A physically injection-resistant high-dynamic robust fall detection method and medium

Abstract

The application relates to a high-dynamic robust fall detection method and medium against physical injection, which comprises the following steps: S1, collecting multi-source observation data of an intelligent walking aid device, and constructing an initial state vector, wherein the multi-source observation data contain real-time acceleration characteristic signals collected by using an accelerometer; S2, judging whether a physical injection attack occurs based on the real-time acceleration characteristic signals, if not, returning to step S1 to re-collect, and if yes, performing filtering processing by using a narrow-band trap wave filtering algorithm based on a human body dynamics mask to obtain pure low-frequency human body motion characteristic signals after cleaning; and S3, when the pure low-frequency human body motion characteristic signals after cleaning trigger a fall threshold, performing fall detection judgment by using a heterogeneous data fusion method based on Kalman filtering based on the pure low-frequency human body motion characteristic signals after cleaning, residual multi-source observation data and the state vector, and outputting a detection result. Compared with the prior art, the application has the advantages of detection reliability and the like.

Description

Technical Field

[0001] This invention relates to the fields of smart wearable safety and digital signal processing, and in particular to a highly dynamic and robust fall detection method and medium resistant to physical injection. Background Technology

[0002] With the accelerating global aging process, smart mobility aids for the elderly (such as smart canes) are playing an increasingly important role in fall detection and emergency rescue. Highly reliable fall detection algorithms are a core technology for ensuring the safety of the elderly. Current fall detection methods largely rely on microelectromechanical systems (MEMS) inertial sensors (such as accelerometers and gyroscopes), which determine danger by capturing the dynamic characteristics of human movement.

[0003] However, in real-world applications, systems often face complex physical environmental interference and potential security threats. Recent security research (such as that published in *Walnut*) has revealed serious physical layer vulnerabilities in inertial sensors: attackers can inject false, raw signals directly into the sensors using methods such as directional sound waves and ultrasonic resonance. These physical attacks can bypass traditional software security protections and directly interfere with the underlying physical sensing logic.

[0004] Existing fall detection technologies mainly fall into the following three categories: 1. Threshold-based detection method: This method determines the detection outcome by setting hard thresholds such as acceleration amplitude and attitude angle change rate. While the algorithm is simple and has good real-time performance, it is highly susceptible to high-frequency vibrations or environmental noise, exhibiting a very high false alarm rate under physical signal injection attacks.

[0005] 2. Deep learning-based detection methods: These methods utilize convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) to identify features in time-domain waveforms. While this approach achieves high accuracy on standard datasets, it is a "black box" model, lacking verification of the physical validity of the input signal. If the underlying sensors are physically compromised, the model will generate incorrect conclusions due to "misleading input."

[0006] 3. Conventional filtering methods: Low-pass filtering, Kalman filtering, and other techniques are used to smooth noise. However, physical injection attacks typically manifest as high-energy narrowband resonant signals, which are fundamentally different from Gaussian white noise. When traditional filters remove this type of specific frequency interference, they often cause phase lag or amplitude attenuation of the actual motion characteristic signal, resulting in a significant decrease in the sensitivity of fall detection.

[0007] Therefore, existing technologies generally face the following technical challenges: (1) Unreliability at the physical perception level: Existing algorithms assume that the data collected by the sensor is the real physical feedback, and lack the identification mechanism for the "physical source" and "frequency fingerprint" of the signal, which makes the system easily blinded in adversarial physical environments.

[0008] (2) The contradiction between detection sensitivity and anti-interference ability: In order to prevent abnormal noise, existing solutions often adopt an over-smoothing strategy, which directly reduces the system's ability to capture the high-frequency impact characteristics of the fall and causes the risk of false alarm.

[0009] (3) Lack of cross-modal verification mechanism: Most schemes rely on a single accelerometer signal and lack the logic of using multi-dimensional physical quantities such as air pressure and geomagnetism to conduct spatiotemporal consistency "cross verification", making it difficult to cope with high-intensity targeted attacks against a single sensor.

[0010] Therefore, a new fall detection technology is urgently needed to solve the above problems. Summary of the Invention

[0011] The purpose of this invention is to provide a highly dynamic and robust fall detection method and medium that coordinates and unifies high dynamic recognition capability and robust safety performance against physical injection.

[0012] The objective of this invention can be achieved through the following technical solutions: A highly dynamic and robust fall detection method resistant to physical injection includes the following steps: S1. Collect multi-source observation data from the intelligent walking assistance device and construct an initial state vector. The multi-source observation data includes real-time acceleration characteristic signals collected using an accelerometer. ; S2, Based on the real-time acceleration characteristic signal If a physical injection attack occurs, return to step S1 to collect data again. If it does, use a narrowband notch filter algorithm based on human dynamics mask to filter the signal and obtain a clean low-frequency human motion feature signal. S3. When the cleaned low-frequency human motion feature signal triggers the fall threshold, the fall detection judgment is performed based on the cleaned low-frequency human motion feature signal, the remaining multi-source observation data, and the state vector using a heterogeneous data fusion method based on Kalman filtering, and the detection result is output.

