Non-contact sensing-based early warning method, device, equipment and medium

Abstract

The application relates to the technical field of intelligent decision-making, and discloses a prewarning method, device and equipment based on non-contact sensing and a medium, which comprises the following steps: collecting non-contact body motion sensing data and thermal distribution imaging data, fusing and constructing a real-time human body digital model; extracting motion trajectory features and posture features based on the real-time human body digital model and performing behavior context reasoning to identify abnormal behavior events; generating vital sign data when the abnormal behavior events are identified; determining a risk first-aid level according to the type of the abnormal behavior events and the vital sign data, and executing corresponding graded emergency dispatch strategies. The application can be applied to home health monitoring and other business scenarios, precise identification and graded response of abnormal behaviors and health risks are realized by fusing body motion and thermal imaging information and combining context reasoning, and the monitoring accuracy and timeliness are improved while privacy is ensured.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to an early warning method, device, equipment and medium based on non-contact perception. Background Technology

[0002] With the accelerating aging process, health and safety monitoring in the home environment has gradually become an important issue in the medical and health field. Falls, sudden physical abnormalities, or prolonged unresponsiveness are relatively common among the elderly in private spaces such as bedrooms and bathrooms. These events are characterized by their suddenness, delayed detection, and rapid escalation of risks. If not identified in time, they can easily lead to serious consequences. Therefore, how to continuously sense the body's condition and promptly identify abnormal behavior in the home setting has become a pressing issue that current home health monitoring technologies need to address.

[0003] While existing video surveillance technologies offer high capabilities in motion recognition, they require continuous image capture of living spaces, severely infringing on personal privacy, especially in highly private areas like bedrooms and bathrooms. This results in extremely low user acceptance and limits the application of such technologies in home healthcare scenarios. Wearable device-based monitoring, while capable of acquiring relatively accurate motion and vital sign data, relies on continuous wear and maintenance. Elderly individuals often experience monitoring interruptions due to forgetting to wear or charge the device, or due to its complex operation, making it difficult to guarantee the continuity and reliability of monitoring.

[0004] Furthermore, existing monitoring systems typically rely on single-dimensional data for abnormal behavior identification, lacking the ability to comprehensively analyze changes in human posture and physiological state. This makes it difficult to effectively distinguish between normal and abnormal behavior, and easily leads to false alarms. Simultaneously, behavior identification and vital sign monitoring are often independent, lacking a correlation analysis mechanism. This makes it difficult to combine physiological state with risk level assessment after abnormal behavior occurs, affecting the accuracy and effectiveness of subsequent emergency response strategies. Summary of the Invention

[0005] The main objective of this invention is to provide an early warning method, device, equipment, and storage medium based on non-contact sensing, aiming to solve the technical problem that existing home health monitoring technologies cannot accurately identify abnormal behavior and determine risk levels based on the correlation analysis of human posture changes and physiological states without relying on video surveillance and wearable devices.

[0006] To achieve the above objectives, the present invention provides an early warning method based on non-contact sensing, comprising: Collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The non-contact motion sensing data and the thermal distribution imaging data are fused to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. Based on the real-time human digital model, the motion trajectory features and posture features of the target object are obtained, and the motion trajectory features and posture features are analyzed by the behavioral context analysis model to identify abnormal behavioral events. When the abnormal behavior event is detected, the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model are locked to generate vital sign data. Based on the type of the abnormal behavioral event and the vital signs data, determine the risk level of the target individual for emergency medical care; The corresponding tiered emergency dispatch strategy shall be executed according to the aforementioned risk and emergency response level.

[0007] Furthermore, to achieve the above objectives, the present invention provides an early warning device based on non-contact sensing, comprising: The multi-source sensing and acquisition module is used to collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The cross-modal fusion modeling module is used to fuse the non-contact motion sensing data and the thermal distribution imaging data to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. The behavioral context reasoning module is used to obtain the motion trajectory features and posture features of the target object based on the real-time human digital model, and to analyze the motion trajectory features and posture features using the behavioral context analysis model to identify abnormal behavioral events. The vital signs analysis module is used to identify the abnormal behavioral event, lock the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model, and generate vital signs data. The risk level determination module is used to determine the risk level of the target object based on the type of the abnormal behavior event and the vital signs data. The emergency dispatch execution module is used to execute the corresponding graded emergency dispatch strategy according to the risk and emergency response level.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a non-contact sensing-based early warning program stored in the memory and executable on the processor, wherein when the non-contact sensing-based early warning program is executed by the processor, it implements the steps of the non-contact sensing-based early warning method as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a non-contact sensing-based early warning program, wherein the non-contact sensing-based early warning program, when executed by a processor, implements the steps of the non-contact sensing-based early warning method as described above.

[0010] Beneficial Effects: This invention relates to the field of intelligent decision-making technology, and discloses an early warning method, device, equipment, and medium based on non-contact perception. The method includes: collecting non-contact motion sensing data and thermal distribution imaging data, fusing them to construct a real-time human digital model; extracting motion trajectory features and posture features based on the real-time human digital model and performing behavioral context reasoning to identify abnormal behavioral events; generating vital sign data when an abnormal behavioral event is identified; determining the risk and emergency response level based on the type of abnormal behavioral event and the vital sign data, and executing the corresponding graded emergency dispatch strategy. This invention can be applied to business scenarios such as home health monitoring. By fusing non-contact motion sensing data and thermal distribution imaging data to construct a real-time human digital model, and combining behavioral context reasoning and vital sign analysis, it achieves accurate identification and graded response to abnormal behaviors and health risks, improving the accuracy and timeliness of home monitoring while protecting privacy and enhancing compliance. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for an early warning method based on non-contact sensing in one embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the early warning method based on non-contact sensing according to the present invention. Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the early warning device based on non-contact sensing of the present invention. Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The early warning method based on non-contact sensing provided in this invention can be applied to applications such as... Figure 1In this application environment, the client communicates with the server via a network. The server can collect non-contact motion sensing data and thermal distribution imaging data from the client, fuse them to construct a real-time human digital model; extract motion trajectory features and posture features based on the real-time human digital model and perform behavioral context reasoning to identify abnormal behavioral events; generate vital sign data when abnormal behavioral events are identified; determine the risk and emergency response level based on the type of abnormal behavioral event and vital sign data, and execute the corresponding graded emergency dispatch strategy. This invention can be applied to business scenarios such as home health monitoring. By fusing non-contact motion sensing data and thermal distribution imaging data to construct a real-time human digital model, and combining behavioral context reasoning and vital sign analysis, it achieves accurate identification and graded response to abnormal behaviors and health risks, improving the accuracy and timeliness of home monitoring while protecting privacy and improving compliance. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the non-contact sensing-based early warning method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0015] like Figure 2 As shown, the early warning method based on non-contact sensing proposed in this invention includes the following steps: S10, collects non-contact motion sensing data and thermal distribution imaging data within the monitoring area; In this embodiment, the monitoring area refers to the spatial range where people spend a long time, including key locations such as the bed area, living area, and hygiene area. Non-contact motion sensing data originates from the reflection changes of electromagnetic waves on the human body surface and micro-movement structures, manifested as spatial coordinate distribution and frequency modulation information that changes over time, reflecting micro-vibration phenomena such as changes in body position, posture, and chest cavity movement. Thermal distribution imaging data originates from the natural intensity differences in infrared radiation emitted by the human body, manifested as a temperature value matrix that changes with spatial location, reflecting the temperature distribution status of different parts of the body and the temperature change trend over time.

[0016] In the specific implementation process, millimeter-wave radar and infrared thermal imaging devices are deployed within the monitoring area. The millimeter-wave radar continuously transmits frequency-modulated continuous wave signals, which are reflected by the human body to form echo signals. Through frequency domain processing, a three-dimensional spatial coordinate sequence and a micro-Doppler spectrum sequence are obtained. The three-dimensional spatial coordinate sequence reflects the macroscopic movement position of the human body, while the micro-Doppler spectrum sequence reflects the subtle periodic vibrations of the human body. The infrared thermal imaging device collects the radiation flux distribution and maps the radiation intensity into a temperature numerical matrix, with each pixel corresponding to a heat source region in space. Both types of data are timestamped to ensure a strict correspondence between spatial changes and thermal changes in the time dimension.

[0017] The three-dimensional spatial coordinates are derived from the calculation of echo time difference and phase difference, while the micro-Doppler spectrum is derived from the modulation effect of human body micro-movements on electromagnetic wave frequencies. The temperature value matrix is ​​derived from the calibration relationship between radiation intensity and temperature. To ensure spatial correspondence between the two types of data, a calibration relationship is established between millimeter-wave radar coordinates and thermal imaging pixel coordinates, allowing spatial coordinates to be mapped to specific pixel locations in the temperature matrix. Through this mapping relationship, the corresponding temperature value can be read at each spatial point, establishing a correspondence between spatial motion information and thermal distribution information at the same spatial location.

[0018] During continuous data acquisition, the movement of the human body in space causes continuous changes in three-dimensional spatial coordinates. Micro-vibrations in the chest and abdominal regions form stable periodic features in the micro-Doppler spectrum, and changes in body surface temperature form local variation regions in the temperature numerical matrix. This multidimensional dataset can fully express the human body's motion and physiological state in space, providing highly reliable raw input data for subsequent processing.

[0019] For example, in home healthcare scenarios, when elderly people get up, sit down, or fall at night, changes in spatial coordinates and temperature distribution are recorded simultaneously, which can distinguish between normal postural adjustments and abnormal falling behaviors. At the same time, changes in respiratory rhythm can be identified through the micro-vibration spectrum of the chest and abdomen region, providing a reliable data foundation for home health monitoring.

[0020] This embodiment acquires body motion reflection information and thermal radiation distribution information simultaneously in the same monitoring area and aligns them in the spatial and temporal dimensions. It can obtain continuous data that simultaneously reflects the human body's motion state and physiological state without relying on video surveillance and wearable devices, providing a highly reliable original basis for health status identification.

[0021] S20, the non-contact motion sensing data and the thermal distribution imaging data are fused to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. In this embodiment, the non-contact motion sensing data includes three-dimensional point cloud coordinates and micro-Doppler spectrum information. The three-dimensional point cloud coordinates describe the geometric contours and positional changes of the human body in space, while the micro-Doppler spectrum information describes the periodic micro-vibrations of the chest and abdomen. Thermal distribution imaging data exists in the form of a two-dimensional temperature numerical matrix, with each pixel corresponding to a heat source region in space. To establish a spatial correspondence between the two types of data, a coordinate transformation relationship needs to be established between the millimeter-wave radar coordinates and the thermal imaging pixel coordinates.

[0022] The coordinate transformation relationship originates from the joint calibration model. The calibration process determines the projection position of spatial coordinates onto the thermal imaging plane, enabling each 3D point cloud coordinate to find its corresponding pixel position in the temperature matrix. This mapping binds temperature values ​​to 3D spatial points, giving each point not only positional attributes but also temperature attributes. Simultaneously, micro-Doppler spectral information is time-matched to the spatial points via timestamps, imbuing each point with micro-vibration spectral attributes. After attribute binding, an attribute-fused point cloud set containing spatial coordinates, temperature values, and micro-vibration spectral information is formed.

[0023] The attribute fusion point cloud dataset is processed through spatial density clustering to identify spatial point sets belonging to the same human body structure. The clustering process is based on spatial distance and attribute similarity thresholds, ensuring that points from the same body part belong to the same cluster. The resulting spatial structure reflects the geometric shape of the human body, the temperature distribution reflects the temperature characteristics of different areas of the body surface, and the micro-motion spectrum reflects the vital sign vibration characteristics of the chest and abdomen. By continuously updating this clustering structure, a dynamically changing spatial structure model is formed over time. This model digitally represents the changes in human posture and vital signs in space, constituting a real-time digital human body model.

