Driver incapacity intervention methods, devices, vehicles and storage media

By using multi-source sensing features and causal reasoning based on causal graph models, the interpretability and reliability of driver disability status determination results are resolved, achieving accuracy in driver disability detection and reliability in intervention, thus ensuring the credibility of determination results and the precision of intervention strategies.

CN123104672APending Publication Date: 2026-11-17CHONGQING CHANGAN AUTOMOBILE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611320391.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, the determination of a driver's disability status relies on the simple splicing or weighting of multi-source data, which cannot reveal the deep causal mechanism between different modal information. This results in a lack of interpretability and reliability of the determination results, a high risk of misjudgment and omission, and an inability to quantify the credibility of the determination results, making it difficult to dynamically adjust intervention strategies.

Method used

By acquiring multi-source perception features of the driver, vehicle motion, and environment, causal reasoning is performed using a pre-set causal graph model to determine the posterior probability of candidate disability status levels. Then, by weighting the causal contribution and confidence, the level with a comprehensive confidence greater than the threshold is selected as the target disability status, and the vehicle is controlled to execute the corresponding intervention strategy.

Benefits of technology

It improves the accuracy and interpretability of driver disability detection, ensures the reliability of the judgment results and the effectiveness of intervention, avoids misjudgment and misintervention under low confidence, and achieves precise matching of intervention measures with the severity of disability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104672A_ABST
    Figure CN123104672A_ABST
Patent Text Reader

Abstract

This application provides a driver disability intervention method, device, vehicle, and storage medium, relating to the field of driver disability intervention technology, aiming to improve the accuracy, interpretability, and reliability of driver disability detection and intervention. The method involves: determining the posterior probability value corresponding to any candidate disability state level from a preset causal graph model based on multi-source sensing features; for any candidate disability state level: determining the causal contribution of each dimension in the multi-source sensing features based on the posterior probability value corresponding to the candidate disability state level, and weighting the causal contribution of each dimension with confidence to determine the comprehensive confidence level; among the candidate disability state levels with a comprehensive confidence level greater than the corresponding preset confidence threshold, selecting the candidate disability state level corresponding to the largest posterior probability value and determining it as the target disability state level to control the vehicle to execute the corresponding intervention strategy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of driver disability intervention technology, specifically to a driver disability intervention method, device, vehicle, and storage medium. Background Technology

[0002] With the rapid development of intelligent driving assistance technology, driver status monitoring systems have evolved from single-vision detection to multi-source data acquisition. However, the relevant multi-source data is mostly limited to the driver's own physiological and behavioral characteristics, lacking sufficient integration of vehicle motion status and environmental information. In related technologies, the determination of driver disability status typically involves simply splicing or weighting multi-source data to directly output the disability level. This approach fails to reveal the deep causal mechanisms between different modalities, resulting in a lack of interpretability and low reliability in the judgment results. This fundamental deficiency further makes it difficult for the system to distinguish between correlation and causality in complex scenarios, leading to a high risk of misjudgment and missed judgment, and the inability to quantitatively assess the reliability of the judgment results. Consequently, subsequent intervention strategies cannot be dynamically adjusted based on the reliability of the judgment. Therefore, how to overcome the limitations of shallow correlation in the fusion of relevant multi-source data and improve the accuracy, interpretability, and reliability of driver disability detection has become an urgent technical problem to be solved. Summary of the Invention

[0003] The purpose of this application is to provide a method, device, vehicle, and storage medium for intervening in driver disability, aiming to improve the accuracy, interpretability, and reliability of driver disability detection and intervention.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, a method for intervening in driver disability is provided. This method includes: acquiring multi-source perception features; the multi-source perception features include features in the following dimensions: driver features, vehicle motion features, and environmental perception features; based on the multi-source perception features, determining the posterior probability value corresponding to any candidate disability state level from a pre-defined causal graph model; the candidate disability state level is used to characterize the severity of driver disability; the pre-defined causal graph model is used to characterize a directed acyclic graph (DAG) connecting environmental nodes, vehicle nodes, driver nodes, and disability state nodes with directed edges, each directed edge representing the causal probability value between two nodes connected in the DAG; the posterior probability value corresponding to any candidate disability state level is based on the path to the disability state node. The causal probability value corresponding to the missing directed edge is determined; for any candidate disability state level among multiple candidate disability state levels: the causal contribution of each dimension in the multi-source perception features is determined based on the posterior probability value corresponding to the candidate disability state level, and the causal contribution of each dimension is weighted by confidence to determine the comprehensive confidence level; the causal contribution is used to characterize the degree of influence of each dimension on the posterior probability value corresponding to the candidate disability state level; the comprehensive confidence level is used to characterize the reliability of the posterior probability value corresponding to the candidate disability state level; among the candidate disability state levels with a comprehensive confidence level greater than the corresponding preset confidence threshold, the candidate disability state level corresponding to the largest posterior probability value is selected and determined as the target disability state level; the vehicle is controlled to execute the intervention strategy corresponding to the target disability state level.

[0005] Based on the aforementioned technical means, this application provides a comprehensive information foundation for cross-modal causal analysis by acquiring multi-source features of driver, vehicle motion, and environmental perception. Subsequently, based on a pre-defined causal graph model, inference is performed using the causal probability values ​​represented by directed edges to determine the posterior probability of each candidate disability state level. This elevates data association from the statistical correlation level to the causal inference level, directly solving the problem that related technologies cannot establish cross-modal causal links, thus enabling the judgment results to have causal interpretability. Next, for any candidate disability state level, the causal contribution of each dimension feature to its posterior probability is calculated, quantifying the specific influence of each modality on the judgment result, making the judgment basis traceable, and further enhancing the transparency of causal inference. Then, the causal contribution is weighted by confidence level to determine the comprehensive confidence level, realizing a quantitative expression of the reliability of the posterior probability, filling the gap in related technologies where the uncertainty of the judgment result cannot be measured, and providing a credible basis for subsequent decision-making. Based on this, the system uses a comprehensive confidence level greater than a threshold as a screening condition and selects the level corresponding to the maximum posterior probability as the target disability state. This ensures that a decision is made only when the judgment result is sufficiently reliable, effectively avoiding misjudgments and misinterventions under low confidence levels. Finally, the system controls the vehicle to execute the intervention strategy corresponding to the level, achieving a precise match between intervention measures and the severity of disability and the reliability of the judgment. This improves the overall accuracy, interpretability, and reliability of driver disability detection and intervention.

[0006] Furthermore, based on multi-source perception features, the posterior probability value corresponding to any candidate disability state level among multiple candidate disability state levels is determined from a pre-defined causal graph model. This includes: determining an initial node set from the pre-defined causal graph model based on multi-source perception features; the initial node set includes at least one initial node, each initial node being any one of an environment node, a vehicle node, or a driver node; determining a set of links from any node in the initial node set as the starting point pointing to the disability state node; the link set includes at least one link; and performing forward propagation inference based on the causal probability values ​​corresponding to all directed edges in the link set to determine the posterior probability value of the disability state node being in any candidate disability state level.

[0007] Based on the aforementioned technical methods, an initial node set is determined using multi-source sensing features. Then, a set of links pointing to the disabled state node is determined from the initial node, and forward propagation reasoning is performed along these links. This achieves end-to-end causal tracing from environmental triggers and vehicle response to changes in driver state. Compared to directly outputting the judgment result, this process ensures that each step of reasoning has a clear causal path to follow, thereby improving the interpretability and accuracy of the disabled state determination.

[0008] Furthermore, the causal contribution of each dimension in the multi-source sensing features is determined based on the posterior probability value corresponding to the candidate disability status level, including: applying a preset increment to the first dimension in the multi-source sensing features, and determining the difference between the posterior probability values ​​before and after applying the preset increment; the first dimension is any dimension in the multi-source sensing features; the ratio of the difference to the preset increment is determined as the causal contribution of the first dimension.

[0009] Based on the aforementioned technical means, the marginal causal impact of a single-dimensional feature on the disability determination result is quantified by applying a preset increment to the feature and calculating the difference in the posterior probability. This method expresses the specific contribution of each modality to the determination result in numerical form, making the determination basis traceable and further enhancing the transparency of causal reasoning.

[0010] Furthermore, the causal contribution of each dimension is weighted by confidence level to determine the comprehensive confidence level, including: obtaining the preset weight of each dimension in the multi-source perception features; and weighting and summing the causal contribution of each dimension with the preset weight of each dimension to obtain the comprehensive confidence level corresponding to the candidate disability status level.

[0011] Based on the aforementioned technical methods, by obtaining preset weights for each dimension of features and performing a weighted summation of causal contributions, the differences in the credibility of different modalities of evidence are incorporated into the calculation of the overall confidence level. This approach enables the overall confidence level to reflect the overall reliability level of evidence across all dimensions under current observation conditions, providing a quantitative basis for determining whether to adopt the results.

[0012] Furthermore, the intervention strategy corresponding to the target disability level is controlled by the vehicle, including: determining the disability detection duration based on the comprehensive confidence level corresponding to the target disability level; the comprehensive confidence level is negatively correlated with the disability detection duration; and executing the intervention strategy corresponding to the target disability level when the duration of monitoring the driver's operation response exceeds the disability detection duration.

[0013] Based on the aforementioned technical methods, by establishing a negative correlation function between disability detection time and overall confidence level, the reliability of the judgment is directly quantified into a specific waiting time. When the overall confidence level is high, the detection time is automatically shortened to quickly trigger intervention; when the overall confidence level is low, the detection time is automatically extended to collect more evidence. This achieves adaptive dynamic adjustment of the confirmation window, effectively improving interaction efficiency while ensuring safety.

[0014] Furthermore, the duration of disability detection satisfies the following relationship:

[0015] in, This refers to the duration of disability detection. The preset detection duration is related to the target disability level. For overall confidence level; To pre-set the reliability threshold; This is the confidence sensitivity coefficient, used to control the degree of influence of the overall confidence level on the window duration; The evidence quality coefficient is determined based on the pre-set reliability of each dimension in the multi-source perception features.

[0016] Based on the aforementioned technical methods, a quantitative calculation model for disability detection duration is constructed by introducing a preset detection duration, a pre-set confidence threshold, a confidence sensitivity coefficient, and an evidence quality coefficient. This allows the adjustment of the window duration to comprehensively consider multiple factors, including the severity of disability, the reliability of the judgment, the sensitivity to changes in confidence, and the quality of sensor data. This model ensures that every adjustment of the window duration has a clear calculation basis, avoiding the rigidity of fixed windows and the blindness of subjective experience-based parameter tuning, further enhancing the adaptability and decision-making transparency of the confirmation mechanism.

[0017] Furthermore, the method also includes: acquiring current multi-source perception features; the current multi-source perception features are the multi-source perception features after implementing the intervention strategy corresponding to the target disability level; determining the comprehensive score of intervention effect based on the deviation between the multi-source perception features and the current multi-source perception features; updating the intervention control parameters through a preset reinforcement learning algorithm with the goal of maximizing the comprehensive score of intervention effect; the intervention control parameters include at least one of the following: preset detection duration, preset reliability threshold, confidence sensitivity coefficient, and evidence quality coefficient.

[0018] Based on the aforementioned technical methods, the intervention effect is evaluated by comparing the deviations in multi-source sensory features before and after the intervention. With the goal of maximizing the comprehensive score of the intervention effect, a reinforcement learning algorithm is used to update the intervention control parameters online. This method upgrades the intervention strategy from open-loop execution to closed-loop optimization, enabling the system to continuously and adaptively adjust based on the actual intervention effect, gradually improving the effectiveness and personalized adaptability of the intervention strategy.

[0019] Furthermore, the method also includes: recalculating the disability detection duration when the change in the overall confidence level exceeds a preset change threshold; and controlling the vehicle to end the disability detection duration and canceling the execution of the intervention strategy corresponding to the target disability level when the duration corresponding to the monitoring of the driver's operation response status is less than or equal to the disability detection duration.

[0020] Based on the aforementioned technical methods, by dynamically monitoring changes in confidence level and driver response status during the disability detection period, the detection period is recalculated when a significant change in confidence level occurs, and intervention is promptly canceled when a valid driver response is detected. This approach allows the confirmation window to be flexibly adjusted according to real-time changes in judgment conditions and driver status, avoiding unnecessary interventions caused by changes in confidence level or the driver regaining control, further improving the system's interactive rationality and user experience.

