Method and system for detecting abnormal behavior of accidental fall of a person in a bathing situation based on millimeter wave radar

Abstract

This invention discloses a method and system for detecting accidental falls during bathing using millimeter-wave radar. The method includes: performing layered preprocessing on the raw signals acquired by the millimeter-wave radar; suppressing dynamic water flow interference through static clutter filtering, water flow feature recognition, and spatiotemporal joint filtering; extracting multi-dimensional physical features such as height trajectory, velocity and acceleration, attitude deformation, scattering point density, and micro-motions from the processed signals in parallel; constructing a decision engine based on multi-level physical rules, sequentially triggering height drops, verifying velocity changes, confirming attitude deformation, and filtering for false alarms; adaptively adjusting the decision threshold and observation time according to the real-time water flow interference intensity; and outputting a fall alarm when all rule conditions are met. This invention does not rely on machine learning models, exhibits strong robustness, high interpretability, and low computational complexity, and achieves reliable, real-time, and non-invasive detection of falls during bathing in environments with strong water flow interference.

Description

Technical Field

[0001] This invention relates to the field of non-contact sensing and intelligent monitoring technology, specifically to a method and system for detecting abnormal behavior such as accidental falls during bathing based on millimeter-wave radar. Background Technology

[0002] Intelligent health monitoring technology, especially proactive safety warnings in high-risk home environments, has become a research focus in the field of smart elderly care and health management. Bathrooms, particularly bathing areas, are among the most dangerous areas for falls among the elderly due to slippery floors, limited space, obstructed views (water mist), and high privacy requirements. The consequences are often severe, necessitating reliable and timely non-contact abnormal behavior detection methods.

[0003] Traditional fall detection technologies primarily include solutions based on vision, wearable sensors, and environmental sensing. While camera-based vision solutions can provide rich posture information, they are difficult for users to accept in highly private settings like bathing, and water vapor can severely obscure the lens, causing malfunction. Wearable devices based on accelerometers and gyroscopes (such as smartwatches and belts) require continuous wear, leading to poor user compliance during bathing and a risk of water damage. Environmental sensors based on sound, vibration, or pressure from carpets are easily affected by continuous water flow, water droplet impact, or humid environments, generating numerous false alarms and lacking reliability.

[0004] Millimeter-wave radar technology, due to its non-contact nature, privacy protection, strong penetration (through clothing, water mist, and non-metallic partitions), and high sensitivity to subtle movements, has shown great potential in the field of human behavior perception in recent years and is considered an ideal choice for solving the aforementioned challenges. Existing research has confirmed the feasibility of using millimeter-wave radar for fall detection in open spaces or bedroom environments. Related methods mostly rely on extracting motion trajectories, velocity profiles, or Doppler features from radar signals and classifying them using threshold judgments or machine learning models.

[0005] However, directly applying existing millimeter-wave radar fall detection solutions to bathing environments faces a core, yet unresolved challenge: strong, dynamic water flow interference. During a shower, the continuous water flow impacts the body and the floor, generating complex micro-Doppler signals and dynamic point clouds that highly overlap with the human motion spectrum. This interference is not simple static noise, but a strong background signal with time-varying and spatially distributed characteristics. It severely overwhelms or distorts key motion features characterizing falls (such as sudden drops in height and abrupt changes in velocity), leading to a sharp decline in the performance of existing fall detection methods based on threshold decisions or general machine learning models, and a significant increase in both false alarm and false negative rates.

[0006] Existing technologies lack systematic modeling and suppression strategies for the physical characteristics of water flow interference, and also lack stable and effective feature extraction and decision-making logic under such interference. To overcome this problem, there is an urgent need for a millimeter-wave radar fall detection method specifically designed for bathing environments and robust to strong water flow interference. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method and system for detecting abnormal behavior of accidental falls in bathing environments based on millimeter-wave radar. This method suppresses water flow interference through layered preprocessing, extracts multi-dimensional physical features in parallel, and adopts a multi-level physical rule decision engine, thus solving the problems of low accuracy and high false alarm rate in fall detection in bathing environments.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for detecting abnormal behavior of accidental falls during bathing based on millimeter-wave radar, comprising the following steps: The raw signals acquired by millimeter-wave radar are preprocessed in layers to suppress water flow interference, resulting in point cloud data mainly composed of human body reflections. Multi-dimensional physical features are extracted in parallel from the point cloud data, including height trajectory, velocity and acceleration, attitude deformation, scattering point density and micro-motion; A decision engine based on multi-level physical rules is constructed. According to the multi-dimensional physical characteristics, the engine sequentially performs the following judgments: height drop trigger judgment, velocity change verification judgment, attitude deformation confirmation judgment, and time sequence correlation and false alarm filtering judgment. Real-time assessment of water flow disturbance intensity ,when When the intensity threshold is exceeded, it automatically enters the high robustness mode, increases the judgment threshold at each level proportionally, and extends the observation time; When all the conditions for judgment are met, a fall confirmation alarm is output.

[0009] Furthermore, the layered preprocessing further includes: For each frame of the original ADC signal, perform a one-dimensional fast Fourier transform and a slow-time fast Fourier transform sequentially to generate a two-dimensional data matrix containing distance and Doppler information. A dynamic target display technique based on recursive averaging is adopted to dynamically subtract the static background component from the current frame signal in order to separate the dynamic target signal; Analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals to automatically identify and mark the interference areas dominated by the water flow generated by the shower head's falling water and its splashing. Spatial weight attenuation is applied to signals marked as water flow-dominant regions, and an adaptive band-stop filter is designed based on the identified water flow characteristic Doppler frequencies to suppress water flow interference components in the frequency domain, resulting in point cloud data dominated by human body reflections.

