Multi-source information fusion automatic driving auxiliary decision method and system

Abstract

The application relates to a multi-source information fusion automatic driving auxiliary decision method and system, the method comprises the following steps: acquiring driver state data and vehicle dynamic data, performing behavior analysis and feature analysis, obtaining behavior feature data and dynamic feature data; acquiring environment perception data, integrating and calculating the behavior feature data and the dynamic feature data, and obtaining a state evaluation value; performing execution level identification according to the state evaluation value, obtaining a driving assistance level; and performing hierarchical decision construction based on the driving assistance level, and obtaining a multi-level driving assistance decision. The application can more accurately identify potential risks and make reasonable responses, and effectively reduces the decision lag problem caused by environmental mutation or abnormal driver state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to a multi-source information fusion automatic driving auxiliary decision-making method and system. BACKGROUND

[0002] With the rapid development of sensor technology and artificial intelligence algorithms, how to achieve accurate driving decision support through multi-source information fusion has become one of the core directions of current research. The existing auxiliary decision-making methods usually perform single-dimensional state evaluation, lack comprehensive evaluation of multiple dynamic correlations, and are difficult to fully reflect the comprehensive risk level of the driving system. This fragmented processing method will lead to a lag in the system's response to complex driving scenarios or decision bias, making it difficult to adapt to different levels of risk intervention needs. SUMMARY

[0003] The main purpose of the present application is to provide a multi-source information fusion automatic driving auxiliary decision-making method and system, which can more accurately identify potential risks and make reasonable responses, effectively reducing the problem of decision lag caused by environmental mutations or abnormal driver states.

[0004] To achieve the above purpose, the present application provides a multi-source information fusion automatic driving auxiliary decision-making method, comprising:

[0005] Obtaining driver state data and vehicle dynamic data, performing behavior analysis and feature analysis to obtain behavior feature data and dynamic feature data;

[0006] Obtaining environmental perception data, integrating and calculating the behavior feature data and the dynamic feature data to obtain a state evaluation value;

[0007] Performing execution level identification according to the state evaluation value to obtain a driving assistance level;

[0008] Based on the driving assistance level, a hierarchical decision-making is constructed to obtain a multi-level driving assistance decision.

[0009] Further, the driver state data and vehicle dynamic data collected by the vehicle-mounted sensor in real time are obtained, behavior analysis and feature analysis are performed, and behavior feature data and dynamic feature data are obtained, comprising:

[0010] The driver state data is subjected to visual signal decomposition to obtain eyelid movement trajectory sequences and head posture angle change sequences;

[0011] The eyelid movement trajectory sequences are subjected to closed-eye duration recording to obtain closed-eye duration;

[0012] The head posture angle change sequences are subjected to pitch angle fluctuation amplitude statistics to obtain a head forward-leaning frequency value;

[0013] performing behavior analysis based on the closed-eye duration and the head-tilt frequency value, to obtain the behavior feature data;

[0014] performing steering wheel angle time collection and lane deviation analysis on the vehicle dynamic data, to obtain the dynamic feature data.

[0015] Further, the performing steering wheel angle time collection and lane deviation analysis on the vehicle dynamic data, to obtain the dynamic feature data, comprises:

[0016] performing steering wheel signal analysis on the vehicle dynamic data, to obtain steering wheel angle time;

[0017] performing steering quantization on the steering wheel angle time, to obtain steering fluctuation feature quantity;

[0018] performing lane line coordinate extraction on the vehicle dynamic data, to obtain lane lateral deviation data;

[0019] performing cointegration relationship analysis on the lane lateral deviation data according to the steering fluctuation feature quantity, to obtain direction deviation control coefficient;

[0020] performing stability classification mapping on the direction deviation control coefficient, to obtain lateral control level;

[0021] performing feature integration on the steering fluctuation feature quantity and the lateral control level, to obtain the dynamic feature data.

[0022] Further, the obtaining environment perception data, integrating and calculating the behavior feature data and the dynamic feature data, to obtain state evaluation value, comprises:

[0023] performing road structure analysis and light intensity identification on the environment perception data, to obtain road structure parameter, obstacle distribution parameter and luminous flux coefficient;

[0024] performing mode recognition and visual compensation processing on the behavior feature data, to generate behavior mode parameter;

[0025] performing environment association on the behavior mode parameter according to the road structure parameter and the obstacle distribution parameter, to obtain behavior environment association parameter;

[0026] performing constraint adjustment on the behavior environment association parameter according to the dynamic feature data and the luminous flux coefficient, to obtain dynamic correction factor;

[0027] performing proportional fusion calculation on the behavior environment association parameter and the dynamic correction factor, to obtain state evaluation value.

[0028] Further, the step of constraining and adjusting the behavioral environment correlation parameters based on the dynamic feature data and the luminous flux coefficient to obtain a dynamic correction factor includes:

[0029] The luminous flux coefficient is used to perform illuminance intensity range compensation calculation to generate illuminance compensation coefficient;

[0030] The vehicle motion trajectory is analyzed from the dynamic feature data to obtain the lateral displacement fluctuation.

[0031] The lateral displacement fluctuation is optically corrected according to the illumination compensation coefficient to generate an illumination correction value.

[0032] Based on the illumination correction amount, displacement-related constraints are applied to the behavioral environment-related parameters to obtain preliminary correction parameters;

[0033] The preliminary correction parameters are subjected to boundary safety verification to obtain the safety boundary parameters;

[0034] The dynamic correction factor is obtained by dynamically synthesizing the initial correction parameter based on the safety boundary parameter.

[0035] Further, the step of identifying the execution level based on the state evaluation value to obtain the driving assistance level includes:

[0036] The state evaluation value is divided into intervals based on a preset security level threshold to obtain security interval information;

[0037] The volatility of the state assessment value is statistically analyzed to obtain a state stability index;

[0038] Based on the state stability index, conflict detection is performed on the safe interval information to obtain a conflict correction level signal;

[0039] Historical driving assistance data is obtained through a preset system storage unit, and the continuity of the conflict correction level signal is verified to obtain the level migration parameters.

[0040] Based on the level migration parameters, the conflict correction level signal is downgraded and verified to obtain the security level output.

[0041] The safety level output is assisted in identifying the urgency level based on the vehicle dynamic data to obtain the driving assistance level.

[0042] Furthermore, the step of acquiring historical driving assistance data through a preset system storage unit, verifying the continuity of the conflict correction level signal, and obtaining level transition parameters includes:

[0043] The historical driving assistance data is subjected to trend fitting to obtain historical driving trend data;

[0044] The conflict correction level signal is analyzed for level discreteness to generate the current level information;

[0045] The current level information is processed by similarity measurement based on the historical driving trend data, and the level trend similarity is output.

[0046] Based on the similarity of the grade trends, a continuity confidence calculation is performed on the current grade information to obtain the migration confidence coefficient;

[0047] The smoothing migration amount is calculated based on the migration confidence coefficient to obtain the level migration parameter.

[0048] Furthermore, the step of constructing a hierarchical decision based on the driving assistance level to obtain multi-level driving assistance decisions includes:

[0049] The driving assistance levels are divided into level ranges to obtain the first-level assistance range, the second-level assistance range, and the third-level assistance range;

[0050] The state evaluation value is weighted according to the first-level auxiliary interval to obtain the first-level control weight set;

[0051] Based on the second-level auxiliary interval, the state evaluation value is dynamically thresholded to obtain the second-level control threshold set;

[0052] The state evaluation values ​​are prioritized according to the third-level auxiliary interval to obtain the third-level control priority sequence.

[0053] The first-level control weight set, the second-level control threshold set, and the third-level control priority sequence are fused and calculated to obtain a multi-level control parameter set;

[0054] The multi-level driving assistance decision is obtained by constructing an auxiliary instruction mapping based on the multi-level control parameter set.

[0055] The present invention also provides an autonomous driving assistance decision-making system based on multi-source information fusion, applied to any one of the above-described autonomous driving assistance decision-making methods based on multi-source information fusion, comprising:

[0056] The acquisition module is used to acquire driver status data and vehicle dynamic data, perform behavior analysis and feature analysis, and obtain behavior feature data and dynamic feature data.

[0057] The analysis module is used to acquire environmental perception data, integrate and calculate the behavioral feature data and the dynamic feature data to obtain a state evaluation value;

[0058] The association module is used to identify the execution level based on the state evaluation value to obtain the driving assistance level;

[0059] The processing module is used to construct hierarchical decisions based on the driving assistance level to obtain multi-level driving assistance decisions.

[0060] The present invention provides an autonomous driving assistance decision-making method and system that integrates multi-source information, which has the following beneficial effects:

[0061] By integrating and comprehensively analyzing driver status data, vehicle dynamic data, and environmental perception data, a more comprehensive assessment of the overall risk level of the driving system can be achieved. This overcomes the assessment bias caused by traditional methods relying solely on single-dimensional data, thereby improving the accuracy and reliability of assisted decision-making. By integrating and calculating behavioral and dynamic feature data, and classifying them based on status assessment values, a hierarchical driving assistance decision-making system is constructed. This system can dynamically adjust intervention strategies according to different risk levels, avoiding the problems of over-intervention or insufficient warnings caused by the coarse granularity of traditional methods. The hierarchical decision-making system based on driving assistance levels enables multi-level intervention strategies that match risk levels, flexibly responding to driving scenarios of varying complexity. This ensures necessary safety interventions while avoiding unnecessary interference, improving the adaptability and user experience of driving assistance. By fusing multi-source information in real time and dynamically adjusting the assistance level, potential risks can be identified more accurately and responded to appropriately, effectively reducing decision-making lag caused by sudden environmental changes or abnormal driver states. Attached Figure Description

[0062] Figure 1 This is a flowchart of an autonomous driving assisted decision-making method based on multi-source information fusion provided by the present invention;

[0063] Figure 2 This is a structural diagram of an autonomous driving assistance decision-making system based on multi-source information fusion, provided by the present invention.