[0024] This real-time human digital model is not a single geometric model, but a composite model that simultaneously possesses geometric morphology, temperature distribution, and micro-vibration spectrum. Geometric morphology is used to describe changes in human posture, temperature distribution is used to describe the thermal characteristics of the body surface, and micro-vibration spectrum is used to describe the rhythms of vital signs. All three are uniformly expressed in the same model using spatial points as carriers, allowing subsequent processing to directly perform posture analysis and vital sign analysis on this model.

[0025] This embodiment binds spatial coordinates, temperature values, and micro-vibration spectra at the same spatial point, which can form a digital expression structure that simultaneously reflects changes in human geometric posture and vital signs, providing a unified data foundation for subsequent posture recognition and physiological state recognition.

[0026] S30, based on the real-time human digital model, obtain the motion trajectory features and posture features of the target object, and apply the behavior context analysis model to analyze the motion trajectory features and posture features to identify abnormal behavior events; In this embodiment, the motion trajectory features and posture features of the target object are obtained based on a real-time human digital model, relying on a continuously updated set of spatial points and their spatial centroid coordinates, 3D bounding box size, temperature distribution, and micro-motion spectrum attributes within the real-time human digital model. The motion trajectory features characterize the displacement and velocity changes of the target object in the time dimension, originating from the time series of the spatial centroid coordinates. The spatial centroid coordinates are obtained by weighting the attribute fusion point cloud set in the real-time human digital model according to spatial coordinates. The weighting method can employ point density weighting or temperature weighting to make the centroid position more sensitive to the main body area of ​​the human body. The spatial centroid coordinates constitute a trajectory sequence over continuous time frames. A velocity vector is obtained by differencing the trajectory sequence, and an acceleration vector is obtained by further differencing the velocity vector. To conform to the kinematic representation of abnormal falls, the vertical velocity component further forms an instantaneous descent velocity, and the vertical acceleration component further forms a ground-contact acceleration. The instantaneous descent velocity and ground-contact acceleration are used as motion trajectory features in subsequent inference.

[0027] Posture features are used to characterize the orientation and posture changes of a target object in space, derived from the 3D bounding box size of a real-time human digital model. The 3D bounding box size is generated from the extreme range of the attribute fusion point cloud set on the 3D coordinate axes, including the long side, short side, and height. The ratio of the long side to the short side reflects the degree of human unfolding in the planar direction. When the ratio meets the criteria for horizontal unfolding, the subject's orientation is determined to be horizontal. The subject's orientation participates in subsequent inference as a posture feature. Posture features can also be extended to determine a static state. The static state is derived from the displacement amplitude of the spatial centroid coordinates within a preset time window and the change amplitude of the 3D bounding box size. When both the displacement amplitude and the size change amplitude are below a threshold, the target object is determined to be in a static state, thus forming a static state determination result in the posture features.

[0028] Behavioral context analysis models are used to consistently judge the characteristics of movement trajectories and postures within the daily behavioral context of a target individual, avoiding misidentification of normal behavior as abnormal behavior. The core of this model can be implemented using a trained contextual reasoning model specifically designed for home health monitoring scenarios. This model learns and models the correlation between the target individual's daily activity patterns and physiological state in a healthcare context. The model training process can employ unsupervised or semi-supervised learning methods, using historical behavioral habit data as the training set, and constructing an individualized baseline distribution of normal behavior through density estimation (such as Gaussian mixture models) or temporal modeling methods.

[0029] The inputs to the behavioral context analysis model include the current time period, current spatial location, movement trajectory characteristics, posture characteristics, and historical behavioral habit data. Historical behavioral habit data originates from the target object's long-term activity records within the monitoring area, including time period distribution, spatial location distribution, posture distribution, and dwell time distribution. The current time period can be mapped from the timestamp of the collected data to intraday time slices (such as typical healthy sleep stages like morning activities, afternoon naps, and nighttime sleep). The current spatial location can be mapped from the spatial centroid coordinates to functional areas within the monitoring area (such as bedroom rest areas, bathrooms, living room activity areas, and other areas closely related to health events).

[0030] Matching probability quantifies the consistency between the current time period and spatial location and historical behavioral data, and is also a key indicator for assessing whether current behavior deviates from healthy norms. In its implementation, matching probability is calculated using a probabilistic model that uses historical behavioral data as training samples to establish a joint probability distribution of multi-dimensional features such as time period, spatial location, and posture. For the current input feature vector, the model calculates its likelihood probability or confidence score under this distribution, which serves as the matching probability value. This process is essentially a detection of abnormal health behaviors; the model learns individualized normal behavioral patterns to identify abnormal deviations that may indicate health risks (such as accidental falls or prolonged inactivity due to sudden discomfort).

[0031] The matching probability, as the output of the behavioral context reasoning model, is fed into the subsequent anomaly detection module along with action trajectory features and posture features. This module can be viewed as a lightweight classifier that, based on preset rules or decision boundaries obtained through machine learning, comprehensively judges whether the current event belongs to "normal behavior," "potential risk," or "emergency anomaly." In healthcare applications, the model's input features are given clear medical meaning: action trajectory features reflect activity level and balance, posture features reflect body posture and state of consciousness, and spatiotemporal features reflect daily routines and activity range. By continuously learning from the user's personalized health baseline, the model achieves accurate and low-false-positive identification of abnormal health events, thus providing a reliable decision-making basis for subsequent vital sign detection and tiered emergency dispatch.

[0032] Abnormal behavior event identification implements branch determination within the behavioral context analysis model. For fall-related anomalies, the instantaneous descent velocity and ground acceleration in the motion trajectory features reflect the combined characteristics of rapid descent and impact, while the subject's orientation in the posture features reflects the post-landing posture direction. When the instantaneous descent velocity exceeds a preset fall threshold and the subject's orientation is horizontal, an abnormal behavior event is confirmed. For prolonged stillness anomalies, the stillness state determination result in the posture features characterizes the continuous immobility of the human body in space, and the matching probability characterizes the degree to which the current time period and current spatial location deviate from historical behavioral habit data. When the stillness state is established and the matching probability is consistently lower than a preset normal behavior confidence level exceeding a preset duration threshold, an abnormal behavior event is confirmed. The determination of a consistently lower confidence level requires continuous sampling within a time window. The time window is derived from a uniform timestamp sequence, and its length is defined by a preset duration threshold, ensuring that the anomaly determination has temporal continuity rather than being triggered by a single frame.

[0033] For example, in a home healthcare scenario, when a target person sits normally on a sofa, the spatial centroid coordinates show a slow descent, with the instantaneous descent speed not exceeding a preset fall threshold. The subject's orientation remains non-horizontal, and the matching probability is consistent with historical behavioral data; therefore, the abnormal behavior event is not confirmed. If the target person suddenly slips in the bathroom area, the spatial centroid coordinates show a rapid descent in the vertical direction, with the instantaneous descent speed exceeding a preset fall threshold. The subject's orientation is determined to be horizontal, and the abnormal behavior event is confirmed. If the target person remains still in the bedroom area for an extended period, and the current time period deviates from historical behavioral data in terms of current spatial location, the stillness is established, and the matching probability remains below a preset normal behavior confidence level exceeding a preset duration threshold; therefore, the abnormal behavior event is confirmed.

[0034] This embodiment extracts motion trajectory features and posture features together and introduces matching probability to judge contextual consistency. It can distinguish between rapid falling anomalies and continuous stillness anomalies on the same real-time human digital model, reducing the probability of normal behavior being misjudged as abnormal behavior.

[0035] S40, when the abnormal behavior event is detected, lock the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model, and generate vital sign data. In this embodiment, a locking action is triggered upon detection of an abnormal behavioral event. This abnormal behavioral event serves as the trigger condition for entering the vital signs extraction process, originating from the analysis results of previous motion trajectory and posture features. The locking action focuses on local areas related to cardiopulmonary activity and their associated data within the real-time human digital model, ensuring that subsequent extraction concentrates on the target object's chest and abdominal movements and surface temperature changes, reducing interference from irrelevant body movements and environmental heat sources on vital signs data. The real-time human digital model contains a set of spatial points and their attribute information. The attribute information is derived from the fusion of non-contact motion sensing data and thermal distribution imaging data, and includes at least micro-Doppler spectral information associated with the spatial point set, local temperature values, and a unified timestamp. Micro-motion features corresponding to cardiopulmonary activity describe subtle surface displacement changes in the chest and abdomen caused by respiratory fluctuations and the mechanical vibration of the heartbeat. These micro-motion features originate from the time-series changes of micro-Doppler spectral information in the chest and abdominal region. Body temperature distribution features describe the spatial temperature distribution in the chest and abdominal region or key parts of the body. These body temperature distribution features originate from the local temperature value set in the target region within the two-dimensional temperature numerical matrix of the thermal distribution imaging data.

[0036] Locating the micro-motion features corresponding to cardiopulmonary activity in a real-time human digital model requires first locating the chest and abdomen region. The location of the chest and abdomen region can be accomplished based on the temperature distribution features and geometric shape of the real-time human digital model. The temperature distribution features provide a relatively stable thermal radiation area on the surface of the chest and abdomen, while the geometric shape provides the scale range of the human body in space. The localization process involves extracting high-temperature connected regions from a two-dimensional temperature numerical matrix and mapping them to a set of spatial points. Then, combining the three-dimensional bounding box dimensions of the spatial point set, the height and lateral ranges of the chest and abdomen region are determined, thus obtaining the spatial range of the chest and abdomen region. Once the chest and abdomen region is determined, the corresponding set of spatial points is marked as the target point set. This target point set is subsequently used to extract micro-motion features and body temperature distribution features, ensuring consistency between references and avoiding redundant calculations in non-target areas.

[0037] The locking and extraction of micro-motion features revolves around the micro-Doppler spectral information associated with the target point set. Micro-Doppler spectral information reflects the minute changes in echo phase over time, forming a spectral sequence in the time dimension. This spectral sequence can be used to generate micro-motion features through frequency band selection and energy aggregation. Frequency band selection distinguishes between low-frequency bands related to respiration and high-frequency bands related to heartbeat, while energy aggregation converts the spectral sequence into a feature sequence that can be used to estimate respiratory and heart rates. To enhance sensitivity to cardiopulmonary activity, micro-motion features can also be demodulated to obtain a displacement sequence, which is then bandpass filtered to separate the respiratory and heartbeat components, ensuring that the micro-motion features simultaneously include respiratory fluctuations and heartbeat vibrations. The time alignment of micro-motion features relies on a unified timestamp, ensuring that the micro-motion features and body temperature distribution features correspond to the same physiological state within the same time window.

[0038] The identification and extraction of body temperature distribution features revolve around the local temperature values ​​of the target point set. These local temperature values ​​originate from the pixel set corresponding to the two-dimensional temperature numerical matrix in the chest and abdominal region. This pixel set can be obtained by mapping the spatial point set in the real-time human digital model to the thermal imaging pixel coordinate system. Body temperature distribution features can be expressed as statistical measures of the local temperature value set, including average temperature, temperature variance, and temperature gradient direction distribution, reflecting both the body temperature level and the uniformity of temperature distribution. To adapt to short-term status assessments following abnormal behavioral events, body temperature distribution features can also form a temperature sequence and calculate the temperature change amplitude within a preset time window. The temperature change amplitude characterizes the trend of body temperature rise or fall within a short period. The time window is determined by a unified timestamp to avoid misjudgments caused by overlapping windows.