[0013] Furthermore, the multi-source observation data also includes real-time altitude characteristic signals collected by barometers, real-time attitude characteristic signals collected by magnetometers, and real-time angular velocity characteristic signals collected by gyroscopes.

[0014] Furthermore, the process of determining whether a physical injection attack has occurred includes: Within a microsecond-level frequency domain monitoring window, the real-time acceleration characteristic signal is analyzed using a sliding window fast Fourier transform method. Convert to frequency domain energy distribution , represented as: , In the formula, For the Hanning window function, The imaginary unit, For frequency variables, for time, This represents the time shift of the sliding window; The normal human kinematic frequency range is set as the first frequency range, and a low-frequency protection mask is applied to the first frequency range. The frequency range in which the sensor physically resonates is set as the second frequency range, wherein the second frequency range is greater than the first frequency range. For the frequency domain energy distribution Calculate the power spectral density; The presence of an abnormally high narrowband resonant peak is detected based on the power spectral density within the second frequency range. If it is, it indicates that a physical injection attack has occurred; otherwise, it indicates that no physical injection attack has occurred.

[0015] Furthermore, the detection steps for the abnormally abruptly increased narrowband resonance peak include: Set a real-time updated environmental background noise threshold; When the power spectral density amplitude at a certain frequency point is detected to exceed the environmental background noise threshold within the second frequency range, and the frequency domain energy distribution... If the bandwidth is extremely narrow, it is determined that an abnormally high narrowband resonance peak has been detected; otherwise, it has not been detected.

[0016] Furthermore, the real-time acceleration characteristic signal The data was acquired at the first sampling frequency. In the event of a physical injection attack, the sampling frequency is switched to the second sampling frequency to acquire real-time acceleration characteristic signals. Utilizing the real-time acceleration characteristic signal at the second sampling frequency Perform filtering.

[0017] Furthermore, the narrowband notch filtering algorithm based on human body dynamics mask employs a transfer function. Filtering is performed, the transfer function Represented as: , In the formula, For the Laplace operator, The center angular frequency, For the attack frequency identified in real time, This is the notch bandwidth.

[0018] Furthermore, the execution steps of the heterogeneous data fusion method based on Kalman filtering include: definition The state vector at time t is ,in, For the height of intelligent walking assistance devices, The vertical velocity, The physical attitude tilt angle; definition The control input vector at time t is ,in, This refers to the clean, low-frequency human motion characteristic signals after cleaning. This is a real-time angular velocity characteristic signal; The time update matrix equation for constructing the prediction model is expressed as: , , , In the formula, for The prior state prediction vector at time t. Here is the system state transition matrix. for The optimal posterior state estimate vector at time t. To control the input matrix, for The control input vector at time t, To control the periodic sampling time interval; Within each control cycle, the optimal posterior state estimation vector from the previous time step is... and the control input vector of the current control cycle Substituting the time update matrix equation into the equation and performing kinematic derivation, we obtain the prior state prediction vector at the current moment. ; The prior error covariance matrix is ​​derived recursively, and is expressed as: , In the formula, Let be the prior error covariance matrix. for The posterior error covariance matrix at time t. The process noise covariance matrix; Construct the actual observation vector based on the multi-source observation data. Using the observation matrix Predict the prior state vector Values ​​are mapped to the observation space to obtain the predicted observation vector. ; Based on the actual observation vector With predicted observation vector Perform residual comparison to obtain the measurement residual vector. ; Based on the prior error covariance matrix The observation noise covariance matrix of the multi-source observation data is combined with Calculate the residual covariance matrix , represented as: , Using the prior error covariance matrix With residual covariance matrix Calculate the Kalman gain matrix for the current control cycle in real time. , represented as: , Using the Kalman gain matrix For the prior state prediction vector Make corrections and output the current result. The optimal posterior state estimation vector at time t And update the posterior error covariance matrix simultaneously. For use in the next control cycle, where the optimal posterior state estimation vector and the updated verification error covariance matrix They are represented as follows: , , In the formula, It is an identity matrix.

[0019] Furthermore, the fall detection judgment is performed by calculating the Mahalanobis distance, wherein the expression for calculating the Mahalanobis distance is: , In the formula, This is the Mahalanobis distance.

[0020] Furthermore, the step of outputting the detection result includes: Determine the Mahalanobis distance Does it meet the requirements? If yes, a physical deception attack is considered to exist, triggering a false alarm interception command, recording an anomaly log, and issuing a defense prompt. If no, a physical deception attack is considered not to exist, and it is further determined whether the cleaned, purified low-frequency human motion characteristic signal exceeds the fall detection threshold. If yes, it is confirmed as a genuine fall, triggering a rescue alarm. If no, it is considered that there is no genuine fall, and no alarm is triggered. Set as the preset robust confidence threshold.

