A method, system, device and medium for detecting the fall of the elderly based on WiFi CSI

Abstract

The present application belongs to the technical field of behavior recognition and fall detection, and particularly relates to a fall detection method, system, device and medium for old people based on WiFi CSI, which collects the CSI sequence of the WiFi channel of the old people space, generates dynamic CSI data through preprocessing, screens candidate violent events relying on short-time motion energy, respectively extracts fall stage features and time-frequency features, obtains fall probability through a double-branch recognition model fusing the features, calculates a comprehensive score in combination with the weight of the incident area and the individual historical motion baseline, rechecks the standing and stationary state of the period after the event, and outputs the fall alarm in stages. The present application adopts non-contact detection, avoids the disadvantages of camera privacy leakage and poor compliance of wearable devices, can effectively distinguish between lying, sitting, bending and other similar actions, greatly reduces false positives, and is modularized and deployed on the edge without uploading original CSI data, has good privacy, and is suitable for home and institutional old-age care scenarios.

Description

Technical Field

[0001] This invention belongs to the field of behavior recognition and fall detection technology, specifically relating to a fall detection method, system, device and medium for elderly people based on WiFi CSI. Background Technology

[0002] Falls are a high-risk event in elderly care facilities, home-based care, and nursing settings. If a fall is not detected promptly, it can lead to prolonged contact with the ground, secondary injuries, hypothermia, dehydration, or other health risks. Therefore, elderly care settings require fall detection solutions that are capable of long-term operation, have low false alarm rates, low false negative rates, and are privacy-friendly.

[0003] Existing fall detection solutions include camera monitoring, wearable devices, ground pressure sensors, infrared sensors, and millimeter-wave radar. Camera monitoring raises privacy concerns and is difficult to use continuously in areas such as bedrooms and bathrooms. Wearable devices require active wearing and maintenance by the elderly, resulting in poor compliance among disabled or cognitively impaired seniors. Ground pressure sensors are complex to deploy and have limited coverage. Infrared sensors and ordinary human body sensors struggle to distinguish falls from similar movements such as sitting, lying down, or bending over.

[0004] WiFi CSI can detect human movement by utilizing the multipath variations of wireless signals during indoor propagation. Fall events typically have phased characteristics such as rapid descent, impact disturbance, and subsequent stillness or inability to get up. These characteristics manifest in the CSI sequence as short-term energy spikes, spectral distribution changes, and a low-motion state after the event. Therefore, WiFi CSI can be used to achieve non-contact fall detection.

[0005] However, existing CSI fall detection methods still suffer from high false alarm rates, insufficient differentiation of similar movements, and inadequate confirmation of post-event status. Especially in elderly care settings, actions such as quickly sitting down, lying down normally, bending over to pick up objects, and being moved with assistance by caregivers can all generate strong wireless channel disturbances. Judging solely based on instantaneous energy peaks can easily lead to false alarms. Furthermore, the risk implications differ across areas. For example, lying down in a bed or sofa area might be considered normal activity, while a rapid descent followed by prolonged stillness in a bathroom, doorway, or passageway area is more likely to indicate a fall risk. Therefore, it is necessary to propose a fall detection method that integrates fall phase characteristics, CSI time-frequency map characteristics, regional weight correction, historical baseline correction, and post-fall status confirmation. Summary of the Invention

[0006] To address the technical problems of existing CSI fall detection methods, this invention provides a fall detection method, system, device, and medium for elderly people based on WiFiCSI.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A fall detection method for elderly people based on WiFi CSI includes the following steps: S1. Channel status information (CSI) is continuously collected by WiFi transceivers deployed in the target elderly care space to obtain the original CSI sequence arranged in time. S2. Preprocess the original CSI sequence to remove invalid interference data and generate a dynamic CSI sequence that only represents human motion perturbation; S3. The short-term motion energy of the dynamic CSI sequence is calculated using a sliding time window, and candidate violent motion events are selected by combining the energy threshold and the energy rise slope. S4. For candidate violent motion events, extract three types of fall stage features: rapid fall features, impact disturbance features, and post-event stillness features. S5. Perform time-frequency transformation on the dynamic CSI sequence corresponding to the candidate violent motion event to generate a CSI time-frequency map, and extract the preset time-frequency features from the time-frequency map; S6. Input the fall stage features and time-frequency features into the fall recognition model with a dual-branch structure for feature fusion, and output the fall probability corresponding to the candidate event; S7. When the probability of falling reaches the preset judgment threshold, continuously monitor the target duration after the event occurs, and complete the subsequent status confirmation of the elderly person getting up, remaining still, and abnormal micro-movements. S8. Combining the fall probability with the subsequent status confirmation results, determine the fall event and selectively output fall alarm information based on the determination results.