[0064] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0067] Reference Figure 1As shown, the present invention provides a multi-source information fusion-based autonomous driving assisted decision-making method, comprising:

[0068] Step S1: Acquire driver status data and vehicle dynamic data, perform behavior analysis and feature analysis, and obtain behavioral feature data and dynamic feature data;

[0069] Step S2: Acquire environmental perception data, integrate and calculate behavioral feature data and dynamic feature data to obtain a state assessment value;

[0070] Step S3: Identify the execution level based on the status assessment value to obtain the driving assistance level;

[0071] Step S4: Construct a hierarchical decision based on the level of driving assistance to obtain multi-level driving assistance decisions.

[0072] Based on the steps described above, the detailed process is as follows:

[0073] Step S1:

[0074] The vehicle-mounted vision sensor captures driver facial images at 30 frames per second, and eyelid movement trajectory sequences are extracted based on machine vision algorithms. The pupil occlusion area ratio is calculated, and when the occlusion area exceeds 80%, it is marked as a closed-eye state. The duration of continuous eye occlusion is used as the core fatigue indicator. Simultaneously, a nine-axis inertial measurement unit captures head posture angle data, and sliding window spectral analysis is performed on the pitch angle to extract the frequency of movements with a forward tilt exceeding 15 degrees per unit time, quantifying head nodding fatigue characteristics.

[0075] In vehicle dynamic data processing, steering wheel angle time is acquired via the controller area network bus at a sampling rate of 100Hz, and steering fluctuation is calculated using a sliding standard deviation algorithm with a window size of 0.5 seconds. Lane recognition cameras output lane line coordinates in real time, and a lateral offset sequence is generated through continuous frame difference calculation. A long-term equilibrium relationship model between steering wheel angle fluctuation and lane offset is established, with the correlation coefficient mapped to the lateral control level. Eye-closing duration, head-nodding frequency, and stability level are timestamped and encapsulated into a dynamic feature data packet.

[0076] Step S2:

[0077] The system synchronously collects illumination intensity (unit: lux) and road point cloud data through a multispectral environmental perception module. The road point cloud is segmented using the RANSAC algorithm to extract curvature radius and obstacle distribution density parameters. Illumination intensity parameters are used for visual feature compensation: when illumination is below 50 lux, an infrared supplementary lighting mode is activated, and the duration of eye closure is compensated with a 1.3-fold weight. Road curvature parameters are used to correct lane lateral offset, eliminating measurement bias caused by centrifugal force in curves through geometric projection.

[0078] Based on road topological complexity (curvature radius < 100m is considered high complexity) and obstacle distribution density (> 0.2 obstacles / meter is considered high density), an environmental correlation matrix for behavioral pattern parameters is established. The stability level and illumination coefficient in the dynamic features are input into the constraint adjustment module. A hyperbolic tangent function is used to generate a safety boundary threshold, providing hard truncation protection for the behavioral-environment correlation parameters. After compensation, the behavioral parameters and constraint adjustment factors are integrated using a weighted fusion algorithm, with a weight allocation ratio of 60% for behavioral features, 35% for dynamic features, and 5% for environmental parameters. The output is a normalized state evaluation value.

[0079] Step S3:

[0080] The status assessment value is input into a three-level threshold divider: the interval [0, 0.3] is marked as monitoring level (code 01), the interval [0.3, 0.6] is warning level (code 02), and the interval [0.6, 1] is intervention level (code 03). The coefficient of variation is calculated for the status values ​​within a continuous 10-second window to generate a stability index. When the assessment value is within the threshold boundary ±0.05 range (e.g., 0.28-0.32), a conflict detection mechanism is activated: if the stability index is below 0.2, the original level is maintained; if it is above 0.2, it is upgraded to the adjacent higher level. 30 minutes of historical driving level data is retrieved from the vehicle's EEPROM memory, and a trend fit is performed using an ARIMA(2, 1, 1) model. A continuous alarm is triggered when the predicted value deviates from the current level by more than one level.

[0081] The level transition parameters are smoothly implemented using filters. The degradation verification module has a safety protection mechanism: if the evaluation value decreases by more than 20% for three consecutive sampling periods, the degradation operation is frozen. Finally, urgency weighting is applied based on real-time vehicle speed parameters; when the vehicle speed exceeds 90 km / h, the control level is automatically upgraded by one level.

[0082] Step S4:

[0083] The driving assistance level is converted into a control strategy through a command mapper: the monitoring level (01) is mapped to the instrument panel status display command; the warning level (02) activates the audible and visual alarm module and outputs a "BI-BI" warning sound at a frequency of 2Hz; the intervention level (03) triggers vehicle control bus commands, including seat vibration (intensity level 3) and forced intervention of the lane keeping system. The driver's gesture signal is obtained from the human-machine interface. The V-shaped gesture covers the audible and visual alarm command for up to 10 minutes, and the palm gesture restores system control. The vehicle speed parameters transmitted in real time by the vehicle bus are used to dynamically adjust the execution intensity and generate a speed optimization control strategy set. The multi-level commands are conflict-resolved through a priority arbitrator. The priority of the emergency braking command is set to 9 (highest level), the lane keeping command to 7, and the audible and visual alarm command to 5. The final output is encapsulated as a CAN protocol data frame, which includes the command type code, execution intensity value, and effective duration field, realizing the system-level deployment of multi-level driving assistance decisions.

[0084] This invention provides a multi-source information fusion-based autonomous driving assistance decision-making method. By integrating driver state data, vehicle dynamic data, and environmental perception data and performing comprehensive analysis, it can more comprehensively assess the overall risk level of the driving system, overcoming the assessment bias caused by traditional methods relying on only single-dimensional data, thereby improving the accuracy and reliability of assisted decision-making. By integrating and calculating behavioral and dynamic feature data and performing hierarchical identification based on state assessment values, a hierarchical driving assistance decision-making system is constructed. This system can dynamically adjust intervention strategies according to different risk levels, avoiding the problems of over-intervention or insufficient warnings caused by the coarse granularity of traditional methods. The hierarchical decision-making based on driving assistance levels realizes multi-level intervention strategies matched to risk levels, flexibly responding to driving scenarios of varying complexity. This ensures necessary safety interventions while avoiding unnecessary interference, improving the adaptability of driving assistance and user experience. By fusing multi-source information in real time and dynamically adjusting the assistance level, potential risks can be identified more accurately and appropriate responses can be made, effectively reducing decision-making lag caused by sudden environmental changes or abnormal driver states.

[0085] In one embodiment, driver state data and vehicle dynamic data are acquired, and behavioral analysis and feature analysis are performed to obtain behavioral feature data and dynamic feature data, including:

[0086] Driver status data is collected synchronously via an onboard near-infrared camera and an inertial measurement unit (IMU). The near-infrared camera is mounted on the inside of the windshield of the cockpit, continuously capturing images of the driver's face at a rate of 30 frames per second to ensure clear imaging under both day and night lighting conditions. The IMU is deployed on top of the driver's seat back, synchronously collecting three-axis head attitude angle data (pitch, yaw, and roll angles) at a sampling rate of 100Hz.

[0087] In the visual signal decomposition process, the original image is converted to grayscale to reduce the data volume, and then Gaussian filtering is used to eliminate high-frequency noise before extracting the facial contour boundaries. Based on pre-defined eye region localization rules (such as the relative positional relationship between the pupil center point and the corner of the eye), the eye region image is extracted from the edge detection results. The pupil center point is tracked on consecutive frames of eye images, and the horizontal and vertical coordinates of the pupil center in each frame are recorded to form time-series data, namely the eyelid movement trajectory sequence—this sequence directly reflects the frequency and amplitude of the driver's eye closing, blinking, and other actions.

[0088] Acquiring the head posture angle change sequence requires Kalman filtering of the raw posture angle data output by the IMU to eliminate interference from vehicle vibration and sensor noise. The filtered pitch angle data (reflecting the forward and backward tilt angle of the head) is extracted as the core parameter for subsequent analysis. To ensure time synchronization, the timestamps of the visual image frames and the IMU data are aligned using a hardware synchronization circuit, with the error controlled within ±1 millisecond, ensuring that the eyelid movement trajectory and head posture angle changes are correlated and analyzed under the same time reference.

[0089] Each frame of the eyelid movement trajectory sequence corresponds to a close-up image of the eye. The proportion of the pupil occlusion area is calculated using image processing algorithms. Each frame of the eye image is binarized, with the grayscale mean as the initial threshold. The optimal segmentation threshold is determined through iterative adjustment, dividing the image into the eyelid region (black) and the pupil region (white). Morphological opening operations are performed on the binary image to remove edge noise. Then, connected component analysis is used to extract the number of pixels in the pupil region, and the proportion of the pupil region to the total number of pixels in the eye is calculated, which is the proportion of the pupil occlusion area.

[0090] When the pupil occlusion area exceeds 80%, the frame is considered a closed-eye state. To avoid misjudging instantaneous blinks (such as brief eye closure caused by changes in light), a threshold for consecutive closed-eye frames is set: only when the occlusion area exceeds 80% for 5 consecutive frames (approximately 0.165 seconds) is it marked as a valid closed-eye event. The duration of closed-eye events is calculated by accumulating the number of valid closed-eye events and multiplying by the frame interval (0.033 seconds). For example, 10 consecutive closed-eye frames correspond to a duration of 0.33 seconds.