[0021] Furthermore, the multiple candidate disability status levels include a first disability level, a second disability level, and a third disability level; wherein, the severity of disability corresponding to the first disability level, the second disability level, and the third disability level increases sequentially; the intervention strategies corresponding to the first disability level include: sending attention reminders to the driver through at least one of the following cabin warning operations: visual reminders, auditory reminders, tactile reminders, and olfactory reminders; the intervention strategies corresponding to the second disability level include: adjusting the vehicle dynamics characteristics through at least one of the following vehicle chassis operations: speed limit control, lane keeping enhancement, and active braking preparation, to assist in stable vehicle driving; the intervention strategies corresponding to the third disability level include: performing at least one of the following remote support operations: fault data upload, emergency contact notification, and remote driving takeover preparation.

[0022] Based on the aforementioned technical means, by classifying disability levels into three progressively more severe levels and configuring progressive intervention strategies for each level, from cockpit sensory alerts to active chassis control and remote cloud support, a precise match between intervention intensity and disability severity is achieved. This layered intervention architecture follows the principle of minimal invasiveness, avoiding excessive intervention that could interfere with driving during mild distractions, and ensuring timely implementation of the highest level of emergency response during severe disability, thus balancing driving safety and user experience.

[0023] Secondly, this application provides a driver disability intervention device, which includes: a feature acquisition module, a probability determination module, a confidence level determination module, a level determination module, and an execution strategy module.

[0024] The feature acquisition module is used to acquire multi-source perception features, which include features in the following dimensions: driver features, vehicle motion features, and environmental perception features. The probability determination module is used to determine the posterior probability value corresponding to any candidate disability state level from a pre-defined causal graph model based on the multi-source perception features. Candidate disability state levels are used to characterize the severity of driver disability. The pre-defined causal graph model represents a directed acyclic graph (DAG) connecting environmental nodes, vehicle nodes, driver nodes, and disability state nodes with directed edges. Each directed edge represents the causal probability value between two nodes connected in the DAG. The posterior probability value corresponding to any candidate disability state level is determined based on the causal probability value corresponding to at least one directed edge in the link pointing to the disability state node. The confidence level is then determined. The confidence level module is used to determine the causal contribution of each dimension in the multi-source perception features based on the posterior probability value corresponding to the candidate disability status level, and to determine the comprehensive confidence level by weighting the causal contribution of each dimension with confidence. The causal contribution is used to characterize the degree of influence of each dimension on the posterior probability value corresponding to the candidate disability status level. The comprehensive confidence level is used to characterize the reliability of the posterior probability value corresponding to the candidate disability status level. The level determination module is used to select the candidate disability status level corresponding to the largest posterior probability value from the candidate disability status levels with a comprehensive confidence level greater than the corresponding preset confidence threshold, and determine it as the target disability status level. The execution strategy module is used to control the vehicle to execute the intervention strategy corresponding to the target disability status level.

[0025] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory storing processor-executable instructions. When the processor is configured to execute the instructions, the electronic device implements the method described in the first aspect.

[0026] Fourthly, this application provides a vehicle that includes the electronic equipment described in the third aspect or the driver disability intervention device described in the second aspect.

[0027] Fifthly, this application provides a computer-readable storage medium that, when the instructions in the computer-readable storage medium are executed by a vehicle's processor, enables the vehicle to perform the methods described in the first aspect and any of their possible implementations.

[0028] Sixthly, this application provides a computer program product including computer instructions that, when executed on a vehicle, cause the vehicle to perform the method described in the first aspect and any possible implementation thereof.

[0029] It should be noted that the technical effects of any of the implementation methods in aspects two through six can be found in the technical effects of the corresponding implementation methods in aspect one, and will not be repeated here.

[0030] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the structure of a driver disability intervention system provided in an embodiment of this application; Figure 2 A flowchart illustrating a driver disability intervention method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a multimodal cause-effect graph provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a factor graph provided in an embodiment of this application; Figure 5 This is a schematic diagram of the composition of a driver disability intervention device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the composition of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] It should be noted that in the embodiments of this application, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplarily" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0034] In the embodiments of this application, the terms "first," "second," "third," "fourth," "fifth," and "sixth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," "fourth," "fifth," and "sixth" may explicitly or implicitly include one or more of that feature.

[0035] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. For "A and / or B," this includes three combinations: A only, B only, and a combination of A and B.

[0036] In related technologies, driver disability intervention methods suffer from numerous problems: multi-source data collection is limited to the driver themselves, lacking effective integration of vehicle motion and environmental information; the judgment process simply splices or weights multi-source data and directly outputs the disability level, failing to reveal the deep causal mechanisms between different modalities, resulting in a lack of interpretability and low reliability of the judgment results; the system cannot quantitatively assess the credibility of the judgment results, making it difficult to dynamically adjust subsequent intervention strategies based on the reliability of the judgment, and leading to a high risk of misjudgment and missed judgment in complex scenarios. Therefore, how to overcome the limitations of shallow correlation in the fusion of relevant multi-source data and improve the accuracy and reliability of disability judgment has become an urgent technical problem to be solved.

[0037] Based on this, this application proposes a method for intervening in driver disability. This method acquires multi-source features of the driver, vehicle motion, and environmental perception, providing a comprehensive information foundation for cross-modal causal analysis. Then, based on a pre-defined causal graph model, inference is performed using the causal probability values ​​represented by directed edges to determine the posterior probability of each candidate disability level. This elevates data association from the statistical correlation level to the causal inference level, directly solving the problem that related technologies cannot establish cross-modal causal links, thus enabling the judgment results to have causal interpretability. Next, for any candidate disability level, the causal contribution of each dimension feature to its posterior probability is calculated, quantifying the specific influence of each modality on the judgment result, making the judgment basis traceable, and further enhancing the transparency of causal inference. Finally, the causal contribution is weighted by confidence level to determine the comprehensive confidence level, realizing a quantitative expression of the reliability of the posterior probability, filling the gap in related technologies where the uncertainty of the judgment result cannot be measured, and providing a credible basis for subsequent decision-making. Based on this, the system uses a comprehensive confidence level greater than a threshold as a screening condition and selects the level corresponding to the maximum posterior probability as the target disability state. This ensures that a decision is made only when the judgment result is sufficiently reliable, effectively avoiding misjudgments and misinterventions under low confidence levels. Finally, the system controls the vehicle to execute the intervention strategy corresponding to the level, achieving a precise match between intervention measures and the severity of disability and the reliability of the judgment. This improves the overall accuracy, interpretability, and reliability of driver disability detection and intervention.

[0038] The embodiments of this application are described below with reference to the accompanying drawings.

[0039] Please see Figure 1 , Figure 1 This is a schematic diagram of a driver disability intervention system provided in an embodiment of this application, applied to a vehicle. The driver disability intervention system includes: a multi-source sensing device 101, a controller 102, and an intervention execution device 103. The controller 102 is connected to both the multi-source sensing device 101 and the intervention execution device 103.

[0040] Among them, the vehicles can be, but are not limited to, pure electric vehicles (PEV / BEV), hybrid electric vehicles (HEV), range-extended electric vehicles (REEV), plug-in hybrid electric vehicles (PHEV), and new energy vehicles.

[0041] The multi-source sensing device 101 is a collection of hardware devices used to acquire multi-source sensing features. In one implementation, the multi-source sensing device 101 may include a driver state sensing device, a vehicle motion sensing device, and an environmental sensing device.

[0042] The driver status perception device includes at least one of the following: a biosensor located in the seat, steering wheel, and seatbelt; an in-vehicle camera; and a hand grip sensor. The biosensor can be installed in the driver's seat cushion and backrest, the steering wheel rim and spokes, or the inside of the seatbelt webbing. The in-vehicle camera can be installed above the dashboard, inside the A-pillar, or at the base of the rearview mirror. The vehicle motion perception device includes at least one of the following: a CAN bus interface and an onboard inertial measurement unit. The CAN bus interface is connected via the vehicle's OBD interface or a gateway. The onboard inertial measurement unit can be installed in the center of the vehicle chassis or inside the shark fin antenna on the roof. The environmental perception device includes at least one of the following: an onboard millimeter-wave radar, a lidar, a forward-facing camera, an ambient light sensor, and a rain sensor. The onboard millimeter-wave radar and lidar can be installed in the front bumper, rear bumper, and fenders. The forward-facing camera can be installed at the top inside the windshield. The ambient light sensor can be installed at the base of the rearview mirror inside the windshield. The rain sensor can be installed at the top inside the windshield. It should be understood that the above are only some embodiments, and this application does not make any specific limitations.

[0043] In one possible implementation, the multi-source sensing device 101 is configured to acquire multi-source sensing features, including driver features, vehicle motion features, and environmental sensing features, and send these features to the controller 102. Specifically, the driver state sensing device collects driver features, including at least one of heart rate, respiratory rate, skin conductance, facial expression, eyelid opening / closing, gaze direction, head posture, steering wheel grip force, and grip position. The vehicle motion sensing device collects vehicle motion features, including at least one of steering angle, steering angular velocity, braking pressure, accelerator pedal opening, vehicle speed, yaw rate, and lateral acceleration. The environmental sensing device collects environmental sensing features, including at least one of road curvature, slope, lane line clarity, traffic participant type and motion state, weather conditions, and lighting conditions.

[0044] The controller 102 can be a vehicle domain controller, a central computing platform, or a standalone electronic control unit, and can be installed in locations such as behind the glove box, below the center console, or under the seats. The controller 102 includes a processor and a data storage unit. The processor integrates at least one of a central processing unit, a graphics processing unit, or a neural network processor. The data storage unit is used to store preset causal graph models, driver profile data, a library of erroneous intervention cases, and intervention control parameters. It should be understood that the above are only some embodiments, and this application does not specifically limit the scope of the invention.

[0045] As one possible implementation, the controller 102 is configured to: determine the posterior probability value corresponding to any candidate disability state level from a preset causal graph model based on multi-source sensing features; for any candidate disability state level, determine the causal contribution of each dimension in the multi-source sensing features based on the posterior probability value, and determine the comprehensive confidence level by weighting the causal contribution with confidence; among the candidate disability state levels whose comprehensive confidence level is greater than the corresponding preset confidence threshold, select the candidate disability state level corresponding to the maximum posterior probability value and determine it as the target disability state level; and send the intervention instruction corresponding to the target disability state level to the intervention execution device 103.

[0046] Intervention execution device 103 is a collection of hardware devices used to execute tiered intervention strategies. Intervention execution device 103 may include a cockpit interaction execution device, a chassis cooperative execution device, and a cloud-based linkage communication device. The cockpit interaction execution device includes at least one of the following: an instrument panel display or head-up display system, in-vehicle speakers, a steering wheel vibration motor, a seatbelt pretensioner motor, a seat vibration motor, and an in-vehicle fragrance system. The chassis cooperative execution device includes at least one of the following: an electronic throttle controller, an electronic stability program controller, an electronic brake booster or brake precharge pump, an electric power steering controller, and a blind spot monitoring radar. The cloud-based linkage communication device includes at least one of the following: an in-vehicle communication module and a high-precision positioning module.

[0047] In one possible implementation, the intervention execution device 103 is configured to receive intervention commands sent by the controller 102 and execute intervention strategies corresponding to the target disability level according to the intervention commands. Specifically, the cockpit interaction execution device is used to execute at least one cockpit warning operation among visual, auditory, tactile, and olfactory alerts corresponding to the first disability level. The chassis coordination execution device is used to execute at least one vehicle chassis operation among speed limit control, lane keeping enhancement, and active braking preparation corresponding to the second disability level. The cloud-based communication device is used to execute at least one remote support operation among fault data uploading, emergency contact notification, and remote driving takeover preparation corresponding to the third disability level.

[0048] It should be noted that, in the embodiments of this application, Figure 1 The illustrated structure does not constitute a limitation on the driver disability intervention system of this application. The system may include more or fewer components than shown in the figures, or combine or separate certain components, or employ different component arrangements. The components shown in the figures can be implemented in hardware, software, or a combination of both.

[0049] For ease of understanding, the driver disability intervention method provided in this application will be described in detail below with reference to the accompanying drawings.

[0050] Figure 2 This is a flowchart illustrating a driver disability intervention method provided in an embodiment of this application, applied to the controller of the aforementioned driver disability intervention system. The specific steps of the driver disability intervention method include: S201. Obtain multi-source sensing features.

