Method for monitoring the state of a driver of a vehicle based on the extraction of physiological signs

Abstract

The present application relates to the technical field of automobile driver state monitoring, and discloses a method for monitoring the state of an automobile driver based on physiological sign extraction. The method comprises continuously capturing the heart rate, breathing and posture movement signal streams of the driver; performing time-frequency analysis on the heart rate signal to obtain a heart rate variability quantification spectrum, decomposing and correcting the breathing signal to extract the breathing depth change envelope line, and performing trajectory reconstruction and frequency division on the posture signal to obtain a body movement characteristic frequency set; constructing a multi-dimensional physiological characteristic matrix based on the above characteristics, performing layer-by-layer matching and difference calculation on the matrix and a preset state template, generating a state deviation degree sequence and a physiological risk label; selecting an initial intervention instruction set according to the label, and dynamically calibrating the parameters of the instruction set by fusing the deviation degree sequence to generate a final intervention instruction set. The method realizes patterned and accurate recognition of the state of the driver and adaptive dynamic intervention, and improves the accuracy of monitoring and the effectiveness of intervention.

Description

Technical Field

[0001] This invention relates to the field of automobile driver condition monitoring technology, specifically to an automobile driver condition monitoring method based on physiological sign extraction. Background Technology

[0002] Currently, driver condition monitoring largely relies on threshold judgments of single or limited signal sources. For example, fatigue is determined by recognizing facial features in images, or vehicle behavior parameters such as steering wheel angle and lane departure are monitored. These methods typically collect and analyze physiological signals such as heart rate and respiration independently, comparing their values ​​only to preset, fixed safety thresholds. Corresponding intervention mechanisms are also mostly static responses, meaning that preset, uniform warning actions are executed when alarm conditions are triggered.

[0003] This judgment model, based on isolated signals and fixed thresholds, struggles to accurately reflect the complex state of a driver comprised of multiple physiological changes. It is susceptible to individual differences and transient interference, leading to frequent false alarms and missed alarms. Furthermore, rigid intervention strategies cannot differentiate responses based on the actual degree and trend of the deterioration, potentially failing due to insufficient warnings or causing interference due to overly aggressive warnings.

[0004] There is a need for a technical solution that can comprehensively utilize multi-dimensional physiological signs to conduct a holistic and patterned assessment of the driver's condition, and can dynamically adjust intervention strategies based on the continuous trend of condition changes, in order to overcome the shortcomings of existing methods in terms of the accuracy of condition identification and the adaptability of intervention measures. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring the state of a car driver based on the extraction of physiological signs, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for monitoring the state of a car driver based on physiological sign extraction, the method comprising:

[0007] Continuously capture the target driver's raw physiological information set, which includes heart rate signal stream, respiratory signal stream, and posture motion signal stream;

[0008] The heart rate signal stream is subjected to time-domain morphological separation and frequency-domain spectral analysis to obtain the baseline waveform of the heart rate rhythm and the quantized spectrum of heart rate variability.

[0009] Perform non-stationary signal decomposition and baseline drift elimination on the respiratory signal stream to extract the respiratory fundamental frequency waveform and the envelope of respiratory depth changes;

[0010] The attitude motion signal stream is subjected to three-dimensional spatial trajectory reconstruction and frequency component division to obtain the distribution spectrum of body kinetic energy on each axis in space and the characteristic frequency set of periodic motion;

[0011] A multidimensional physiological feature matrix is ​​constructed based on the quantized spectrum, the changing envelope, and the set of characteristic frequencies.

[0012] The multidimensional physiological feature matrix is ​​matched and the difference is calculated layer by layer with the preset driver state template to generate a state deviation sequence and multiple physiological risk markers.

[0013] Based on the physiological risk markers, select the corresponding initial intervention instruction set from the preset intervention strategy library;

[0014] The parameters of the initial intervention instruction set are dynamically calibrated by integrating the state deviation sequence to generate the final intervention instruction set.

[0015] Preferably, the step of performing time-domain morphological separation and frequency-domain spectral analysis on the heart rate signal stream to obtain the baseline waveform of the heart rate rhythm and the quantized spectrum of heart rate variability includes:

[0016] A multi-scale sliding window is applied to the heart rate signal stream to obtain a series of heart rate signal segments;

[0017] For each heart rate signal segment, local extreme points are detected to determine the time coordinates of the start point, peak point, and end point of each heartbeat cycle;

[0018] Based on the time coordinates, the waveforms of all heartbeat cycles are aligned, superimposed, and averaged to obtain the reference waveform of the heart rate rhythm.

[0019] The heart rate residual signal is obtained by subtracting the reference waveform from each of the heart rate signal segments;

[0020] The power spectral density of the heart rate residual signal is estimated, and its energy proportion in the ultra-low frequency, low frequency, and high frequency bands is calculated to form a quantitative spectrum of the heart rate variability.

[0021] Preferably, the step of performing non-stationary signal decomposition and baseline drift elimination on the respiratory signal stream, and extracting the change envelope of the respiratory fundamental frequency waveform and respiratory depth, includes:

[0022] Adaptive mode decomposition is applied to the respiratory signal stream, decomposing it into multiple intrinsic mode function components and a residual trend component;

[0023] Identify the target component in the intrinsic mode function components whose energy is concentrated near the respiratory rate, and reconstruct the target component into a pure respiratory signal;

[0024] Subtracting the residual trend component from the respiratory signal stream yields a respiratory correction signal with the slow-varying trend removed;

[0025] Calculate the instantaneous frequency of the pure respiratory signal to obtain the waveform of the respiratory fundamental frequency changing over time;

[0026] The respiratory correction signal is subjected to absolute value processing and low-pass filtering to obtain the change envelope of the respiratory depth, which characterizes the slow change of respiratory amplitude.