[0010] Furthermore, the water flow feature identification and labeling identifies and labels the dominant water flow area by calculating the energy and variance within the region. : If a certain area satisfy (Energy threshold) and (Velocity variance threshold) is then marked as the dominant flow zone. .

[0011] If a certain region satisfies and Then it is marked as the dominant water flow area, where Energy threshold This is the velocity variance threshold.

[0012] in, This represents the energy within the region Ω; This represents the value of the dynamic target matrix at coordinates (r, v) in the k-th frame. Represents the distance unit in the distance-Doppler spectrum; This represents the velocity element in the range-Doppler spectrum; Indicates the monitoring area; This indicates the average speed within the monitored area; It indicates the degree of dispersion of velocity distribution within the monitoring area; Furthermore, the posture deformation feature extraction specifically involves: performing principal component analysis on the three-dimensional coordinates of the human body point group, calculating the eigenvalues ​​of its covariance matrix, and calculating the ratio of the maximum eigenvalue to the minimum eigenvalue as the posture extension ratio. .

[0013] Furthermore, the posture extension ratio The calculation formula is: in, Let represent the largest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; Let represent the smallest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; To prevent division by zero for an extremely small positive number; When standing, the human body extends in the vertical direction (first principal component). Larger; When falling and landing in a lying / prone position, the body unfolds horizontally, with similar dimensions in all three dimensions. Close to 1.

[0014] Furthermore, the scattering point density and micro-motion analysis specifically involves calculating the density ratio of the number of human body point clouds to the total number of point clouds. Furthermore, the entropy value was calculated from the local Doppler spectrum to analyze the micro-Doppler entropy abrupt changes caused by unconscious limb swinging.

[0015] Furthermore, the decision-making process of the decision engine based on multi-level physical rules includes: Level 1 altitude drop trigger: Determines whether the altitude drop exceeds a preset threshold; Level 2 velocity mutation verification: Determine whether there is a velocity peak exceeding the peak threshold during the descent and whether the velocity rapidly drops below the resting threshold after impact; Level 3 Attitude Deformation Confirmation: After the velocity mutation verification, determine whether the attitude stretch ratio decreases and remains below the low attitude threshold, and whether the duration exceeds the static confirmation time. Level 4 temporal correlation and false alarm filtering: comprehensively judge whether the entire event sequence conforms to the temporal logic of "instability, fall, impact, and stillness", and exclude the situation of large-scale autonomous movement in a short period of time after the impact.

[0016] Furthermore, the intensity of the water flow interference The dynamic total energy is estimated in real time by the ratio of the energy in the dominant water flow zone, and the formula is as follows: in, This indicates the intensity of water flow disturbance estimated by the system in real time; The total dynamic energy. Energy in the water flow-dominant zone.

[0017] The present invention provides an abnormal behavior detection system for accidental falls during bathing based on millimeter-wave radar, comprising: The layered preprocessing module is used to perform layered preprocessing on the raw signals acquired by the millimeter-wave radar, suppress water flow interference, and obtain point cloud data mainly based on human body reflections. The feature extraction module is used to extract multi-dimensional physical features in parallel from the point cloud data. The multi-dimensional physical features include height trajectory, velocity and acceleration, attitude deformation, scattering point density, and micro-motion. The decision engine module is used to construct a decision engine based on multi-level physical rules. According to the multi-dimensional physical characteristics, it sequentially performs judgments on sudden height drop triggers, velocity change verifications, attitude deformation confirmations, and temporal correlation and false alarm filtering, and evaluates the intensity of water flow interference in real time. ,when When the intensity threshold is exceeded, it automatically enters the high robustness mode, increases the judgment threshold at each level proportionally, and extends the observation time; The alarm output module is used to output a fall confirmation alarm when all the judgment conditions are met.

[0018] Furthermore, the layered preprocessing module further includes: The range-Doppler analysis unit is used to sequentially perform one-dimensional fast Fourier transform and slow-time-dimensional fast Fourier transform on each frame of the original ADC signal to generate a two-dimensional data matrix containing range and Doppler information. The static clutter filtering unit is used to dynamically subtract the static background component from the current frame signal using a moving target display technique based on recursive averaging, so as to separate the moving target signal. The water flow feature recognition unit is used to analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals, and automatically identify and mark the water flow interference area dominated by the water flow generated by the shower head falling water flow and its splashing. The joint filtering unit is used to perform spatial weight attenuation on the signal marked as the water flow-dominant region, and to design an adaptive band-stop filter based on the identified water flow characteristic Doppler frequency to suppress water flow interference components in the frequency domain, thereby obtaining point cloud data dominated by human body reflection.

[0019] The beneficial effects of this invention are as follows: This invention provides a method and system for detecting accidental falls during bathing using millimeter-wave radar. The method includes: performing layered preprocessing on the raw signals acquired by the millimeter-wave radar; suppressing dynamic water flow interference through static clutter filtering, water flow feature recognition, and spatiotemporal joint filtering; extracting multi-dimensional physical features such as height trajectory, velocity and acceleration, attitude deformation, scattering point density, and micro-motions from the processed signals in parallel; constructing a decision engine based on multi-level physical rules, sequentially triggering sudden height drops, verifying sudden velocity changes, confirming attitude deformation, and filtering for false alarms; adaptively adjusting the decision threshold and observation time according to the real-time water flow interference intensity; and outputting a fall alarm when all rule conditions are met. This invention does not rely on machine learning models, possesses strong robustness, high interpretability, and low computational complexity, and achieves reliable, real-time, and non-invasive detection of falls during bathing in environments with strong water flow interference. Compared with existing technologies, it has the following advantages: 1. Strong resistance to water flow interference: Through static clutter filtering, water flow feature recognition and spatiotemporal joint filtering in the layered preprocessing, it can effectively suppress strong dynamic water flow interference in the bathing environment, significantly improve the purity of human body point cloud data, and solve the problem of detection failure caused by water flow echo masking in the shower scene in the existing method.

