Fall detection method and system using cross-validation of time-series barometric pressure trajectories and inertial data

Abstract

The application discloses a fall detection method and system based on cross verification of time sequence air pressure trajectory and inertial data, and belongs to the technical field of fall detection. The application comprises the following steps: S1, system initialization; S2, synchronous acquisition of motion data and air pressure data, and giving all data a unified timestamp through a time sequence synchronization module; S3, preliminary fall suspicion identification; S4, dynamic division of a time sequence window, taking the preliminary fall suspicion point as a time reference, and dynamically dividing a forward reference window, a process dynamic window and a post-reference window; S5, cross verification calculation, calculating a first average acceleration and a second average acceleration, and an absolute difference value between the two; S6, judging whether the preliminary fall suspicion point is a real fall event according to the absolute difference value. The application adopts the above method and system, realizes more accurate and reliable identification of a fall event, and significantly reduces the false positive rate in a complex daily activity scene.

Description

Technical Field

[0001] This invention relates to the field of fall detection technology, and in particular to a fall detection method and system that cross-validates time-series air pressure trajectories and inertial data. Background Technology

[0002] With the increasing aging of the global population, the safety monitoring of elderly people living alone has become an important social issue. Falls are one of the most common accidental injuries among the elderly, making timely and accurate detection crucial. Against this backdrop, wearable fall detection devices based on inertial measurement units (IMUs) have emerged and become a current research and application hotspot.

[0003] Currently, mainstream fall detection methods primarily rely on data collected by IMUs (which typically include a three-axis accelerometer and a gyroscope). These methods can be broadly categorized into two types: first, threshold-based methods, which identify the intensity of movement by setting thresholds for acceleration, angular velocity, or attitude angles; once the data exceeds the preset threshold, a fall is determined. Second, pattern recognition-based methods, which distinguish fall-related activities from daily activities (such as walking, sitting, and jumping) by training machine learning models (such as support vector machines and convolutional neural networks).

[0004] However, the aforementioned IMU-based methods have inherent limitations, resulting in a consistently high false alarm rate in real-world scenarios. The fundamental reason is that many everyday activities (such as sitting down quickly, landing after a jump, or bending down to pick up an object) have movement patterns similar to actual fall heights, making it difficult to effectively distinguish them using only inertial data. Although some studies have attempted to introduce multi-sensor fusion strategies to improve accuracy, such as using barometers as auxiliary sensors, these approaches still have significant limitations in application: most solutions only use barometers for static height compensation or as a simple data overlay for instantaneous judgment (e.g., using a judgment of whether the height has significantly decreased to aid decision-making). This simple fusion fails to fully utilize the continuous changes in barometric pressure data over time, and cannot construct a dynamic verification mechanism that reflects the physical characteristics of the entire fall process. Summary of the Invention

[0005] The purpose of this invention is to provide a fall detection method and system that cross-validates time-series air pressure trajectories and inertial data. From the perspective of time dynamic evolution, it performs deeper fusion and collaborative analysis of multi-source sensor data, thereby achieving more accurate and reliable identification of fall events and significantly reducing the false alarm rate in complex daily activity scenarios.

[0006] To achieve the above objectives, this invention provides a fall detection method based on cross-validation of time-series barometric trajectories and inertial data, comprising the following steps:

[0007] S1, System initialization;

[0008] S2. Multi-source data synchronous acquisition: Motion data and air pressure data are acquired synchronously through inertial measurement unit and barometer, and a unified timestamp is assigned to all data through time synchronization module;

[0009] S3, Primary fall suspicion identification: Based on motion data, a lightweight convolutional neural network deployed on an embedded platform is used to identify primary fall suspicion points;

[0010] S4. Dynamic division of time window: taking the initial suspected fall point as the time reference, and combining the preset time offset and air pressure sampling period, dynamically divide the forward reference window, process dynamic window and backward reference window.