[0021] The present invention also provides a storage medium storing a program thereon, characterized in that, when the program is executed, it implements a highly dynamic and robust fall detection method resistant to physical injection as described above.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses real-time acceleration feature signals to identify and extract the "frequency domain fingerprint" of non-natural motion features, thereby establishing an illegal signal filtering mechanism at the physical perception level. It also uses a narrowband notch filter algorithm to accurately remove high-frequency physical injection interference without damaging the real motion feature signals, thus solving the contradiction between detection sensitivity and anti-interference capability. At the same time, it introduces a heterogeneous data fusion method based on Kalman filtering to fuse multi-source heterogeneous data and quantitatively evaluate the consistency of each sensor data in physical logic evolution. This ensures high real-time performance of fall detection while achieving a coordinated unity of high dynamic recognition capability and robust safety performance in adversarial physical environments.

[0023] (2) In the signal processing stage, this invention employs a combination of sliding window fast Fourier transform and adaptive narrowband notch filtering to map the abnormal resonant waveform generated by the sensor to the frequency domain in real time for "fingerprint" extraction. By locking and protecting the biological motion components in the first frequency range through human dynamics masking, malicious injection signals in the extremely high frequency band (second frequency range) are accurately identified and eliminated. This mechanism enhances the perception system's ability to identify and eliminate physical attacks at specific frequencies, solves the fatal flaw of traditional defense methods that cause the loss of instantaneous features during falls due to "over-smoothing," and ensures the accuracy of the algorithm in capturing highly dynamic movements.

[0024] (3) This invention constructs a heterogeneous state space integrating accelerometers, barometers, and magnetometers. Utilizing the physical inconsistency and spatiotemporal coupling logic among the sensor data, a unified verification input is constructed through Kalman filtering and Mahalanobis distance calculation. This architecture enables the system to simultaneously consider the consistency of sensor dynamics and physical evolution when assessing fall risk, improving robustness and anti-spoofing capabilities in environments where a single sensor is hijacked or blinded. Furthermore, by introducing Mahalanobis distance to establish a "cross-verification" logic for multi-source sensors, this invention does not rely on complex nonlinear optimization solutions, exhibiting extremely high computational efficiency and meeting the requirements of embedded terminals for real-time fall detection response.

[0025] (4) This invention addresses the challenges of aging-related mobility assistance and adversarial physical environments by constructing a real-time safe fall detection structure that incorporates frequency domain fingerprint recognition and cross-modal spatiotemporal verification. During the execution phase, this structure performs physical validity verification on the underlying inertial signals and outputs a true action judgment that excludes physical hijacking interference through a two-level filtering mechanism, thereby achieving real-time safety and high dynamic performance for fall monitoring in complex environments.

[0026] (5) In adversarial scenario tests involving high-intensity physical injection attacks such as ultrasonic resonance (second frequency range), the present invention demonstrated extremely high defensive effectiveness. Experimental data shows that: the false alarm rate is reduced: the false alarm rates of traditional thresholding methods and deep learning methods are as high as 85.0% and 72.0% respectively when encountering physical injection; while the present invention, through a dual defense mechanism, significantly reduces the false alarm rate induced by the attack to 3.0%; the high dynamic sensitivity is maintained: in a strong interference environment, the present invention still maintains a true fall detection sensitivity of 97.0%, ensuring that while intercepting malicious attacks, it does not miss the real distress signals of the elderly, achieving a coordinated unity of safety and performance. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a system structure diagram of the present invention; Figure 3 This is a waveform comparison diagram of the domain fingerprint recognition and adaptive signal filtering modules of the present invention; Figure 4 This is a decision curve diagram of the cross-modal spatiotemporal consistency verification module of the present invention; Figure 5 This is a bar chart illustrating the overall system defense performance evaluation of the present invention. Figure 6 This is a schematic diagram of the multi-level defense decision-making process of the present invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0029] This embodiment provides a highly dynamic and robust fall detection method resistant to physical injection. This method is based on a corresponding detection system (such as...). Figure 2 As shown in the diagram, this system constructs a dual-defense architecture based on "frequency domain cleaning" and "cross-modal verification" to achieve real-time identification and interception of physical injection attacks. Specifically, it includes multiple heterogeneous sensors (including accelerometers, barometers, and magnetometers) and a main control chip (such as an ESP32) connected to these sensors. The main control chip includes a Spectral-Imprint Filtering Module and a Cross-Modal Spatiotemporal Consistency Verification Module, used for physical attack interception and actual fall detection. Specifically, as shown in the diagram... Figure 1 As shown, the steps the system takes to perform fall detection include: 1. Signal preprocessing, initialization calibration, and frequency domain feature extraction: During system startup or reset, the main control chip first performs initialization calibration, establishing a zero-bias reference for the sensors by collecting static environmental background data, providing a reliable physical reference for subsequent consistency verification. After calibration, the system operates at the first sampling frequency. (i.e., the base sampling rate) receives real-time acceleration characteristic signals from the accelerometer in the smart mobility aid device. (Real-time domain signal). To capture high-frequency attack fingerprints while also meeting the system's low-power normal monitoring requirements, the first sampling frequency... The optimal sampling frequency is set above the Nyquist sampling law, which satisfies the normal human body dynamics frequency range (e.g., 50Hz~100Hz). A second sampling frequency switching logic is triggered only when a high-frequency injection attack causes aliasing anomalies in the low-frequency signal. The system constructs a microsecond-level sliding window and uses a sliding-window Fast Fourier Transform (Sliding-Window FFT) to convert the time-domain signal into a frequency-domain energy distribution. : , in, This is a Hanning Window function used to reduce spectral leakage.

