Fall detection method, device and computer readable storage medium

Abstract

This invention discloses a fall detection method, device, and computer-readable storage medium. The method includes: acquiring echo signals collected by millimeter-wave radar from a monitored object, and determining whether a suspected fall event has occurred based on the echo signals; in response to determining that a suspected fall event has occurred, extracting a target audio data stream from a continuously cached audio data stream based on the time of occurrence of the suspected fall event; encapsulating the correlation feature data between the target audio data stream and the suspected fall event to obtain a behavior verification data packet; in response to a verification request from a verification terminal, outputting the behavior verification data packet, and receiving the event verification result of the behavior verification data packet from the verification terminal; and performing corresponding control operations based on the event verification result, adaptively updating the judgment conditions when a false alarm is detected. This invention balances high-dimensional multimodal verification accuracy with absolute indoor privacy isolation and achieves adaptive iterative upgrading of the judgment conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing data processing technology, and in particular to a fall detection method, device, and computer-readable storage medium. Background Technology

[0002] With the increasing aging of the population, technologies for 24 / 7, non-contact health monitoring of monitored individuals have been widely adopted. Among existing non-contact monitoring solutions, radar sensors are used to collect target echo signals and identify behavioral states. Due to their proactive sensing capabilities and strong environmental adaptability, this has become the mainstream development direction.

[0003] However, traditional non-contact status monitoring technologies face a deep-seated technical contradiction between behavioral recognition accuracy and user privacy protection in practical deployments. On the one hand, pure radar sensing solutions are prone to high false alarm and false alarm rates when independently performing status determination due to underlying physical factors such as indoor environmental clutter interference, multi-target spatial aliasing, and ambiguity of static and dynamic human posture features, resulting in unnecessary waste of monitoring resources. On the other hand, to suppress such high-frequency false alarms, traditional technical solutions often choose to introduce high-density, 24 / 7 video surveillance or continuous environmental sound recording as a multimodal auxiliary verification method. While this auxiliary method improves the judgment accuracy in specific scenarios to some extent, it inevitably brings massive network transmission of continuous data streams and background storage load, and fundamentally breaks the privacy and security boundaries of high-density private scenarios such as bedrooms and bathrooms. Because existing monitoring solutions have not performed targeted topology optimization for transient multimodal verification scenarios triggered by abnormal events, traditional solutions struggle to achieve high-precision state variation verification and false alarm convergence while strictly meeting the constraints of privacy security and low-power data transmission. Summary of the Invention

[0004] This application provides a fall detection method that aims to address the problem that traditional non-contact monitoring technologies struggle to achieve high-precision fall detection and false alarm suppression under strict privacy protection regulations.

[0005] To achieve the above objectives, embodiments of this application provide a fall detection method, including:

[0006] Acquire the echo signal collected by the millimeter-wave radar from the monitored object, and determine whether a suspected fall event has occurred based on the echo signal;

[0007] In response to the determination of a suspected fall event, the target audio data stream is extracted from the continuously cached audio data stream based on the time of occurrence of the suspected fall event;

[0008] The target audio data stream and the associated feature data of the suspected fall event are encapsulated to obtain a behavior verification data packet;

[0009] In response to the review request from the review terminal, the behavior review data packet is output, and the event review result for the behavior review data packet is received from the review terminal.

[0010] Based on the event review results, the corresponding control operations are executed, wherein...

[0011] When the event review result indicates that the event is a real event, an alarm command is triggered;

[0012] When the event review result indicates a false alarm, the criteria used to determine the suspected fall event are updated.

[0013] In one embodiment, determining whether a suspected fall event has occurred based on the echo signal includes:

[0014] Perform analytical transformation on the echo signal to extract the spatial point cloud distribution parameters and kinematic evolution parameters of the monitored object;

[0015] A fusion assessment is performed based on the spatial point cloud distribution parameters and the kinematic evolution parameters to obtain the behavioral state indicators of the monitored object.

[0016] The behavioral status indicators are matched and compared with preset behavioral boundary judgment conditions;

[0017] When the behavioral state indicator meets the preset behavioral limit judgment condition, the suspected fall event is determined to have occurred.

[0018] In one embodiment, the spatial point cloud distribution parameters include an absolute height feature index and a spatial three-dimensional divergence feature index, the kinematic evolution parameters include a vertical axial velocity change rate feature index, and the preset behavioral limit judgment condition is a preset confidence judgment threshold.

[0019] A fusion assessment is performed based on the spatial point cloud distribution parameters and the kinematic evolution parameters to obtain behavioral state indicators of the monitored object, including:

[0020] The absolute height feature index, the spatial three-dimensional divergence feature index, and the vertical axis velocity change rate feature index are input into a preset machine learning mapping model to perform feature aggregation operations.

[0021] Based on the output dimension of the machine learning mapping model, the corresponding classification mapping metric is extracted, and the classification mapping metric is used as the behavior state indicator.

[0022] In one embodiment, in response to determining that a suspected fall event has occurred, based on the time of occurrence of the suspected fall event, a target audio data stream is extracted from a continuously cached audio data stream, including:

[0023] In response to the determination of a suspected fall event, the audio acquisition module is switched to the event response state, and the sequence timestamp corresponding to the time of occurrence is established as the reference anchor point from the audio data stream continuously cached by the audio acquisition module.

[0024] Backtracking to the historical time dimension, extract the preceding audio data segment of the first preset span, and forward to the future time dimension, extract the subsequent audio data segment of the second preset span;

[0025] The preceding audio data segment and the subsequent audio data segment are concatenated in a time domain to generate the target audio data stream.

[0026] In one embodiment, the fall detection method further includes:

[0027] During the normal monitoring period when no suspected fall event is detected, the audio acquisition module is controlled to enter a low-power silent control mode, and the time-domain cyclic overwrite buffer operation of the audio data stream is maintained.

[0028] In one embodiment, the target audio data stream and the associated feature data of the suspected fall event are encapsulated to obtain a behavior verification data packet, including:

[0029] The target audio data stream is subjected to noise reduction and data compression processing to obtain a normalized target audio data stream;

[0030] Obtain the associated feature data of the suspected fall event, wherein the associated feature data includes the event timestamp, the behavioral data collected by radar, the coordinates of the event location, and the confidence index of the suspected event;

[0031] The normalized target audio data stream and the associated feature data are subjected to multidimensional field alignment and concatenation encapsulation, and then encrypted using a preset encryption mapping sequence to obtain the encrypted behavior verification data packet.

[0032] In one embodiment, when the event review result indicates a false alarm, the criteria for determining the suspected fall event are updated, including:

[0033] Extract the false trigger feedback identifier carried in the event review result;

[0034] In response to the false trigger feedback flag, an initial judgment threshold corresponding to the vertical axial velocity change rate characteristic index is extracted;

[0035] Perform incremental compensation operation on the initial judgment threshold to obtain the updated judgment threshold;

[0036] The updated judgment threshold is used to replace the initial judgment threshold and serves as the judgment benchmark for the next monitoring cycle.

[0037] In one embodiment, after executing the corresponding control operation based on the event review result, the fall detection method further includes:

[0038] The event review results are tagged and bound together with the corresponding behavior review data packets to generate sample review record data.

