Fall detection method for elderly living alone based on Kalman filtering and sparrow search optimization

Abstract

The invention discloses a kalman filtering and sparrow search optimization-based fall detection method for elderly people living alone. The method comprises the steps of S1, obtaining calibrated original motion data; s2, inputting the calibrated original motion data into a Kalman filtering algorithm to obtain attitude feature data; s3, training an action recognition model by using a pre-established labeled reference data set and the attitude feature data, and outputting preliminary action category judgment sequence data; s4, collecting historical motion behavior data, judging sequence data in combination with the preliminary motion category, initializing a sparrow search optimization algorithm, and forming a self-adaptive motion recognition model; and S5, inputting the real-time posture feature data into the self-adaptive action recognition model, generating fall event judgment result data, and when the fall event judgment result data indicates that a fall event occurs, triggering an alarm mechanism. According to the method, the problem that the transition states of suspicious falling and slow falling cannot be recognized clearly in a traditional method is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fall detection, and in particular to a fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization. Background Art

[0002] With the increasing trend of population aging, the health and safety of elderly people living alone are receiving increasing attention from the society, especially in the early identification and response of falls. Falls are one of the main causes of injury and death among the elderly. If falls in people living alone are not discovered in time, it is very likely to cause serious consequences. Therefore, fall detection technology has become an important research direction in smart elderly care and remote health monitoring systems.

[0003] Existing fall detection methods mainly rely on two means: one is an indoor monitoring system based on video image analysis, which uses visual recognition algorithms to identify the falling behavior of the elderly; the other is a data monitoring method based on wearable devices, which usually combines accelerometers and gyroscope sensors to collect the elderly's movement information and uses traditional machine learning models for classification and identification. Although the indoor monitoring system has high accuracy, it is extremely dependent on the camera coverage environment and is easily affected by occlusion and lighting changes. There is also great controversy in terms of privacy protection; although the data monitoring method is more suitable for daily use by elderly people living alone, it also has some key technical bottlenecks.

[0004] First, traditional fall detection methods based on wearable devices mostly use low-complexity filtering methods such as sliding average and simple threshold judgment in data processing. They lack the ability to comprehensively model multi-source posture feature information, which easily leads to misjudgment and missed judgment. Secondly, most motion recognition models use fixed parameters that are universal for all individuals, and fail to fully consider the differences in physical characteristics and movement patterns of different elderly individuals, resulting in insufficient robustness in practical applications. Thirdly, the judgment rules of existing fall discrimination models are mostly based on static thresholds, which are difficult to adapt to parameter drift caused by behavioral changes in long-term use. In addition, the optimization methods for model parameters in existing research often rely on manual debugging or simple grid search, which makes it difficult to efficiently complete the exploration of large-scale parameter space and find the optimal solution.

[0005] In summary, there is an urgent need to propose a detection method that integrates efficient data filtering, personalized modeling and intelligent optimization to better meet the real-time, accuracy and adaptability requirements of fall detection for elderly people living alone. Summary of the Invention

[0006] One purpose of the present invention is to propose a fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization. The present invention effectively solves the problem of unclear identification of the transition states of "suspicious fall" and "slow fall" in traditional methods.

[0007] According to an embodiment of the present invention, a fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization includes the following steps:

[0008] S1. Collect raw motion data output by wearable sensor devices of elderly people living alone, and continuously record it at a preset sampling frequency. Perform sensor self-calibration processing on the raw motion data to obtain calibrated raw motion data;

[0009] S2. The calibrated raw motion data is input into the Kalman filter algorithm, and the Kalman filter processing result motion data is generated based on the state prediction and observation update, and the posture feature extraction is performed based on the Kalman filter processing result motion data to obtain the posture feature data;

[0010] S3. Using the pre-established labeled reference data set and posture feature data to train the action recognition model, obtain an initial action recognition model, input the real-time posture feature data into the initial action recognition model, and output the preliminary action category determination sequence data;

[0011] S4. Collect historical motion behavior data and combine it with preliminary action category determination sequence data to initialize the sparrow search optimization algorithm, perform global and local searches on the fall discrimination parameter set, obtain the optimized fall discrimination parameter set, and update the initial action recognition model based on the optimized fall discrimination parameter set to form an adaptive action recognition model;

[0012] S5. Input the real-time posture feature data into the adaptive motion recognition model to generate fall event determination result data. When the fall event determination result data indicates that a fall event has occurred, trigger the alarm mechanism and send an alarm message to the preset rescue contact through the communication module.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Collect raw motion data through wearable sensor devices of elderly people living alone. The raw motion data includes acceleration raw motion data A raw (t)=[a x (t),a y (t),a z (t)] T and the original motion data of angular velocity G raw (t) = [ω x (t),ω y (t),ω z (t)] T , where a x (t),a y (t),a z (t) represents the original acceleration motion data along the X, Y, and Z directions at time t, ω x (t),ωy (t),ω z (t) represents the original motion data of angular velocity along the X, Y, and Z directions at time t. The original motion data is sampled at the preset sampling frequency f s Continuous recording, sampling time interval is Δt=1 / f s ;

[0015] S12. Execute sensor self-calibration processing, perform scale factor correction and bias correction on the raw acceleration motion data and the raw angular velocity motion data, respectively, to obtain calibrated acceleration motion data and calibrated angular velocity motion data. The calibrated acceleration motion data is obtained by multiplying the raw acceleration motion data by the acceleration scale factor of the corresponding axis and then adding the acceleration bias term of the axis. The calibrated angular velocity motion data is obtained by multiplying the raw angular velocity motion data by the angular velocity scale factor of the corresponding axis and then adding the angular velocity bias term of the axis. The scale factor is used to correct the amplitude deviation of the sensor output in different axes, and the bias term is used to compensate for the static output offset of the sensor. The calibration results are used to improve the accuracy of subsequent attitude estimation.