[0030] 2. Narrowband notch filtering algorithm and dynamic sampling rate switching logic based on human body dynamics mask: The system backend maintains a low-frequency protection mask (Biomechanical Mask) for a first frequency range (e.g., the human biomechanical frequency range), defined as the "natural biological movement safety zone." A second frequency range (e.g., the sensor's physical resonant frequency band) is also defined, where the second frequency range is larger than the first. The purpose of defining the first frequency range is to delineate the legitimate biological movement frequency bands that must be fully preserved (to prevent accidental injury); the purpose of the second frequency range is to delineate the physical attack frequency bands that require focused monitoring (for precision strikes).

[0031] During operation, the system analyzes the frequency domain energy distribution obtained above. Power spectral density (PSD) calculation is performed. The specific method for detecting abnormally high narrowband resonant peaks is as follows: the system updates the ambient background noise threshold in real time. When the PSD amplitude of a certain frequency point is detected to exceed the noise threshold within the second frequency range, and its energy distribution exhibits an extremely narrow bandwidth (i.e., a very high quality factor Q), the system determines that an abnormally high narrowband resonant peak has been detected, thus extracting an illegal frequency domain fingerprint and confirming that a physical injection attack has occurred. Otherwise, it is considered that no abnormally high narrowband resonant peak has been detected, and no physical injection attack has occurred.

[0032] At this point, the system immediately triggers an early warning mechanism, issuing an abnormal status alert through the background log or communication interface, and executes dynamic sampling rate switching logic: the main control chip automatically increases the sampling frequency from the first sampling frequency to a higher-resolution second sampling frequency to obtain high-precision attack feature fingerprints. Subsequently, the system strictly uses the high-resolution real-time acceleration feature signal obtained after switching to the second sampling frequency as the input data stream, automatically tuning the adaptive narrowband notch filter, whose transfer function... Based on the identified attack frequency Perform precise leveling: , in, The center angular frequency, This is the notch bandwidth. This process ensures that while eliminating attack components in a specific frequency band, the true motion waveform in the first frequency range is fully preserved.

[0033] In summary, this step triggers an adaptive narrowband notch filter algorithm upon detecting an illegal resonance peak. This algorithm directly uses the abnormal resonance peak frequency extracted in real-time by a pre-sliding window Fast Fourier Transform (FFT) as the attack frequency, and dynamically tunes the center frequency and stopband width of the notch filter accordingly. After processing by this algorithm, the system accurately eliminates attack components in specific frequency bands, outputting a cleaned, low-frequency human motion characteristic signal, while fully preserving the real motion characteristic data within the first frequency range (e.g., the human dynamics frequency range), providing a reliable data source for subsequent fall detection.

[0034] Steps 1-2 above are integrated into the domain fingerprint recognition and adaptive signal filtering module, which serves to filter out the surge in sensor noise and false waveforms caused by physical attacks such as ultrasonic waves and lasers.

[0035] 3. Cross-modal spatiotemporal consistency verification decision: This step primarily uses the cleaned, purified low-frequency human motion feature signal obtained in the previous step as the core observation component input (or as the dynamic driving term input for state estimation) of the Kalman filter in the downstream cross-modal spatiotemporal consistency verification method. Furthermore, during fusion, the Kalman filter simultaneously receives the cleaned, purified low-frequency human motion feature signal, the real-time altitude feature signal acquired by the barometer, and the real-time attitude feature signal acquired by the magnetometer to jointly construct a multi-dimensional heterogeneous state vector. This provides reliable underlying data support for the state estimation of the post-Kalman filter and the cross-modal consistency verification of Mahalanobis distance, achieving hierarchical collaborative defense from 'signal-level purification' to 'system-level verification'. Specifically, this step includes: When the cleaned acceleration characteristics (pure low-frequency human motion characteristic signals) trigger the fall threshold, the system executes a heterogeneous data fusion method based on Kalman filtering to perform fall detection and judgment: (1) Definition of state vector and control input definition The state vector of the smart walking aid device is ,in, The height of the intelligent walking aid device (corresponding to the space estimated by the barometer). The vertical velocity, The physical attitude tilt angle; The system's control input vector is defined as ,in, This represents the cleaned, low-frequency human motion characteristic signal (vertical axis component). This refers to the real-time angular velocity characteristic signal acquired by the gyroscope.