[0051] Among them, multi-source perception features include features in the following dimensions: driver features, vehicle motion features, and environmental perception features.

[0052] Driver characteristics are observational data characterizing a driver's physiological state and behavioral performance. One implementation method involves acquiring driver characteristics using biosensors located in the seat, steering wheel, and seatbelt, and / or in-vehicle cameras and hand-grip sensors.

[0053] As one possible implementation, driver characteristics include at least one of the following: heart rate, respiratory rate, skin conductance response, facial expression, eyelid opening and closing, gaze direction, head posture, steering wheel grip strength and grip position, which can be used as observational evidence for driver nodes in the causal graph model in subsequent steps. The aforementioned characteristics include driver behavioral characteristics, physiological characteristics, etc., and this application does not impose specific limitations on this classification. It should be understood that observational evidence refers to real-time observations used to update the probability distribution of nodes in the causal graph model, that is, instantiating the state of the corresponding node in the causal graph model based on the actually collected driver characteristic data, thereby driving the subsequent causal inference process.

[0054] One approach involves preprocessing, denoising, and feature extraction of the collected raw data to output standardized driver features, which can then be directly used by the causal graph model.

[0055] Vehicle motion characteristics are observational data used to characterize the real-time dynamic state and control inputs of a vehicle. In one implementation, vehicle motion characteristics can be acquired via the vehicle's CAN bus and / or an onboard inertial measurement unit.

[0056] As one possible implementation, vehicle motion characteristics include at least one of the following: steering angle, steering angular velocity, braking pressure, accelerator pedal opening, vehicle speed, yaw rate, and lateral acceleration, which are used as observational evidence for vehicle nodes in the causal graph model in subsequent steps.

[0057] One implementation also acquires information such as the status of the vehicle's driver assistance system, powertrain system, and chassis system. The status of the driver assistance system includes the activation status and alarm information of systems such as adaptive cruise control, lane keeping assist, and automatic emergency braking. The status of the powertrain system includes engine or motor torque and speed. The status of the chassis system includes suspension mode and drive mode.

[0058] One implementation method can parse and format raw vehicle motion data in real time, outputting vehicle motion features for direct use by causal graph models.

[0059] Environmental perception features are observational data used to characterize the vehicle's external driving scene and traffic situation. One implementation method involves acquiring environmental perception features using at least one of onboard millimeter-wave radar, lidar, cameras, and ambient light sensors.

[0060] As one possible implementation, environmental perception features include at least one of the following: road curvature, slope, lane line clarity, traffic participant type and motion state, weather conditions, and lighting conditions. For example, traffic participant type and motion state includes data such as the type, location, and speed of vehicles, pedestrians, cyclists, etc., and lighting conditions include, for example, daytime, nighttime, or tunnels. In one implementation, environmental perception features are used as observational evidence for environmental nodes in a causal graph model in subsequent steps. In another implementation, the raw environmental perception data undergoes target detection, tracking, and classification processing to output environmental state features for direct use by the causal graph model.

[0061] Based on S201, by simultaneously collecting perceptual features from three dimensions—driver physiological behavior, vehicle dynamics, and the external environment—this approach overcomes the limitations of related technologies where data sources are confined to the driver while ignoring vehicle and environmental information. It constructs multimodal observational data covering all elements of the driver, vehicle, and environment. This multi-source perceptual feature enables subsequent causal inference to trace the entire chain from environmental triggers and vehicle responses to changes in driver state, providing a complete evidentiary basis for cross-modal causal modeling and ensuring the comprehensiveness and accuracy of disability assessment from the data source.

[0062] S202. Based on the multi-source perception features, determine the posterior probability value corresponding to any candidate disability level from the preset causal graph model.

[0063] Among them, the candidate disability status level is used to characterize the severity of driver disability.

[0064] As one possible implementation, candidate disability levels may include a first disability level, a second disability level, and a third disability level. A first disability level indicates that the driver is in a state of mild distraction, such as a brief deviation of vision from the road ahead but still possessing some environmental awareness. A second disability level indicates that the driver is in a state of moderate disability, such as slow reaction time, sluggish operation, and a significant decrease in vehicle control. A third disability level indicates that the driver is in a state of severe disability, such as closed eyes, loss of control, or body tilting, having essentially lost control of the vehicle.

[0065] The severity of disability increases sequentially from the first disability level to the second disability level and the third disability level.

[0066] It should be understood that the classification of candidate disability status levels is not limited to the above examples. In practical applications, more or fewer levels can be flexibly set according to factors such as vehicle model positioning and safety strategies.

[0067] In this embodiment of the application, a preset causal graph model is used to characterize a directed acyclic graph of environment nodes, vehicle nodes, driver nodes and disabled state nodes connected by directed edges. Each directed edge represents the causal probability value between two nodes connected in the directed acyclic graph.

[0068] The environment node may include at least one of the following: weather condition sub-node, road condition sub-node, and traffic participant sub-node. The vehicle node may include at least one of the following: steering state sub-node, speed state sub-node, and braking state sub-node. The driver node may include at least one of the following: driver physiology sub-node and driver behavior sub-node. The driver physiology sub-node includes heart rate, respiratory rate, and skin conductance response, while the driver behavior sub-node includes gaze direction, head posture, and steering wheel grip strength.

[0069] As one possible implementation, the environmental perception features obtained in step S201 can be input into the corresponding environmental node, the vehicle motion features into the corresponding vehicle node, and the driver features into the corresponding driver node. For example, weather condition features can be input into the weather condition sub-node, road curvature and slope features into the road condition sub-node, and traffic participant type and motion state features into the traffic participant sub-node; vehicle speed and yaw rate features can be input into the speed state sub-node, steering angle and steering angular velocity features into the steering state sub-node, and braking pressure features into the braking state sub-node; heart rate and respiratory rate features can be input into the driver's physiological sub-node, and gaze direction and head posture features into the driver's behavior sub-node.

[0070] It should be understood that the process of inputting the multi-source sensing features obtained in step S201 into the corresponding nodes in the causal graph model is to instantiate the real-time observation data into observation evidence for each node in the causal graph, so that the original prior probability distribution can be updated, thereby driving subsequent causal inference and posterior probability calculation.

[0071] As one possible implementation, the direction of the directed edges represents the direction of causal influence transmission. Each directed edge is associated with a causal probability value, which represents the conditional probability that a child node will be in a certain state when the parent node is in a specific state. For example, a directed edge from a weather condition child node to a road condition child node represents the causal influence of weather changes on road conditions; a directed edge from a driver's physiology child node to a driver's behavior child node represents the causal influence of changes in physiological state on behavioral performance; and a directed edge from a driver's behavior child node to a disability state node represents the direct influence of abnormal behavior on the disability state.

[0072] It should be understood that the pre-defined causal graph model can be constructed through a combination of domain knowledge guidance and data-driven approaches. First, the node types and initial graph structure are defined based on prior knowledge of the assisted driving domain. Then, causal dependencies between nodes are learned from historical data using structure learning methods such as conditional independence testing. Finally, the topological structure of the causal graph and the causal probability values ​​of each edge are determined.

[0073] In this embodiment of the application, the posterior probability value corresponding to any candidate disability state level is determined based on the causal probability value corresponding to at least one directed edge in the link pointing to the disability state node.

[0074] It should be understood that a link here refers to a directed path that starts from an initial node, propagates step by step along the direction of directed edges, and finally reaches a disabled node. The causal probability values ​​of all directed edges in a link collectively determine the degree of influence of that link on the disabled node. The calculation of the posterior probability value takes into account the influence of all activated links.

[0075] As one possible implementation, based on multi-source sensing features, the posterior probability value corresponding to any candidate disability state level among multiple candidate disability state levels is determined from a preset causal graph model. This includes: determining an initial node set from the preset causal graph model based on multi-source sensing features; determining a set of links from the preset causal graph model with any node in the initial node set as the starting point and pointing to the disability state node; and performing forward propagation reasoning based on the causal probability values ​​corresponding to all directed edges in the link set to determine the posterior probability value of the disability state node being in any candidate disability state level.

[0076] The initial node set includes at least one initial node, and each initial node is any one of the following: an environment node, a vehicle node, and a driver node. The link set includes at least one link.

[0077] One implementation method involves determining an initial node set from a preset causal graph model based on multi-source sensing features, including: comparing the observation data of each dimension in the multi-source sensing features with the normal value range of the corresponding nodes in the preset causal graph model, and determining the nodes whose observation data deviates from the normal value range by more than a preset deviation threshold as initial nodes.

[0078] For example, when a driver's heart rate data exceeds their personalized normal fluctuation range, the heart rate node in the driver's physiological sub-node is determined as the initial node; when the vehicle speed data exceeds the safe speed range under current road conditions, the speed status sub-node in the vehicle node is determined as the initial node; when the lane line clarity in the environmental perception features is lower than a preset threshold, the road condition sub-node in the environment node is determined as the initial node. Through these methods, real-time observation data is transformed into activated initial evidence nodes in the causal graph model.

[0079] One implementation method involves determining a set of links from any node in the initial node set as the starting point to a disabled node in a preset causal graph model, based on an initial node set. This includes: in the preset causal graph model, taking each initial node in the initial node set as the starting point, performing a depth-first or breadth-first traversal along the direction of the directed edges to search for all paths that can reach the disabled node, treating each path as a link, and the set of all links constitutes a link set.

[0080] For example, starting with the weather condition sub-node, a complete causal link can be formed by following the path from weather condition to road condition, road condition to vehicle status, vehicle status to driver behavior, and driver behavior to the disability status node. Through this method, all activation paths that may have a causal impact on the disability status can be extracted.

[0081] One implementation method involves forward propagation reasoning based on the causal probability values ​​corresponding to all directed edges in the link set to determine the posterior probability value of a disabled node in any candidate disabled state level. This includes: employing a belief propagation algorithm, using the observation values ​​of each initial node in the initial node set as evidence, and propagating probability information forward level by level along each link in the link set. In each level of propagation, the state probability distribution of the next-level node is calculated based on the current node's state probability distribution and the causal probability values ​​on the corresponding directed edges. Finally, the probability information transmitted by all links is aggregated at the disabled node to obtain the posterior probability value of the disabled node in each candidate disabled state level.

[0082] For example, suppose there are two activation links. Link A has a probability of 0.9 for slippery roads due to heavy rain, a probability of 0.8 for increased vehicle speed fluctuations due to slippery roads, a probability of 0.6 for increased vehicle speed fluctuations leading to visual deviation, and a probability of 0.7 for visual deviation leading to mild disability. The probability of this link contributing to mild disability is 0.9 multiplied by 0.8 multiplied by 0.6 multiplied by 0.7, resulting in 0.3024. Link B has a probability of 0.8 for abnormal heart rate leading to abnormal head posture, and a probability of 0.7 for abnormal head posture leading to mild disability. The probability of this link contributing to mild disability is 0.8 multiplied by 0.7, resulting in 0.56. When aggregating the two links, multiplying the contribution probabilities of each link yields an unnormalized joint support of 0.3024 multiplied by 0.56, approximately equal to 0.1693. Similarly, after calculating the joint support for moderate and severe disability, the levels are normalized to obtain the posterior probability values ​​of the disabled state node in each candidate disability state level.

[0083] It should be understood that forward propagation reasoning refers to the process of inferring from known evidence nodes step by step along the causal direction towards the target node. Unlike general correlation calculations, forward propagation reasoning in causal graphs follows a strict causal direction, which can distinguish causal effects in different directions (e.g., environmental changes lead to driver incapacity, and driver incapacity leads to abnormal vehicle conditions), thereby avoiding the interference of spurious associations on the judgment results.

