Real-time freezing gait detection system based on inertial sensor and multi-perception feedback

Abstract

The invention discloses a real-time frozen gait detection system based on an inertial sensor and multi-perception feedback, and particularly relates to the field of gait detection, which comprises an acquisition module, a signal preprocessing module, a feature construction module, a classification module and a judgment module, the acquisition module is used for acquiring an acceleration signal and an angular velocity signal under a preset sampling frequency through a first inertial measurement device fixed at a waist position and second inertial measurement devices respectively fixed at two feet positions; and combining the acceleration signal and the angular velocity signal acquired at the waist position and the acceleration signal and the angular velocity signal acquired at the two foot positions to construct a gait data sequence. Inertial sensors are arranged at the waist and the two feet to collect multi-source gait signals, filtering processing and time domain / frequency domain feature construction and classification judgment are combined, multi-mode feedback of touch, vision and hearing is output when a freezing event is triggered, and real-time detection and effective intervention on the freezing gait are achieved.

Description

Technical Field

[0001] This invention relates to the field of gait detection technology, and more specifically, to a real-time frozen gait detection system based on inertial sensors and multi-sensory feedback. Background Technology

[0002] In existing gait-assisted studies of Parkinson's disease, frozen gait (FOG) has always been one of the most challenging symptoms; patients often experience sudden cessation of gait or extremely reduced stride length during walking, which not only severely affects their mobility but also significantly increases the risk of falls.

[0003] To address this issue, existing technologies primarily rely on single sensing methods for detection, such as plantar pressure sensors or video monitoring. However, these methods suffer from three main drawbacks: First, a single sensor captures limited gait features, and algorithms are prone to misjudgment in complex environments, leading to detection delays. Second, common feedback modes are limited to sound or vibration cues, lacking multi-sensory intervention, resulting in slow patient responses to cues and limited intervention effectiveness. Third, many existing devices require fixed scenarios or cumbersome wearing methods to operate, lacking convenience for daily use and severely limiting their practicality.

[0004] Based on the above situation, the existing technologies can be summarized as follows: the existing technologies have not yet formed a solution that takes into account real-time detection, effective intervention and portable application, and therefore there are still obvious deficiencies in the recognition and handling of frozen gait in real life scenarios. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a real-time frozen gait detection system based on inertial sensors and multi-sensory feedback. By deploying inertial sensors at the waist and feet to collect multi-source gait signals, and combining filtering, time-domain / frequency-domain feature construction and classification, the system outputs multi-modal feedback of touch, vision and hearing when a frozen event is triggered, thereby achieving real-time detection and effective intervention of frozen gait and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a real-time frozen gait detection system based on inertial sensors and multi-sensory feedback, comprising an acquisition module, a signal preprocessing module, a feature construction module, a classification module, and a determination module;

[0007] The acquisition module acquires acceleration and angular velocity signals at a preset sampling frequency using a first inertial measurement device fixed at the waist and a second inertial measurement device fixed at both feet. The acceleration and angular velocity signals acquired at the waist and the acceleration and angular velocity signals acquired at both feet are combined to construct a gait data sequence.

[0008] The signal preprocessing module is used to receive gait data sequences, perform bandpass filtering on acceleration signals to remove gravity components and noise, perform low-pass filtering on angular velocity signals to suppress drift, and construct a sliding window according to preset window duration, step and overlap parameters to divide the filtering results into windowed signal segments.

[0009] The feature construction module is used to receive the windowed signal segment, perform gait event recognition and feature calculation on the windowed signal segment, generate a feature vector, and output the feature vector to the classification module;

[0010] The classification module is used to input the feature vector into the support vector machine classifier, solve the frozen gait label and the unfrozen gait label in the classification operation, and form the corresponding classification result for subsequent judgment;

[0011] The determination module is used to receive the classification results and construct a discrimination sequence, perform continuous window statistics on the discrimination sequence, and when three consecutive windows in the discrimination sequence are identified as frozen gait, a frozen event signal is generated and output to the intervention feedback module.

[0012] In a preferred embodiment, the execution of the intervention feedback module includes:

[0013] Upon receiving a freezing event signal, the vibration feedback device is controlled to output a vibration signal within a fixed frequency range and maintain the vibration signal for a preset time period to form tactile stimulation of the dorsum of the foot; the vibration feedback device is integrated inside a second inertial measurement device for the position of both feet.

[0014] Upon receiving a freeze event signal, the laser guidance device fixed in front of the waist is controlled to project a straight optical guidance signal in the area in front of the user's toes, and a visual reference line is formed by setting the position, distance and angle of the light spot to guide the stepping action.

[0015] Upon receiving a freeze event signal, the sound output module of the control terminal generates a sound prompt signal with a fixed frequency and volume, enabling the user to unfreeze the state through auditory stimulation.

[0016] During the execution of vibration feedback, laser guidance, and sound cues, joint feedback is generated simultaneously, and the parameters of the joint feedback are statistically analyzed and recorded.