[0091] Using the maximum exposed pupil area (i.e., pupil area without eyelid obstruction) during the driver's initial awake state as a benchmark, the proportion of obstructed area in subsequent frames is converted into a relative proportion to the benchmark value. The output duration of closed eyes is averaged using a sliding window (window size is 3 cycles, i.e., 30 seconds) to smooth out short-term fluctuations and reflect the driver's fatigue accumulation trend over medium to long periods.

[0092] Pitch angle data in the head posture angle change sequence reflects the degree of forward and backward tilt of the driver's head. Under fatigue, drivers often exhibit a natural forward tilt of the head due to decreased attention, manifesting as regular fluctuations in the pitch angle over a certain period. By performing low-frequency filtering (cutoff frequency 1Hz) on the pitch angle data, high-frequency noise interference such as vehicle bumps is removed, retaining the low-frequency fluctuation component of 0.1-0.5Hz, which is related to the driver's active head posture adjustments.

[0093] Waveform analysis is used to detect the peaks and troughs of pitch angle fluctuations: when the pitch angle changes from a positive value to a negative value (or vice versa), it is recorded as one fluctuation. A fluctuation amplitude threshold is set: when the difference between a single peak and trough exceeds 15 degrees, it is determined as a valid forward tilt movement (this threshold is based on head movement statistics under fatigue driving scenarios; in a normal, conscious state, the forward tilt amplitude is usually less than 10 degrees). The number of valid forward tilt movements within a unit of time (60 seconds) is counted, which is the head tilt frequency.

[0094] The significance of forward tilting is quantified by setting an amplitude weighting mechanism. The amplitude value of each effective forward tilt (the difference between the peak and trough) is divided by 15 degrees to obtain the weighting coefficient (range 0.67-1). The total forward tilt intensity is obtained by summing the weighted values. The final output head tilt frequency value is the ratio of the total forward tilt intensity to 60 seconds, normalized to the interval [0, 1]. This head tilt frequency value is used as a key indicator for fatigue early warning.

[0095] Vehicle dynamic data is acquired through collaborative acquisition via the vehicle's Controller Area Network (CAN) bus and lane recognition camera. The CAN bus transmits steering wheel angle signals in real time at a sampling rate of 100Hz. This signal is decoded to generate continuous time-series data, recording the trajectory of the driver's steering wheel angle changes. Simultaneously deployed on the vehicle's front bumper, the lane recognition camera captures road surface images at a rate of 25 frames per second. Lane line position coordinates are extracted using edge detection algorithms, and lane lateral offset data (i.e., the vehicle's left or right offset relative to the lane centerline) is generated through continuous frame differential calculation.

[0096] A sliding window cointegration analysis method was used to establish the correlation between steering wheel angle and lane departure. A linear regression analysis was performed on steering variability (standard deviation of steering wheel angle) and lane departure using a 0.5-second window size, and the correlation coefficient between the two was calculated. If the correlation coefficient exceeded 0.7 (a set threshold), a significant correlation was determined, and steering variability was included as an influencing factor on lane departure in the dynamic feature data.

[0097] The encapsulation of dynamic feature data requires two key processing steps: first, normalizing the steering fluctuation to eliminate dimensional differences caused by varying driver operating habits; second, smoothing and filtering the lane departure (with a window size of 3 cycles, or 7.5 seconds) to remove instantaneous deviation interference caused by vehicle bumps. The output dynamic feature data includes three dimensions: steering fluctuation, lane departure, and correlation coefficient, reflecting the stability of vehicle lateral control and the consistency of driver operation.

[0098] Environmental perception data is collected synchronously via a multispectral sensor and a light intensity sensor. The multispectral sensor is mounted on the top of the vehicle's windshield, acquiring road point cloud data at a rate of 10 frames per second. After segmentation processing, the road curvature radius and obstacle distribution density are extracted. The light intensity sensor is deployed above the dashboard, monitoring ambient light intensity (unit: lux) in real time. The data is averaged and filtered through a sliding window to eliminate interference from instantaneous light changes.

[0099] When the light intensity is below 50 lux, the infrared supplementary light mode is activated to compensate for the weight of the closed-eye duration feature value in the behavioral feature data (based on the fatigue detection experience threshold); when the road curvature radius is less than 100 meters (high curvature road section), the lane offset in the dynamic features is scaled (to eliminate the measurement deviation caused by the centrifugal force of the curve).

[0100] Based on road topological complexity (curvature radius < 100m is considered high complexity) and obstacle distribution density (> 0.2 obstacles / meter is considered high density), a behavior-environment coupling coefficient is established. This coefficient is determined using a lookup table: the coupling coefficient is 1.5 for high complexity + high density scenarios, and 1.0 for other scenarios. Steering fluctuations and lane deviations in the dynamic feature data need to be multiplied by this coefficient to generate constrained and adjusted dynamic parameters.

[0101] The compensated behavioral characteristic data (duration of eye closure, frequency of forward lean) and the constraint-adjusted dynamic parameters (steering fluctuation, lane deviation) are weighted and summed according to the weight ratio (behavioral characteristics 60%, dynamic characteristics 35%, environmental parameters 5%), and the state evaluation value is output.

[0102] This embodiment effectively improves the accuracy and reliability of fatigue driving warning by using multi-source information fusion for multi-dimensional feature extraction and dynamic state assessment. Specifically, feature analysis integrating driver state data and vehicle dynamic data avoids the information limitations of single sensor data, comprehensively reflecting the static and dynamic manifestations of driver fatigue. The introduction of environmental perception data reduces the impact of environmental interference on feature extraction, minimizing the risk of misjudgment. The state assessment value, through constraint adjustment and smoothing filtering, eliminates instantaneous fluctuation interference, ensuring that the assessment results stably reflect the actual driving state.

[0103] In one embodiment, vehicle dynamic data is subjected to steering wheel angle time acquisition and lane departure analysis to obtain dynamic feature data, including:

[0104] The steering wheel angle signal is a digitally encoded angle value (unit: degrees), containing real-time angle changes in response to the driver's steering wheel operation. The steering sensor, located at the steering wheel shaft, converts the mechanical rotation of the steering wheel into an electrical signal using electromagnetic induction or optical signal modulation. After analog-to-digital conversion, it outputs the angle data in analog form.

[0105] The effective payload of the turning signal is extracted using a CAN bus protocol parser, eliminating erroneous frames caused by bus interference. A precise timestamp (accuracy ±1 millisecond) is added to each frame of turning data using a GPS or onboard clock module to ensure consistency of the time base in subsequent analyses. The final output turning time series is an array of angle values ​​arranged in chronological order, directly reflecting the dynamic process of the driver's steering wheel operation, providing raw data support for subsequent steering fluctuation analysis.

[0106] Steering quantization aims to transform continuous steering angle time series into quantifiable characteristic parameters reflecting steering stability. The steering angle data is segmented into sliding windows with a window size of 0.5 seconds (corresponding to 25 frames of data) to ensure coverage of typical timescales of driver operation. The standard deviation of the steering angle data within each window is calculated to obtain the steering variability (unit: degrees), which characterizes the dispersion of the driver's steering wheel operation over a short period.

[0107] The steering angle data is low-pass filtered (cutoff frequency 1Hz) to eliminate high-frequency noise interference such as vehicle bumps, retaining the low-frequency fluctuation component of 0.1-0.5Hz, which is related to the driver's active steering wheel adjustments. Subsequently, the extreme values ​​of the fluctuation amplitude (difference between peaks and troughs) are extracted through waveform analysis, and combined with the fluctuation frequency (unit: Hz) to generate a steering fluctuation characteristic. This characteristic comprehensively reflects the smoothness and frequency characteristics of the driver's steering operation. Under fatigue conditions, drivers often exhibit increased steering fluctuation (fluctuation > 2 degrees) or abnormal frequency (> 0.3Hz) due to distraction; therefore, this parameter can serve as an important input for fatigue warning.

[0108] Lane coordinate extraction is achieved using an onboard camera or millimeter-wave radar. The camera is mounted on the top of the windshield and captures road surface images at a rate of 25 frames per second. After grayscale conversion and Gaussian filtering for noise reduction, the lane line edges are extracted using the Canny edge detection algorithm. Optical flow tracking is performed on consecutive frames of images, and the longitudinal position of the lane lines is determined by feature point matching. Then, the lateral coordinates (in meters) of the lane lines are extracted using Hough transform, generating a time series of lane line positions.

[0109] Millimeter-wave radar is deployed on the front bumper of the vehicle, emitting electromagnetic waves at a frequency of 77 GHz. It detects the lateral offset of lane lines (in meters) through the Doppler effect of the echo signal. After range-Doppler imaging processing, the radar data extracts the lateral position information of the lane lines, which is then fused and calibrated with camera data to eliminate single-sensor errors.

[0110] The output lateral offset sequence is an array of lane line lateral coordinates arranged over time, reflecting the real-time offset of the vehicle relative to the lane centerline. When driving while fatigued, the driver's reaction time is delayed or attention is reduced, often resulting in increased lane deviation (>0.3 meters) or increased deviation frequency (>0.1 Hz). This sequence provides crucial data for subsequent steering and offset correlation analysis.

[0111] The cointegration analysis of steering undulation characteristics and lane lateral offset data needs to be achieved through time window alignment and correlation calculation. A time window of 0.5 seconds (corresponding to 25 frames of data) is selected, and the steering undulation characteristics (unit: degrees) and lane lateral offset data (unit: meters) are matched one-to-one with the timestamps to form synchronized data pairs.

