A cockpit safety strategy calibration method and system based on pilot status monitoring

Abstract

This invention relates to the field of intelligent connected vehicle safety technology, and discloses a cockpit safety strategy calibration method and system based on driver state monitoring. The method includes: collecting multimodal monitoring data of the driver, extracting state features, and performing multimodal feature fusion; collecting driving scenario parameters in real time and quantifying the risks of the driving scenario; constructing a driver individual feature model for new users through transfer learning and extracting individual feature parameters; calculating and classifying a comprehensive risk assessment value; adaptively calibrating intervention thresholds and intensity parameters; executing safety intervention strategies and performing multidimensional effect evaluation; and optimizing the adaptive safety strategy configuration. This invention achieves differentiated strategy calibration for different driver sensitivities and behavioral characteristics, avoiding the problems of over-intervention and under-intervention caused by fixed thresholds.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle safety technology, and more specifically, to a cockpit safety strategy calibration method and system based on driver status monitoring. Background Technology

[0002] With the rapid development of intelligent connected vehicle technology, driver status monitoring systems have become an important technical means to improve driving safety. Existing driver status monitoring systems mainly collect physiological and behavioral data of drivers through devices such as cameras and sensors, identify abnormal states such as fatigue and distraction, and trigger corresponding safety warnings or intervention measures.

[0003] However, existing technologies have significant technical shortcomings: the safety strategy calibration uses a fixed threshold mode, which cannot adapt to the individual differences of different drivers, resulting in over-intervention for highly sensitive drivers and under-intervention for less sensitive drivers; it lacks the ability to dynamically perceive and adaptively adjust the risks of driving scenarios, failing to lower the intervention threshold in advance in high-risk scenarios and potentially causing unnecessary interference in low-risk scenarios; the strategy calibration parameters are fixed and cannot be continuously optimized based on the actual intervention effect, resulting in the system performance not improving with usage time and data accumulation.

[0004] These problems seriously affect the practicality and effectiveness of driver condition monitoring systems, and there is an urgent need for a safety strategy calibration method that can achieve personalization and scenario adaptation. Summary of the Invention

[0005] This invention provides a cockpit safety strategy calibration method and system based on driver status monitoring, which solves the technical problem that safety strategy calibration in related technologies lacks personalized adaptation and scenario adaptation capabilities.

[0006] This invention provides a cockpit safety strategy calibration method based on driver status monitoring, comprising: Multimodal monitoring data of the driver is collected, state features are extracted and multimodal feature fusion is performed to obtain the driver state feature vector; Real-time collection of driving scenario parameters; quantification of driving scenario risks based on driving scenario parameters; and obtaining a comprehensive scenario risk index. Based on the driver state feature vector, a driver individual feature model for new users is constructed through transfer learning, and individual feature parameters are extracted to obtain the driver individual feature parameter set; Based on the driver's state feature vector, the scenario comprehensive risk index and the driver's individual feature parameter set, the comprehensive risk assessment value is calculated and classified to obtain the comprehensive risk assessment value and risk level classification results. Based on the risk level classification results, the comprehensive risk index of the scenario, and the driver's individual characteristic parameter set, the intervention threshold and intensity parameters are adaptively calibrated to obtain the adaptive safety strategy configuration; Based on adaptive security policy configuration, the intervention threshold and comprehensive risk assessment value are compared, the security intervention policy is executed and a multi-dimensional effect evaluation is performed to obtain an intervention effect evaluation report and a comprehensive effect score. Based on the intervention effect evaluation report and the comprehensive effect score, the adaptive security policy configuration is optimized to obtain the optimized adaptive security policy configuration and performance improvement report.

[0007] In a preferred embodiment, facial visual data of the driver is collected and eye and head features are extracted. Blink frequency, proportion of prolonged eye closure, gaze deviation angle and head posture angle parameters are calculated to obtain facial visual features. Collect steering wheel operation data and extract driving behavior features, calculate the standard deviation of steering wheel angle, peak angular velocity, grip torque fluctuation amplitude and spectral energy distribution to obtain steering wheel operation features; Collect vehicle motion state data and extract driving performance characteristics, calculate vehicle speed fluctuation coefficient, root mean square value of lateral offset distance, number of lane departures, vehicle distance change rate and pedal operation frequency to obtain driving performance characteristics; A multimodal feature fusion network is constructed, which maps facial visual features, steering wheel operation features, and driving performance features into hidden layer representations and then concatenates them. The network is then input into a bidirectional long short-term memory network for temporal modeling and outputs a driver state feature vector.

[0008] In a preferred embodiment, the real-time acquisition of driving scenario parameters and the quantification of driving scenario risks based on these parameters include: Traffic flow information is collected using millimeter-wave radar and cameras to calculate vehicle density, distance to the vehicle in front, and collision time, and to calculate traffic density risk scores. The road curvature is calculated based on steering wheel angle and vehicle speed data, the slope angle is obtained from the inertial measurement unit, the lane width is calculated through lane line detection, and the road geometric risk score is obtained. By integrating traffic density risk score, road geometry risk score, and vehicle speed risk coefficient, a comprehensive risk index for the scenario is calculated.

[0009] In a preferred embodiment, the step of constructing a driver individual feature model for new users through transfer learning and extracting individual feature parameters includes: Load a pre-trained baseline model of general driver features; Collect initial driving data from new users and perform feature annotation; continuously collect status monitoring data; automatically annotate the collected driver status feature vectors; and distinguish between normal state samples and abnormal state samples. Update the weight parameters of the last two fully connected layers of the general driver feature baseline model and perform parameter updates to obtain the individual driver feature model; Driver individual characteristic parameters and characteristic baselines are extracted. The characteristic baseline vector and covariance matrix are calculated from the normal state samples. The sensitivity parameters are evaluated by analyzing the driver's response to the test prompts. The fatigue accumulation parameters are obtained by fitting the fatigue accumulation model, and the driver individual characteristic parameter set is obtained.

[0010] In a preferred embodiment, the calculation and classification of the comprehensive risk assessment value includes: Mahalanobis distance was used to calculate the deviation of the driver's state from the individual baseline, and then normalized. The adjustment coefficient is calculated based on the individual sensitivity parameters to obtain the adjusted driver state risk value; A weighted fusion method is used to calculate the comprehensive assessment value of driver state risk and scenario risk; By setting risk thresholds, the comprehensive risk assessment value is divided into low risk, medium risk, high risk, and emergency risk.

[0011] In a preferred embodiment, the adaptive calibration intervention threshold and intensity parameters include: A three-tiered intervention strategy system was established based on the risk level classification results; The intervention threshold is adjusted according to the individual driver's sensitivity. Based on the sensitivity parameters in the set of individual driver characteristic parameters, the threshold adjustment coefficient is calculated, and the baseline threshold is adjusted in a personalized manner. The intervention threshold is dynamically adjusted based on the scenario risk index. The scenario risk adjustment coefficient is calculated based on the scenario comprehensive risk index, and the personalized intervention threshold is dynamically adjusted. The intervention intensity parameters are calibrated and strategy configurations are generated. The intensity parameters of each level of intervention measures are calibrated based on the comprehensive risk assessment value.

[0012] In a preferred embodiment, comparing the intervention threshold and the comprehensive risk assessment value, implementing the safety intervention strategy, and conducting a multidimensional effect evaluation includes: The comprehensive risk assessment value is compared with the intervention thresholds at each level to generate corresponding level intervention control instructions; By analyzing facial images, steering wheel operations, and seat pressure data, driver response behavior is detected, and response delay and quality are recorded. State feature vectors were extracted at multiple time points after intervention, and the state improvement index was calculated. Analyze the changes in lane keeping, speed stability, and following safety before and after the intervention, and calculate the improvement in safety margin. The overall intervention effect score is calculated based on response behavior, state improvement effect, and safety margin improvement.

[0013] In a preferred embodiment, optimizing the adaptive security policy configuration includes: Construct an intervention effect dataset and perform data preprocessing; Construct a policy optimization model based on reinforcement learning, model the policy labeling problem as a reinforcement learning problem, and build a policy network and a value network; To ensure optimization safety, a safety constraint mechanism is introduced, which sets upper and lower limits for the intervention threshold, adds a constraint projection layer to the output layer of the policy network, and introduces a constraint penalty term into the optimization objective function. Perform parameter optimization and verify the optimization effect. Use the near-end policy optimization algorithm to iteratively update the policy parameters and evaluate the performance metrics on the test set. Deploy the optimized strategy model and perform canary release verification, then use a canary release strategy to test on a small number of vehicles.

[0014] In a preferred embodiment, the multimodal feature fusion network processes visual features, operational features, and performance features respectively. The visual feature branch maps facial visual features to hidden layer representations, the operational feature branch maps steering wheel operation features to hidden layer representations, and the performance feature branch maps driving performance features to hidden layer representations.

[0015] In a preferred embodiment, a cockpit safety strategy calibration system based on driver status monitoring is used to execute the above-described cockpit safety strategy calibration method based on driver status monitoring, including: The data acquisition module is used to collect multimodal monitoring data of the driver, extract state features, and perform multimodal feature fusion to obtain the driver state feature vector; The scene perception module is used to collect driving scene parameters in real time, quantify the risk of the driving scene based on the driving scene parameters, and obtain the comprehensive scene risk index. The individual modeling module, based on the driver state feature vector, constructs a driver individual feature model for new users through transfer learning and extracts individual feature parameters to obtain a set of driver individual feature parameters; The risk assessment module calculates and classifies the comprehensive risk assessment value based on the driver's state feature vector, the scenario comprehensive risk index, and the driver's individual feature parameter set, thus obtaining the comprehensive risk assessment value and risk level classification results. The strategy calibration module adaptively calibrates the intervention threshold and intensity parameters based on the risk level classification results, the comprehensive risk index of the scenario, and the driver's individual characteristic parameter set, to obtain an adaptive safety strategy configuration. The intervention execution module, based on adaptive security policy configuration, compares intervention thresholds and comprehensive risk assessment values, executes security intervention strategies, performs multi-dimensional effect evaluation, and obtains an intervention effect evaluation report and a comprehensive effect score. The strategy optimization module optimizes the adaptive security strategy configuration based on the intervention effect evaluation report and the comprehensive effect score, resulting in an optimized adaptive security strategy configuration and a performance improvement report.