[0017] The parameters of the joint feedback are established as a joint mapping relationship between the freeze event trigger time, vibration frequency, laser projection distance, and sound frequency.

[0018] In a preferred embodiment, the interaction management module is used to guide the user to complete a ten-step walk during the first run, construct baseline gait parameters based on the data collected from the ten-step walk, and statistically analyze the time and duration of the frozen events during subsequent runs, and record the corresponding feature vectors and intervention signals, and finally generate a frozen event log and output rehabilitation progress information.

[0019] In a preferred embodiment, the signal preprocessing module performs the following operations after receiving the gait data sequence:

[0020] Bandpass filtering is performed on the acceleration signal in the gait data sequence: the acceleration signal is represented as a time series a(t), processed by a digital Butterworth bandpass filter, retaining frequency components within the range of a lower limit frequency of 0.5Hz and an upper limit frequency of 15Hz, removing DC components and gravity terms below 0.5Hz, and attenuating random noise above 15Hz, thereby generating the filtered acceleration signal.

[0021] A low-pass filtering operation is performed on the angular velocity signal in the gait data sequence: the angular velocity signal is represented as a time series ω(t), processed by a low-pass filter, retaining frequency components below the cutoff frequency of 10Hz, and attenuating fast fluctuation components above 10Hz to suppress high-frequency drift and sensor noise, thereby generating a filtered angular velocity signal.

[0022] The filtered acceleration signal With the filtered angular velocity signal Perform time synchronization and align sampling timestamps to form a unified data pair. And construct a continuous set of processed signals, the set of processed signals being denoted as... Where t i Indicates the sampling time; i represents the sampling point number; N represents the total number of sampling points;

[0023] Construct a sliding window sequence on the processed signal set: set the window duration parameter T. w =4s, step parameter T s =0.5s, overlap parameter T0=2s, and a length of T is extracted from the starting time t0 on the time axis. w The signal segment, and after each segment, the start time is advanced forward by T. s This ensures that adjacent windows maintain a time overlap of T0 until the entire sampling interval is covered;

[0024] The processed signal set is segmented according to the sliding window sequence to form windowed signal segments one by one. Where t k The starting time point of the k-th window is used as the final output of the windowed signal segment to the feature construction module as input.

[0025] In a preferred embodiment, the execution of the feature construction module after receiving the windowed signal fragment includes:

[0026] Gait events at the left and right foot positions are identified in the windowed signal segment, a stride length sequence is obtained, and the mean stride length μ is calculated. SL with step size standard deviation σ SL The step size variation coefficient (SLVC) is formed, and the step size variation coefficient (SLVC) is defined as follows:

[0027]

[0028] Identify the left and right foot swing phases in the windowed signal segment and solve for the proportion of the left foot swing phase SW. left The proportion of the right foot swing to SW right And calculate the gait asymmetry index (GAI), which is defined as follows:

[0029]

[0030] In the windowed signal segment, a frequency domain transformation is performed on the acceleration signal. The energy of the frozen characteristic frequency band (3–8 Hz) and the energy of the total gait frequency band (0.5–15 Hz) are statistically analyzed. The dominant frequency band energy ratio (FBER) is calculated, and the definition of the dominant frequency band energy ratio (FBER) is as follows:

[0031]

[0032] Where P(f) represents the power spectrum of the acceleration signal at frequency f;

[0033] The step size variation coefficient (SLVC), gait asymmetry index (GAI), and main band energy ratio (FBER) are combined to form a feature vector, and the feature vector is output to the classification module.

[0034] In a preferred embodiment, after receiving the current window feature vector output by the feature construction module, the classification module sequentially performs the following steps to extract the frozen state label and the unfrozen state label:

[0035] The feature vector of the current window is compared with the baseline gait parameters output by the interaction management module dimension by dimension. The difference between the feature value of each dimension and the baseline mean is calculated and the difference is standardized. When the difference of any dimension exceeds the preset upper and lower limits, the feature vector of the window is marked as needing to be realigned and the standardization process is returned until all the differences of all dimensions are within the target range before proceeding to the next operation.

[0036] In the feature vector after standardization, class weights are constructed based on the distribution of data labeled as frozen and data labeled as unfrozen during the training phase. When the proportion of the two types of labeled data does not meet the preset threshold range, the weights are corrected. The corrected feature vector and the baseline difference result are input into the support vector machine classifier, so that the classifier can refer to both the individual baseline difference and the class distribution difference when making a judgment.

[0037] When the support vector machine (SVM) classifier performs a discrimination operation on the input feature vector: first, it calculates the distance of the feature vector relative to the separating boundary established by the SVM classifier and records the positive and negative directions of the distance; when the distance is positive and falls into a preset frozen region, the feature vector is labeled as frozen; when the distance is negative and falls into a preset unfrozen region, the feature vector is labeled as unfrozen; at the same time, the absolute value of the distance is used as the discrimination strength output to characterize the confidence level of the solved label.