[0011] S5. Cross-validation calculation: Extract time-series air pressure trajectory from the air pressure data in the process dynamic window and calculate the first average acceleration based on the air pressure-altitude conversion and uniform acceleration motion model. At the same time, calculate the second average acceleration from the motion data in the process dynamic window and calculate the absolute difference between the first average acceleration and the second average acceleration.

[0012] S6. Determine whether the primary suspected fall is a real fall event based on the absolute difference.

[0013] Preferably, step S2 specifically includes the following steps:

[0014] S21, using the inertial measurement unit with motion sampling period Collect human motion data, including triaxial acceleration and triaxial angular velocity data;

[0015] S22. Simultaneously, the barometer is used to sample the air pressure periodically. Collect ambient air pressure data ;

[0016] S23, motion data, and barometric pressure data are all assigned a unified timestamp through a timing synchronization module. .

[0017] Preferably, in step S3, after preprocessing the motion data, it is input into a lightweight convolutional neural network model for analysis. When the confidence level of the model indicating a fall exceeds a preset threshold, the current moment is recorded as a primary suspected fall point. .

[0018] Preferably, in step S4, the forward reference window The time range is The forward reference point is Extract 2 from this window NSeveral barometric pressure data points form a forward time-series barometric pressure trajectory, Trace_front(t), which represents the stable height of the human body before the fall. This is the preset time offset;

[0019] The process dynamic window The time range is Extract the contents of this window Several barometric pressure data points are used to construct a time-series barometric pressure trajectory, Trace_process(t), to capture drastic changes in altitude during a fall. The minimum time interval, and its relationship with the preset time offset are as follows: ;

[0020] The rear reference window The time range is The rear reference point is Extract 2 from this window N The barometric pressure data points form the post-time barometric pressure trajectory Trace_rear(t), which is used to represent the static height of the human body after the fall.

[0021] Preferably, step S5 specifically includes the following steps:

[0022] S51. Extract the time-series air pressure trajectory from the air pressure data in the process dynamic window and calculate the first average acceleration based on the air pressure-altitude conversion and uniformly accelerated motion model. ;

[0023] S52. Calculate the second average acceleration from the motion data in the process dynamics window. From the motion data in step S1, extract the data in the process dynamic window. Vertical acceleration data over a time period And calculate the second average acceleration within that window. The formula is as follows:

[0024] ;

[0025] S53, Calculation and absolute difference The formula is as follows:

[0026] .

[0027] Preferably, step S51 specifically includes the following steps:

[0028] S511. Extract each pressure value from the process time-series pressure trajectory Trace_process(t). Convert to relative altitude using the international standard barometric altitude formula. The formula is as follows:

[0029] ;

[0030] in, At sea level, atmospheric pressure is the standard pressure. The standard temperature at sea level This represents the vertical temperature lapse rate. R Let be the ideal gas constant. It is the acceleration due to gravity. M The molar mass of dry air;

[0031] S512, Calculation Process Dynamic Window The start time and the end time The height difference between The formula is as follows:

[0032] ;

[0033] S513. Based on the formula for uniformly accelerated linear motion, calculate the dynamic window during this process. Inside, the first average acceleration derived from the change in altitude. The formula is as follows:

[0034] ;

[0035] in, For process dynamic window Total duration .

[0036] Preferably, in step S6, the specific determination method is as follows:

[0037] Preset tolerance threshold The absolute difference With tolerance threshold In comparison, when At that time, the primary suspected fall point was determined. This is a real fall event, triggering a subsequent alarm mechanism;

[0038] when If the alarm is false, the system will not issue an alarm and will resume normal monitoring.

[0039] A fall detection system based on cross-validation of time-series barometric trajectories and inertial data includes:

[0040] The main control module provides sufficient computing power support for time-series air pressure trajectory analysis and cross-validation algorithms;

[0041] The time series analysis module, integrated within the main control module, is used to execute the dynamic partitioning of the time series window and the cross-validation calculation step.