[0016] The beneficial effects of this invention are as follows: This invention effectively solves the problems of lack of personalization and scenario adaptability in safety strategy calibration in the prior art by using a technical solution of multi-source data fusion, individual feature adaptive learning and scenario risk perception. Through rapid modeling of individual features based on transfer learning, the system can build a personalized feature model after only 3 driving sessions for new users, realize differentiated strategy calibration for different drivers’ sensitivity and behavioral characteristics, and avoid the problems of over-intervention or under-intervention caused by fixed thresholds.

[0017] This invention employs dynamic scenario risk assessment and forward-looking risk prediction technologies, which can quantify scenario risks in real time based on multi-dimensional parameters such as traffic density, road geometry, and vehicle speed. It also combines high-precision map information to identify high-risk road sections in advance, enabling preventative strategy adjustments. Compared to the static calibration methods of existing technologies, the adaptive calibration method of this invention can reduce the intervention threshold in advance in high-risk scenarios, thereby improving the initiative and effectiveness of safety protection.

[0018] This invention establishes a policy continuous optimization mechanism based on reinforcement learning. Through multi-dimensional intervention effect evaluation and causal inference analysis, the policy parameters achieve self-evolution. The system performance continuously improves with data accumulation, the effective intervention rate is improved, and the false alarm rate is reduced, providing an efficient, accurate, and personalized technical solution for driver status monitoring of intelligent connected vehicles. Attached Figure Description

[0019] Figure 1 This is a flowchart of the main process of a cockpit safety strategy calibration method based on driver status monitoring in this invention. Figure 2 This is a detailed flowchart of a cockpit safety strategy calibration method based on driver status monitoring in this invention; Figure 3 This is a block diagram of a cockpit safety strategy calibration system based on driver status monitoring in this invention. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses a cockpit safety strategy calibration method based on driver state monitoring, such as Figures 1 to 2 As shown, it includes the following steps: Step 1: Collect multimodal monitoring data of the driver, extract state features and perform multimodal feature fusion to obtain the driver state feature vector; Based on multimodal data collected by an onboard sensor network, including driver facial images, steering wheel operation behavior, and vehicle motion status, deep neural networks are used to extract and fuse features. The specific steps include: Step 1.1: Collect driver's facial visual data and extract eye and head features; An infrared camera is installed in the center of the vehicle's dashboard, with the camera's field of view covering the driver's face area. The frame rate is set to 30 frames per second. The system acquires driver's facial image frames at 30 time points per second and transmits the image data to the onboard computing unit for processing.

[0022] A facial landmark detection algorithm based on a deep convolutional neural network is used to analyze each frame of the image, locating 68 facial landmarks, with a focus on extracting the coordinates of landmarks in the eye region. The vertical distance between the landmarks of the upper and lower eyelids is calculated to determine the eye's closed state; when the vertical distance is less than a threshold for the open eye state, the eye is considered closed. The number of times the eye closes within a consecutive 30-second time window is counted to calculate the blink frequency. The number of frames for each eye closure is counted, and cases where the closure duration exceeds a threshold are marked as prolonged eye closure, calculating the proportion of prolonged eye closure. The pupil landmark coordinates are then used as the basis for further analysis. The system calculates the pupil center position by tracking the pupil center's displacement trajectory between consecutive frames, calculates the gaze direction vector, compares the gaze direction vector with the normal gaze direction in front of the vehicle, and calculates the gaze deviation angle. When the gaze deviation angle exceeds a set threshold, it is determined to be a distracted state. Based on facial key points, the system calculates the head posture in three-dimensional space, including pitch angle, yaw angle, and roll angle. The three angle parameters of the head posture are obtained through perspective transformation and geometric calculation. When the pitch angle continues to drop and exceeds a set threshold, it is determined to be a head-down state. When the yaw angle continues to deviate from the front and exceeds a set threshold, it is determined to be a head-tilt state.

[0023] Six parameters, including blinking frequency, percentage of prolonged eye closure, angle of gaze deviation, and three angular parameters of head posture, are used as facial visual features for output.

[0024] Step 1.2: Collect steering wheel operation data and extract driving behavior features; Angle and torque sensors are installed on the steering column to extract driving behavior characteristics. The angle sensor measures the steering wheel's rotation angle relative to the center position, with a measurement range of -720 degrees to +720 degrees and a sampling frequency of 100 Hz. The torque sensor measures the grip torque and steering torque applied by the driver to the steering wheel, with a sampling frequency of 100 Hz. The system collects steering wheel angle and torque data every 10 milliseconds.

[0025] Within a continuous 30-second time window, the standard deviation of the steering wheel angle is calculated. The standard deviation reflects the smoothness of the driver's steering wheel control; a larger standard deviation indicates more unstable steering wheel operation. The steering wheel angular velocity is calculated, which is the rate of change of the angle between adjacent sampling points. The peak value of the angular velocity within the time window is statistically analyzed, and the peak value reflects the frequency of the driver's rapid correction operations. The fluctuation amplitude of the steering wheel grip torque within the time window is calculated. The fluctuation amplitude parameter is obtained by calculating the variance of the torque signal. Under normal driving conditions, the driver maintains a stable grip on the steering wheel, and the grip torque fluctuation amplitude is small. Under fatigue or distraction, the driver's grip force is unstable, and the fluctuation amplitude increases.

[0026] Spectral analysis was performed on the steering wheel angle timing signal. The main frequency component of the signal was extracted using Fast Fourier Transform. Under normal driving conditions, steering wheel operation has a certain rhythm, and the main frequency is concentrated in a specific frequency band. Under abnormal conditions, the operation rhythm is disordered and the spectrum distribution changes. The energy distribution characteristics of the spectrum are extracted as a supplement to the driving behavior characteristics.

[0027] The steering wheel operation characteristics are output as four parameters: standard deviation of steering wheel angle, peak angular velocity, grip torque fluctuation amplitude, and spectral energy distribution.

[0028] Step 1.3: Collect vehicle motion state data and extract driving performance features; The vehicle's motion status data, including parameters such as vehicle speed, lateral acceleration, longitudinal acceleration, and yaw rate, is acquired in real time through the vehicle controller local area network bus, with a sampling frequency of 50 Hz. At the same time, the lateral offset distance of the vehicle relative to the lane centerline is obtained from the lane keeping assist system, and the longitudinal distance to the vehicle in front is obtained from the forward collision warning system.

[0029] Within a 30-second time window, the standard deviation and coefficient of variation of vehicle speed are calculated to reflect the stability of speed maintenance. Under normal driving conditions, drivers can maintain a relatively stable speed with minimal fluctuations. When fatigued or distracted, the ability to control speed decreases, and speed fluctuations increase. The root mean square value of the vehicle's lateral deviation distance is calculated to reflect lane keeping ability. The larger the root mean square value of the lateral deviation distance, the more obvious the lateral sway of the vehicle within the lane, and the worse the lane keeping ability. The number of times the vehicle touches the lane boundary line and the duration of the collision are recorded to identify lane departures. The minimum and average following distances to the vehicle in front are extracted from the forward distance data, and the rate of change of following distance is calculated. By analyzing the pattern of the rate of change of following distance, the driver's following stability is identified. Under normal conditions, drivers can maintain a safe following distance and adjust smoothly with the speed of the vehicle in front. Under abnormal conditions, the following distance is poorly controlled, resulting in frequent approaching or widening of the following distance. The operation frequency data of the brake pedal and accelerator pedal are collected, and the number of times the pedals are pressed and the average duration of the presses within the time window are statistically analyzed. Under fatigued conditions, the driver's ability to predict road conditions decreases, and the frequency and rhythm of braking and acceleration operations change.

[0030] Five parameters—vehicle speed fluctuation coefficient, root mean square value of lateral offset distance, number of lane departures, vehicle distance change rate, and pedal operation frequency—are used as driving performance characteristics output.

[0031] Step 1.4: Fuse multimodal features and extract high-level state representation vectors; Based on the facial visual features output in step 1.1, the steering wheel operation features output in step 1.2, and the driving performance features output in step 1.3, a total of 15 original feature parameters are used for feature fusion and high-level representation extraction using a deep neural network.

[0032] A multimodal feature fusion network is constructed, comprising three branches that handle visual features, operational features, and performance features, respectively. The visual feature branch uses a fully connected layer to map the 6-dimensional input to a 32-dimensional hidden layer representation; the operational feature branch maps the 4-dimensional input to a 16-dimensional hidden layer representation; and the performance feature branch maps the 5-dimensional input to a 16-dimensional hidden layer representation. The hidden layer representations of the three branches are concatenated to form a 64-dimensional fused feature.

[0033] The 64-dimensional fused features are input into a Bi-directional Long Short-Term Memory (BiLSTM) network for temporal modeling. The 30 fused feature vectors calculated per second over 30 consecutive seconds are arranged in chronological order to form a 30x64 temporal matrix. The BiLSTM network processes the temporal matrix simultaneously from both forward and backward directions to capture the evolution patterns and dependencies of features in the temporal dimension. The hidden layer of the BiLSTM network is set to 128 units, and the hidden state at the last moment is output as the temporal encoding result.

[0034] The temporal coding results are transformed by a fully connected layer and mapped to a driver state feature vector. This comprehensively encodes the driver's current state information and temporal evolution features in the three dimensions of vision, operation, and performance, and can robustly represent the driver's real-time state.

[0035] The parameters of the multimodal feature fusion network and the bidirectional long short-term memory network were obtained through pre-training on large-scale driving data. The pre-training dataset contains driving data from drivers of different ages, genders, and driving experiences in various scenarios, along with corresponding driver state annotations. The network is trained through supervised learning, enabling it to learn the mapping relationship from raw features to state representations. To improve the robustness of the network, data augmentation techniques are employed during training to add noise perturbations to the input features, simulating sensor measurement errors and environmental interference. An adversarial training mechanism is also used to generate adversarial examples and add them to the training set, enhancing the network's resistance to abnormal inputs.

[0036] In some embodiments, due to differences in data quality and reliability among different sensors, some sensors may malfunction or experience data loss. An adaptive feature fusion method based on an attention mechanism can be employed to dynamically adjust the fusion weights according to the real-time quality of each modality's data, thereby improving the system's fault tolerance. Specifically, an attention module is introduced in the feature fusion stage. This module calculates attention weights based on the inherent consistency and confidence of each modality's features. For each modality's feature vector, its Mahalanobis distance to the historical normal feature distribution is calculated. A smaller Mahalanobis distance indicates a more reliable current feature, which is assigned a higher attention weight; a larger Mahalanobis distance indicates a potential anomaly in the current feature, which is then weighted lower or excluded. After normalization using a softmax function, the attention weights are weighted and fused with the features of each modality to obtain an adaptively weighted fused feature. In this way, when the data quality of a certain sensor deteriorates, the system automatically reduces the impact of that modality's data, relying on other reliable sensor data for state judgment, ensuring the robustness and continuity of the overall system.