[0039] The sample verification record data is transmitted to the cloud storage array to perform historical sample accumulation operation;

[0040] When the number of sample verification records in the cloud storage array reaches the preset iteration activation scale limit, the preset model retraining program is triggered, and the updated judgment conditions are output.

[0041] To achieve the above objectives, this application also proposes a fall detection device, including a memory, a processor, and a fall detection program stored in the memory and executable on the processor. When the processor executes the fall detection program, it implements the fall detection method as described in any of the above claims.

[0042] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a fall detection program, which, when executed by a processor, implements the fall detection method as described in any of the preceding claims.

[0043] The fall detection method of this application has the following beneficial effects:

[0044] First, it balances the "accuracy of anomaly verification" with "absolute indoor privacy";

[0045] Second, it has overcome the bottleneck of "false alarms in the environment" in traditional non-contact sensing algorithms;

[0046] Third, a high-efficiency heterogeneous data fusion and transmission paradigm with low communication load and strong anti-encryption was constructed;

[0047] Fourth, drive "adaptive iterative upgrade of judgment conditions" to achieve long-term self-evolution of equipment. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0049] Figure 1 This is a modular structure diagram of an embodiment of the fall detection device of the present invention;

[0050] Figure 2 This is a flowchart illustrating an embodiment of the fall detection method of the present invention.

[0051] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0053] It should be noted that when ordinal numbers such as "first" and "second" are mentioned in the embodiments of this application, they are only used to distinguish different objects and do not indicate a specific order or degree of importance, unless the context clearly specifies otherwise. Furthermore, the "connection" or "coupling" described in the embodiments of this application includes not only direct physical connections but also indirect connections or electrical / communication connections via an intermediate medium.

[0054] like Figure 1 As shown, Figure 1 This is a schematic diagram of the structure of the fall detection device 1 in the hardware operating environment involved in the embodiment of the present invention.

[0055] The fall detection device 1 (hereinafter referred to as "the device") in this application embodiment can be physically manifested as, but is not limited to, a server (including cloud server, server cluster, edge computing node), high-performance workstation, personal computer (PC), mobile terminal, IoT gateway, or dedicated embedded processing device. The device is configured to execute the fall detection method provided in this application embodiment.

[0056] like Figure 1 As shown, the device may include a memory 11, a processor 12, a communication interface 13, and a system bus 14.

[0057] The memory 11 is used to store computer programs (or instructions) and data required for the operation of the device.

[0058] The memory 11 includes at least one type of readable storage medium. The readable storage medium includes non-volatile memory (NVM), such as solid-state drive (SSD), hard disk drive (HDD), flash memory, optical disk, or other magnetic / optical storage media; the readable storage medium may also include volatile memory, such as random access memory (RAM) or cache.

[0059] More importantly, the memory 11 stores the operating system, the database, and the fall detection program 10 involved in this application.

[0060] Processor 12 is the core of the device's operation and control center.

[0061] Specifically, processor 12 may be one or more central processing units (CPUs), microprocessors (MCUs), digital signal processors (DSPs), or field-programmable gate arrays (FPGAs). In embodiments involving artificial intelligence, big data processing, or image rendering, processor 12 may also include an artificial intelligence acceleration chip (such as an NPU, TPU) or a graphics processing unit (GPU) for performing parallel vector or tensor operations.

[0062] The processor 12 uses the system bus 14 to read the fall detection program 10 in the memory 11, and implements the various steps of the fall detection method provided in this application embodiment by parsing and executing the program instructions.

[0063] Communication interface 13 (or network interface) is used to enable communication and interaction between the device and other electronic devices (such as clients, third-party servers, and sensor nodes).

[0064] Specifically, the communication interface 13 may optionally include a wired interface (such as an Ethernet interface, fiber optic interface, or USB interface) or a wireless interface (such as a Wi-Fi module, cellular mobile communication module, Bluetooth module, or NFC module). This interface supports various standard communication protocols, including but not limited to TCP / IP, HTTP / HTTPS, UDP, MQTT, and RPC.

[0065] System bus 14 can be a Peripheral Component Interconnect Standard (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus is used to transfer instruction and data streams between processor 12, memory 11 and communication interface 13.

[0066] Optionally, the device 1 may also include a user interface (not shown) for human-computer interaction. The user interface may include a display unit (such as an LCD screen, OLED screen, or touch screen) and an input unit (such as a keyboard, mouse, or microphone).

[0067] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a physical limitation on the fall detection device 1. Depending on the specific application scenario, the device may include fewer or more components than shown, or combine certain components, or use different component arrangements.

[0068] exist Figure 1 In the operating environment shown, processor 12 calls the fall detection program 10 stored in memory 11 and is configured to perform the following operations:

[0069] Acquire the echo signal collected by the millimeter-wave radar from the monitored object, and determine whether a suspected fall event has occurred based on the echo signal;

[0070] In response to the determination of a suspected fall event, the target audio data stream is extracted from the continuously cached audio data stream based on the time of occurrence of the suspected fall event;

[0071] The target audio data stream and the associated feature data of the suspected fall event are encapsulated to obtain a behavior verification data packet;

[0072] In response to the review request from the review terminal, the behavior review data packet is output, and the event review result for the behavior review data packet is received from the review terminal.

[0073] Based on the event review results, the corresponding control operations are executed, wherein...

[0074] When the event review result indicates that the event is a real event, an alarm command is triggered;

[0075] When the event review result indicates a false alarm, the criteria used to determine the suspected fall event are updated.

[0076] Furthermore, the processor 12 may also be configured to perform refined steps of the fall detection method in any of the following embodiments.

[0077] Based on the hardware architecture of the aforementioned fall detection device, an embodiment of the fall detection method of the present invention is proposed. The fall detection method of the present invention aims to solve the problem that traditional non-contact monitoring technologies, under strict privacy protection regulations, struggle to achieve high-precision fall detection and false alarm suppression.

[0078] Reference Figure 2 , Figure 2In one embodiment of the fall detection method of the present invention, the fall detection includes the following steps:

[0079] S10. Acquire the echo signal collected by the millimeter-wave radar from the monitored object, and determine whether a suspected fall event has occurred based on the echo signal.

[0080] Specifically, this step aims to establish an all-weather, non-contact proactive sensing barrier for health monitoring. Low-power, continuously operating millimeter-wave radar collects multi-dimensional reflected echoes within the monitored space, extracting high-precision combinations of dynamic and static spatial behavioral characteristic parameters. The local processor then performs a preliminary judgment of the behavioral state while ensuring user privacy, efficiently filtering out suspected time points with potential fall tendencies. This provides a precise data foundation and trigger signals for subsequent transient sound correlation verification.

[0081] Specifically, the radar sensor module continuously transmits frequency-modulated continuous wave signals into the target monitoring space. These signals propagate across the target space and are reflected upon encountering the surface of the monitored object. The processor acquires the reflected low-delay analog electrical signals via the echo receiving unit and converts them into a digital echo sequence via an analog-to-digital converter. Next, the data processing chip reads this digital echo sequence and calls a preset core algorithm to perform multi-dimensional feature analysis, extracting geometric topological and dynamic evolutionary physical parameters that quantitatively characterize the monitored object's body posture. By inputting these physical parameters into a classification decision unit for boundary matching, the system can adaptively identify whether the monitored object experiences a sudden drop in height and abnormal speed due to instability, thereby outputting a judgment result indicating a suspected fall.