[0039] Vital signs data are generated jointly from micro-motion characteristics and body temperature distribution characteristics. This data, as a structured output, is used for subsequent risk and emergency response level determination. Vital signs data may include respiratory rate, heart rate, and body temperature status indicators. The respiratory rate is derived from the estimated dominant frequency or period of the respiratory component in the micro-motion characteristics; the heart rate is derived from the estimated dominant frequency or peak interval of the heartbeat component in the micro-motion characteristics; and the body temperature status indicators are derived from a combination of the average temperature value and the amplitude of temperature change in the body temperature distribution characteristics. Body temperature status indicators can be expressed using interval labels, ensuring the accuracy and interpretability of the vital signs data when subsequently compared with life safety thresholds and normal health fluctuation thresholds.

[0040] This embodiment regionally locks the micro-motion features and body temperature distribution features in the real-time human digital model after an abnormal behavior event is triggered, and generates structured vital sign data based on a unified timestamp. It can obtain key indicators related to cardiopulmonary activity under privacy-friendly non-contact perception conditions, providing a consistent and comparable data foundation for subsequent risk and emergency response level determination.

[0041] S50, determine the risk level of the target object based on the type of the abnormal behavior event and the vital signs data; In this embodiment, the type of abnormal behavioral event, serving as the first input condition for determining the risk level for emergency medical treatment, is derived from the analysis results of movement trajectory and posture characteristics. Abnormal behavioral events are categorized into fall-type events characterized by rapid movement and stagnation-type events characterized by prolonged stillness. These two types of events differ significantly in their physiological risk implications; fall-type events typically indicate uncontrolled displacement within a short period, while stagnation-type events typically indicate a prolonged lack of voluntary movement. This distinction is used to assign different weights when subsequently combined with vital sign data.

[0042] Vital signs data, serving as the second input for determining the risk level of emergency care, includes respiratory rate, heart rate, and body temperature. Respiratory rate and heart rate reflect the target individual's respiratory and cardiac rhythms, while body temperature reflects any abnormal fluctuations in body surface temperature. The vital signs data are derived from a comprehensive calculation of micromotion characteristics and body temperature distribution characteristics, with a unified timestamp ensuring that all three correspond to the same physiological state within the same time window. At this stage, vital signs data are no longer subjected to spectral or spatial calculations but are directly used as structured numerical values ​​for threshold comparison.

[0043] The risk level assessment is based on a dual-threshold system: a life safety limit threshold and a normal health fluctuation threshold. The life safety limit threshold determines whether a life-threatening extreme state has been reached, while the normal health fluctuation threshold determines whether the state deviates from the normal physiological fluctuation range. Respiratory rate, heart rate, and body temperature are compared with their corresponding life safety limit and normal health fluctuation thresholds to determine three possible outcomes: outside the safety limit, outside the normal fluctuation range but not reaching the safety limit, and within the normal fluctuation range.

[0044] The type of abnormal behavioral event and the three types of outcomes together constitute the basis for determining the risk emergency response level. When the abnormal behavioral event is a fall and any vital sign reaches the life safety limit threshold, the risk emergency response level is determined to be Level 1 Critical Emergency Response. This determination reflects a high-risk combination of fall accompanied by physiological abnormalities. When the abnormal behavioral event is a fall and vital signs do not reach the life safety limit threshold but deviate from the normal health fluctuation threshold, the risk emergency response level is determined to be Level 2 Abnormal Warning Level. This determination reflects the existence of potential physiological abnormalities after the fall but not yet reaching a critical state. When the abnormal behavioral event is a fall and all vital signs are within the normal health fluctuation threshold range, the risk emergency response level is determined to be Level 3 Routine Monitoring Level.

[0045] When the abnormal behavior event is a detention event and any vital sign touches the life safety limit threshold, the risk first aid level is determined as the first-level critical first aid level. This determination reflects the situation of long-term immobility accompanied by dangerous vital signs. When the abnormal behavior event is a detention event and the vital signs do not touch the life safety limit threshold but there are deviations from the healthy normal fluctuation threshold, the risk first aid level is determined as the second-level abnormal warning level. When the abnormal behavior event is a detention event and all vital signs are within the healthy normal fluctuation threshold range, the risk first aid level is determined as the third-level daily monitoring level. Through the above combined determination, a complete mapping relationship is formed between the type of abnormal behavior event and the vital sign data, avoiding the state of being unable to classify.

[0046] In this implementation, by jointly determining the type of abnormal behavior event and the vital sign data in two dimensions and using a dual-threshold system to divide the physiological state, the determination of the risk first aid level reflects both the degree of behavioral abnormality and the degree of physiological abnormality, improving the accuracy and interpretability of risk identification.

[0047] S60, execute the corresponding hierarchical emergency dispatching strategy according to the risk first aid level.

[0048] In this embodiment, the risk first aid level, as the dispatching trigger condition, comes from the result of the comprehensive determination of the type of abnormal behavior event and the vital sign data. The risk first aid level is divided into the first-level critical first aid level, the second-level abnormal warning level, and the third-level daily monitoring level. Different levels correspond to different external interaction behaviors and local response behaviors. The execution of the hierarchical emergency dispatching strategy depends on this determination result of the risk first aid level, and no longer repeats the posture analysis or vital sign calculation, avoiding repeated processing.

[0049] When the risk first aid level is the first-level critical first aid level, the system enters the high-priority linkage state. In this state, it is necessary to establish the external communication ability. The communication behaviors include initiating a communication connection request to the remote first aid dispatching center and pushing the identity file information and vital sign data of the target object. The identity file information comes from the pre-stored personal health information, and the vital sign data comes from the generated structured respiratory rate value, heart rate value, and body temperature status index. After the communication connection request is established, a stable data transmission state is maintained to ensure that the first aid dispatching center can obtain the current physiological state and abnormal information of the target object in real time.

[0050] When the risk emergency response level reaches Level 2 (anomaly warning level), the system enters a local interactive confirmation state. In this state, the intelligent interactive terminal located within the monitoring area plays a status inquiry voice command to inquire whether the target object is in a normal state. Simultaneously, the voice acquisition channel is activated to receive feedback from the target object. The voice acquisition channel continuously monitors the target object's audio feedback within a preset waiting time threshold. If no valid user feedback signal is detected within the waiting time threshold, an anomaly alarm notification is generated and sent to the preset guardian's mobile terminal, completing the transition from local confirmation to remote notification.

[0051] When the risk emergency response level is Level 3 (daily monitoring level), the system enters a recording and continuous data acquisition state. In this state, no external alarms are triggered. Instead, the timestamps and types of currently identified abnormal behavioral events are written to the system's operation log for subsequent health behavior analysis. Simultaneously, continuous acquisition of non-contact motion sensing data and thermal distribution imaging data is maintained, ensuring uninterrupted monitoring and providing continuous data support for potential risk changes.

[0052] This implementation directly maps risk and emergency response levels to different external linkage behaviors and local interaction behaviors, enabling anomaly identification results to be instantly transformed into specific emergency response actions, improving the timeliness and accuracy of emergency response, while avoiding false alarms in low-risk states.

[0053] In one embodiment, step S10 above includes: S101 controls the millimeter-wave radar set in the monitoring area to transmit frequency-modulated continuous wave signals and receive echo signals reflected by the target object; S102, perform fast Fourier transform processing and clutter suppression processing on the echo signal in the distance and velocity dimensions, and extract the point cloud sequence containing three-dimensional spatial coordinates and micro Doppler spectrum as non-contact motion sensing data. S103 controls the infrared thermal imaging sensor to collect information on the distribution of infrared radiation flux within the monitoring area; S104, Based on the infrared radiation flux distribution information, temperature quantization mapping processing is performed to generate a two-dimensional temperature numerical matrix as thermal distribution imaging data. S105, mark the non-contact motion sensing data and the thermal distribution imaging data with a unified timestamp to achieve time synchronization.

[0054] In this embodiment, the monitoring area is used to define the sensor coverage and data sampling range. The boundary of the monitoring area can be determined by the room's planar coordinates, the available passageway mask, or the sensor's effective field of view parameters, so that subsequent data processing is carried out only around the target object's activity space. Non-contact motion sensing data emphasizes acquiring motion information without physical contact with the target object. The data source adopts the electromagnetic wave echo measurement mechanism of millimeter-wave radar, which uses the transmission of frequency-modulated continuous wave signals and the reception of echo signals reflected by the target object to form raw measurement data. The frequency of the frequency-modulated continuous wave signal changes with time. The beat frequency component between the echo signal and the transmitted signal is used to characterize the target object's distance information, and the change in the echo phase over time is used to characterize the target object's velocity information. Together, they form the input basis for subsequent extraction of three-dimensional spatial coordinates and micro-Doppler spectrum.

[0055] The echo signal undergoes Fast Fourier Transform (FFT) processing in both the range and velocity dimensions to map the time-domain sampling sequence to the frequency-domain feature space. The range-dimensional FFT mapping maps the beat frequency component to a range spectrum, allowing for the separation of reflection intensities at different range units. The velocity-dimensional FFT performing frequency-domain expansion of the phase changes of continuous pulses or frames allows for the separation of different radial velocity components. Clutter suppression is used to reduce interference from static background reflections and low-frequency fixed echoes on the target object. Clutter sources include fixed scatterers such as walls, furniture, and the ground, as well as slowly drifting environmental components. Clutter suppression can be achieved through background modeling, mean removal, sliding window high-pass filtering, or constant false alarm rate (CFAR) detection thresholds, resulting in subsequent point cloud sequences and micro-Doppler spectra that more effectively represent the motion and micro-movements of the target object.

[0056] Point cloud sequences, as part of non-contact motion sensing data, are crucial for representing the three-dimensional spatial coordinates of points. These coordinates can be determined by range estimation, azimuth estimation, and elevation estimation. Range estimation originates from the peak position of the range spectrum, while elevation estimation comes from the phase difference or beamforming output of the antenna array. The point cloud sequence organizes a set of points from multiple sampling frames in temporal order. Each frame reflects the spatial scattering distribution of the target object at the same sampling moment, and the changes in point distribution between consecutive frames characterize the motion process. Micro-Doppler spectra, another component of non-contact motion sensing data, are used to characterize the frequency shift modulation features caused by the target object's minute periodic motion. The micro-Doppler spectrum is derived from the energy distribution matrix obtained by short-time frequency domain analysis or time-frequency transformation of the echo signal. Low-frequency and high-frequency components in the spectrum correspond to slow motion and rapid micro-motion, respectively, and the changes in spectral energy over time support subsequent attitude change recognition and abnormal behavior event determination.

[0057] Thermal distribution imaging data emphasizes reflecting the surface temperature distribution of a target object using thermal radiation information. Infrared thermal imaging sensors acquire infrared radiation flux distribution information, which represents the radiation intensity corresponding to different pixel directions within the monitored area. Radiation flux is affected by factors such as the emissivity of the target object's surface, ambient temperature, distance attenuation, and the sensor's response curve. Therefore, temperature quantization mapping processing is used to convert the radiation flux distribution information into a computable temperature numerical expression. Temperature quantization mapping processing can include nonlinear calibration curve mapping, emissivity compensation, ambient temperature compensation, and dead pixel repair, ensuring that each pixel corresponds to a temperature value. A two-dimensional temperature numerical matrix serves as the thermal distribution imaging data; the matrix's row and column indices correspond to the thermal imaging pixel coordinates, and the matrix elements are the temperature values ​​of those pixels. This two-dimensional temperature numerical matrix can be directly used for subsequent regional temperature statistics, body temperature distribution feature extraction, and target object temperature anomaly identification.