[0037] Step 2: Collect driving scenario parameters in real time, quantify the risk of the driving scenario based on the driving scenario parameters, and obtain the comprehensive risk index of the scenario; Step 2.1: Collect traffic flow status information ahead and calculate traffic density risk parameters; A millimeter-wave radar mounted on the vehicle's front bumper scans targets within a 200-meter range ahead. The radar operates at 77 GHz, with an angular resolution of 1 degree and a range resolution of 0.5 meters. It outputs a target list every 50 milliseconds, containing the distance, azimuth, and relative speed information for each detected target. The radar's target list is then clustered and tracked to identify vehicle targets in the lane ahead, filtering out stationary objects and non-vehicle targets on the roadside. Detection results for the same vehicle at consecutive time points are correlated to establish a target tracking trajectory, obtaining stable distance and speed information for each vehicle ahead.

[0038] The system counts the number of vehicles in the same lane within a 200-meter radius ahead, calculating the vehicle density parameter. Vehicle density is defined as the number of vehicles per unit distance; higher density indicates denser traffic flow and a greater potential collision risk. It also calculates the longitudinal distance between the current vehicle and the nearest vehicle ahead, defined as the distance to the vehicle in front. Based on the current speed and the distance to the vehicle in front, it calculates the collision time, which is equal to the distance to the vehicle in front divided by the relative speed between the two vehicles; a shorter collision time indicates a higher risk of a rear-end collision. Finally, it collects images of the road ahead using a forward-facing camera and employs a deep learning-based target detection algorithm to identify vehicles, pedestrians, non-motorized vehicles, and other traffic participants in the images as part of the traffic complexity. The system also counts the number of each type of target detected per unit time as a supplementary parameter to the traffic complexity.

[0039] The traffic density risk score is calculated by taking into account parameters such as vehicle density, distance to the vehicle in front, collision time, and traffic complexity. The weighting coefficients are determined by statistical analysis of historical data based on the correlation between each parameter and actual accident risk. The traffic density risk score is normalized to the range of 0 to 1, with a higher value indicating a higher traffic risk.

[0040] Step 2.2: Extract road geometric features and assess road risk parameters; The steering wheel angle and vehicle speed data are obtained from the vehicle chassis control system. Combined with the vehicle's steering geometry parameters, the curvature of the current road is calculated. The road curvature is defined as the reciprocal of the radius of curvature of the vehicle's driving trajectory. The larger the curvature, the sharper the curve and the higher the driving difficulty. The curvature over a continuous time period is filtered by a moving average to obtain a smooth road curvature curve.

[0041] Longitudinal and lateral acceleration data are obtained from the vehicle's inertial measurement unit. Under non-acceleration and non-braking conditions, the longitudinal acceleration reflects the longitudinal slope of the road. The longitudinal acceleration is negative when going uphill and positive when going downhill. The longitudinal slope angle of the road is calculated based on the longitudinal acceleration and gravitational acceleration.

[0042] Lane lines are detected using images from a forward-facing camera to identify the current lane's boundary lines. Based on the lane line image coordinates and camera calibration parameters, the lane width is calculated. Narrower lanes offer less driving margin for error and pose a higher risk.

[0043] If the vehicle is equipped with a high-precision map and positioning system, it can query the attribute information of the road segment ahead at the current location from the high-precision map database, including road grade, speed limit, number of lanes, intersection distribution, tunnel location, etc.; determine the road type based on the road grade and speed limit, with the risk level of highways, urban expressways, ordinary urban roads, and rural roads increasing in that order; and identify whether there are special road segments such as sharp bends, steep slopes, tunnels, and bridges ahead, as these road segments have a higher risk level than ordinary road segments.

[0044] If high-precision map data is unavailable, risk assessment relies entirely on local information perceived in real time by onboard sensors. Based on currently calculated parameters such as curvature, slope, and lane width, a rule-based approach is used to determine the road risk level. Curvature exceeding a set threshold is identified as a curve, slope exceeding a set threshold is identified as a slope, and lane width below a set threshold is identified as a narrow road, with corresponding risk scores increased accordingly.

[0045] Based on parameters such as road curvature, slope, lane width, risk level, and road type, a fuzzy logic reasoning method is used to calculate the road geometric risk score. The fuzzy logic system outputs a comprehensive risk assessment result based on the fuzzy membership degree of each parameter and preset reasoning rules. The fuzzy logic reasoning rule base includes the following specific rules: when the road curvature radius is greater than 500 meters, the slope angle is less than 3 degrees, and the lane width is greater than 3.7 meters, the road geometric risk is low; when the road curvature radius is between 200 and 500 meters, the slope angle is less than 3 degrees, and the lane width is between 3.2 and 3.7 meters, the road geometric risk is medium-low; when the road curvature radius is less than 200 meters or the slope angle is greater than 8 degrees, regardless of the lane width, the road geometric risk is at least medium-high; when the road curvature radius is less than 200 meters, the slope angle is greater than 8 degrees, and the lane width is less than 3.2 meters, the road geometric risk is high; when a tunnel section is detected, the risk level is increased by one level; when a bridge section is detected and crosswinds are strong, the risk level is increased by one level.

[0046] Step 2.3: Integrate multi-dimensional parameters and calculate the comprehensive risk index of the scenario; Based on the traffic density risk score output in step 2.1 and the road geometry risk score output in step 2.2, and combined with the current vehicle speed information, a scenario risk fusion model is used to calculate the comprehensive risk index.

[0047] Under the same traffic density and road conditions, the higher the speed, the greater the risk. The speed risk coefficient is calculated based on the ratio of the current speed to the road speed limit. When the speed is close to or exceeds the speed limit, the speed risk coefficient is close to or greater than 1. When the speed is significantly lower than the speed limit, the speed risk coefficient is less than 1.

[0048] A nonlinear fusion function is used to calculate the comprehensive risk index of a scenario. The fusion function considers the coupling effect of traffic risk, road risk, and vehicle speed risk. The fusion function is in the form of a weighted product of the three risk scores, and the weight coefficients are dynamically adjusted according to different scenario types. In highway scenarios, traffic density and vehicle speed have larger weights; in mountainous road scenarios, road geometric risk has a larger weight.

[0049] In some embodiments, since the risks of driving scenarios have spatiotemporally dynamic characteristics, considering only the instantaneous scenario information at the current moment may not be sufficient to reflect the risk evolution trend over a future period. A forward-looking scenario risk prediction method based on information about road segments ahead can be adopted. The aim is to identify high-risk road segments to be entered in advance, providing support for the calibration of preventative strategies. Specifically, when the vehicle is equipped with a high-precision map and positioning system, attribute information of road segments 10 to 15 kilometers ahead is queried from the high-precision map database, including the distribution of continuous curves, tunnel locations, bridge sections, accident-prone areas, etc. The arrival time of the vehicle at each road segment is estimated based on the current vehicle speed, constructing a road segment risk time series for the next 10 to 15 minutes. For identified high-risk road segments, the scenario risk index is gradually increased starting 5 to 10 minutes before the vehicle arrives at the segment, lowering the intervention threshold of safety strategies in advance and achieving preventative safety protection. For road segments with a history of frequent accidents, additional risk weighting coefficients are set based on the accident frequency and severity of the segment in the accident database to further strengthen the risk assessment value of the segment. By predicting scenarios and risks in advance, the system can intervene in advance when the driver's condition has not deteriorated significantly but is about to enter a high-risk section of road, thus avoiding accidents caused by the superposition of deterioration of condition and high risk of scenario.

[0050] Step 3: Based on the driver state feature vector, construct the driver individual feature model of the new user through transfer learning and extract the individual feature parameters to obtain the driver individual feature parameter set; Step 3.1: Load the pre-trained general driver feature baseline model; Download the pre-trained general driver feature baseline model from the cloud model library. It is trained on a large-scale driving dataset containing driving data from more than 1,000 drivers of different ages, genders and driving experience in various scenarios. The data duration for each driver is no less than 20 hours. The dataset includes the driver's state features, driving behavior features and corresponding state labels.

[0051] The general driver feature baseline model adopts a deep neural network structure. The input is the driver state feature vector from step 1, and the output is the driver's state classification result and state confidence. The general driver feature baseline model is trained through supervised learning to learn the mapping relationship from state features to state categories.

[0052] The universal driver feature baseline model learns the feature distribution patterns of various drivers under normal driving conditions, the feature change patterns under fatigue conditions, and the range of feature differences between individuals during training. This prior knowledge is encoded in the weight parameters of the universal driver feature baseline model and can serve as the initialization basis for personalized modeling of new users. The universal driver feature baseline model is then loaded into the memory of the onboard computing unit as the foundation model for rapid learning of individual features.

[0053] Step 3.2: Collect initial driving data from new users and perform feature annotation; When a new user uses the vehicle's driver status monitoring system for the first time, the system enters an individual characteristic learning mode. During the user's first three driving sessions, the system continuously collects driver status monitoring data, with each driving session lasting no less than 30 minutes, to ensure the sufficiency and diversity of the collected data.

[0054] During each driving session, the driver's state feature vector is extracted in real time and recorded at a frequency of once per second to form a time-series feature dataset. At the same time, the driving scenario information and vehicle state information at the corresponding moment are also recorded.

[0055] The collected driver state feature vectors are automatically labeled. In the first 10 minutes of the initial driving phase, assuming the driver is in a normal state, the feature data of this period is labeled as normal state samples. By analyzing the temporal changes of the features, time periods in which the features deviate significantly from the distribution of the initial phase are identified. Time periods with deviations exceeding a set threshold are labeled as suspected abnormal states. These time periods are further verified in conjunction with driving performance indicators. If abnormal driving behaviors such as lane departure or increased speed fluctuations occur, they are confirmed as abnormal state samples.

[0056] After three driving tests, approximately 5400 feature sample points were collected, of which about 80% were in normal condition and about 20% were in abnormal condition. The collected labeled dataset was divided into a training set and a validation set, with the training set accounting for 80% and the validation set accounting for 20%.