[0082] In some embodiments, determining whether a suspected fall event has occurred based on the echo signal includes the following steps S11-S14.

[0083] S11. Perform analytical conversion on the echo signal to extract the spatial point cloud distribution parameters and kinematic evolution parameters of the monitored object.

[0084] Specifically, the processor performs a two-dimensional Fast Fourier Transform (FFT) on the received digitized echo signal, converting the distance dimension and Doppler dimension, and projects the time-domain sampling sequence into a range-Doppler resolution map. Subsequently, the system applies a constant false alarm rate (CFAR) operator to perform matrix noise stripping on the map, accurately filtering out static environmental clutter and multipath interference, and identifying the dynamic Doppler cell group representing the physical boundary of the monitored object. The processor further performs phase difference analysis on this cell group using an azimuth-elevation joint angle measurement matrix to extract the current three-dimensional spatial coordinates of the monitored object in a three-dimensional Cartesian coordinate system. By continuously aggregating these spatial coordinates within a sliding time window, the system can generate high-fidelity, measurable spatial point cloud distribution parameters and kinematic evolution parameters that quantify its motion inertial characteristics in real time.

[0085] S12. Perform a fusion evaluation based on the spatial point cloud distribution parameters and the kinematic evolution parameters to obtain the behavioral state indicators of the monitored object.

[0086] Specifically, this step leverages the strong physical coupling between geometric deformation and spatiotemporal dynamic energy mutation during a fall. The processor, acting as the execution entity, invokes the internally stored evaluation and control module, cross-fusing spatial point cloud distribution parameters representing static posture with kinematic evolution parameters representing dynamic temporal changes as multivariate inputs. These spatial point cloud distribution parameters include absolute height characteristic indicators and spatial three-dimensional divergence characteristic indicators, while the kinematic evolution parameters include vertical axial velocity change rate characteristic indicators. The processor performs nonlinear feature mapping on these heterogeneous parameters, projecting feature data from different physical dimensions into a unified high-dimensional space for behavioral risk assessment, and calculating behavioral state indicators to comprehensively quantify the fall probability of the monitored individual.

[0087] In some embodiments, a fusion evaluation is performed based on the spatial point cloud distribution parameters and the kinematic evolution parameters to obtain behavioral state indicators of the monitored object, including the following steps S121-S122:

[0088] S121. Input the absolute height feature index, the spatial three-dimensional divergence feature index, and the vertical axial velocity change rate feature index into a preset machine learning mapping model to perform feature aggregation operation.

[0089] Specifically, the processor reads the absolute height feature index, spatial three-dimensional divergence feature index, and vertical axial velocity change rate feature index obtained from the current cycle calculation, and encapsulates them into a normalized behavioral feature vector. The processor uses this behavioral feature vector as input to a preset machine learning mapping model. In some optional embodiments, this machine learning mapping model can be a lightweight convolutional neural network model or a deep fully connected neural network. Through multi-layer forward propagation and cascaded transformations of nonlinear activation functions within the model's hidden layers, high-order feature association and adaptive aggregation are performed on the aforementioned physical quantities representing the spatial divergence pattern, the proportion of absolute low-level nodes, and the gravitational axial velocity acceleration. This mitigates the negative impact of feature fragmentation caused by occlusion from indoor furniture obstacles or multi-target aliasing interference, outputting a high-order feature aggregation matrix.

[0090] S122. Based on the output dimension of the machine learning mapping model, extract the corresponding classification mapping metric and use the classification mapping metric as the behavior state indicator.

[0091] Specifically, the output layer of the machine learning mapping model is configured with a confidence regression operator, and an exemplary calculation formula is as follows:

[0092]

[0093] in, The preset model weight matrix, The higher-order feature aggregation matrix calculated in the above steps, The output result is calculated using a preset bias scalar. This is the classification mapping metric for the output layer dimension of the machine learning mapping model. The processor extracts this classification mapping metric and uses it as a behavioral state indicator representing the probability of falling in the current cycle.

[0094] S13. Match and compare the behavior status indicators with the preset behavior boundary judgment conditions.

[0095] Specifically, the preset behavioral boundary judgment condition is a preset confidence judgment threshold. In some specific embodiments, this confidence judgment threshold is exemplarily set to 0.75. The processor calls the system's built-in logical comparison operator to compare the online calculated behavioral state index with the confidence judgment threshold by the magnitude of the calculated value. This process establishes a dynamic boundary between normal daily human activities and abnormal extreme situations within the system, aiming to effectively filter out slight characteristic fluctuations caused by routine behaviors of the monitored subject, such as quickly sitting down, lying down, or bending over to pick up items, at the physical level, ensuring high robustness of the comparison process at the logical level.

[0096] S14. When the behavior status indicator meets the preset behavior limit judgment condition, it is determined that the suspected fall event has occurred.

[0097] Specifically, when the processor determines that the behavior status indicator is greater than or equal to the confidence threshold (for example, the current real-time behavior status indicator is calculated to be 0.83, exceeding the predetermined threshold of 0.75), the system determines that the current physical motion characteristics and spatial distribution pattern are highly likely to conform to the underlying physical evolution law of human instability and fall, thereby triggering the determination of the suspected fall event. At this time, the processor immediately locks the current system clock, generates a trigger control command carrying a unique event ID and timestamp, and immediately terminates the low-power silent state of the sound data stream acquisition module, starts the subsequent target audio segment interception, and completes the seamless handover of the control chain from radar all-weather main judgment to short-term sound verification.

[0098] Furthermore, if the behavioral state indicator is less than the confidence threshold, the system determines that the current situation is only a benign fluctuation of normal behavioral state and no suspected fall event has occurred, and the radar sensor continues the monitoring cycle with the original overhead.

[0099] It is understandable that by performing the cascaded refinement process from step S11 to step S14, this application thoroughly decomposes the behavior recognition task into a physical mapping of spatial geometric topology (absolute height and divergence) and dynamic temporal evolution (vertical axis velocity change rate), and integrates the high-order feature aggregation capability of machine learning models. This effectively overcomes the technical limitations of traditional non-contact monitoring schemes, which are susceptible to indoor background clutter, multi-target aliasing, and fuzzy special attitude features. Without the physical constraint of introducing high-overhead, privacy-sensitive all-weather video or continuous audio recording data, it significantly improves the confidence and sensitivity of the radar front-end's initial judgment, laying a highly deterministic feature boundary for the further convergence of the overall false alarm rate and missed alarm rate of the entire system.

[0100] S20. In response to determining that a suspected fall event has occurred, based on the time of occurrence of the suspected fall event, extract the target audio data stream from the continuously cached audio data stream.