[0112] For each data pair within a window, the covariance is calculated to reflect the coordinated trend of change between the two factors; simultaneously, the standard deviation is calculated to eliminate dimensional differences. The ratio of covariance to standard deviation is the directional deviation control coefficient, characterizing the degree of linear correlation between steering wheel operation and lane departure. This coefficient is used to traverse the entire time series through a sliding window, generating a continuous correlation coefficient curve, providing a quantitative basis for subsequent stability grading.

[0113] The stability classification of the lane departure control coefficient requires a clearly defined threshold range. Based on statistical data from measured tests under fatigue driving scenarios, when the correlation coefficient > 0.7, steering wheel operation and lane departure exhibit strong coordination, classified as "high stability"; when 0.5 < correlation coefficient ≤ 0.7, the coordination is moderate, classified as "medium stability"; and when the correlation coefficient ≤ 0.5, the coordination is weak, and there is a high risk of loss of steering control, classified as "low stability". This classification is achieved through a lookup table method, with each level corresponding to different control strategy weights, providing level identifiers for subsequent feature integration.

[0114] Feature integration requires weighted fusion of steering fluctuation characteristics (reflecting operational activity) and lateral control levels (reflecting operational reliability). In practice, steering fluctuation characteristics are normalized (range 0-1), and lateral control levels are converted into numerical weights (high stability = 1.2, medium stability = 1.0, low stability = 0.8).

[0115] The two factors are weighted and summed according to their respective weight ratios (60% for fluctuation characteristics and 40% for stability level) to generate a dynamic feature value in the 0-1 range. This value is then smoothed by a sliding window mid-range filter (window size is 3 cycles, i.e., 7.5 seconds) to eliminate short-term fluctuation interference, and finally serves as the quantitative output of the vehicle's dynamic characteristics, reflecting the overall state of the driver's lateral control.

[0116] This embodiment comprehensively captures the dynamic details of driver operation (such as steering frequency and amplitude) by collecting steering wheel angle time and performing steering quantification analysis, avoiding the information limitations of single sensor data; the extraction and cointegration analysis of lane lateral deviation data establishes the correlation between steering operation and lane control, reducing the impact of environmental interference such as curves and bumps on feature extraction; stability grading mapping identifies high-risk scenarios with unreliable operation by quantifying the degree of coordination between steering and deviation; finally, the integration of dynamic feature data comprehensively reflects the overall state of driver lateral control.

[0117] In one embodiment, environmental perception data is acquired, and behavioral feature data and dynamic feature data are integrated and calculated to obtain a state assessment value, including:

[0118] The processing of environmental perception data begins with the raw data acquisition from vehicle-mounted sensors, encompassing multimodal information such as road images, laser point clouds, and illumination intensity. After grayscale conversion and Gaussian filtering for noise reduction, road images are processed using edge detection algorithms to extract terrain contours such as lane lines, curves, and slopes. The curvature radius of the lane lines (reflecting the degree of road curvature) and slope angle (reflecting the road undulation) are also identified to generate road structure parameters.

[0119] Laser point cloud data is clustered to extract the coordinates of obstacles within 50 meters ahead, and the number of obstacles per unit area (obstacle distribution density) is calculated to generate obstacle distribution parameters. Illumination intensity data is collected in real-time by a light sensor mounted above the dashboard. After being averaged using a sliding window (window size 5 seconds) to eliminate interference from instantaneous light changes, a stable luminous flux coefficient (unit: lux) is obtained, reflecting the direct impact of ambient light on visual characteristics. These parameters collectively constitute a physical characteristic description of the vehicle's driving environment, providing environmental background support for subsequent analysis.

[0120] Behavioral feature data originates from real-time monitoring of the driver's state, including two core data categories: eyelid movement trajectory and head posture angle. Eyelid movement trajectory is analyzed over time to extract the duration of eye closure and blinking frequency, reflecting the driver's visual fatigue level. Head posture angle is analyzed by spectrum to extract the forward tilt frequency and backward tilt amplitude, reflecting the driver's level of concentration. Pattern recognition is performed using a pre-set fatigue feature database: the duration of eye closure and the forward tilt frequency are compared with historical fatigue samples to determine the current level as "mild fatigue" (eye closure duration < 2 seconds, forward tilt frequency < 0.5 times / second), "moderate fatigue" (2 seconds ≤ eye closure duration < 4 seconds, 0.5 times / second ≤ forward tilt frequency < 1 time / second), or "severe fatigue" (eye closure duration ≥ 4 seconds, forward tilt frequency ≥ 1 time / second).

[0121] Visual compensation processing adjusts feature weights based on luminous flux coefficient: when luminous flux coefficient < 50 lux (low-light environment), the weight of the duration of eye closure is increased (from 0.6 to 0.8), as low light easily leads to visual misjudgment; when luminous flux coefficient > 100 lux (strong-light environment), the weight of the forward tilt frequency is increased (from 0.4 to 0.6), as drivers are more likely to adjust their head posture due to glare under strong light. The final generated behavioral pattern parameters integrate fatigue level and environmentally adaptive adjusted feature values.

[0122] Environmental correlation analysis requires dynamically linking driver behavior characteristics with the road physical environment. For the curve radius (R) in the road structure parameters, when R < 100 meters (sharp curves), the driver needs to frequently adjust the steering wheel, and the frequency of head tilting forward will increase due to eye tracking of the curve. Therefore, the threshold for head tilting forward frequency is adjusted to 1.2 times the normal value (e.g., the normal threshold of 1 time / second is adjusted to 1.2 times / second). When R > 200 meters (gentle curves), the driver's operation is more stable, and the threshold for the duration of closing the eyes can be relaxed to 3 seconds (normal 2 seconds).

[0123] Regarding the obstacle density (ρ) in the obstacle distribution parameters, when ρ > 0.2 obstacles / meter (dense obstacles), the driver's attention is highly focused, and the duration of eye closure is shortened (threshold adjusted to 1.5 seconds), while the frequency of head tilting forward increases due to frequent observation of obstacles (threshold adjusted to 1.1 times / second). When ρ < 0.1 obstacles / meter (open road sections), the driver is prone to prolonged eye closure due to relaxation, and the eye closure duration threshold is tightened to 1.8 seconds. Through the above environmental association rules, behavioral pattern parameters (fatigue level) are dynamically bound to road environment features (topology, obstacles), generating behavioral environment association parameters that reflect the driver's true fatigue state in different road scenarios, avoiding misjudgments caused by single features due to environmental differences.

[0124] Road structure parameters (such as curve radius and gradient) and obstacle distribution parameters (such as obstacle density) together constitute the environmental constraints for driver operation. For sharp curves with a radius less than 100 meters, drivers need to frequently adjust the steering wheel to maintain their trajectory. In this case, the frequency of head tilting forward increases significantly due to visual tracking of the curve (normal frequency 0.5-1 times / second, increasing to 1.2-1.5 times / second in sharp curve scenarios). Therefore, the forward tilting frequency threshold in the behavioral pattern parameters needs to be dynamically adjusted according to the curve radius: when the curve radius is <100 meters, the forward tilting frequency threshold increases from 1 time / second to 1.2 times / second to avoid frequent adjustments being misjudged as fatigue.

[0125] Regarding obstacle distribution parameters, when the obstacle density is greater than 0.2 obstacles / meter (such as in construction zones or congested traffic), drivers are highly focused, and the duration of eye-closing is shortened due to frequent observation of obstacles (normal threshold 2 seconds, tightened to 1.5 seconds in dense obstacle scenarios). Conversely, in open road sections (obstacle density < 0.1 obstacles / meter), drivers are more likely to close their eyes for extended periods due to relaxation, and the eye-closing duration threshold is relaxed to 2.5 seconds. By linearly weighting the behavioral pattern parameters (eye-closing duration, forward leaning frequency) with road topology and obstacle distribution parameters (weights of 0.4 and 0.6 respectively), a behavioral environment correlation parameter is generated. This parameter reflects the driver's true fatigue state in different road scenarios, effectively avoiding misjudgments caused by single features due to environmental differences.

[0126] Dynamic characteristic data (such as steering sway and lane departure) reflect the stability of vehicle lateral control, while the luminous flux coefficient (ambient light intensity) directly affects the reliability of visual characteristics. When steering sway exceeds 2 degrees (high-frequency adjustment) or lane departure is greater than 0.3 meters (deviation from lane center), it indicates abnormal driver operation. In this case, the weight of the behavioral environment-related parameters should be reduced (from 0.7 to 0.5) to reflect the interference of operational instability on fatigue state.

[0127] When the luminous flux coefficient is below 50 lux (low-light environment), the visual sensor is susceptible to noise, and the reliability of features such as eye-closing duration decreases. Therefore, the dynamic correction factor needs to introduce an illumination compensation coefficient (1.2 times) to improve the reliability of behavioral environment-related parameters. When the luminous flux coefficient is above 100 lux (strong-light environment), the driver frequently adjusts their head posture due to glare, reducing the reliability of forward tilt frequency features. The dynamic correction factor introduces an attenuation coefficient (0.8 times). Finally, the dynamic correction factor is calculated using the formula K=α×β (α is the operational stability weight, and β is the illumination compensation coefficient) to adjust the behavioral environment-related parameters to reliable values ​​adapted to the current environment.

[0128] The calculation of the state assessment value requires a combination of behavioral environment-related parameters (reflecting the essence of fatigue) and dynamic correction factors (reflecting environmental interference). These two factors are weighted and summed according to their respective weight ratios (60% for behavioral environment-related parameters and 40% for dynamic correction factors) to generate a state assessment value in the 0-1 range. To eliminate short-term fluctuations, a sliding window midpoint filter (10-second window size) is used to smooth the assessment value, ensuring that the result stably reflects the actual driving state. For example, when the behavioral environment-related parameter is 0.8 (high-risk fatigue) and the dynamic correction factor is 1.1 (low-light compensation), the weighted calculation yields 0.8 × 0.6 + 1.1 × 0.4 = 0.92. After filtering, the output state assessment value is 0.9, indicating a highly dangerous state. This value directly serves as the core input for subsequent classification decisions, providing a quantitative basis for determining the level of driver assistance.