[0057] Step 3.3: Perform rapid adaptive personalization of the model based on meta-learning algorithms; A model-independent meta-learning algorithm is used to fine-tune the general driver feature baseline model. The core idea of ​​the meta-learning algorithm is to use only a small amount of personal data and a few gradient updates to quickly adapt to the feature distribution of new users on the basis of the general driver feature baseline model, so as to obtain the individual driver feature model.

[0058] The parameters of a baseline model for general driver features are used as initialization parameters and fine-tuned on a training set of new users. During fine-tuning, only the weight parameters of the last two fully connected layers are updated, while the parameters of the preceding convolutional layers and long short-term memory layers remain frozen. This strategy enables the rapid learning of personalized feature mappings for new users while preserving general feature extraction capabilities.

[0059] Mini-batch gradient descent was used for parameter updates, with a batch size of 32 and a learning rate of 0.001, for 10 training iterations. After each iteration, model performance was evaluated on the validation set, monitoring classification accuracy and loss function values. Training was stopped when validation set performance no longer improved, resulting in a driver individual feature model, which could more accurately identify the driver's state.

[0060] Step 3.4: Extract individual driver feature parameters and feature baselines; Based on personalized models and collected individual driving data, individual characteristic parameters of drivers are extracted. A feature baseline vector for the driver is extracted from normal-state samples. Statistical analysis is performed on the feature vectors of all normal-state samples, calculating the mean of features in each dimension to obtain a 128-dimensional feature baseline vector. This feature baseline vector represents the typical feature distribution center of the driver under normal driving conditions. The covariance matrix of the normal-state samples is calculated to characterize the fluctuation range of features in each dimension and the correlation between dimensions. The covariance matrix is ​​used to subsequently calculate the deviation of real-time features from the baseline.

[0061] Individual characteristic parameters of the driver are defined, reflecting the driver's acceptance and reaction intensity to safety interventions. Sensitivity levels are assessed by analyzing the driver's responses to system test prompts during the first three driving sessions. During the individual characteristic learning phase, the system issues mild test prompts at different times, such as low-volume voice reminders, recording the driver's response time and actions. Drivers with short response times and obvious actions are considered highly sensitive, while those with long response times or no obvious response are considered low sensitive. The sensitivity parameter is quantified into values ​​between 0.5 and 1.5, with 1 representing medium sensitivity, greater than 1 representing high sensitivity, and less than 1 representing low sensitivity.

[0062] The test prompts are triggered during periods when the driver is relatively stable and road conditions are safe. Specific triggering conditions are: the current speed is consistently between 70% and 90% of the speed limit for more than 30 seconds; there are no other vehicles or obstacles within 200 meters ahead; lane markings are clear and the vehicle is centered in the lane; the driver's face is facing forward and their eyes are open. The triggering method employs a progressive design. The first test uses the slightest visual prompt, displaying a light blue dot with a diameter of 2 cm in the upper right corner of the instrument panel for 3 seconds. The second test adds an auditory prompt, playing a gentle prompt at 50% of the normal voice navigation volume. The third test combines visual and auditory prompts, but the intensity is maintained at a level that does not affect normal driving. Each test is spaced at least 10 minutes apart to ensure the driver does not develop any preconceived notions. If an emergency or abnormal driver condition is detected during the test, the test is immediately stopped and normal monitoring mode is resumed.

[0063] The fatigue accumulation characteristic parameters of the driver were extracted. By analyzing the evolution curves of the state characteristics with driving time during three driving processes, a fatigue accumulation model was fitted. The fatigue accumulation model describes the relationship between the driver's fatigue level and driving time. There are individual differences in the fatigue accumulation rate among different drivers. An exponential function model was used to fit the fatigue accumulation curve to obtain the fatigue accumulation rate parameter, which is used to predict the driver's fatigue state at future moments.

[0064] The extracted feature baseline vectors, covariance matrices, sensitivity parameters, and fatigue accumulation parameters are integrated into a set of individual driver feature parameters and stored in the local database of the vehicle system as personalized input for subsequent risk assessment and strategy calibration.

[0065] In some embodiments, because a driver's physiological and psychological state may change over time, such as with age, accumulated driving experience, and changes in health, the initially constructed individual feature model may gradually lose accuracy. A continuous learning mechanism can be used to dynamically update the individual feature model, aiming to ensure that the model always accurately reflects the driver's current characteristics. Specifically, after completing the initial individual feature model construction, the system enters continuous learning mode. During each subsequent driving session, the system continues to collect driver state data, compares the newly collected data with the current individual feature baseline, and calculates the degree of drift in the feature distribution. When a certain amount of new data has accumulated, the model update process is triggered. The model update uses an incremental learning method, adding new data to the training set, recalculating the feature baseline and covariance matrix, and fine-tuning the personalized model parameters using an online gradient descent algorithm. To prevent new data from over-covering historical features, an experience playback mechanism is used, sampling some historical data during the update to maintain a balance between the old and new data. Through continuous learning, the individual feature model can adaptively adjust to follow the long-term changes in the driver, maintaining the model's long-term effectiveness.

[0066] Step 4: Based on the driver's state feature vector, the scenario comprehensive risk index, and the driver's individual feature parameter set, calculate the comprehensive risk assessment value and classify it to obtain the comprehensive risk assessment value and risk level classification results; Step 4.1, calculate the deviation of the driver's state from the individual baseline; Based on the driver state feature vector output in step 1 and the driver's feature baseline vector and covariance matrix extracted in step 3, calculate the difference vector between the current state feature vector and the feature baseline vector. Each dimension of the difference vector represents the deviation of that dimension's feature from the normal baseline.

[0067] Mahalanobis distance is used to measure the overall deviation of state characteristics from the baseline. The Mahalanobis distance is calculated using the difference vector and the inverse of the covariance matrix, yielding a scalar value through matrix operations. The calculated Mahalanobis distance is then normalized, mapping it to the interval between 0 and 1. Normalization employs a quantile mapping method based on historical data statistics. According to the deviation distribution of a large number of drivers, the current deviation value is mapped to the corresponding quantile. A larger deviation indicates that the driver's current state deviates more from the normal state, and the higher the abnormal risk. The normalized state deviation value is output as a quantitative indicator of the driver's inherent risk.

[0068] Step 4.2: Adjust the state risk weights based on individual sensitivity; Based on the driver individual characteristic parameter set obtained in step 3, which reflects the driver's risk tolerance and response characteristics to intervention, the weight of state deviation is adjusted according to the individual characteristic parameters. For highly sensitive drivers, the same state deviation corresponds to a higher subjective risk perception and should be given a higher weight in risk assessment; for low-sensitive drivers, the same state deviation corresponds to a lower subjective risk and the weight should be reduced accordingly.

[0069] The adjusted driver state risk value is calculated as the state deviation multiplied by the sensitivity adjustment coefficient. The sensitivity adjustment coefficient is obtained through a linear mapping based on the sensitivity parameter. When the sensitivity parameter is 1.0, the adjustment coefficient is 1.0; when the sensitivity parameter is greater than 1.0, the adjustment coefficient is greater than 1.0; and when the sensitivity parameter is less than 1.0, the adjustment coefficient is less than 1.0. The adjusted driver state risk value is output, which integrates the objective state deviation degree and subjective individual difference factors.

[0070] Step 4.3: Integrate driver state risk with scenario environment risk; Based on the scenario comprehensive risk index obtained in step 2, which reflects the objective risk level of the current driving environment, a weighted fusion method is used to calculate the comprehensive assessment value of driver state risk and scenario environment risk. The formula for calculating the comprehensive risk assessment value is the adjusted driver state risk value multiplied by a first weighting coefficient plus the scenario comprehensive risk index multiplied by a second weighting coefficient. The weighting coefficients reflect the contribution of driver factors and environmental factors to the overall risk. In this embodiment, the first weighting coefficient for driver state risk is set to 0.6, and the second weighting coefficient for scenario environment risk is set to 0.4, reflecting the dominant role of driver factors in risk formation.

[0071] Based on weighted summation, the coupling effect of driver state risk and scenario risk is further considered. When the driver's state is poor and the scenario risk is high, the actual risk generated by the combination of the two is significantly higher than that of simple linear addition. A nonlinear coupling term is introduced, which is equal to the driver state risk value multiplied by the scenario risk index and then multiplied by the coupling coefficient. The coupling term is added to the comprehensive risk assessment value to reflect the risk amplification effect. The value of the coupling coefficient is dynamically determined according to the interaction strength of driver state risk and scenario risk. When both the driver state risk value and the scenario risk index are below 0.5, the interaction is weak, and the coupling coefficient is set to 0.1. When one risk value is between 0.5 and 0.7 and the other is above 0.3, the interaction is enhanced, and the coupling coefficient is set to 0.2. When both risk values ​​are above 0.6, a significant risk amplification effect is produced, and the coupling coefficient is set to 0.3 to 0.5.

[0072] The specific calculation method is as follows: Calculate the product of the two risk values, determine the basic coupling coefficient based on the size of the product, and then fine-tune it according to the distribution characteristics of the risk values. When the driver's state risk is significantly higher than the scenario risk, it indicates that the main risk originates from the driver's internal factors, and the coupling coefficient is appropriately reduced. When the scenario risk is significantly higher than the driver's state risk, it indicates that the external environment is the main risk source, and the coupling coefficient is increased accordingly. The final value of the coupling term is obtained by multiplying the driver's state risk value by the scenario risk index and then by the dynamically determined coupling coefficient.

[0073] The calculated comprehensive risk assessment value is saturated and limited to the range of 0 to 1. The output comprehensive risk assessment value comprehensively reflects the driver's inherent state risk, individual sensitivity differences, environmental scenario risk, and the coupling effect between the two.

[0074] Step 4.4: Determine the risk level classification based on the comprehensive risk value; Based on the comprehensive risk assessment value, three risk thresholds are set, dividing the risk assessment value range of 0 to 1 into four levels. A comprehensive risk assessment value between 0 and 0.3 is considered low risk, indicating that the driver is in a normal state and the scenario is safe, requiring no intervention; a value between 0.3 and 0.6 is considered medium risk, indicating that the driver's state is slightly abnormal or the scenario has some risk, requiring a Level 1 warning; a value between 0.6 and 0.8 is considered high risk, indicating that the driver's state is significantly abnormal or the scenario has significant risk, requiring a Level 2 warning; and a value between 0.8 and 1 is considered emergency risk, indicating that the driver's state is severely abnormal and the scenario is high-risk, requiring a Level 3 mandatory intervention. The risk level classification results are output based on this classification process.