[0101] Specifically, this step aims to precisely slice the temporal boundaries of the multimodal verification mechanism's data. When the millimeter-wave radar determines that the monitored object's behavior has undergone an abnormal transition, the system immediately cuts off the silent or only temporarily overwritten non-permanent storage state through this step, and selectively extracts a short, high-value acoustic segment spanning the period before and after the abnormal event from the continuously cached audio data stream. This processing ensures the complete retention of the acoustic characteristics of the moment of the fall and collision and its preceding warning signs, while minimizing the permanent storage overhead and processing bandwidth of useless non-event audio, thereby constructing a highly efficient and privacy-preserving event verification data source at the system architecture level.

[0102] In some embodiments, step S20 can be implemented by the following steps S21-S23.

[0103] S21. In response to the determination that a suspected fall event has occurred, the audio acquisition module is switched to the event response state, and the sequence timestamp corresponding to the time of occurrence is established as the reference anchor point from the audio data stream continuously cached by the audio acquisition module.

[0104] Specifically, when the processor performs a state comparison and determines that the output shows a suspected behavioral anomaly, the control core immediately sends a state transition interrupt instruction to the audio acquisition module. In response to this interrupt instruction, the module changes the control bits in its internal configuration register, switching its operating mode from normal to event response mode and activating the high sampling rate processing channel or peripheral communication bus. Simultaneously, the processor reads the time of the suspected fall event and performs a timestamp index search in the continuously buffered audio data stream, matching a specific audio data frame that is perfectly aligned with or closest to the event in the time domain. The system establishes the monotonically increasing sequence timestamp carried by this specific audio data frame as the reference anchor point for subsequent segmentation mechanisms, thereby locking the absolute center of the data slice on the timeline.

[0105] In some embodiments, during normal monitoring periods when no suspected fall event is determined to have occurred, the audio acquisition module is controlled to enter a low-power silent control mode, and the temporal cyclic overwrite buffer operation of the audio data stream is maintained.

[0106] Specifically, during the steady-state monitoring period when the radar front-end does not detect any abnormal behavior, the control core maintains a low-power silent control mode for the audio acquisition module. In this mode, the microphone acquisition peripheral disables the write path to the permanent non-volatile memory, and the related digital signal processing operators or cloud communication protocol stack remain dormant. However, the audio acquisition module maintains a time-domain cyclic overwrite buffer operation on the audio data stream, that is, it uses the direct memory access controller to continuously write the raw audio sampling data output from the underlying analog-to-digital converter into a first-in-first-out volatile circular queue with a preset space. The storage depth of this circular queue is topologically designed so that its maximum retention time is at least greater than a subsequently defined first preset span. When the buffered data reaches the queue boundary, the newly written audio sampling frame automatically overwrites the oldest historical sampling frame. This dynamic cyclic overwrite mechanism under normal conditions physically blocks the risks of data leakage and permanent retention, achieving absolute privacy isolation for the user's steady-state living environment.

[0107] S22. Backtrack to the historical time dimension to extract the preceding audio data segment of the first preset span, and forward to the future time dimension to extract the subsequent audio data segment of the second preset span.

[0108] Specifically, the processor uses the sequence timestamp corresponding to the reference anchor point as the origin of the time domain, and calculates the memory address offset by backtracking along the negative direction of the time axis, thereby extracting and locking a segment of preceding audio data up to the length of the first preset span from the circular queue. In some specific embodiments, the first preset span can be exemplarily set to 8 to 12 seconds, preferably 10 seconds, which corresponds to the complete physical and dynamic evolution cycle of the human body from loss of balance and tilting to final collision with the ground. Simultaneously, the processor performs a timed count forward along the positive direction of the time axis from the reference anchor point, continuing the normal recording behavior of the audio acquisition module after the event occurs until the boundary of the second preset span is reached. In some specific embodiments, the second preset span can be exemplarily set to 5 to 12 seconds, preferably 10 seconds, to fully capture the breathing, groans, cries for help, and the recovery of ambient background sound after the monitored object falls to the ground. The formula for the time domain boundary of the truncated interval is as follows:

[0109]

[0110] in, The sequence timestamp that represents this reference anchor point Corresponding to the first preset span, Corresponding to the second preset span, the calculated result is This defines the absolute time domain window of the target audio data stream.

[0111] S23. Perform a time-domain sequence concatenation operation on the preceding audio data segment and the subsequent audio data segment to generate the target audio data stream.

[0112] Specifically, the processor allocates an independent linear contiguous storage space in volatile memory and reads the end data frame pointer of the preceding audio data segment and the start data frame pointer of the subsequent audio data segment, respectively. The system uses a low-level high-speed memory copy operator to perform a time-domain sequence concatenation operation on these two independent data segments according to their timestamp order. To eliminate potential phase discontinuities or voltage abrupt noise at the splicing boundaries, the processor adaptively introduces a microsecond-level fade-in / fade-out window function at the concatenation node for smooth correction. This allows the processor to generate a single audio file with an exemplary total length of 20 seconds and high acoustic feature density without disrupting the continuity of the acoustic waveform. This audio file is then output as the target audio data stream to the next-level encapsulation module.

[0113] It is understandable that by executing steps S21 to S23 above, along with silent overlay caching control during normal monitoring cycles, this application resolves the sharp technical contradiction between "traditional technology-triggered recording leading to the loss of critical acoustic precursors to collisions" and "continuous recording and permanent storage infringing on user's steady-state privacy" in data processing. By constructing an architecture of "normal silent cyclic overlay + bidirectional temporal slicing triggered by anomalies," the system implements "erase and hide" volatile management of data when no event occurs. At the moment of an anomaly, it accurately extracts the local acoustic context containing the entire fall cycle through state switching and bidirectional temporal backtracking. This not only achieves ironclad privacy isolation protection under physical constraints but also greatly reduces the system's useless power consumption and network bandwidth loss, significantly improving the information certainty of subsequent multimodal verification.

[0114] S30. Encapsulate the target audio data stream and the associated feature data of the suspected fall event to obtain a behavior verification data packet.

[0115] Specifically, this step aims to achieve the fusion, packaging, and secure isolation of multimodal heterogeneous data. Since the acoustic data captured by the front end and the trajectory features output by radar belong to different dimensions of physical information, the system uses this step to perform frequency domain purification and volume trimming of the audio stream, and to perform time-series alignment and binary reassembly of its core digital parameters related to the event. Combined with the underlying encryption mapping rules, this operation not only maximizes the compression of the network transmission load of outgoing data at the physical level, but also builds a solid security barrier against eavesdropping and tampering in subsequent transmission links, thereby outputting a highly structured, high-security behavior verification carrier.

[0116] Specifically, after completing the preceding and subsequent time-domain interception and concatenation, the local processor extracts the physical pointer of the target audio data stream and synchronously reads multi-dimensional state data strongly correlated with the current suspected event from the radar data processing module. The control core, through its built-in protocol stack, maps and packages this continuous unstructured acoustic signal with discrete structured radar parameters in both time and field dimensions, aggregating the originally fragmented sensor information into a single composite data frame. This results in a behavior verification data packet that combines real-time sound data with physical behavior trajectories, facilitating a one-time push to the verification terminal.

[0117] In some embodiments, step S30 can be implemented by the following steps S31-S33.