[0058] The time synchronization of non-contact motion sensing data and thermal imaging data relies on a unified timestamp marking mechanism. A unified timestamp is used to align millimeter-wave radar sampling frames with infrared thermal imaging sampling frames, ensuring that different sensors output correlateable data pairs under the same time reference. The timestamp can originate from the system clock, external synchronization pulses, or network time protocol synchronization results; the timestamp accuracy needs to cover the alignment requirements at the sampling period level. After marking non-contact motion sensing data and thermal imaging data with a unified timestamp, subsequent fusion processing can select corresponding frames based on the same timestamp, avoiding deviations in attitude and temperature feature correlation caused by cross-time mismatches, and providing a consistent time reference for abnormal behavior event identification and vital sign data generation.

[0059] This embodiment transmits frequency-modulated continuous wave signals and receives echo signals using millimeter-wave radar. Combining fast Fourier transform processing in the range and velocity dimensions with clutter suppression processing, it generates non-contact motion sensing data containing a point cloud sequence with three-dimensional spatial coordinates and a micro-Doppler spectrum. Simultaneously, it collects infrared radiation flux distribution information through an infrared thermal imaging sensor and performs temperature quantization mapping processing to generate thermal distribution imaging data with a two-dimensional temperature numerical matrix. It also marks the two types of data with a unified timestamp to achieve time synchronization, so that motion information and temperature information can be aligned under the same time reference. This provides a consistent data foundation for subsequent fusion to construct a real-time human digital model and reduces feature deviations introduced by cross-sensor frame mismatch.

[0060] In one embodiment, step S20 above includes: S201, Establish a joint calibration model that includes the millimeter-wave radar coordinate system and the thermal imaging pixel coordinate system, and determine the coordinate transformation matrix from the millimeter-wave radar coordinate system to the thermal imaging pixel coordinate system. S202, Analyze the three-dimensional point cloud coordinates and micro-Doppler spectrum information in the non-contact motion sensing data; S203, using the coordinate transformation matrix to determine the projected pixel coordinates of each three-dimensional point cloud coordinate on the thermal distribution imaging data plane; S204, based on the projection pixel coordinates, index the local temperature value in the thermal distribution imaging data, bind the local temperature value and the micro-Doppler spectrum information spatiotemporally associated with the three-dimensional point cloud coordinates as attribute features to the corresponding three-dimensional point cloud coordinates, and generate an attribute fusion point cloud set; S205, Perform spatial density clustering analysis on the attribute fusion point cloud set to generate a real-time human digital model that includes the geometric shape, temperature distribution and micro-motion spectrum features of the target object.

[0061] In this embodiment, the fusion of non-contact motion sensing data and thermal distribution imaging data is based on the premise of cross-sensor coordinate unification. A joint calibration model is used to describe the geometric correspondence between the millimeter-wave radar coordinate system and the thermal imaging pixel coordinate system. The millimeter-wave radar coordinate system carries the spatial meaning of three-dimensional point cloud coordinates, while the thermal imaging pixel coordinate system carries the pixel meaning of the two-dimensional temperature numerical matrix. The joint calibration model constrains the relative pose of the two types of coordinates through extrinsic parameter relationships, enabling the same spatial location to be mapped between the three-dimensional point cloud coordinates and the projected pixel coordinates. The coordinate transformation matrix is ​​used to express the transformation operator from the millimeter-wave radar coordinate system to the thermal imaging pixel coordinate system. The matrix can consist of rotational and translational components and, in conjunction with the intrinsic parameters of the thermal imaging model, completes the three-dimensional to two-dimensional projection calculation. The matrix can be obtained based on the corresponding point set formed by a calibration plate, corner reflector array, or multi-point co-view target, by minimizing the reprojection error.

[0062] The analysis of 3D point cloud coordinates and micro-Doppler spectrum information from non-contact motion sensing data is used to separate spatial and dynamic spectral information. The 3D point cloud coordinates represent the spatial distribution of scattering points of the target object in the millimeter-wave radar coordinate system. These coordinates can carry the 3D position derived from range, azimuth, and elevation angles. The micro-Doppler spectrum information represents the modulation characteristics of the echo frequency shift caused by the target object's minute periodic motion. The spectrum information can be organized as a time-frequency two-dimensional matrix or a frequency band energy vector, and bound to the sampling frame timestamp to support subsequent spatiotemporal alignment. The analysis process requires frame-level depackaging of the point cloud sequence, noise point removal, and point intensity threshold filtering. Simultaneously, frequency band selection and energy normalization are performed on the micro-Doppler spectrum information to facilitate dimensional fusion or feature juxtaposition with temperature information.

[0063] The determination of projected pixel coordinates relies on the combined calculation of the coordinate transformation matrix and the projection model. Each 3D point cloud coordinate is first transformed to the equivalent coordinates of the thermal imaging sensor using the coordinate transformation matrix, and then mapped to the pixel coordinates of the thermal distribution imaging data plane using the thermal imaging model. The projection calculation needs to handle field-of-view clipping and occlusion invalid points; when the projected pixel coordinates fall outside the valid range of the thermal distribution imaging data, the corresponding point cloud point does not participate in subsequent attribute binding. To reduce alignment errors caused by cross-sensor sampling differences, the projected pixel coordinates can be combined with a timestamp neighbor frame matching strategy, selecting the 2D temperature value matrix that has the same timestamp as the point cloud frame or the smallest difference for indexing.

[0064] Local temperature indexes are used to write temperature information from thermal distribution imaging data into the attribute space of the point cloud. The local temperature value corresponding to the projected pixel coordinates can be directly taken from the single-pixel value of the two-dimensional temperature matrix, or a more stable temperature estimate can be obtained by neighborhood aggregation centered on the projected pixel coordinates. Neighborhood aggregation can use mean, weighted mean, or median methods to suppress thermal imaging noise and bad pixels. The micro-Doppler spectral information spatiotemporally correlated with the 3D point cloud coordinates is used to attach micro-motion spectral attributes to the same spatial scattering point. Spatiotemporal correlation includes timestamp consistency and spatial proximity. Timestamp consistency ensures synchronization between the spectral source and the point cloud frame, while spatial proximity establishes a matching relationship between point cloud points and spectral resolution units, such as allocating spectral energy to corresponding point cloud points based on distance or angle thresholds. Attribute feature binding writes the local temperature value and micro-Doppler spectral information into the corresponding 3D point cloud coordinates, forming attributed point records. The attribute-fused point cloud set consists of multiple attributed points, containing both geometric spatial structures and multimodal attributes such as temperature distribution and micro-motion spectrum.

[0065] Spatial density clustering analysis is used to separate the main structure of a target object from an attribute-fused point cloud set and form a stable digital representation that can be used for subsequent identification. Based on the spatial proximity of point clouds, clustering analysis forms cluster structures according to the distance between points and the density of their neighborhoods. This can separate the target object point cloud from background scattering points and maintain the continuity of the main object's outline even under conditions of sparse, occluded, or locally missing point clouds. Attribute constraints can be introduced during the clustering process, using temperature distribution features as a screening condition to enhance focus on human body regions. For example, clusters with temperatures higher than the environmental baseline can be preferentially retained, while static background clusters can be filtered out using micro-Doppler spectral energy distribution. Clustering results are used to generate real-time human digital models. The model takes the geometric shape of the target object as the main framework. The geometric shape can be expressed by the boundary of point cloud clusters, skeleton curves, voxel occupancy grids or parameterized bounding volumes. Temperature distribution is expressed by point attribute mapping or projected thermal distribution imaging data plane to form thermal distribution. The micro-motion spectrum features of vital signs are expressed by point attribute frequency band energy, spectrum peak trajectory or energy of selected frequency bands changing over time. Thus, the real-time human digital model has both observable posture changes and extractable vital sign features.

[0066] This embodiment determines the coordinate transformation matrix from the millimeter-wave radar coordinate system to the thermal imaging pixel coordinate system through a joint calibration model. This matrix is ​​then used to map the three-dimensional point cloud coordinates to projected pixel coordinates. The projected pixel coordinates are then used to index the local temperature values, and the local temperature values ​​are bound to the spatiotemporally correlated micro-Doppler spectrum information to the corresponding three-dimensional point cloud coordinates, forming an attribute fusion point cloud set. Subsequently, spatial density clustering analysis is performed on the attribute fusion point cloud set to generate a real-time human digital model that simultaneously includes geometric morphology, temperature distribution, and micro-motion spectrum features of vital signs. This allows spatial posture expression, temperature expression, and micro-motion spectrum expression to be synchronously organized within the same object structure, reducing the risk of inconsistency between posture and vital sign expression caused by cross-modal information separation, and improving the consistency and usability of the inputs required for subsequent abnormal behavior event recognition and vital sign data generation.

[0067] In one embodiment, step S30 above includes: S301, continuously analyze the spatial centroid coordinates of the real-time human digital model, and determine the instantaneous descent velocity and ground-touching acceleration of the spatial centroid in the vertical direction as motion trajectory features; S302, determine the size of the three-dimensional bounding box of the real-time human digital model, and determine the subject orientation as a posture feature based on the ratio of the long side to the short side of the three-dimensional bounding box. S303, Obtain historical behavioral habit data of the target object, and determine the matching probability between the current time period and current spatial location and the historical behavioral habit data as the basis for the judgment of the behavioral situation analysis model; S304, when the instantaneous descent speed exceeds a preset fall threshold and the subject is facing a horizontal direction, or when the target object is determined to be in a stationary state based on the posture characteristics and the matching probability is continuously lower than a preset normal behavior confidence level for more than a preset duration threshold, an abnormal behavior event is confirmed to have occurred.

[0068] In this embodiment, the real-time human digital model is used to carry the spatial morphology and temporal evolution information of the target object. The spatial centroid coordinates are used to describe the geometric center position of the overall distribution of the target object. The spatial centroid coordinates can be calculated from the point set or voxel set belonging to the target object in the real-time human digital model. The calculation process can be obtained by averaging the three-dimensional coordinates, finding the weighted center of the voxel occupancy distribution, or finding the center point of the cluster boundary. Continuous analysis of the spatial centroid coordinates is used to form a time series. The vertical displacement change is obtained by differentiating the spatial centroid coordinates of adjacent sampling times. The vertical direction can be defined by the gravity direction of the monitoring area or the preset vertical axis of the coordinate system. The instantaneous descent velocity is used to characterize the vertical displacement change amplitude per unit time. During calculation, the vertical displacement difference can be divided by the time difference of adjacent sampling times, and the velocity series is smoothed by a window, outlier removal, and sampling interval calibration are performed to suppress the interference of measurement noise on the threshold determination. Ground contact acceleration is used to characterize the intensity of velocity changes during descent. It can be obtained by further differentiating the instantaneous descent velocity sequence and combined with the near-ground height condition or velocity change condition of the ground contact event to identify the acceleration peak at the moment of ground contact. Ground contact acceleration, as a feature of the motion trajectory, is used in conjunction with instantaneous descent velocity to distinguish between slow sitting down, lying down and sudden falling behavior.

[0069] The 3D bounding box size describes the circumscribed range of the geometric shape of a target object in a real-time human digital model. The bounding box is generated by calculating the minimum and maximum values ​​of the 3D coordinates of the target object's point set, yielding length parameters along each axis. The ratio of the long side to the short side expresses the extension direction of the shape. The long side can be the maximum of the three axial lengths of the bounding box, while the short side can be the minimum or second minimum, selected based on the required pose recognition accuracy. The subject's orientation maps the ratio to an orientation category. The mapping rule can be based on threshold intervals; for example, a ratio exceeding a certain threshold might be classified as horizontal, while a ratio within another threshold interval might be classified as vertical or tilted. To enhance the stability of orientation determination, the bounding box size can be modified by incorporating the principal component orientations of the target object's point set, or a time window voting process can be applied to the bounding box size sequence to avoid orientation fluctuations caused by short-term occlusion affecting subsequent abnormal behavior event determination.