[0008] The preprocessing operation of S2 includes at least one of the following: abnormal data packet removal, sampling time alignment, effective subcarrier filtering, CSI amplitude normalization, phase correction, static background elimination, and filtering noise reduction; the static background elimination is achieved by moving average, high-pass filtering, or principal component analysis.

[0009] The short-term motion energy in S3 is calculated by accumulating the squared difference between the dynamic CSI within the window and the mean of the window; the energy threshold is an adaptive threshold, generated by superimposing multiple times the standard deviation of the background energy on the mean of the background motion energy, and a multi-link channel synchronization disturbance verification condition is added when screening candidate violent motion events.

[0010] In S4, the rapid descent features include energy rise slope, Doppler frequency shift change, and high-frequency energy ratio; the impact disturbance features include CSI energy peak value, peak decay rate, and multi-link synchronization peak coefficient; and the post-event static features include static duration, human body micro-motion amplitude, and standing action identification mark.

[0011] The S5 generates a CSI time-frequency map using short-time Fourier transform or wavelet transform, and the extracted time-frequency features include Doppler frequency shift, spectral peak, time-frequency energy distribution coefficient, high-frequency energy ratio, and spectral entropy.

[0012] In the S6, the first branch of the fall recognition model receives dynamic CSI time-series data to extract time-series features, and the second branch receives CSI time-frequency maps to extract image spectral features. The two features are then fused through feature splicing or attention mechanisms and fed into a classifier to obtain the fall probability.

[0013] Before outputting the alarm information, a comprehensive score calculation step is also included. The formula for calculating the comprehensive fall score is: F=αPfall+βQimpact+γQstatic+δQregion+ηQhistory-ρQnormal In the formula, Pfall is the model output fall probability, Qimpact is the impact disturbance intensity, Qstatic is the degree of post-event static confirmation, Qregion is the risk correction term for the incident area, Qhistory is the correction term for the individual's historical action baseline, Qnormal is the probability of normal actions such as sitting, bending over, and lying down, and α, β, γ, δ, η, and ρ are preset weight coefficients; a three-level alarm threshold is set based on the comprehensive score, and three types of alarm information are output in a graded manner: suspected fall, confirmed fall, and high-risk fall.

[0014] The S7 subsequent status confirmation is as follows: if the movement energy is consistently below the static threshold during the target period, there is no standing up movement and only sporadic abnormal micro-movements, then a high risk of fall is identified; if the elderly person is detected to get up and resume normal activities, then it is marked as ordinary vigorous movement and no high-level alarm is triggered.

[0015] A fall detection system for elderly people based on WiFi CSI is provided. The system is used to execute a fall detection method for elderly people based on WiFi CSI, including a CSI acquisition module, a preprocessing module, a candidate event detection module, a fall stage feature extraction module, a time-frequency feature extraction module, a fall recognition module, a comprehensive scoring module, a regional and historical baseline correction module, a subsequent status confirmation module, and an alarm module. The data output of the CSI acquisition module is connected to the input of the preprocessing module. The preprocessing module outputs a dynamic CSI sequence to the candidate event detection module. One end of the candidate event detection module is connected to the fall phase feature extraction module, and the other end is connected to the time-frequency feature extraction module. The feature outputs of the fall phase feature extraction module and the time-frequency feature extraction module are connected to the fall recognition module. The fall recognition module outputs the fall probability to the comprehensive scoring module. The regional and historical baseline correction module inputs regional risk and historical baseline correction parameters to the comprehensive scoring module. The comprehensive scoring module outputs the scoring result to the subsequent status confirmation module. The status determination data of the subsequent status confirmation module is sent back to the comprehensive scoring module to complete the secondary correction. The comprehensive scoring module finally links with the alarm module to output an alarm. All modules are integrated and deployed on the edge computing device. The original CSI data is locally stored, and only the alarm data and statistical feature data are transmitted externally.