[0075] Step 5: Based on the risk level classification results, the comprehensive risk index of the scenario, and the driver's individual characteristic parameter set, adaptively calibrate the intervention threshold and intensity parameters to obtain the adaptive safety strategy configuration; Step 5.1: Establish a baseline threshold system for the multi-level intervention strategy; Based on the risk level classification results output in step 4, a three-tiered intervention strategy system is established, including Level 1 warning, Level 2 warning, and Level 3 mandatory intervention. Specifically: Level 1 warning corresponds to a medium-risk level, with intervention methods including visual cues and mild voice reminders, aiming to attract the driver's attention without causing strong interference. The baseline trigger threshold for Level 1 warning is set at a comprehensive risk assessment value of 0.3. Level 2 warning corresponds to a high-risk level, with intervention methods including strong visual cues, high-volume voice warnings, and seat vibration, aiming to forcibly arouse the driver's alertness. The baseline trigger threshold for Level 2 warning is set at a comprehensive risk assessment value of 0.6. Level 3 mandatory intervention corresponds to an emergency risk level, with intervention methods including all-around alarms, strong seat vibration, steering wheel force feedback, and driver assistance system takeover, aiming for the system to directly intervene and control the vehicle when the driver's response is insufficient. The baseline trigger threshold for Level 3 intervention is set at a comprehensive risk assessment value of 0.8. The baseline threshold system is determined based on statistical analysis of large-scale driving and accident data to ensure that each level of intervention matches the actual risk level.

[0076] Step 5.2: Adjust the intervention threshold according to the individual driver's sensitivity; Based on the driver individual characteristic parameter set output in step 3, the baseline threshold is adjusted individually according to sensitivity. For highly sensitive drivers, who are more receptive to safety interventions and respond quickly, a relatively lenient threshold setting can be used to avoid excessive intervention causing resentment. The threshold adjustment coefficient for highly sensitive drivers is set to a value greater than 1, making the actual trigger threshold higher than the baseline threshold and the intervention triggering conditions more stringent. For low-sensitive drivers, who are less responsive to mild interventions and require earlier intervention to achieve a warning effect, the threshold adjustment coefficient is set to a value less than 1, making the actual trigger threshold lower than the baseline threshold and the intervention triggering conditions more lenient. The threshold adjustment coefficient has a linear relationship with the sensitivity parameter, calculated as 2 minus the sensitivity parameter. When the sensitivity parameter is 1.5, the adjustment coefficient is 0.5; when the sensitivity parameter is 0.5, the adjustment coefficient is 1.5. Multiplying each level of baseline threshold by the threshold adjustment coefficient yields the personalized intervention threshold. After personalized adjustment, different drivers may trigger different levels of intervention under the same risk assessment value, reflecting the adaptability to individual differences.

[0077] Step 5.3: Dynamically adjust the intervention threshold based on the scenario risk index; Based on the comprehensive scenario risk index output in step 2, the personalized threshold is dynamically adjusted according to the scenario risk. When the scenario risk index is high, it indicates that the current driving environment itself is high-risk. Even if the driver's condition is acceptable, the intervention threshold should be lowered to intervene earlier and enhance preventative protection. The scenario risk threshold correction coefficient is calculated by adding 1 to the scenario adjustment parameter multiplied by the scenario risk index. The scenario adjustment parameter is set to 0.5. When the scenario risk index is low, it indicates that the driving environment is safe. The intervention threshold can be appropriately increased to give the driver more autonomy and reduce unnecessary interference.

[0078] Dividing the personalized intervention threshold by the scenario risk correction coefficient yields the dynamically corrected final intervention threshold. The higher the scenario risk index, the larger the correction coefficient, the lower the final threshold, and the easier it is to trigger intervention. Through dynamic scenario correction, the security strategy can be flexibly adjusted according to real-time changes in environmental risks, providing more proactive protection in high-risk scenarios and focusing more on comfort in low-risk scenarios.

[0079] Step 5.4: Calibrate the intervention intensity parameters and generate the strategy configuration; Based on the comprehensive risk assessment value output in step 4.3, the specific intensity parameters of each level of intervention measures are calibrated. For visual cues of Level 1 warnings, the size and color brightness of the warning icon on the display screen are calibrated; the higher the risk assessment value, the larger the icon size and the more vivid the color. For voice reminders, the playback volume and speech rate are calibrated; a low volume and slow speech rate are used when the risk assessment value is near the threshold, and the volume and speech rate are increased when the risk assessment value significantly exceeds the threshold. For seat vibrations of Level 2 warnings, the vibration frequency and vibration amplitude are calibrated; the vibration frequency setting range is 5 to 15 Hz, and the vibration amplitude setting range corresponds to 20% to 60% of the motor drive voltage; the greater the degree to which the risk assessment value exceeds the Level 2 threshold, the higher the vibration frequency and amplitude. For steering wheel force feedback of Level 3 intervention, the magnitude and duration of the applied corrective torque are calibrated; the corrective torque is calculated based on the current lane departure direction and departure distance, and the torque magnitude is proportional to the risk assessment value. The trigger condition for assisted driving takeover is that the risk assessment value exceeds the Level 3 threshold and the driver does not respond effectively within 5 seconds.

[0080] The calibrated intervention thresholds and intensity parameters at all levels are integrated into a complete security policy configuration data structure, which includes information such as triggering conditions, intervention types, and execution parameters. The resulting adaptive security policy configuration integrates individual sensitivity, scenario risk, and real-time status to achieve personalized and scenario-based accurate calibration.

[0081] In some embodiments, because drivers may become adapted to and desensitized to repeated, identical warning methods during long-distance driving, leading to a decline in warning effectiveness over time, a dynamic adjustment strategy of diverse intervention methods can be adopted. The aim is to maintain the driver's alertness and sensitivity by varying the intervention methods. Specifically, the system records the number of times and time intervals of various intervention measures triggered by the driver in the current driving session. When the same intervention method is detected to be repeatedly triggered in a short period of time, such as more than 5 triggers of a Level 1 warning within 10 minutes, the system determines that the driver may have adapted to the current warning method. The system automatically switches the sensory channel or presentation of the warning, such as switching from visual cues to auditory cues, from fixed-frequency voice to voice with varying pitch, and from single seat vibration to a combination of multi-point vibrations. At the same time, the system adopts a progressive intervention intensity increase strategy, gradually increasing the intervention intensity in continuous warnings. The first warning uses the lowest intensity, and subsequent warnings progressively increase the intensity level until the driver's condition improves or the highest intensity is reached. Through the diversity of intervention methods and the gradual increase in intensity, the system avoids the driver from developing a habitual response to a single warning mode, maintaining the long-term effectiveness of the warning system.

[0082] Step 6: Based on the adaptive security policy configuration, compare the intervention threshold and the comprehensive risk assessment value, execute the security intervention policy and conduct a multi-dimensional effect evaluation to obtain the intervention effect evaluation report and the comprehensive effect score. Step 6.1: Determine the intervention trigger conditions and generate control commands; Based on the comprehensive risk assessment value output in step 4 and the intervention threshold output in step 5, the comprehensive risk assessment value is compared with the Level 1, Level 2, and Level 3 intervention thresholds in sequence. If the risk assessment value is lower than the Level 1 threshold, it is determined that no intervention is required, and the system continues to monitor but does not perform any intervention actions; if the risk assessment value exceeds the Level 1 threshold but is lower than the Level 2 threshold, it is determined that a Level 1 warning is triggered; if the risk assessment value exceeds the Level 2 threshold but is lower than the Level 3 threshold, it is determined that a Level 2 warning is triggered; if the risk assessment value exceeds the Level 3 threshold, it is determined that a Level 3 mandatory intervention is triggered.

[0083] Based on the judgment results, the corresponding level of intervention measures and execution parameters are extracted from the safety policy configuration. For Level 1 warnings, visual prompts and voice playback commands are generated, including display content, icon parameters, voice text, and volume parameters. For Level 2 warnings, strong visual prompts, high-volume voice commands, and seat vibration commands are generated, including vibration frequency, amplitude, and duration parameters. For Level 3 interventions, omnidirectional alarm commands, strong seat vibration commands, steering wheel torque commands, and driver assistance takeover requests are generated.

[0084] The generated control commands are sent to the corresponding execution units via the vehicle controller LAN bus. Visual commands are sent to the vehicle display controller, voice commands to the multimedia audio controller, seat vibration commands to the seat control module, steering wheel torque commands to the electric power steering system, and driver assistance takeover requests to the advanced driver assistance system controller. Upon receiving the control commands, each execution unit drives the corresponding hardware to perform intervention actions according to the command parameters. The vehicle display shows warning icons and text information, the speakers play voice warnings, the seat's built-in vibration motor generates tactile stimulation, the steering wheel motor applies corrective torque, and the driver assistance system takes over the vehicle's steering and speed control.

[0085] Record the time when the intervention is triggered, the risk assessment value at the time of triggering, the risk level, the selected intervention level, and the execution parameters as an intervention event log for subsequent effect evaluation.

[0086] Step 6.2: Monitor driver response behavior and extract response features; Based on the method in step 1, the system continuously extracts driver state features. After the intervention measures are implemented, the system continuously monitors the driver's response behavior and assesses whether the driver perceives the intervention and takes corresponding actions.

[0087] By analyzing the driver's facial image sequence, the system detects whether the driver's gaze turns to the warning information display location. A pupil tracking algorithm determines whether the driver's gaze falls on the warning area of ​​the display screen and whether the gaze duration exceeds 1 second, thus determining whether the driver perceives the visual warning. By analyzing steering wheel operation data, the system detects whether the driver actively adjusts the steering wheel after intervention; the criterion is that the steering wheel angle changes by more than a set threshold within 3 seconds after intervention. Finally, by analyzing seat pressure sensor data, the system detects whether the driver's body posture is adjusted, such as changes in sitting position or leaning forward.