[0027] Preferably, the step of performing three-dimensional spatial trajectory reconstruction and frequency component segmentation on the attitude motion signal stream to obtain the distribution spectrum of body kinetic energy on each axis in space and the characteristic frequency set of periodic motion includes:

[0028] Three-dimensional linear acceleration data and three-dimensional angular velocity data are separated from the attitude motion signal stream;

[0029] The displacement trajectory of the driver's body parts in three-dimensional space is obtained by performing a second integral on the three-dimensional linear acceleration data and combining it with the three-dimensional angular velocity data through attitude calculation.

[0030] The projection components of the displacement trajectory on each axis of the spatial rectangular coordinate system are subjected to Fourier transform to obtain the spectrum on each axis. The distribution spectrum of the body kinetic energy on each axis of space is composed of the spectrum.

[0031] Significant peaks in the spectrum along each axis are detected, and the frequencies corresponding to the peaks are compared with a preset range of typical body motion frequencies to filter out frequency points belonging to periodic motion, thus forming the characteristic frequency set.

[0032] Preferably, the step of constructing a multidimensional physiological feature matrix based on the quantized spectrum, the changing envelope, and the set of feature frequencies includes:

[0033] The ultra-low frequency energy ratio, the low frequency to high frequency energy ratio, and the total power value are extracted from the quantized spectrum and used as the first set of feature vectors.

[0034] Extract the mean, variance, slope, and zero-crossing rate from the changing envelope to form the second set of feature vectors;

[0035] Extract the main frequency value, frequency concentration, and frequency stability coefficient from the set of characteristic frequencies to form the third set of feature vectors;

[0036] The first set of feature vectors, the second set of feature vectors, and the third set of feature vectors are concatenated in a predetermined order to form a row vector;

[0037] Repeat the above steps over multiple consecutive time frames, and stack the resulting row vectors in chronological order to form the multidimensional physiological feature matrix, where each row of the matrix corresponds to the multidimensional feature of a time frame.

[0038] Preferably, the step of performing layer-by-layer matching and difference calculation between the multidimensional physiological feature matrix and the preset driver state template to generate a state deviation sequence and multiple physiological risk markers includes:

[0039] Obtain the preset driver status template, which includes a normal state feature template, a fatigue state feature template, and a disease risk state feature template.

[0040] Calculate the Euclidean distance between each row of the feature vector of the multidimensional physiological feature matrix and the normal state feature template to form a basic deviation sequence;

[0041] Calculate the similarity between each row of the feature vector of the multidimensional physiological feature matrix and the fatigue state feature template, and convert the similarity into fatigue deviation.

[0042] Calculate the matching score between the feature vector of each row of the multidimensional physiological feature matrix and a specific risk pattern in the disease risk state feature template;

[0043] When the value of the basic deviation sequence continuously exceeds the warning threshold, an attention distraction risk marker is generated;

[0044] When the fatigue deviation is consistently higher than the fatigue threshold, a fatigue accumulation risk marker is generated.

[0045] When the score matching a specific risk pattern exceeds a risk threshold, a sudden health risk marker is generated.

[0046] Preferably, the step of selecting a corresponding initial intervention instruction set from a preset intervention strategy library based on the physiological risk marker includes:

[0047] In response to the attention distraction risk marker, an attention reawakening strategy is invoked from the intervention strategy library. The attention reawakening strategy includes initial parameters for acoustic stimulus type, visual cue content, and trigger duration.

[0048] In response to the fatigue accumulation risk marker, an anti-fatigue strategy is invoked from the intervention strategy library, the anti-fatigue strategy including initial parameters for suggested rest intervals, seat massage modes, and air conditioning adjustment targets;

[0049] In response to the sudden health risk marker, an emergency response strategy is invoked from the intervention strategy library. The emergency response strategy includes initial parameters for vehicle control recommendations, contact priorities, and information broadcast templates.

[0050] The initial parameters of the selected strategy are summarized to form the initial intervention instruction set.

[0051] Preferably, the step of dynamically calibrating the parameters of the initial intervention instruction set by fusing the state deviation sequence to generate the final intervention instruction set includes:

[0052] Read the latest values ​​and trends of the state deviation sequence;

[0053] Based on the magnitude of the latest value, the intensity parameter of the acoustic stimulus or the frequency parameter of the visual cue in the initial intervention instruction set are linearly adjusted;

[0054] Based on the slope of the changing trend, the duration parameter of the rest interval suggested in the initial intervention instruction set is adjusted non-linearly.

[0055] The intensity curve of the seat massage mode and the specific temperature value of the air conditioning adjustment target are weighted and corrected by combining the current time, historical driving time and ambient light intensity.

[0056] All parameters, after being adjusted for intensity, frequency, duration, and specific target values, are integrated to form a final set of intervention commands that can be directly parsed by the vehicle's execution system.

[0057] Preferably, after generating the final set of intervention instructions, a model adaptive update step is also included:

[0058] Real-time reception of the target driver's feedback physiological information set after the intervention measures are implemented;

[0059] The heart rate variability quantification spectrum, respiratory depth change envelope and body movement characteristic frequency set for subsequent time periods are extracted from the feedback physiological information set to form a feedback feature matrix.

[0060] Calculate the feature drift between the feedback feature matrix and the multidimensional physiological feature matrix on which the final intervention instruction set was generated;

[0061] If the feature drift is greater than the adaptive update threshold, the feedback feature matrix is ​​used to incrementally update the corresponding state feature template in the preset driver state template.

[0062] Preferably, the continuously captured set of raw physiological information of the target driver includes:

[0063] A piezoelectric sensor array is deployed beneath the surface of the vehicle seat to generate the raw electrical signals of the heart rate signal stream and the respiratory signal stream by detecting minute pressure fluctuations caused by heartbeats and breathing.