[0129] This embodiment effectively improves the accuracy and environmental adaptability of fatigue driving warnings by integrating environmental perception and behavioral features through multi-source information fusion. Specifically, the analysis of environmental perception data and the identification of light intensity provide a dynamic environmental background for behavioral feature analysis, avoiding the information limitations of single sensor data; the generation of behavioral pattern parameters, combined with environmental association rules to adjust thresholds, reduces misjudgments caused by differences in road scenarios; and the dynamic correction factor, through the coordinated calibration of steering fluctuations and luminous flux coefficients, eliminates the noise impact of environmental interference on features. The final state evaluation value, after proportional fusion and smoothing filtering, comprehensively reflects the driver's true fatigue state and environmental constraints, significantly enhancing the robustness of the warning, reducing the probability of false triggering, and providing more reliable decision support for the active safety control of high-level autonomous driving.

[0130] In one embodiment, the behavioral environment-related parameters are constrained and adjusted based on dynamic feature data and luminous flux coefficient to obtain a dynamic correction factor, including:

[0131] The luminous flux coefficient reflects the degree of influence of ambient light on the visual sensor. Illumination intensity range compensation calculations must be based on the reliability differences of visual features under different lighting scenarios. For each range of luminous flux coefficient, the error coefficient of the visual features is determined through experimental statistics: in low light conditions, the visual sensor is susceptible to noise interference, so the error coefficient is set to 1.3 (the feature value needs to be amplified to compensate for noise); in medium light conditions, the error coefficient is 1.0 (no significant interference); in strong light conditions, the driver adjusts their head posture due to glare, so the error coefficient is set to 0.8 (the feature value needs to be reduced to decrease misjudgment). The illumination compensation coefficient is generated through range mapping. For example, when the luminous flux coefficient is 30 lux (low light), the compensation coefficient is 1.3; when the luminous flux coefficient is 100 lux (medium light), the compensation coefficient is 1.0; and when the luminous flux coefficient is 200 lux (strong light), the compensation coefficient is 0.8. This coefficient is directly used for subsequent optical correction of lateral displacement fluctuations, ensuring the consistency of feature values ​​under different lighting scenarios.

[0132] Dynamic feature data includes steering wheel angle time and lane lateral offset data. Vehicle trajectory analysis requires extracting the actual change in lateral displacement from these two types of data. The steering wheel angle time is differentiated to obtain the steering wheel angular velocity (unit: degrees / second), reflecting how quickly the driver adjusts the direction. Subsequently, the lane lateral offset data is calculated using a sliding window difference (window size 0.5 seconds) to obtain the change in lateral displacement per unit time (unit: meters / second).

[0133] By combining vehicle speed sensor data (unit: km / h), the steering wheel operation is converted into actual lateral displacement fluctuation using the kinematic formula: displacement fluctuation = angular velocity × wheel radius × time. This fluctuation comprehensively reflects the degree of matching between the driver's operation and the vehicle's actual movement. Under fatigue conditions, due to delayed driver reaction, the lateral displacement fluctuation may abnormally increase or decrease, providing basic data for subsequent optical correction.

[0134] The core function of the illumination compensation coefficient is to correct the measurement error of lateral displacement by the visual sensor under different lighting conditions. The lateral displacement fluctuation is multiplied by the illumination compensation coefficient generated in step 1 to obtain the illumination correction amount. For example, when the luminous flux coefficient is 30 lux (weak light, compensation coefficient 1.3), if the original lateral displacement fluctuation is 0.4 m / s, it becomes 0.4 × 1.3 = 0.52 m / s after correction (amplifying the error to compensate for insufficient measurement under weak light); when the luminous flux coefficient is 200 lux (strong light, compensation coefficient 0.8), the original fluctuation is 0.4 m / s, which becomes 0.4 × 0.8 = 0.32 m / s after correction (reducing the error to decrease glare interference). This corrected displacement more realistically reflects the actual state of driver operation and vehicle movement, avoiding the misleading influence of illumination changes on subsequent constraint adjustments and ensuring the accuracy of the dynamic correction factor.

[0135] The illumination correction reflects the degree of matching between the driver's operation and the vehicle's actual movement. It is combined with behavioral environment parameters, and the parameter weights are adjusted through displacement-related constraint rules. A safe threshold range for displacement fluctuation is set: under normal driving scenarios, lateral displacement fluctuation should be between 0.1 and 0.5 m / s; fluctuations exceeding 0.5 m / s or falling below 0.1 m / s are considered abnormal displacements.

[0136] For abnormal displacement scenarios, if the illumination correction is greater than 0.5 m / s, it indicates aggressive driver operation. In this case, the weight of "eyes-closed duration" in the behavioral environment-related parameters should be reduced, as aggressive operation masks fatigue characteristics. If the displacement is less than 0.1 m / s (such as prolonged constant-speed driving), it indicates driver relaxation and inattention. The weight of "forward lean frequency" should be increased (from 0.4 to 0.6), as forward leaning better reflects attention status. Through the above displacement-related constraint rules, the behavioral environment-related parameters are adjusted to preliminary correction parameters adapted to the current operating state, ensuring consistency between the parameters and actual driving behavior.

[0137] Boundary safety verification requires setting safe ranges for parameters to avoid misjudgments due to excessive adjustments. First, based on historical fatigue driving data statistics, determine reasonable fluctuation ranges for behavioral environment-related parameters: the safe range for eye-closing duration is 0.5-4 seconds, and the safe range for forward-leaning frequency is 0.3-1.5 times / second. If the initially corrected parameters exceed this range (e.g., the corrected eye-closing duration is 4.5 seconds, exceeding the upper limit of 4 seconds), then safety boundary correction is triggered: the excess portion is reduced proportionally (e.g., 4.5 seconds is corrected to 4 seconds) to ensure the parameters remain within the safe range.

[0138] The safety boundaries are dynamically adjusted based on vehicle speed sensor data (unit: km / h): In high-speed scenarios (>80 km / h), the driver's operational error tolerance decreases, and the safety boundaries narrow (the upper limit for eye-closing time is reduced to 3.5 seconds, and the upper limit for forward tilting frequency is reduced to 1.2 times / second); in low-speed scenarios (<40 km / h), the error tolerance increases, and the safety boundaries widen (the upper limit for eye-closing time is increased to 4.5 seconds, and the upper limit for forward tilting frequency is increased to 1.8 times / second). The final safety boundary parameters are the adjusted values ​​of the parameters within the safe range, ensuring the reliability of subsequent dynamic factor synthesis.

[0139] The dynamic factor synthesis requires combining safety boundary parameters with preliminary correction parameters, and generating the final dynamic correction factor through weighted fusion. In practice, the fusion weights of each parameter are first determined: safety boundary parameters account for 40% (ensuring parameters are within a safe range), and preliminary correction parameters account for 60% (preserving core information related to the environment). For example, if the eye-closing duration in the preliminary correction parameters is 3.8 seconds (within a safe range) and the safety boundary parameter is 0 (not exceeding the limit), then the dynamic correction factor = 3.8 × 0.6 + 0 × 0.4 = 2.28; if the forward tilt frequency in the preliminary correction parameters is 1.6 times / second (exceeding the upper limit of 1.5 times / second) and the safety boundary parameter is -0.1 times / second (reduced to 1.5 times / second), then the dynamic correction factor = 1.5 × 0.6 + (-0.1) × 0.4 = 0.9 - 0.04 = 0.86.

[0140] The output dynamic correction factor is smoothed by sliding window averaging filter to eliminate short-term fluctuation interference, reflecting the driver's real fatigue state and environmental constraints under different operating scenarios, and providing reliable input for the calculation of state assessment value.

[0141] This embodiment effectively improves the reliability and environmental adaptability of fatigue driving warning through multi-dimensional adjustment and verification of dynamic correction factors. Specifically, the illumination compensation coefficient dynamically adjusts the feature weights according to the ambient light intensity, solving the problem of visual feature measurement errors in low-light and high-light scenarios and avoiding misjudgments caused by changes in illumination; the optical correction of lateral displacement fluctuations, combined with the actual vehicle movement state, corrects the measurement deviation of the visual sensor caused by illumination interference, ensuring a more realistic matching relationship between operation and movement; displacement-related constraints and boundary safety verification limit the parameter fluctuation range, preventing abnormal operations from misleading fatigue characteristics and enhancing the anti-interference ability of parameters; dynamic factor synthesis, through weighted fusion of safety boundaries and correction parameters, retains the core information related to the environment while ensuring that the parameters are within a reasonable range, and the output dynamic correction factor more accurately reflects the driver's true state.

[0142] In one embodiment, the driver assistance level is determined by identifying the execution level based on the state assessment value, including:

[0143] The status assessment value is a quantitative indicator generated by the fusion of multi-source information, comprehensively reflecting the driver's current fatigue risk level. Its numerical range typically covers a continuous interval from 0 to 1, where 0 represents the lowest safe state (no signs of fatigue) and 1 represents the highest risk state (severe fatigue). The preset safety level thresholds are scientifically set based on the typical characteristics of fatigued driving and traffic safety requirements, and are usually divided into three core intervals: low-risk interval (0-0.3, corresponding to "monitoring level"), medium-risk interval (0.3-0.6, corresponding to "early warning level"), and high-risk interval (e.g., above 0.6, corresponding to "intervention level").