[0088] The time it takes for the driver to make their first noticeable response is recorded, and the response delay is calculated. Response delay is defined as the time difference between the intervention trigger moment and the driver's first response. The shorter the response delay, the higher the driver's sensitivity to the intervention and the better the intervention effect. The threshold for response delay is determined based on human factors engineering research and a large amount of driving behavior data statistics. For visual warnings, a fast response is defined as a response delay of less than 1.5 seconds, a normal response as a response delay of 1.5 to 3 seconds, a slow response as a response delay of 3 to 6 seconds, and no response as a response delay of more than 6 seconds. For auditory warnings, because auditory stimuli are processed faster, a fast response is defined as a response delay of less than 1 second, a normal response as a response delay of 1 to 2.5 seconds, a slow response as a response delay of 2.5 to 5 seconds, and no response as a response delay of more than 5 seconds. For tactile warnings such as seat vibration, a fast response is defined as a response delay of less than 0.8 seconds, a normal response as a response delay of 0.8 to 2 seconds, a slow response as a response delay of 2 to 4 seconds, and no response as a response delay of more than 4 seconds. In the response quality score, a fast response scores 1.0, a normal response scores 0.8, a slow response scores 0.4, and no response scores 0.0.

[0089] The system detects whether the driver actively turns off or ignores the warning information, such as canceling the warning prompt via steering wheel buttons. If the driver actively turns off the warning and the condition does not improve or even worsens within a short period of time after turning it off, it is marked as intervention resistance behavior, reflecting that the driver has low acceptance of the intervention method.

[0090] The characteristic parameters of driver response behavior are extracted to obtain driver response behavior characteristics, including whether the driver perceives the behavior, response delay, response action type, and whether the driver resists the behavior. These characteristics serve as behavioral dimension data for effect evaluation.

[0091] Step 6.3: Track the trajectory of changes in driver state and evaluate the improvement effect; After the intervention is triggered, the driver's state features are continuously extracted according to the method in step 1, and the trajectory of changes in the state features is tracked. At time points of 5 seconds, 10 seconds, 30 seconds, and 60 seconds after the intervention, the driver's state feature vector is extracted and compared with the state feature vector at the time of intervention triggering to calculate the feature change vector.

[0092] Using the method in step 4, the deviation of the state characteristics at each time point from the individual baseline is calculated, and a curve of the deviation over time is constructed. If the deviation gradually decreases after the intervention, it indicates that the driver's state is returning to a normal state, and the intervention has a positive effect. If the deviation remains high or continues to rise after the intervention, it indicates that the intervention has not produced the expected effect.

[0093] Define a state improvement index, calculating the difference between the deviation at 60 seconds after intervention and the deviation at the intervention trigger time. A positive difference indicates state improvement, while a negative difference indicates state deterioration. Normalize the state improvement index to the range of 0 to 1 as a quantitative indicator of improvement effectiveness. Set criteria for effective improvement: if the state improvement index exceeds 0.3 and the deviation shows a stable decreasing trend within 30 seconds, the intervention is considered effective. If the state improvement index is below 0.1 or the deviation shows no significant change, the intervention is considered ineffective. Output the state improvement effect evaluation result, including the improvement index value and the effectiveness judgment.

[0094] Step 6.4: Assess changes in objective safety indicators and calculate the safety margin increase. In addition to improving driver condition, the impact of the intervention on objective driving safety indicators was further evaluated. Vehicle motion data before and after the intervention were obtained, with a focus on analyzing changes in lane keeping performance, speed stability, and following safety. The root mean square (RMS) values ​​of lane deviation were calculated within 30 seconds before and 60 seconds after the intervention, and the differences were compared. A significant decrease in the RMS value of lane deviation after several outcomes indicated improved lane keeping performance and more stable vehicle driving. The standard deviation of vehicle speed before and after the intervention was calculated to assess changes in speed stability; a decrease in the standard deviation of vehicle speed after several outcomes indicated improved speed control. The average and minimum following distances before and after the intervention were calculated to assess changes in following safety; an increase in the average following distance and an improvement in the minimum following distance after several outcomes indicated an increase in the following safety margin.

[0095] The improvement in three dimensions—lane keeping, speed stability, and following safety—is used to calculate the safety margin improvement index, which serves as an objective safety dimension assessment result of the intervention effect. The safety margin improvement index is a weighted average of the improvement in the three dimensions, with the weights determined based on the correlation between each index and accident risk.

[0096] Step 6.5: Integrate multi-dimensional evaluation indicators and generate an effectiveness evaluation report; Based on the driver response behavior characteristics output in step 6.2, the state improvement effect evaluation results output in step 6.3, and the safety margin improvement index output in step 6.4, a multi-dimensional evaluation model is used to calculate the comprehensive intervention effect score. The comprehensive effect score is calculated using a weighted summation method, with the response behavior dimension weighted at 0.2, the state improvement dimension weighted at 0.5, and the safety margin dimension weighted at 0.3. The response behavior dimension score is determined based on response delay and whether there is resistance; the shorter the response delay, the higher the score, and points are deducted for resistance. The state improvement dimension score directly uses the state improvement value; the safety margin dimension score uses the safety margin improvement index. A weighted comprehensive score is calculated, ranging from 0 to 1, with a higher score indicating a better intervention effect. A threshold of 0.5 is set for effective intervention; a comprehensive score above 0.5 is considered effective intervention, and below 0.5 is considered ineffective intervention. A structured intervention effect evaluation report is generated, including the time, location, triggering reason, risk assessment value, intervention level, execution parameters, driver response, state improvement, changes in safety indicators, comprehensive effect score, and effectiveness determination.

[0097] The effectiveness evaluation report is stored in the vehicle's database and uploaded to the cloud data center when the vehicle has a network connection. This process accumulates the intervention effectiveness dataset, outputs the intervention effectiveness evaluation report and a quantified comprehensive effectiveness score, and completes the closed-loop evaluation of this intervention, providing data support for subsequent strategy optimization.

[0098] In some embodiments, because the intervention effect is influenced by various confounding factors, such as driver self-adjustment, changes in road conditions, and the behavior of other vehicles, simple before-and-after comparisons may not accurately separate the intervention's effect. Therefore, an intervention effect attribution analysis method based on causal inference can be employed to more accurately quantify the net effect of the intervention. Specifically, a counterfactual inference model is constructed to simulate the evolutionary trajectory of the driver's state and vehicle movement without intervention. The prediction of the counterfactual trajectory is based on the historical evolution pattern of the driver's state and scenario characteristics. A trained state prediction model is used to generate a state development curve under the uninterrupted condition. The observed post-intervention state trajectory is compared with the counterfactual predicted trajectory; the difference between the two trajectories represents the causal effect of the intervention. By using causal inference, the influence of confounding factors such as natural recovery and scenario changes can be eliminated, allowing for a more accurate assessment of the true effect of the intervention and providing high-quality feedback signals for strategy optimization.

[0099] Step 7: Based on the intervention effect evaluation report and the comprehensive effect score, optimize the adaptive security policy configuration to obtain the optimized adaptive security policy configuration and performance improvement report; Step 7.1: Construct the intervention effect dataset and perform data preprocessing; Based on the intervention effect evaluation report output in step 6, an intervention effect dataset is constructed. Each record in the intervention effect dataset contains complete information such as the driver's state characteristics at the time of intervention triggering, scenario risk parameters, individual characteristic parameters, calibrated intervention threshold and intensity parameters, actual intervention measures implemented, and intervention effect evaluation results.

[0100] An optimization trigger condition is set: when the total number of valid and invalid intervention records accumulated in the dataset reaches more than 100, a parameter optimization process is initiated. During the data accumulation process, the system continuously monitors data quality and filters out abnormal and incomplete data. The dataset is preprocessed, including data cleaning, feature normalization, and sample balancing. Data cleaning removes abnormal records caused by sensor malfunctions and missing records caused by network transmission errors. Feature normalization maps features of different dimensions to the 0-1 range, eliminating the impact of dimension differences on the optimization algorithm. The sample distribution of valid and invalid interventions in the dataset is analyzed. If the number of samples of the two classes is severely imbalanced, oversampling or undersampling methods are used to balance the samples to ensure that the optimization process does not favor the majority class. The preprocessed dataset is divided into a training set and a test set. The training set is used for parameter optimization, and the test set is used to verify the optimization effect, with a division ratio of 8:2.

[0101] Step 7.2: Construct a policy optimization model based on reinforcement learning; The adaptive safety strategy configuration calibration problem is modeled as a reinforcement learning problem. The state space is defined as a combination of driver state characteristics, scenario risk parameters, and individual characteristic parameters. The action space is defined as the adjustment of calibration parameters, including the adjustment coefficients of intervention thresholds at all levels, the adjustment coefficients of intervention intensity parameters, the sensitivity adjustment coefficient, and the scenario adjustment coefficient. The reward function is defined as the comprehensive score of intervention effect, with positive rewards for effective intervention and negative rewards for ineffective intervention. The reward value is proportional to the effect score.

[0102] A proximal policy optimization algorithm is used to construct a policy optimization model, which adopts a dual-network architecture comprising two core components: a policy network and a value network. The policy network is responsible for learning the optimal parameter adjustment strategy. Based on the current driver state characteristics, scenario risk parameters, and individual characteristic parameters, it outputs the probability distribution of adjustment actions for each calibrated parameter, guiding the system on how to adjust intervention thresholds and intensity parameters to achieve better intervention effects. The value network is responsible for evaluating the value of the current state, predicting the long-term cumulative reward obtainable under a given state, and providing a value benchmark for the policy network's learning. The two networks interact through a shared underlying feature extraction layer. The policy network's output actions interact with the environment to generate reward signals. The value network uses these reward signals to calculate an advantage function, which guides the policy network's parameter update direction, enabling the policy network to learn better parameter adjustment strategies.

[0103] The policy network adopts a multi-layer fully connected neural network structure. The input is the feature vector of the state space, and the output is the probability distribution of the adjustment coefficients of each dimension of the action space. It contains 3 hidden layers, each with 128 neurons. The activation function is the modified linear unit function. The value network adopts the same network structure. The input is the state features, and the output is the state value estimate, which is used to calculate the advantage function and guide the policy update direction.

[0104] Policy optimization iterations are performed on the training set data. In each iteration, a batch of intervention events is sampled from the training set, and the state features of the events are input into the policy network to obtain the adjustment actions of the calibration parameters. The adjusted intervention threshold and intensity parameters are calculated. The reward signal is calculated based on the actual intervention effect score, and the advantage function is calculated using the temporal difference method. The advantage function represents the degree of superiority or inferiority of the current action relative to the average level.