[0020] 2. High reliability and low false alarm rate: A decision engine based on multi-level physical rules is constructed to sequentially judge sudden drops in height, sudden changes in speed, and attitude deformation. False alarm filtering is combined with temporal logic. At the same time, the decision threshold and observation time are adaptively adjusted according to the intensity of real-time water flow interference, which enables accurate differentiation between falling behavior and non-falling behavior (such as bending over, sitting down, etc.), greatly reducing false alarms and missed alarms.

[0021] 3. Not dependent on machine learning models: The entire method adopts classical signal processing and rule reasoning, which does not require a large amount of labeled data for model training. This avoids the problems of high data collection costs and poor model generalization ability. It has high interpretability, strong robustness and low computational complexity, making it particularly suitable for resource-constrained embedded real-time systems.

[0022] 4. Non-invasive and privacy-protected: Based on millimeter-wave radar non-contact sensing technology, it does not require users to wear any devices and does not collect image or sound information. While achieving all-weather active monitoring, it strictly protects user privacy in high-privacy scenarios such as restrooms.

[0023] 5. Wide range of practical value: This invention provides a reliable, practical and low-cost solution for bathing safety monitoring in fields such as smart elderly care and healthy housing. It can effectively reduce the risk of failure to provide timely assistance to high-risk groups such as the elderly after a fall in the bathroom, and has significant social and economic benefits.

[0024] The above and other objects, advantages, and features of the present invention will be more fully set forth and demonstrated through the following detailed description of specific embodiments in conjunction with the accompanying drawings. Those skilled in the art, upon referring to the following detailed description and the accompanying drawings, will be able to better understand and realize the above advantages of the present invention. Other objects, features, and advantages of the present invention will become clearer after being described in detail in the detailed description section in conjunction with the accompanying drawings. Attached Figure Description

[0025] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0026] Figure 1 This is a flowchart of the core detection algorithm in this embodiment; Figure 2 This is a schematic diagram of the overall system deployment and architecture in this embodiment; Figure 3This is a schematic diagram of the test environment and radar deployment for the bathing scenario in this embodiment; Figure 4 This is an example of point cloud comparison before and after layered signal processing in this embodiment; Figure 5 This is an example diagram showing the multi-dimensional physical feature extraction results of this embodiment; Figure 6 This is a flowchart illustrating the logic of the multi-level rule decision engine in this embodiment. Figure 7 This is an example diagram of the final alarm output and host computer display interface of the system in this embodiment. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example 1

[0028] like Figure 1 As shown, Figure 1 The diagram shown is a core flowchart of the detection algorithm of the present invention. The abnormal behavior detection method and system for accidental falls in the bathing situation based on millimeter-wave radar provided in this embodiment is characterized by including the following steps: The raw signals acquired by millimeter-wave radar are preprocessed in layers to suppress water flow interference, resulting in point cloud data mainly composed of human body reflections. Multi-dimensional physical features are extracted in parallel from the point cloud data, including height trajectory, velocity and acceleration, attitude deformation, scattering point density and micro-motion; A decision engine based on multi-level physical rules is constructed. According to the multi-dimensional physical characteristics, the engine sequentially performs the following judgments: height drop trigger judgment, velocity change verification judgment, attitude deformation confirmation judgment, and time sequence correlation and false alarm filtering judgment. Real-time assessment of water flow disturbance intensity ,when When the intensity threshold is exceeded, it automatically enters the high robustness mode, increases the judgment threshold at each level proportionally, and extends the observation time; When all the conditions for judgment are met, a fall confirmation alarm is output.

[0029] The layered preprocessing described in this embodiment further includes: For each frame of the original ADC signal, perform a one-dimensional fast Fourier transform and a slow-time fast Fourier transform sequentially to generate a two-dimensional data matrix containing distance and Doppler information. A dynamic target display technique based on recursive averaging is adopted to dynamically subtract the static background component from the current frame signal in order to separate the dynamic target signal; Analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals to automatically identify and mark the interference areas dominated by the water flow generated by the shower head's falling water and its splashing. Spatial weight attenuation is applied to signals marked as water flow-dominant regions, and an adaptive band-stop filter is designed based on the identified water flow characteristic Doppler frequencies to suppress water flow interference components in the frequency domain, resulting in point cloud data dominated by human body reflections.

[0030] In this embodiment, the water flow feature identification and labeling is performed by calculating the energy and variance within the region to label the dominant water flow area. : If a certain area satisfy (Energy threshold) and (Velocity variance threshold) is then marked as the dominant flow zone. .

[0031] If a certain region satisfies and Then it is marked as the dominant water flow area, where Energy threshold This is the velocity variance threshold.

[0032] in, This represents the energy within the region Ω; This represents the value of the dynamic target matrix at coordinates (r, v) in the k-th frame. Represents the distance unit in the distance-Doppler spectrum; This represents the velocity element in the range-Doppler spectrum; Indicates the monitoring area; This indicates the average speed within the monitored area; It indicates the degree of dispersion of velocity distribution within the monitoring area; In this embodiment, the posture deformation feature extraction specifically involves: performing principal component analysis on the three-dimensional coordinates of the human body point group, calculating the eigenvalues ​​of its covariance matrix, and calculating the ratio of the maximum eigenvalue to the minimum eigenvalue as the posture extension ratio. .