[0016] An electronic device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements a fall detection method for elderly people based on WiFi CSI.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements all the steps of a WiFi CSI-based fall detection method for elderly people.

[0018] Compared with the prior art, the beneficial effects of this invention are: 1. This invention uses WiFi CSI for non-contact fall detection, which eliminates the need for cameras and wearable devices, reducing privacy intrusion and dependence on wearing devices. It is suitable for privacy-sensitive elderly care scenarios such as bedrooms, bathrooms, and nursing beds.

[0019] 2. This invention divides fall events into stages such as rapid descent, impact disturbance, and post-event stillness, and integrates CSI time series features and CSI time-frequency graph features, which can enhance the interpretability and accuracy of fall recognition and improve the ability to distinguish similar actions such as sitting down quickly, lying down normally, and bending over to pick up objects.

[0020] 3. This invention introduces a fall status confirmation mechanism, which continues to determine whether there is a prolonged state of stillness, failure to get up, or abnormal micro-movement after detecting a candidate violent movement event, thereby reducing false alarms caused by short-term violent movements.

[0021] 4. This invention combines spatial area correction and individual historical movement baseline correction to improve fall risk assessment results. It can adapt to different elderly care scenarios such as bed area, chair area, toilet area, doorway area, and passageway area, as well as the differences in movement habits of different elderly people, reducing false alarms or missed alarms caused by regional and individual differences. Attached Figure Description

[0022] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0023] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0024] Figure 1 This is a schematic diagram of the detection method of the present invention; Figure 2 This is a flowchart of the detection method of the present invention.

[0025] Wherein: A represents the target elderly care space; B represents edge processing; C represents remote analysis; D represents abnormal results; B1 represents CSI data collection; B2 represents preprocessing; B3 represents candidate event detection; B4 represents the extraction of fall stage features and time-frequency features; B5 represents the output of fall probability and comprehensive score; C1 represents the visualization of event statistics; C2 represents the correction of historical baselines; C3 represents the comprehensive score; C4 represents the alarm management and push notification; D1 represents the monitoring / management end; D2 represents the family end; and D3 represents the caregiver end. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. These descriptions are only for further illustrating the features and advantages of the present invention, and not for limiting the claims of the present invention. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] This invention provides a fall detection method for elderly people based on WiFi CSI, comprising the following steps: Step 1: Collect WiFi Channel Status Information (CSI) sequences within the target senior living space. Specifically, by setting up WiFi signal transmitters and receivers within the target senior living space, continuously acquire changes in channel amplitude, phase, or subcarrier response caused by the activities of the elderly residents, forming a time-ordered CSI sequence.

[0029] Step 2: Preprocess the CSI sequence to obtain a dynamic CSI sequence. Specifically, perform one or more of the following processes on the CSI sequence: outlier removal, time alignment, subcarrier filtering, amplitude normalization, phase correction, static background removal, and filtering noise reduction, to obtain a dynamic CSI sequence that primarily characterizes human motion disturbances.

[0030] Step 3: Detect candidate violent motion events based on the dynamic CSI sequence. Specifically, the system calculates short-term motion energy based on a sliding time window, and combines this with energy thresholds, energy rise slopes, multi-link disturbance synchronicity, or event duration to segment candidate violent motion events from the dynamic CSI sequence.

[0031] Step 4: Extract fall phase features from the candidate violent motion events. Specifically, the system extracts rapid fall features to characterize the rapid descent of the human body, impact disturbance features to characterize the contact process between the human body and the ground or object, and post-event stillness features to characterize the state of stillness, not getting up, or abnormal micro-movement after the event.

[0032] Step 5: Generate a CSI time-frequency map and extract time-frequency features based on the candidate violent motion events. Specifically, perform a short-time Fourier transform or wavelet transform on the dynamic CSI sequence corresponding to the candidate violent motion events to obtain the CSI time-frequency map, and extract one or more time-frequency features from the CSI time-frequency map, including Doppler frequency shift, spectral peak, time-frequency energy distribution, high-frequency energy ratio, and spectral entropy.