[0038] The label of the current window obtained by the support vector machine classifier is compared with the labels of adjacent windows in the time buffer obtained by the support vector machine classifier. The percentage of times the labels of adjacent windows match the current window is counted and used as the consistency ratio. When the consistency ratio reaches a preset lower limit and the discrimination strength of the current window reaches a preset lower limit, the label of the current window is directly confirmed as the output result. When the consistency ratio is insufficient or the discrimination strength of the current window is insufficient, backtracking logic is triggered. The backtracking logic includes inputting the feature vector of the current window and the feature vector of the adjacent windows into the support vector machine classifier to re-execute the discrimination operation, and replacing the original label of the current window with the label obtained by the new discrimination operation.

[0039] After completing the cross-match or backtracking discrimination, a dual threshold decision is made on the final discrimination strength and consistency ratio: when the discrimination strength reaches the preset upper limit and the consistency ratio reaches the preset upper limit, the final label and corresponding confidence level are output and the result is passed to the judgment module; when the discrimination strength and consistency ratio do not meet the conditions at the same time, the result is kept as undetermined and the cross-match and backtracking discrimination process is re-entered after the next window of data arrives.

[0040] In a preferred embodiment, the execution of the determination module after receiving the classification result output by the classification module includes:

[0041] The classification results of the current window are stored into the discrimination sequence in chronological order, and the discrimination sequence is kept updated so that the latest classification result is always appended to the end of the discrimination sequence;

[0042] The classification results of consecutive windows are statistically analyzed in the discriminant sequence to detect whether the labels between adjacent windows are consistent, and the statistical results of consecutive windows are generated.

[0043] When the continuous window statistics result shows that the three most recent windows are all identified as frozen, a freeze event signal is generated and the freeze event signal is output to the intervention feedback module;

[0044] When the statistical results of the continuous window do not meet the conditions for the frozen state, the discrimination sequence is kept in the update state and waits for the classification results of the subsequent window to enter the next round of statistics.

[0045] The technical effects and advantages of this invention are as follows:

[0046] 1. By fusing gait data collected through multi-position inertial measurement devices, and forming a real-time recognition closed loop through preprocessing, feature extraction, and classification, anomalies can be captured in a timely manner at the early stage of frozen gait, fundamentally solving the problems of detection delay and high false judgment rate of existing methods;

[0047] 2. By employing bandpass filtering, lowpass filtering, and sliding window segmentation in the signal preprocessing stage, gravity components and high-frequency drift are effectively removed, ensuring the stability of the time-series signal. This provides high signal-to-noise ratio data input for subsequent feature calculations, thereby improving detection robustness.

[0048] 3. The step size variation coefficient, gait asymmetry index and main frequency band energy ratio are extracted in the time domain, phase domain and frequency domain through the feature construction module, and combined into a unified feature vector, so that the discrimination basis of frozen gait is comprehensive and the discrimination reliability of the classifier is enhanced.

[0049] 4. By performing support vector machine discrimination under the dual reference of individualized baseline parameters and training data distribution through the classification module, and combining window consistency statistics and backtracking mechanism within the time buffer, instantaneous misjudgment is suppressed, and the stability and reliability of label output are improved.

[0050] 5. The intervention feedback module simultaneously outputs vibration, laser, and sound signals after a freezing event is triggered, realizing multi-sensory feedback of touch, vision, and hearing, effectively shortening the patient's reaction time to unfreeze, improving the intervention effect, and enhancing the practicality of daily applications. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the system modules of the present invention.

[0052] Figure 2 This is a schematic diagram of the system hardware layout of the present invention (with the positions of the sensors and laser modules and the laser projection range marked).

[0053] Figure 3 This is an example diagram illustrating the laser guidance effect of the present invention (showing the projection position of the laser line in front of the toes in a frozen state). Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Refer to the instruction manual appendix Figure 1-3 An embodiment of the present invention provides a real-time frozen gait detection system based on inertial sensors and multi-sensory feedback, comprising an acquisition module, a signal preprocessing module, a feature construction module, a classification module, and a determination module.

[0056] The acquisition module acquires acceleration and angular velocity signals at a preset sampling frequency using a first inertial measurement device fixed at the waist and a second inertial measurement device fixed at both feet. The acceleration and angular velocity signals acquired at the waist and the acceleration and angular velocity signals acquired at both feet are combined to construct a gait data sequence.

[0057] The signal preprocessing module is used to receive gait data sequences, perform bandpass filtering on acceleration signals to remove gravity components and noise, perform low-pass filtering on angular velocity signals to suppress drift, and construct a sliding window according to preset window duration, step and overlap parameters to divide the filtering results into windowed signal segments.

[0058] The feature construction module is used to receive the windowed signal segment, perform gait event recognition and feature calculation on the windowed signal segment, generate a feature vector, and output the feature vector to the classification module;

[0059] The classification module is used to input the feature vector into the support vector machine classifier, solve the frozen gait label and the unfrozen gait label in the classification operation, and form the corresponding classification result for subsequent judgment;

[0060] The determination module is used to receive the classification results and construct a discrimination sequence, perform continuous window statistics on the discrimination sequence, and when three consecutive windows in the discrimination sequence are identified as frozen gait, a frozen event signal is generated and output to the intervention feedback module.