[0033] The method according to claim 4, characterized in that the posture extension ratio The calculation formula is: in, Let represent the largest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; Let represent the smallest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; To prevent division by zero for an extremely small positive number; When standing, the human body extends in the vertical direction (first principal component). Larger; When falling and landing in a lying / prone position, the body unfolds horizontally, with similar dimensions in all three dimensions. Close to 1.

[0034] In this embodiment, the scattering point density and micro-motion analysis specifically involves calculating the density ratio of the human body point cloud to the total point cloud. Furthermore, the entropy value was calculated from the local Doppler spectrum to analyze the micro-Doppler entropy abrupt changes caused by unconscious limb swinging.

[0035] The decision-making process of the multi-level physical rule-based decision engine described in this embodiment includes: Level 1 altitude drop trigger: Determines whether the altitude drop exceeds a preset threshold; Level 2 velocity mutation verification: Determine whether there is a velocity peak exceeding the peak threshold during the descent and whether the velocity rapidly drops below the resting threshold after impact; Level 3 Attitude Deformation Confirmation: After the velocity mutation verification, determine whether the attitude stretch ratio decreases and remains below the low attitude threshold, and whether the duration exceeds the static confirmation time. Level 4 temporal correlation and false alarm filtering: comprehensively judge whether the entire event sequence conforms to the temporal logic of "instability, fall, impact, and stillness", and exclude the situation of large-scale autonomous movement in a short period of time after the impact.

[0036] The water flow interference intensity described in this embodiment The dynamic total energy is estimated in real time by the ratio of the energy in the dominant water flow zone, and the formula is as follows: in, This indicates the intensity of water flow disturbance estimated by the system in real time; The total dynamic energy. Energy in the water flow-dominant zone.

[0037] The abnormal behavior detection system for accidental falls during showering based on millimeter-wave radar provided in this embodiment is characterized by comprising: The layered preprocessing module is used to perform layered preprocessing on the raw signals acquired by the millimeter-wave radar, suppress water flow interference, and obtain point cloud data mainly based on human body reflections. The feature extraction module is used to extract multi-dimensional physical features in parallel from the point cloud data. The multi-dimensional physical features include height trajectory, velocity and acceleration, attitude deformation, scattering point density, and micro-motion. The decision engine module is used to construct a decision engine based on multi-level physical rules. According to the multi-dimensional physical characteristics, it sequentially performs judgments on sudden height drop triggers, velocity change verifications, attitude deformation confirmations, and temporal correlation and false alarm filtering, and evaluates the intensity of water flow interference in real time. ,when When the intensity threshold is exceeded, it automatically enters the high robustness mode, increases the judgment threshold at each level proportionally, and extends the observation time; The alarm output module is used to output a fall confirmation alarm when all the judgment conditions are met.

[0038] The layered preprocessing module described in this embodiment further includes: The range-Doppler analysis unit is used to sequentially perform one-dimensional fast Fourier transform and slow-time-dimensional fast Fourier transform on each frame of the original ADC signal to generate a two-dimensional data matrix containing range and Doppler information. The static clutter filtering unit is used to dynamically subtract the static background component from the current frame signal using a moving target display technique based on recursive averaging, so as to separate the moving target signal. The water flow feature recognition unit is used to analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals, and automatically identify and mark the water flow interference area dominated by the water flow generated by the shower head falling water flow and its splashing. The joint filtering unit is used to perform spatial weight attenuation on the signal marked as the water flow-dominant region, and to design an adaptive band-stop filter based on the identified water flow characteristic Doppler frequency to suppress water flow interference components in the frequency domain, thereby obtaining point cloud data dominated by human body reflection. Example 2

[0039] This embodiment details the implementation process of a method and system for detecting accidental falls during showering using millimeter-wave radar, specifically including the following steps: S1. Perform layered preprocessing on the raw signals acquired by the millimeter-wave radar to suppress environmental interference such as water flow. Specifically: S101. Perform one-dimensional fast Fourier transform (1DFFT) and slow time dimension FFT on each frame of the original ADC signal to generate a two-dimensional data matrix containing distance and Doppler information.

[0040] S102. Implement static clutter filtering: Using recursive averaging-based moving target display technology, the static background components generated by fixed facilities such as walls and shower heads are dynamically subtracted from the current frame signal to initially separate the dynamic target signal.

[0041] S103. Perform water flow feature identification and marking: Analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals, and automatically identify and mark the dominant interference area generated by the shower head's falling water flow and its splashing.

[0042] S104. Perform spatial-temporal joint filtering: Spatial weight attenuation is applied to the signal marked as the water flow-dominant region; at the same time, an adaptive band-stop filter is designed based on the identified water flow characteristic Doppler frequency to further suppress water flow interference components that overlap with the human motion spectrum in the frequency domain, resulting in filtered point cloud data dominated by human reflection.

[0043] S2. Extract multi-dimensional physical features from the preprocessed data in parallel, specifically: S201. Height Trajectory Analysis: Apply density clustering algorithm to the filtered point cloud to separate the human target point group; calculate and track the height coordinate sequence H(t) of its three-dimensional centroid; calculate the relative height change in real time.

[0044] S202. Velocity and acceleration feature extraction: Within the distance-Doppler cell corresponding to the human body point group, calculate its average radial velocity V(t) and its time derivative (radial acceleration a(t)); capture the typical "rapid increase in velocity (instability) and sudden drop in velocity (impact)" pattern during the fall.