[0036] (2) Time update matrix equation of the prediction model The time update matrix equation of the prediction model is expressed as: , in, for The prior state prediction vector at time t; for The optimal posterior state estimate vector at time t; To control the periodic sampling time interval; Here is the system state transition matrix. To control the input matrix, its specific matrix structure is defined as follows: , , (3) Prediction model processing: At the beginning of each control cycle, the main control chip executes the predicted model: First, the optimal posterior state estimation vector from the previous time step... With the clean control input of the current cycle Substituting the above time update matrix equation, we complete the kinematic derivation and output the prior prediction vectors of the device's height, velocity, and attitude at the current moment. ; Subsequently, the prior error covariance matrix is ​​recursively calculated. (in (This is the process noise covariance matrix).

[0037] The prior prediction value output by the prediction model The data will be directly transmitted to subsequent measurement update steps and compared with the actual height measurement value of the barometer and the actual attitude measurement value of the magnetometer to provide a benchmark reference for multi-source data correlation detection and Mahalanobis distance calculation.

[0038] The specific steps for measurement update (correction) and residual comparison are as follows: Obtain the prior state prediction vector at the current moment. With the prior error covariance matrix Subsequently, when a new round of actual sampling data arrives from the barometer (providing altitude observation) and the magnetometer (providing attitude observation), the system executes a measurement update process, the specific steps of which are as follows: Step A (Residual Comparison and Calculation) The system reads the actual observation vectors from the heterogeneous sensors. Using the observation matrix Mapping the prior state predictions to the observation space yields the predicted observation vector. The system performs residual comparison and calculates the measurement residual vector (Innovation): , The residual vector The physical significance lies in quantifying the deviation between the motion state deduced based on pure acceleration and the environmental state actually observed by air pressure / geomagnetic field.

[0039] Step B (Calculation of residual covariance matrix and Kalman gain) At this point, the prior error covariance matrix output from the aforementioned prediction stage is invoked. Combined with the observation noise covariance matrix of heterogeneous sensors The system first calculates the residual covariance matrix. : , Subsequently, using ,and Real-time calculation of the Kalman gain matrix for the current period The reliability of dynamic tradeoff model predictions compared to actual measurements: , Step C (Post-hoc state update): Using the calculated Kalman gain It corrects the prior predicted state and outputs the optimal posterior state estimate vector at the current time. And update the posterior error covariance matrix simultaneously. For use in the next control cycle: , , In the formula, It is an identity matrix.

[0040] After the above measurement update steps, the system extracts the measurement residual vector. With residual covariance matrix This will be directly used as the benchmark parameter for subsequent multi-source data correlation detection, in order to calculate the Mahalanobis distance between multi-source observations (acceleration, barometric altitude, geomagnetic attitude) in real time. : .

[0041] The detection and judgment logic is as follows: like , To establish a robust confidence threshold, if the accelerometer output is found to logically conflict with the physical evolution of air pressure and geomagnetism, it is confirmed that a single sensor (such as the accelerometer) has been subjected to a high-intensity targeted physical injection attack or malicious hijacking. In this case, a forced interception mechanism will be executed to directly suppress and refuse to trigger a fall alarm, thereby cutting off false alarm commands caused by physical deception, recording anomaly logs, and continuously sending defense prompts to the user or terminal.

[0042] like The system determines that the multi-source heterogeneous data possesses spatiotemporal physical consistency, confirming that the sensor data from each sensor of the intelligent walking aid device is authentic and reliable. At this point, a data trust pass is issued to the system. Under this pass status, the system substitutes the cleaned, purified acceleration characteristic signal into the fall detection rules: if this motion characteristic triggers the fall detection threshold (combined with the sudden drop in air pressure and posture rollover characteristics), the system ultimately confirms that the target has actually fallen and triggers an emergency distress alarm command; otherwise, if the fall detection threshold is not triggered, it is determined to be normal walking or minor daily fluctuations, and the system does not issue an alarm.

[0043] By setting an adjustable safety factor (A preset robust confidence threshold) is used to adjust the strength of the safety constraints after multi-source sensor fusion. Mahalanobis distance is used to quantify the degree to which observed values ​​deviate from the normal physical evolution trajectory, and alarm suppression is only triggered when severe conflicts occur in multi-dimensional logic. This mechanism, while ensuring defense strength, maximizes the retention of rapid response sensitivity in real-world fall scenarios, achieving a harmonious balance between defense robustness and real-time detection.

[0044] Through the above-mentioned serial verification control logic, the Mahalanobis distance verification result serves as a necessary control prerequisite for the release of the fall detection alarm. This ensures that an emergency distress call is only issued under the dual closed-loop conditions of 'data and physical consistency' and 'motion characteristics conforming to fall characteristics'. This fundamentally balances the system's robustness against physical deception with its sensitivity to real fall rescue.