[0042] The sensor module includes a six-axis inertial measurement unit and a high-precision barometer, which are used to collect motion data and barometric pressure data, respectively.

[0043] The time synchronization module is used to provide a unified timestamp for all collected motion and air pressure data, ensuring the synchronization of multi-source data in the time dimension.

[0044] The decision-making module is used to determine whether a fall is genuine and to control the alarm mechanism.

[0045] The human-computer interaction module is used to display user status and provide fall alarm functions;

[0046] The communication module is used to send alarm information to the monitoring terminal when a real fall event is detected;

[0047] The power management module is used to provide a stable operating voltage for all modules in the system.

[0048] Preferably, the time series analysis module includes:

[0049] The window division unit is used to dynamically divide the forward reference window, the process dynamic window, and the backward reference window based on the primary fall suspicion point.

[0050] The trajectory extraction unit is used to extract the corresponding time-series air pressure trajectory from each time window;

[0051] The physical quantity calculation unit is used to calculate the first average acceleration and the second average acceleration.

[0052] Therefore, the fall detection method and system of the present invention, which employs the above-mentioned cross-validation of time-series air pressure trajectory and inertial data, has the following advantages:

[0053] (1) In this invention, time-series air pressure trajectory analysis is introduced to realize dynamic perception. By dividing the time windows into forward, process and backward, a time-series air pressure trajectory that can fully reflect the dynamic changes in height before and after the fall event is constructed. This in-depth mining of air pressure data in the time dimension provides a richer and more reliable dynamic information dimension for judgment.

[0054] (2) In this invention, deep cross-validation at the physical level is achieved, which significantly reduces false alarms. By comparing the first average acceleration derived from the air pressure trajectory with the second average acceleration directly measured by the inertial sensor, deep data fusion between sensors with different physical principles is achieved. This is fundamentally different from simple data superposition or threshold judgment. It can effectively identify those daily activities that mimic falling in inertial data but do not actually undergo drastic changes in height, thereby greatly reducing the false alarm rate of the system.

[0055] (3) In this invention, a primary fall suspicion identification is set up, and then through fine trajectory extraction and cross-validation, it is determined whether a fall has actually occurred. This multi-level decision mechanism significantly improves the reliability of the fall detection system and the user's trust.

[0056] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0057] Figure 1 This is a flowchart of a fall detection method based on cross-validation of time-series air pressure trajectory and inertial data according to the present invention;

[0058] Figure 2 A temporal window partitioning diagram of a fall detection method based on cross-validation of temporal barometric trajectory and inertial data according to the present invention;

[0059] Figure 3 This is a schematic diagram of the structure of a fall detection system based on cross-validation of time-series air pressure trajectory and inertial data according to the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0061] Example

[0062] like Figures 1-2 As shown, a fall detection method based on cross-validation of time-series barometric trajectories and inertial data includes the following steps:

[0063] S1. System Initialization: After the system is powered on, the real-time operating system is initialized, and the following tasks are created: sensor data acquisition task, convolutional neural network inference task, time series analysis cross-validation task, LVGL interface refresh task, and data transfer between tasks is achieved through message queue to ensure real-time system response.

[0064] S2. Multi-source data synchronous acquisition: Motion data and air pressure data are acquired synchronously through inertial measurement unit and barometer, and a unified timestamp is assigned to all data through time synchronization module;

[0065] S21, using the inertial measurement unit with motion sampling period Collect human motion data, including triaxial acceleration. and triaxial angular velocity data ;

[0066] S22. Simultaneously, the barometer is used to sample the air pressure periodically. Collect ambient air pressure data ;

[0067] S23, motion data, and barometric pressure data are all assigned a unified timestamp through a timing synchronization module. .