[0105] Step 7.3 introduces a security constraint mechanism to ensure optimized security; During strategy optimization, a safety constraint mechanism is introduced, setting lower limits for intervention thresholds: the Level 3 mandatory intervention threshold must not exceed 0.75 to ensure the system will inevitably trigger mandatory intervention in extremely high-risk situations; the Level 2 warning threshold must not exceed 0.55; and the Level 1 warning threshold must not exceed 0.25 to guarantee basic warning coverage. Upper limits for intervention thresholds are also set to prevent excessive intervention due to excessively low thresholds: the Level 1 warning threshold must not be lower than 0.15, the Level 2 warning threshold must not be lower than 0.45, and the Level 3 intervention threshold must not be lower than 0.65, providing drivers with basic autonomous driving space.

[0106] A constraint projection layer is added to the output layer of the policy network to map the adjustment coefficients of the network output to a legal range. A truncation function is used to truncate parameters that exceed the constraint range, ensuring that the output calibration parameters always meet the safety constraints. The specific numerical ranges of the safety constraints are set as follows: the adjustment range of the first-level warning threshold is limited to 0.15 to 0.35 to ensure that mild risks can be detected in a timely manner without being overly sensitive; the adjustment range of the second-level warning threshold is limited to 0.45 to 0.65 to ensure effective warnings for medium risks; and the adjustment range of the third-level mandatory intervention threshold is limited to 0.65 to 0.85 to ensure mandatory protection in high-risk situations. The constraint ranges of the intervention intensity parameters are as follows: the brightness adjustment range of visual cues is 50% to 150% of the standard brightness; the voice volume adjustment range is 70% to 130% of the standard volume; the seat vibration intensity adjustment range is 60% to 140% of the standard intensity; and the steering wheel torque adjustment range is 80% to 120% of the standard torque. The penalty mechanism employs a tiered penalty strategy: a mild penalty of 10% of the current reward value is applied when the optimization algorithm attempts to adjust the parameters to the constraint boundary; a moderate penalty of 30% of the current reward value is applied when the parameters exceed the constraint range; and a severe penalty of 100% of the current reward value is applied when the parameters severely violate safety constraints, such as when the level 3 intervention threshold exceeds 0.9, forcibly resetting the parameters to the safe range. The penalty mechanism also includes cumulative penalties; consecutive violations of constraints lead to progressively increasing penalty values, guiding the optimization algorithm away from the constraint boundaries.

[0107] A constraint penalty term is introduced into the objective function. When the policy tends to violate safety constraints, a penalty signal is given to guide the policy away from the constraint boundary. The weight of the penalty term gradually increases with each optimization iteration, strengthening the hard requirements of the safety constraints. Through the safety constraint mechanism, the optimization process improves the intervention effect while ensuring that the policy always meets the safety requirements, avoiding the safety risks brought about by optimization.

[0108] Step 7.4: Perform parameter optimization and verify the optimization effect; The policy parameters are iteratively updated on the training set using a proximal policy optimization algorithm. The policy gradient is calculated in each iteration, and the weight parameters of the policy network and value network are updated using a mini-batch gradient ascent algorithm. The learning rate is set to 0.0003, and the number of iterations is 50, processing all training samples in each iteration. After each iteration, the performance of the current policy is evaluated on the test set. Intervention events from the test set are input into the optimized policy optimization model to calculate the adjusted calibration parameters, simulating the intervention effect when using the new parameters.

[0109] Calculate performance metrics such as average intervention effectiveness score, effective intervention rate, false positive rate, and false negative rate on the test set, and compare them with the baseline strategy before optimization. The effective intervention rate is defined as the proportion of effective interventions to the total number of interventions, the false positive rate is defined as the proportion of ineffective interventions to the total number of interventions, and the false negative rate is obtained by analyzing cases where no intervention occurred but a risk event subsequently occurred.

[0110] If the optimized strategy improves the effective intervention rate and reduces the false positive rate on the test set, the optimization is deemed effective, and the optimized strategy and model parameters are saved. If the optimization effect is not obvious or the performance deteriorates, the number of training iterations is increased or the hyperparameters are adjusted, and the optimization process is re-executed. The changes in performance indicators before and after each optimization are recorded, and an optimization effect report is generated, which includes information such as the increase in effective intervention rate, the decrease in false positive rate, and changes in strategy parameters.

[0111] Step 7.5: Deploy the optimized strategy model and perform canary-scale verification; The optimized strategy model parameters are packaged into an update package and pushed to vehicles via a vehicle-to-everything (V2X) remote upgrade mechanism. To ensure the security and reliability of the optimization strategy, a canary release strategy is adopted, conducting canary testing on a small number of vehicles. The number of canary vehicles is set to 5% of the total fleet size. The canary vehicles install the optimized strategy model, while the remaining vehicles retain the original strategy.

[0112] During the gray-scale testing period, the system operation status and user feedback of the gray-scale vehicles are continuously monitored, and intervention effect data of the gray-scale vehicles are collected. Key indicators such as effective intervention rate, false alarm rate, and user complaint rate are statistically analyzed. The gray-scale verification period is set at 2 weeks. If the performance indicators of the gray-scale vehicles are better than those of the non-gray-scale vehicles and there are no abnormalities during the verification period, the gray-scale test is considered to have passed. After the gray-scale test is passed, the optimization strategy is pushed to all vehicles to complete the full update of the strategy model. In the first month after the full update, system performance and user feedback continue to be monitored to ensure that the optimization strategy is stable and reliable in large-scale applications.

[0113] A strategy version management mechanism is established to record the version number, optimization time, parameter changes, and performance improvements for each optimization. This supports strategy rollback and version switching; if a new strategy encounters problems in practical application, it can be quickly rolled back to the previous stable version. Through continuous data accumulation, parameter optimization, and canary-scale verification, the safety strategy calibration system possesses self-evolution capabilities. Its performance continuously improves with usage time and data accumulation, providing drivers with increasingly accurate safety protection. It outputs optimized adaptive safety strategy configurations and performance improvement reports, completing this round of optimization iterations.

[0114] In some embodiments, due to differences in sensor configuration, cockpit hardware capabilities, and user group characteristics among different vehicle models, directly migrating strategy parameters optimized for one model to a new model may not yield good results. A meta-learning-based cross-vehicle strategy migration method can be employed to accelerate strategy calibration for new models and reduce development costs. Specifically, a multi-vehicle strategy knowledge base is built in the cloud, storing strategy parameters, vehicle attribute features, user group characteristics, and strategy performance data for different models. When a new model needs to deploy a security strategy system, existing models with similar attributes to the new model are retrieved from the knowledge base, and their strategy parameters are extracted as initialization parameters. A meta-learning algorithm is used to train a cross-vehicle strategy adaptation model, which learns how to adjust strategy parameters based on differences in vehicle attributes. The attribute features of the new model are input into the adaptation model to obtain parameter adjustment coefficients for the new model. These coefficients are then used to adjust the initialization parameters, resulting in adapted strategy parameters. A small-scale trial run is conducted on the new model using the adapted strategy to collect performance data and perform rapid fine-tuning and optimization, completing the strategy calibration for the new model. By migrating across vehicle models, new models can quickly obtain a high-performing initial strategy without having to accumulate data and optimize parameters from scratch, thus shortening the development cycle.

[0115] A cockpit safety strategy calibration system based on pilot status monitoring, such as Figure 3 As shown, a cockpit safety strategy calibration method based on driver status monitoring, as described above, includes: The data acquisition module is used to collect multimodal monitoring data of the driver, extract state features, and perform multimodal feature fusion to obtain the driver state feature vector; The scene perception module is used to collect driving scene parameters in real time, quantify the risk of the driving scene based on the driving scene parameters, and obtain the comprehensive scene risk index. The individual modeling module, based on the driver state feature vector, constructs a driver individual feature model for new users through transfer learning and extracts individual feature parameters to obtain a set of driver individual feature parameters; The risk assessment module calculates and classifies the comprehensive risk assessment value based on the driver's state feature vector, the scenario comprehensive risk index, and the driver's individual feature parameter set, thus obtaining the comprehensive risk assessment value and risk level classification results. The strategy calibration module adaptively calibrates the intervention threshold and intensity parameters based on the risk level classification results, the comprehensive risk index of the scenario, and the driver's individual characteristic parameter set, to obtain an adaptive safety strategy configuration. The intervention execution module, based on adaptive security policy configuration, compares intervention thresholds and comprehensive risk assessment values, executes security intervention strategies, performs multi-dimensional effect evaluation, and obtains an intervention effect evaluation report and a comprehensive effect score. The strategy optimization module optimizes the adaptive security strategy configuration based on the intervention effect evaluation report and the comprehensive effect score, resulting in an optimized adaptive security strategy configuration and a performance improvement report.

[0116] In one embodiment of the present invention, a specific example is provided: A driver is driving a smart connected vehicle equipped with the safety strategy calibration system of this invention on a long-distance highway. The driving starts at 8:00 AM, the planned mileage is 400 kilometers, and the estimated driving time is 4 hours.

[0117] At the start of driving, the system loads the driver's individual characteristic model. The driver's individual characteristic parameter is 0.8, which belongs to the medium to low sensitivity type. The fatigue accumulation rate parameter shows that the driver's fatigue characteristics begin to appear after driving continuously for 2 hours.

[0118] During the first hour of driving, the driver's condition was normal. Real-time monitoring by the system showed that the driver's blink frequency, steering wheel operation, and lane-keeping performance were all within the normal baseline range. The current scenario was a straight section of highway with moderate traffic density, and the scenario risk index was 0.25. The overall risk assessment value was 0.18, indicating a low-risk level, and the system did not trigger any intervention.

[0119] The monitoring data for a typical moment in the first hour are shown in Table 1: Table 1: Monitoring data at typical times in the first hour: During the second hour of driving, the driver began to show signs of mild fatigue, with decreased blinking frequency, increased instances of prolonged eye closure, and decreased steering wheel stability. The system's calculated deviation from the driver's condition gradually increased, reaching a comprehensive risk assessment value of 0.32 at 10.10 minutes, exceeding the Level 1 warning threshold of 0.28. The system triggered a Level 1 warning, displaying a fatigue alert icon on the instrument panel and playing a gentle voice prompt. After hearing the prompt, the driver adjusted their seating position and opened the windows for ventilation. The driver's condition improved within 3 minutes, and the comprehensive risk assessment value dropped to 0.24, indicating the system deemed the intervention effective.