[0084] For example, please refer to Figure 3This is a schematic diagram of a multimodal causal graph provided in an embodiment of this application. The multimodal causal graph includes a disability state node D, a driver state causal chain, a vehicle state causal chain, and an environmental state causal chain. The disability state node D has values ​​for normal, first disability level, second disability level, and third disability level. The driver state causal chain includes a driver feature node E_driver, which contains visual feature sub-nodes, physiological feature sub-nodes, and behavioral feature sub-nodes. The vehicle state causal chain includes a vehicle motion feature node E_vehicle, which contains chassis feature sub-nodes, electrical feature sub-nodes, and steering feature sub-nodes. The environmental state causal chain includes an environmental perception feature node E_env, which contains weather feature sub-nodes, road condition feature sub-nodes, and illumination feature sub-nodes. Cross-modal causal interactions are represented by directed edges. For example, a weather feature sub-node pointing to a physiological feature sub-node indicates the causal relationship of severe weather causing driver stress, and a physiological feature sub-node pointing to the disability state node D indicates the direct impact of abnormal physiological indicators on the disability state. Similarly, a steering feature sub-node pointing to a behavioral feature sub-node indicates the causal relationship of steering failure causing the driver to correct their behavior, and a behavioral feature sub-node pointing to the disability state node D indicates the direct impact of abnormal steering wheel operation on the disability state. The directed edges between nodes form connections from E_driver, E_vehicle, and E_env to the disability state node D, respectively, with each directed edge representing the causal probability value between the two connected nodes.

[0085] It should be understood that, Figure 3 This is merely an exemplary structure provided for the embodiments of this application. In actual applications, the connection relationships of nodes and directed edges can be increased or decreased according to specific needs.

[0086] Based on S202, and using a pre-defined causal graph model, inference is performed based on the causal probability values ​​represented by directed edges, elevating the correlation between multi-source sensing features and disability states from the statistical correlation level to the causal reasoning level. By determining the set of links pointing to disability state nodes based on the initial node set and performing forward propagation inference along the links, a full-link causal tracing is achieved from environmental triggers and vehicle response to changes in driver state. This makes the judgment results causally interpretable, solving the problem in related technologies that simply splice or weight multi-source data without revealing cross-modal causal mechanisms. Simultaneously, the posterior probability values ​​of each candidate disability state level are output instead of a single discrete label, providing a foundation for quantifying the reliability of the judgment in subsequent steps.

[0087] S203. For any candidate disability status level among multiple candidate disability status levels: determine the causal contribution of each dimension in the multi-source perception features based on the posterior probability value corresponding to the candidate disability status level, and weight the causal contribution of each dimension with confidence to determine the comprehensive confidence.

[0088] Among them, the causal contribution degree is used to characterize the degree of influence of each dimension on the posterior probability value corresponding to the candidate disability status level; the comprehensive confidence degree is used to characterize the reliability of the posterior probability value corresponding to the candidate disability status level.

[0089] It should be understood that the posterior probability value itself only reflects the probability that a disabled node is at a certain level under the current observational evidence, but does not reveal which dimensions of features drive this result, nor can it express the credibility of the probability value itself.

[0090] Therefore, for each of the multiple candidate disability status levels, the causal contribution of each dimension is calculated and weighted by confidence level to obtain the comprehensive confidence level corresponding to that candidate disability status level. The calculation method for one candidate disability status level is as follows, and the calculation method for other candidate disability status levels is the same. The causal contribution can be used to trace the root cause of the judgment result and clarify the contribution of each dimension's characteristics to the judgment conclusion; the comprehensive confidence level can be used to assess the reliability of the disability status judgment, providing a basis for subsequent decision-making on whether to adopt the judgment result.

[0091] As one possible implementation method, the causal contribution of each dimension in the multi-source sensing features is determined based on the posterior probability value corresponding to the candidate disability status level, including: applying a preset increment to the first dimension in the multi-source sensing features, and determining the difference between the posterior probability values ​​before and after applying the preset increment; the first dimension is any dimension in the multi-source sensing features; the ratio of the difference to the preset increment is determined as the causal contribution of the first dimension.

[0092] The preset increment can be a small positive value, such as one percent of the current value of the feature in that dimension or a fixed step size, but this application does not make any specific limitation on it.

[0093] One implementation method allows for the following relationships to be satisfied by the causal contribution of each dimension in multi-source sensing features:

[0094] in, Indicates the first Causal contribution of each dimension feature; Indicates the first One candidate disability status level; Indicates the first Features in one dimension; Indicates keeping except the first The other dimensions besides the first dimension are used as the baseline value; the meaning of this partial derivative is that, while keeping the other dimensions unchanged, the disabled state node is at the first dimension. The posterior probability value of the candidate disability status level for the th candidate disability status level The partial derivative of the i-th dimension feature, characterizing the i-th The change in posterior probability value caused by a unit change in a feature dimension.

[0095] In practical calculations, the aforementioned partial derivatives can be approximated by applying a preset increment and calculating the difference between the posterior probability values. That is, the causal contribution of each dimension in multi-source perceptual features can satisfy the following relationships:

[0096] in, Indicates the first Causal contribution of each dimension feature; Indicates the first One candidate disability status level; Indicates the first Current observations of each dimension of features; Indicates the first Preset increments applied to each dimension of features; Indicates keeping except the first Features in all dimensions except for one dimension remain unchanged; This indicates that, given the current feature observations across all dimensions, the disabled state node is at the [number]th [position]. The posterior probability value of each candidate disability status level; Indicates only for the first After applying a preset increment to the feature dimension, the disabled state node is in the first position. The posterior probability value of each candidate disability status level.

[0097] It should be understood that the numerator of this formula is the change in the posterior probability value before and after applying the preset increment, and the denominator is the preset increment, which represents the overall change in the posterior probability value before and after applying the preset increment. The dimensional feature for the th The marginal influence of the posterior probability of each candidate disability level. The larger the absolute value of the causal contribution, the stronger the causal influence of this feature on the determination of the candidate disability level; the sign of the causal contribution indicates the direction of influence, a positive value means that increasing the value of this feature will increase the posterior probability of the disability level, and a negative value means that increasing the value of this feature will suppress the posterior probability of the disability level.

[0098] In this embodiment, by applying incremental perturbations to each dimension of features and observing the changes in posterior probabilities, the marginal causal effect of each mode on the judgment result can be accurately quantified, making the basis for disability judgment traceable and interpretable. This solves the problem in related technologies where the judgment process is like a black box and the contribution of each factor cannot be known.

[0099] As one possible implementation method, the causal contribution corresponding to each dimension is weighted by confidence level to determine the comprehensive confidence level, including: obtaining the preset weights corresponding to each dimension in the multi-source perception features; and weighting and summing the causal contribution corresponding to each dimension with the preset weights corresponding to each dimension to obtain the comprehensive confidence level corresponding to the candidate disability status level.

[0100] In one implementation, the preset weight corresponding to each dimension of the multi-source sensing features can be determined based on at least one of the following: the historical detection accuracy of the sensor corresponding to that dimension, the signal-to-noise ratio under the current operating conditions, or the prior correlation strength between that dimension and the disability state. For example, under conditions of insufficient light at night, the preset weight of visual features (such as gaze direction and facial expression) can be reduced, while the preset weight of physiological features (such as heart rate and respiratory rate) can be increased accordingly.

[0101] Another implementation approach is to dynamically update the preset weights corresponding to each dimension in the multi-source sensing features based on the confidence level of each dimension. The updated preset weights satisfy the following relationship:

[0102] in, Indicates the updated number Preset weights for each dimension; Indicates the number before the update Preset weights for each dimension; Indicates the first The confidence values ​​for each dimension range from 0 to 1 and are determined based on at least one of the following: the current operating condition of the sensor corresponding to the feature of that dimension, the historical detection accuracy, or the prior correlation strength between the feature of that dimension and the disability state. The denominator represents the total number of dimensions of the multi-source sensing features; This represents the sum of weighted confidence scores across all dimensions, used for normalization.

[0103] It should be understood that this formula weights the original weights by using the confidence level of each dimension as an adjustment factor, and then normalizes the weights so that dimensions with high confidence levels receive larger weights and dimensions with low confidence levels receive smaller weights, with the sum of all updated weights being 1. This dynamic weight adjustment mechanism enables adaptive fusion of cross-modal evidence, automatically adjusting the contribution ratio of each dimension in the overall confidence calculation based on changes in the real-time data quality of each sensor.

[0104] For example, assume the preset weight of the driver feature dimension before update is 0.8 with a confidence level of 0.75; the preset weight of the vehicle motion feature dimension before update is 0.15 with a confidence level of 0.6; and the preset weight of the environmental perception feature dimension before update is 0.05 with a confidence level of 0.5. Then the sum of the numerators is 0.8 × 0.75 + 0.15 × 0.6 + 0.05 × 0.5 = 0.6 + 0.09 + 0.025 = 0.715. The pre-set weights for the updated driver feature dimension are (0.8 × 0.75) / 0.715 = 0.6 / 0.715, resulting in a weight of approximately 0.839; the pre-set weights for the updated vehicle motion feature dimension are (0.15 × 0.6) / 0.715 = 0.09 / 0.715, resulting in a weight of approximately 0.126; and the pre-set weights for the updated environmental perception feature dimension are (0.05 × 0.5) / 0.715 = 0.025 / 0.715, resulting in a weight of approximately 0.035. It is evident that the driver feature dimension, with the highest confidence, receives a larger weight, while the environmental perception feature dimension, with its lower confidence, receives a correspondingly smaller weight, thus tilting the fusion weights towards the higher confidence modality.

[0105] One implementation involves weighting and summing the causal contribution of each dimension with its corresponding preset weight, which satisfies the following relationship: .in, Indicates the first The overall confidence level corresponding to each candidate disability status level; This represents the total number of dimensions of the multi-source sensing features; Indicates the first Preset weights for each dimension of features; Indicates the first The absolute value of the causal contribution of each dimension feature. The higher the overall confidence level, the more reliable the posterior probability value corresponding to the candidate disability level.

[0106] For example, assuming the preset weight for the driver feature dimension is 0.5, the preset weight for the vehicle motion feature dimension is 0.3, and the preset weight for the environmental perception feature dimension is 0.2, for the first candidate disability level, if the absolute value of the causal contribution of the driver feature is 0.4, the absolute value of the causal contribution of the vehicle motion feature is 0.2, and the absolute value of the causal contribution of the environmental perception feature is 0.1, then the overall confidence level of this candidate disability level is 0.5 multiplied by 0.4 plus 0.3 multiplied by 0.2 plus 0.2 multiplied by 0.1, resulting in 0.28.

[0107] As another possible implementation, each dimension includes at least one sub-dimension. For example, the driver characteristic dimension includes sub-dimensions such as heart rate, gaze direction, and steering wheel grip strength. One implementation method is to first determine the causal contribution of each sub-dimension according to the above method, then sum the causal contributions of each sub-dimension according to the sub-dimension weights to obtain the causal contribution of that dimension, and then sum it with the preset weights at the dimension level to determine the overall confidence level.

[0108] Based on S203, this implementation method achieves attribution analysis of the judgment results by calculating the causal contribution of each dimension feature to the posterior probability value, making the judgment basis for each candidate disability status level transparent and traceable. Furthermore, by weighting the causal contribution with confidence levels, the method incorporates the differences in credibility among different modalities, outputting a comprehensive confidence level that quantifies the reliability of the judgment. This process addresses the dual shortcomings of related technologies where disability judgment results cannot explain their origin or measure their credibility, providing a quantitative basis for subsequent reliability-based screening and decision-making.

[0109] S204. Among the candidate disability status levels whose overall confidence level is greater than the corresponding preset confidence threshold, select the candidate disability status level corresponding to the maximum posterior probability value and determine it as the target disability status level.

[0110] The pre-set confidence threshold can be set according to factors such as vehicle model positioning, safety strategy level or regulatory requirements, and this application does not impose specific restrictions on it.

[0111] As one possible implementation method, the target disability status level can be one of the candidate disability status levels. The candidate disability status level with a comprehensive confidence level greater than the corresponding preset confidence threshold is determined as a reliable candidate level, thereby obtaining a set of reliable candidate levels. Then, the candidate disability status level with the largest posterior probability value is selected from the set of reliable candidate levels as the target disability status level.

[0112] For example, suppose the candidate disability status levels include a first disability level, a second disability level, and a third disability level. The overall confidence level of the first disability level is 0.85 and its posterior probability is 0.6; the overall confidence level of the second disability level is 0.72 and its posterior probability is 0.3; and the overall confidence level of the third disability level is 0.45 and its posterior probability is 0.1. If the preset confidence thresholds for the first, second, and third disability levels are all 0.7, then the overall confidence levels of both the first and second disability levels are greater than their respective preset confidence thresholds, while the overall confidence level of the third disability level is less than its corresponding preset confidence threshold and is therefore filtered out. Among the remaining first and second disability levels, the posterior probability of the first disability level (0.6) is greater than the posterior probability of the second disability level (0.3), therefore the first disability level is determined as the target disability status level. The above embodiments are for illustrative purposes only and do not impose specific limitations on the specific solutions of this application.

