A WIFI fall monitoring algorithm combined with human posture evaluation

Abstract

The embodiment of the application provides a WIFI fall monitoring algorithm combined with human body posture evaluation. The method comprises the following steps: collecting WIFI signals and multi-cascading posture sensing signals when a human body falls; extracting CSI data from the WIFI signals; extracting human body posture information from the multi-cascading posture sensing signals; performing amplitude and phase preprocessing on the CSI data; dividing the preprocessed CSI data into a training set, a verification set and a test set; introducing the training set into an algorithm model, and performing feature extraction, classification and training in combination with the human body posture information; and performing fall identification test through the test set. The application improves the cross-environment detection capability of the model, so that the model still has the same detection accuracy under the condition of cross-environment, thereby overcoming the problem that it is difficult to adapt to environmental changes in the traditional machine learning method.

Description

Technical Field

[0001] This invention relates to the technical field of computer analysis applications, and more particularly to a WIFI fall detection algorithm that incorporates human posture assessment. Background Technology

[0002] As living standards improve, many countries are facing serious aging population problems. Falls are one of the most common health issues affecting the elderly, as they can potentially cause irreversible damage. Contactless and private WiFi has attracted research attention due to its ubiquity and non-invasive nature.

[0003] However, the performance of existing research is usually based on data collection and model training in specific environments. Furthermore, trained models often contain environment-specific information, making them difficult to adapt to environmental changes and unable to generalize to new environments. Therefore, improving the model's ability to detect falls across different environments has become a pressing technical problem for fall detection research. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a WIFI fall detection algorithm that combines human posture assessment to at least partially solve the above problems.

[0005] According to a first aspect of the present invention, a WIFI fall detection algorithm combining human posture assessment is provided, comprising: collecting WIFI signals and multi-cascade posture sensing signals when a person falls; extracting CSI data from the WIFI signals; extracting human posture information from the multi-cascade posture sensing signals; preprocessing the amplitude and phase of the CSI data; dividing the preprocessed CSI data into a training set, a validation set, and a test set; importing the training set into the algorithm model and performing feature extraction and classification training in conjunction with the human posture information; and conducting fall recognition tests using the test set.

[0006] In one implementation, the CSI data undergoes amplitude and phase preprocessing, including standardization, filtering, and other processing of the amplitude data; and phase error elimination and subcarrier noise component elimination of the phase data.

[0007] In another implementation, human posture information is extracted from multi-cascaded posture sensing signals, including obtaining the acceleration and angle of the human body when performing actions from the multi-cascaded posture sensing signals of multi-cascaded posture sensors; by processing the acceleration and angle of the human body when performing actions, the real-time acceleration and direction of the human body's movement are obtained as human posture information.

[0008] In another implementation, the algorithm model consists of an amplitude sensing module and a phase sensing module. The algorithm model combines human posture information and performs feature extraction and classification training through the amplitude sensing module and the phase sensing module.

[0009] The beneficial effects of the present invention are as follows:

[0010] (1) The present invention uses WiFi signals as a fall detection method. WiFi signals do not record image and audio information, nor do they require wearing other sensors, thus protecting the privacy of the test subject and facilitating the test subject's activities.

[0011] (2) This invention uses ordinary commercial WiFi devices to establish a fall detection environment without the need for expensive hardware devices and complicated deployment, which increases ease of use and practicality, and has good promotion and adaptability.

[0012] (3) This invention applies deep learning technology to fall detection and adds two types of information in human motion posture to the WiFi signal, namely acceleration and direction, which overcomes the problem that traditional machine learning methods are difficult to adapt to environmental changes and has the ability to have the same detection accuracy under cross-environment conditions. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a flowchart illustrating the steps of a WIFI fall detection algorithm that incorporates human posture assessment, according to an embodiment of the present invention.

[0015] Figure 2 This is a structural diagram of the amplitude sensing module and the phase sensing module according to another embodiment of the present invention.

[0016] Figure 3 This is a structural diagram of the human posture extractor in the amplitude sensing module and phase sensing module of another embodiment of the present invention.

[0017] Figure 4 This is an overall framework diagram of a WIFI fall detection algorithm that combines human posture assessment, according to another embodiment of the present invention. Detailed Implementation

[0018] To provide a clearer understanding of the technical features, objectives, and effects of the embodiments of the present invention, specific implementation methods of the embodiments of the present invention will now be described with reference to the accompanying drawings.