[0064] Distributed inertial sensing units are embedded in the backrest and cushion of the vehicle seat to generate the raw motion data of the attitude motion signal stream by measuring the three-dimensional acceleration and angular velocity of the driver's torso and hips.

[0065] A high-precision analog-to-digital converter is used to synchronously sample and digitize the original electrical signals and the original motion data to ensure that the heart rate signal stream, the respiratory signal stream, and the posture motion signal stream have a unified and synchronized time reference.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] By fusing multidimensional heterogeneous physiological features such as heart rate variability spectrum, respiratory depth envelope, and body movement characteristic frequencies into a feature matrix, and performing layer-by-layer matching and difference calculation with a preset driver state template, pattern recognition of driver state is achieved. This method shifts from isolated threshold judgment to multidimensional, holistic pattern similarity assessment, which can more accurately capture complex states characterized by the coordinated changes of multiple vital signs, reduce the misjudgment rate caused by fluctuations in a single physiological signal or environmental interference, and improve the robustness and accuracy of state recognition. The generated continuous state deviation sequence can quantitatively reflect the trend and degree of state deviation from the normal pattern, providing continuous parameter basis for dynamic state assessment.

[0068] The initial intervention command parameters selected from the strategy library are dynamically calibrated based on the state deviation sequence, giving the final executed intervention command adaptive characteristics. This allows the intensity, frequency, or other parameters of the intervention to be non-linearly adjusted according to the driver's real-time state deviation and its changing trend, rather than executing fixed commands. The dynamic calibration mechanism enables the personalization and gradation of the intervention strategy. For mild state deviations, gentle prompts can be given, while for sharp or severe deviations, the intervention intensity is automatically increased. This ensures the effectiveness of the warning while reducing unnecessary interference to the driver, improving the intelligence and acceptability of human-machine interaction. Attached Figure Description

[0069] Figure 1 This is a schematic diagram illustrating the working principle of the vehicle driver state monitoring method based on physiological sign extraction described in this invention.

[0070] Figure 2 A flowchart for heart rate signal processing and quantization spectrum generation;

[0071] Figure 3 Flowchart for constructing a multidimensional physiological feature matrix;

[0072] Figure 4 This is a graph showing the nonlinear relationship between the suggested rest interval and the slope of the state deviation.

[0073] Figure 5This is a graph showing the relationship between the feature drift over time and the adaptive update threshold. Detailed Implementation

[0074] 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.

[0075] Please see Figure 1 This invention provides a method for monitoring the state of a car driver based on physiological sign extraction. The method includes: continuously capturing a set of raw physiological information of the target driver, including heart rate signal stream, respiratory signal stream, and posture motion signal stream; performing time-domain morphological separation and frequency-domain spectral analysis on the heart rate signal stream, obtaining a baseline waveform of the heart rate rhythm and a quantized spectrum of heart rate variability; performing non-stationary signal decomposition and baseline drift elimination on the respiratory signal stream, extracting the respiratory fundamental frequency waveform and the envelope of respiratory depth changes; and performing three-dimensional spatial trajectory reconstruction and frequency component segmentation on the posture motion signal stream. The process yields the distribution spectrum of body kinetic energy along each axis in space and the characteristic frequency set of periodic motion. Based on the obtained quantized spectrum, changing envelope, and characteristic frequency set, a multidimensional physiological feature matrix is ​​constructed. The constructed multidimensional physiological feature matrix is ​​matched and the difference is calculated layer by layer with a preset driver state template. This calculation generates a state deviation sequence and multiple physiological risk markers. Based on the generated physiological risk markers, the corresponding initial intervention instruction set is selected from a preset intervention strategy library. The parameters of the initial intervention instruction set are dynamically calibrated by fusing the state deviation sequence, and finally, an executable final intervention instruction set is generated.

[0076] Example 1: See Figure 2A multi-scale sliding window is applied to the heart rate signal stream to obtain a series of heart rate signal segments. Local extrema are detected for each heart rate signal segment to determine the time coordinates of the start, peak, and end points of each heartbeat cycle. Based on the determined time coordinates, the waveforms of all heartbeat cycles are aligned and averaged to obtain a baseline waveform for the heart rate rhythm. This baseline waveform is subtracted from each heart rate signal segment to obtain the heart rate residual signal. Power spectral density estimation is performed on the heart rate residual signal to calculate its energy proportion in the ultra-low frequency, low frequency, and high frequency bands, forming a quantitative spectrum of heart rate variability. Adaptive mode decomposition is applied to the respiratory signal stream, decomposing it into multiple intrinsic mode function components and a residual trend component. Target components with energy concentrated near the respiratory frequency are identified among the intrinsic mode function components, and these target components are reconstructed into a pure respiratory signal. The residual trend component is subtracted from the original respiratory signal stream to obtain a respiratory correction signal with the slow-varying trend removed. The instantaneous frequency of the pure respiratory signal is calculated to obtain the waveform of the respiratory fundamental frequency changing over time. The respiratory correction signal is subjected to absolute value processing and low-pass filtering to obtain the change envelope of respiratory depth, which characterizes the slow change of respiratory amplitude.

[0077] Three-dimensional linear acceleration data and three-dimensional angular velocity data are separated from the attitude motion signal stream. The three-dimensional linear acceleration data is integrally divided twice, and combined with the three-dimensional angular velocity data, the displacement trajectory of the driver's body parts in three-dimensional space is obtained through attitude calculation. Fourier transforms are performed on the projection components of the displacement trajectory on each axis of the Cartesian coordinate system to obtain the spectrum for each axis. The distribution spectrum of body kinetic energy on each axis is composed of these spectra. Significant peaks in the spectrum of each axis are detected, and the frequencies corresponding to these peaks are compared with a preset range of typical body motion frequencies to filter out frequency points belonging to periodic motion, forming a characteristic frequency set.