[0016] S13. Perform data integrity check on the calibrated acceleration motion data and the calibrated angular velocity motion data, identify and eliminate abnormal sampling points caused by signal loss, device oscillation or communication interruption, retain the data that meets the sampling continuity requirements to form a complete data stream, and combine the calibrated acceleration motion data after eliminating the abnormal sampling points with the calibrated angular velocity motion data to form a calibrated original motion data vector. The calibrated original motion data vector includes one acceleration correction value and one angular velocity correction value in the three directions of X, Y, and Z to ensure that the input dimensions of the posture estimation are consistent.

[0017] Optionally, S2 includes the following steps:

[0018] S21. The ratio of the height H of the elderly living alone divided by the sum of their average stride length L and the constant ε is the individual physical sign coefficient c. p ;

[0019] S22. Construct a state vector containing seven elements, write the quaternion elements q0(k), q1(k), q2(k), q3(k) in sequence, and the three-axis gyroscope zero bias estimate and individual sign coefficient c p , forming the state vector at this moment;

[0020] S23. Establish the state transfer equation and observation equation, and use the state transfer equation to map the state vector of the previous moment to the predicted state vector of the current moment The prediction error covariance matrix Pk|k-1 at this moment is calculated using the state transfer matrix and the prediction error covariance matrix at the previous moment;

[0021] S24. Calculate the difference between two consecutive sampled accelerations for each of the three axes, square them, and then sum them to obtain the energy index E at this moment. k , calculate the energy sliding average according to the energy index The energy sliding average is the energy sliding average of the previous moment multiplied by the smoothing factor λ E , plus the current energy index multiplied by 1-λ E ;

[0022] S25. Multiply the prediction error covariance matrix by the transpose of the observation matrix and multiply it by the inverse matrix of the observation residual covariance to obtain the initial Kalman gain Use the drift threshold of the previous moment plus the adjustment coefficient λ K Multiply the Frobenius norm of the initial Kalman gain minus the drift threshold of the previous moment. If the Frobenius norm exceeds the drift threshold, the new Kalman gain K is obtained by scaling it proportionally. k , otherwise the initial Kalman gain is directly used;

[0023] S26. Using Kalman gain K k Complete the state update and obtain the adaptive extended Kalman filter estimation result at this moment Contains quaternion elements, three-axis gyroscope bias estimates, and individual sign coefficients;

[0024] S27. Extract quaternion elements from the adaptive extended Kalman filter estimation results, calculate the pitch angle, roll angle, and yaw angle; then calculate the acceleration jerk value J k :

[0025] The difference between the adaptive accelerations of two consecutive samples is calculated for each of the three axes, squared, summed, and squared. The individual sign coefficient c is used. p Multiplying the integral value of the adaptive vertical acceleration in this sampling interval to obtain the sign-weighted vertical displacement D k , the pitch angle, roll angle, yaw angle, acceleration jerk value, and weighted vertical displacement of the body sign are combined in time sequence to form the posture feature data sequence T feat .

[0026] Optionally, the state transfer equation stipulates that the state vector at this moment is equal to the state vector at the previous moment after being mapped by the state transfer matrix and superimposed with process noise; the observation equation stipulates that the observation vector is the original motion data vector after calibration, which is obtained after being mapped by the observation matrix and superimposed with observation noise.

[0027] Optionally, S3 includes the following steps:

[0028] S31. Build a labeled reference dataset in, Represents the posture feature data sequence of the i-th sample, y (i) ∈Y represents the standard action category label corresponding to the i-th sample, and the action category label set Y = {standing, walking, bending, sitting, falling};

[0029] S32. Constructing action recognition model f θ :T feat →Y, the action recognition model is characterized by the model parameters θ, which adopts a dual-branch architecture combining convolutional units and gated recurrent units, and uses the softmax function to define the output layer to output the predicted probability distribution of each category:

[0030]

[0031] in, represents the probability distribution of the action category of the i-th sample, h (i) is the temporal feature representation after model encoding, W o 、b o are the output weight matrix and bias term respectively;

[0032] S33. Based on the reference dataset D ref Train the model parameters θ and define the cross entropy loss function with regularization:

[0033]

[0034] in, represents the predicted probability that the i-th sample is classified as category c, δ(·) is the indicator function, and λ θ is the regularization coefficient, which is used to suppress overfitting of the model. After the training is completed, the initial action recognition model f is obtained. θ* ;

[0035] S34. Collect real-time posture feature data of elderly people living alone and input it into the initial action recognition model f θ* , and obtain the real-time output of the preliminary action category determination result sequence:

[0036]

[0037] in, Represents the action category prediction sequence after identifying the posture feature data at each time point.