[0045] S203. Posture Change Feature Extraction: Principal component analysis is performed on the three-dimensional coordinates of the human body point group to calculate the eigenvalues ​​of its covariance matrix, and the ratio of the maximum to the minimum eigenvalue (posture extension ratio λ) is calculated; the significant change of this ratio from standing posture (relatively large) to lying posture (close to 1) is a key indicator of a fall.

[0046] Scattering point density and micro-motion analysis: Calculate the density ratio ρ(t) of the number of human point clouds to the total number of point clouds; calculate the entropy value of the local Doppler spectrum and analyze the micro-Doppler entropy value mutation caused by unconscious limb swinging.

[0047] S3. Construct a decision engine based on multi-level physical rules to perform feature fusion and judgment, specifically as follows: S301, Level 1: Sudden Drop in Height Triggered. Determine if the following condition is met: the drop in height ΔH exceeds a preset threshold θH. This step serves as an initial screening for fall events.

[0048] S302, Level Two: Velocity Abrupt Change Verification. This step determines whether, during the descent, the peak velocity vpeak exceeds the peak threshold θP, and the velocity rapidly drops below the rest threshold θS after impact. This step verifies the dynamic anomaly of the fall event.

[0049] S303, Level 3: Posture Change Confirmation. After the velocity mutation verification, check whether the posture extension ratio λ decreases and remains below the low posture threshold θL for a duration exceeding the static confirmation time Ts. This step confirms the body posture after the fall.

[0050] S304, Level 4: Timing Correlation and False Alarm Filtering. A comprehensive assessment is made to determine whether the entire event sequence conforms to the temporal logic of "instability, fall, impact, stillness," excluding situations where significant spontaneous movement (such as attempting to stand up) occurs shortly after an impact, in order to distinguish between a genuine fall and similar activities such as bending over or sitting down quickly.

[0051] S305, Adaptive Threshold Adjustment: Real-time assessment of the current water flow disturbance intensity W. When W exceeds the intensity threshold, it automatically enters a high robustness mode, proportionally increasing the judgment thresholds at each level (such as θH, θV, etc.), extending the observation time Ts, and requiring all four conditions mentioned above to be strictly met before a final judgment can be made.

[0052] S4. Output the final detection result: If and only if all four levels of rules (S301 to S304) after adaptive adjustment by S305 are satisfied, the decision engine determines that a "fall" event has occurred, immediately generates and outputs an alarm message containing the event type and timestamp; otherwise, it is determined to be normal activity and is continuously monitored. Example 3

[0053] like Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the overall deployment and architecture of the system. This embodiment further elaborates on the deployment and architecture of the system with specific illustrations and implementation processes. In practical applications, an FMCW (Frequency Modulated Continuous Wave) millimeter-wave radar sensor operating in the 60GHz band is installed in the center of the ceiling above the shower area in the bathroom, at a height H of approximately 2.4 meters, vertically downwards, to ensure that its beam completely covers the entire shower area. The radar housing must meet the IP67 waterproof rating.

[0054] The overall system architecture provided in this embodiment is deployed according to a hierarchical structure of "physical environment, sensing, signal processing, core algorithm, output, and cloud / monitoring terminal". The specific process from data acquisition to final alarm is as follows: 1. The Physical World: The Bathroom Environment Application scenarios include: shower areas, showerheads, and wet floors. These scenarios are high-risk environments for falls due to strong, dynamic water flow disturbances (falling water from the showerhead, splashing water) and slippery surfaces.

[0055] 2. Sensing Layer: Millimeter-wave Radar A millimeter-wave radar, mounted from the center of the ceiling looking down, is used as the only non-contact sensor.

[0056] The radar continuously transmits frequency-modulated continuous wave (FMCW) signals and receives echoes reflected from the human body and the environment. This layer enables non-invasive, privacy-friendly raw data collection of human activities.

[0057] 3. Signal Processing Layer Basic digital signal processing of radar echoes: ADC sampling, Fast Fourier Transform (FFT), and point cloud generation.

[0058] The output contains raw point cloud data containing information such as distance, Doppler velocity, angle, and signal-to-noise ratio, providing basic input for subsequent core algorithms.

[0059] 4. The core algorithm layer consists of three tightly coupled modules: Layered signal processing (suppressing water flow interference): Performing static clutter filtering, water flow feature identification and labeling, and spatial-temporal joint filtering, effectively suppressing strong dynamic interference generated by shower water flow and splashing, and extracting a pure point cloud mainly composed of human body reflections.

[0060] Multi-dimensional feature extraction: Extract physical features such as height trajectory, velocity and acceleration, posture deformation (such as posture stretch ratio), scattering point density and micro-motion (such as micro-Doppler entropy) from the purified point cloud in parallel to comprehensively depict the kinematic and morphological changes during the fall process.

[0061] Multi-level rule decision engine: Based on multi-level physical rules (sudden drop in height trigger, speed change verification, attitude deformation confirmation, and time-series correlation to prevent false alarms), it makes decisions step by step; at the same time, it adaptively adjusts the threshold and observation time according to the real-time assessment of the water flow interference intensity, and finally outputs a fall confirmation signal.

[0062] 5. The output layer includes Local sound and light alarm: When a fall is detected, an sound and light alarm will be triggered immediately on site to alert cohabitants or caregivers.

[0063] Network gateway: Uploads alarm information (event type, timestamp, emergency level, etc.) to the cloud / monitoring terminal via wireless / wired network.

[0064] 6. Cloud / Monitoring Terminal Mobile App: Pushes fall alarm notifications to family members or the user's mobile phone, allowing them to view real-time status and historical records.

[0065] Nursing center platform: integrated into professional elderly care or medical monitoring systems to achieve functions such as centralized management, multi-user monitoring, and emergency dispatch.