[0113] It should be understood that this implementation method ensures the reliability and accuracy of the judgment results through a dual screening mechanism. First, candidate disability status levels with insufficient reliability are filtered out based on a comprehensive confidence level greater than a preset confidence threshold, avoiding high-risk decisions when evidence is insufficient. Then, the target disability status level with the highest posterior probability value is selected from the remaining reliable candidate status levels, ensuring that the output is the most likely and credible disability status under the current observational evidence. This mechanism effectively avoids misjudgments and misinterpretations under low confidence levels, improving the safety and robustness of disability determination.

[0114] S205. Control the vehicle to execute the intervention strategy corresponding to the target disability level.

[0115] It should be understood that after determining the target disability level, the vehicle will automatically perform corresponding operations according to the preset intervention strategy corresponding to that level to ensure driving safety.

[0116] As one possible implementation, the multiple candidate disability status levels include a first disability level, a second disability level, and a third disability level; wherein the severity of disability corresponding to the first disability level, the second disability level, and the third disability level increases sequentially.

[0117] It should be understood that the first level of disability corresponds to mild distraction, the second level to moderate disability, and the third level to severe disability. One approach is to design intervention strategies that prioritize low-invasive interventions, gradually escalating the intensity if these are ineffective.

[0118] One implementation method involves an intervention strategy corresponding to the first level of disability, which includes sending an attention alert to the pilot through at least one of the following cockpit alert operations: visual alerts, auditory alerts, tactile alerts, and olfactory alerts.

[0119] The visual alerts include warning icons and text prompts displayed on the dashboard screen, showing the disability risk level and suggested coping strategies. The auditory alerts include warning sounds or voice announcements played through the vehicle's speakers, with the voice content customized according to the type of disability. The tactile alerts include tactile feedback provided through seatbelt vibration, steering wheel vibration, and seat vibration, with vibration patterns adjusted according to the disability level. The olfactory alerts include releasing invigorating scents, such as mint or lemon, through the vehicle's fragrance system to stimulate the driver's senses.

[0120] For example, when the target disability level is the first disability level, an attention reminder icon can be displayed on the instrument panel, while a voice prompt "Your attention may be distracted, please concentrate on driving" can be played through the speaker, and the driver can be alerted by slight vibration of the steering wheel.

[0121] One implementation method involves an intervention strategy corresponding to the second disability level, which includes adjusting vehicle dynamics through at least one of the following vehicle chassis operations: speed limit control, lane keeping enhancement, and active braking preparation, to assist in stable vehicle operation.

[0122] Speed ​​limit control includes automatically limiting the vehicle's maximum speed to prevent speeding due to slow driver reaction. Lane keeping assist enhancement includes increasing the intervention of the lane keeping assist system to improve lane centering ability and prevent the vehicle from deviating from its lane. Active braking preparation includes a pre-charge braking system to shorten response time during emergency braking and provide the driver with a braking backup.

[0123] For example, when the target disability level is the second disability level, the system can automatically limit the vehicle speed to within 80% of the current road speed limit, while increasing the steering torque of the lane keeping assist and pre-charging the brake fluid pressure in preparation for emergency braking.

[0124] Intervention strategies for the second level of disability also include power steering backup and enhanced blind spot monitoring. Power steering backup involves switching to Sport mode power steering to improve steering response and assist the driver in steering maneuvers. Enhanced blind spot monitoring involves increasing the sensitivity and range of the blind spot monitoring system's alarms to alert the driver to surrounding vehicles.

[0125] For example, when the target disability level is the second disability level, the system can automatically limit the vehicle speed to within 80% of the current road speed limit, while increasing the steering torque of the lane keeping assist and pre-charging the brake fluid pressure in preparation for emergency braking.

[0126] One implementation method includes an intervention strategy corresponding to the third level of disability, which includes performing at least one remote support operation among fault data uploading, emergency contact notification, and remote driving takeover preparation.

[0127] The fault data upload includes encrypting and uploading disability detection data and intervention records to a cloud server for subsequent analysis and algorithm optimization. Emergency contact notification includes automatically sending a notification containing the vehicle's location and status to pre-set emergency contacts when severe disability is detected and the driver remains unresponsive. Remote driving takeover preparation includes, when technically feasible, preparing to enter remote driving mode for a safety driver in the background to take over vehicle control.

[0128] For example, when the target disability level is the third level, the system can automatically call the emergency contact, upload the current vehicle location and driver status data to the cloud, and enter a remote takeover standby state.

[0129] Intervention strategies for the third level of disability also include contacting customer service centers and providing road condition information. Contacting customer service centers involves automatically dialing the customer service hotline, where customer service personnel communicate with the driver via the in-vehicle voice system to provide remote assistance. Providing road condition information includes sending relevant road condition updates based on the driver's condition, such as the location of the next service area or gas station, to suggest rest stops for the driver.

[0130] For example, when the target disability level is the third level, the system can automatically call the emergency contact, upload the current vehicle location and driver status data to the cloud, and enter a remote takeover standby state.

[0131] As one possible implementation method, controlling the vehicle to execute the intervention strategy corresponding to the target disability level includes: determining the disability detection duration based on the comprehensive confidence level corresponding to the target disability level; and executing the intervention strategy corresponding to the target disability level when the duration of monitoring the driver's operation response exceeds the disability detection duration.

[0132] Among these findings, the overall confidence level is negatively correlated with the duration of disability detection. It should be understood that a higher overall confidence level indicates a higher reliability of the judgment result and a shorter confirmation waiting time, allowing for rapid intervention to ensure safety; conversely, a lower overall confidence level indicates greater uncertainty in the judgment result, requiring a longer waiting time to collect more evidence and avoid misinterpretation.

[0133] The driver's operational response status includes at least one of the following: driver operation of the steering wheel, operation of the brake pedal, operation of the accelerator pedal, pressing of the confirmation button, or touching of the intervention cancellation control. It should be understood that when any of the above operational responses is detected, it indicates that the driver has the ability to actively control the vehicle, and the disability detection duration should be reset or the current intervention procedure should be cancelled.

[0134] One implementation involves recalculating the disability detection duration when the change in overall confidence exceeds a preset change threshold; and controlling the vehicle to end the disability detection duration and canceling the intervention strategy corresponding to the target disability level when the duration corresponding to the monitored driver's operational response is less than or equal to the disability detection duration.

[0135] It should be understood that the overall confidence level may change due to new observation data. When the change exceeds the preset threshold, it indicates that the uncertainty of the current judgment result has changed significantly, and the disability detection duration needs to be recalculated to adapt to the new judgment conditions. During the disability detection duration, the system continuously monitors whether the driver makes any active actions. If the driver responds effectively within the window period, it indicates that the driver still has the ability to actively control the vehicle. At this point, ending the wait and canceling the intervention can avoid unnecessary disturbance to the driver, while maximizing both driving safety and interactive experience.

[0136] One implementation method stipulates that the disability detection duration satisfies the following relationship:

[0137] in, This refers to the duration of disability detection. The preset detection duration is related to the target disability level. For overall confidence level; To pre-set the reliability threshold; This is the confidence sensitivity coefficient, used to control the degree of influence of the overall confidence level on the window duration; The evidence quality coefficient is determined based on the pre-set reliability of each dimension in the multi-source perception features.

[0138] For example, the preset detection duration is related to the target disability level, that is, the higher the disability level, the shorter the preset detection duration. For example, the first disability level corresponds to 5 seconds, the second disability level corresponds to 3 seconds, and the third disability level corresponds to 0 seconds for immediate triggering. The preset reliability threshold is the pre-defined standard for judgment, with a value range of 0.5 to 0.9, and a default value of 0.7. This is the confidence sensitivity coefficient, used to control the degree of influence of the overall confidence level on the window duration. The value range can be from 0.5 to 2.0, with a default value of 1.0. The larger the sensitivity coefficient, the more significant the impact of the confidence level change on the window duration. The evidence quality coefficient is determined based on the confidence level of each dimension in the multi-source perception features. It reflects the reliability of the input data and can range from 0.5 to 2.0. The higher the evidence quality, the smaller the coefficient and the shorter the window duration.

[0139] It should be understood that the above formula allows the disability detection duration to be dynamically adjusted based on overall confidence level, evidence quality, and disability level. At high confidence levels, the window is shortened for a rapid response, while at low confidence levels, the window is extended to collect more information, achieving a balance between safety and interaction efficiency. During the confirmation window, the driver's response status and confidence level changes are continuously monitored. When a driver response is detected, the confirmation window is immediately terminated and intervention is canceled. The window duration is recalculated when the overall confidence level changes significantly.

[0140] As one possible implementation method, the evidence quality coefficient The following relationship must be satisfied:

[0141] in, The quality of evidence coefficient, ranging from 0.5 to 2.0, reflects the reliability of the input data. Higher quality evidence indicates greater reliability. The smaller the value, the shorter the corresponding disability detection time. The evidence consistency index, with a value ranging from 0 to 1, is used to measure the degree of consistency among evidence in various dimensions of multi-source perception features. The sufficiency of evidence index has a value range of 0 to 1 and is used to measure the proportion of valid evidence to the total amount of evidence.

[0142] One implementation method is the evidence consistency index. It can be determined based on the entropy of the evidence probability distribution, satisfying... ,in The entropy of the probability distribution of evidence in each dimension. The maximum entropy is denoted as σ. When the evidence across all dimensions is highly consistent, the entropy value is small, and the consistency index approaches 1; when the evidence across all dimensions is contradictory, the entropy value is large, and the consistency index approaches 0.

[0143] One implementation method is the evidence sufficiency index. It can be determined based on the ratio of the number of valid pieces of evidence to the total number of pieces of evidence, satisfying the following conditions. ,in This refers to the number of valid pieces of evidence in the current multi-source sensing features that meet the preset conditions for data quality. This represents the total number of dimensions of the multi-source sensing features. The more valid evidence there is, the closer the sufficiency index is to 1.

[0144] It should be understood that by decomposing the evidence quality coefficient into the reciprocal of the product of the consistency index and the sufficiency index, the calculation of the evidence quality coefficient comprehensively considers the degree of consistency among evidence in each dimension and the coverage ratio of valid evidence. When the evidence in each dimension is highly consistent and there is sufficient valid evidence, The value approaches 1, and the window duration is close to the baseline value; when the evidence is contradictory or there is insufficient valid evidence, As the value increases, the window duration is extended accordingly to collect more observational data and avoid making misjudgments when the evidence is unreliable.

[0145] For example, an adaptive confirmation window calculation and human-computer interaction process provided in this application embodiment is as follows: First, the comprehensive confidence score is calculated based on the consistency index of evidence in each dimension of multi-source perception features. and sufficiency index The overall confidence level is obtained by weighting and summing the causal contribution of each dimension feature with a preset weight. Then, the preset detection duration is obtained, and the corresponding level of disability is determined based on the target disability status. The first disability level corresponds to 5 seconds, the second disability level corresponds to 3 seconds, and the third disability level corresponds to 0 seconds (instant trigger). This example uses the first disability level. Seconds. Next, the disability detection duration is calculated, taking into account the overall confidence level and the preset confidence threshold. and confidence sensitivity coefficient Substitute into the formula The detection time is shortened by a fraction of a second compared to the preset detection duration, demonstrating the adaptive adjustment effect of rapid response under high confidence levels. Subsequently, cockpit interaction is performed within the disability detection duration. The interaction method is selected based on the overall confidence level range. In this example, the overall confidence level is 0.83, which is greater than 0.75, belonging to the high confidence range, and a combination of voice inquiry and instrument text prompts is used. Finally, the system waits for the driver's response. If the driver provides a valid response, the status is determined to be normal and the warning is exited. If there is no response after the timeout, disability is confirmed and the corresponding intervention strategy is triggered. If the response is abnormal, a self-check is triggered, thus achieving adaptive adjustment of the confirmation window and coordinated hierarchical interaction while ensuring safety.

[0146] As one possible embodiment, the method further includes: acquiring current multi-source sensing features; determining a comprehensive intervention effect score based on the deviation between the multi-source sensing features and the current multi-source sensing features; and updating the intervention control parameters through a preset reinforcement learning algorithm with the goal of maximizing the comprehensive intervention effect score.