[0045] Step 3 above is integrated into the cross-modal spatiotemporal consistency verification module, which can prevent a single sensor (such as an accelerometer) from being completely blinded or hijacked by high-intensity directional sound waves.

[0046] Based on the fall detection method described above, the following verifications were performed in this embodiment: (1) Normal recognition verification under complex walking postures This embodiment is used to verify the system's ability to retain "biological motion signals" under different daily behaviors. In the experiment, the subjects wore smart canes integrating ESP32 main control chip and heterogeneous sensor array, and performed three typical activities: brisk walking on flat ground, climbing stairs, and walking on gravel roads.

[0047] Experimental setup: The system was set to a sliding sampling window of 50ms and a 0-20Hz human dynamics mask was enabled.

[0048] Experimental Results: In a continuous dynamic walking test lasting 120 minutes, the system accurately identified normal low-frequency motion components using a sliding window FFT, and the waveform closely matched the actual physical movements. Under normal conditions, the system achieved a success rate of over 98.5% in recognizing real falls. These results demonstrate that this approach effectively filters out high-frequency mechanical vibration noise in everyday environments while maintaining high dynamic recognition sensitivity.

[0049] (2) Safety threshold Verification of the impact on the robustness of the judgment This embodiment adjusts the Mahalanobis distance safety threshold in the cross-modal spatiotemporal consistency verification module. This verifies the system's ability to adjust between defense strength and detection sensitivity.

[0050] Experimental scheme: Simulate directional sound waves to perform a high-intensity resonant attack on the accelerometer.

[0051] Experimental results: When When set to 5.0, the security constraint triggering conditions are relatively sensitive. Although all physical attacks are intercepted, there is a very low probability of an "interception alarm" in transient scenarios involving violent collisions with the ground.

[0052] when When upgraded to 8.0 (preferred parameter), the Mahalanobis distance can perfectly quantify the physical contradiction between acceleration and barometric altitude.

[0053] Statistical data: In Under these conditions, the false positive rate against 27kHz physical injection attacks was significantly reduced from 85% of the baseline method to 3%, while maintaining 97% detection sensitivity. This demonstrates that by adjusting... The parameters enable extremely high defensive robustness in complex adversarial environments.

[0054] (3) Noise robustness verification of heterogeneous sensors This embodiment is used to evaluate the operational stability of the system when sensor accuracy decreases or when there is random environmental noise (such as air pressure fluctuations or geomagnetic interference).

[0055] Experimental design: Inject 10% random Gaussian noise into the output of the heterogeneous sensor to simulate the reading deviation of the sensor under aging or extreme weather conditions.

[0056] Experimental Results: With the help of the cross-modal spatiotemporal consistency verification module, the Kalman filter can effectively smooth sensor drift through residual analysis of predicted and observed values. Experimental results show that even with a 10% decrease in sensing accuracy, the accuracy fluctuation of Mahalanobis distance determination is less than 4%. This result demonstrates that the screening structure combining the security value function with physical state logic has extremely strong fault tolerance.

[0057] (4) Interception performance verification under high-intensity physical hijacking This embodiment verifies the system's interception performance when encountering extreme physical attacks such as Walnut and adversarial process control.

[0058] Experimental scheme: The attacker used an ultrasonic device to send a modulated resonant waveform, attempting to force the accelerometer to generate a sinusoidal characteristic signal that satisfies the fall threshold.

[0059] Experimental procedure: The original acceleration signal was severely distorted when injected at 27 kHz, and subfigure (a) shows that the amplitude instantaneously exceeded the traditional threshold.

[0060] The first line of defense uses frequency domain fingerprinting to identify extremely high frequency abnormal resonance peaks and activate adaptive notch filtering.

[0061] The second line of defense detected acceleration and reported "descent," but the pressure altitude remained constant, and the Mach-scale distance... It surged exponentially, exceeding the safety threshold.

[0062] Figure 3 The effectiveness of the domain fingerprint recognition and adaptive signal filtering module in the embodiments of the present invention has been verified. Figure 3 (a) shows the severe distortion of a conventional time-domain acceleration signal when subjected to a 27 kHz physical injection attack; Figure 3 (b) in the figure reveals the extremely high frequency anomalous resonance peaks that were accurately identified in the frequency domain using sliding window FFT; Figure 3 (c) in the figure proves that after adaptive narrowband notch filtering, the system successfully eliminated high-frequency attack components and output a pure biological motion characteristic waveform that fits the actual fall height.

[0063] Figure 4 This demonstrates the dynamic decision-making process of the cross-modal spatiotemporal consistency verification module in preventing single-sensor hijacking. When encountering a physical deception attack (red line), the Mahalanobis distance spikes exponentially and exceeds the system's safety threshold due to physical logic contradictions arising from changes in acceleration, air pressure, and geomagnetism, triggering the interception mechanism. However, in a real fall scenario (green line), the spatiotemporal data from multiple sensors remain consistent, and the Mahalanobis distance stabilizes within the safe zone, thus ensuring the reliable transmission of genuine distress signals.