[0038] Optionally, the S4 includes the following steps:

[0039] S41. Collect historical movement behavior datasets during long-term use by elderly people living alone in, represents the posture feature data sequence at the jth moment, Predict preliminary action category labels output in the action category prediction sequence;

[0040] S42. Constructing a fall discrimination parameter set Θ = {θ1, θ2, ..., θ n}, where each θ i Represents the key parameters used for action judgment or threshold classification, including attitude angle change threshold, acceleration jerk critical value, acceleration duration window, and vital sign weighted vertical displacement threshold index;

[0041] S43. Construct a target optimization function F(Θ) suitable for fall detection. The target optimization function is used to minimize the weighted combination of the false positive rate and the missed positive rate:

[0042] F(Θ)=α·FPR(Θ)+β·FNR(Θ)+γ·(1-ACC hist (Θ));

[0043] Among them, FPR(Θ) represents the false alarm rate of non-falling actions under the current parameters, FNR(Θ) represents the missed alarm rate, and ACC hist (Θ) represents the model's response to the historical data set D under the parameter set Θ hist The prediction accuracy of , α, β, γ are the balance coefficients;

[0044] S44. Initialize the population set P of the sparrow search optimization algorithm = {Θ (1) ,Θ (2) ,…,Θ (S)}, where each individual Θ (s) is a candidate solution for the fall discrimination parameter set, S is the population size, and according to the objective optimization function F(Θ (s) )Evaluate the fitness of all individuals;

[0045] S45. In the search iteration process, the behavior strategies of the discoverer and joiner are dynamically guided according to the historical action intensity characteristics and the current posture change pattern, and the energy sliding average is integrated. Control the search radius, adaptively adjust the search weight and step size, perform global search and local search alternating iterations, and obtain the optimized fall discrimination parameter set that meets the optimal fitness.

[0046] S46. The optimized fall discrimination parameter set Θ * Applied to the initial action recognition model The parameter adjustment process is to write the key threshold of the parameter set into the feature branch path control through the embedded gate control unit to form an adaptive action recognition model Dynamically match the response threshold and discrimination strategy of the adaptive action recognition model to the posture feature data with the individual behavioral characteristics of elderly people living alone;

[0047] S47. The real-time posture feature data sequence T of the elderly living alone feat (t) Input adaptive action recognition model Output updated action category prediction sequence

[0048] Optionally, the S5 includes the following steps:

[0049] S51. Extract the posture change rate, acceleration jerk peak, vertical displacement amplitude and state sequence as the fall determination input features, and define the fall event determination result data;

[0050] S52. When the fall event determination result data is confirmed to be a fall, an alarm message is immediately sent to the preset rescue contact through the communication module. The alarm message includes the time of the fall, the fall event determination result, the real-time coordinates of the sensor and the status sequence summary.

[0051] Optionally, the fall event determination result data output rules are as follows:

[0052] Normal state: If the attitude angle change rate Δθ k ≤θ thr 、Acceleration jerk value J k ≤J thr , vertical displacement D k ≤D thr , and the action category label does not show falling, the result is judged as normal;

[0053] Suspected fall state: If any of the following conditions are met: Attitude angle change rate Δθ k >θ thr 、Acceleration jerk value J k >J thr , vertical displacement D k >D thr , but the action category is not stably output as a fall or the change is not continuous, and the judgment result is a suspected fall;

[0054] Confirm the fall state: If the following conditions are met at the same time: falls appear continuously in the action category label and Δθ is satisfied k >θ thr 、J k >J thr 、D k >D thr, the result is confirmed to be a fall.

[0055] The beneficial effects of the present invention are:

[0056] (1) The present invention expands on the traditional Kalman filter and introduces "individual vital sign coefficients" as part of the state vector. The height and stride length information of elderly people living alone are quantified into personalized parameters and embedded into the filter modeling process. By constructing a seven-dimensional state vector containing quaternion attitude elements, gyroscope offset estimates and vital sign coefficients, and combining it with a dynamic update mechanism of the energy sliding average, high-precision fusion and noise suppression of multi-source sensor data are achieved. In addition, the Kalman gain adaptive adjustment strategy with drift threshold judgment is adopted to maintain system convergence and numerical stability when the sensor status is abnormal.

[0057] (2) In the back-end of the action recognition stage, the present invention proposes an adaptive learning mechanism for fall discrimination parameters based on sparrow search optimization, couples the optimization algorithm with the gated action recognition model, constructs an optimization space containing posture angle change threshold, acceleration jerk critical value, and vertical displacement multi-dimensional parameters, and jointly optimizes the fall discrimination parameter set with the misjudgment rate, missed judgment rate and historical accuracy as the objective function, effectively avoiding the problems of "lack of generalization" and "frequent false alarms" caused by the traditional static threshold mechanism. The optimized parameters are injected into the model branch path through the gated unit to form an adaptive action recognition model with strong individual adaptability.

[0058] (3) The present invention constructs a "two-stage action recognition-parameter optimization" fusion framework: in the first stage, a convolution-gated recurrent unit combination model is used to preliminarily classify the posture feature sequence output by the sensor, and identify typical actions such as standing, walking, bending, and falling; in the second stage, the parameters are optimized through the sparrow search algorithm driven by historical motion behavior, thereby adjusting the model discrimination threshold and realizing continuous dynamic adjustment, which effectively solves the problem of unclear recognition of the transition state between "suspicious fall" and "slow fall" in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0060] Figure 1 This is a flow chart of a fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization proposed by the present invention. DETAILED DESCRIPTION

[0061] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0062] refer to Figure 1 A fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization includes the following steps:

[0063] S1. Collect raw motion data output by wearable sensor devices of elderly people living alone, and continuously record it at a preset sampling frequency. Perform sensor self-calibration processing on the raw motion data to obtain calibrated raw motion data;