[0061] The execution of the intervention feedback module includes:

[0062] Upon receiving a freezing event signal, the vibration feedback device is controlled to output a vibration signal within a fixed frequency range and maintain the vibration signal for a preset time period to form tactile stimulation of the dorsum of the foot; the vibration feedback device is integrated inside a second inertial measurement device for the position of both feet.

[0063] Upon receiving a freeze event signal, the laser guidance device fixed in front of the waist is controlled to project a straight optical guidance signal in the area in front of the user's toes, and a visual reference line is formed by setting the position, distance and angle of the light spot to guide the stepping action.

[0064] The vibration feedback device is a structural component integrated inside the inertial measurement unit at the position of both feet. When frozen gait is detected, it can output vibration signals within a preset frequency range to generate tactile stimulation on the instep of the user.

[0065] The laser guidance device is a transmitting module fixed in front of the waist. It can project a straight optical guidance signal in the area in front of the user's toes, and help the user cross the light spot by forming a visible reference line, thereby unfreezing the gait.

[0066] Upon receiving a freeze event signal, the sound output module of the control terminal generates a sound prompt signal with a fixed frequency and volume, enabling the user to unfreeze the state through auditory stimulation.

[0067] During the execution of vibration feedback, laser guidance, and sound cues, joint feedback is generated simultaneously, and the parameters of the joint feedback are statistically analyzed and recorded.

[0068] The parameters of the joint feedback are established as a joint mapping relationship between the freeze event trigger time, vibration frequency, laser projection distance, and sound frequency.

[0069] The interactive management module is used to guide the user to complete ten normal steps during the first run, construct baseline gait parameters based on the data collected from the ten normal steps, and statistically analyze the time and duration of the frozen events during subsequent runs, and record the corresponding feature vectors and intervention signals, and finally generate a frozen event log and output rehabilitation progress information.

[0070] The calculation of baseline gait parameters includes solving the mean and standard deviation of step length from data collected by ten normal walking steps to form a step length stability benchmark, solving the gait symmetry benchmark from the time ratio of the left and right foot swing phases, and calculating the frequency domain energy ratio benchmark from the energy distribution of acceleration signals in the frozen characteristic frequency band and the total gait frequency band, thereby constructing individualized baseline gait parameters.

[0071] The signal preprocessing module performs the following operations after receiving the gait data sequence:

[0072] Bandpass filtering is performed on the acceleration signal in the gait data sequence: the acceleration signal is represented as a time series a(t), processed by a digital Butterworth bandpass filter, retaining frequency components within the range of a lower limit frequency of 0.5Hz and an upper limit frequency of 15Hz, removing DC components and gravity terms below 0.5Hz, and attenuating random noise above 15Hz, thereby generating the filtered acceleration signal.

[0073] A low-pass filtering operation is performed on the angular velocity signal in the gait data sequence: the angular velocity signal is represented as a time series ω(t), processed by a low-pass filter, retaining frequency components below the cutoff frequency of 10Hz, and attenuating fast fluctuation components above 10Hz to suppress high-frequency drift and sensor noise, thereby generating a filtered angular velocity signal.

[0074] The filtered acceleration signal With the filtered angular velocity signal Perform time synchronization and align sampling timestamps to form a unified data pair. And construct a continuous set of processed signals, the set of processed signals being denoted as... Where t i Indicates the sampling time; i represents the sampling point number; N represents the total number of sampling points;

[0075] Construct a sliding window sequence on the processed signal set: set the window duration parameter T. w =4s, step parameter T s =0.5s, overlap parameter T0=2s, and a length of T is extracted from the starting time t0 on the time axis. w The signal segment, and after each segment, the start time is advanced forward by T. s This ensures that adjacent windows maintain a time overlap of T0 until the entire sampling interval is covered;

[0076] The processed signal set is segmented according to the sliding window sequence to form windowed signal segments one by one. Where t kThe starting time point of the k-th window is used as the final output of the windowed signal segment to the feature construction module as input.

[0077] The execution of the feature construction module after receiving the windowed signal fragment includes:

[0078] Gait events at the left and right foot positions are identified in the windowed signal segment, a stride length sequence is obtained, and the mean stride length μ is calculated. SL with step size standard deviation σ SL The step size variation coefficient (SLVC) is formed, and the step size variation coefficient (SLVC) is defined as follows:

[0079]

[0080] Identify the left and right foot swing phases in the windowed signal segment and solve for the proportion of the left foot swing phase SW. left The proportion of the right foot swing to SW right And calculate the gait asymmetry index (GAI), which is defined as follows:

[0081]

[0082] In the windowed signal segment, a frequency domain transformation is performed on the acceleration signal. The energy of the frozen characteristic frequency band (3–8 Hz) and the energy of the total gait frequency band (0.5–15 Hz) are statistically analyzed. The dominant frequency band energy ratio (FBER) is calculated, and the definition of the dominant frequency band energy ratio (FBER) is as follows:

[0083]

[0084] Where P(f) represents the power spectrum of the acceleration signal at frequency f;

[0085] The step size variation coefficient (SLVC), gait asymmetry index (GAI), and main band energy ratio (FBER) are combined to form a feature vector, and the feature vector is output to the classification module.