[0144] After the status assessment value is output from the pre-feature fusion module, it is input to the interval segmentation module. This module has a built-in preset threshold comparison logic, which performs preliminary classification by comparing the status value with the threshold of each interval in real time. When the status value falls in the low-risk interval, it is marked as "monitoring level," prompting the system to continue to observe the driver's status; when the status value enters the medium-risk interval, it is marked as "warning level," triggering an audible and visual alarm or information prompt; when the status value exceeds the high-risk interval threshold, it is marked as "intervention level," initiating active control measures (such as limiting vehicle speed or assisted braking). This process is implemented through the efficient computation of digital logic circuits or microcontrollers, ensuring the real-time nature and accuracy of the classification, and providing a basic basis for the level division in subsequent steps.

[0145] The state stability index is used to quantify the degree of fluctuation of state assessment values ​​over time and is a key basis for judging the authenticity of fatigue state. Its calculation logic is based on the analysis of changes within a sliding time window: a continuous time window is selected, all state assessment values ​​within the window are extracted, the absolute value of the difference between adjacent time points is calculated, and then the average of these differences is obtained, ultimately yielding the average fluctuation amplitude. This value is normalized and mapped to the 0-1 interval, where 0 represents a completely stable state and 1 represents a drastic fluctuation.

[0146] The physical significance of stability indicators lies in reflecting the persistence of driver fatigue: when awake, the driver's physiological characteristics are actively regulated by the brain, and the state assessment value fluctuates less; when fatigued, due to distraction and slowed neural response, the state value is easily affected by environmental interference (such as sudden changes in light or road bumps), resulting in a significant increase in the fluctuation range.

[0147] Stability indicators can serve as the core basis for screening "pseudo-fatigue" states. If the state value is in the medium-to-high risk range but the stability indicator is extremely low (close to 0), it is a misjudgment caused by environmental interference. If the state value is in the low-risk range but the stability indicator is high (close to 1), then the accumulation of potential fatigue risks should be guarded against.

[0148] The core objective of conflict detection is to identify abnormal fluctuations in the state assessment value near the boundary of the primary interval, avoiding misclassification of the level due to transient interference or measurement errors. In practice, the criteria for determining boundary conflicts are first set: when the state assessment value approaches the boundary of the primary interval and the stability index exceeds a preset fluctuation threshold, it is determined to be a "boundary conflict." At this point, the primary marker is dynamically corrected based on the magnitude of the stability index.

[0149] For example, if the status assessment value is 0.29 (close to the upper boundary of low risk 0.3), but the stability index is 0.25 (above the threshold 0.2), it indicates that the status value is overestimated due to environmental interference (such as visual misjudgment under low light). In this case, the initial label "monitoring level" should be corrected to "warning level". Conversely, if the status assessment value is 0.31 (just entering the medium risk range), but the stability index is 0.1 (below the threshold 0.2), it indicates that the status value is stable and reliable, and the "warning level" label should be maintained.

[0150] The corrected signal is processed by a logic gating unit, which outputs a "maintain," "upgrade," or "downgrade" correction command, providing more accurate input for the calculation of subsequent level transition parameters. This process is implemented through a preset decision table or conditional judgment logic in the microcontroller, ensuring that the correction result conforms to the physiological laws of fatigued driving while effectively filtering out interference from environmental noise.

[0151] The system storage unit typically uses non-volatile memory to store the driver's historical driving assistance data for long periods. This data includes key information such as a sequence of status assessment values ​​from the past 30 minutes to 2 hours, primary safety level markers, and conflict correction records, forming continuous time-series data indexed by timestamps.

[0152] The core of continuity verification is to determine the rationality of the level adjustment by analyzing the matching degree between the trend of historical data and the current conflict correction signal. In specific implementation, firstly, the historical state evaluation value sequence corresponding to the current time window (such as the most recent 10 minutes) is extracted from the storage unit, and its mean, variance and other statistics are calculated to construct a historical state distribution model. Subsequently, the current conflict-corrected level signal (such as "monitoring level" or "early warning level") is compared with this model: if the current level is consistent with the mainstream level of the same historical time period (such as being at "monitoring level" for 90% of the time in the past), it is judged as "continuous"; if the current level deviates significantly from the historical distribution (such as being at "early warning level" for only 10% of the time in the past, but currently appearing continuously), it is judged as "abnormal".

[0153] The grading transition parameter is generated based on the continuity verification results and is used to quantify the trend strength of grading adjustments. For example, if the continuity verification shows that the current grading is consistent with the historical trend, the transition parameter is set to "stable"; if there are abnormal deviations, the transition parameter is set to "fluctuating". This parameter is smoothed by a sliding window averaging filter to eliminate the interference of single abnormal data and outputs a transition parameter that reflects the grading change trend, providing a historical reference for subsequent grading verification.

[0154] The core of downgrade verification is to avoid excessive adjustments to the security level based on conflict correction and historical trends, thereby ensuring the authenticity and reliability of the security level. In practice, a level migration decision table is first established, and verification rules are set according to the type of migration parameters (stable / fluctuating) and the direction of conflict correction (upgrade / downgrade).

[0155] If the migration parameter is "stable" and the conflict correction is "maintain", then the safety level output is consistent with the conflict correction signal;

[0156] If the migration parameter is "stable" but the conflict correction is "upgraded", the duration of this level in historical data needs to be checked: if the duration of the current level does not exceed the historical average (e.g., the historical average duration of "warning level" is 5 minutes), then the original level is maintained; if it exceeds the average, then upgrade is allowed.

[0157] If the migration parameter is "volatility" and the conflict correction is "downgrade", the volatility of the current state assessment value must be considered: if the volatility is lower than the historical threshold (e.g., 0.2), the downgrade is maintained; if the volatility is too high (e.g., above 0.3), the downgrade is postponed to avoid misjudgment.

[0158] During the verification process, the matching logic of the decision table is executed by a microcontroller or dedicated logic chip to ensure the real-time performance of the verification rules. The final output security level is the result of the combined effect of the conflict correction signal and the migration parameters. It reflects the immediate risk of the current state while avoiding misadjustment of the level due to short-term fluctuations, providing a more robust input for subsequent auxiliary urgency identification.

[0159] Vehicle dynamic data (such as vehicle speed, steering angle, braking signal, and lateral acceleration) is crucial information reflecting the urgency of driving scenarios and needs to be combined with safety level output to generate the final driver assistance level. Key features extracted from vehicle dynamic data include: higher vehicle speeds indicate shorter driver reaction time and lower tolerance for fatigue risk; larger steering angles indicate more focused driver attention and higher reliability of fatigue characteristics; and greater lateral acceleration indicates poorer lateral stability and a greater potential risk of fatigue.

[0160] The core of assisted urgency recognition is the weighted fusion of vehicle dynamic data and safety level output. For example, when the safety level is "warning level" (0.3-0.6) and the vehicle speed is >90km / h, the urgency level increases by 20%; when the safety level is "intervention level" (>0.6) and the lateral acceleration is >0.5g, the urgency level increases by 30%. The fused urgency value is calculated using a fuzzy logic algorithm and ultimately mapped to a driving assistance level (such as "light assistance," "moderate assistance," and "emergency intervention"). This level directly controls the vehicle's active safety functions (such as lane keeping strength and braking response speed), ensuring that the assistance measures match the risk level in different driving scenarios, maximizing driving safety.

[0161] This embodiment effectively improves the accuracy and scenario adaptability of fatigue driving warning through multi-dimensional state assessment and dynamic level verification. Specifically, the interval segmentation of the preset safety level threshold divides the state assessment value into three levels: monitoring, early warning, and intervention, providing a clear benchmark for risk classification; the combination of volatility statistics and state stability indicators identifies the persistence of fatigue state and reduces misjudgments caused by environmental interference; and conflict detection and correction are used to correct abnormal fluctuations near the signal processing boundary.

[0162] In one embodiment, historical driving assistance data is obtained through a preset system storage unit, and the continuity of the conflict correction level signal is verified to obtain level transition parameters, including:

[0163] Historical driver assistance data is stored in non-volatile memory, containing a continuous sequence of state assessment values ​​from the past 30 minutes to 2 hours, primary safety level markers (monitoring, warning, and intervention levels), and corresponding vehicle dynamic parameters (such as vehicle speed and steering angle). The core of trend fitting is to extract long-term patterns of data through time series analysis. Specifically, a sliding window averaging method is used: with a 5-minute window, the mean of the state assessment values ​​within the window is calculated, generating a series of smooth mean points; then, linear regression or multinomial fitting is performed on these mean points to obtain a trend vector reflecting the state's change over time. This vector, with timestamps on the horizontal axis and state assessment values ​​on the vertical axis, visually presents the long-term evolution trend of driver fatigue. The length of the trend vector is consistent with the time span of the historical data, ensuring sufficient coverage of driving scenario samples and providing a reliable benchmark for subsequent similarity measurements.

[0164] The conflict correction level signal is a correction signal containing "maintain," "escalate," and "degrade" instructions. Its level distribution characteristics need to be analyzed to reflect recent fatigue state changes. In practice, the distribution of instruction types in the conflict correction signal over the past 10 minutes is first statistically analyzed: the percentage of "maintain" instructions (reflecting state stability), the percentage of "escalate" instructions (reflecting risk accumulation trends), and the percentage of "degrade" instructions are recorded. Combining the state assessment value ranges corresponding to each instruction (e.g., "escalate" often corresponds to adjustments from warning to intervention levels), a feature vector of the current level information is generated, including parameters such as level transition frequency and state interval distribution density. This distribution vector is visualized using a histogram or probability density function, intuitively showing the concentration trend and dispersion of recent level adjustments, providing a quantitative basis for comparison with historical trends.