[0033] Step 6: Input the fall stage features and the time-frequency features into the fall recognition model to obtain the fall probability. Specifically, the fall recognition model is used to process the dynamic CSI time series and CSI time-frequency map respectively, and the two types of features are fused to output the fall probability corresponding to the candidate violent movement event.

[0034] Step 7: When the fall probability meets the preset conditions, perform subsequent status confirmation on the target time period after the candidate strenuous movement event. Specifically, determine whether the target elderly person has been in a static state for more than a preset duration, whether no standing up movement was detected, whether there is an abnormal micro-movement state, or whether normal activity has resumed.

[0035] Step 8: Determine the fall alarm result based on the fall probability and subsequent status confirmation results, and output the fall alarm information when the alarm conditions are met.

[0036] In one optional implementation, before outputting fall alarm information, the system generates a comprehensive fall score based on the fall probability, impact disturbance intensity, post-event static confirmation degree, regional risk correction item, individual historical action baseline correction item, and normal similar action probability, and determines the alarm level based on the comprehensive fall score.

[0037] The present invention also provides an electronic device and a computer-readable storage medium for performing the above-described WiFiCSI-based fall detection method for the elderly.

[0038] Example 1: System Deployment Environment See Figure 1 and Figure 2 In one embodiment, a WiFi transmitter and a WiFi receiver are installed within the target elderly care space. The WiFi transmitter can be a router or access point, and the WiFi receiver can be a terminal device that supports CSI data acquisition. The receiver continuously acquires CSI sequences and transmits the data to an edge computing device for processing.

[0039] Targeted senior living spaces can include one or more of the following: senior bedroom, senior living facility room, home care room, nursing bed area, bathroom area, corridor area, entrance area, living room area, or activity room area. For larger senior living spaces, multiple WiFi transmission and reception links can be set up to cover the main activity areas of the seniors.

[0040] The acquired CSI sequence can be represented as: Where t represents the sampling time, k represents the subcarrier number, m represents the transmitting antenna number, n represents the receiving antenna number, |H| represents the CSI amplitude, and φ represents the CSI phase.

[0041] Example 2: CSI Preprocessing During the method execution, the system first preprocesses the CSI sequence. Preprocessing includes removing abnormal data packets, aligning sampling times, selecting stable subcarriers, normalizing the CSI amplitude, expanding and linearly correcting the CSI phase, removing static background components, and filtering for noise reduction.

[0042] Abnormal data packets can be removed based on sudden changes in CSI amplitude, missing timestamps, abnormal sampling intervals, or link quality anomalies. Static background components can be removed using moving averages, high-pass filtering, principal component analysis, or background template subtraction. The preprocessed dynamic CSI sequence primarily reflects wireless channel changes caused by human movement.

[0043] Example 3: Detection of Candidate Vigorous Exercise Events During the candidate event detection phase, the system uses a sliding window to calculate short-time motion energy: Ew(t)=Στ∈WΣkΣmΣn|X(τ,k,m,n)-μW(k,m,n)| 2 Where W represents the sliding time window, X represents the preprocessed dynamic CSI sequence, and μW represents the mean CSI feature within the window.

[0044] If Ew(t) exceeds the adaptive energy threshold and the energy rise slope exceeds the preset slope threshold, the system marks that time period as a candidate violent motion event. The adaptive energy threshold can be determined based on the background energy mean and background energy standard deviation, for example, set as the sum of the background energy mean and a certain multiple of the background energy standard deviation.

[0045] In one alternative implementation, candidate violent motion events can also be confirmed by combining conditions such as whether multiple CSI links experience synchronous disturbances within the same time window, whether the proportion of high-frequency energy within the target time window exceeds a preset ratio, and whether the duration of the event is within the preset fall candidate time range.

[0046] Example 4: Feature Extraction of the Fall Stage Fall events typically consist of several phases. The first phase is the rapid descent phase, during which the body's center of gravity drops rapidly, and the principal components of the CSI, Doppler shift, and high-frequency energy percentage change rapidly. The second phase is the impact disturbance phase, during which the body comes into contact with the ground or other objects, and a short-term spike in CSI kinetic energy occurs. The third phase is the post-event state phase, during which the elderly person may be stationary, unable to get up, or exhibit only slight micro-movements.