[0147] Among them, the current multi-source perception features are the multi-source perception features after implementing the intervention strategy corresponding to the target disability status level.

[0148] One implementation method involves determining a comprehensive intervention effect score based on the deviation between multi-source perception features and the current multi-source perception features. This score includes: assessing indicators of driver recovery after intervention, such as whether the driver's gaze returned to the front, whether the head posture was normal, and whether the operation was timely; assessing indicators of vehicle improvement after intervention, such as whether the vehicle returned to the center of the lane, whether the speed returned to normal, and whether dangerous maneuvers were avoided; and assessing indicators of driver response quality, including response time, response method, and response completeness. These indicators are weighted and fused, then normalized to calculate the comprehensive intervention effect score, which ranges from 0 to 1, with higher scores indicating better intervention effects.

[0149] One implementation method pre-determines the reinforcement learning algorithm, including a proximal policy optimization algorithm or a deep deterministic policy gradient algorithm. The comprehensive score of the intervention effect is used as the reward signal, and the parameters of the intervention policy are used as the weights of the policy network. The policy parameters are updated using a policy gradient method, with the formula being: the new policy parameters equal to the old policy parameters plus the learning rate multiplied by the gradient of the reward function with respect to the policy parameters. Through online learning, the policy parameters can be dynamically adjusted based on actual results, gradually optimizing the performance of the intervention policy.

[0150] For example, in one implementation, the reinforcement learning algorithm is pre-defined to satisfy the following relationship:

[0151] in, This indicates the updated policy parameters; This indicates the strategy parameters before the update; This represents the learning rate, used to control the step size of parameter updates, and its value can range from 0.001 to 0.1. Representing the reward function For strategy parameters The gradient is used to determine the direction and magnitude of parameter updates. Policy parameters. Includes at least one of the following: preset detection duration, preset reliability threshold, confidence sensitivity coefficient, and evidence quality coefficient.

[0152] One implementation method, reward function Based on the comprehensive score of intervention effect, the following relationship is satisfied:

[0153] in, The comprehensive score for the intervention effect ranges from 0 to 1 and is calculated by weighting the driver's post-intervention recovery index, vehicle condition improvement index, and driver response quality index. This is the cost coefficient for erroneous intervention, used to characterize the negative impact of erroneous intervention. This is the cost coefficient for missed detection, used to characterize the safety risks caused by the failure to detect disability status in a timely manner; and These are the cost weighting coefficients for false intervention and false detection, used to balance the trade-off between false intervention and false detection.

[0154] It should be understood that this formula updates the policy parameters along the gradient direction of the reward function using the policy gradient method, allowing the policy parameters to be dynamically adjusted based on the actual effect of each intervention. When the overall score of the intervention effect is high, the reward function value increases, and the policy parameters continue to be optimized in the current direction; when a false intervention occurs, the cost of the false intervention increases, the reward function value decreases, and the policy parameters are adjusted in the direction of reducing the probability of false intervention; when a false negative occurs, the cost of the false negative increases, the reward function value decreases, and the policy parameters are adjusted in the direction of improving detection sensitivity. Through the above online learning process, the policy parameters are gradually optimized, achieving adaptive evolution from a general policy to a personalized policy.

[0155] The intervention control parameters include at least one of the following: preset detection duration, preset reliability threshold, confidence sensitivity coefficient, and evidence quality coefficient.

[0156] One implementation involves updating the preset detection time in the direction of shortening the preset detection time if several post-prognosis driver state recovery indicators show a long response time, so that subsequent interventions can be triggered earlier; if several post-prognosis driver response quality indicators show that the driver actively cancels the intervention, the preset detection time is extended to give the driver more time for autonomous response.

[0157] If, for a pre-set confidence threshold, several prognostic driver state recovery indicators show that the driver state has not improved or has not improved significantly, indicating that there may be a false intervention, then the system is updated along the gradient direction of increasing the pre-set confidence threshold, so that subsequent judgments need to reach a higher confidence standard to trigger intervention; if several prognostic vehicle state improvement indicators show that the vehicle state has improved significantly, indicating that the intervention is effective, then the pre-set confidence threshold can be appropriately reduced to improve the system's response sensitivity.

[0158] Regarding the confidence sensitivity coefficient, if the fluctuation range of the overall confidence level is large in practical applications, causing the window duration to change frequently, then the coefficient should be updated in the direction of decreasing the confidence sensitivity coefficient to make the window duration change smoother; if the window duration is still too long when the overall confidence level is high, resulting in untimely intervention, then the coefficient should be updated in the direction of increasing the confidence sensitivity coefficient to enhance the control of the confidence level on the window duration.

[0159] Regarding the evidence quality coefficient, if a post-intervention assessment reveals significant noise in sensor data for a certain dimension, the coefficient is updated in the direction of increasing the evidence quality coefficient for that dimension to reduce the impact of noisy data on the window duration. If the overall data quality for each dimension is high, the coefficient is updated in the direction of decreasing the evidence quality coefficient to shorten the window duration and improve response speed.

[0160] It should be understood that the above-mentioned intervention control parameters are only some embodiments of this application. In actual applications, the intervention control parameters may also include other adjustable parameters related to the intervention strategy, and this application does not impose specific limitations on them.

[0161] Through the online learning process described above, each intervention control parameter can be dynamically adjusted along the gradient direction of the reward function based on the actual effect of each intervention, gradually optimizing the performance of the intervention strategy and achieving adaptive evolution from a general strategy to a personalized strategy.

[0162] As one possible embodiment, the method further includes a self-correction step for misintervention: detecting misintervention conditions and triggering self-correction measures when the misintervention conditions are met, so as to avoid the impact of misintervention on user experience.

[0163] The conditions for detecting erroneous intervention include at least one of the following: Condition 1: The driver's condition returns to normal after intervention and no obvious signs of disability are detected, indicating that the erroneous intervention may be due to misjudgment; Condition 2: The driver does not respond to the intervention, but the subsequent judgment results are consistently normal for more than a preset threshold, such as being judged as normal for 10 consecutive frames, indicating that the erroneous intervention may be due to an occasional anomaly; Condition 3: The driver has a negative response to the intervention, such as a voice complaint or manually canceling the intervention operation, indicating that the intervention method may be inappropriate.

[0164] One implementation involves triggering self-correction measures when any of the aforementioned erroneous intervention conditions are detected. These self-correction measures include at least one of the following: revoking the current intervention, i.e., immediately stopping ongoing intervention operations such as visual alerts, auditory alerts, and speed limit control, and restoring the vehicle to normal control; adjusting the parameters of subsequent intervention strategies, such as increasing the preset confidence threshold to reduce the probability of future misjudgments, extending the preset detection duration to increase evidence collection time, or decreasing the confidence sensitivity coefficient to make window duration changes smoother; apologizing to the driver and explaining the reason, explaining to the driver via in-vehicle voice or display screen that the intervention was erroneously triggered and expressing apology; and recording the erroneous intervention case for subsequent analysis, storing the multi-source perception features, judgment process data, and the reason for the erroneous intervention in a case library as negative samples for a preset reinforcement learning algorithm to optimize strategy parameters.

[0165] It should be understood that the self-correction mechanism for misintervention enables the system to automatically identify and correct erroneous interventions, learn from individual misjudgments, continuously reduce the rate of misintervention, and increase drivers' trust and acceptance of the system. Together with the closed-loop feedback mechanism for intervention effects, it constitutes a complete self-learning optimization system.

[0166] As one possible embodiment, the method also includes a driver profile creation and updating step: creating and continuously updating a personalized driver profile to achieve an adaptive evolution of intervention strategies from general to personalized.

[0167] One implementation method involves establishing a driver profile that includes at least one of the following: Physiological baseline acquisition: By monitoring the driver's physiological indicators under normal conditions over a long period, a personalized physiological baseline is established. Different drivers have different physiological baselines; for example, the normal fluctuation range of heart rate and the baseline value of respiratory rate vary from person to person. By collecting physiological data from multiple normal driving sessions, the driver's personalized normal range is determined, which serves as a benchmark for judging whether physiological characteristics are abnormal in subsequent steps.

[0168] Behavioral habit analysis: This involves analyzing drivers' behavioral habits during daily driving, including eye scanning patterns, steering wheel grip, and pedal operation habits. For example, some drivers habitually check their rearview mirrors frequently, while others habitually grip the steering wheel lightly with one hand. By establishing personalized behavioral habit models through long-term observation, we can avoid misjudging normal driving habits as signs of disability.

[0169] Response pattern recognition: Records drivers' response patterns to various prompts and interventions, including response time distribution and preferred response methods. For example, some drivers respond quickly to visual prompts, while others respond quickly to auditory prompts. By identifying drivers' response preferences, the fastest response method can be prioritized in subsequent interventions, improving intervention efficiency.

[0170] Anti-interference ability assessment: This assesses the driver's anti-interference ability and fatigue sensitivity under different environmental conditions. For example, some drivers are more prone to fatigue when driving at night, and some drivers experience significant changes in attention in rainy weather. By assessing the driver's anti-interference ability in different scenarios, a basis is provided for adjusting judgment thresholds and intervention strategies in different scenarios.

[0171] Personalized parameter optimization: Based on the driver profile described above, adjust the parameter configurations for each step. For example, adjust the normal range of the driver's physiological sub-nodes in the causal graph model according to the physiological baseline, adjust the abnormal judgment threshold of the driver's behavior sub-nodes according to behavioral habits, adjust the preset detection duration and preset reliability threshold according to the response mode, and adjust the confidence sensitivity coefficient and evidence quality coefficient under different environmental conditions according to the anti-interference capability.

[0172] One implementation involves an incremental update mechanism for the driver profile. After each driving session, the profile is updated based on the driving data. Discrepancies between anomaly detection results and actual conditions are also considered in the profile update. For example, if the system determines the driver is disabled during a driving session but the driver manually cancels intervention, the multi-source perception features from that session are used as negative samples to adjust the anomaly detection threshold. This continuous learning process improves the accuracy of personalized adaptation.

[0173] It should be understood that the driver profiling module enables the system to adaptively adjust the threshold for disability assessment and the parameters of intervention strategies based on the physiological characteristics, behavioral habits, and response preferences of different drivers. This represents a qualitative leap from general fixed parameters to personalized dynamic adaptation, further improving the accuracy of disability assessment and the effectiveness of intervention.

[0174] In summary, exemplarily, the graded intervention execution and closed-loop optimization process provided in this application embodiment is as follows: First, a graded intervention strategy is executed, triggering corresponding level intervention measures according to the target disability level. For the first disability level, cockpit interaction is implemented with voice prompts and instrument panel alerts; for the second disability level, chassis coordination is implemented with active speed limiting and seatbelt vibration; and for the third disability level, cloud-based linkage is implemented with emergency takeover and automatic parking. Subsequently, the intervention effect is evaluated in real time, and a driver recovery score is calculated. Vehicle condition improvement score Driver response quality score The weighted average of the intervention effects yields a comprehensive score. Next, online learning of the policy parameters is performed, using the reward function... As the optimization objective, among which The cost coefficient for incorrect intervention. This is the cost coefficient for missed detections. and The corresponding cost weight coefficients are used, and the policy gradient update formula is applied. The intervention control parameters are updated. Then, the driver profile is updated, including updating the physiological baseline based on benchmark values ​​such as heart rate and blink frequency, updating behavioral habits based on steering wheel operation frequency and head posture, and updating the response pattern based on response speed and preference to intervention methods. Simultaneously, a self-correction detection for erroneous intervention is performed. Detection conditions include recovery detection when the driver's state returns to normal after intervention, no response detection if no response is generated within a time limit, and result rollback detection if the result is judged as normal after a consecutive preset number of frames. Self-correction measures are triggered when the erroneous intervention conditions are met. Finally, the optimized strategy parameters are output, with the preset weights for each dimension updated to 0.72 for driver features, 0.18 for vehicle motion features, and 0.10 for environmental perception features. The disability detection duration is adjusted to 3.5 seconds, and personalized configurations are optimized based on the driver profile, achieving a complete closed-loop optimization from intervention execution, effect evaluation, parameter learning to personalized adaptation and erroneous intervention self-correction.

[0175] As one possible embodiment, before step S202, the method further includes a step of constructing a preset causal graph model.