[0078] In specific implementation, a multi-scale sliding window is applied to the heart rate signal stream to obtain a series of heart rate signal segments. In some embodiments, the length of the sliding window can be set to cover at least ten consecutive heartbeat cycles. Local extrema are detected for each heart rate signal segment to determine the time coordinates of the start point, peak point, and end point of each heartbeat cycle. Local extrema detection is achieved by comparing the amplitude relationship between the signal point and its neighboring points. Based on the determined time coordinates, the waveforms of all heartbeat cycles are aligned and averaged to obtain the reference waveform of the heart rate rhythm. Waveform alignment is performed with the start point as the reference for time offset correction. The reference waveform of the heart rate rhythm is subtracted from each heart rate signal segment to obtain the heart rate residual signal, which mainly contains information about heart rate variability. Power spectral density estimation is performed on the heart rate residual signal to calculate its energy proportion in the ultra-low frequency, low frequency, and high frequency bands, forming a quantized spectrum of heart rate variability. The power spectral density estimation uses the Welch method.

[0079] Adaptive mode decomposition (ADD) is applied to the respiratory signal stream, decomposing it into multiple intrinsic mode function (IMF) components and a residual trend component. ADD uses an iterative filtering process to decompose the signal into components of different scales. Target components with energy concentrated near the respiratory frequency are identified among the IMF components, and these target components are reconstructed into a clean respiratory signal. The energy concentration is determined by calculating the integral of the power spectrum of each component within a preset respiratory frequency band. The residual trend component is subtracted from the original respiratory signal stream to obtain a respiratory correction signal with the slow-varying trend removed; the residual trend component represents the slow baseline drift in the signal. The instantaneous frequency of the clean respiratory signal is calculated, yielding the waveform of the respiratory fundamental frequency changing over time. The instantaneous frequency is obtained through Hilbert transform and phase differentiation. The respiratory correction signal is then subjected to absolute value processing and low-pass filtering to obtain the envelope representing the slow changes in respiratory amplitude, with the cutoff frequency of the low-pass filter lower than the respiratory fundamental frequency.

[0080] Three-dimensional linear acceleration data and three-dimensional angular velocity data are separated from the attitude motion signal stream. These data are output from the accelerometer and gyroscope of the inertial measurement unit, respectively. The three-dimensional linear acceleration data is integrated twice, and combined with the three-dimensional angular velocity data, attitude calculation is performed to obtain the displacement trajectory of the driver's body parts in three-dimensional space. The attitude calculation uses complementary filtering or Kalman filtering algorithms to fuse multi-sensor data to compensate for integration errors. Fourier transforms are performed on the projection components of the displacement trajectory on each axis of the spatial rectangular coordinate system to obtain the spectrum for each axis. The distribution spectrum of body kinetic energy on each axis is composed of these spectra. In some embodiments, the spectrum for each axis is normalized for comparison. Significant peaks in the spectrum of each axis are detected, and the frequencies corresponding to these peaks are compared with a preset typical body motion frequency range. Frequency points belonging to periodic motion are selected to form a characteristic frequency set. It can be understood that the detection of significant peaks is achieved by setting an amplitude threshold, for example, satisfying the following relationship:

[0081] ,

[0082] in: For frequency Spectral amplitude at that location This is the mean of the axial spectral amplitude. Standard deviation, This is the preset amplitude coefficient.

[0083] Example 2: See Figure 3The following steps are performed: First, the proportion of ultra-low frequency energy, the ratio of low-frequency to high-frequency energy, and the total power value are extracted from the quantized spectrum; these values ​​are used as the first set of feature vectors. Second, the mean, variance, slope, and zero-crossing rate are extracted from the changing envelope; these values ​​are used as the second set of feature vectors. Third, the dominant frequency value, frequency concentration, and frequency stability coefficient are extracted from the feature frequency set; these values ​​are used as the third set of feature vectors. The first, second, and third set of feature vectors are then concatenated in a predetermined order to form a row vector. This process is repeated over multiple consecutive time frames, and the resulting row vectors are stacked chronologically to form a multidimensional physiological feature matrix, where each row of the matrix corresponds to a multidimensional feature of a time frame.

[0084] In specific implementation, the following parameters are extracted from the quantized spectrum: the proportion of ultra-low frequency energy, the ratio of low-frequency to high-frequency energy, and the total power value. The low-frequency to high-frequency energy ratio is the ratio of low-frequency energy to high-frequency energy, and the total power value is the integral of the spectral energy within a preset frequency band. These values ​​serve as the first set of feature vectors, which can be understood as representing the activity state of the autonomic nervous system. The following parameters are extracted from the changing envelope: the mean, variance, slope, and zero-crossing rate of the changing envelope. The slope of the changing envelope is obtained by linear fitting, and the zero-crossing rate is the ratio of the number of times the changing envelope crosses its mean level to the time length. These values ​​serve as the second set of feature vectors. The following parameters are extracted from the feature frequency set: the dominant frequency, frequency concentration, and frequency stability coefficient. The dominant frequency is the frequency point with the highest amplitude in the feature frequency set. The frequency concentration is used to quantify the degree of clustering of frequency points around the dominant frequency. The frequency stability coefficient is the reciprocal of the standard deviation of the dominant frequency values ​​over multiple consecutive time windows. These values ​​serve as the third set of feature vectors.