[0047] The system extracts fall phase features from candidate violent motion events. Features of rapid falls may include the slope of kinetic energy rise, the direction of principal component changes, the duration of the action, changes in the spectral centroid, changes in Doppler frequency shift, the proportion of high-frequency energy, and changes in the direction of human motion.

[0048] Impact disturbance characteristics may include short-term energy peak value of CSI, peak duration, peak width, energy mutation rate, multi-link synchronization peak value, spectral width, spectral entropy, and energy decay rate after impact. Post-event static characteristics may include mean kinetic energy, kinetic energy variance, static duration, standing motion detection results, posture maintenance characteristics, and human micro-motion characteristics.

[0049] Example 5: CSI Time-Frequency Plot and Time-Frequency Characteristics In the time-frequency feature extraction stage, the system performs short-time Fourier transform or wavelet transform on the dynamic CSI sequences corresponding to candidate violent motion events to obtain the CSI time-frequency map. Fall events in the time-frequency map typically show a short period of concentrated energy enhancement, increased spectral width, followed by a transition to a low-energy static region.

[0050] The system extracts features such as Doppler frequency shift, time-frequency energy distribution, spectral peak value, high-frequency energy proportion, low-frequency static characteristics, and spectral entropy from the CSI time-frequency map. These time-frequency features are used to distinguish falls from non-falling actions that have similar instantaneous energy changes, such as sitting down quickly, lying down normally, or bending over to pick up an object.

[0051] Example 6: Fall Recognition Model The fall detection model can employ a two-branch structure. The first branch takes a dynamic CSI time series as input to extract time-varying features; the second branch takes a CSI time-frequency map as input to extract spectral and time-frequency energy distribution features. The outputs of the two branches are fused through an attention fusion module or a feature concatenation module and then input into a classifier to obtain the fall probability Pfall.

[0052] In one alternative implementation, the first branch may employ a convolutional neural network, a recurrent neural network, a gated recurrent network, a long short-term memory network, a temporal convolutional network, or a Transformer network; the second branch may employ a convolutional neural network, a visual Transformer network, or a lightweight image classification network.

[0053] Example 7: Fall Assessment In one embodiment, the fall assessment score F can be expressed as: F=αPfall+βQimpact+γQstatic+δQregion+ηQhistory-ρQnormal Where Pfall represents the probability of falling from the model output, Qimpact represents the impact disturbance intensity, Qstatic represents the degree of post-event static confirmation, Qregion represents the regional risk correction term, Qhistory represents the individual historical action baseline correction term, Qnormal represents the probability of similar actions such as sitting down normally, lying down normally, bending over or picking up objects, and α, β, γ, δ, η and ρ are weight coefficients.

[0054] When the fall score F is greater than or equal to the first alarm threshold, the system outputs a suspected fall alarm; when the fall score F is greater than or equal to the second alarm threshold and no getting up action is detected within the target time period after the event, the system outputs a confirmed fall alarm; when the fall score F is greater than or equal to the third alarm threshold and the duration of stillness after the event exceeds the preset time, the system outputs a high-risk fall alarm.

[0055] Example 8: Subsequent Status Confirmation In the subsequent status confirmation phase, the system continues to monitor the target time period after the candidate violent movement event. If no standing up movement is detected within the target time period, and the kinetic energy remains below the resting threshold, or only abnormal micro-movements are detected without significant displacement, the system confirms a high risk of fall.

[0056] If the system detects that an elderly person quickly gets up and resumes normal activity, it can record the event as a suspected fall or violent movement, without triggering a high-level alarm. Through subsequent status confirmation, this invention can avoid directly classifying short-term violent movements as falls, thereby reducing the false alarm rate.

[0057] Example 9: Region Correction and Similar Action Differentiation During the area correction phase, the system determines the area where the event occurred based on CSI positioning or area fingerprint recognition. If the event occurs in the bed area, chair area, or sofa area, and the movement pattern matches normal lying down or sitting, the fall probability or fall score is reduced. If the event occurs in the toilet area, passageway area, or doorway area, and is accompanied by rapid falling, impact disturbance, and subsequent stillness, the fall alarm level is increased.