[0165] The core of similarity measurement is to quantify the degree of matching between current level information and historical state trends. In practice, a dynamic time warping algorithm is used: aligning the feature vectors of historical driving trend data with those of the current level information, calculating the optimal matching path between the two on the time axis; and using the reciprocal of the accumulated path distance as the similarity score. For example, if historical trends show a "stable-slightly fluctuating" driver fatigue state, and the instruction types of the current level information (maintain 70%, upgrade 25%, downgrade 5%) are highly consistent with the historical high-frequency pattern, then the similarity score is close to 1; if the current distribution shows an abnormal "frequent upgrades," the similarity score decreases significantly. This similarity value, after normalization to eliminate the influence of dimensions, directly reflects the degree of fit between the current level adjustment and the long-term trend, providing a key input for subsequent continuity confidence calculations.

[0166] The similarity score reflects the degree of matching between the current level information and the historical trend, and needs to be further converted into a confidence coefficient reflecting the continuity of level adjustments. In specific implementation, firstly, by setting a similarity threshold range, the similarity score is divided into three levels: "highly consistent", "moderately consistent", and "lowly consistent". When the similarity score is close to 1, it is determined that the current level distribution is highly consistent with the historical trend, and the continuity risk is low; when the similarity score is in the medium range, it is determined to be a partial match, and the continuity risk is moderate; when the similarity score is low, it is determined to be a trend deviation, and the continuity risk is high.

[0167] Similarity levels are mapped to transfer confidence coefficients: highly consistent confidence coefficients represent strong continuity in level adjustments, moderately consistent coefficients correspond to medium continuity, and lowly consistent coefficients correspond to weak continuity. These coefficients are smoothed using a sliding window averaging filter to eliminate interference from single-time similarity fluctuations. The final output is a transfer confidence coefficient reflecting the reliability of the continuity of level adjustments, providing confidence weights for subsequent calculations of smoothed transfer amounts. This ensures that level adjustments conform to both current conflict detection results and long-term driving patterns.

[0168] The calculation of smooth migration amount needs to combine the migration confidence coefficient with the adjustment requirements of the current level information. In specific implementation, the instruction type characteristics of the current level information are analyzed first: if the proportion of "upgrade" instructions is significantly higher than the historical average, it indicates that the fatigue state is accumulating, and the migration direction is "upward adjustment"; if the proportion of "downgrade" instructions is higher, it indicates that there is a need for misjudgment correction, and the migration direction is "downward adjustment"; if the two are balanced, the current level is maintained.

[0169] The migration amplitude is adjusted based on the migration confidence coefficient: at high confidence coefficients, the migration amplitude is calculated as a higher proportion of the historical average adjustment (enhancing the adjustment strength to match strong continuity); at medium confidence coefficients, the migration amplitude is calculated as a standard proportion of the historical average adjustment (normal adjustment to balance continuity and current risk); at low confidence coefficients, the migration amplitude is calculated as a lower proportion of the historical average adjustment (weakening the adjustment to avoid misjudgment). A smoothed migration amount is obtained through weighted summation. This migration amount is then superimposed on the current conflict correction level signal to generate the final level migration parameter. This ensures that the level adjustment reflects both the immediate risk of current conflict detection and consistency with historical trends, providing a more robust input for subsequent auxiliary urgency identification.

[0170] This embodiment effectively improves the accuracy and continuity of fatigue driving warning level adjustments through multi-dimensional historical data verification and smooth migration parameter calculation. Specifically, the fitting of historical driving trend data captures the long-term evolution of driver fatigue state, providing a stable benchmark reference for current level adjustments; level dispersion analysis and similarity measurement processing identify the degree of matching between the distribution of the level to be adjusted in the near future and historical trends, avoiding misjudgments caused by short-term fluctuations; the calculation of the continuity confidence coefficient combines historical patterns and current risks, quantifying the reliability of level adjustments and reducing unfounded level jumps; and the dynamic calculation of the smooth migration amount balances immediate risks and long-term trends, ensuring that level adjustments reflect both current conflict detection results and are consistent with the driver's long-term driving habits.

[0171] In one embodiment, a hierarchical decision-making process is constructed based on the level of driving assistance to obtain multi-level driving assistance decisions, including:

[0172] The classification of driver assistance levels is the foundation for multi-level decision-making, and its core basis is the correlation between fatigue driving risk and typical characteristics of driving scenarios. In practice, the distribution patterns of driver state assessment values ​​(combining physiological indicators such as eye closure duration and head tilt frequency with vehicle dynamic data such as steering fluctuations and lane departure) are first analyzed. Combined with traffic safety requirements and actual driving scenarios, the state assessment values ​​are divided into three core intervals: the first-level assistance interval corresponds to low-risk states (such as being alert or mildly fatigued), the second-level assistance interval corresponds to medium-risk states (such as moderately fatigued), and the third-level assistance interval corresponds to high-risk states (such as severely fatigued).

[0173] The boundaries of the interval divisions were determined through historical fatigue driving data statistics and real-world scenario verification, ensuring that each interval matches the driver's actual reaction ability and vehicle control requirements. For example, the low-risk interval covers the low-fluctuation area of ​​the state assessment value, corresponding to scenarios where the driver is focused and operates smoothly; the medium-risk interval covers the medium-fluctuation area, corresponding to scenarios where attention begins to wander and minor operational abnormalities occur; and the high-risk interval covers the high-fluctuation area, corresponding to scenarios where attention is severely wandered and the risk of loss of control increases significantly. After the division, the three auxiliary intervals correspond to three core auxiliary strategies: "basic monitoring," "active early warning," and "emergency intervention," providing a structured framework for subsequent weight allocation and threshold calculation.

[0174] The core objective of the first-level auxiliary zone (low risk) is to maintain continuous monitoring of the driver's condition; therefore, weight allocation focuses on the differentiated configuration of the "monitoring" attribute. In practice, the distribution characteristics of the condition assessment values ​​within the first-level zone are first analyzed (e.g., stability concentrated in low-fluctuation areas). Combined with the priority of monitoring functions (e.g., information recording, lightweight prompts), different weight coefficients are assigned to each sub-zone of the condition assessment values. For example, higher weights are assigned to the portion of the condition assessment values ​​close to the safety threshold to strengthen basic condition recording; medium weights are assigned to the portion in the middle of the zone to activate regular prompts (e.g., on-screen text alerts); and lower weights are assigned to the portion near the zone boundary to avoid false triggering in low-risk areas.

[0175] Weight allocation is achieved through a preset weight mapping strategy, which dynamically adjusts based on the functional requirements of low-risk intervals in historical data (e.g., monitoring most of the time and alerts for a small portion). The resulting first-level control weight set contains multiple sub-weight values, each corresponding to a different position of the state evaluation value within the first-level interval. This ensures a balance between monitoring and alert functions, provides a weight basis for low-risk levels in subsequent multi-level fusion, and avoids neglecting medium- and high-risk changes due to excessive focus on low-risk areas.

[0176] The core objective of the second-level auxiliary zone (medium risk) is proactive early warning. This requires adjusting the warning threshold based on real-time conditions and environmental factors to adapt to different driving scenarios. Historical distribution characteristics of state assessment values ​​within the second-level zone (such as the central tendency of moderate fluctuation areas) are obtained, and dynamic thresholds are calculated by combining these with current environmental parameters (such as vehicle speed, light intensity, and road conditions). For example, in high-speed scenarios, driver reaction time is shorter, so the warning threshold needs to be adjusted towards lower risk (reducing the risk of false alarms); in low-speed scenarios, reaction time is longer, so the threshold can be adjusted towards higher risk (avoiding missed alarms).

[0177] Dynamic threshold calculation uses a sliding window to statistically analyze the mean and fluctuation characteristics of the current state assessment value. Environmental parameters are then used to correct the mean deviation, generating a threshold adapted to the current scenario. For example, the mean and standard deviation of the current window state assessment value reflect the concentration of medium-risk areas, and vehicle speed parameters correct the mean deviation, making the threshold more closely match actual driving needs. This threshold is updated in real time to ensure it matches the medium-risk characteristics of the second-level interval. The resulting second-level control threshold set contains multiple dynamically adjusted threshold points, covering different state positions within the second-level interval, providing precise triggering criteria for proactive warnings and balancing the timeliness and accuracy of warnings.

[0178] The fusion of multi-level control parameter sets requires a comprehensive consideration of the multi-dimensional needs of low-risk monitoring, medium-risk early warning, and high-risk intervention. This is achieved through hierarchical weighting and complementary characteristics to optimize the parameters. Specifically, the low-risk monitoring weights (emphasizing state stability) of the first-level control weight set are first used as basic parameters and then fused with the medium-risk dynamic thresholds (emphasizing scenario adaptability) of the second-level control threshold set using a linear weighting method to generate intermediate parameters reflecting the risk transition zone. Subsequently, the high-risk emergency priority (emphasizing intervention timeliness) of the third-level control priority sequence is used as a correction factor to dynamically adjust the intermediate parameters, ensuring that control parameters in high-risk scenarios prioritize emergency responses.

[0179] During the integration process, the weight allocation of parameters at each level is based on the risk distribution of driving scenarios: in low-risk scenarios, the first level has a higher weight to ensure the stability of basic monitoring; in medium-risk scenarios, the second level threshold has a higher weight to adapt to dynamic early warning needs; in high-risk scenarios, the third level priority dominates to enhance the timeliness of emergency intervention. The resulting multi-level control parameter set contains hierarchical control parameters, which not only preserves the monitoring continuity in low-risk scenarios but also takes into account the early warning accuracy in medium-risk scenarios and the urgency of intervention in high-risk scenarios, providing comprehensive parameter support for the construction of subsequent auxiliary commands.