[0019] In this document, “exemplary” means “serving as an example, illustration or description”, and any illustrations or implementations described herein as “illustrative” should not be construed as a more preferred or advantageous technical solution.

[0020] To keep the drawings concise, only the parts relevant to the invention are shown schematically in each figure, and they do not represent the actual structure of the product. Furthermore, for ease of understanding, some figures show only one or more components with the same structure or function, or only one or more are labeled.

[0021] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0022] To facilitate understanding, before describing the specific embodiments of the present invention in detail, the prior art of the WIFI fall detection algorithm that combines human posture assessment will be described by way of example.

[0023] Current technologies for fall detection generally fall into two categories: wearable device-based and camera-based methods. Wearable devices, including accelerometers, gyroscopes, and RFID, have been studied for use in fall detection. However, for both wearable device-based and camera-based fall detection methods, wearable devices are not convenient for automatically protecting the elderly anytime and anywhere, while fall detection methods based on indoor camera deployment pose serious privacy issues for the elderly in their home lives.

[0024] To overcome the limitations of wearable devices and cameras, WiFi-based systems have attracted attention in recent years. Research on WiFi fall detection mainly falls into two categories.

[0025] One approach uses RSSI to characterize indoor activity on WiFi or other wireless devices; for example, PAWS explored the RSSI fingerprints of WiFi ambient signals for different activities. However, because RSSI measures the overall amplitude response of multiple superimposed subcarriers, it loses phase information and detailed information about each frequency component.

[0026] Another type uses CSI to characterize indoor activities on WiFi or other wireless devices. CSI can extract standard information from commercial WiFi devices and simultaneously measure amplitude and phase information at different frequency components, providing more granular information about the propagation environment. This has become a research hotspot in the field of wireless sensing in recent years. Currently, CSI-based fall detection algorithms are mainly divided into learning algorithms with manually extracted features and learning algorithms with automatically extracted features. The former relies on manually extracted CSI matrices to extract statistical features of CSI, such as mean, standard deviation, and variance, and uses machine learning models to learn the feature differences between CSI measurement matrices under different action conditions to identify fall actions. WiFall and RT-Fall are early works using CSI data to build fall detection stepping stones. Both methods use SVM-based classifiers to detect fall events. FallDeFi improves the performance of WiFall and RT-Fall by applying short-time Fourier transform (STFT) spectrograms and power burst curves (PBCs) to extract features. Sensing-Fi detects falls based on WiFi signals and measurements from ground-mounted accelerometers to capture floor vibrations. While traditional machine learning algorithms are fast, their drawback is that the performance of most machine learning algorithms depends on the accuracy of the extracted features. Feature processing is time-consuming and requires professional knowledge to label features. Furthermore, the experience-based errors caused by manually selecting features are difficult to control.

[0027] Deep learning eliminates the need for manual feature selection. The amplitude and phase information of the CSI matrix only requires preprocessing (segmentation, noise reduction, etc.) before being input into a neural network model (such as a deep neural network, convolutional neural network (CNN), recurrent neural network, etc.) for supervised training. The model can then automatically calculate weight biases and extract data features for classification. For example, a stacked CNN model includes a classifier for identifying rooms and a detection model for identifying activities in each room. During detection, it first identifies which room the activity occurred in, and then determines which activity occurred in that room.

[0028] However, because the features extracted in the aforementioned existing WiFi-based FD systems are environment-dependent, the classifiers trained in these works are affected by environmental changes and cannot generalize well to new environments without performance degradation. These systems require retraining when environmental settings change, making them impractical as it is not suitable to require users to fall down and collect training data every time furniture placement or equipment deployment is changed.

[0029] Current research has also proposed solutions to cross-scenario problems. For example, DeFall compares patterns of real-time velocity and acceleration estimates with templates to detect falls. However, DeFall uses humanoid dummies to learn fall models, so it can only detect simulated "hard falls," i.e., falling from a stationary position at a certain height. This property greatly limits its practical application.

[0030] Therefore, improving the model's ability to detect fall damage across environments has become a pressing technical problem to be solved in fall detection research.