[0068] S3. Basic fall detection: After preprocessing the motion data, it is input into a lightweight convolutional neural network model for analysis. When the confidence level of the model in detecting a fall exceeds a preset threshold... At that time, the current moment is recorded as the initial suspected fall point. ;

[0069] S4. Dynamic division of time window: taking the initial suspected fall point as the time reference, and combining the preset time offset and air pressure sampling period, dynamically divide the forward reference window, process dynamic window and backward reference window.

[0070] Forward reference window The time range is The forward reference point is Extract 2 from this window N Several barometric pressure data points form a forward time-series barometric pressure trajectory, Trace_front(t), which represents the stable height of the human body before the fall. This is the preset time offset;

[0071] Process Dynamic Window The time range is Extract the contents of this window Several barometric pressure data points are used to construct a time-series barometric pressure trajectory, Trace_process(t), to capture drastic changes in altitude during a fall. The minimum time interval, given by the scheduler interrupt, is typically 1ms. Its relationship with the preset time offset is as follows: ;

[0072] Rear reference window The time range is The rear reference point is Extract 2 from this window N The barometric pressure data points form the post-time barometric pressure trajectory Trace_rear(t), which is used to represent the static height of the human body after the fall.

[0073] S5. Cross-validation calculation: Extract time-series air pressure trajectory from the air pressure data in the process dynamic window and calculate the first average acceleration based on the air pressure-altitude conversion and uniform acceleration motion model. At the same time, calculate the second average acceleration from the motion data in the process dynamic window and calculate the absolute difference between the first average acceleration and the second average acceleration.

[0074] S51. Extract the time-series air pressure trajectory from the air pressure data in the process dynamic window and calculate the first average acceleration based on the air pressure-altitude conversion and uniformly accelerated motion model. ;

[0075] S511. Extract each pressure value from the process time-series pressure trajectory Trace_process(t). Convert to relative altitude using the international standard barometric altitude formula. The formula is as follows:

[0076] ;

[0077] in, The standard atmospheric pressure at sea level is taken as 1013.25 hPa. The standard temperature at sea level is 288.15 K. The vertical temperature lapse rate is taken as 0.0065 K / m. R The ideal gas constant is given by a value of 8.31432 J / (mol·K). The acceleration due to gravity is 9.80665. , M The molar mass of dry air is taken as 0.0289644 kg / mol;

[0078] S512, Calculation Process Dynamic Window The start time and the end time The height difference between The formula is as follows:

[0079] ;

[0080] S513. Based on the formula for uniformly accelerated linear motion, calculate the dynamic window during this process. Inside, the first average acceleration derived from the change in altitude. The formula is as follows:

[0081] ;

[0082] in, For process dynamic window Total duration ;

[0083] S52. Calculate the second average acceleration from the motion data in the process dynamics window. From the motion data in step S1, extract the data in the process dynamic window. Vertical acceleration data over a time period And calculate the second average acceleration within that window. The formula is as follows:

[0084] ;

[0085] S53, Calculation and absolute difference The formula is as follows:

[0086] ;

[0087] S6. Determine whether a primary suspected fall is a genuine fall event based on the absolute difference, with a preset tolerance threshold. The absolute difference With tolerance threshold In comparison, when At that time, the initial suspicion of a fall was determined. This is a real fall event, triggering a subsequent alarm mechanism;

[0088] when If the alarm is false, the system will not issue an alarm and will resume normal monitoring.

[0089] like Figure 3 As shown, a fall detection system that cross-validates time-series barometric pressure trajectories and inertial data includes:

[0090] The main control module uses the STM32H743IIT6 chip as the core processor. This chip has a built-in double-precision floating-point arithmetic unit with a main frequency of 480MHz, providing sufficient computing power support for time-series barometric trajectory analysis and cross-validation algorithms.