[0066] The system provided in this embodiment addresses the strong water flow interference in bathing scenarios by designing a layered signal processing and adaptive decision mechanism, achieving non-contact fall detection with high robustness, high interpretability, and low computational complexity.

[0067] like Figure 3 The diagram shown illustrates the test environment and radar deployment for the bathing scenario in this invention. Figure 3 This tool is used to visually demonstrate the installation location, orientation, and monitoring coverage of millimeter-wave radar in a real bathroom environment, and specifically includes the following: 1. Bathroom layout: The diagram depicts a typical shower area, including: Shower head: Usually mounted on the wall, at a height of approximately 1.8 to 2.0 meters, with water flowing vertically downwards or at a certain angle. Floor: The wet and slippery shower floor is a place where people may fall.

[0068] 2. Millimeter-wave radar installation method and location: The radar is fixed in the center of the bathroom ceiling, directly above the shower area. Installation height: Approximately 2.4 meters (as stated in the instruction manual, "Installation height H is approximately 2.4 meters"). Installation posture: The radar is positioned vertically downwards (i.e., the radar antenna plane is parallel to the ground, and the normal points towards the Earth's center) to ensure that its beam can completely cover the entire shower area.

[0069] 3. Radar beam coverage: Using a 60GHz FMCW radar with a narrow beam angle, vertically downward installation can form a roughly conical or rectangular coverage area, completely covering the space where a person is standing, bending over, or falling.

[0070] 4. Human activity area: Show the range of movement during daily bathing (such as standing under the showerhead), as well as the area where one might lie down after a fall. The radar beam should cover both standing height (approximately 0.5~1.9 meters) and ground height (0~0.5 meters).

[0071] This method can accurately identify accidental falls caused by slippery conditions in a shower environment and effectively distinguish them from normal bathing activities (such as bending over, squatting, turning around, etc.), achieving highly reliable, low false alarm rate non-invasive safety monitoring.

[0072] The processing steps of this invention are as follows, based on actual measurement data of falls in bathing scenarios: Step 1: Perform layered preprocessing on the raw millimeter-wave radar signal to suppress water flow interference. First, the radar front-end mixes the transmitted linear frequency modulated signal with the received echo to obtain the beat signal (intermediate frequency signal). Let the transmitted signal be... in, For carrier frequency, For frequency modulation slope, The amplitude is [value]. The distance is [distance]. After the target is reflected, the received signal experiences a delay. ,in The speed of light. The expression for the intermediate frequency signal obtained after mixing and filtering is: Among them, beat frequency It contains the distance information of the target. This is the initial phase. After sampling and digitizing the intermediate frequency signal frame by frame, the following processing is performed: Distance and Doppler analysis: A one-dimensional fast Fourier transform is performed on the fast time dimension of each frame of data to obtain the distance-dimensional spectrum. Distance resolution. By signal bandwidth Decide: For continuous The signals in frames within the same range cell are subjected to a second-dimensional FFT to obtain the Doppler spectrum, thereby obtaining the radial velocity of the target. Information. Maximum unambiguous speed measurement range. for: in, For radar wavelength, The frame period is denoted as . After a two-dimensional FFT, a two-dimensional range-Doppler matrix (RDM) containing target range and radial velocity information is obtained, denoted as . ,in For distance cell index, This is the index for the velocity unit.

[0073] Static clutter filtering: To eliminate reflections from static backgrounds such as walls and fixed structures, a recursive averaging method is used to process the RDM. Let the... The RDM of the frame is Then the estimated static background and dynamic target matrix Update using the following formula: in, This is the forgetting factor, used to control the rate of background updates.

[0074] Water flow feature identification and labeling: analysis of dynamic matrix The spatiotemporal characteristics are then analyzed. The dominant flow regions are then identified by calculating the energy and variance within the region. : If a certain area satisfy (Energy threshold) and (Velocity variance threshold) is then marked as the dominant flow zone. .

[0075] in, This represents the energy within the region Ω; This represents the value of the dynamic target matrix at coordinates (r, v) in the k-th frame. This indicates the average speed within the monitored area; Space-time joint filtering: Spatial filtering: In subsequent point cloud generation and feature calculation, for points belonging to... The contribution weight of point cloud data in a region It was reduced.

[0076] Frequency domain filtering: based on the identified Internal mean Doppler velocity Design a center frequency in the frequency domain. band-stop filter ,right Filtering is performed to obtain a "purified" matrix that suppresses the principal components of the water flow. : in, This refers to a band-stop filter designed in the frequency domain. Through the Point cloud data is obtained by performing CFAR detection and angle FFT. The processed point cloud effect is as follows: Figure 4 As shown.

[0077] Step 2: Extract multi-dimensional physical features in parallel from the preprocessed data. Through the CFAR detection and angle FFT are performed to obtain point cloud data, and the following four types of features are extracted simultaneously: Height trajectory features: Apply density-based clustering algorithm to each frame of point cloud to separate human target point groups. ,in Calculate its three-dimensional centroid coordinates. And track its height component Changes over time.

[0078] This represents the separated group of human target points; Represents the three-dimensional coordinates of the i-th point cloud; Represents the three-dimensional centroid coordinate components of the point cloud in the k-th frame; Velocity and acceleration characteristics: in human body point groups Within the corresponding range-Doppler cell, calculate its energy-weighted average radial velocity. : Calculate radial acceleration .

[0079] Posture deformation characteristics: for human body point groups Principal component analysis is performed on the three-dimensional coordinates of the point cloud. First, the covariance matrix of the point cloud is calculated. : right Perform eigenvalue decomposition to obtain eigenvalues. .