[0031] The specific implementation of the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0032] See Figure 1 This is a flowchart illustrating the steps of a WIFI fall detection algorithm combining human posture assessment according to an embodiment of the present invention, mainly including the following steps:

[0033] Step S101: Collect WIFI signals and multi-cascaded attitude sensor signals when a person falls;

[0034] Step S102: Extract CSI data from the WIFI signal;

[0035] Step S103: Extract human posture information from multi-cascaded posture sensing signals;

[0036] Step S104: Perform amplitude and phase preprocessing on the CSI data;

[0037] Step S105: Import the training set into the algorithm model and perform feature extraction and classification training by combining human posture information.

[0038] Step S106: Divide the preprocessed CSI data into training set, validation set and test set;

[0039] Step S107: Perform a fall recognition test using a test set.

[0040] In summary, the beneficial effects of the present invention are as follows:

[0041] (1) The present invention uses WiFi signals as a fall detection method. WiFi signals do not record image and audio information, nor do they require wearing other sensors, thus protecting the privacy of the test subject and facilitating the test subject's activities.

[0042] (2) This invention uses ordinary commercial WiFi devices to establish a fall detection environment without the need for expensive hardware devices and complicated deployment, which increases ease of use and practicality, and has good promotion and adaptability.

[0043] (3) This invention applies deep learning technology to fall detection and adds two types of information in human motion posture to the WiFi signal, namely acceleration and direction, which overcomes the problem that traditional machine learning methods are difficult to adapt to environmental changes and has the ability to have the same detection accuracy under cross-environment conditions.

[0044] In one implementation, the CSI data undergoes amplitude and phase preprocessing, including standardization, filtering, and other processing of the amplitude data; and phase error elimination and subcarrier noise component elimination of the phase data.

[0045] For example, CSI data undergoes amplitude and phase preprocessing, including CSI amplitude data preprocessing and CSI phase data preprocessing. The specific CSI amplitude preprocessing procedure is as follows:

[0046] ① Standardization processing: Perform layer-based standardization processing on the data of each transmit / receive antenna pair. At this time, the data matrix size is N. rt ×T w The antennas are arranged logarithmically, separating the number of subcarriers, and the resulting matrix is ​​of size N. rt ×N sc ×T w The data. Where N... rt N is the product of the number of receiving antennas and the number of transmitting antennas. sc T represents the number of subcarriers. w The length of the signal acquired within a specified time period. Therefore, the standardized amplitude data is:

[0047]

[0048] Among them, H amp and H amp These are the amplitude matrices before and after normalization for a given channel, μ. amp and σ amp These are the mean and standard deviation of the magnitude matrix, respectively.

[0049] ② Filtering: Hampel filtering is performed on the subcarrier samples in each pair of transmit and receive antennas to eliminate peak values ​​and outliers. A sliding window of length 30 is used to calculate the difference between the data point on the subcarrier and the median value of the data in the window. If the difference exceeds the product of the threshold and the standard deviation, the median value of the data in this window will replace the outlier in the interval.

[0050] ③ Other processing: First, linear interpolation is performed on the standardized CSI data, which estimates the value based on the two nearest neighbors of the point to be interpolated in the data sequence. Then, linear regression is used to smooth the trend of CSI and zero-padding is applied to the CSI data.

[0051] The specific CSI phase data preprocessing process is as follows:

[0052] ① Eliminating Phase Errors: Firstly, phase differences are sensitive to environmental changes. However, raw CSI phase data contains two main types of noise: carrier frequency offset (CFO) and sampling time offset (STO). The variance of the phase difference between two antenna pairs is the sum of the individual variances of each antenna. Different antennas in a WiFi network card share the same RF oscillator, resulting in identical time-varying phase offsets. Therefore, the phase difference between the transmitting and receiving antenna pairs can be calculated using the following formula, effectively eliminating phase errors.

[0053]

[0054] Where, θ ij This represents the phase difference between the i-th transmit / receive antenna pair and the j-th transmit / receive antenna pair.