[0086] After receiving the current window feature vector output by the feature construction module, the classification module sequentially performs the following steps to extract the frozen state label and the unfrozen state label:

[0087] The feature vector of the current window is compared with the baseline gait parameters output by the interaction management module dimension by dimension. The difference between the feature value of each dimension and the baseline mean is calculated and the difference is standardized. When the difference of any dimension exceeds the preset upper and lower limits, the feature vector of the window is marked as needing to be realigned and the standardization process is returned until all the differences of all dimensions are within the target range before proceeding to the next operation.

[0088] In the feature vector after standardization, class weights are constructed based on the distribution of data labeled as frozen and data labeled as unfrozen during the training phase. When the proportion of the two types of labeled data does not meet the preset threshold range, the weights are corrected. The corrected feature vector and the baseline difference result are input into the support vector machine classifier, so that the classifier can refer to both the individual baseline difference and the class distribution difference when making a judgment.

[0089] When the support vector machine (SVM) classifier performs a discrimination operation on the input feature vector: first, it calculates the distance of the feature vector relative to the separating boundary established by the SVM classifier and records the positive and negative directions of the distance; when the distance is positive and falls into a preset frozen region, the feature vector is labeled as frozen; when the distance is negative and falls into a preset unfrozen region, the feature vector is labeled as unfrozen; at the same time, the absolute value of the distance is used as the discrimination strength output to characterize the confidence level of the solved label.

[0090] The label of the current window obtained by the Support Vector Machine (SVM) classifier is compared with the labels of adjacent windows in the time buffer obtained by the SVM classifier. The percentage of times the labels of adjacent windows match the current window is counted and used as the consistency ratio. When the consistency ratio reaches a preset lower limit and the discrimination strength of the current window reaches a preset lower limit, the label of the current window is directly confirmed as the output result. When the consistency ratio is insufficient or the discrimination strength of the current window is insufficient, backtracking logic is triggered. The backtracking logic includes inputting the feature vector of the current window and the feature vector of the adjacent windows into the SVM classifier to re-execute the discrimination operation, and replacing the original label of the current window with the label obtained by the new discrimination operation. In addition, the time buffer here refers to the cache area set up during the classification process to store the discrimination results of multiple consecutive adjacent time windows, which is used to call the results of the previous and next windows for consistency comparison when the current window is discriminated.

[0091] After completing the cross-match or backtracking discrimination, a dual threshold decision is made on the final discrimination strength and consistency ratio: when the discrimination strength reaches the preset upper limit and the consistency ratio reaches the preset upper limit, the final label and corresponding confidence level are output and the result is passed to the judgment module; when the discrimination strength and consistency ratio do not meet the conditions at the same time, the result is kept as undetermined and the cross-match and backtracking discrimination process is re-entered after the next window of data arrives.

[0092] The execution of the determination module after receiving the classification result output by the classification module includes:

[0093] The classification results of the current window are stored sequentially into the discrimination sequence in chronological order, and the discrimination sequence is kept updated so that the latest classification result is always appended to the end of the discrimination sequence. The discrimination sequence is a set of classification results arranged in chronological order, used to record the frozen or unfrozen state labels of multiple consecutive windows. Keeping the discrimination sequence updated means that whenever a new window classification result is generated, it is appended to the end of the sequence, and the oldest result is deleted at the same time, so that the sequence always retains only the consecutive discrimination results within the most recent time range.

[0094] The classification results of consecutive windows are statistically analyzed in the discriminant sequence to detect whether the labels between adjacent windows are consistent, and the statistical results of consecutive windows are generated.

[0095] When the continuous window statistics result shows that the three most recent windows are all identified as frozen, a freeze event signal is generated and the freeze event signal is output to the intervention feedback module;

[0096] When the statistical results of the continuous window do not meet the conditions for the frozen state, the discrimination sequence is kept in the update state and waits for the classification results of the subsequent window to enter the next round of statistics.

[0097] The actual implementation of this plan includes:

[0098] 1. Hardware components:

[0099] The sensor module includes:

[0100] Waist sensor (attached to the back of the waist): Built-in three-axis accelerometer and gyroscope, used to monitor the torso tilt angle and overall motion trajectory, with a sampling rate greater than 100Hz.

[0101] Foot sensors (attached to the instep): Each foot is equipped with a six-axis inertial measurement unit (IMU) to collect foot acceleration, angular velocity and vibration frequency data, with a sampling rate greater than 100Hz.

[0102] Vibration feedback module: Integrated inside the sensor, it triggers high-frequency vibration (frequency 10-50Hz) when gait freeze is detected.