[0118] S31. Perform noise reduction and data compression processing on the target audio data stream to obtain a normalized target audio data stream.

[0119] Specifically, for common steady-state background noise in indoor environments, such as air conditioning fan noise and appliance operation noise, the processor calls its built-in digital signal processing operators to perform a spectrum subtraction frequency domain filtering operation on the target audio data stream. An exemplary implementation is as follows: the processor projects the time-domain audio frame into the frequency domain through a short-time Fourier transform, estimates the noise power spectrum of the non-speech segment, and subtracts this noise power spectrum from the noisy speech power spectrum. The calculation logic is exemplarily represented as follows:

[0120]

[0121] in, The power spectrum of the noisy signal. For the estimated noise power spectrum, The preset over-subtraction factor, This is the power spectrum of the clean speech / collision sound after noise reduction. After recovering the time-domain waveform through inverse frequency domain transformation, the processor further calls the audio codec (such as Opus or AAC encoding standard) to perform scale compression on the noise-reduced data stream, eliminating redundant information below the human hearing masking threshold. Thus, while ensuring that the high-frequency transient features such as impact and distress calls are not lost, the audio data volume is compressed to an extremely low bit rate to generate the normalized target audio data stream.

[0122] S32. Obtain the associated feature data of the suspected fall event, wherein the associated feature data includes the event occurrence timestamp, radar-collected behavioral data, event location coordinates, and suspected event confidence index.

[0123] Specifically, the processor accesses the radar system status register via the internal data bus to capture the associated feature data frozen within the current trigger cycle. Among these features, the event timestamp (e.g., UTC timestamp) establishes an absolute time reference for the entire event; the behavioral data collected by the radar includes a sequence of three-dimensional point cloud movement trajectories of the monitored object before and after the fall, used to reconstruct the physical evolution of the action on the verification terminal; the event location coordinates characterize the specific three-dimensional topological orientation of the monitored object in the target space (e.g., bedroom, bathroom), used to guide possible subsequent precise rescue; and the suspected event confidence index quantifies the fall probability calculated by the front-end machine learning mapping model, used to provide priority scheduling reference for the cloud or terminal when multiple events occur concurrently.

[0124] S33. Perform multi-dimensional field alignment and concatenation encapsulation on the normalized target audio data stream and the associated feature data, and encrypt them using a preset encryption mapping sequence to obtain the encrypted behavior verification data packet.

[0125] Specifically, the processor constructs a standard binary data frame structure in memory. The system first uses the event timestamp as a synchronization reference to align the continuous, normalized target audio data stream with discrete radar behavior data, position coordinates, and confidence indices, performing multi-dimensional field alignment of the frame header, payload, and checksum. After alignment, the processor concatenates and encapsulates the heterogeneous data into a unified verification payload. Next, the system introduces a preset encryption mapping sequence (in some specific embodiments, this may exemplarily employ the AES-256 symmetric encryption algorithm or the RSA asymmetric encryption mechanism), performing obfuscation and diffusion operations on each bit of this encryption mapping sequence and the verification payload. Through this high-intensity mathematical mapping transformation, the system forcibly converts the plaintext verification payload into a scrambled ciphertext bitstream, ultimately generating an encrypted behavior verification data packet. This data packet appears as meaningless high-entropy noise at third-party nodes that do not possess the specific decryption key.

[0126] It is understandable that by performing the data purification, feature aggregation, and obfuscation encryption processes described in steps S31 to S33, this application overcomes the technical bottlenecks of "difficulty in synchronizing heterogeneous multimodal data" and "extended data being highly susceptible to man-in-the-middle eavesdropping." By introducing spectral subtraction and high compression ratio processing, the system forcibly removes invalid environmental noise that masks early signs of falls and significantly reduces network transmission latency. Simultaneously, through rigorous multi-dimensional field alignment and industrial-grade encryption mapping, this application binds multimodal physical features into a closed-loop evidence chain that cannot be reverse-analyzed or tampered with. This mechanism ensures both the efficiency and integrity of transmitted data and establishes an unbreakable data privacy barrier in the cloud-based verification link.

[0127] S40. In response to the review request from the review terminal, output the behavior review data packet and receive the event review result of the behavior review data packet from the review terminal.

[0128] Specifically, this step aims to establish a closed-loop interactive channel between the machine-based front-end judgment and the back-end human listening and verification. After the system completes the local aggregation and encryption of multimodal heterogeneous data, the highly compressed on-site situational data must be pushed to an authorized verification terminal through a standardized interface. Through point cloud trajectory rendering and audio decoding playback on the terminal interface, the system entrusts complex scenes that are difficult for radar to identify to human listening and verification for secondary authentication, thereby forming a highly deterministic event qualitative closed loop at the end of the physical link, providing a benchmark scalar for subsequent alarm scheduling and adaptive iteration.

[0129] Specifically, the backend server or the bound guardian's terminal device (such as a smartphone app or mini-program) sends a handshake and data retrieval request to the local radar main control board via a wireless communication network. In response to this request, the system establishes an encrypted transmission tunnel and pushes the encrypted behavior verification data packet residing in local memory to the verification terminal. Upon receiving the data, the verification terminal uses a pre-distributed asymmetric key to perform decryption and parsing, simultaneously rendering the 3D point cloud motion trajectory of the monitored object before and after the fall on the screen UI layer, and playing the decoded transient sounds from the scene (such as impact sounds, groans, or normal background walking sounds) through the speaker. Based on the bimodal presentation of visual trajectory and auditory acoustics, the verification personnel input confirmation or rejection commands through the human-computer interaction interface. The local processor receives the command via the downlink and parses it into a binary event verification result (i.e., a Boolean value of "true" or "false").

[0130] In some alternative embodiments, the verification terminal can be replaced by a cloud-based automated verification server, thereby providing a completely automated verification solution that requires no human intervention.

[0131] Specifically, the cloud-based collaborative verification server asynchronously sends a request to retrieve a behavior verification data packet to the local processor at the edge via a high-speed network. In response to this request, the local processor pushes the encrypted behavior verification data packet to the cloud-based collaborative verification server via a wireless network. Upon receiving the data packet, the cloud-based collaborative verification server invokes an internally deployed parallel decryption operator to restore it to a multi-dimensional aligned decrypted payload, and uses a field stripping procedure to thoroughly separate the normalized target audio data stream from radar feature data containing suspected event confidence indicators and behavioral data (such as point cloud trajectory sequences).

[0132] Next, the cloud-based collaborative verification server initiates a dual-path parallel machine pre-screening mechanism to perform unsupervised cross-validation: On the acoustic verification path, the server inputs the demodulated normalized target audio data stream into a preset acoustic event detection (AED) neural network model. By extracting the time-frequency features of the log-Mel spectrogram and performing time-series convolution operations, it automatically identifies whether there are specific acoustic fingerprints representing human impact, fall, violent breathing, groaning, or calls for help, and then calculates the first acoustic confidence component; On the radar verification path, the server calls a long short-term memory (LSTM) network or a three-dimensional temporal convolutional network to perform global temporal deformation fitting on the point cloud trajectory sequence, and calculates the second radar confidence component representing the topological evolution of the actual human fall action.