[0085] The first, second, and third sets of feature vectors are concatenated in a predetermined order to form a row vector. In some embodiments, the predetermined order may be [ultra-low frequency energy ratio, low-frequency to high-frequency energy ratio, total power value, mean of the changing envelope, variance of the changing envelope, slope of the changing envelope, zero-crossing rate of the changing envelope, dominant frequency value, frequency concentration, and frequency stability coefficient]. The steps of extracting the ultra-low frequency energy ratio, low-frequency to high-frequency energy ratio, and total power value from the quantized spectrum, extracting its mean, variance, slope, and zero-crossing rate from the changing envelope, and extracting the dominant frequency value, frequency concentration, and frequency stability coefficient from the feature frequency set are repeated over multiple consecutive time frames. It can be understood that the length of each time frame is consistent with the length of the analysis window used to generate the quantized spectrum, changing envelope, and feature frequency set. The obtained row vectors are stacked in chronological order to form a multidimensional physiological feature matrix, where each row of the matrix corresponds to a multidimensional feature of a time frame. In some embodiments, a row vector containing 10 features stacked over 100 consecutive time frames will form a 100-row, 10-column multidimensional physiological feature matrix.

[0086] Frequency concentration can be calculated using information entropy, for example, by quantifying it using the following relationship:

[0087] ,

[0088] in: Indicates frequency concentration. This represents the total number of frequency points in the characteristic frequency set. Indicates the first The normalized amplitude weight for each frequency point is calculated as the proportion of the amplitude at that frequency point to the sum of the amplitudes at all points. Optional, frequency concentration. The smaller the value, the more concentrated the frequency energy is at a few frequency points. The frequency stability coefficient is calculated by taking the reciprocal of the standard deviation of the dominant frequency values ​​over multiple consecutive time windows. If the standard deviation is... Then the frequency stability coefficient is Standard deviation The calculation is based on a sequence of dominant frequency values ​​over a historical period.

[0089] Example 3: Obtain a preset driver state template, which includes a normal state feature template, a fatigue state feature template, and a disease risk state feature template. Calculate the Euclidean distance between each row of the multidimensional physiological feature matrix's feature vectors and the normal state feature template to form a basic deviation sequence. Calculate the similarity between each row of the multidimensional physiological feature matrix's feature vectors and the fatigue state feature template, and convert the similarity into a fatigue deviation. Calculate the matching score between each row of the multidimensional physiological feature matrix's feature vectors and a specific risk pattern in the disease risk state feature template. When the value of the basic deviation sequence continuously exceeds a warning threshold, an attention distraction risk label is generated. When the fatigue deviation is consistently higher than the fatigue threshold, a fatigue accumulation risk label is generated. When the matching score with a specific risk pattern exceeds a risk threshold, a sudden health risk label is generated.

[0090] In response to an attention distraction risk marker, an attention arousal strategy is retrieved from the intervention strategy library. This strategy includes initial parameters for acoustic stimulus type, visual cue content, and trigger duration. In response to a fatigue accumulation risk marker, an anti-fatigue strategy is retrieved from the intervention strategy library. This strategy includes initial parameters for suggested rest intervals, seat massage modes, and air conditioning adjustment targets. In response to a sudden health risk marker, an emergency response strategy is retrieved from the intervention strategy library. This strategy includes initial parameters for vehicle control suggestions, contact priorities, and information broadcast templates. The initial parameters of the selected strategies are then aggregated to form an initial intervention instruction set.

[0091] In practical implementation, a preset driver state template is obtained. This template includes a normal state feature template, a fatigue state feature template, and a disease risk state feature template. These templates are obtained through statistical learning of the physiological characteristics of corresponding states from a large amount of historical driving data. The Euclidean distance between each row of the multidimensional physiological feature matrix and the normal state feature template is calculated to form a basic deviation sequence. The formula for calculating the Euclidean distance is:

[0092] ,

[0093] in: Represents the fundamental deviation of a single feature vector. This represents the eigenvector of the current row in the multidimensional physiological feature matrix. One element, The first element representing the feature template vector of the normal state One element, This represents the total dimension of the feature vectors. The similarity between each row of the feature vectors in the multidimensional physiological feature matrix and the fatigue state feature template is calculated, and the similarity is converted into a fatigue deviation. It can be understood that the similarity calculation uses cosine similarity, and the fatigue deviation is defined as the difference between 1 and the cosine similarity value. The matching score between each row of the feature vectors in the multidimensional physiological feature matrix and a specific risk pattern in the disease risk state feature template is calculated. When the value of the baseline deviation sequence continuously exceeds the warning threshold, an attention distraction risk marker is generated. It can be understood that the judgment of continuous exceedance requires the baseline deviation sequence to be consistently greater than the warning threshold within a preset time window. When the fatigue deviation is consistently higher than the fatigue threshold, a fatigue accumulation risk marker is generated. In some embodiments, the judgment of consistently higher is based on the moving average of the fatigue deviation over multiple consecutive analysis periods being higher than the fatigue threshold. When the matching score with a specific risk pattern exceeds the risk threshold, a sudden health risk marker is generated. The risk threshold is a pre-set high threshold.