[0103] Laser guidance module (attached to the front of the waist): emits a red linear laser (wavelength 650nm, power <5mW, compliant with IEC60825-1 safety standards), and when triggered, projects a straight light spot 20cm in front of the toes to guide the patient across the "virtual obstacle".

[0104] Control terminal (Pad / Mobile phone): Built-in Bluetooth 5.0 module, receives sensor data and runs detection algorithms to synchronously control laser and vibration feedback.

[0105] 2. Software Algorithm

[0106] The following is a design scheme for a gait freeze (FOG) detection algorithm based on waist and bipedal IMUs (Inertial Measurement Units), which combines signal processing, feature extraction and machine learning classification techniques to meet the requirements of real-time performance and accuracy.

[0107] Algorithm flow:

[0108] Step 1: Signal preprocessing:

[0109] 1-1 Filtering and Noise Reduction:

[0110] Acceleration signal: Butterworth bandpass filter (0.5–15Hz) to eliminate high-frequency noise and gravitational interference;

[0111] Gyroscope signal: low-pass filter (cutoff frequency 10Hz) to suppress high-frequency drift;

[0112] 1-2 Data Segmentation:

[0113] Sliding window: 4-second window (2-second overlap), samples updated every 0.5 seconds;

[0114] Step 2: Spatiotemporal Feature Extraction:

[0115] 2-1 Extracting multidimensional features from sensors, with a focus on gait rhythm anomalies:

[0116] Temporal characteristics:

[0117] A. Step size variation coefficient (change in distance between the two IMU feet);

[0118]

[0119] B. Gait asymmetry (time difference in the swing phase of the left and right feet);

[0120]

[0121] Frequency domain characteristics:

[0122] A. Main frequency band energy ratio (FFT or wavelet transform): Typical high-frequency jitter during the FOG period (3–8Hz);

[0123]

[0124] Where P(f) represents the power spectrum of the acceleration signal at frequency f, i.e., the acceleration power spectrum (wavelet transform coefficients); the numerator of FBER represents the frozen characteristic frequency band (3-8Hz), and the denominator of FBER represents the total gait frequency band (0.5-15Hz).

[0125] Step 3: Real-time classification model; Considering the requirements of recognition rate and real-time performance, SVM is selected as the classifier here;

[0126] Step 4: Post-processing and decision-making;

[0127] Event confirmation: If three consecutive windows are categorized as FOG, a freeze event is triggered (to reduce false alarms);

[0128] Output: Real-time alarms (laser / vibration / sound alerts) and freeze event logs (timestamp, duration, characteristic parameters);

[0129] 3. Human-computer interaction design

[0130] Adaptive calibration: When a user uses the Pad for the first time, the software guides them through 10 normal steps to automatically calibrate the baseline of individualized gait parameters.

[0131] Historical data visualization: Records the time, duration, and intervention effects of frozen events, and generates rehabilitation progress reports for doctors' reference.

[0132] It should be noted that the execution principle of this solution is as follows: When the interactive management module runs for the first time, it guides the user to complete ten normal steps and solves the mean step length and standard deviation of step length, the proportion of left and right swing phase time, and the energy ratio of frozen feature frequency band to total gait frequency band, thereby constructing individualized baseline gait parameters.

[0133] Subsequently, the acquisition module synchronously acquires acceleration and angular velocity signals at a preset sampling frequency through a first inertial measurement device fixed at the waist and a second inertial measurement device fixed at both feet, and synthesizes them into a gait data sequence. The signal preprocessing module performs bandpass filtering on the acceleration signal to remove gravity components and noise, performs low-pass filtering on the angular velocity signal to suppress drift, and constructs a sliding window according to preset window duration, step size and overlap parameters to form windowed signal segments.

[0134] The feature construction module completes gait event recognition and calculates the stride variation coefficient, gait asymmetry index based on the difference in the proportion of swing phases, and main frequency band energy ratio within each windowed signal segment. It then combines the above three types of features into a feature vector and outputs it.

[0135] After receiving the feature vector, the classification module first compares it with the baseline gait parameters dimension by dimension to complete individual alignment and standardization. Then, it sets the class weights by combining the data distribution of the corresponding frozen state and the corresponding unfrozen state that have been labeled during the training phase. The standardized feature vector and the class weights are then input into the support vector machine classifier to calculate the signed distance relative to the separating boundary. The sign of the distance is used to solve for the frozen state label or the unfrozen state label, and the absolute value of the distance is used as the discrimination strength.

[0136] The judgment module writes the classification results of consecutive windows into the discrimination sequence in chronological order and performs consecutive window statistics. When the three most recent windows are all identified as frozen, a freezing event signal is generated and output to the intervention feedback module. Otherwise, the discrimination sequence is updated and the module waits for subsequent windows to enter the next round of statistics.