[0180] The core of auxiliary command mapping is to convert multi-level control parameters into specific execution commands. Its mapping rules are based on the direct correlation between fatigue driving risk and vehicle control needs. In practice, firstly, "basic monitoring" commands (such as continuously recording driver physiological data and updating status logs) are generated based on the low-risk weight set in the multi-level control parameter set; secondly, "active warning" commands (such as tiered audio-visual prompts and instrument panel status displays) are generated based on the medium-risk threshold set; and finally, "emergency intervention" commands (such as limiting maximum speed and activating lane keeping assist) are generated based on the third-level priority sequence.

[0181] The command mapping must satisfy a positive correlation between risk level and command intensity: low risk corresponds to "hint-level" commands, medium risk corresponds to "alert-level" commands, and high risk corresponds to "mandatory-level" commands (such as automatic deceleration and partial control takeover). A preset command priority table ensures the execution order of multi-level commands: emergency intervention commands take precedence over active warning commands, and active warning commands take precedence over basic monitoring commands, avoiding execution confusion caused by command conflicts. The final multi-level driver assistance decision includes a hierarchical command set, with each command labeled with its corresponding driver assistance level and execution conditions, directly guiding the execution of the vehicle's active safety functions.

[0182] This embodiment effectively improves the accuracy and scenario adaptability of fatigue driving warnings through the hierarchical construction and parameter fusion of multi-level driving assistance decision-making. The division of driving assistance levels into intervals structures the state assessment values ​​into three risk levels: low, medium, and high, providing a clear decision-making framework for different risk scenarios. The first-level control weight set strengthens the continuous monitoring capability of low-risk areas through differentiated weight allocation, avoiding the hidden danger of neglecting basic conditions due to excessive focus on high risks. The second-level control threshold set adjusts the warning triggering conditions in conjunction with dynamic environmental parameters, ensuring timely and accurate warnings in medium-risk scenarios and reducing the risk of false alarms and missed alarms. The third-level control priority sequence ensures the priority execution of intervention commands in high-risk scenarios through urgency ranking, improving the timeliness of critical operations. The fusion calculation of multi-level control parameter sets integrates the multi-dimensional needs of monitoring, warning, and intervention, balancing the control intensity of different risk levels. The auxiliary command mapping converts parameters into specific execution actions, directly guiding the precise implementation of vehicle active safety functions.

[0183] Reference Figure 2 As shown, the present invention also provides an autonomous driving assistance decision-making system based on multi-source information fusion, applicable to any of the above-mentioned autonomous driving assistance decision-making methods based on multi-source information fusion, comprising:

[0184] The data acquisition module is used to acquire driver status data and vehicle dynamic data, perform behavior analysis and feature analysis, and obtain behavioral feature data and dynamic feature data.

[0185] The analysis module is used to acquire environmental perception data, integrate and calculate behavioral and dynamic feature data, and obtain a state assessment value.

[0186] The association module is used to identify the execution level based on the status assessment value to obtain the driving assistance level;

[0187] The processing module is used to construct hierarchical decisions based on the level of driving assistance, resulting in multi-level driving assistance decisions.

[0188] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the system and each module described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0189] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for assisting decision-making in autonomous driving by fusing multi-source information, characterized in that, include: The system acquires real-time driver status data and vehicle dynamic data from onboard sensors, performs behavioral and feature analysis, and obtains behavioral and dynamic feature data. Acquire environmental perception data, integrate and calculate the behavioral feature data and the dynamic feature data to obtain a state evaluation value; The execution level is identified based on the state assessment value to obtain the driving assistance level; Based on the aforementioned driving assistance levels, a hierarchical decision-making process is constructed to obtain multi-level driving assistance decisions. The process involves acquiring real-time driver status data and vehicle dynamic data from onboard sensors, performing behavioral and feature analysis to obtain behavioral and dynamic feature data, including: Visual signal decomposition was performed on the driver state data to obtain eyelid movement trajectory sequence and head posture angle change sequence; The duration of eye closure is recorded by analyzing the eyelid movement trajectory sequence to obtain the duration of eye closure. The pitch angle fluctuation amplitude of the head posture angle change sequence is statistically analyzed to obtain the head forward tilt frequency value; Behavioral analysis is performed based on the duration of eye closure and the frequency of head tilting forward to obtain the behavioral characteristic data; The vehicle dynamic data is obtained by collecting steering wheel angle time and analyzing lane departure. The dynamic feature data obtained by collecting steering wheel angle time and analyzing lane departure from the vehicle dynamic data includes: The vehicle dynamic data is analyzed to obtain the steering wheel turn time by analyzing the steering wheel signal. The steering wheel turning time is quantized to obtain steering fluctuation characteristic quantities; Lane line coordinates are extracted from the vehicle dynamic data to obtain lane lateral offset data; Based on the steering fluctuation characteristic, the lane lateral offset data is analyzed for cointegration relationship to obtain the directional offset control coefficient. The lateral control level is obtained by performing a stability-level mapping on the directional offset control coefficients. The dynamic feature data is obtained by integrating the steering fluctuation characteristic and the lateral control level. The process of acquiring environmental perception data, integrating and calculating the behavioral feature data and the dynamic feature data to obtain a state evaluation value includes: The environmental perception data is analyzed for road structure and identified for light intensity to obtain road structure parameters, obstacle distribution parameters and light flux coefficient; The behavioral feature data is subjected to pattern recognition and visual compensation processing to generate behavioral pattern parameters; Based on the road structure parameters and the obstacle distribution parameters, the behavior pattern parameters are correlated with the environment to obtain behavior environment correlation parameters; Based on the dynamic feature data and the light flux coefficient, the behavioral environment correlation parameters are constrained and adjusted to obtain a dynamic correction factor; The state evaluation value is obtained by proportionally fusing the behavioral environment-related parameters and the dynamic correction factor.

2. The autonomous driving assisted decision-making method based on multi-source information fusion according to claim 1, characterized in that, The step of adjusting the behavioral environment correlation parameters based on the dynamic feature data and the luminous flux coefficient to obtain a dynamic correction factor includes: The luminous flux coefficient is used to perform illuminance intensity range compensation calculation to generate illuminance compensation coefficient; The vehicle motion trajectory is analyzed from the dynamic feature data to obtain the lateral displacement fluctuation. The lateral displacement fluctuation is optically corrected according to the illumination compensation coefficient to generate an illumination correction value. Based on the illumination correction amount, displacement-related constraints are applied to the behavioral environment-related parameters to obtain preliminary correction parameters; The preliminary correction parameters are subjected to boundary safety verification to obtain the safety boundary parameters; The dynamic correction factor is obtained by dynamically synthesizing the initial correction parameter based on the safety boundary parameter.

3. The autonomous driving assisted decision-making method based on multi-source information fusion according to claim 1, characterized in that, The step of identifying the execution level based on the state evaluation value to obtain the driving assistance level includes: The state evaluation value is divided into intervals based on a preset security level threshold to obtain security interval information; The volatility of the state assessment value is statistically analyzed to obtain a state stability index; Based on the state stability index, conflict detection is performed on the safe interval information to obtain a conflict correction level signal; Historical driving assistance data is obtained through a preset system storage unit, and the continuity of the conflict correction level signal is verified to obtain the level migration parameters. Based on the level migration parameters, the conflict correction level signal is downgraded and verified to obtain the security level output. The safety level output is assisted in identifying the urgency level based on the vehicle dynamic data to obtain the driving assistance level.

4. The autonomous driving assistance decision-making method based on multi-source information fusion according to claim 3, characterized in that, The process of acquiring historical driving assistance data through a preset system storage unit, verifying the continuity of the conflict correction level signal, and obtaining level transition parameters includes: The historical driving assistance data is subjected to trend fitting to obtain historical driving trend data; The conflict correction level signal is analyzed for level discreteness to generate the current level information; The current level information is processed by similarity measurement based on the historical driving trend data, and the level trend similarity is output. Based on the similarity of the grade trends, a continuity confidence calculation is performed on the current grade information to obtain the migration confidence coefficient; The smoothing migration amount is calculated based on the migration confidence coefficient to obtain the level migration parameter.

5. The autonomous driving assistance decision-making method based on multi-source information fusion according to claim 1, characterized in that, The hierarchical decision-making based on the driving assistance level results in multi-level driving assistance decisions, including: The driving assistance levels are divided into level ranges to obtain the first-level assistance range, the second-level assistance range, and the third-level assistance range; The state evaluation value is weighted according to the first-level auxiliary interval to obtain the first-level control weight set; Based on the second-level auxiliary interval, the state evaluation value is dynamically thresholded to obtain the second-level control threshold set; The state evaluation values ​​are prioritized according to the third-level auxiliary interval to obtain the third-level control priority sequence. The first-level control weight set, the second-level control threshold set, and the third-level control priority sequence are fused and calculated to obtain a multi-level control parameter set; The multi-level driving assistance decision is obtained by constructing an auxiliary instruction mapping based on the multi-level control parameter set.

6. An autonomous driving assistance decision-making system based on multi-source information fusion, characterized in that, The autonomous driving assistance decision-making method applied to the multi-source information fusion of any one of claims 1-5 includes: The acquisition module is used to acquire driver status data and vehicle dynamic data, perform behavior analysis and feature analysis, and obtain behavior feature data and dynamic feature data. The analysis module is used to acquire environmental perception data, integrate and calculate the behavioral feature data and the dynamic feature data to obtain a state evaluation value; The association module is used to identify the execution level based on the state evaluation value to obtain the driving assistance level; The processing module is used to construct hierarchical decisions based on the driving assistance level to obtain multi-level driving assistance decisions.

Images