[0055] ② Eliminating subcarrier noise components: Because the phase information obtained by WiFi devices is not very reliable, the carrier frequencies between the transmitter and receiver cannot be perfectly synchronized. Therefore, subcarrier noise components can be eliminated using the following formula:

[0056]

[0057] in,

[0058]

[0059]

[0060]

[0061] Where, ψ K It is the phase after linear transformation calibration, K is the number of subcarriers, and Φ K It is the phase after unwinding, φ K δ is the original phase, β is the timing error offset (delay), β is the random phase offset caused by the device, Z is the random noise during the measurement process, and φ is the random phase offset caused by the device. m It is the phase of the m-th subcarrier, η K It is the subcarrier index, η m Let m be the m-th subcarrier, and N be the Fast Fourier Transform. After the transformation using the above formula, the changes in phase over time can reflect changes in motion.

[0062] In another implementation, human posture information is extracted from multi-cascaded posture sensing signals, including obtaining the acceleration and angle of the human body when performing actions from the multi-cascaded posture sensing signals of multi-cascaded posture sensors; by processing the acceleration and angle of the human body when performing actions, the real-time acceleration and direction of the human body's movement are obtained as human posture information.

[0063] In another implementation, the algorithm model consists of an amplitude sensing module and a phase sensing module. The algorithm model combines human posture information and performs feature extraction, classification, and training through the amplitude sensing module and the phase sensing module.

[0064] As another example, combined Figure 2 , Figure 3 and Figure 4 The WIFI fall detection algorithm combining human posture assessment in this embodiment of the invention is implemented through the following steps:

[0065] 1. Data Preprocessing

[0066] Because the amplitude of Channel State Information (CSI) is affected by environmental noise interference, and the phase data exhibits random phase shifts, this invention processes the CSI data in two ways. First, the amplitude data extracted from the CSI is filtered; second, the phase data of the CSI is calibrated. Furthermore, this invention utilizes multi-cascaded attitude sensors to collect crucial attitude information related to human activity: acceleration and orientation.

[0067] (1) CSI amplitude data preprocessing

[0068] Due to the influence of WiFi network card hardware and the physical environment, the amplitude measurements in the raw CSI measurement data are usually very noisy. Noise reduction is necessary to minimize the overfitting problem caused by noise in the subsequent feature extraction model and to obtain more accurate signal features of actions such as falls. The specific CSI amplitude preprocessing process is as follows:

[0069] ① Standardization processing: Perform layer-based standardization processing on the data of each transmit / receive antenna pair. At this time, the data matrix size is N. rt ×T w The antennas are arranged logarithmically, separating the number of subcarriers, and the resulting matrix is ​​of size N. rt ×N sc ×T w The data. Where N... rt N is the product of the number of receiving antennas and the number of transmitting antennas. sc T represents the number of subcarriers. w The length of the signal acquired within a specified time period. Therefore, the standardized amplitude data is:

[0070]

[0071] Among them, H amp and These are the amplitude matrices before and after normalization for a given channel, μ. amp and σ amp These are the mean and standard deviation of the magnitude matrix, respectively.

[0072] ② Filtering: Hampel filtering is performed on the subcarrier samples in each pair of transmit and receive antennas to eliminate peak values ​​and outliers. A sliding window of length 30 is used to calculate the difference between the data point on the subcarrier and the median value of the data in the window. If the difference exceeds the product of the threshold and the standard deviation, the median value of the data in this window will replace the outlier in the interval.

[0073] ③ Other processing: First, linear interpolation is performed on the standardized CSI data. This is done by estimating the numerical value based on the two nearest data points to the point to be interpolated in the data sequence. Then, linear regression is used to smooth the trend of CSI and zero-filling is performed on the CSI data.

[0074] (2) CSI amplitude data preprocessing

[0075] ① Eliminating Phase Error: Firstly, phase difference is sensitive to environmental changes. However, the original CSI phase data contains two main types of noise: carrier frequency offset (CFO) and sampling time offset (STO). The variance of the phase difference between two antenna pairs is the sum of the individual variances of each antenna. Different antennas in a WiFi network card share the same RF oscillator, resulting in identical time-varying phase offsets. Therefore, the phase difference between the transmitting and receiving antenna pairs can be calculated using the following formula, effectively eliminating phase error:

[0076]

[0077] Where, θ ij This represents the phase difference between the i-th transmit / receive antenna pair and the j-th transmit / receive antenna pair.