[0137] Upon receiving a freezing event signal, the intervention feedback module controls the vibration feedback device integrated into the second inertial measurement device for both feet to output a fixed frequency vibration to generate tactile stimulation of the dorsum of the foot. At the same time, it controls the laser guidance device fixed in front of the waist to project a linear optical guidance signal in front of the toes and outputs a sound prompt with a fixed frequency and volume from the terminal. The three feedbacks work synchronously to promote the recovery of stepping movements. The parameters of the combined feedback are statistically analyzed and recorded to be included in the freezing event log and rehabilitation progress information generated by the interactive management module.

[0138] By executing the steps of “individualized baseline, filtering and windowing, time-domain / phase / frequency-domain features, weighted discrimination, sequence statistics, multi-sensor intervention and log closure” in sequence, this solution demonstrates real-time performance, discrimination stability and intervention effectiveness.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time frozen gait detection system based on inertial sensors and multi-sensor feedback, comprising an acquisition module, a signal preprocessing module, a feature construction module, a classification module, and a judgment module, characterized in that: The acquisition module acquires acceleration and angular velocity signals at a preset sampling frequency using a first inertial measurement device fixed at the waist and a second inertial measurement device fixed at both feet. The acceleration and angular velocity signals acquired at the waist and the acceleration and angular velocity signals acquired at both feet are combined to construct a gait data sequence. The signal preprocessing module is used to receive gait data sequences, perform bandpass filtering on acceleration signals to remove gravity components and noise, perform low-pass filtering on angular velocity signals to suppress drift, and construct a sliding window according to preset window duration, step and overlap parameters to divide the filtering results into windowed signal segments. The feature construction module is used to receive the windowed signal segment, perform gait event recognition and feature calculation on the windowed signal segment, generate a feature vector, and output the feature vector to the classification module; The classification module is used to input the feature vector into the support vector machine classifier, solve for the frozen gait label and the unfrozen gait label in the classification operation, and form the corresponding classification result for subsequent judgment; The determination module is used to receive the classification results and construct a discrimination sequence, perform continuous window statistics on the discrimination sequence, and when three consecutive windows in the discrimination sequence are identified as frozen gait, a frozen event signal is generated and output to the intervention feedback module.

2. The real-time frozen gait detection system based on inertial sensors and multi-sensory feedback according to claim 1 further includes an intervention feedback module, characterized in that: The execution of the intervention feedback module includes: Upon receiving a freezing event signal, the vibration feedback device is controlled to output a vibration signal within a fixed frequency range and maintain the vibration signal for a preset time period to form tactile stimulation of the dorsum of the foot; the vibration feedback device is integrated inside a second inertial measurement device for the position of both feet. Upon receiving a freeze event signal, the laser guidance device fixed in front of the waist is controlled to project a straight optical guidance signal in the area in front of the user's toes, and a visual reference line is formed by setting the position, distance and angle of the light spot to guide the stepping action. Upon receiving a freeze event signal, the sound output module of the control terminal generates a sound prompt signal with a fixed frequency and volume, enabling the user to unfreeze the state through auditory stimulation. During the execution of vibration feedback, laser guidance, and sound cues, joint feedback is generated simultaneously, and the parameters of the joint feedback are statistically analyzed and recorded. The parameters of the joint feedback are established as a joint mapping relationship between the freeze event trigger time, vibration frequency, laser projection distance, and sound frequency.

3. The real-time frozen gait detection system based on inertial sensors and multi-sensor feedback according to claim 2 further includes an interactive management module, characterized in that: The interactive management module is used to guide the user to complete a ten-step walk during the first run, construct baseline gait parameters based on the data collected from the ten-step walk, and statistically analyze the time and duration of the frozen events during subsequent runs, and record the corresponding feature vectors and intervention signals, and finally generate a frozen event log and output rehabilitation progress information.

4. The real-time frozen gait detection system based on inertial sensors and multi-sensor feedback according to claim 3, characterized in that: The signal preprocessing module performs the following operations after receiving the gait data sequence: Bandpass filtering is performed on the acceleration signal in the gait data sequence: the acceleration signal is represented as a time series a(t), processed by a digital Butterworth bandpass filter, the frequency components are retained in the range of the lower limit frequency 0.5Hz and the upper limit frequency 15Hz, the DC component and gravity term below 0.5Hz are removed, and the random noise above 15Hz is weakened, thereby generating the filtered acceleration signal a(t); A low-pass filtering operation is performed on the angular velocity signal in the gait data sequence: the angular velocity signal is represented as a time series ω(t), processed by a low-pass filter, retaining frequency components below the cutoff frequency of 10Hz, and attenuating fast fluctuation components above 10Hz to suppress high-frequency drift and sensor noise, thereby generating a filtered angular velocity signal. The filtered acceleration signal With the filtered angular velocity signal Perform time synchronization and align sampling timestamps to form a unified data pair. And construct a continuous set of processed signals, the set of processed signals being denoted as... Where t i Indicates the sampling time; i represents the sampling point number; N represents the total number of sampling points; Construct a sliding window sequence on the processed signal set: Set the window duration parameter T w =4s, step parameter T s =0.5s, overlap parameter T0=2s, and a length of T is extracted from the starting time t0 on the time axis. w The signal segment, and after each segment, the start time is advanced forward by T. s This ensures that adjacent windows maintain a time overlap of T0 until the entire sampling interval is covered; The processed signal set is segmented according to the sliding window sequence to form windowed signal segments one by one. Where t k The starting time point of the k-th window is used as the final output of the windowed signal segment to the feature construction module as input.