[0176] One implementation method defines the node types and initial structure of the cause-effect graph based on prior knowledge in the field of assisted driving. Node types include environment nodes, vehicle nodes, driver nodes, and disability state nodes. Environment nodes include weather condition sub-nodes, road condition sub-nodes, etc.; vehicle nodes include steering state sub-nodes, speed state sub-nodes, etc.; and driver nodes include driver physiological sub-nodes and driver behavior sub-nodes. Driver physiological sub-nodes include heart rate and blood pressure, while driver behavior sub-nodes include gaze direction and head posture, etc. The initial connections between possible directed edges are determined based on prior knowledge.

[0177] One implementation method involves using structure learning methods such as conditional independence testing and PC algorithms to learn the conditional dependencies between nodes from historical data, thereby determining the structure of the causal graph. The historical data includes multiple sets of multi-source sensory feature samples labeled with driver disability status. Through statistical analysis of the sample data and causal discovery algorithms, the existence and direction of causal edges between nodes are automatically learned, thus correcting and supplementing the initial structure defined by prior knowledge.

[0178] One approach involves cross-modal feature alignment, mapping features of different dimensions from multi-source perception features to a unified feature space to establish cross-modal causal relationships. For example, a causal relationship can be established between rain / snow weather features from environmental perception features and blurred vision and slow reaction behavior from driver features; similarly, an abnormal steering feature from vehicle motion features can be established between emergency maneuvering behavior from driver features. Cross-modal feature alignment enables connections between similar features from different perception modules in the causal graph, forming cross-modal causal links.

[0179] The pre-defined causal graph model uses a directed acyclic graph (DAG), where nodes represent variables and directed edges represent causal relationships. The direction of the directed edges indicates the direction of causal influence transmission, and the causal probability value of the directed edges represents the strength of the causal relationship.

[0180] It should be understood that the construction steps of the pre-defined causal graph model can be completed offline. After completion, it can be deployed in the vehicle system for real-time inference, or the model structure can be updated periodically based on new data during vehicle operation.

[0181] As one possible implementation, before weighting the causal contribution of each dimension with confidence in step S203, an evidence update and confidence transfer step is also included to dynamically determine the preset weights of the features of each dimension.

[0182] One implementation method includes an evidence update step comprising: modeling the observation data of each dimension in multi-source sensing features using a probability distribution model; dynamically adjusting the parameters of the probability distribution based on the statistical characteristics of the observation data; calculating the likelihood of the observation data of each dimension under the hypothesis of the candidate disability level; and fusing evidence from different sources using an evidence aggregation algorithm on multiple frames of data to output the evidence strength of each dimension. The probability distribution model includes at least one of Gaussian, Bernoulli, or Poisson distributions. The evidence aggregation algorithm can employ DS evidence theory. Higher evidence strength indicates more reliable support of the observation data of that dimension for disability determination.

[0183] For example, for heart rate data in the driver feature dimension, Gaussian distribution modeling can be used to determine the distribution parameters of the normal state based on the historical mean and variance of heart rate, and the likelihood of the current heart rate observation under this distribution can be calculated as the strength of evidence; for the binary feature of whether the gaze direction deviates from the front, Bernoulli distribution modeling can be used; for count features such as the number of gaze deviations per unit time, Poisson distribution modeling can be used.

[0184] One implementation method includes the following confidence propagation steps: converting a predefined causal graph model into a factor graph representation, where factor nodes represent causal relationships between nodes of each dimension; employing a confidence propagation algorithm to perform message passing on the factor graph, the message passing including two directions: likelihood messages from observation nodes to variable nodes and prior messages from variable nodes to observation nodes; iterating message passing until the confidence scores of each node converge, obtaining the posterior confidence scores corresponding to each dimension feature. The posterior confidence scores of each dimension feature are used as predefined weights corresponding to that dimension for weighted summation in step S203.

[0185] For example, factor nodes This represents the probability distribution of causal relationships between nodes in each dimension and nodes in a disabled state. During message passing, the message from the observation node to the factor node is first calculated. That is, the likelihood of the observed data in each dimension under the assumption of disability; then based on the factor function And input messages, compute messages from factor nodes to variable nodes:

[0186] in, Indicates from factor node To variable node The message represents the variable under factor constraints. Values Confidence level; For factor functions, representing variables and The probability distribution of causal relationships between them, that is, the causal probability values ​​of the corresponding directed edges in the pre-defined causal graph model; For evidence variables, representing characteristic variables in each dimension; For the disability status variable, the value takes the level of each candidate disability status; To start from the evidence nodes To factor node The message indicates the degree of support of the observational evidence for the factor. Message passing iterates in both directions until the confidence scores of each node converge, ultimately yielding the disabled state node. The approximate estimates of the posterior probability values ​​for each candidate disability level are obtained, and the posterior confidence of each evidence node is also converged.

[0187] It should be understood that the evidence update step transforms the original observation data from each dimension into likelihood evidence with a unified dimension, resolving the problem of difficulty in direct fusion caused by the heterogeneity of data from different modalities. The confidence transfer step utilizes the causal relationship structure in the causal graph model to achieve dynamic fusion of cross-modal information through message passing on the factor graph, enabling the preset weights of each dimension to adaptively adjust according to the current observation evidence, rather than remaining fixed. Together, these steps provide a dynamic and reliable weighting basis for calculating the overall confidence score.

[0188] In summary, for example, please refer to Figure 4This is a schematic diagram of a factor graph structure provided in an embodiment of this application. The factor graph is obtained by converting a preset causal graph model and includes factor nodes, variable nodes, and a disability state node D. The factor nodes include a first factor node f_driver, a second factor node f_vehicle, and a third factor node f_env, each representing the joint probability distribution function of the corresponding dimension's features. The first factor node f_driver connects to the variable nodes corresponding to the driver's features, including visual feature variable node E_v, physiological feature variable node E_p, and behavioral feature variable node E_b. The second factor node f_vehicle connects to the variable nodes corresponding to the vehicle's motion features, including chassis feature variable node E_c, electrical feature variable node E_s, and steering feature variable node E_t. The third factor node f_env connects to the variable nodes corresponding to the environmental perception features, including weather feature variable node E_w, road condition feature variable node E_r, and illumination feature variable node E_f.

[0189] All the variable nodes E_v, E_p, E_b, E_c, E_s, E_t, E_w, E_r, and E_f are connected to the disability status node D. The values ​​of the disability status node D include normal, first disability level, second disability level, and third disability level. Specifically, E_v represents visual characteristics, E_p represents physiological characteristics, E_b represents behavioral characteristics, E_c represents chassis characteristics, E_s represents electrical characteristics, E_t represents steering characteristics, E_w represents weather characteristics, E_r represents road condition characteristics, and E_f represents lighting characteristics.

[0190] It should be understood that message passing on the factor graph includes likelihood messages from each variable node to the corresponding factor node, and messages from the factor node to the disabled state node D. By iterating message passing until the confidence of each node converges, the posterior probability value of the disabled state node D in each candidate disabled state level is obtained.

[0191] For example, an embodiment of this application provides a dynamic confidence propagation and weight adaptive adjustment process as follows: First, initialize the confidence of each node, and convert the observation data of each dimension into initial likelihood messages to satisfy... The initial confidence scores of each sub-node in the driver feature dimension are respectively the visual features. Physiological characteristics The initial confidence levels of each sub-node in the vehicle motion feature dimension are as follows: chassis features Electrical characteristics The initial confidence levels of each sub-node in the environmental perception feature dimension are as follows: weather features Road condition characteristics .

[0192] Then, confidence propagation from factor nodes to variable nodes is performed, according to the formula. Calculate the first factor node respectively Second factor node and the third factor node To the disabled state node The message. Then, confidence propagation from variable nodes to factor nodes is performed, satisfying... ,in This is the normalization constant. Iterative confidence propagation is then performed until convergence, with the convergence condition being... If the conditions are not met, the confidence levels of each node are iteratively updated. Upon convergence, a cross-modal consensus is reached, according to the formula... The preset weights for each dimension are dynamically adjusted, resulting in a preset weight of 0.75 for the driver feature dimension, 0.15 for the vehicle motion feature dimension, and 0.10 for the environmental perception feature dimension. The final output is the fusion result, with a comprehensive confidence level. The preset weights for each dimension are 0.75 for driver features, 0.15 for vehicle motion features, and 0.10 for environmental perception features. This enables dynamic fusion of cross-modal information and adaptive adjustment of weights based on belief propagation, providing a reliable basis for subsequent disability level determination.

[0193] As one possible implementation, after determining the causal contribution of each dimension feature in step S203, the method may further include performing causal effect quantification analysis on the causal probability value to generate a causal effect report with complete causal explanation basis, providing support for the optimization of subsequent intervention strategies.

[0194] As one possible implementation, the causal effect report includes at least one of the following: a ranking of the causal contribution of each dimension of features, direct causal effect values, indirect causal effect values, total causal effect values, and a list of key causal paths.

[0195] One implementation method, causal contribution ranking, refers to arranging the causal contribution of each dimension feature in the multi-source perception features from largest to smallest according to the absolute value, so as to clarify the ranking of the degree of influence of each dimension feature on the current disability determination.

[0196] The direct causal effect value refers to the causal effect value generated when the node corresponding to this dimension directly acts on the disabled node through a directed edge. One implementation method is to set the node corresponding to this dimension to the current observation value and the current observation value plus a preset increment in a preset causal graph model, respectively, while fixing the values ​​of all intermediate nodes to block indirect paths. The posterior probability values ​​of the disabled node under the two settings are determined, and the difference between the two is the direct causal effect value.

[0197] The indirect causal effect value refers to the causal effect value generated by the node corresponding to this dimension indirectly acting on the disabled node through at least one intermediate node. One implementation method is to subtract the direct causal effect value from the causal contribution of this dimension.

[0198] The total causal effect value refers to all causal effects that the node corresponding to this dimension has on the disabled node. The total causal effect value is the sum of the direct causal effect value and the indirect causal effect value, and is numerically equal to the causal contribution of this dimension.

[0199] The critical causal path list refers to a list consisting of at least one link with the highest causal contribution selected from the link set determined in step S202. One implementation method is to determine the causal contribution of a single link based on the product of the causal probability values ​​of each directed edge in that link; a larger product indicates a stronger causal influence of the link on the disabled node. All links in the link set are sorted from highest to lowest causal contribution, and a predetermined number of links with the highest ranking are selected as critical causal paths and included in the critical causal path list.

[0200] As one possible implementation, the method further includes: generating a causal effect probability distribution, describing the posterior probability distribution under different causal hypotheses. Here, different causal hypotheses refer to hypothetical scenarios formed by setting different values ​​for one or more nodes in a pre-defined causal graph model. One implementation involves, given multi-source sensory features as observational evidence, determining the posterior probability value of a disabled state node at each candidate disabled state level when a node corresponding to a certain dimension is set to different values. The posterior probability values ​​of each candidate disabled state level collectively constitute the posterior probability distribution under that causal hypothesis. By generating multiple sets of posterior probability distributions under causal hypotheses, the dependence of the judgment result on key dimensional features can be intuitively displayed, achieving a quantitative expression of uncertainty.

[0201] This causal effect report can serve as input for an adaptive confirmation window mechanism and a tiered intervention strategy, enabling the entire disability detection and intervention process to have complete causal interpretability and traceability.

[0202] Based on S201 to S205, this application embodiment provides a comprehensive information foundation for cross-modal causal analysis by acquiring multi-source perception features covering driver, vehicle motion, and environmental perception. Based on a preset causal graph model, inference is performed using the causal probability values ​​represented by directed edges to determine the posterior probability value of each candidate disability state level, elevating data association from the statistical correlation level to the causal inference level, thus enabling the judgment results to have causal interpretability. By calculating the causal contribution of each dimension's features and performing confidence weighting to determine the comprehensive confidence level, a quantitative expression of the reliability of the posterior probability value is achieved. Reliable candidate levels are screened based on a comprehensive confidence level greater than a preset confidence threshold, and the level corresponding to the highest posterior probability value is selected as the target disability state level, ensuring that a decision is made only when the judgment result is sufficiently reliable. Ultimately, the duration of disability detection was dynamically adjusted based on comprehensive confidence levels, and a tiered intervention strategy was implemented. This achieved a precise match between intervention measures and the severity of disability and the reliability of the determination. Furthermore, through closed-loop feedback and reinforcement learning optimization of the intervention effect, the adaptability and effectiveness of the intervention strategy were continuously improved, thereby enhancing the overall accuracy, interpretability, and safety of driver disability detection and intervention.