[0064] S2. The calibrated raw motion data is input into the Kalman filter algorithm, and the Kalman filter processing result motion data is generated based on the state prediction and observation update, and the posture feature extraction is performed based on the Kalman filter processing result motion data to obtain the posture feature data;

[0065] S3. Using the pre-established labeled reference data set and posture feature data to train the action recognition model, obtain an initial action recognition model, input the real-time posture feature data into the initial action recognition model, and output the preliminary action category determination sequence data;

[0066] S4. Collect historical motion behavior data and combine it with preliminary action category determination sequence data to initialize the sparrow search optimization algorithm, perform global and local searches on the fall discrimination parameter set, obtain the optimized fall discrimination parameter set, and update the initial action recognition model based on the optimized fall discrimination parameter set to form an adaptive action recognition model;

[0067] S5. Input the real-time posture feature data into the adaptive motion recognition model to generate fall event determination result data. When the fall event determination result data indicates that a fall event has occurred, trigger the alarm mechanism and send an alarm message to the preset rescue contact through the communication module.

[0068] In this embodiment, S1 includes the following steps:

[0069] S11. Collect raw motion data through wearable sensor devices of elderly people living alone. The raw motion data includes acceleration raw motion data A raw (t)=[a x (t),a y (t),a z (t)] T and the original motion data of angular velocity G raw (t) = [ω x (t),ω y (t),ω z (t)] T , where a x (t),a y (t),a z(t) represents the original acceleration motion data along the X, Y, and Z directions at time t, ω x (t),ω y (t),ω z (t) represents the original motion data of angular velocity along the X, Y, and Z directions at time t. The original motion data is sampled at the preset sampling frequency f s Continuous recording, sampling time interval is Δt=1 / f s ;

[0070] S12. Execute sensor self-calibration processing, perform scale factor correction and bias correction on the raw acceleration motion data and the raw angular velocity motion data, respectively, to obtain calibrated acceleration motion data and calibrated angular velocity motion data. The calibrated acceleration motion data is obtained by multiplying the raw acceleration motion data by the acceleration scale factor of the corresponding axis and then adding the acceleration bias term of the axis. The calibrated angular velocity motion data is obtained by multiplying the raw angular velocity motion data by the angular velocity scale factor of the corresponding axis and then adding the angular velocity bias term of the axis. The scale factor is used to correct the amplitude deviation of the sensor output in different axes, and the bias term is used to compensate for the static output offset of the sensor. The calibration results are used to improve the accuracy of subsequent attitude estimation.

[0071] S13. Perform data integrity check on the calibrated acceleration motion data and the calibrated angular velocity motion data, identify and eliminate abnormal sampling points caused by signal loss, device oscillation or communication interruption, retain the data that meets the sampling continuity requirements to form a complete data stream, and combine the calibrated acceleration motion data after eliminating the abnormal sampling points with the calibrated angular velocity motion data to form a calibrated original motion data vector. The calibrated original motion data vector includes one acceleration correction value and one angular velocity correction value in the three directions of X, Y, and Z to ensure that the input dimensions of the posture estimation are consistent.

[0072] In this embodiment, S2 includes the following steps:

[0073] S21. The ratio of the height H of the elderly living alone divided by the sum of their average stride length L and the constant ε is the individual physical sign coefficient c. p ;

[0074] S22. Construct a state vector containing seven elements, write the quaternion elements q0(k), q1(k), q2(k), q3(k) in sequence, and the three-axis gyroscope zero bias estimate and individual sign coefficient c p , forming the state vector at this moment;

[0075] S23. Establish the state transfer equation and observation equation, and use the state transfer equation to map the state vector of the previous moment to the predicted state vector of the current moment The prediction error covariance matrix Pk|k-1 at this moment is calculated using the state transfer matrix and the prediction error covariance matrix at the previous moment;

[0076] S24. Calculate the difference between two consecutive sampled accelerations for each of the three axes, square them, and then sum them to obtain the energy index E at this moment. k , calculate the energy sliding average according to the energy index The energy sliding average is the energy sliding average of the previous moment multiplied by the smoothing factor λ E , plus the current energy index multiplied by 1-λ E ;

[0077] S25. Multiply the prediction error covariance matrix by the transpose of the observation matrix and multiply it by the inverse matrix of the observation residual covariance to obtain the initial Kalman gain Use the drift threshold of the previous moment plus the adjustment coefficient λ K Multiply the Frobenius norm of the initial Kalman gain minus the drift threshold of the previous moment. If the Frobenius norm exceeds the drift threshold, the new Kalman gain K is obtained by scaling it proportionally. k , otherwise the initial Kalman gain is directly used;

[0078] S26. Using Kalman gain K k Complete the state update and obtain the adaptive extended Kalman filter estimation result at this moment Contains quaternion elements, three-axis gyroscope bias estimates, and individual sign coefficients;

[0079] S27. Extract quaternion elements from the adaptive extended Kalman filter estimation results, calculate the pitch angle, roll angle, and yaw angle; then calculate the acceleration jerk value J k :

[0080] The difference between the adaptive accelerations of two consecutive samples is calculated for each of the three axes, squared, summed, and squared. The individual sign coefficient c is used. p Multiplying the integral value of the adaptive vertical acceleration in this sampling interval to obtain the sign-weighted vertical displacement D k , the pitch angle, roll angle, yaw angle, acceleration jerk value, and weighted vertical displacement of the body sign are combined in time sequence to form the posture feature data sequence T feat .