[0080] Define posture extension ratio for: in, Let represent the largest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; Let represent the smallest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; To prevent division by zero for a very small positive number.

[0081] When standing, the human body extends in the vertical direction (first principal component). Larger; When falling and landing in a lying / prone position, the body unfolds horizontally, with similar dimensions in all three dimensions. Close to 1.

[0082] Scattering point density and micro-motion characteristics: Calculate the point cloud density ratio of the human body: ,in, This represents the total number of point clouds in the current frame.

[0083] The Shannon entropy of the local Doppler spectrum is calculated to quantify the degree of micro-motion disturbance. This is done within the unit of distance from the human body. Nearby, extract velocity dimension spectrum The probability distribution is obtained after normalization. Then the micro-Doppler entropy for: This indicates that the purified dynamic target matrix is ​​within a specific distance cell. The velocity-dimensional spectrum is obtained by calculating the square of the amplitude at a given point. This represents the distance unit where the human target is located; Indicates the Doppler velocity of a human target; The involuntary limb movements during a fall may cause spectral diffusion, thereby leading to... Mutations.

[0084] like Figure 5 The image shown is an example of the feature extraction results obtained by the millimeter-wave radar sensor in an embodiment of the present invention.

[0085] Step 3: Build and execute a decision engine based on multi-level physics rules Decision engine Figure 6 The illustrated logic flow employs a progressive decision-making process. All decisions are based on physical thresholds and include an adaptive mechanism. A threshold set is defined. The standard values ​​are shown in Table 1.

[0086] Table 1: Example of System Core Parameter Configuration The system estimates the intensity of water flow disturbance in real time: in, This indicates the intensity of water flow disturbance estimated by the system in real time; The total dynamic energy. Energy in the water flow-dominant zone.

[0087] when When the threshold is adjusted to a high robustness mode, the threshold is then adjusted to... (scaling factor) ),and (Extendance Factor) ).

[0088] This represents the adjusted set of physical thresholds; The logic of multi-level judgment is as follows: Level 1: Triggered by sudden drop in altitude.

[0089] Condition 1 is the logic for triggering a "sudden descent in altitude." It requires that within a specified time window, the target's altitude difference (i.e., the difference between the maximum and minimum altitude) must exceed an altitude threshold, and simultaneously, the maximum absolute value of its vertical velocity must also exceed a vertical velocity threshold. Both conditions must be met simultaneously to determine whether the target has undergone a rapid and significant vertical descent.

[0090] Level 2: Velocity mutation verification. A typical "stall-impact" velocity profile needs to be detected within the time window of a sudden drop in altitude.

[0091] Within the time window of sudden altitude drop, there must exist two moments, ta and tb (with ta being later than tb), such that the velocity V(ta) at moment ta is greater than a preset velocity threshold θp, and the absolute value of the velocity at moment tb, |V(tb)|, is less than a preset post-impact velocity threshold θS, in order to verify whether there are typical "stall-impact" velocity profile characteristics.

[0092] Level 3: Posture deformation confirmation. After impact, the posture must remain in a falling position.

[0093] For all times t within the entire time interval [tb, tb+Ts] from the impact time tb to the time Ts, the target's attitude parameter Rλ(t) must be less than the preset attitude threshold θz, thus confirming that the target's attitude after the impact continues to be a falling posture (i.e., the attitude deformation conforms to the characteristics of falling).

[0094] Level 4: Timing Correlation and False Alarm Prevention. This involves comprehensively assessing the physical plausibility and consistency of the event sequence.

[0095] exist Internal, displacement of the center of mass It should be less than the activity threshold. And it did not reappear. or The situation.

[0096] Step 4: Output final detection results and performance verification To verify the effectiveness of the method of the present invention, a test was conducted in a bathing environment, and the test results are as follows. Figure 7 As shown. If and only if Upon system activation, a "fall" event is detected. An alarm signal is immediately generated, which can be displayed locally and reported to the monitoring platform. The alarm information includes the event type, timestamp, and fall location. An example of the host computer interface is shown below. Figure 7 As shown in the figure. The test results show that the system can maintain a high detection rate and keep the false alarm rate at a low level even under strong water flow interference.

[0097] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for detecting abnormal behavior of accidental falls during bathing based on millimeter-wave radar, characterized in that, Includes the following steps: The raw signals acquired by millimeter-wave radar are preprocessed in layers to suppress water flow interference, resulting in point cloud data mainly composed of human body reflections. Multi-dimensional physical features are extracted in parallel from the point cloud data, including height trajectory, velocity and acceleration, attitude deformation, scattering point density and micro-motion; A decision engine based on multi-level physical rules is constructed. According to the multi-dimensional physical characteristics, the engine sequentially performs the following judgments: height drop trigger judgment, velocity change verification judgment, attitude deformation confirmation judgment, and time sequence correlation and false alarm filtering judgment. Real-time assessment of water flow disturbance intensity ,when When the intensity threshold is exceeded, it automatically enters the high robustness mode, increases the judgment threshold at each level proportionally, and extends the observation time; When all the conditions for judgment are met, a fall confirmation alarm is output.

2. The abnormal behavior detection method for accidental falls during bathing based on millimeter-wave radar as described in claim 1, characterized in that, The layered preprocessing further includes: For each frame of the original ADC signal, perform a one-dimensional fast Fourier transform and a slow-time fast Fourier transform sequentially to generate a two-dimensional data matrix containing distance and Doppler information. A dynamic target display technique based on recursive averaging is adopted to dynamically subtract the static background component from the current frame signal in order to separate the dynamic target signal; Analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals to automatically identify and mark the interference areas dominated by the water flow generated by the shower head's falling water and its splashing. Spatial weight attenuation is applied to signals marked as water flow-dominant regions, and an adaptive band-stop filter is designed based on the identified water flow characteristic Doppler frequencies to suppress water flow interference components in the frequency domain, resulting in point cloud data dominated by human body reflections.