[0070] Historical behavioral habit data is used to express the behavioral distribution patterns of the target object in the temporal and spatial dimensions. Data sources may include statistical results such as activity time distribution, dwelling location distribution, common posture change patterns, sleep periods, and daily routines within the historical monitoring period. The current time period is used to represent the intraday interval to which the current moment belongs, and the current spatial location is used to represent the location area division result of the target object within the monitoring area. The spatial location can be obtained by mapping the spatial centroid coordinates to the area grid or functional area labels. The matching probability is used to characterize the degree of consistency between the current time period and the current spatial location in the historical behavioral habit data. The matching probability can be formed by conditional frequency, Bayesian posterior probability, kernel density estimation score, or similarity normalization score, and decay weighting is applied to different historical windows to reflect the update of habits over time. The judgment criterion of the behavioral context analysis model is based on the matching probability as input, used to distinguish between dwellings that conform to daily patterns and abnormal dwellings that deviate from the patterns. A matching probability that is consistently lower than the preset normal behavior confidence level indicates that the deviation between the current context and historical habits has reached the attention threshold.

[0071] The confirmation of abnormal behavior events is based on a joint condition of motion trajectory features, posture features, and matching probability. The fall-related determination branch uses a combination of instantaneous descent velocity and the subject's orientation. An instantaneous descent velocity exceeding a preset fall threshold indicates a rapid vertical fall, while a horizontal orientation indicates a change in shape from vertical to horizontal. This combined constraint suppresses misjudgments caused by a single velocity threshold. The stationary determination branch uses a combination of static state and matching probability. A static state can be determined by indicators such as the displacement amplitude of the spatial centroid coordinates within a time window being lower than a static threshold, the change amplitude of the bounding box size being lower than a change threshold, and the micro-motion spectrum energy in the body motion frequency band being lower than a body motion threshold. Posture features provide shape stability constraints. A matching probability consistently lower than a preset normal behavior confidence level exceeding a preset duration threshold is used to introduce a persistent judgment. The preset duration threshold can be defined by the number of consecutive sampling frames, cumulative time, or a sliding window meeting a proportional requirement, preventing brief deviations from habits or short periods of stillness from being judged as abnormal behavior events. The two types of branches are combined using OR logic to cover both rapid, sudden events and slow, persistent events, ensuring that abnormal behavior events have clear triggering conditions and a reproducible judgment process.

[0072] This embodiment obtains instantaneous descent velocity and ground-touching acceleration as motion trajectory features by continuously analyzing the spatial centroid coordinates of a real-time human digital model. It then determines the subject's orientation as a posture feature based on the size of the three-dimensional bounding box and the ratio of the long side to the short side. Furthermore, it incorporates historical behavioral habit data to calculate the matching probability between the current time period and the current spatial location as the basis for the behavioral context analysis model. Abnormal behavioral events are confirmed when the instantaneous descent velocity exceeds a preset fall threshold and the subject's orientation is horizontal; or when the subject is in a static state and the matching probability remains below a preset normal behavior confidence level for a preset duration threshold. This allows abnormal identification to simultaneously incorporate kinematic change constraints, geometric shape constraints, and contextual consistency constraints, reducing the confusion probability between normal sitting and accidental falls, and between normal sleep and abnormal stagnation. This improves the distinguishability and stability of abnormal behavioral event determination, providing a more reliable triggering basis for subsequent vital sign data generation and risk emergency response level determination.

[0073] In one embodiment, step S40 above includes: S401, in response to the detection of the abnormal behavior event, locate the chest and abdomen region with the highest thermal radiation intensity in the real-time human digital model, and extract the spatial coordinates and average temperature value of the chest and abdomen region as body temperature distribution features. S402, according to the spatial coordinates, control the detection beam of the millimeter-wave radar to focus on the chest and abdomen region, and continuously collect the micro-Doppler phase modulation signal of the chest and abdomen region as micro-motion characteristics; S403, perform adaptive mode decomposition and spectrum analysis on the micro-motion feature to separate the first signal component corresponding to the respiratory frequency and the second signal component corresponding to the heart rate. S404, based on the first signal component, the second signal component, and the average temperature value, generate vital sign data including respiratory rate, heart rate, and body temperature status indicators.

[0074] In this embodiment, abnormal behavior events serve as triggering conditions to define the start time and processing window for vital sign data generation. The trigger time can be determined by the confirmation timestamp of the abnormal behavior event, and forward and backward sampling time windows are configured around this timestamp to capture synchronous segments of millimeter-wave radar echo signals and infrared radiation flux distribution information, avoiding the mixing of body movement or temperature changes during non-abnormal periods into the vital sign data. The real-time human digital model carries a fusion representation of spatial geometry, temperature distribution, and micro-Doppler spectrum. The locking operation relies on the spatial and temporal consistency of the real-time human digital model, limiting the micro-motion characteristics of cardiopulmonary activity and body temperature distribution characteristics to the same chest and abdominal region. The locking process includes five continuous actions: region localization, beam focusing, signal acquisition, signal separation, and parameter generation.

[0075] The chest and abdomen region is used to define the spatial range of body temperature distribution and micromotion characteristics. In a real-time human digital model, this region can be represented by a set of projected pixel coordinates, a subset of 3D point clouds, or voxel sub-blocks. The highest thermal radiation intensity is used to determine candidate locations for the chest and abdomen region. This intensity is derived from the temperature quantization mapping results of thermal distribution imaging data. The identification method can be to perform sliding window aggregation on the thermal distribution imaging data, calculate the temperature energy or mean temperature for each window, select an extreme value window, and then map the extreme value window back to the corresponding spatial unit in the real-time human digital model. Spatial coordinates describe the positioning results of the chest and abdomen region in a millimeter-wave radar coordinate system or a unified coordinate system. These coordinates can be obtained by calculating the spatial centroid of a subset of the chest and abdomen region's point cloud, the geometric center of a voxel sub-block, or by fitting the minimum bounding box center to the region's boundary points. The average temperature value is used to describe the temperature level in the chest and abdomen region. It can be calculated from the set of local temperature values ​​corresponding to the chest and abdomen region. This set of local temperature values ​​is obtained from the pixel index of the thermal distribution imaging data. The calculation process may include outlier pixel removal, thermal noise correction, emissivity compensation, and robust mean calculation within a time window to avoid temperature drift caused by thermal imaging jitter. Body temperature distribution characteristics, expressed by both spatial coordinates and the average temperature value, represent the location and temperature of the region, making subsequent vital sign data both localizable and interpretable.

[0076] The millimeter-wave radar's detection beam focusing is used to concentrate the energy of the micro-Doppler phase-modulated signal onto the chest and abdomen region. Control can be achieved through beamforming parameter distribution, phased array phase weight configuration, or selecting the direction of the strongest echo after multi-beam scanning. Spatial coordinates serve as the beam control input, requiring a coordinate transformation from the real-time human digital model coordinates to the millimeter-wave radar coordinate system. This transformation can be achieved using a coordinate transformation matrix provided by the joint calibration model, yielding beam pointing parameters such as azimuth, elevation, and range thresholds. Continuous acquisition covers multiple respiratory and cardiac cycles. The acquisition duration can be determined by a time window set after an abnormal behavioral event is triggered. The sampling rate can be correlated with the millimeter-wave radar's frequency-modulated continuous wave configuration to meet the micro-Doppler spectral resolution requirements. Micro-Doppler phase-modulated signals are used to characterize phase changes caused by minute displacements in the chest and abdominal region. Signal formation can be achieved by phase extraction, phase expansion, phase detrending, and bandpass pre-filtering of the echo signal. The filtered frequency band can cover the frequency range of respiratory and heart rate, and the body movement frequency band is suppressed to reduce the interference of large movements on micro-motion characteristics. Micro-motion characteristics are expressed as time-domain sequences or time-frequency representations of micro-Doppler phase-modulated signals, providing a computational object for subsequent separation of respiratory and heart rate components.

[0077] Adaptive mode decomposition (ADD) is used to decompose micro-motion features into multiple intrinsic mode components with narrowband characteristics. Adaptability is reflected in the dynamic determination of the number of decomposition levels, stopping criteria, and mode selection thresholds based on the energy distribution, noise level, and non-stationarity of the micro-motion features. In implementation, endpoint extension can be applied to the micro-motion features to suppress boundary effects, a convergence threshold can be set for the residual energy, and an interval constraint can be set for the mode center frequency to reduce mode aliasing. Spectral analysis is used to calculate the power spectral density, instantaneous frequency, or short-time Fourier spectrum for each mode component, and to select the mode or mode combination corresponding to the main peak frequency within the candidate ranges for respiratory frequency and heartbeat frequency, respectively. The first and second signal components are used to carry the separation results. The first signal component corresponds to the respiratory frequency, and its selection can be based on the stability and periodicity of the low-frequency main peak. The second signal component corresponds to the heartbeat frequency, and its selection can be based on the significance of the mid-to-high frequency main peak and the consistency of the harmonic structure. The separation process may include phase consistency verification of the first signal component and the second signal component, peak confidence assessment and abnormal peak removal, to avoid misjudging the residual body motion frequency as respiratory frequency or heart rate.

[0078] Respiratory rate, heart rate, and body temperature status indicators constitute the structured output of vital sign data. Respiratory rate can be calculated from the dominant frequency of the first signal component, mapping the frequency unit to frequencies per minute. Robust statistics can be performed on the respiratory rate sequence within a time window to suppress transient disturbances. Heart rate can be calculated from the dominant frequency of the second signal component, and the selection of the dominant peak can be corrected by combining the consistency checks of the harmonics and half-frequency of the heartbeat. Body temperature status indicators are mapped from the average temperature value. The mapping method can be threshold interval mapping, piecewise linear mapping, or deviation mapping based on individual baseline temperature. Individual baseline temperature can be obtained statistically from historical thermal distribution imaging data during non-abnormal periods. The values ​​of body temperature status indicators can be discrete labels or continuous scores, so that they, along with respiratory rate and heart rate, form a unified format of vital sign data. Vital signs data can structurally include field names, field values, timestamps, spatial coordinates of the chest and abdomen region, and confidence information. The confidence information can be generated by the signal-to-noise ratio of the micro-Doppler phase modulation signal, the significance of the main modal peak, and the stability of the temperature value, which facilitates the use of reliable input for subsequent risk and emergency treatment level determination.

[0079] This embodiment locates the chest and abdomen region with the highest thermal radiation intensity in a real-time human digital model after an abnormal behavioral event is triggered, and extracts its spatial coordinates and average temperature value to form body temperature distribution characteristics. Then, the spatial coordinates are used to control the focusing of the detection beam of the millimeter-wave radar and continuously collect the micro-Doppler phase modulation signal of the chest and abdomen region to form micro-motion characteristics. Subsequently, adaptive mode decomposition and spectrum analysis are performed on the micro-motion characteristics to separate the first signal component and the second signal component. Based on the first signal component, the second signal component and the average temperature value, vital sign data including respiratory rate, heart rate and body temperature status indicators are generated. This makes the vital sign parameters spatially bound to the chest and abdomen region, temporally synchronized with the abnormal behavioral event, and achieves the separable expression of respiratory and heartbeat components in terms of signal. This reduces the impact of large-scale body movement and temperature noise on vital sign estimation and improves the stability and usability of vital sign data as input for subsequent risk emergency response levels.