[0081] In this embodiment, the state transfer equation stipulates that the state vector at this moment is equal to the state vector at the previous moment after being mapped by the state transfer matrix and superimposed with process noise; the observation equation stipulates that the observation vector is the original motion data vector after calibration, which is obtained after being mapped by the observation matrix and superimposed with observation noise.

[0082] In this embodiment, S3 includes the following steps:

[0083] S31. Build a labeled reference dataset in, Represents the posture feature data sequence of the i-th sample, y (i) ∈Y represents the standard action category label corresponding to the i-th sample, and the action category label set Y = {standing, walking, bending, sitting, falling};

[0084] S32. Constructing action recognition model f θ :T feat →Y, the action recognition model is characterized by the model parameters θ, which adopts a dual-branch architecture combining convolutional units and gated recurrent units, and uses the softmax function to define the output layer to output the predicted probability distribution of each category:

[0085]

[0086] in, represents the probability distribution of the action category of the i-th sample, h (i) is the temporal feature representation after model encoding, W o 、b o are the output weight matrix and bias term respectively;

[0087] S33. Based on the reference dataset D ref Train the model parameters θ and define the cross entropy loss function with regularization:

[0088]

[0089] in, represents the predicted probability that the i-th sample is classified as category c, δ(·) is the indicator function, and λ θ is the regularization coefficient, which is used to suppress overfitting of the model. After training, the initial action recognition model is obtained.

[0090] S34. Collect real-time posture feature data of elderly people living alone and input it into the initial action recognition model Get the real-time output of the preliminary action category determination result sequence:

[0091]

[0092] in, Represents the action category prediction sequence after identifying the posture feature data at each time point.

[0093] In this embodiment, S4 includes the following steps:

[0094] S41. Collect historical movement behavior datasets during long-term use by elderly people living alone in, represents the posture feature data sequence at the jth moment, Predict preliminary action category labels output in the action category prediction sequence;

[0095] S42. Constructing a fall discrimination parameter set Θ = {θ1, θ2, ..., θ n}, where each θ i Represents the key parameters used for action judgment or threshold classification, including attitude angle change threshold, acceleration jerk critical value, acceleration duration window, and vital sign weighted vertical displacement threshold index;

[0096] S43. Construct a target optimization function F(Θ) suitable for fall detection. The target optimization function is used to minimize the weighted combination of the false positive rate and the missed positive rate:

[0097] F(Θ)=α·FPR(Θ)+β·FNR(Θ)+γ·(1-ACC hist (Θ));

[0098] Among them, FPR(Θ) represents the false alarm rate of non-falling actions under the current parameters, FNR(Θ) represents the missed alarm rate, and ACC hist (Θ) represents the model's response to the historical data set D under the parameter set Θ hist The prediction accuracy of , α, β, γ are the balance coefficients;

[0099] S44. Initialize the population set P of the sparrow search optimization algorithm = {Θ (1) ,Θ (2) ,…,Θ (S)}, where each individual Θ (s) is a candidate solution for the fall discrimination parameter set, S is the population size, and according to the objective optimization function F(Θ (s) )Evaluate the fitness of all individuals;

[0100] S45. In the search iteration process, the behavior strategies of the discoverer and joiner are dynamically guided according to the historical action intensity characteristics and the current posture change pattern, and the energy sliding average is integrated. Control the search radius, adaptively adjust the search weight and step size, perform global search and local search alternating iterations, and obtain the optimized fall discrimination parameter set that meets the optimal fitness.

[0101] S46. The optimized fall discrimination parameter set Θ * Applied to the initial action recognition model The parameter adjustment process is to write the key threshold of the parameter set into the feature branch path control through the embedded gate control unit to form an adaptive action recognition model Dynamically match the response threshold and discrimination strategy of the adaptive action recognition model to the posture feature data with the individual behavioral characteristics of elderly people living alone;

[0102] S47. The real-time posture feature data sequence T of the elderly living alone feat (t) Input adaptive action recognition model Output updated action category prediction sequence

[0103] In this embodiment, S5 includes the following steps:

[0104] S51. Extract the posture change rate, acceleration jerk peak, vertical displacement amplitude and state sequence as the fall determination input features, and define the fall event determination result data;

[0105] S52. When the fall event determination result data is confirmed to be a fall, an alarm message is immediately sent to the preset rescue contact through the communication module. The alarm message includes the time of the fall, the fall event determination result, the real-time coordinates of the sensor and the status sequence summary.

[0106] In this embodiment, the fall event determination result data output rules are as follows:

[0107] Normal state: If the attitude angle change rate Δθ k ≤θ thr 、Acceleration jerk value J k ≤J thr , vertical displacement D k ≤D thr , and the action category label does not show falling, the result is judged as normal;

[0108] Suspected fall state: If any of the following conditions are met: Attitude angle change rate Δθ k >θ thr 、Acceleration jerk value J k >J thr , vertical displacement D k >D thr , but the action category is not stably output as a fall or the change is not continuous, and the judgment result is a suspected fall;

[0109] Confirm the fall state: If the following conditions are met at the same time: falls appear continuously in the action category label and Δθ is satisfied k >θ thr 、J k >J thr 、D k >D thr, the result is confirmed to be a fall.