3. The abnormal behavior detection method for accidental falls during bathing based on millimeter-wave radar as described in claim 1, characterized in that, The water flow feature identification and labeling identifies and labels the dominant water flow areas by calculating the energy and variance within the region. : If a certain area satisfy and Then mark it as the dominant water flow area. ;in, Indicates the energy threshold; Indicates the velocity variance threshold; If a certain region satisfies and Then it is marked as the dominant water flow area, where Energy threshold This is the velocity variance threshold; in, This represents the energy within the region Ω; This represents the value of the dynamic target matrix at coordinates (r, v) in the k-th frame; Represents the distance unit in the distance-Doppler spectrum; This represents the velocity element in the range-Doppler spectrum; Indicates the monitoring area; This indicates the average speed within the monitored area; It indicates the degree of dispersion of velocity distribution within the monitoring area.

4. The abnormal behavior detection method for accidental falls during bathing based on millimeter-wave radar as described in claim 1, characterized in that, The posture deformation feature extraction specifically involves: performing principal component analysis on the three-dimensional coordinates of the human body point group, calculating the eigenvalues ​​of its covariance matrix, and calculating the ratio of the maximum eigenvalue to the minimum eigenvalue as the posture extension ratio. .

5. The method for detecting abnormal behavior of accidental falls during bathing based on millimeter-wave radar according to claim 4, characterized in that, The posture stretch ratio The calculation formula is: in, Let represent the largest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; Let represent the smallest eigenvalue obtained after eigenvalue decomposition of the covariance matrix C; To prevent division by zero for a very small positive number.

6. The abnormal behavior detection method for accidental falls during bathing based on millimeter-wave radar as described in claim 1, characterized in that, The scattering point density and micro-motion analysis specifically involves calculating the density ratio of the human body point cloud to the total point cloud. Furthermore, the entropy value was calculated from the local Doppler spectrum to analyze the micro-Doppler entropy abrupt changes caused by unconscious limb swinging.

7. The method for detecting abnormal behavior of accidental falls during bathing based on millimeter-wave radar as described in claim 1, characterized in that, The decision-making process of the decision engine based on multi-level physical rules includes: Level 1 altitude drop trigger: Determines whether the altitude drop exceeds a preset threshold; Level 2 velocity mutation verification: Determine whether there is a velocity peak exceeding the peak threshold during the descent and whether the velocity rapidly drops below the resting threshold after impact; Level 3 Attitude Deformation Confirmation: After the velocity mutation verification, determine whether the attitude stretch ratio decreases and remains below the low attitude threshold, and whether the duration exceeds the static confirmation time. Level 4 temporal correlation and false alarm filtering: comprehensively judge whether the entire event sequence conforms to the temporal logic of "instability, fall, impact, and stillness", and exclude the situation of large-scale autonomous movement in a short period of time after the impact.

8. The method for detecting abnormal behavior of accidental falls during bathing based on millimeter-wave radar as described in claim 1, characterized in that, The intensity of water flow interference The dynamic total energy is estimated in real time by the ratio of the energy in the dominant water flow zone, and the formula is as follows: in, This indicates the intensity of water flow disturbance estimated by the system in real time; The total dynamic energy. Energy in the water flow-dominant zone.

9. A system for detecting abnormal behavior such as accidental falls during showering based on millimeter-wave radar, characterized in that, include: The layered preprocessing module is used to perform layered preprocessing on the raw signals acquired by the millimeter-wave radar, suppress water flow interference, and obtain point cloud data mainly based on human body reflections. The feature extraction module is used to extract multi-dimensional physical features in parallel from the point cloud data. The multi-dimensional physical features include height trajectory, velocity and acceleration, attitude deformation, scattering point density, and micro-motion. The decision engine module is used to construct a decision engine based on multi-level physical rules. According to the multi-dimensional physical characteristics, it sequentially performs judgments on sudden height drop triggers, velocity change verifications, attitude deformation confirmations, and temporal correlation and false alarm filtering, and evaluates the intensity of water flow interference in real time. ,when When the intensity threshold is exceeded, it automatically enters the high robustness mode, increases the judgment threshold at each level proportionally, and extends the observation time; The alarm output module is used to output a fall confirmation alarm when all the judgment conditions are met.

10. The abnormal behavior detection system for accidental falls during bathing based on millimeter-wave radar as described in claim 9, characterized in that, The layered preprocessing module further includes: The range-Doppler analysis unit is used to sequentially perform one-dimensional fast Fourier transform and slow-time-dimensional fast Fourier transform on each frame of the original ADC signal to generate a two-dimensional data matrix containing range and Doppler information. The static clutter filtering unit is used to dynamically subtract the static background component from the current frame signal using a moving target display technique based on recursive averaging, so as to separate the moving target signal. The water flow feature recognition unit is used to analyze the spatiotemporal continuity, Doppler distribution concentration, and spatial distribution clustering of dynamic signals, and automatically identify and mark the water flow interference area dominated by the water flow generated by the shower head falling water flow and its splashing. The joint filtering unit is used to perform spatial weight attenuation on the signal marked as the water flow-dominant region, and to design an adaptive band-stop filter based on the identified water flow characteristic Doppler frequency to suppress water flow interference components in the frequency domain, thereby obtaining point cloud data dominated by human body reflection.

Images