[0133] Subsequently, the cloud-based collaborative verification server performs nonlinear weighted aggregation on the suspected event confidence index, the first acoustic confidence component, and the second radar confidence component using a multimodal decision fusion operator. An evaluation formula for a fully automated review decision is as follows:

[0134]

[0135] in, This is the confidence index for the suspected front-end radar event carried in the data packet. This is the first acoustic confidence component, independently calculated in the cloud. This is the confidence component of the second radar. These are the preset decision weight coefficients for each feature path. The cloud-based collaborative verification server will calculate the final fusion confidence score. Compare the execution magnitude with the preset secondary verification security threshold: when When the value is greater than or equal to the secondary verification security threshold, the server automatically generates a judgment result representing the real event; when... If the value is below the secondary verification security threshold, the server automatically generates a judgment result representing a false alarm event. Finally, the cloud-based collaborative verification server uses this fully automated verification conclusion as the event's review result and asynchronously feeds it back to the local processor via the network to drive subsequent alarm triggering or local threshold compensation updates.

[0136] It is understandable that by introducing this cloud-based collaborative verification server to perform non-manual automated multimodal cross-verification, this application not only achieves zero human intervention and millisecond-level closed-loop response across the entire chain, greatly saving the scheduling resources and labor costs of back-end monitoring personnel, but also elevates the privacy isolation boundary at the data processing endpoint from the "human eye / ear barrier" to the level of a pure "machine mathematical closed loop," giving the fall detection method extremely high all-weather automated anti-interference defense capability.

[0137] S50. Execute corresponding control operations based on the event review result, wherein when the event review result indicates a real event, an alarm command is triggered; when the event review result indicates a false alarm event, the judgment conditions used to determine the suspected fall event are updated.

[0138] Specifically, the processor executes branch routing logic on the received event verification results. When the boolean value corresponding to the event verification result is true, indicating that a real fall event has occurred, the system immediately attaches a high-priority emergency response thread and generates an alarm command.

[0139] The alarm command includes: pushing an emergency alarm notification to the verification terminal of each contact in the preset emergency contact list, which includes the identity information of the monitored person, the time and location of the fall; sending a request for help to the dispatch system of the community service center or emergency center associated with the monitored person; if the system is equipped with a smart speaker or indoor intercom, it can also initiate a two-way voice call to the monitored person through the device, so that the guardian can remotely confirm the actual situation of the monitored person and provide comfort.

[0140] When the Boolean value corresponding to the event verification result is false, indicating that the front-end radar is misled by environmental clutter and generates a false alarm, the system abandons the alarm and switches to the local adaptive calibration thread. For the specific feature channel that caused this false alarm (such as the low vertical velocity judgment benchmark), the judgment conditions used to determine the suspected fall event are dynamically modified, so as to adaptively block similar interference sources in subsequent cycles.

[0141] In some embodiments, updating the criteria used to determine the suspected fall event can be achieved through the following steps S51-S54:

[0142] S51. Extract the false trigger feedback identifier carried in the event review result.

[0143] Specifically, after receiving the event verification results from the verification terminal or the fully automated verification server, the processor calls the underlying communication message parser to retrieve its payload field. The system reads the dedicated diagnostic control bits carried in the data packet, extracts and removes the false alarm feedback identifier specifically used to characterize the cause of the false alarm. This false alarm feedback identifier not only qualitatively declares that the current event belongs to a normal disturbance other than a fall, but also includes the specific waveform type that triggered the false alarm (e.g., the misjudged cause is sitting down quickly, hitting the mattress, or environmental wind noise), providing a highly confident classification guide label for subsequent directional parameter compensation.

[0144] S52. In response to the false trigger feedback flag, extract the initial judgment threshold corresponding to the vertical axial velocity change rate characteristic index.

[0145] Specifically, in response to the extracted false trigger feedback flag, the processor activates the parameter retrieval thread in the internal register. Since different types of false triggers correspond to different dimensions of physical feature overload, when the false trigger feedback flag indicates that the current misjudgment is caused by sudden velocity changes such as violent limb axial swaying or rapid lying down, the processor addresses the algorithm configuration area in non-volatile memory and reads the currently running initial judgment threshold corresponding to the vertical axis velocity change rate feature index. Through this refined cross-domain correlation, the system can accurately locate the underlying feature vulnerabilities that cause false alarm judgments and logical failures in the front-end radar, avoiding the secondary negative impact on the normal monitoring false alarm rate caused by blindly adjusting other irrelevant feature parameters (such as divergence or height thresholds).

[0146] S53. Perform incremental compensation operation on the initial judgment threshold to obtain the updated judgment threshold.

[0147] Specifically, the processor invokes the accumulation unit in the digital signal processing core, using the read initial judgment threshold as the first addend input. Simultaneously, based on a preset false alarm penalty mechanism, the system retrieves a single-step penalty compensation amount corresponding to the false trigger feedback flag. An exemplary incremental compensation calculation formula is as follows:

[0148]

[0149] in, The initial threshold corresponding to this vertical axial velocity change rate characteristic index, The preset single-step velocity change rate increment compensation constant (in some specific optional embodiments, this constant exemplarily corresponds to the upward adjustment of the vertical Z-axis velocity change rate) is used. ), This is the attenuation adjustment factor adaptively converged by the system based on the historical cumulative false alarm frequency of the monitored object over multiple cycles. The processor performs incremental compensation cascaded addition or nonlinear weighted mapping operations on these two factors to calculate an updated judgment threshold that can tolerate similar interference energy in the current scenario but with a higher magnitude.

[0150] S54. Replace the initial judgment threshold with the updated judgment threshold as the judgment benchmark for the next monitoring cycle.

[0151] Specifically, the processor acquires the binary data corresponding to the updated judgment threshold after calculation, applies a memory write operation operator to lock the static memory address where the original judgment threshold resides, and performs an overwrite operation. After completing the memory data refresh, the processor generates a parameter update synchronization signal, forcibly notifying the radar signal main processing thread to perform boundary reload. At this point, the updated judgment threshold officially takes effect, replacing the initial judgment threshold as the hard boundary benchmark for determining whether the monitored object has triggered a suspected fall event in the next sensing and monitoring cycle. In subsequent operations, the radar must generate kinematic characteristic changes exceeding this new threshold amplitude before it can release the interception command to the audio module again.

[0152] It is understandable that by performing the adaptive threshold incremental compensation processing from steps S51 to S54, this application overcomes the technical bottleneck of traditional non-contact sensor parameters being fixed, making it difficult to adapt to specific home interference scenarios at the underlying algorithm mechanism level. By establishing this reverse feedback loop between the radar's main judgment and the verification response, the system transforms each "false alarm" into a learning inspiration for edge-side self-calibration, enabling the system to perform adaptive topological deformation and attenuation of feature judgment boundaries at the perception level for specific monitoring space noise and interference. This design greatly eliminates the technical problems of excessive consumption of monitoring resources and loss of user trust caused by high-frequency false alarms in similar devices at the physical level.