[0110] Example 1:

[0111] At 7:30 am on October 15, 2024, Ms. Zhang, an elderly woman living alone in City A, got up at home as usual to prepare breakfast. Ms. Zhang is 78 years old and suffers from mild osteoporosis and chronic hypotension. Once a fall occurs, it is very likely to cause serious consequences. In order to ensure her daily safety, in August 2024, the community health service center and the smart elderly care platform deployed a system based on the present invention on her. The system includes a six-axis inertial measurement unit sensor worn on the waist, a local processing terminal and a remote alarm communication module.

[0112] At 7:42 that morning, when Ms. Zhang bent down in the kitchen to pick up a dropped spatula, she suddenly lost her balance due to the slippery ground. She fell sideways and backwards, landing on her waist in a typical fall. The system completed the entire process from data collection to fall recognition and alarm triggering in less than 3 seconds. Finally, the emergency contact person and community caregiver preset by the platform arrived at the scene within 5 minutes to deal with the situation.

[0113] The specific technical implementation of the entire detection process is as follows:

[0114] The sensor collects Ms. Zhang's acceleration and angular velocity data at a frequency of 100 Hz. In the method of the present invention, the collected data first undergoes a self-calibration process of proportional factor and bias correction to eliminate initial oscillation data and communication interruption points to form a stable original motion data vector. On this basis, the system performs adaptive Kalman filtering based on the extended state vector. The state vector contains Ms. Zhang's attitude quaternion, three-axis gyroscope deviation estimate and its individual vital sign coefficient, so that the attitude estimation has the ability of individualized adjustment.

[0115] After the filtering module outputs the posture estimation sequence, the system further extracts the pitch angle, roll angle, jerk value and vertical displacement data to form a time series feature sequence. The model uses a convolution-gated loop structure to extract the spatiotemporal representation of the feature sequence and preliminarily identifies it as a "bend-fall" transition action sequence. At the same time, the system calls the historical behavior database for approximately 28 hours of action sequence data collected by Ms. Zhang over the past 8 weeks, and fuses it with the real-time data for calculation.

[0116] Traditional methods often use static preset thresholds to detect falls (acceleration peak exceeding 2.5g and posture angle change exceeding 60° trigger an alarm), which can easily misjudge violent but non-falling movements, such as rapid sitting or jumping. The present invention introduces a sparrow search optimization algorithm to jointly evaluate the false alarm rate and missed alarm rate based on historical behavioral characteristics and current sequence change trends. By dynamically adjusting the target optimization function through weights, it automatically finds the optimal discrimination parameter set that meets individual characteristics. For example, the jerk threshold is 1.92m / s. 3 , vertical displacement threshold = 0.26m, attitude angle change threshold = 52°, and embed them into the gating module to update the classification model parameters in real time.

[0117] In this incident, the total delay from the occurrence of the fall to the system judging it as a "confirmed fall" state and issuing an alarm was only 2.85 seconds, of which sensor data acquisition and filtering processing took about 0.84 seconds, posture feature extraction and classification judgment took 1.21 seconds, and optimization parameter retrieval and judgment update took 0.8 seconds. The alarm information included "Fall timestamp: 2024-10-1507:42:17", "Location coordinates (derived from the cross-positioning of the wearable device and the home WiFi location): 31.673N, 119.963E", "Judgment category: confirmed fall" and "Status summary: jerk peak = 2.03m / s 3 , pitch angle = 67.3°, vertical displacement = 0.34m", ensuring the accuracy and timeliness of subsequent rescue response.

[0118] To verify the advantages of the present invention in actual scenarios, we selected 6 elderly people living alone to participate in comparative tests between September 5 and October 5, 2024. Each elderly person wore the detection equipment for no less than 8 hours a day, and the cumulative monitoring data was about 960 hours. We selected 8 actions including "falling", "falling down", "sitting down quickly", and "turning around violently" as the standard action library. In the same scenario and equipment, we used the traditional static threshold detection method and the method of the present invention to test them respectively. The accuracy, false positive rate, and missed positive rate of each method are shown in the following table:

[0119] Table 1 Data comparison between traditional static threshold detection method and the method of the present invention

[0120]

[0121] One of the participants, Mr. Wang (72 years old), was doing a quick squat at home on September 21. The traditional method misjudged it as a fall and issued a false alarm. However, the method of the present invention accurately determined it as a "sitting down" action and did not trigger the alarm, thus avoiding unnecessary intervention. On the same evening, Mr. Wang actually slipped while taking an object from a high place, triggering a fall alarm and receiving timely rescue, thus avoiding further deterioration of the head trauma.

[0122] In summary, this Example 1 comprehensively verifies the advantages of the method of the present invention in posture estimation accuracy, fall recognition accuracy and individual adaptability through real elderly home scenes, traceable behavioral data, system response logs and misjudgment comparison tests. It solves the key problems of the existing technology in frequent false alarms, static parameter inadaptability, and large influence of posture noise, and has significant practical value and promotion potential.

[0123] This invention expands on the traditional Kalman filter and introduces "individual vital sign coefficients" as part of the state vector. The height and stride length information of elderly people living alone is quantified into personalized parameters and embedded into the filtering modeling process. By constructing a seven-dimensional state vector containing quaternion attitude elements, gyroscope offset estimates and vital sign coefficients, and coordinating it with a dynamic update mechanism of the energy sliding average, high-precision fusion and noise suppression of multi-source sensor data are achieved. In addition, a Kalman gain adaptive adjustment strategy with drift threshold judgment is adopted to maintain system convergence and numerical stability even when the sensor status is abnormal.