[0080] In one embodiment, step S50 above includes: S501, determine whether the abnormal behavior event is a fall event with rapid movement characteristics or a lingering event with static and continuous characteristics. S502, Obtain respiratory rate, heart rate and body temperature status indicators from the vital signs data; S503, compare the respiratory rate value, the heart rate value and the body temperature status index with the corresponding life safety limit threshold and the normal health fluctuation threshold respectively; S504, when it is determined that the abnormal behavior event is a fall-related event, and at least one of the respiratory rate value, the heart rate value, or the body temperature status index is lower than or exceeds the corresponding life safety limit threshold, the risk emergency level is determined to be Level 1 critical emergency level. S505, when it is determined that the abnormal behavior event is a fall-related event, and the respiratory rate value, the heart rate value and the body temperature status index have not reached the corresponding life safety limit threshold, but at least one of them deviates from the corresponding normal health fluctuation threshold, the risk emergency level is determined to be a level two abnormal warning level. S506, when it is determined that the abnormal behavior event is a fall-related event, and the respiratory rate, heart rate and body temperature indicators are all within the corresponding normal health fluctuation threshold range, the risk emergency level is determined to be the third level of daily monitoring. S507, when it is determined that the abnormal behavior event is a lingering event, and at least one of the respiratory rate value, the heart rate value, or the body temperature status index is lower than or exceeds the corresponding life safety limit threshold, the risk emergency level is determined to be Level 1 critical emergency level. S508, when it is determined that the abnormal behavior event is a lingering event, and the respiratory rate value, the heart rate value and the body temperature status index have not reached the corresponding life safety limit threshold, but at least one of them deviates from the corresponding normal health fluctuation threshold, the risk emergency level is determined to be a level two abnormal warning level. S509, when the abnormal behavior event is determined to be a lingering event, and the respiratory rate, heart rate and body temperature are all within the corresponding normal health fluctuation threshold range, the risk emergency level is determined to be Level 3 daily monitoring level.

[0081] In this embodiment, the type of abnormal behavior event is used to limit the grading entry point for risk emergency response, allowing the same set of vital sign data to be judged using different combinations of judgment thresholds and level mapping relationships under different event contexts. The type of abnormal behavior event is obtained from the classification conclusions of action trajectory features and posture features in the judgment results of abnormal behavior events. Fall events with rapid movement features correspond to a high amplitude change in the instantaneous descent velocity and ground acceleration of the spatial centroid coordinate in the vertical direction, accompanied by a change in the subject's orientation from the vertical to the horizontal direction. Stagnation events with static and continuous features correspond to a relatively stable subject orientation and a low amplitude change in the spatial centroid coordinate, while the matching probability with historical behavioral habit data remains lower than the preset normal behavior confidence level within a preset duration threshold. Fall events and stagnation events are classified and output as mutually exclusive. The classification output can be recorded as event type labels and type confidence levels. The type confidence level can be generated by the consistency of instantaneous descent velocity, ground acceleration, subject orientation judgment, and the degree of deviation in matching probability, and is used for the confidence weighting of subsequent threshold comparison results.

[0082] Vital signs data, serving as the input data carrier for risk and emergency response level determination, includes three elements: respiratory rate, heart rate, and body temperature status indicators. These three elements are associated with the spatial coordinates and average temperature values ​​of the chest and abdomen region during the generation phase and are consistent with the trigger time of the abnormal behavioral event. When extracting respiratory rate values ​​from vital signs data, unit standardization can be selected, mapping the respiratory rate to a real value or fixed-point value per minute, and simultaneously extracting the corresponding confidence field for subsequent anomaly confirmation. Heart rate value extraction can include range validation and smoothing. Range validation filters out obviously distorted values ​​through a preset physiologically reasonable interval, while smoothing robustly aggregates continuous heart rate estimates within a short time window after the abnormal behavioral event is triggered, outputting a representative heart rate value. Body temperature status indicators are either discrete labels or continuous scores. During extraction, the original expression of the body temperature status indicator is preserved, and it can be associated with a numerical version of the average temperature value, enabling threshold comparison to support both label-based mapping determination and numerical interval determination.

[0083] A two-tiered threshold system, consisting of life safety limit thresholds and routine health fluctuation thresholds, maps respiratory rate, heart rate, and body temperature indicators to three levels: critical, warning, and routine monitoring. Life safety limit thresholds mark the critical range where vital signs threaten safety, while routine health fluctuation thresholds mark the boundaries of vital signs deviating from their normal stable range. The correspondence requires that each parameter has a corresponding life safety limit threshold and a corresponding routine health fluctuation threshold. The life safety limit threshold for respiratory rate can be composed of a lower and upper limit threshold; the life safety limit threshold for heart rate can be composed of a lower and upper limit threshold; and the life safety limit threshold for body temperature indicators can be expressed as a critical range or a set of critical labels. Routine health fluctuation thresholds can also be expressed as upper and lower limits or a set of intervals, allowing correlation with individual baseline differences. Individual baseline differences can be statistically obtained from historical vital sign data during non-abnormal periods, with the results expressed as individual baseline respiratory rate, individual baseline heart rate, and individual baseline temperature intervals, thus allowing for individualized configuration of routine health fluctuation thresholds. Thresholds can be stored in a threshold configuration table, which includes fields such as parameter name, threshold type, threshold boundary, applicable age group, and applicable monitoring period, making it easy to select the corresponding threshold for different target objects and different time periods.

[0084] The comparison operation performs two types of judgments for each parameter: comparison of life safety limit thresholds outputs the life safety limit judgment result, and comparison of routine health fluctuation thresholds outputs the routine health fluctuation judgment result. Values ​​below or above the corresponding life safety limit threshold are used to uniformly describe situations where the parameter falls into the critical range on either side; values ​​deviating from the corresponding routine health fluctuation threshold are used to describe situations where the parameter exceeds the daily fluctuation range but has not yet entered the life safety limit range; values ​​within the corresponding routine health fluctuation threshold range are used to describe situations where the parameter remains within the daily fluctuation range. The body temperature status indicator supports two implementation paths during comparison: when the body temperature status indicator is a continuous score, it is directly compared with the threshold range; when the body temperature status indicator is a discrete label, a label-to-interval mapping table is used for comparison. The mapping table can map labels such as low fever, high fever, and hypothermia to a set of intervals corresponding to the life safety limit threshold or the routine health fluctuation threshold. The comparison results are expressed as a structured Boolean set, including critical respiratory rate markers, abnormal respiratory rate markers, critical heart rate markers, abnormal heart rate markers, critical body temperature markers, and abnormal body temperature markers, and retains the threshold boundary values ​​that cause the markers to be valid for subsequent record auditing.

[0085] The risk emergency response level is determined using a combination of event type priority and parameter anomaly priority mapping method. The mapping process requires coverage of all input combinations for fall-related and lingering events, while maintaining mutual exclusivity. Level 1 Critical Emergency Response Level is used to output when the life safety limit assessment result is true. The triggering condition remains consistent for both fall-related and lingering events, based on at least one of the following indicators: respiratory rate, heart rate, or body temperature, being lower than or exceeding the corresponding life safety limit threshold. The judgment logic for at least one indicator corresponds to a Boolean set OR operation, ensuring that any parameter triggers the critical status flag to output Level 1 Critical Emergency Response Level. Level 2 Abnormal Warning Level is used to output when the life safety limit assessment result is not triggered but the health routine fluctuation assessment result is true. The triggering condition remains consistent for both fall-related and lingering events, based on the following pre-constraint: respiratory rate, heart rate, and body temperature indicators not reaching the corresponding life safety limit threshold, and at least one of them deviating from the corresponding health routine fluctuation threshold. The pre-constraint corresponds to all critical status flags being false, and the post-constraint corresponds to the existence of a true value in the abnormal status flag. The Level 3 daily monitoring level is used to handle outputs where all critical status markers and all abnormal status markers are false. The triggering conditions remain consistent for both fall-related and lingering events, based on the respiratory rate, heart rate, and body temperature all falling within their corresponding normal health fluctuation thresholds, and both critical and abnormal status markers in the corresponding Boolean set being false. The level output can be expressed using both a level code and a level name. The level code is used for rapid matching of subsequent scheduling strategies, while the level name is used for recording and display. The code and name maintain a fixed mapping within the system.

[0086] Fall-related events and lingering events each correspond to three level branches, maintaining a priority order among them: the life safety limit threshold branch takes precedence over the routine health fluctuation threshold branch, and the routine health fluctuation threshold branch takes precedence over branches within the routine health fluctuation threshold range. This priority order is guaranteed by a conditional judgment structure, which can be implemented using short-circuit logic or a state machine transition table. The state machine transition table contains four columns: event type state, critical condition marker state, abnormality marker state, and output level state. The critical condition marker state prioritizes matching the Level 1 critical emergency care level, the abnormality marker state matches the Level 2 abnormal warning level, and the rest match the Level 3 routine monitoring level. This structure ensures that the same input will not trigger multiple level outputs, and also ensures that all inputs have a unique level output.

[0087] This embodiment categorizes abnormal behavioral events into mutually exclusive types: fall-related events and lingering events. It extracts respiratory rate, heart rate, and body temperature indicators from vital sign data, and then performs a two-layer comparison between these indicators and their corresponding life safety limit thresholds and normal health fluctuation thresholds. This establishes three judgment branches—life safety limit violation, deviation from normal health fluctuations, and within the normal health fluctuation range—for both fall-related and lingering events, outputting a Level 1 critical emergency response level, a Level 2 abnormal warning level, and a Level 3 daily monitoring level. This enables the risk emergency response level output to jointly respond to the event context and the degree of abnormality in multiple vital signs, improving the coverage completeness and clarity of judgment boundaries at the input combination level, and providing definite level-driven conditions for subsequent tiered emergency dispatch strategies.

[0088] In one embodiment, step S60 above includes: S601, when the risk emergency level is determined to be Level 1 critical emergency level, a communication connection is established with the remote emergency dispatch center, and the identity file information and vital sign data of the target object are pushed through the communication connection. S602, when the risk emergency response level is determined to be a level 2 abnormal warning level, the intelligent interactive terminal located in the monitoring area is driven to play a status inquiry voice command and open the voice acquisition channel to wait for user feedback. S603, if no valid user feedback signal is detected through the voice acquisition channel within the preset waiting time threshold, an abnormal alarm notification is generated and sent to the preset guardian's mobile terminal. S604, when the risk emergency response level is determined to be a Level 3 daily monitoring level, the timestamp and type of the currently identified abnormal behavior event are written into the system operation log, and the continuous collection of non-contact motion sensing data and thermal distribution imaging data is maintained.

[0089] In this embodiment, the risk and emergency response level serves as the trigger condition for the tiered emergency dispatch strategy. The input to the dispatch decision comes from the determination result of the risk and emergency response level, and the output is a combination of external communication actions, local interaction actions, notification actions, and log retention actions that correspond one-to-one with the risk and emergency response level. The tiered emergency dispatch strategy can be implemented using a strategy table-driven approach. The strategy table uses the risk and emergency response level as the index key and associates communication connection request templates, push data field sets, interactive terminal control command sets, waiting time thresholds, notification receiver identifiers, and log writing field sets, making the execution actions corresponding to different risk and emergency response levels configurable and traceable.