[0153] After executing the corresponding control operation based on the event review result, the fall detection method of this application also includes a closed-loop iteration step, which aims to establish a global long-term self-evolution mechanism in which the system is driven from application layer feedback to deep algorithm model.

[0154] Local radar hardware, limited by computing power, storage, and power consumption, cannot independently perform gradient backpropagation training of large-scale deep neural networks at the edge. This step establishes topological coupling between the edge sensing end and the distributed cloud computing cluster, enabling the reverse flow and asynchronous aggregation of "high-confidence event labels" generated by manual listening or automatic cloud verification with their corresponding "multimodal original physical feature vectors." Through this, the system can utilize a continuous stream of real-world scene data to build a data-driven immune barrier, achieve iterative evolution of the global algorithm model through a large-sample retraining program in the cloud, and finally distribute the newly converged network weights to each terminal, fundamentally overcoming the technical bottleneck of traditional health monitoring equipment's fixed algorithms being unable to cope with complex and ever-changing home interference characteristics.

[0155] Specifically, the closed-loop iteration steps include the following steps S110-S130.

[0156] S110. The event review result is tagged and bound to the corresponding behavior review data packet to generate sample review record data.

[0157] Specifically, in response to the received event verification result, the local processor starts a sample labeling thread. The processor allocates a temporary structured data storage space in memory, reads the original behavior verification data packet that triggered this verification process (this data packet, as an unstructured feature vector, contains a normalized target audio data stream in log-Mel spectrum, and radar behavior data including vertical axis velocity change rate features and a 3D point cloud trajectory sequence), and simultaneously extracts the true / false status indication bit represented by the event verification result (e.g., mapping "real fall" to a positive sample label "1", and "environmental false alarm / false alarm" to a negative sample label "0"). Next, the processor calls the data field concatenation operator, using this true / false status indication bit as a supervision label for strongly supervised learning, and forcibly writes it to the end of the data frame of the behavior verification data packet or a specific metadata field. This achieves a tagged binding association between feature data and the verification qualitative result, combining them to generate sample verification record data that possesses both high-dimensional physical features and standard qualitative conclusions.

[0158] S120. The sample verification record data is transmitted to the cloud storage array to perform historical sample accumulation operation.

[0159] Specifically, to avoid consuming bandwidth resources during periods of high network load, the local processor mounts the generated sample verification record data to a low-priority background asynchronous transmission queue. When the processor detects that the current wireless communication link (such as Wi-Fi or 5G cellular network) is in an idle state or triggers a preset timed reporting cycle, it establishes a secure outgoing channel between the control peripheral communication interface and the remote cloud server. The sample verification record data is asynchronously uploaded to the cloud server via the network. After the cloud server's receiving gateway performs verification and comparison on the data, it routes it to a specific category in a distributed cloud storage array (such as a sample cluster library built on the Hadoop Distributed File System HDFS or a non-relational database) to perform historical sample accumulation operations. The cloud storage array adopts a multi-replica redundancy mechanism to perform full-time, long-term throughput accumulation of acoustic and point cloud composite samples tagged with real / false alarms from hundreds or thousands of edge radar nodes, continuously expanding the capacity of the full-scene fall dynamics database.

[0160] S130. When the number of sample verification record data in the cloud storage array reaches the preset iteration activation scale limit, the preset model retraining program is triggered, and the updated judgment condition is output.

[0161] Specifically, a large-scale monitoring daemon resides within the cloud server. This daemon continuously counts the total absolute number of newly accumulated sample review records in the cloud storage array through periodic polling or event triggering. When discovered When the preset iteration activation scale limit is exceeded (e.g., the cumulative number of new samples reaches 1000, or the tiered threshold of 500 samples in a specific false positive scenario), the daemon releases a retraining activation signal to the computing power cluster. In response to this signal, the cloud computing power cluster automatically calls upon GPU / TPU computing resources and mounts the preset model retraining program.

[0162] During retraining, the retraining program divides the large-scale historical sample dataset accumulated in the cloud storage array into training and validation sets. It uses the multimodal physical features of the samples as input vectors and the bound labels as the expected output, re-injecting them into the lightweight CNN machine learning model or high-dimensional spatiotemporal feature alignment model used in the front end to perform suspected fall detection. The retraining program employs backpropagation and stochastic gradient descent (SGD) or the Adam optimizer to calculate the gradient of the global loss function in a higher-dimensional sample space. This drives the bias and weight matrices of neurons in each layer of the network to perform adaptive updates, penalizing feature combination pathways that cause false alarms or missed alarms in the radar. When the network model's loss function converges to within a preset convergence tolerance range on the validation set, and its generalization precision and recall exceed those of the previous generation model, the model retraining program terminates. The cloud server exports and outputs the updated global network weight file, which is then used as the updated judgment criterion. The updated judgment criteria are subsequently distributed to each edge radar node via OTA (Over-The-Air) download technology, either through broadcast or targeted distribution, covering the old algorithm model on each node's local machine. This completes a fully automatic closed-loop upgrade of the algorithm boundary at the top level of the system topology.

[0163] It is understandable that by implementing the cloud-based large-sample accumulation and fully automated model retraining mechanism from steps S110 to S130, this application addresses the technical bottlenecks of "insufficient generalization ability of fixed models due to the one-sidedness of single monitoring environment features" and "the inability of edge devices to self-evolve due to limited computing power" at a deeper level of data processing. This application treats the results of each manual or automated machine review as extremely valuable data nutrients, transforming them into the cornerstone of strongly supervised learning through label binding, driving cloud computing power to continuously correct and converge the hyperplane boundary of the lightweight CNN machine learning model. This gives the system an adaptive anti-interference evolution mechanism similar to a biological immune system, enabling it to spontaneously suppress the global false positive and false negative rates to extremely low levels as the deployment scale and usage cycle extend, greatly improving the robustness of this fall detection method for industrial-grade deployment in heterogeneous home spaces.

[0164] In summary, the present application has the following beneficial effects based on the above embodiments:

[0165] First, it balances "accurate anomaly verification" with "absolute indoor privacy." This application employs an architecture of "radar-triggered main judgment + edge-based silent cyclic buffering + transient bidirectional slicing," ensuring the microphone remains in a low-power, volatile buffer state that is easily erased and hidden under normal conditions. The system only precisely captures short-term acoustic vectors before and after the event when the radar detects a suspected anomaly. This mechanism physically blocks the possibility of normal privacy leakage, eliminating resistance from monitored individuals to deployment in private settings such as bathrooms and bedrooms.

[0166] Secondly, this application overcomes the bottleneck of "false alarms" in traditional non-contact sensing algorithms. Traditional solutions, due to their fixed parameters, are highly susceptible to interference from environmental clutter such as rapid lying down indoors, pets jumping, and heavy objects falling. This application establishes a "local boundary incremental compensation based on false alarm feedback" mechanism, which transforms the false alarm tags output by manual or automated machine verification into dynamic penalties and corrections for judgment thresholds such as the local vertical axial velocity change rate. This self-calibrating architecture, which becomes more accurate with use, significantly reduces the overall false alarm rate of the system.