[0078] ② Eliminating subcarrier noise components: Because the phase information obtained by WiFi devices is not very reliable, the carrier frequencies between the transmitter and receiver cannot be perfectly synchronized. Therefore, subcarrier noise components can be eliminated using the following formula:

[0079]

[0080] in,

[0081]

[0082]

[0083]

[0084] Where, ψ K It is the phase after linear transformation calibration, K is the number of subcarriers, and Φ K It is the phase after unwinding, φ K δ is the original phase, β is the timing error offset (delay), β is the random phase offset caused by the device, Z is the random noise during the measurement process, and φ is the random phase offset caused by the device. m It is the phase of the m-th subcarrier, η K It is the subcarrier index, η m Let m be the m-th subcarrier, and N be the Fast Fourier Transform. After the transformation using the above formula, the changes in phase over time can reflect changes in motion.

[0085] (3) Attitude information collection

[0086] This invention uses a multi-cascaded attitude sensor to obtain the acceleration and angle of the human body when it moves from the multi-cascaded attitude sensing signals of the multi-cascaded attitude sensor, and finally obtains the real-time acceleration and direction of the human body movement through processing.

[0087] 2. Model Training

[0088] The algorithm model of this invention consists of an amplitude sensing module and a phase sensing module. The model first performs single-modal inference with CSI phase and CSI amplitude respectively. During the training phase, human posture information, namely the acceleration of the subject during the activity and the direction of human movement, is added during the training of the two single-modal models. Then, multimodal fusion is performed, and the prediction results are used to determine whether the subject has fallen.

[0089] Two unimodal multitasking learning modules possess two modes of CSI data, representing and identifying whether a subject has fallen, respectively. Given a set of N subjects, with M samples for each subject, the dataset is denoted as D. T .

[0090]

[0091] in, This represents the j-th sample from the i-th subject using the CSI amplitude sensing module. This represents the sample corresponding to the CSI phase sensing module, y ij The label indicates whether the subject has fallen. The goal of this invention is to detect whether the corresponding signal of the subject indicates a fall, and therefore it can be represented as:

[0092]

[0093] Wherein, Model(·) is the algorithm model of this invention.

[0094] To perform fall detection based on CSI phase and amplitude data, this invention needs to extract spatial features across subcarriers and temporal features across carriers. Therefore, this invention uses a model called WiFi-LRCN in the amplitude sensing and phase sensing modules to extract temporal features from the data, and adds an attention mechanism to improve the model's understanding of the data. After parameter tuning experiments, this invention updated its main blocks and related parameters, as shown in Table 1:

[0095] Table 1 LRCN Structure

[0096]

[0097] The WiFi-LRCN model is cited in this invention because it considers model training efficiency and reduces model complexity while preserving capacity. WiFi-LRCN includes convolutional blocks, residual blocks, batch normalization layers, ReLU layers, flattening layers, and fully connected layers. The CSI phase and amplitude input to the network first pass through a convolutional block with a 7×21 kernel, a 5×5 stride, and 64 channels. A batch normalization ReLU layer is used for better convergence. To extract features from deeper layers, this invention designs three convolutional blocks with 64, 128, and 512 channels respectively, each followed by a residual block. Compared to the classic ResNet-18, this invention has fewer parameters and fewer floating-point operations.

[0098] In WiFi-LRCN, convolutional layers and residual blocks are used to extract global features of falls. The mathematical formula for the residual block can be expressed as:

[0099] F(x A / P )=C(x A / P )+x A / P

[0100] Where x represents the input feature, F(·) is the residual block, and C(·) is the convolutional layer. As the depth of the deep neural network increases, the residual design alleviates the degradation problem of the deep neural network, and the batch normalization layer further solves the gradient vanishing problem and helps to obtain better performance. The ReLU layer is used as an activation layer to increase nonlinearity to obtain better model capacity.

[0101] This invention adds attention mechanisms to the WiFi-LRCN of the amplitude sensing module and the phase sensing module to enable them to better understand the data. By leveraging the similarity between the attention mechanism and the human observation mechanism, the attention mechanism tends to focus on important local information of the object when extracting features. Then, it combines information from different regions to form an overall impression of the observed object. Furthermore, the different levels of shallow feature information contained in the feature pool make the global feature extraction module more targeted when performing calculations, thereby establishing a more reliable global dependency relationship.