[0120] In the third hour of driving, the vehicle entered a mountainous highway section, encountering a series of curves ahead with increased road curvature. The system retrieved road information from a high-precision map and identified a high-risk, continuous curve section 5 kilometers ahead, with the scenario risk index rising to 0.52. Based on proactive scenario risk prediction, the system began gradually lowering the intervention threshold 8 minutes before the vehicle reached this section, reducing the Level 1 warning threshold from 0.28 to 0.22. At this point, the driver's deviation was 0.28, which, while not yet meeting the warning criteria under normal conditions, triggered a Level 1 warning under the reduced threshold. The system proactively alerted the driver to the high-risk section ahead, urging them to concentrate. Upon receiving the alert, the driver proactively reduced speed and adjusted their driving posture, successfully navigating the continuous curve section without lane departure or other dangerous situations.

[0121] The monitoring data for high-risk road sections in the third hour are shown in Table 2: Table 2: Monitoring data of high-risk road sections in the 3rd hour: In the fourth hour of driving, the driver had been driving continuously for nearly three hours, resulting in a significant cumulative fatigue effect. At 11:35 AM, the comprehensive risk assessment value reached 0.63, exceeding the level two warning threshold of 0.58. The system triggered a level two warning, the seat began to vibrate, the voice warning volume increased, and the driver was advised to rest at a service area as soon as possible. After receiving the level two warning, the driver remained alert and continued driving, pulling into a highway service area to rest five minutes later, completing the first half of this long-distance drive.

[0122] This application example demonstrates the application effect of the method of the present invention in a real driving scenario. The system can perform adaptive strategy calibration based on the driver's individual characteristics and real-time status, as well as the dynamic risks of the driving scenario, to achieve accurate early warning intervention and improve driving safety.

[0123] This invention achieves precise calibration and continuous evolution of cockpit safety strategies through multi-source data fusion, individual feature adaptive learning, dynamic perception of scenario risks, and closed-loop optimization of effect feedback. It solves the problems of poor applicability, lack of personalization and scenario adaptability caused by fixed threshold calibration in existing technologies, and provides an effective technical solution for driver status monitoring and active safety protection of intelligent connected vehicles.

[0124] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for calibrating a cockpit safety strategy based on driver state monitoring, characterized in that, Includes the following steps: Multimodal monitoring data of the driver is collected, state features are extracted and multimodal feature fusion is performed to obtain the driver state feature vector; Real-time collection of driving scenario parameters; quantification of driving scenario risks based on driving scenario parameters; and obtaining a comprehensive scenario risk index. Based on the driver state feature vector, a driver individual feature model for new users is constructed through transfer learning, and individual feature parameters are extracted to obtain the driver individual feature parameter set; Based on the driver's state feature vector, the scenario comprehensive risk index and the driver's individual feature parameter set, the comprehensive risk assessment value is calculated and classified to obtain the comprehensive risk assessment value and risk level classification results. Based on the risk level classification results, the comprehensive risk index of the scenario, and the driver's individual characteristic parameter set, the intervention threshold and intensity parameters are adaptively calibrated to obtain the adaptive safety strategy configuration; Based on adaptive security policy configuration, the intervention threshold and comprehensive risk assessment value are compared, the security intervention policy is executed and a multi-dimensional effect evaluation is performed to obtain an intervention effect evaluation report and a comprehensive effect score. Based on the intervention effect evaluation report and the comprehensive effect score, the adaptive security policy configuration is optimized to obtain the optimized adaptive security policy configuration and performance improvement report.

2. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The process of collecting multimodal monitoring data from drivers, extracting state features, and fusing multimodal features includes: The driver's facial visual data was collected and the eye and head features were extracted. The blinking frequency, the proportion of prolonged eye closure, the angle of gaze deviation, and the head posture angle parameters were calculated to obtain the facial visual features. Collect steering wheel operation data and extract driving behavior features, calculate the standard deviation of steering wheel angle, peak angular velocity, grip torque fluctuation amplitude and spectral energy distribution to obtain steering wheel operation features; Collect vehicle motion state data and extract driving performance characteristics, calculate vehicle speed fluctuation coefficient, root mean square value of lateral offset distance, number of lane departures, vehicle distance change rate and pedal operation frequency to obtain driving performance characteristics; A multimodal feature fusion network is constructed, which maps facial visual features, steering wheel operation features, and driving performance features into hidden layer representations and then concatenates them. The network is then input into a bidirectional long short-term memory network for temporal modeling and outputs a driver state feature vector.

3. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The real-time collection of driving scenario parameters, and the quantification of driving scenario risks based on these parameters, include: Traffic flow information is collected using millimeter-wave radar and cameras to calculate vehicle density, distance to the vehicle in front, and collision time, and to calculate traffic density risk scores. The road curvature is calculated based on steering wheel angle and vehicle speed data, the slope angle is obtained from the inertial measurement unit, the lane width is calculated through lane line detection, and the road geometric risk score is obtained. By integrating traffic density risk score, road geometry risk score, and vehicle speed risk coefficient, a comprehensive risk index for the scenario is calculated.

4. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The step of constructing a driver individual characteristic model for new users through transfer learning and extracting individual characteristic parameters includes: Load a pre-trained baseline model of general driver features; Collect initial driving data from new users and perform feature annotation; continuously collect status monitoring data; automatically annotate the collected driver status feature vectors; and distinguish between normal state samples and abnormal state samples. Update the weight parameters of the last two fully connected layers of the general driver feature baseline model and perform parameter updates to obtain the individual driver feature model; Driver individual characteristic parameters and characteristic baselines are extracted. The characteristic baseline vector and covariance matrix are calculated from the normal state samples. The sensitivity parameters are evaluated by analyzing the driver's response to the test prompts. The fatigue accumulation parameters are obtained by fitting the fatigue accumulation model, and the driver individual characteristic parameter set is obtained.

5. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The calculation and classification of the comprehensive risk assessment value includes: Mahalanobis distance was used to calculate the deviation of the driver's state from the individual baseline, and then normalized. The adjustment coefficient is calculated based on the individual sensitivity parameters to obtain the adjusted driver state risk value; A weighted fusion method is used to calculate the comprehensive assessment value of driver state risk and scenario risk; By setting risk thresholds, the comprehensive risk assessment value is divided into low risk, medium risk, high risk, and emergency risk.

6. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The adaptive calibration intervention threshold and intensity parameters include: A three-tiered intervention strategy system was established based on the risk level classification results; The intervention threshold is adjusted according to the individual driver's sensitivity. Based on the sensitivity parameters in the set of individual driver characteristic parameters, the threshold adjustment coefficient is calculated, and the baseline threshold is adjusted in a personalized manner. The intervention threshold is dynamically adjusted based on the scenario risk index. The scenario risk adjustment coefficient is calculated based on the scenario comprehensive risk index, and the personalized intervention threshold is dynamically adjusted. The intervention intensity parameters are calibrated and strategy configurations are generated. The intensity parameters of each level of intervention measures are calibrated based on the comprehensive risk assessment value.

7. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The comparison of intervention thresholds and comprehensive risk assessment values, the execution of safety intervention strategies, and the multi-dimensional effectiveness evaluation include: The comprehensive risk assessment value is compared with the intervention thresholds at each level to generate corresponding level intervention control instructions; By analyzing facial images, steering wheel operations, and seat pressure data, driver response behavior is detected, and response delay and quality are recorded. State feature vectors were extracted at multiple time points after intervention, and the state improvement index was calculated. Analyze the changes in lane keeping, speed stability, and following safety before and after the intervention, and calculate the improvement in safety margin; The overall intervention effect score is calculated based on response behavior, state improvement effect, and safety margin improvement.

8. The cabin safety policy calibration method based on driver state monitoring according to claim 1, characterized in that, The optimization of adaptive security policy configuration includes: Construct an intervention effect dataset and perform data preprocessing; Construct a policy optimization model based on reinforcement learning, model the policy labeling problem as a reinforcement learning problem, and build a policy network and a value network; To ensure optimization safety, a safety constraint mechanism is introduced, which sets upper and lower limits for the intervention threshold, adds a constraint projection layer to the output layer of the policy network, and introduces a constraint penalty term into the optimization objective function. Perform parameter optimization and verify the optimization effect. Use the near-end policy optimization algorithm to iteratively update the policy parameters and evaluate the performance metrics on the test set. Deploy the optimized strategy model and perform canary release verification, then use a canary release strategy to test on a small number of vehicles.

9. The cabin safety policy calibration method based on driver state monitoring according to claim 2, characterized in that, The multimodal feature fusion network processes visual features, operational features, and performance features respectively. The visual feature branch maps facial visual features to hidden layer representations, the operational feature branch maps steering wheel operation features to hidden layer representations, and the performance feature branch maps driving performance features to hidden layer representations.

10. A cockpit safety policy calibration system based on driver state monitoring, characterized in that, A cockpit safety strategy calibration method based on driver status monitoring as described in any one of claims 1-9 includes: The data acquisition module is used to collect multimodal monitoring data of the driver, extract state features, and perform multimodal feature fusion to obtain the driver state feature vector; The scene perception module is used to collect driving scene parameters in real time, quantify the risk of the driving scene based on the driving scene parameters, and obtain the comprehensive scene risk index. The individual modeling module, based on the driver state feature vector, constructs a driver individual feature model for new users through transfer learning and extracts individual feature parameters to obtain a set of driver individual feature parameters; The risk assessment module calculates and classifies the comprehensive risk assessment value based on the driver's state feature vector, the scenario comprehensive risk index, and the driver's individual feature parameter set, thus obtaining the comprehensive risk assessment value and risk level classification results. The strategy calibration module adaptively calibrates the intervention threshold and intensity parameters based on the risk level classification results, the comprehensive risk index of the scenario, and the driver's individual characteristic parameter set, to obtain an adaptive safety strategy configuration. The intervention execution module, based on adaptive security policy configuration, compares intervention thresholds and comprehensive risk assessment values, executes security intervention strategies, performs multi-dimensional effect evaluation, and obtains an intervention effect evaluation report and a comprehensive effect score. The strategy optimization module optimizes the adaptive security strategy configuration based on the intervention effect evaluation report and the comprehensive effect score, resulting in an optimized adaptive security strategy configuration and a performance improvement report.

Images