5. A real-time frozen gait detection system based on inertial sensors and multi-sensor feedback according to claim 4, characterized in that: The execution of the feature construction module after receiving the windowed signal fragment includes: Gait events at the left and right foot positions are identified in the windowed signal segment, a stride length sequence is obtained, and the mean stride length μ is calculated. SL with step size standard deviation σ SL The step size variation coefficient (SLVC) is formed, and the step size variation coefficient (SLVC) is defined as follows: Identify the left and right foot swing phases in the windowed signal segment and solve for the proportion of the left foot swing phase SW. left The proportion of the right foot swing to SW right And calculate the gait asymmetry index (GAI), which is defined as follows: In the windowed signal segment, a frequency domain transformation is performed on the acceleration signal. The energy of the frozen characteristic frequency band (3–8 Hz) and the energy of the total gait frequency band (0.5–15 Hz) are statistically analyzed. The dominant frequency band energy ratio (FBER) is calculated, and the definition of the dominant frequency band energy ratio (FBER) is as follows: Where P(f) represents the power spectrum of the acceleration signal at frequency f; The step size variation coefficient (SLVC), gait asymmetry index (GAI), and main band energy ratio (FBER) are combined to form a feature vector, and the feature vector is output to the classification module.

6. A real-time frozen gait detection system based on inertial sensors and multi-sensor feedback according to claim 5, characterized in that: After receiving the current window feature vector output by the feature construction module, the classification module sequentially performs the following steps to extract the frozen state label and the unfrozen state label: The feature vector of the current window is compared with the baseline gait parameters output by the interaction management module dimension by dimension. The difference between the feature value of each dimension and the baseline mean is calculated and the difference is standardized. When the difference of any dimension exceeds the preset upper and lower limits, the feature vector of the window is marked as needing to be realigned and the standardization process is returned until all the differences of all dimensions are within the target range before proceeding to the next operation. In the feature vector after standardization, class weights are constructed based on the distribution of data labeled as frozen and data labeled as unfrozen during the training phase. When the proportion of the two types of labeled data does not meet the preset threshold range, the weights are corrected. The corrected feature vector and the baseline difference result are input into the support vector machine classifier, so that the classifier can refer to both the individual baseline difference and the class distribution difference when making a judgment. When the support vector machine (SVM) classifier performs a discrimination operation on the input feature vector: first, it calculates the distance of the feature vector relative to the separating boundary established by the SVM classifier and records the positive and negative directions of the distance; when the distance is positive and falls into a preset frozen region, the feature vector is labeled as frozen; when the distance is negative and falls into a preset unfrozen region, the feature vector is labeled as unfrozen; at the same time, the absolute value of the distance is used as the discrimination strength output to characterize the confidence level of the solved label. The label of the current window obtained by the support vector machine classifier is compared with the labels of adjacent windows in the time buffer obtained by the support vector machine classifier. The percentage of times the labels of adjacent windows match the current window is counted and used as the consistency ratio. When the consistency ratio reaches a preset lower limit and the discrimination strength corresponding to the current window reaches a preset lower limit, the label of the current window is directly confirmed as the output result. When the consistency ratio is insufficient or the current window's discrimination strength is insufficient, backtracking logic is triggered. The backtracking logic includes inputting the feature vector of the current window and the feature vectors of the adjacent windows into the support vector machine classifier to re-execute the discrimination operation, and replacing the original label of the current window with the label solved by the discrimination operation. After completing cross-comparison or backtracking discrimination, a dual threshold decision is performed on the final discrimination strength and consistency ratio: when the discrimination strength reaches the preset upper limit and the consistency ratio reaches the preset upper limit, the final label and corresponding confidence score are output and the result is passed to the judgment module. If the discrimination strength and consistency ratio do not meet the conditions at the same time, the result remains undetermined, and the cross-comparison and backtracking discrimination process will be re-entered after the next window of data arrives.

7. A real-time frozen gait detection system based on inertial sensors and multi-sensor feedback according to claim 6, characterized in that: The execution of the determination module after receiving the classification result output by the classification module includes: The classification results of the current window are stored into the discrimination sequence in chronological order, and the discrimination sequence is kept updated so that the latest classification result is always appended to the end of the discrimination sequence; The classification results of consecutive windows are statistically analyzed in the discriminant sequence to detect whether the labels between adjacent windows are consistent, and the statistical results of consecutive windows are generated. When the continuous window statistics result shows that the three most recent windows are all identified as frozen, a freeze event signal is generated and the freeze event signal is output to the intervention feedback module; When the statistical results of the continuous window do not meet the conditions for the frozen state, the discrimination sequence is kept in the update state and waits for the classification results of the subsequent window to enter the next round of statistics.

Images