[0058] During the similarity action differentiation phase, the system distinguishes candidate vigorous movement events from one or more actions such as sitting down normally, lying down normally, bending over, turning quickly, picking up objects, and nursing-assisted movement. If the probability of a candidate event being similar to one of the above normal actions is high, the fall score is reduced using Qnormal.

[0059] Example 10: Individual Historical Movement Baseline Correction In another embodiment, the system may incorporate an individual's historical movement baseline for correction. This baseline may include normal sitting speed, normal lying duration, daily activity intensity, post-event recovery time, or historical suspected fall records.

[0060] If an elderly person usually sits down quickly, the system can reduce false alarms for rapid sitting down by using historical baselines. If an elderly person usually moves slowly, but experiences a short period of strong impact followed by a prolonged period of stillness, the system will increase the alarm level.

[0061] Example 11: Alarm Output Alarm information may include the time of the incident, the area where the incident occurred, the probability of a fall, the overall fall score, the duration of stillness after the incident, whether any standing motion was detected, whether any subtle human movements were detected, the alarm level, and the recommended handling method.

[0062] For example, the system can categorize alarm levels into suspected falls, confirmed falls, and high-risk falls. Suspected falls alert caregivers; confirmed falls send real-time alarms; and high-risk falls trigger emergency procedures when prolonged stillness or abnormal micro-movements are detected.

[0063] Example 12: Edge Deployment and Privacy Protection This invention can be deployed in edge computing devices to perform CSI processing, fall detection, comprehensive scoring, and alarms in real time. The system can upload only the event time, alarm level, fall comprehensive score, and feature summary, without uploading the original CSI data or video images, thereby reducing privacy risks.

[0064] In one alternative implementation, the edge computing device can also cache low-confidence events as verification events, which can be manually verified by caregivers or the elderly care monitoring platform.

[0065] Example 13: System Module This invention also provides a fall detection system for elderly people based on WiFi CSI, including a CSI acquisition module, a preprocessing module, a candidate event detection module, a fall stage feature extraction module, a time-frequency feature extraction module, a fall recognition module, a comprehensive scoring module, a regional and historical baseline correction module, a subsequent status confirmation module, a regional correction module, and an alarm module.

[0066] The CSI acquisition module collects WiFi channel status information (CSI) sequences within the target elderly care space. The preprocessing module preprocesses the CSI sequences to obtain dynamic CSI sequences. The candidate event detection module calculates short-term motion energy based on the dynamic CSI sequences and detects candidate violent motion events. The fall phase feature extraction module extracts rapid fall features, impact disturbance features, and post-event stationary features. The time-frequency feature extraction module generates a CSI time-frequency map and extracts time-frequency features. The fall identification module calculates the fall probability based on fall phase features and time-frequency features. The comprehensive scoring module generates a comprehensive fall score. The subsequent status confirmation module confirms the target time period after the event. The area and historical baseline correction module corrects for risk based on the event location and individual historical action baselines. The alarm module outputs fall alarm information.

[0067] Example 14: Electronic device and storage medium The present invention also provides an electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the aforementioned fall detection method for elderly people based on WiFi CSI. The electronic device may be an edge computing device, a server, an elderly care monitoring terminal, an embedded device, or a combination thereof.

[0068] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned fall detection method for elderly people based on WiFi CSI. The computer-readable storage medium may include a read-only memory, a random access memory, a magnetic disk, an optical disk, flash memory, or other media capable of storing program code.