[0091] The time series analysis module, integrated within the main control module, is used to execute the dynamic partitioning step of the time series window and the cross-validation calculation step. It includes: a window partitioning unit, used to dynamically partition the forward reference window, the process dynamic window, and the backward reference window based on the primary fall suspect point; a trajectory extraction unit, used to extract the corresponding time series air pressure trajectory from each time window; and a physical quantity calculation unit, used to calculate the first average acceleration and the second average acceleration.

[0092] The sensor module, including a six-axis inertial measurement unit (IMU) and a high-precision barometer, has sampling periods of [missing information]. and Collect motion data and air pressure data;

[0093] The timing synchronization module uses the DS3231 high-precision real-time clock chip to provide a unified timestamp for all collected motion data and barometric pressure data, ensuring the synchronization of multi-source data in the time dimension.

[0094] The decision-making module is used to determine whether a fall is genuine and to control the alarm mechanism.

[0095] The human-computer interaction module includes a 1.3-inch TFT touch screen based on the LVGL graphics library, a vibration motor, and an alarm module including a sound alarm, which realizes user status display and fall alarm functions;

[0096] The communication module uses an HC-05 Bluetooth module to send alarm information to the monitoring terminal when a real fall event is detected.

[0097] The power management module uses the TP4056 lithium battery management chip to provide stable operating voltage for all modules in the system.

[0098] In terms of software architecture, a layered design is adopted, including:

[0099] At the operating system layer, multi-task scheduling is implemented based on FreeRTOS, and shared resources are protected by semaphores to ensure the real-time performance of sensor data acquisition, neural network inference, and time series analysis.

[0100] In the artificial intelligence layer, the STM32Cube.AI toolchain is used to convert the pre-trained Keras model into highly optimized C code, and 8-bit integer quantization inference is implemented on the STM32H743IIT6 chip.

[0101] At the application layer, the LVGL graphics library is used to build the user interface displayed on the touch screen, implementing a low-power refresh mechanism with the interface refresh rate controlled at 30fps.

[0102] Therefore, the present invention adopts the fall detection method and system of cross-validation of time-series air pressure trajectory and inertial data, and performs deeper fusion and collaborative analysis of multi-source sensor data from the dimension of time dynamic evolution, thereby achieving more accurate and reliable identification of fall events and significantly reducing the false alarm rate in complex daily activity scenarios.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fall detection method of cross-verification of time series barometric pressure trajectory and inertial data, characterized by: The method comprises the following steps: S1, system initialization; S2, multi-source data synchronous acquisition, synchronous acquisition of motion data and air pressure data through an inertial measurement unit and an air pressure gauge, and all data are given a unified timestamp through a timing synchronization module; S3, preliminary fall suspicion identification, identification of a preliminary fall suspicion point based on motion data through a lightweight convolutional neural network deployed on an embedded platform; S4, dynamic division of a timing window, dynamic division of a forward reference window, a process dynamic window and a post-reference window based on the preliminary fall suspicion point as a time reference, in combination with a preset time offset and an air pressure sampling period; S5, cross-validation calculation, extraction of a timing air pressure trajectory from air pressure data in the process dynamic window and calculation of a first average acceleration based on air pressure-height conversion and a uniform speed motion model, while a second average acceleration is calculated from motion data in the process dynamic window, and an absolute difference value between the first average acceleration and the second average acceleration is calculated; Step S5 specifically comprises the following steps: S51, extracting a time-series air pressure trajectory from the air pressure data within the process dynamic window and calculating a first average acceleration based on the air pressure-height conversion and the uniform motion model ; S52. Calculate a second average acceleration from motion data within a process dynamic window ; from the motion data of step S2, extract vertical acceleration data for a process dynamic window time period , and calculate a second average acceleration within that window , as follows: ； wherein, is a gas pressure sampling period, is a minimum time interval, which has a relationship with a preset time offset is ; S53, calculate with the absolute difference , as follows: ； Step S51 specifically comprises the following steps: S511. Extract each pressure value from the process time-series pressure trajectory Trace_process(t). Convert to relative altitude using the international standard barometric altitude formula. The formula is as follows: ； wherein is the sea level standard atmospheric pressure, is the sea level standard temperature, is the temperature lapse rate, R is the ideal gas constant, is the acceleration of gravity, M is the molar mass of dry air; S512, compute process dynamic window of the start time and the end time between the start time of the end time ； S513、Based on the uniform velocity straight line motion formula, the first average acceleration derived from the height change in the process dynamic window the first average acceleration derived from the height change , the formula is as follows: ； wherein is the total duration of the process dynamic window of the process dynamic window ; S6, judging whether the preliminary fall suspicion point is a real fall event according to the absolute difference value.