[0094] In response to a risk marker indicating distraction, an attention-awakening strategy is invoked from the intervention strategy library. This strategy includes initial parameters for acoustic stimulus type, visual cue content, and trigger duration. Initial parameters for acoustic stimulus type include the pitch and rhythm of the cue tone, and initial parameters for visual cue content include the text or icon content displayed on the dashboard. In response to a risk marker indicating accumulated fatigue, an anti-fatigue strategy is invoked from the intervention strategy library. This strategy includes initial parameters for suggested rest intervals, seat massage mode, and air conditioning adjustment targets. Initial parameters for seat massage mode specify the massage area and base intensity level, and initial parameters for air conditioning adjustment targets specify the target temperature and airflow. In response to a risk marker indicating sudden health problems, an emergency response strategy is invoked from the intervention strategy library. This strategy includes initial parameters for vehicle control suggestions, contact priority, and information broadcast templates. Initial parameters for vehicle control suggestions include "suggest slowly pulling over to the side of the road," initial parameters for contact priority specify prioritizing contact with emergency services, and initial parameters for information broadcast templates include pre-recorded voice warning content. The initial parameters of the selected strategy, including acoustic stimulus type, visual cue content, trigger duration, suggested rest interval, seat massage mode, air conditioning adjustment target, vehicle control suggestions, communication priority, and information broadcast template, are summarized to form an initial intervention instruction set. Optionally, the initial intervention instruction set can be transmitted within the system in the form of a structured data dictionary or list.

[0095] Example 4: Read the latest values ​​and trends of the state deviation sequence. Based on the magnitude of the latest values, linearly adjust the intensity parameters of the acoustic stimuli or the frequency parameters of the visual cues in the initial intervention command set. Based on the slope of the trend, non-linearly adjust the duration parameters of the suggested rest intervals in the initial intervention command set. Combining the current time, historical driving time, and ambient light intensity, weighted corrections are made to the intensity curve of the seat massage mode and the specific temperature value of the air conditioning adjustment target. Integrate all parameters after adjustments for intensity, frequency, duration, and specific target values ​​to form the final intervention command set that can be directly parsed by the vehicle's execution system.

[0096] In practice, the latest values ​​and trends of the state deviation sequence are read. The trend is obtained by calculating the slope of a linear regression of the state deviation sequence within a recent time window. Based on the magnitude of the latest value, the intensity parameter of the acoustic stimulus or the frequency parameter of the visual cue in the initial intervention instruction set is linearly adjusted. It can be understood that the intensity parameter of the acoustic stimulus and the frequency parameter of the visual cue are directly proportional to the latest value. Based on the slope of the trend, the duration parameter of the suggested rest interval in the initial intervention instruction set is non-linearly adjusted. In some embodiments, when the duration parameter is... At that time, the final adjusted duration It can be calculated using the following formula:

[0097] ,

[0098] in: This indicates the recommended rest interval duration after dynamic calibration. This parameter represents the duration of the suggested rest interval in the initial intervention instruction set. The slope representing the trend of the state deviation sequence. This is the proportionality coefficient. This is a non-linear factor used to control the sensitivity of the adjustment. The intensity curve of the seat massage mode and the specific temperature value of the air conditioning target are weighted and corrected by combining the current time, historical driving time, and ambient light intensity. The intensity curve correction of the seat massage mode increases the base intensity level as historical driving time increases, while the correction of the specific temperature value of the air conditioning target is based on a combination of the current time and ambient light intensity. It can be understood that the ambient light intensity is obtained through the vehicle's light sensor, and its value is used to fine-tune the target temperature of the air conditioning. In some embodiments, refer to Table 1 for the ambient light intensity and air conditioning temperature correction values.

[0099] Table 1: Mapping Table of Ambient Light Intensity and Air Conditioning Temperature Correction Values

[0100]

[0101] Table 1 shows that during periods of strong daylight, the target temperature is tended to be lowered to provide a cooler environment. The weighted correction process combines the coefficients corresponding to the current time, the coefficients corresponding to the historical driving duration, and the temperature correction values ​​corresponding to the ambient light intensity obtained from Table 1 to obtain the final parameters of the seat massage mode intensity curve and the specific temperature value of the air conditioning adjustment target. All parameters adjusted for intensity, frequency, duration, and specific target value are integrated to form a final intervention command set that can be directly parsed by the vehicle's execution system. This final intervention command set is typically encapsulated and transmitted using a standardized in-vehicle network data frame format.

[0102] See Figure 4 This graph is the core visualization result of the dynamic calibration process for intervention instructions, focusing on the relationship between the slope of the deviation and the recommended rest interval. The horizontal axis represents the slope of the deviation, reflecting the rate of deterioration of the driver's physiological state from the normal pattern; the vertical axis represents the recommended rest interval, a key parameter in the anti-fatigue intervention strategy. This trend aligns with the logic of non-linearly adjusting the rest interval based on the trend of deviation changes, ensuring that rest time is not excessively increased during mild deterioration, while extending the rest interval during rapid deterioration. This allows the intervention strategy to better adapt to the driver's real-time state, reflecting the personalized and tiered characteristics of the dynamic calibration mechanism.

[0103] Example 5: Real-time reception of the target driver's feedback physiological information set after intervention measures are implemented. From the feedback physiological information set, the quantitative spectrum of heart rate variability, the envelope of respiratory depth changes, and the set of body movement characteristic frequencies for subsequent time periods are extracted to construct a feedback feature matrix. The feature drift between the feedback feature matrix and the multidimensional physiological feature matrix used to generate the final intervention instruction set is calculated. If the feature drift is greater than the adaptive update threshold, the corresponding state feature template in the preset driver state template is incrementally updated using the feedback feature matrix.

[0104] A piezoelectric sensor array is deployed beneath the surface of the vehicle seat to generate raw electrical signals for heart rate and respiration by detecting minute pressure fluctuations caused by heartbeats and respiration. Distributed inertial sensing units are embedded in the backrest and cushion of the vehicle seat to generate raw motion data for posture motion signal streams by measuring the three-dimensional acceleration and angular velocity of the driver's torso and hips. A high-precision analog-to-digital converter is used to synchronously sample and digitize the raw electrical signals and raw motion data to ensure that the heart rate, respiration, and posture motion signal streams have a unified and synchronized time reference.