[0069] The above description only illustrates the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. A fall detection method for elderly people based on WiFi CSI, characterized in that: Includes the following steps: S1. Channel status information (CSI) is continuously collected by WiFi transceivers deployed in the target elderly care space to obtain the original CSI sequence arranged in time. S2. Preprocess the original CSI sequence to remove invalid interference data and generate a dynamic CSI sequence that only represents human motion perturbation; S3. The short-term motion energy of the dynamic CSI sequence is calculated using a sliding time window, and candidate violent motion events are selected by combining the energy threshold and the energy rise slope. S4. For candidate violent motion events, extract three types of fall stage features: rapid fall features, impact disturbance features, and post-event stillness features. In S4, rapid fall features include energy rise slope, Doppler frequency shift change, and high-frequency energy ratio; impact disturbance features include CSI energy peak value, peak decay rate, and multi-link synchronization peak coefficient; post-event stillness features include stillness duration, human micro-movement amplitude, and standing action recognition marker. S5. Perform time-frequency transformation on the dynamic CSI sequence corresponding to the candidate violent motion event to generate a CSI time-frequency map, and extract the preset time-frequency features from the time-frequency map; S6. Input the fall stage features and time-frequency features into the fall recognition model with a dual-branch structure for feature fusion, and output the fall probability corresponding to the candidate event; S7. When the probability of falling reaches the preset judgment threshold, continuously monitor the target duration after the event occurs, and complete the subsequent status confirmation of the elderly person getting up, remaining still, and abnormal micro-movements. Before outputting the alarm information, a comprehensive score calculation step is also included. The formula for calculating the comprehensive fall score is: F=αPfall+βQimpact+γQstatic+δQregion+ηQhistory-ρQnormal In the formula, Pfall is the model output fall probability, Qimpact is the impact disturbance intensity, Qstatic is the degree of post-event static confirmation, Qregion is the risk correction term for the incident area, Qhistory is the correction term for the individual's historical action baseline, Qnormal is the probability of normal actions such as sitting, bending over, and lying down, and α, β, γ, δ, η, and ρ are preset weight coefficients; a three-level alarm threshold is set based on the comprehensive score, and three types of alarm information are output in a graded manner: suspected fall, confirmed fall, and high-risk fall; S8. Combining the fall probability with the subsequent status confirmation results, determine the fall event and selectively output fall alarm information based on the determination results.

2. The fall detection method for elderly people based on WiFi CSI according to claim 1, characterized in that: The preprocessing operation of S2 includes at least one of the following: abnormal data packet removal, sampling time alignment, effective subcarrier filtering, CSI amplitude normalization, phase correction, static background elimination, and filtering noise reduction; the static background elimination is achieved by moving average, high-pass filtering, or principal component analysis.

3. The fall detection method for elderly people based on WiFi CSI according to claim 1, characterized in that: The short-term motion energy in S3 is calculated by accumulating the squares of the difference between the dynamic CSI within the window and the mean of the window; the energy threshold is an adaptive threshold, generated by superimposing multiple times the standard deviation of the background energy on the mean of the background motion energy, and a multi-link channel synchronization disturbance verification condition is added when screening candidate violent motion events.

4. The fall detection method for elderly people based on WiFi CSI according to claim 1, characterized in that: The S7 subsequent status confirmation is as follows: if the movement energy is consistently below the static threshold during the target period, there is no standing up movement and only sporadic abnormal micro-movements, then a high risk of fall is identified; if the elderly person is detected to get up and resume normal activities, then it is marked as ordinary vigorous movement and no high-level alarm is triggered.

5. A fall detection system for elderly people based on WiFi CSI, characterized in that: The system is used to execute the fall detection method for elderly people based on WiFi CSI as described in any one of claims 1-4, including a CSI acquisition module, a preprocessing module, a candidate event detection module, a fall stage feature extraction module, a time-frequency feature extraction module, a fall recognition module, a comprehensive scoring module, a regional and historical baseline correction module, a subsequent status confirmation module, and an alarm module. The data output terminal of the CSI acquisition module is connected to the input terminal of the preprocessing module, and the preprocessing module outputs a dynamic CSI sequence to the candidate event detection module. The candidate event detection module is connected to the fall stage feature extraction module at one end and the time-frequency feature extraction module at the other end; the feature outputs of the fall stage feature extraction module and the time-frequency feature extraction module are connected to the fall recognition module; the fall recognition module outputs the fall probability to the comprehensive scoring module, and the regional and historical baseline correction module inputs regional risk and historical baseline correction parameters to the comprehensive scoring module; The comprehensive scoring module outputs the scoring results to the subsequent status confirmation module; the status judgment data of the subsequent status confirmation module is sent back to the comprehensive scoring module for secondary correction, and the comprehensive scoring module finally links with the alarm module to output an alarm; all modules are integrated and deployed on the edge computing device, the original CSI data is locally sealed, and only alarm data and statistical feature data are transmitted externally.

6. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, characterized in that: When the computer program is executed by the processor, it implements the fall detection method for elderly people based on WiFi CSI as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements all the steps of the fall detection method for elderly people based on WiFi CSI as described in any one of claims 1-4.

Images