2. The fall detection method of claim 1, wherein: Step S2 specifically comprises the following steps: S21, using the inertial measurement unit with motion sampling period Collect human motion data, including triaxial acceleration and triaxial angular velocity data; S22, simultaneously, by the barometer with a barometric sampling period collecting ambient barometric data ; S23, the motion data and the air pressure data are both given a unified time stamp by a time sequence synchronization module .

3. The time-series barometric pressure trajectory and inertial data cross-validated fall detection method of claim 2, wherein: In step S3, after the motion data is preprocessed, the motion data is input into a lightweight convolutional neural network model for analysis. When the confidence of the model falling exceeds a preset threshold, the current time is recorded as a primary fall suspect point .

4. The time-series gait trajectory and inertial data cross-validated fall detection method according to claim 3, wherein: In step S4, the forward reference window has a time range of , a forward reference point is , 2 N pressure data points in the window are extracted to form a forward time-series pressure trace Trace_front(t) for representing the stable height of the human body before the fall occurs, wherein is a preset time offset. The process dynamic window The time range of the process dynamic window is The process dynamic window is used to extract the air pressure data points within the window to form a process time sequence air pressure track Trace_process(t), which is used to capture the sharp change of height during the fall process, wherein is the minimum time interval, and the relationship between and the preset time offset is ; The rear reference window The time range is The rear reference point is Extract 2 from this window N The barometric pressure data points form the post-time barometric pressure trajectory Trace_rear(t), which is used to represent the static height of the human body after the fall.

5. The time sequence of air pressure trajectory and inertial data cross verification fall detection method according to claim 4, characterized in that: In step S6, the specific judgment method is as follows: Preset tolerance threshold The absolute difference With tolerance threshold In comparison, when At that time, the primary suspected fall point was determined. This is a real fall event, triggering a subsequent alarm mechanism; When a false positive is determined, the system does not alarm and returns to normal monitoring.

6. A fall detection system for performing the time-series barometric pressure trajectory and inertial data cross-validated fall detection method of any one of claims 1-5, characterized by: It comprises: A main control module provides sufficient computing power support for timing air pressure trajectory analysis and cross-validation algorithms; A timing analysis module integrated in the main control module is used to perform the steps of dynamic division of a timing window and cross-validation calculation; A sensor module comprising a six-axis inertial measurement unit and a high-precision air pressure gauge is used to collect motion data and air pressure data respectively; A timing synchronization module is used to provide a unified timestamp for all collected motion data and air pressure data, ensuring the synchronization of multi-source data in the time dimension; A decision-making module is used to judge whether it is a real fall and control the alarm mechanism; A human-computer interaction module is used to realize user state display and fall alarm functions; A communication module is used to send alarm information to a monitoring terminal when a real fall event is detected; A power management module is used to provide stable working voltage for each module of the system.

7. The time sequence air pressure trajectory and inertial data cross-verification fall-detecting system according to claim 6, characterized in that: The timing analysis module comprises: A window division unit is used to dynamically divide a forward reference window, a process dynamic window and a post-reference window based on the preliminary fall suspicion point as a reference; A trajectory extraction unit is used to extract corresponding timing air pressure trajectories from each time window; A physical quantity calculation unit is used to calculate the first average acceleration and the second average acceleration.

Images