072] according to Figure 3 This demonstration showcases a Wi-Fi signal receiver system based on a general-purpose software-defined radio peripheral. A signal transmission system is established, where signals transmitted from a router are received by the receiver across various human activity scenarios, obtaining orthogonal frequency division multiplexing (OFDM) symbol stream data of the Wi-Fi signal. The obtained stream data is then grouped into sets of 4000 samples each. These samples include training set samples and test samples.

[0073] For each sample of the orthogonal frequency division multiplexing symbol stream data segment, the amplitude and phase of the data are normalized separately. Amplitude and phase statistical histogram features are extracted for each sample at 0.01 intervals. Fast Fourier Transform (FFT) amplitude features are then extracted. A total of 100 amplitude and 100 phase statistical histogram features are obtained, resulting in 4000 FFT amplitude features. These are combined to obtain a coarse data representation of 4200 dimensions.

[0074] Through multiple tests and dataset division, three scenarios were identified: one person standing, one person sitting, and one person falling. 1200 training samples and 100 test samples were obtained for each scenario.

[0075] During the initialization process, principal component analysis feature extractor is used to extract the coarse features of the training samples. Principal components with cumulative variance greater than the pre-configured cumulative feature proportion threshold are selected to obtain fine features. The pre-configured cumulative feature proportion threshold is 99.9%.

[0076] The sample types and fine features of the training set samples are input into the feedforward neural network classifier for recognition and detection training.

[0077] The feedforward neural network classifier used in this embodiment has the following configuration:

[0078] The network learning and training algorithm is the quantized conjugate gradient method;

[0079] The activation function of the hidden layer is set to the hyperbolic tangent function;

[0080] The output layer activation function is set to a linear function;

[0081] The number of hidden layers is 1;

[0082] The number of neurons in the hidden layer is 50;

[0083] The learning rate is 0.02;

[0084] The optimization target value is 0.0002.

[0085] In the non-initialization process, the coarse data of the test sample is characterized, and the same linear transformation is performed through principal component analysis of the test sample to extract the fine features of the test sample.

[0086] The fine features of the sample to be tested are input into the trained feedforward neural network classifier for recognition and detection, and the recognition and detection results are obtained.

[0087] Experimental results show that the system provided by this invention can be used for human activity detection and danger warning. After experimental scenario verification, when distinguishing between three scenarios—one person standing, one person sitting, and one person falling—based on the detection results of 100 test samples, the system can achieve a 100% recognition accuracy and a 0% false alarm rate for the one-person falling scenario; a 75% recognition accuracy for the one-person sitting scenario; and a 94% recognition accuracy for the one-person standing scenario.

[0088] It should be understood that the above embodiments are merely examples, and the technical solutions of the present invention are not limited. Based on the equivalent structures and processes described in the specification and drawings of the present invention, those skilled in the art can make settings according to actual needs.

[0089] Through the above description of the embodiments, those skilled in the art can understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media capable of storing program code.

[0090] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for human activity detection and early warning based on Wi-Fi signals, characterized in that, The method comprises the following steps: receiving a Wi-Fi signal, obtaining system requirements, and determining whether the system requirements are initialization, if yes, collecting first orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to a training set sample, extracting first sampling data of a preconfigured number of sampling points from the collected first orthogonal frequency division multiplexing symbol stream data, and obtaining first rough data representation by counting amplitude and phase information of the first sampling data; performing principal component analysis on the first rough data representation, selecting principal components with cumulative variance greater than a preconfigured cumulative feature proportion threshold, obtaining first fine features, and adding the first fine features to a first feature set of the training set sample, and training a parameter-adjustable feedforward neural network classifier based on the first feature set; if no, collecting second orthogonal frequency division multiplexing symbol stream data of the Wi-Fi signal corresponding to a to-be-tested sample, extracting second sampling data of a preconfigured number of sampling points from the collected second orthogonal frequency division multiplexing symbol stream data, and obtaining second rough data representation by counting amplitude and phase information of the second sampling data; performing the same linear transformation on the second rough data representation by using a principal component analysis linear transformation matrix of the training set sample to obtain second fine features and add the second fine features to a second feature set of the to-be-tested sample; inputting the second feature set into the trained feedforward neural network classifier to obtain an identification detection result; outputting early warning information based on the identification detection result; During the principal component analysis, the expression of the covariance matrix is: wherein, is a covariance matrix, is a number of samples, is data after mean removal for each dimension feature based on d-dimensional original data containing N samples, d is a dimension number; During the principal component analysis, the fine features are obtained based on the covariance matrix, and the expression of the fine features is: wherein, for fine features, for feature vector groups, by the eigenvectors corresponding to the first q largest eigenvalues of the covariance matrix , q is the number of retained principal components; The identification detection result is divided into three cases of a standing posture scene, a sitting and lying posture scene, and a falling posture scene, and the early warning information is divided into three types of standing early warning information, falling early warning information, and sitting and lying early warning information corresponding to the identification detection result; The rough data representation includes an amplitude histogram feature, a phase histogram feature, and a fast Fourier transform amplitude feature. The collection process of the orthogonal frequency division multiplexing symbol stream data is as follows: The receiver receives the Wi-Fi signal sent by the Wi-Fi signal source, and sequentially performs frame detection, frequency offset correction, symbol alignment, phase offset correction, and channel estimation on the Wi-Fi signal to obtain the orthogonal frequency division multiplexing symbol stream data.

2. The human activity detection and warning method based on Wi-Fi signal according to claim 1, characterized in that, After the identification detection result is the standing posture scene, when a continuous preconfigured number of to-be-tested samples are all determined to be the standing posture scene, the standing early warning information is outputted; After the identification detection result is the falling posture scene, when a continuous preconfigured number of to-be-tested samples are all determined to be the falling posture scene, the falling early warning information is outputted; After the identification detection result is the sitting and lying posture scene, it is determined whether it is in a daytime period, if no, no early warning information is outputted, if yes, when a continuous preconfigured number of to-be-tested samples are all determined to be the sitting and lying posture scene, the sitting and lying early warning information is outputted. 3.The human activity detection and early warning method based on Wi-Fi signal according to claim 1, characterized in that, The Wi-Fi signal is transmitted from a transmitting end, passes through a sample space, and is received by a receiving end, and the sample space is a human activity scene space. 4.The human activity detection and early warning method based on Wi-Fi signal according to claim 1, characterized in that, The feedforward neural network classifier comprises an input layer, one or more hidden layers comprising a plurality of neurons, and an output layer comprising a plurality of neurons, and full connection between adjacent layers, and the model of each neuron comprises an activation function.

5. The human activity detection and warning method based on Wi-Fi signal according to claim 4, characterized in that, In the training process of the feedforward neural network classifier, the algorithm used is the quantized conjugate gradient method.

6. The human activity detection and warning method based on Wi-Fi signal according to claim 4, characterized in that, The activation function of the hidden layer is the hyperbolic tangent function.

Images