[0090] When the emergency risk level is determined to be Level 1 critical, a communication connection request is triggered by the remote emergency dispatch center. This request establishes a session channel to the remote emergency dispatch center and can be carried using cellular networks, broadband networks, or dedicated IoT access. The connection establishment phase may include target address resolution, session key negotiation, message sequence number initialization, and retransmission window settings to ensure that subsequent data delivery can still be confirmed even under network jitter. The payload of the communication connection request carries the target object's identity profile information and vital signs data. The identity profile information is used for object matching and resource allocation on the emergency dispatch side. Fields may include target object identifier, residential address text, emergency contact identifier, allergy history label, past risk warning label, and monitoring area identifier. Field organization can use structured key-value pairs with accompanying version numbers for easy version-based parsing by the remote emergency dispatch center. When vital sign data and identity profile information are pushed within the same connection session, a packet-splitting strategy can be adopted. This strategy distinguishes between static and dynamic fields based on message type. Identity profile information, as a static field, is pushed once after the session is established or resent at preset intervals. Vital sign data, as a dynamic field, is pushed at the event trigger time and can include a collection timestamp and data confidence field, facilitating the remote emergency dispatch center's assessment of data timeliness and reliability. An arrival receipt mechanism can be set during the push process. This mechanism confirms successful reception using the message sequence number. If no receipt is received, a retransmission count and backoff interval are triggered. The backoff interval can increase with the number of consecutive failures, thereby reducing congestion risk in unstable network scenarios. To reduce the resource consumption of falsely triggered outbound calls, the communication connection request can also include a risk emergency level field and a timestamp field for the abnormal behavior event, enabling the remote emergency dispatch center to merge and dispatch multiple times the same object is triggered.

[0091] When the risk emergency response level is determined to be a Level II abnormal warning level, the status inquiry voice command and voice acquisition channel of the intelligent interactive terminal within the monitoring area are triggered. Control of the intelligent interactive terminal can be achieved through LAN control messages, Bluetooth control messages, or commands forwarded by the home gateway. The control commands include playback content identifiers, volume levels, playback counts, playback intervals, and interaction window duration, ensuring the status inquiry voice command remains perceptible under different environmental noise levels. The content of the status inquiry voice command can be generated using a template, with template fields including greetings, confirmation statements, and response prompts to avoid misunderstandings caused by playing only a single prompt. The voice acquisition channel is used to receive voice feedback signals from the target object. Channel activation actions may include microphone array power supply, sampling rate setting, echo cancellation switch setting, and channel buffer initialization. The buffer stores short audio segments in a circular queue and provides a data basis for subsequent detection of valid user feedback signals. The determination of valid user feedback signals can be based on a combination of multiple conditions. These conditions can include a voice energy threshold, a duration threshold, a keyword hit indicator, and a non-environmental noise feature threshold. The keyword hit indicator can be implemented using a preset keyword list, which includes confirmation words and help requests, avoiding misjudgments based solely on sound triggers. To accommodate the feedback methods of people with mobility impairments, the voice acquisition channel can also be compatible with non-voice feedback signal inputs, such as touch or button confirmation inputs from smart interactive terminals, and these inputs can be mapped to valid user feedback signals, thereby reducing missed judgments caused by difficulties in speaking.

[0092] When no valid user feedback signal is detected through the voice acquisition channel within a preset waiting time threshold, an abnormal alarm notification is triggered and sent to the preset guardian's mobile terminal. The waiting time threshold is used to limit the interaction confirmation window. The waiting time threshold can be dynamically configured by time period. A shorter threshold is set during nighttime to reduce the exposure time to disability risks, and a longer threshold is set during daytime to reduce invalid notifications. The abnormal alarm notification can be generated using a structured alarm message. The message fields can include the target object identifier, risk emergency level, timestamp of the abnormal behavior event, type of abnormal behavior event, vital sign data summary, and interaction failure reason identifier. The interaction failure reason identifier is used to mark the failure to detect a valid user feedback signal. The guardian's mobile terminal can send the notification through SMS, in-app push, or voice outbound call channels. Channel selection can be configured by priority queue. Each channel in the queue is configured with a sending timeout and retry count, and switches to the next channel when consecutive failures occur. To reduce interference caused by duplicate notifications, a suppression window can be set for notification sending. When the same target object and the same risk emergency level are triggered consecutively, the suppression window merges the alarms and updates the latest vital sign data summary, presenting a single aggregated notification to the guardian's mobile terminal.

[0093] When the risk emergency response level is determined to be Level 3 (daily monitoring level), system operation log writing and data acquisition status maintenance are triggered. The fields written to the system operation log include the timestamp and type of the abnormal behavior event. Field writing can be appended and include a log index field, which contains the target object identifier and monitoring area identifier for easy subsequent retrieval and auditing. Log records can also include risk emergency response level fields and vital sign data summary fields, making low-risk events statistically significant over long-term observation. Maintaining continuous acquisition of non-contact motion sensing data and thermal distribution imaging data is crucial for monitoring continuity. Continuous acquisition can include sampling period parameters, buffer queue length parameters, frame loss handling strategies, and time synchronization marking strategies. In low-risk scenarios, the sampling period parameter can be set to a lower frequency to reduce equipment load while still retaining the time resolution required for event re-identification. Maintenance of continuous acquisition status can also include sensor health detection fields. These fields record the online status, signal strength statistics, and temperature matrix integrity statistics of millimeter-wave radar and infrared thermal imaging sensors. When an abnormality occurs in the health detection fields, a maintenance alarm is triggered instead of emergency dispatch, thus distinguishing between equipment failure and health risk alarms.

[0094] For example, in a healthy home setting, the monitoring subjects are elderly people living alone or with partial disabilities. Their daily routines include high-risk segments such as getting out of bed at night, using the toilet, washing up in the morning, activities in the living room, and afternoon naps. Nursing services focus more on situations such as slipping in the bathroom, falling at night without being discovered, and prolonged stillness accompanied by abnormal breathing. The monitoring area covers the area around the bed and the passageway to the bathroom, extending to the area near the bathroom door and sink. The space is divided using commonly used semantic areas for home health management, such as "bed-off area," "corridor area," "toilet area," "washing area," and "rest area," so that subsequent judgments can be linked to specific care contexts such as "risk of using the toilet at night," "prolonged stay in the bathroom," and "abnormal falls after getting out of bed." Millimeter-wave radar continuously senses changes in body movement and micro-vibrations in the chest and abdomen, while thermal imaging sensors continuously provide information on the body surface heat distribution. The two are aligned with a unified timestamp to form a data foundation that can be used for home condition assessment. The absence of image acquisition in bedrooms and bathrooms better meets the privacy requirements of home care. The elderly are more accepting of this during private times such as washing, changing, and using the toilet, and caregivers are more willing to extend monitoring coverage to high-risk areas.

[0095] During a stable operating period, the system summarizes daily routines such as "nighttime bedtime, nighttime bed-leaving frequency, average bathroom stay duration, and living room activity time after waking up," forming historical behavioral habit data. One night, an elderly person went to sleep as usual and briefly left bed to go to the bathroom. The spatial centroid coordinates moved slowly along the corridor, the three-dimensional spatial bounding box remained vertical, and the matching probability remained at a high level, consistent with previous nighttime toilet-going behavior. When the elderly person sat down or stood up, there was a change in vertical displacement, but the instantaneous descent speed did not reach the fall threshold, the ground-touching acceleration did not show a peak, and the body orientation did not change to a horizontal direction. The behavioral context analysis model judged this action as normal sitting or standing up and did not generate an abnormal behavior event. Similar false alarm suppression is crucial in the home environment. For example, when an elderly person bends down to pick up an item, sits on the edge of the bed to put on shoes, or briefly stays against a wall, relying solely on the amplitude of body movement can easily lead to misjudgment. Adding the body orientation and matching probability makes it closer to the "autonomous and controllable actions" in the context of home care.

[0096] Suppose an elderly person slips and falls after entering the bathroom. Their spatial centroid coordinates rapidly decrease vertically, resulting in a significant abrupt change in ground acceleration. The three-dimensional bounding box transitions from vertical to horizontal, and the behavioral scenario analysis model confirms this abnormal behavioral event. Upon triggering the event, the system locates the chest and abdomen region in the real-time human digital model. Thermal imaging reveals this region as having significant thermal radiation intensity and provides an average temperature value. Millimeter-wave radar focuses its detection beam based on the spatial coordinates of the chest and abdomen region, continuously acquiring micro-Doppler phase modulation signals. It separates the first signal component related to respiratory rate and the second signal component related to heart rate, generating vital sign data including respiratory rate, heart rate, and body temperature indicators. Common dangerous combinations in home care scenarios include respiratory arrest after a fall, shallow and slow breathing after a fall, and abnormal body temperature deviation after a fall. After vital sign data are compared with thresholds, when the respiratory rate or heart rate reaches the life safety limit threshold, the risk emergency level is directly set to Level 1 critical emergency level. The tiered emergency dispatch strategy then initiates a communication connection request to the remote emergency dispatch center, pushing identity information and vital sign data. The caregiver obtains a clear summary of "fall location, trigger time, and vital sign status," which facilitates rapid assessment and configuration of treatment measures by emergency personnel upon arrival. The home service station can also coordinate on-site response resources based on this information.

[0097] Another common high-risk aspect of home-based elderly care is "prolonged immobility" accompanied by physiological abnormalities, especially occurring in the bathroom or bedside. For example, an elderly person may enter the bathroom at 3 AM and remain motionless for an extended period, exhibiting extremely low spatial centroid displacement and a stable orientation. The probability of this behavior consistently falls below the confidence level for normal behavior and exceeds the duration threshold, thus confirming the abnormal behavior event. In nursing contexts, this typically corresponds to "risk of prolonged immobility in the toilet" or "suspected fainting risk." If, after generating vital sign data, the risk level does not reach the critical threshold for life safety but deviates from the normal health fluctuation threshold, the emergency risk level is determined to be a Level II abnormal warning. A tiered emergency dispatch strategy triggers the smart interactive terminal to play a status inquiry voice command and activate the voice acquisition channel to await feedback. When the elderly person is able to respond, the voice acquisition channel detects a valid feedback signal, subsequent notification actions are not triggered, and the home care record retains an entry for "nighttime immobility but able to respond," facilitating health management personnel to track chronic disease-related risks such as difficulty defecating and frequent nighttime urination. If no valid feedback signal is detected within the waiting time threshold, an abnormal alarm notification is sent to the guardian's mobile terminal. Family members can then confirm by phone or visit the guardian by following the care script for "unresponsive overnight bathroom visit". The health manager can then arrange follow-up visits and care suggestions to reduce frequent false alarms caused by relying solely on movement alarms.

[0098] For low-risk conditions, home health management emphasizes continuous observation and recording of behavioral trends. For example, if an elderly person changes from standing to sitting on a sofa, takes a short break to lie on their side, has stable body temperature indicators, and their respiratory and heart rate values ​​are within the normal healthy fluctuation threshold range, the risk emergency level is determined to be Level 3 daily monitoring. The tiered emergency dispatch strategy writes the timestamps and types of abnormal behavioral events into the system operation log and maintains continuous collection of non-contact motion sensing data and thermal distribution imaging data. In home health follow-ups, the log can be interpreted as care signals such as "decreased activity level," "increased number of times getting out of bed at night," and "longer time spent in the bathroom," providing a basis for subsequent home rehabilitation training plans, chronic disease management reminders, and environmental modification suggestions for high-risk fall groups. At the same time, it avoids directly escalating minor postural changes into external notifications, reducing the ineffective use of care resources.

[0099] This embodiment maps risk emergency response levels to tiered emergency dispatch strategies. It solidifies the remote emergency dispatch center communication connection request and the push of identity file information and vital sign data corresponding to the Level 1 critical emergency response level into an executable external linkage process. It also solidifies the intelligent interactive terminal status inquiry voice command and voice acquisition channel activation action corresponding to the Level 2 abnormal warning level into an executable human-machine confirmation process. When no valid user feedback signal is detected within the waiting time threshold, an abnormal alarm notification is generated and sent to the guardian's mobile terminal. Furthermore, it writes the timestamps and types of abnormal behavior events corresponding to the Level 3 daily monitoring level into the system operation log and maintains continuous collection of non-contact motion sensing data and thermal distribution imaging data. This allows different risk emergency response levels to have differentiated action sets and clear trigger boundaries, reducing external linkage delays in high-risk scenarios, improving the confirmation loop in medium-risk scenarios, and enhancing the traceability and continuous monitoring continuity in low-risk scenarios.