[0124] In the back-end of the action recognition stage, the present invention proposes an adaptive learning mechanism for fall discrimination parameters based on sparrow search optimization, couples the optimization algorithm with the gated action recognition model, constructs an optimization space including posture angle change threshold, acceleration jerk critical value, and vertical displacement multi-dimensional parameters, and jointly optimizes the fall discrimination parameter set with the false positive rate, missed positive rate, and historical accuracy as objective functions, effectively avoiding the problems of "insufficient generalization" and "frequent false alarms" caused by the traditional static threshold mechanism. The optimized parameters are injected into the model branch path through the gated unit to form an adaptive action recognition model with strong individual adaptability.

[0125] This paper constructs a "two-stage action recognition-parameter optimization" fusion framework: in the first stage, a convolution-gated recurrent unit combination model is used to preliminarily classify the posture feature sequence output by the sensor, identifying typical actions such as standing, walking, bending, and falling; in the second stage, parameter optimization is performed through a sparrow search algorithm driven by historical motion behavior, thereby adjusting the model discrimination threshold and realizing continuous dynamic adjustment, effectively solving the problem of unclear recognition of the transition state between "suspicious fall" and "slow fall" in traditional methods.

[0126] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization, characterized in that: The steps include: S1. Collect and calibrate the raw motion data output by the wearable sensor device of the elderly living alone to obtain the calibrated raw motion data; S2. The calibrated raw motion data is input into the Kalman filter algorithm, and the Kalman filter processing result motion data is generated based on the state prediction and observation update, and the posture feature extraction is performed based on the Kalman filter processing result motion data to obtain the posture feature data; S3. Using the pre-established labeled reference data set and posture feature data to train the action recognition model, obtain an initial action recognition model, input the real-time posture feature data into the initial action recognition model, and output the preliminary action category determination sequence data; S4. Collect historical motion behavior data and combine it with preliminary action category determination sequence data to initialize the sparrow search optimization algorithm, perform global and local searches on the fall discrimination parameter set, obtain the optimized fall discrimination parameter set, and update the initial action recognition model based on the optimized fall discrimination parameter set to form an adaptive action recognition model; S5. Input the real-time posture feature data into the adaptive motion recognition model to generate fall event determination result data. When the fall event determination result data indicates that a fall event has occurred, trigger the alarm mechanism and send an alarm message to the preset rescue contact through the communication module.

2. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 1 is characterized in that: Said S1 comprises the following steps: S11. Collect raw motion data through wearable sensor devices of elderly people living alone. The raw motion data includes acceleration raw motion data A raw (t)=[a x (t),a y (t),a z (t)] T and the original motion data of angular velocity G raw (t) = [ω x (t),ω y (t),ω z (t)] T , where a x (t),a y (t),a z (t) represents the original acceleration motion data along the X, Y, and Z directions at time t, ω x (t),ω y (t),ω z (t) represents the original motion data of angular velocity along the X, Y, and Z directions at time t. The original motion data is sampled at the preset sampling frequency f s Continuous recording, sampling time interval is Δt=1 / f s ; S12. Execute sensor self-calibration processing, perform scale factor correction and bias correction on the original acceleration motion data and the original angular velocity motion data, respectively, to obtain calibrated acceleration motion data and calibrated angular velocity motion data; S13. Perform data integrity check on the calibrated acceleration motion data and the calibrated angular velocity motion data, identify and eliminate abnormal sampling points caused by signal loss, device oscillation or communication interruption, retain the data that meets the sampling continuity requirements to form a complete data stream, and combine the calibrated acceleration motion data after eliminating the abnormal sampling points with the calibrated angular velocity motion data to form a calibrated original motion data vector.

3. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 2 is characterized in that: The S2 comprises the following steps: S21. The ratio of the height H of the elderly living alone divided by the sum of their average stride length L and the constant ε is the individual physical sign coefficient c. p ; S22. Construct a state vector containing seven elements, write the quaternion elements q0(k), q1(k), q2(k), q3(k) in sequence, and the three-axis gyroscope zero bias estimate and individual sign coefficient c p , forming the state vector at this moment; S23. Establish the state transfer equation and observation equation, and use the state transfer equation to map the state vector of the previous moment to the predicted state vector of the current moment The prediction error covariance matrix Pk|k-1 at this moment is calculated using the state transfer matrix and the prediction error covariance matrix at the previous moment; S24. Calculate the difference between two consecutive sampled accelerations for each of the three axes, square them, and then sum them to obtain the energy index E at this moment. k , calculate the energy sliding average according to the energy index The energy sliding average is the energy sliding average of the previous moment multiplied by the smoothing factor λ E , plus the current energy index multiplied by 1-λ E ; S25. Multiply the prediction error covariance matrix by the transpose of the observation matrix and multiply it by the inverse matrix of the observation residual covariance to obtain the initial Kalman gain Use the drift threshold of the previous moment plus the adjustment coefficient λ K Multiply the Frobenius norm of the initial Kalman gain minus the drift threshold of the previous moment. If the Frobenius norm exceeds the drift threshold, the new Kalman gain K is obtained by scaling it proportionally. k , otherwise the initial Kalman gain is directly used; S26. Using Kalman gain K k Complete the state update and obtain the adaptive extended Kalman filter estimation result at this moment Contains quaternion elements, three-axis gyroscope bias estimates, and individual sign coefficients; S27. Extract quaternion elements from the adaptive extended Kalman filter estimation results, calculate the pitch angle, roll angle, and yaw angle; then calculate the acceleration jerk value J k : The difference between the adaptive accelerations of two consecutive samples is calculated for each of the three axes, squared, summed, and squared. The individual sign coefficient c is used. p Multiplying the integral value of the adaptive vertical acceleration in this sampling interval to obtain the sign-weighted vertical displacement D k , the pitch angle, roll angle, yaw angle, acceleration jerk value, and weighted vertical displacement of the body sign are combined in time sequence to form the posture feature data sequence T feat .

4. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 3 is characterized in that: The state transfer equation stipulates that the state vector at this moment is equal to the state vector at the previous moment after being mapped by the state transfer matrix and superimposed with process noise; the observation equation stipulates that the observation vector is the original motion data vector after calibration, which is obtained after being mapped by the observation matrix and superimposed with observation noise.

5. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 3 is characterized in that: The S3 includes the following steps: S31. Build a labeled reference dataset in, Represents the posture feature data sequence of the i-th sample, y (i) ∈Y represents the standard action category label corresponding to the i-th sample, and the action category label set Y = {standing, walking, bending, sitting, falling}; S32. Constructing action recognition model f θ :T feat →Y, the action recognition model is characterized by the model parameters θ, which adopts a dual-branch architecture combining convolutional units and gated recurrent units, and uses the softmax function to define the output layer to output the predicted probability distribution of each category: in, represents the probability distribution of the action category of the i-th sample, h (i) is the temporal feature representation after model encoding, W o 、b o are the output weight matrix and bias term respectively; S33. Based on the reference dataset D ref Train the model parameters θ and define the cross entropy loss function with regularization: in, represents the predicted probability that the i-th sample is classified as category c, δ(·) is the indicator function, and λ θ is the regularization coefficient, which is used to suppress overfitting of the model. After the training is completed, the initial action recognition model f is obtained. θ* ; S34. Collect real-time posture feature data of elderly people living alone and input it into the initial action recognition model f θ* , and obtain the real-time output of the preliminary action category determination result sequence: in, Represents the action category prediction sequence after identifying the posture feature data at each time point.

6. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 5 is characterized in that: The S4 comprises the following steps: S41. Collect historical movement behavior datasets during long-term use by elderly people living alone in, represents the posture feature data sequence at the jth moment, Predict preliminary action category labels output in the action category prediction sequence; S42. Constructing a fall discrimination parameter set Θ = {θ1, θ2, ..., θ n }, where each θ i Represents the key parameters used for action judgment or threshold classification, including attitude angle change threshold, acceleration jerk critical value, acceleration duration window, and vital sign weighted vertical displacement threshold index; S43. Construct a target optimization function F(Θ) suitable for fall detection. The target optimization function is used to minimize the weighted combination of the false positive rate and the missed positive rate: F(Θ)=α·FPR(Θ)+β·FNR(Θ)+γ·(1-ACC hist (I)); Among them, FPR(Θ) represents the false alarm rate of non-falling actions under the current parameters, FNR(Θ) represents the missed alarm rate, and ACC hist (Θ) represents the model's response to the historical data set D under the parameter set Θ hist The prediction accuracy of , α, β, γ are the balance coefficients; S44. Initialize the population set P of the sparrow search optimization algorithm = {Θ (1) ,Θ (2) ,…,Θ (S) }, where each individual Θ (s) is a candidate solution for the fall discrimination parameter set, S is the population size, and according to the objective optimization function F(Θ (s) )Evaluate the fitness of all individuals; S45. In the search iteration process, the behavior strategies of the discoverer and joiner are dynamically guided according to the historical action intensity characteristics and the current posture change pattern, and the energy sliding average is integrated. Control the search radius, adaptively adjust the search weight and step size, perform global search and local search alternating iterations, and obtain the optimized fall discrimination parameter set that meets the optimal fitness. S46. The optimized fall discrimination parameter set Θ * Applied to the initial action recognition model f θ* The parameter adjustment process is to write the key threshold of the parameter set into the feature branch path control through the embedded gate control unit to form an adaptive action recognition model Dynamically match the response threshold and discrimination strategy of the adaptive action recognition model to the posture feature data with the individual behavioral characteristics of elderly people living alone; S47. The real-time posture feature data sequence T of the elderly living alone feat (t) Input adaptive action recognition model Output updated action category prediction sequence 7. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 6 is characterized in that: The S5 comprises the following steps: S51. Extract the posture change rate, acceleration jerk peak, vertical displacement amplitude and state sequence as the fall determination input features, and define the fall event determination result data; S52. When the fall event determination result data is confirmed to be a fall, an alarm message is immediately sent to the preset rescue contact through the communication module. The alarm message includes the time of the fall, the fall event determination result, the real-time coordinates of the sensor and the status sequence summary.

8. The fall detection method for elderly people living alone based on Kalman filtering and sparrow search optimization according to claim 7 is characterized in that: The output rules of the fall event determination result data are as follows: Normal state: If the attitude angle change rate Δθ k ≤θ thr 、Acceleration jerk value J k ≤J thr , vertical displacement D k ≤D thr , and the action category label does not show falling, the result is judged as normal; Suspected fall state: If any of the following conditions are met: Attitude angle change rate Δθ k >θ thr 、Acceleration jerk value J k >J thr , vertical displacement D k >D thr , but the action category is not stably output as a fall or the change is not continuous, and the judgment result is a suspected fall; Confirm the fall state: If the following conditions are met at the same time: falls appear continuously in the action category label and Δθ is satisfied k >θ thr 、J k >J thr 、D k >D thr , the result is confirmed to be a fall.

Images