[0105] In practice, the system receives a set of feedback physiological information from the target driver in real time after the intervention. This set is acquired using the same method as the original physiological information set, and includes heart rate signal streams, respiratory signal streams, and posture movement signal streams for subsequent time periods. The system extracts a quantitative spectrum of heart rate variability, a respiratory depth change envelope, and a set of body movement characteristic frequencies for subsequent time periods from the feedback physiological information set. The extraction method is the same as that used to extract these parameters from the original physiological information set, forming a feedback feature matrix with dimensions consistent with the multidimensional physiological feature matrix. The system calculates the feature drift between the feedback feature matrix and the multidimensional physiological feature matrix used to generate the final intervention instruction set. In some embodiments, the feature drift is calculated by calculating the average Euclidean distance between the eigenvectors of corresponding rows of the two matrices.

[0106] ,

[0107] in: Indicates the characteristic drift amount. This indicates the number of rows in the feature matrix (number of time frames). The dimension of the feature vector. The feedback feature matrix represents the first... Line 1 The characteristic values ​​of the column, The multidimensional physiological feature matrix used to generate the final set of intervention instructions is represented by the first... Line 1 The eigenvalues ​​of the column. If the feature drift is greater than the adaptive update threshold, the corresponding state feature template in the preset driver state template is incrementally updated using the feedback feature matrix. The incremental update adopts the moving average or exponential weighted average method. It can be understood that the update process only modifies the state feature template in the preset driver state template that corresponds to the currently detected physiological risk marker.

[0108] A piezoelectric sensor array is deployed beneath the surface of the vehicle seat. This array detects minute pressure fluctuations caused by heartbeats and respiration, generating raw electrical signals for heart rate and respiration. The piezoelectric sensor array is distributed in a matrix to facilitate the capture of physiological signals from different locations in the torso. Distributed inertial sensing units are embedded in the backrest and cushion of the vehicle seat. These units measure the three-dimensional acceleration and angular velocity of the driver's torso and hips, generating raw motion data for attitude motion signal streams. Each distributed inertial sensing unit includes a three-axis accelerometer and a three-axis gyroscope. A high-precision analog-to-digital converter (ADC) is used to synchronously sample and digitize the raw electrical signals and motion data. In some embodiments, the ADC synchronously acquires multiple analog signals at a sampling rate of at least 1 kHz. Optionally, synchronization is ensured by a unified hardware clock trigger, guaranteeing that the heart rate, respiration, and attitude motion signal streams have a unified and synchronized time reference.

[0109] See Figure 5 This figure is a core visualization of the model's adaptive update process, focusing on the degree of change in physiological characteristics after intervention and the template update triggering logic. The horizontal axis represents time, the vertical axis represents feature drift, and the dashed line represents the adaptive update threshold. Feature drift reflects the magnitude of change in the driver's physiological characteristics after intervention. When this value exceeds the threshold, the system automatically performs incremental updates to the preset driver state template. This figure clearly presents the triggering timing of template updates, demonstrating the mechanism of dynamically optimizing the state template based on feedback physiological information. By monitoring the dynamic changes in feature drift, adaptive iteration of the state template is achieved, improving the accuracy of subsequent state recognition.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the state of a driver of a vehicle based on the extraction of physiological signs, characterized in that, The method comprises the following steps: Continuously capturing a set of raw physiological information of a target driver, the set of raw physiological information comprising a heart rate signal stream, a respiration signal stream, and a posture motion signal stream; Performing time-domain morphology separation and frequency-domain spectrum analysis on the heart rate signal stream to obtain a reference waveform of heart rate rhythm and a quantized spectrum graph of heart rate variability; Performing non-stationary signal decomposition and baseline drift elimination on the respiration signal stream to extract a respiration fundamental frequency waveform and a change envelope line of respiration depth; Performing three-dimensional space trajectory reconstruction and frequency component division on the posture motion signal stream to obtain a distribution spectrum of body movement energy on each axis in space and a characteristic frequency set of periodic motion; Constructing a multi-dimensional physiological feature matrix based on the quantized spectrum graph, the change envelope line, and the characteristic frequency set; Performing layer-by-layer matching and difference calculation on the multi-dimensional physiological feature matrix and a preset driver state template to generate a state deviation sequence and a plurality of physiological risk markers; According to the physiological risk markers, selecting a corresponding initial intervention instruction set from a preset intervention strategy library; Fusing the state deviation sequence to dynamically calibrate the parameters of the initial intervention instruction set to generate a final intervention instruction set; The method comprises the following steps: Extracting an ultra-low frequency energy proportion, a low frequency to high frequency energy ratio, and a total power value from the quantized spectrum graph as a first group of feature vectors; Extracting a mean value, a variance, a slope, and a zero-crossing rate from the change envelope line as a second group of feature vectors; Extracting a main frequency value, a frequency concentration degree, and a frequency stability coefficient from the characteristic frequency set as a third group of feature vectors; Concatenating the first group of feature vectors, the second group of feature vectors, and the third group of feature vectors in a predetermined order to form a row vector; Repeating the above steps on a plurality of consecutive time frames, stacking the obtained plurality of row vectors in time sequence to form the multi-dimensional physiological feature matrix, wherein each row of the matrix corresponds to the multi-dimensional features of a time frame.

2. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 1, characterized in that, The method comprises the following steps: Applying a multi-scale sliding window to the heart rate signal stream to obtain a series of heart rate signal segments; Performing local extreme point detection on each heart rate signal segment to determine the time coordinates of the start point, peak point, and end point of each heart beat cycle; Aligning and superimposing the waveforms of all heart beat cycles according to the time coordinates to obtain the reference waveform of heart rate rhythm; Subtracting the reference waveform from each heart rate signal segment to obtain a heart rate residual signal; Performing power spectrum density estimation on the heart rate residual signal to calculate the energy proportion in the ultra-low frequency, low frequency, and high frequency bands, forming the quantized spectrum graph of heart rate variability.

3. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 2, characterized in that, The method comprises the following steps: applying adaptive mode decomposition to the respiratory signal stream to decompose it into a plurality of intrinsic mode function components and a residual trend component; identifying a target component from the intrinsic mode function components in which energy is concentrated around the respiratory frequency, and reconstructing the target component into a pure respiratory signal; subtracting the residual trend component from the respiratory signal stream to obtain a respiratory corrected signal with slow varying trend removed; calculating the instantaneous frequency of the pure respiratory signal to obtain a waveform of the respiratory fundamental frequency varying over time; performing absolute value processing and low-pass filtering on the respiratory corrected signal to obtain a respiratory depth envelope representing slow variation of the respiratory amplitude.

4. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 1, characterized by, the three-dimensional space trajectory reconstruction and frequency component division performed on the posture motion signal stream to obtain a body motion energy distribution spectrum on each spatial axis and a characteristic frequency set of periodic motion, including: separating three-dimensional linear acceleration data and three-dimensional angular velocity data from the posture motion signal stream; integrating the three-dimensional linear acceleration data twice and combining the three-dimensional angular velocity data to obtain a displacement trajectory of the driver's body part in three-dimensional space through attitude calculation; performing Fourier transform on the projection components of the displacement trajectory on each axis of the spatial rectangular coordinate system to obtain a frequency spectrum on each axis, and the body motion energy distribution spectrum on each spatial axis is composed of the frequency spectra; detecting significant peaks in the frequency spectrum on each axis, comparing the frequencies corresponding to the peaks with a preset typical body motion frequency range, and screening out frequency points belonging to periodic motion to form the characteristic frequency set.

5. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 1, characterized in that, the multi-dimensional physiological feature matrix and the preset driver state template are matched layer by layer and the difference is calculated to generate a state deviation sequence and a plurality of physiological risk markers, including: obtaining the preset driver state template, the preset driver state template including a normal state feature template, a fatigue state feature template, and a disease risk state feature template; calculating the Euclidean distance between each row feature vector of the multi-dimensional physiological feature matrix and the normal state feature template to form a basic deviation sequence; calculating the similarity between each row feature vector of the multi-dimensional physiological feature matrix and the fatigue state feature template, and converting the similarity into a fatigue deviation; calculating the matching score between each row feature vector of the multi-dimensional physiological feature matrix and a specific risk pattern in the disease risk state feature template; when the value of the basic deviation sequence continuously exceeds a warning threshold, generating an attention distraction risk marker; when the fatigue deviation continuously exceeds a fatigue threshold, generating a fatigue accumulation risk marker; when the matching score with the specific risk pattern exceeds a risk threshold, generating a sudden health risk marker.

6. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 5, characterized in that, the physiological risk markers are used to select corresponding initial intervention instruction sets from a preset intervention strategy library, including: in response to the attention distraction risk marker, calling an attention awakening strategy from the intervention strategy library, the attention awakening strategy including initial parameters of acoustic stimulation type, visual prompt content, and trigger duration; In response to the fatigue accumulation risk flag, an anti-fatigue strategy is called from an intervention strategy library, the anti-fatigue strategy including initial parameters of a recommended rest interval, a seat massage mode, and an air conditioning regulation target; In response to the sudden health risk flag, an emergency response strategy is called from the intervention strategy library, the emergency response strategy including initial parameters of a vehicle control recommendation, a contact priority, and an information broadcast template; The initial parameters of the selected strategy are summarized to form the initial intervention instruction set.

7. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 6, characterized in that, The parameters of the initial intervention instruction set are dynamically calibrated by fusing the state deviation degree sequence to generate a final intervention instruction set, including: Reading the latest value and change trend of the state deviation degree sequence; According to the size of the latest value, linearly adjusting the intensity parameter of the acoustic stimulus or the frequency parameter of the visual prompt in the initial intervention instruction set; According to the slope of the change trend, non-linearly adjusting the time length parameter of the recommended rest interval in the initial intervention instruction set; Combining the current time, historical driving time, and ambient light intensity, the intensity curve of the seat massage mode and the specific temperature value of the air conditioning regulation target are weighted and corrected; After the final intervention instruction set is generated, the model is updated adaptively, including:

8. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 1, characterized in that, Real-time receiving a feedback physiological information set of the target driver after the intervention measure is implemented; Extracting the heart rate variability quantization spectrum, the respiration depth change envelope, and the body motion characteristic frequency set of the subsequent time period from the feedback physiological information set to form a feedback feature matrix; Calculating the feature drift between the feedback feature matrix and the multi-dimensional physiological feature matrix on which the final intervention instruction set is generated; If the feature drift is greater than an adaptive update threshold, the feedback feature matrix is used to incrementally update the corresponding state feature template in the preset driver state template. The original physiological information set of the target driver is continuously captured, including:

9. The method for monitoring the state of a driver of a vehicle based on extraction of a physiological sign according to claim 1, characterized in that, A piezoelectric sensing array is arranged under the surface of the vehicle seat to detect the original electrical signals of the heart rate signal stream and the respiration signal stream caused by the slight pressure fluctuations due to heartbeats and respiration; A distributed inertial sensing unit is embedded in the backrest and cushion of the vehicle seat to measure the three-dimensional acceleration and angular velocity of the driver's torso and hips to generate the original motion data of the attitude motion signal stream; High-precision analog-to-digital converters are used to synchronously sample and digitize the original electrical signals and the original motion data to ensure that the heart rate signal stream, the respiration signal stream, and the attitude motion signal stream have a unified and synchronous time reference. ​

Images