[0102] In the attention mechanism, the Queries, Keys, and Values ​​implementations first select feature maps processed by 1x1 convolutions as the mapping for Keys. This aims to minimize the difference between the key information during the attention operation and the input data. Then, the feature maps extracted after 1x1 convolutions are used as the mapping for Queries, allowing more abstract features to be considered during the weight update calculation with Keys. Finally, the feature maps with more abstract global information are mapped to Values, ensuring that they contain shallow global features at the starting point of the update process, providing a foundation for the entire global feature extraction process. The data is then input into a normalization layer, generating attention scores for each signal feature. The temporal average of two signal features is then calculated, and the corresponding attention scores are weighted to obtain the global human pose trajectory feature. The calculation process is as follows:

[0103]

[0104]

[0105] Where Q, K, and V represent matrix-form Queries, Keys, and Values, respectively, d k Indicates the dimension of the vector Keys.

[0106] See Figure 2 The structural diagrams of the amplitude sensing module and the phase sensing module are shown. Figure 3The diagram shows the structure of the human pose extractor in the amplitude sensing module and the phase sensing module. During the training phase, in the inference of the two single-modal models, to avoid the model learning environmental features from the CSI signal, this invention introduces a human pose extractor as an auxiliary task in both the amplitude sensing module and the phase sensing module. This helps the main feature extraction module extract the human pose features during activity, ignoring environmental features, thereby achieving cross-domain detection. The human pose extractors in the amplitude sensing module and the phase sensing module use the outputs of different stages of WiFi-LRCN as signal inputs. The amplitude sensing module uses human motion acceleration as the predicted output, and the phase sensing module uses human motion direction as the predicted output. The human pose extractor consists of three 1D deconvolutional blocks (upsampling the extracted features to the same temporal resolution as the acceleration / direction change) and two fully connected layers. Each 1D deconvolutional block contains three deconvolutional layers, followed by batch normalization, rectified linear unit activation, and residual connections. The present invention also uses skip connections by connecting the depth convolutional features in the WiFi-LRCN portions Res-layer_1, Res-layer_2 and Res-layer_3 of the two modules to the deconvolutional layers in the human pose extractor, which follows the UNet structure.

[0107] After extracting CSI amplitude and phase feature vectors through the amplitude sensing and phase sensing modules respectively, this invention concatenates the two feature vectors obtained by the amplitude sensing and phase sensing modules into a multimodal feature vector. This concatenation correlates the feature information of the two modalities, resulting in a higher-dimensional and more discriminative feature space. Then, a classification module is constructed using three fully connected layers and one sigmoid layer. The fall probability value output by the classification module is a number between 0 and 1; if the score exceeds 0.5, the subject is considered to have fallen.

[0108] During training, the loss function of this system consists of two parts. The first part, fall detection classification, uses the cross-entropy loss function to measure the difference between the true probability distribution and the predicted probability distribution, which is expressed as follows:

[0109]

[0110] in, Y represents the loss in fall detection classification. fall This represents the true classification result. This indicates the predicted classification result.

[0111] The auxiliary task (i.e., the human pose extractor) in both the amplitude sensing module and the phase sensing module uses the mean squared error loss function to measure the squared average difference between the model's predicted output and the true label. Their representations are as follows:

[0112]

[0113]

[0114] in, The squared value of the average difference between the acceleration change predicted by the auxiliary task model in the amplitude sensing module and the actual acceleration change; This measures the squared average difference between the direction change predicted by the auxiliary task model in the phase sensing module and the actual direction change. n represents the length of the change in acceleration and direction, and t∈(1,T) represents the time variable, where T represents the total duration. X acc and X deg These represent the actual changes in acceleration and direction, respectively. These represent the acceleration changes predicted by the auxiliary task model in the amplitude sensing module and the directional changes predicted by the auxiliary task model in the phase sensing module, respectively.

[0115] The total loss is expressed as follows:

[0116]

[0117] In addition, this experiment used two laptops with three external transmit / receive antennas as transmitters and receivers, both equipped with Intel 5300 network cards for communication. The distance between the transmitter and receiver was approximately 3 meters, and their height above the ground was approximately 1.5 meters to ensure a large environmental coverage.