[0167] Third, a highly efficient heterogeneous data fusion and transmission paradigm with low communication load and strong anti-encryption capabilities was constructed. The local processor uses a spectral subtraction operator to strip away the steady-state environmental noise masking collision precursors and performs scale volume compression. Subsequently, through clock synchronization, the purified acoustic stream is concatenated and encapsulated with radar trajectory, position coordinates, and confidence level using multi-dimensional field strong alignment, and obfuscation operations are performed using a high-order encrypted mapping sequence. This mechanism maximizes the compression of the network transmission load of outgoing data, reduces transmission latency, and ensures data protection against eavesdropping and tampering on the outgoing link.

[0168] Fourth, it drives "adaptive iterative upgrades of judgment conditions" to achieve long-term self-evolution of the device. This application establishes a dynamic flow topology between the edge sensing end and the cloud computing array, asynchronously storing sample record data with clear true / false qualitative labels in the cloud. When the sample size reaches a limit, model retraining is automatically activated, dynamically correcting the hyperplane boundary of the machine learning model, and issuing updated judgment conditions via OTA. This process eliminates the bottleneck of fixed models being unable to cope with heterogeneous home characteristics, enabling the device to have long-term anti-interference iterative upgrade capabilities.

[0169] Furthermore, this application embodiment also provides a computer-readable storage medium (or a non-volatile computer-readable storage medium) storing a computer program (or instructions). When the computer program is executed by a processor, it implements the various steps in the above-described fall detection method embodiment. The computer-readable storage medium may include any medium capable of storing program code, including but not limited to: read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk, flash memory, hard disk (HDD), or solid-state drive (SSD). This storage medium may exist independently or be integrated into a processor or server.

[0170] The embodiments described herein may be provided as methods, systems, or computer program products. Therefore, this application may be implemented entirely in hardware, entirely in software, or a combination of hardware and software. Furthermore, this application may also be embodied as a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, optical storage, flash memory, etc.).

[0171] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this embodiment. It should be understood that each flow, block, and combination thereof in the flowchart illustrations and / or block diagrams can be implemented by computer program instructions. These instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device for execution, thereby producing a machine for implementing a specified function. Simultaneously, these instructions can also be stored in a computer-readable storage medium or loaded onto a computer device, causing the device to perform a series of operational steps to produce a computer-implemented process.

[0172] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. If such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A fall detection method, characterized by, include: Acquire the echo signal collected by the millimeter-wave radar from the monitored object, and determine whether a suspected fall event has occurred based on the echo signal; In response to the determination of a suspected fall event, the target audio data stream is extracted from the continuously cached audio data stream based on the time of occurrence of the suspected fall event; The target audio data stream and the associated feature data of the suspected fall event are encapsulated to obtain a behavior verification data packet; In response to the review request from the review terminal, the behavior review data packet is output, and the event review result for the behavior review data packet is received from the review terminal. Based on the event review results, the corresponding control operations are executed, wherein... When the event review result indicates that the event is a real event, an alarm command is triggered; When the event review result indicates a false alarm, the criteria used to determine the suspected fall event are updated.

2. The fall detection method as described in claim 1, characterized in that, Determining whether a suspected fall event has occurred based on the echo signal includes: Perform analytical transformation on the echo signal to extract the spatial point cloud distribution parameters and kinematic evolution parameters of the monitored object; A fusion assessment is performed based on the spatial point cloud distribution parameters and the kinematic evolution parameters to obtain the behavioral state indicators of the monitored object. The behavioral status indicators are matched and compared with preset behavioral boundary judgment conditions; When the behavioral state indicator meets the preset behavioral limit judgment condition, the suspected fall event is determined to have occurred.

3. The fall detection method as described in claim 2, characterized in that, The spatial point cloud distribution parameters include absolute height feature index and spatial three-dimensional divergence feature index; the kinematic evolution parameters include vertical axial velocity change rate feature index; and the preset behavior limit judgment condition is a preset confidence judgment threshold. A fusion assessment is performed based on the spatial point cloud distribution parameters and the kinematic evolution parameters to obtain behavioral state indicators of the monitored object, including: The absolute height feature index, the spatial three-dimensional divergence feature index, and the vertical axis velocity change rate feature index are input into a preset machine learning mapping model to perform feature aggregation operations. Based on the output dimension of the machine learning mapping model, the corresponding classification mapping metric is extracted, and the classification mapping metric is used as the behavior state indicator.

4. The fall detection method as described in claim 1, characterized in that, In response to determining that a suspected fall event has occurred, based on the time of occurrence of the suspected fall event, a target audio data stream is extracted from the continuously cached audio data stream, including: In response to the determination of a suspected fall event, the audio acquisition module is switched to the event response state, and the sequence timestamp corresponding to the time of occurrence is established as the reference anchor point from the audio data stream continuously cached by the audio acquisition module. Backtracking to the historical time dimension, extract the preceding audio data segment of the first preset span, and forward to the future time dimension, extract the subsequent audio data segment of the second preset span; The preceding audio data segment and the subsequent audio data segment are concatenated in a time domain to generate the target audio data stream.

5. The fall detection method as described in claim 1, characterized in that, The fall detection method also includes: During the normal monitoring period when no suspected fall event is detected, the audio acquisition module is controlled to enter a low-power silent control mode, and the time-domain cyclic overwrite buffer operation of the audio data stream is maintained.

6. The fall detection method as described in claim 1, characterized in that, The target audio data stream and the associated feature data of the suspected fall event are encapsulated to obtain a behavior verification data packet, including: The target audio data stream is subjected to noise reduction and data compression processing to obtain a normalized target audio data stream; Obtain the associated feature data of the suspected fall event, wherein the associated feature data includes the event timestamp, the behavioral data collected by radar, the coordinates of the event location, and the confidence index of the suspected event; The normalized target audio data stream and the associated feature data are subjected to multidimensional field alignment and concatenation encapsulation, and then encrypted using a preset encryption mapping sequence to obtain the encrypted behavior verification data packet.

7. The fall detection method as described in claim 3, characterized in that, When the event review result indicates a false alarm, the criteria used to determine the suspected fall event are updated, including: Extract the false trigger feedback identifier carried in the event review result; In response to the false trigger feedback flag, an initial judgment threshold corresponding to the vertical axial velocity change rate characteristic index is extracted; Perform incremental compensation operation on the initial judgment threshold to obtain the updated judgment threshold; The updated judgment threshold is used to replace the initial judgment threshold and serves as the judgment benchmark for the next monitoring cycle.

8. The fall detection method according to any one of claims 1 to 7, characterized in that, After executing the corresponding control operation based on the event review result, the fall detection method further includes: The event review results are tagged and bound together with the corresponding behavior review data packets to generate sample review record data. The sample verification record data is transmitted to the cloud storage array to perform historical sample accumulation operation; When the number of sample verification records in the cloud storage array reaches the preset iteration activation scale limit, the preset model retraining program is triggered, and the updated judgment conditions are output.

9. A fall detection device, characterized in that, The method includes a memory, a processor, and a fall detection program stored in the memory and executable on the processor, wherein the processor, when executing the fall detection program, implements the fall detection method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a fall detection program, which, when executed by a processor, implements the fall detection method as described in any one of claims 1-8.

Images