[0100] In one embodiment, a non-contact sensing-based early warning device is provided, which corresponds one-to-one with the non-contact sensing-based early warning method described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the non-contact sensing-based early warning device of the present invention. The modules include a multi-source sensing and acquisition module 10, a cross-modal fusion modeling module 20, a behavioral context reasoning module 30, a vital sign analysis module 40, a risk level determination module 50, and an emergency dispatch execution module 60. Detailed descriptions of each functional module are as follows: The multi-source sensing and acquisition module 10 is used to acquire non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The cross-modal fusion modeling module 20 is used to fuse the non-contact motion sensing data and the thermal distribution imaging data to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. The behavioral context reasoning module 30 is used to obtain the motion trajectory features and posture features of the target object based on the real-time human digital model, and to analyze the motion trajectory features and posture features using the behavioral context analysis model to identify abnormal behavioral events. The vital signs analysis module 40 is used to lock the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model when the abnormal behavior event is detected, and generate vital signs data. The risk level determination module 50 is used to determine the risk emergency response level of the target object based on the type of the abnormal behavior event and the vital sign data. The emergency dispatch execution module 60 is used to execute the corresponding graded emergency dispatch strategy according to the risk and emergency response level.

[0101] For specific limitations regarding the early warning device based on non-contact sensing, please refer to the aforementioned limitations on early warning methods based on non-contact sensing, which will not be repeated here. Each module in the aforementioned early warning device based on non-contact sensing can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0102] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a non-contact sensing-based early warning method on the server side.

[0103] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a non-contact sensing-based early warning method.

[0104] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The non-contact motion sensing data and the thermal distribution imaging data are fused to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. Based on the real-time human digital model, the motion trajectory features and posture features of the target object are obtained, and the behavioral context analysis model is applied to analyze the motion trajectory features and posture features to identify abnormal behavioral events. When the abnormal behavior event is detected, the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model are locked to generate vital sign data. Based on the type of the abnormal behavioral event and the vital signs data, determine the risk level of the target individual for emergency medical care; The corresponding tiered emergency dispatch strategy shall be executed according to the aforementioned risk and emergency response level.

[0105] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The non-contact motion sensing data and the thermal distribution imaging data are fused to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. Based on the real-time human digital model, the motion trajectory features and posture features of the target object are obtained, and the behavioral context analysis model is applied to analyze the motion trajectory features and posture features to identify abnormal behavioral events. When the abnormal behavior event is detected, the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model are locked to generate vital sign data. Based on the type of the abnormal behavioral event and the vital signs data, determine the risk level of the target individual for emergency medical care; The corresponding tiered emergency dispatch strategy shall be executed according to the aforementioned risk and emergency response level.

[0106] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0107] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0108] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0109] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0110] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A pre-warning method based on non-contact sensing, characterized in that, Includes the following steps: Collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The non-contact motion sensing data and the thermal distribution imaging data are fused to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. Based on the real-time human digital model, the motion trajectory features and posture features of the target object are obtained, and the motion trajectory features and posture features are analyzed by the behavioral context analysis model to identify abnormal behavioral events. When the abnormal behavior event is detected, the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model are locked to generate vital sign data. Based on the type of the abnormal behavioral event and the vital signs data, determine the risk level of the target individual for emergency medical care; The corresponding tiered emergency dispatch strategy shall be executed according to the aforementioned risk and emergency response level.

2. The early warning method based on non-contact sensing as described in claim 1, characterized in that, Collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area, including: The millimeter-wave radar set up in the monitoring area is controlled to transmit frequency-modulated continuous wave signals and receive echo signals reflected by the target object. The echo signal is processed by fast Fourier transform in the distance and velocity dimensions and clutter suppression processing to extract point cloud sequences containing three-dimensional spatial coordinates and micro-Doppler spectra as non-contact motion sensing data. Control the infrared thermal imaging sensor to collect information on the distribution of infrared radiation flux within the monitoring area; Temperature quantization mapping is performed based on the infrared radiation flux distribution information to generate a two-dimensional temperature numerical matrix as thermal distribution imaging data. The non-contact motion sensing data and the thermal distribution imaging data are marked with a unified timestamp to achieve time synchronization.

3. The early warning method based on non-contact sensing as described in claim 1, characterized in that, The non-contact motion sensing data and the thermal distribution imaging data are fused to construct a real-time human digital model reflecting the target object's posture changes and vital signs, including: A joint calibration model comprising a millimeter-wave radar coordinate system and a thermal imaging pixel coordinate system is established, and the coordinate transformation matrix from the millimeter-wave radar coordinate system to the thermal imaging pixel coordinate system is determined. Analyze the three-dimensional point cloud coordinates and micro-Doppler spectrum information in the non-contact motion sensing data; The coordinate transformation matrix is ​​used to determine the projected pixel coordinates of each 3D point cloud coordinate on the thermal distribution imaging data plane; The local temperature value in the thermal distribution imaging data is indexed according to the projected pixel coordinates. The local temperature value and the micro-Doppler spectrum information spatiotemporally associated with the three-dimensional point cloud coordinates are bound as attribute features to the corresponding three-dimensional point cloud coordinates to generate an attribute fusion point cloud set. Spatial density clustering analysis is performed on the attribute fusion point cloud set to generate a real-time human digital model that includes the geometric shape, temperature distribution, and micro-motion spectrum features of the target object.

4. The early warning method based on non-contact sensing as described in claim 1, characterized in that, Based on the real-time human digital model, the motion trajectory features and posture features of the target object are obtained. The behavioral context analysis model is then applied to analyze the characteristics of the movement trajectory and posture to identify abnormal behavioral events, including: The spatial centroid coordinates of the real-time human digital model are continuously analyzed to determine the instantaneous descent velocity and ground-touching acceleration of the spatial centroid in the vertical direction as motion trajectory features. The size of the three-dimensional bounding box of the real-time human digital model is determined, and the orientation of the subject is determined as a posture feature based on the ratio of the long side to the short side of the three-dimensional bounding box. Acquire historical behavioral habit data of the target object, and determine the matching probability between the current time period and current spatial location and the historical behavioral habit data as the basis for the judgment of the behavioral context analysis model; An abnormal behavior event is confirmed when the instantaneous descent speed exceeds a preset fall threshold and the subject is facing a horizontal direction, or when the target object is determined to be stationary based on the posture characteristics and the matching probability is continuously lower than a preset normal behavior confidence level for more than a preset duration threshold.

5. The early warning method based on non-contact sensing as described in claim 1, characterized in that, Upon detecting the abnormal behavioral event, the system locks the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model, and generates vital sign data, including: In response to the detection of the abnormal behavior event, the chest and abdomen region with the highest thermal radiation intensity is located in the real-time human digital model, and the spatial coordinates and average temperature value of the chest and abdomen region are extracted as body temperature distribution features. The detection beam of the millimeter-wave radar is controlled to focus on the chest and abdomen region according to the spatial coordinates, and the micro-Doppler phase modulation signal of the chest and abdomen region is continuously collected as micro-motion characteristics. Adaptive mode decomposition and spectral analysis are performed on the micro-motion characteristics to separate the first signal component corresponding to the respiratory frequency and the second signal component corresponding to the heart rate. Based on the first signal component, the second signal component, and the average temperature value, vital sign data including respiratory rate, heart rate, and body temperature status indicators are generated.

6. The early warning method based on non-contact sensing as described in claim 1, characterized in that, Based on the type of the abnormal behavioral event and the vital sign data, determine the risk level of the target individual for emergency medical care, including: Determine whether the abnormal behavior event is a fall event characterized by rapid movement or a lingering event characterized by static persistence. The respiratory rate, heart rate, and body temperature status are obtained from the vital signs data. The respiratory rate, heart rate, and body temperature indicators are compared with their corresponding life safety limit thresholds and normal health fluctuation thresholds, respectively. When the abnormal behavior event is determined to be a fall-related event, and at least one of the respiratory rate, heart rate, or body temperature indicators is lower than or exceeds the corresponding life safety limit threshold, the risk emergency treatment level is determined to be Level 1 critical emergency treatment level. When the abnormal behavior event is determined to be a fall-related event, and the respiratory rate, heart rate and body temperature indicators have not reached the corresponding life safety limit threshold, but at least one of them deviates from the corresponding normal health fluctuation threshold, the risk emergency level is determined to be a Level II abnormal warning level. When the abnormal behavior event is determined to be a fall-related event, and the respiratory rate, heart rate and body temperature are all within the corresponding normal health fluctuation threshold range, the risk emergency level is determined to be Level 3 daily monitoring level. When the abnormal behavior event is determined to be a lingering event, and at least one of the respiratory rate, heart rate, or body temperature indicators is lower than or exceeds the corresponding life safety limit threshold, the risk emergency level is determined to be Level 1 critical emergency level. When the abnormal behavior event is determined to be a lingering event, and the respiratory rate, heart rate and body temperature indicators have not reached the corresponding life safety limit threshold, but at least one of them deviates from the corresponding normal health fluctuation threshold, the risk emergency level is determined to be a level two abnormal warning level. When the abnormal behavior event is determined to be a lingering event, and the respiratory rate, heart rate, and body temperature are all within the corresponding normal health fluctuation threshold range, the risk emergency response level is determined to be Level 3 daily monitoring level.

7. The early warning method based on non-contact sensing as described in claim 1, characterized in that, According to the aforementioned risk and emergency response level, the corresponding tiered emergency dispatch strategy is executed, including: When the risk emergency response level is determined to be Level 1 critical emergency response level, a communication connection is established with the remote emergency dispatch center, and the target's identity file information and vital sign data are pushed through the communication connection. When the risk emergency response level is determined to be a Level II abnormal warning level, the intelligent interactive terminal located in the monitoring area is driven to play a status inquiry voice command and open the voice acquisition channel to wait for user feedback. If no valid user feedback signal is detected through the voice acquisition channel within the preset waiting time threshold, an abnormal alarm notification is generated and sent to the preset guardian's mobile terminal. When the risk emergency response level is determined to be Level 3 daily monitoring level, the timestamp and type of the currently identified abnormal behavior events are written into the system operation log, and the continuous collection of non-contact motion sensing data and thermal distribution imaging data is maintained.

8. A warning device based on non-contact sensing, characterized in that, The early warning device based on non-contact sensing includes: The multi-source sensing and acquisition module is used to collect non-contact motion sensing data and thermal distribution imaging data within the monitoring area; The cross-modal fusion modeling module is used to fuse the non-contact motion sensing data and the thermal distribution imaging data to construct a real-time human digital model that reflects the posture changes and vital signs of the target object. The behavioral context reasoning module is used to obtain the motion trajectory features and posture features of the target object based on the real-time human digital model, and to analyze the motion trajectory features and posture features using the behavioral context analysis model to identify abnormal behavioral events. The vital signs analysis module is used to identify the abnormal behavioral event, lock the micro-motion features and body temperature distribution features corresponding to cardiopulmonary activity in the real-time human digital model, and generate vital signs data. The risk level determination module is used to determine the risk level of the target object based on the type of the abnormal behavior event and the vital signs data. The emergency dispatch execution module is used to execute the corresponding graded emergency dispatch strategy according to the risk and emergency response level.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a non-contact sensing-based early warning program stored in the memory and executable on the processor, wherein the non-contact sensing-based early warning program, when executed by the processor, implements the steps of the non-contact sensing-based early warning method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a warning program based on non-contact sensing, which, when executed by a processor, implements the steps of the warning method based on non-contact sensing as described in any one of claims 1-7.

Images