[0118] Both devices ran Ubuntu 18.04 and used the linux csitool tool to collect CSI data at the receiving end. The network cards on both devices operated on the 2.4GHz band, with a signal sampling rate of 250 pkt / s, occupying a 40MHz bandwidth and using 30 subcarriers. A total of nine action categories were collected: backward fall, bending over, forward fall, jumping, lying down, remaining still, side fall, sitting down, and standing up.

[0119] This invention employs an algorithm to obtain a reliable and stable model, effectively improving its learning ability and making the learned model more robust and effective. To avoid overfitting, a 4-fold cross-validation method is used. The dataset is divided into training, validation, and test sets in a 14:3:3 ratio. The learning scheme is implemented in PyTorch, and the model is trained on an NVIDIA A40. The Adam optimizer is used to achieve better convergence. The batch size is set to 32, and the learning rate is 10. -3， The system was trained for a total of 30 epochs. The results were evaluated using accuracy (ACC), with cross-environment test ACCs of 94.5%, 95.4%, 98.2%, and 96.3%, and an average ACC of 96.1%.

[0120] In summary, the beneficial effects of the present invention are as follows:

[0121] (1) The present invention uses WiFi signals as a fall detection method. WiFi signals do not record image and audio information, nor do they require wearing other sensors, thus protecting the privacy of the test subject and facilitating the test subject's activities.

[0122] (2) This invention uses ordinary commercial WiFi devices to establish a fall detection environment without the need for expensive hardware devices and complicated deployment, which increases ease of use and practicality, and has good promotion and adaptability.

[0123] (3) This invention applies deep learning technology to fall detection and adds two types of information in human motion posture to the WiFi signal, namely acceleration and direction, which overcomes the problem that traditional machine learning methods are difficult to adapt to environmental changes and has the ability to have the same detection accuracy under cross-environment conditions.

[0124] Specific embodiments of the invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing can be advantageous.

[0125] It should be noted that all directional indications (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0126] In the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include at least one of that feature.

[0127] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0128] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0129] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A WIFI fall detection method combining human posture assessment, characterized in that, include: Collects WIFI signals and multi-cascaded attitude sensor signals when a person falls; Extract CSI data from the WIFI signal; Human posture information is extracted from the multi-cascaded posture sensing signals; The CSI data is preprocessed for amplitude and phase. The preprocessed CSI data is divided into training set, validation set and test set; The training set is imported into the algorithm model, and feature extraction and classification training are performed in conjunction with the human posture information. A fall recognition test was conducted using the test set. The algorithm model consists of an amplitude sensing module and a phase sensing module. The WiFi-LRCN model is used in the amplitude sensing module and the phase sensing module to extract the temporal features in the data, and an attention mechanism is added to improve the model's understanding of the data. The algorithm model performs single-modal inference using CSI phase and CSI amplitude respectively. During the training phase, human posture information is added to the two single-modal training processes, and then multimodal fusion is performed. Subsequently, the prediction results are used to determine whether the subject has fallen. The WiFi-LRCN model includes convolutional blocks, residual blocks, batch normalization layers, ReLU layers, flattening layers, and fully connected layers; the convolutional blocks and residual blocks are used to extract global features of falls; each convolutional block and residual block consists of three elements. The algorithm model uses the outputs of different stages in WiFi-LRCN as signal inputs for the human pose extractors in the amplitude sensing module and phase sensing module, respectively. The amplitude sensing module uses human motion acceleration as the predicted output, and the phase sensing module uses the human motion direction as the predicted output. The human pose extractor consists of three one-dimensional deconvolutional blocks and two fully connected layers. The algorithm model also uses skip connections by connecting the deep convolutional features in the three residual blocks of the WiFi-LRCN part in the amplitude sensing module and phase sensing module to the deconvolutional block in the human pose extractor. This connection follows the UNet structure.

2. The method according to claim 1, characterized in that, The preprocessing of the CSI data in terms of amplitude and phase includes: The amplitude data in the CSI data is standardized, filtered, and subjected to other processing. Phase errors and subcarrier noise components are eliminated from the phase data in the CSI data.

3. The method according to claim 2, characterized in that, Extracting human posture information from the multi-cascaded posture sensing signals includes: The acceleration and angle of the human body when it performs movements are obtained from the multi-cascade attitude sensing signals of the multi-cascade attitude sensor. By processing the acceleration and angle of the human body during its movements, the real-time acceleration and direction of the human body's movement are obtained as human posture information.

Images