[0064] Figure 5This paper compares and evaluates the comprehensive performance of the method of this invention with existing baseline methods under adversarial physical attack environments. When attacked, traditional thresholding methods and deep learning methods produce extremely high false alarm rates (85.0% and 72.0%, respectively); while this system maintains a high real fall detection sensitivity of 97.0% while significantly reducing the false alarm rate to 3.0%, achieving a balance between defensive robustness and high detection dynamism.

[0065] Figure 6 This diagram illustrates the multi-level defense decision-making process of an embodiment of the present invention, fully presenting the logical process of the system from sensor initialization calibration. The system first performs initialization calibration to establish a zero-bias reference. In real-time monitoring, abnormal resonances are identified through a first-level decision bifurcation: if an attack is detected, dynamic sampling frequency adaptation and adaptive notch filtering are simultaneously triggered. Subsequently, the system enters the second line of defense, using heterogeneous data to calculate the Mahalanobis distance. Logical verification is performed, and an alarm is output only when the spatiotemporal logic is consistent and the features meet the threshold. The clear decision branches and loop logic in the diagram ensure the system's high robustness in adversarial environments.

[0066] In summary, the aforementioned detection method constructs a real-time frequency domain monitoring window based on sliding window Fast Fourier Transform (FFT) and utilizes a biomechanical mask to identify and extract the "frequency domain fingerprint" of unnatural motion features, thereby establishing an illegal signal filtering mechanism at the physical perception level. This invention employs an adaptive narrowband notch filter algorithm to accurately eliminate high-frequency physical injection interference without damaging the real motion feature signals (first frequency range), resolving the contradiction between detection sensitivity and anti-interference capability. Simultaneously, the system introduces a cross-modal spatiotemporal consistency verification mechanism, using Kalman filtering to fuse multi-source heterogeneous data such as acceleration, air pressure, and geomagnetism, and quantitatively evaluating the consistency of each sensor's observations in the physical logic evolution by calculating the Mahalanobis distance. This mechanism effectively suppresses hijacking or blinding attacks targeting a single sensor through "cross-verification" of multi-dimensional data. While ensuring high real-time performance in fall detection, it achieves a coordinated balance between the high dynamic recognition capability and robust security performance of the perception system in adversarial physical environments.

[0067] The method described in this embodiment provides tangible economic and social value for intelligent mobility assistance: Smart mobility assistance safety for an aging society: Directly integrated with smart elderly care scenarios, it provides "system-level" safety protection for mobility aids such as smart canes used by the elderly. By effectively resisting environmental noise and potential malicious physical interference, it provides more reliable monitoring support for the elderly's home and outdoor activities, resulting in significant social safety benefits.

[0068] Reducing the ineffective use of medical resources and public services: By reducing the false alarm rate caused by physical injection from over 70% to 3%, this technology can significantly reduce ineffective emergency response and waste of medical emergency resources caused by "false fall alarms." For nursing homes and community monitoring centers, this significantly reduces management costs and improves the overall operational efficiency of emergency response systems.

[0069] Promoting the engineering application of high-security smart wearable devices: The "frequency domain cleaning + cross-modal verification" architecture proposed in this invention is not only applicable to fall detection, but can also be extended to fields with extremely high requirements for the integrity of sensing data, such as industrial robots and medical exoskeletons. This technological advancement fundamentally solves the physical vulnerability problem of MEMS sensors, providing key technical support for the large-scale application of next-generation high-security smart hardware.

[0070] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0076] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A highly dynamic and robust fall detection method resistant to physical injection, characterized in that, Includes the following steps: S1. Collect multi-source observation data from the intelligent walking assistance device and construct an initial state vector. The multi-source observation data includes real-time acceleration characteristic signals collected using an accelerometer. ; S2, Based on the real-time acceleration characteristic signal If a physical injection attack occurs, return to step S1 to collect data again. If it does, use a narrowband notch filter algorithm based on human dynamics mask to filter the signal and obtain a clean low-frequency human motion feature signal. S3. When the cleaned low-frequency human motion feature signal triggers the fall threshold, the fall detection judgment is performed based on the cleaned low-frequency human motion feature signal, the remaining multi-source observation data, and the state vector using a heterogeneous data fusion method based on Kalman filtering, and the detection result is output.

2. The highly dynamic robust fall detection method resistant to physical injection as described in claim 1, characterized in that, The multi-source observation data also includes real-time altitude characteristic signals collected by barometers, real-time attitude characteristic signals collected by magnetometers, and real-time angular velocity characteristic signals collected by gyroscopes.

3. The highly dynamic robust fall detection method resistant to physical injection according to claim 1, characterized in that, The process of determining whether a physical injection attack has occurred includes: Within a microsecond-level frequency domain monitoring window, the real-time acceleration characteristic signal is analyzed using a sliding window fast Fourier transform method. Convert to frequency domain energy distribution , is represented as: ， In the formula, For the Hanning window function, The imaginary unit, For frequency variables, for time, This represents the time shift of the sliding window; The normal human kinematic frequency range is set as the first frequency range, and a low-frequency protection mask is applied to the first frequency range. The frequency range in which the sensor physically resonates is set as the second frequency range, wherein the second frequency range is greater than the first frequency range. For the frequency domain energy distribution Calculate the power spectral density; The presence of an abnormally high narrowband resonant peak is detected based on the power spectral density within the second frequency range. If it is, it indicates that a physical injection attack has occurred; otherwise, it indicates that no physical injection attack has occurred.

4. The highly dynamic robust fall detection method against physical injection according to claim 3, characterized in that, The detection steps for the abnormally abruptly increased narrowband resonance peak include: Set a real-time updated environmental background noise threshold; When the power spectral density amplitude at a certain frequency point is detected to exceed the environmental background noise threshold within the second frequency range, and the frequency domain energy distribution... If the bandwidth is extremely narrow, it is determined that an abnormally high narrowband resonance peak has been detected; otherwise, it has not been detected.

5. The highly dynamic robust fall detection method resistant to physical injection according to claim 1, characterized in that, The real-time acceleration characteristic signal The data was acquired at the first sampling frequency. In the event of a physical injection attack, the sampling frequency is switched to the second sampling frequency to acquire real-time acceleration characteristic signals. Using the real-time acceleration characteristic signal at the second sampling frequency Perform filtering.

6. The highly dynamic robust fall detection method against physical injection according to claim 1, characterized in that, The narrowband notch filtering algorithm based on human dynamics masking employs a transfer function. Filtering is performed, the transfer function Represented as: ， In the formula, For the Laplace operator, The center angular frequency, For the attack frequency identified in real time, This is the bandwidth of the notch filter.

7. The highly dynamic robust fall detection method against physical injection according to claim 1, characterized in that, The execution steps of the heterogeneous data fusion method based on Kalman filtering include: definition The state vector at time t is ,in, For the height of intelligent walking assistance devices, The vertical velocity, This refers to the physical attitude tilt angle; definition The control input vector at time t is ,in, This refers to the clean, low-frequency human motion characteristic signals after cleaning. This is a real-time angular velocity characteristic signal; The time update matrix equation for constructing the prediction model is expressed as: ， ， ， In the formula, for The prior state prediction vector at time t. Here is the system state transition matrix. for The optimal posterior state estimate vector at time t. To control the input matrix, for The control input vector at time t, To control the periodic sampling time interval; Within each control cycle, the optimal posterior state estimation vector from the previous time step is... and the control input vector of the current control cycle Substituting the time update matrix equation into the equation and performing kinematic derivation, we obtain the prior state prediction vector at the current moment. ; The prior error covariance matrix is ​​derived recursively, and is expressed as: ， In the formula, Let be the prior error covariance matrix. for The posterior error covariance matrix at time t. The process noise covariance matrix; Construct the actual observation vector based on the multi-source observation data. Using the observation matrix Predict the prior state vector Values ​​are mapped to the observation space to obtain the predicted observation vector. ; Based on the actual observation vector With predicted observation vector Perform residual comparison to obtain the measurement residual vector. ; Based on the prior error covariance matrix The observation noise covariance matrix of the multi-source observation data is combined with Calculate the residual covariance matrix , is represented as: ， Using the prior error covariance matrix With residual covariance matrix Calculate the Kalman gain matrix for the current control cycle in real time. , is represented as: ， Using the Kalman gain matrix For the prior state prediction vector Make corrections and output the current result. The optimal posterior state estimation vector at time t And update the posterior error covariance matrix simultaneously. For use in the next control cycle, where the optimal posterior state estimation vector and the updated verification error covariance matrix They are represented as follows: ， ， In the formula, It is an identity matrix.

8. The highly dynamic robust fall detection method against physical injection according to claim 7, characterized in that, The fall detection is performed by calculating the Mahalanobis distance, wherein the expression for calculating the Mahalanobis distance is: ， In the formula, This is the Mahalanobis distance.

9. The highly dynamic robust fall detection method against physical injection according to claim 8, characterized in that, The steps for outputting the detection results include: Determine the Mahalanobis distance Does it meet the requirements? If yes, a physical deception attack is considered to exist, triggering a false alarm interception command, recording an anomaly log, and issuing a defense prompt. If no, a physical deception attack is considered not to exist, and it is further determined whether the cleaned, purified low-frequency human motion characteristic signal exceeds the fall detection threshold. If yes, it is confirmed as a genuine fall, triggering a rescue alarm. If no, it is considered that there is no genuine fall, and no alarm is triggered. This is the preset robust confidence threshold.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements a highly dynamic and robust fall detection method resistant to physical injection as described in any one of claims 1-9.

Images