[0203] This application constructs a multimodal causal reasoning engine, establishing cross-modal causal relationships between the environment, vehicle, and driver using a directed acyclic graph. This achieves deep feature fusion and synergistic enhancement, overcoming the limitations of traditional methods that rely solely on shallow fusion based on statistical correlation. Furthermore, a dynamic confidence propagation network quantifies the reliability of the judgment results, making the confidence level of disability determination quantifiable and traceable, providing a clear and credible basis for intervention decisions. An adaptive confirmation window mechanism dynamically adjusts the confirmation time based on the overall confidence level, rapidly triggering intervention at high confidence levels and extending observation at low confidence levels, improving interaction efficiency by approximately 40%. The synergistic effect of multimodal causal reasoning and a self-correction mechanism for misinterpretation effectively reduces unnecessary interventions caused by misjudgments, lowering the misinterpretation rate by approximately 60%. A closed-loop feedback mechanism for intervention effects uses the improvement in driver and vehicle status after intervention as a feedback signal to drive continuous optimization of strategy parameters, achieving a qualitative leap from open-loop control to closed-loop self-learning. The driver profile module establishes personalized profiles through long-term learning, evolving intervention strategies from general fixed parameters to personalized customization, achieving a qualitative leap from general to personalized. The aforementioned technical effects are not simply the sum of the modules, but rather the result of the organic synergy of causal reasoning, confidence calibration, closed-loop optimization, and adaptive profiling, which produces system property changes that go beyond simple summation.

[0204] The foregoing primarily describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the aforementioned functions, the driver disability intervention device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0205] This application embodiment can, based on the above method, exemplarily divide a driver disability intervention device or electronic device into functional modules. For example, the driver disability intervention device or electronic device may include functional modules corresponding to each functional division, or two or more functions may be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0206] Please see Figure 5 This application provides a driver disability intervention device, which includes: a feature acquisition module 501, a probability determination module 502, a confidence level determination module 503, a level determination module 504, and an execution strategy module 505.

[0207] The feature acquisition module 501 is used to acquire multi-source perception features. These multi-source perception features include features in the following dimensions: driver features, vehicle motion features, and environmental perception features. The probability determination module 502 is used to determine the posterior probability value corresponding to any candidate disability state level from a preset causal graph model based on the multi-source perception features. The candidate disability state level is used to characterize the severity of driver disability. The preset causal graph model is used to characterize a directed acyclic graph (DAG) connecting environmental nodes, vehicle nodes, driver nodes, and disability state nodes with directed edges. Each directed edge characterizes the causal probability value between two nodes connected in the DAG. The posterior probability value corresponding to any candidate disability state level is determined based on the causal probability value corresponding to at least one directed edge in the link pointing to the disability state node. The confidence level is determined. Block 503 is used for any candidate disability status level among multiple candidate disability status levels: determining the causal contribution of each dimension in the multi-source perception features based on the posterior probability value corresponding to the candidate disability status level, and weighting the causal contribution of each dimension with confidence to determine the comprehensive confidence; the causal contribution is used to characterize the degree of influence of each dimension on the posterior probability value corresponding to the candidate disability status level; the comprehensive confidence is used to characterize the reliability of the posterior probability value corresponding to the candidate disability status level; the level determination module 504 is used to select the candidate disability status level corresponding to the largest posterior probability value among the candidate disability status levels with a comprehensive confidence greater than the corresponding preset confidence threshold, and determine it as the target disability status level; the execution strategy module 505 is used to control the vehicle to execute the intervention strategy corresponding to the target disability status level.

[0208] As one possible embodiment, the probability determination module 502 is specifically used to: determine the posterior probability value corresponding to any candidate disability state level among multiple candidate disability state levels from a preset causal graph model based on multi-source perception features, including: determining an initial node set from the preset causal graph model based on multi-source perception features; the initial node set includes at least one initial node, each initial node being any one of an environment node, a vehicle node, or a driver node; determining, based on the initial node set, any node in the initial node set as a link set pointing from the starting point to the disability state node from the preset causal graph model; the link set includes at least one link; and performing forward propagation reasoning based on the causal probability values ​​corresponding to all directed edges in the link set to determine the posterior probability value of the disability state node being in any candidate disability state level.

[0209] As one possible embodiment, the confidence determination module 503 is specifically used to: determine the causal contribution of each dimension in the multi-source sensing features based on the posterior probability value corresponding to the candidate disability status level, including: applying a preset increment to the first dimension in the multi-source sensing features, and determining the difference between the posterior probability values ​​before and after applying the preset increment; the first dimension is any dimension in the multi-source sensing features; and determining the ratio of the difference to the preset increment as the causal contribution of the first dimension.

[0210] As one possible implementation, the confidence determination module 503 is specifically used to: perform confidence weighting on the causal contribution corresponding to each dimension to determine the comprehensive confidence, including: obtaining the preset weights corresponding to each dimension in the multi-source perception features; and weighting and summing the causal contribution corresponding to each dimension with the preset weights corresponding to each dimension to obtain the comprehensive confidence corresponding to the candidate disability status level.

[0211] As one possible embodiment, the execution strategy module 505 is specifically used to control the vehicle to execute the intervention strategy corresponding to the target disability status level, including: determining the disability detection duration based on the comprehensive confidence level corresponding to the target disability status level; the comprehensive confidence level is negatively correlated with the disability detection duration; and executing the intervention strategy corresponding to the target disability status level when the duration of monitoring the driver's operation response status exceeds the disability detection duration.

[0212] As one possible implementation, the disability detection duration satisfies the following relationship:

[0213] in, This refers to the duration of disability detection. The preset detection duration is related to the target disability level. For overall confidence level; To pre-set the reliability threshold; This is the confidence sensitivity coefficient, used to control the degree of influence of the overall confidence level on the window duration; The evidence quality coefficient is determined based on the pre-set reliability of each dimension in the multi-source perception features.

[0214] As one possible embodiment, the driver disability intervention device further includes a self-learning module, used to: acquire current multi-source perception features; the current multi-source perception features are the multi-source perception features after implementing the intervention strategy corresponding to the target disability level; determine a comprehensive score for the intervention effect based on the deviation between the multi-source perception features and the current multi-source perception features; update the intervention control parameters through a preset reinforcement learning algorithm with the goal of maximizing the comprehensive score for the intervention effect; the intervention control parameters include at least one of the following: preset detection duration, preset confidence threshold, confidence sensitivity coefficient, and evidence quality coefficient.

[0215] As one possible embodiment, the execution strategy module 505 is further configured to: recalculate the disability detection duration when the change in the overall confidence level exceeds a preset change threshold; and control the vehicle to end the disability detection duration and cancel the execution of the intervention strategy corresponding to the target disability level when the duration corresponding to the monitoring of the driver's operation response status is less than or equal to the disability detection duration.

[0216] As one possible embodiment, the multiple candidate disability status levels include a first disability level, a second disability level, and a third disability level; wherein the severity of disability corresponding to the first disability level, the second disability level, and the third disability level increases sequentially; the intervention strategy corresponding to the first disability level includes: sending an attention alert to the driver through at least one of the following cabin warning operations: visual alert, auditory alert, tactile alert, and olfactory alert; the intervention strategy corresponding to the second disability level includes: adjusting the vehicle dynamics characteristics through at least one of the following vehicle chassis operations: speed limit control, lane keeping enhancement, and active braking preparation, to assist in stable vehicle driving; the intervention strategy corresponding to the third disability level includes: performing at least one of the following remote support operations: fault data upload, emergency contact notification, and remote driving takeover preparation.

[0217] like Figure 6 As shown, the electronic device 600 provided in this application embodiment includes, but is not limited to, a processor 601 and a memory 602.

[0218] The memory 602 described above is used to store the executable instructions of the processor 601. It is understood that the processor 601 is configured to execute instructions to implement the methods in the above embodiments.

[0219] It should be noted that those skilled in the art will understand that Figure 6 The electronic device structure shown does not constitute a limitation on electronic device 600; electronic device may include, but is not limited to, other electronic devices. Figure 6 This may indicate more or fewer components, or combinations of certain components, or different component arrangements.

[0220] Processor 601 is the control center of electronic device 600. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in memory 602, and by calling data stored in memory 602, it performs various functions and processes data of electronic device 600, thereby providing overall monitoring of electronic device 600. Processor 601 may include one or more processing units. Optionally, processor 601 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into processor 601.

[0221] The memory 602 can be used to store software programs and various data. The memory 602 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required by at least one functional module (such as a determination unit, processing unit, etc.), etc. Furthermore, the memory 602 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0222] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 602 including instructions, which can be executed by a processor 601 of an electronic device 600 to implement the methods in the above embodiments.

[0223] Optionally, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0224] In an exemplary embodiment, this application also provides a computer program product including one or more instructions, which can be executed by the processor 601 of the electronic device 600 to perform the methods described above.

[0225] It should be noted that when one or more instructions in the computer-readable storage medium or computer program product are executed by the processor of an electronic device, they implement the various processes of the above method embodiments and achieve the same technical effect as the above method. To avoid repetition, they will not be described again here.

[0226] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0227] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0228] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0229] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0230] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, essentially, or the parts that contribute to related technologies, or all or part of the technical solutions, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0231] The above embodiments are merely preferred embodiments provided to fully illustrate this application, and the scope of protection of this application is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on this application are all within the scope of protection of this application.

Claims

1. A driver incapacitation intervention method characterized by, Executed on the vehicle controller, the driver disability intervention method includes: Acquire multi-source perception features; the multi-source perception features include features in the following dimensions: driver features, vehicle motion features, and environmental perception features; Based on the multi-source sensing features, a posterior probability value corresponding to any candidate disability state level is determined from a preset causal graph model. The candidate disability state level is used to characterize the severity of driver disability. The preset causal graph model is used to characterize a directed acyclic graph of environment nodes, vehicle nodes, driver nodes, and disability state nodes connected by directed edges. Each directed edge represents the causal probability value between two nodes connected in the directed acyclic graph. The posterior probability value corresponding to any candidate disability state level is determined based on the causal probability value corresponding to at least one directed edge in the link pointing to the disability state node. For any candidate disability status level among the multiple candidate disability status levels: Based on the posterior probability value corresponding to the candidate disability status level, determine the causal contribution of each dimension in the multi-source perception features, and weight the causal contribution of each dimension with confidence to determine the comprehensive confidence level; the causal contribution level is used to characterize the degree of influence of each dimension on the posterior probability value corresponding to the candidate disability status level; the comprehensive confidence level is used to characterize the reliability of the posterior probability value corresponding to the candidate disability status level. Among the candidate disability status levels whose overall confidence level is greater than the corresponding preset confidence threshold, the candidate disability status level corresponding to the maximum posterior probability value is selected and determined as the target disability status level. Control the vehicle to execute the intervention strategy corresponding to the target disability level.

2. The method of claim 1, wherein, The step of determining the posterior probability value corresponding to any candidate disability state level from a preset causal graph model based on the multi-source sensing features includes: Based on the multi-source perception features, an initial node set is determined from the preset causal graph model; the initial node set includes at least one initial node, and each initial node is any one of the environment node, the vehicle node, and the driver node; Based on the initial node set, a set of links is determined from the preset causal graph model, with any node in the initial node set as the starting point, pointing to the disabled node; the set of links includes at least one link. Based on the causal probability values ​​corresponding to all directed edges in the link set, forward propagation reasoning is performed to determine the posterior probability value of the disabled state node being in any candidate disabled state level.

3. The method of claim 1, wherein, The step of determining the causal contribution of each dimension in the multi-source sensing features based on the posterior probability value corresponding to the candidate disability state level includes: A preset increment is applied to the first dimension of the multi-source sensing features, and the difference between the posterior probability values ​​before and after applying the preset increment is determined; the first dimension is any dimension of the multi-source sensing features. The ratio of the difference to the preset increment is determined as the causal contribution of the first dimension.

4. The method of claim 1, wherein, The step of weighting the causal contribution of each dimension with confidence to determine the overall confidence level includes: Obtain the preset weights corresponding to each dimension in the multi-source sensing features; The causal contribution of each dimension is weighted and summed with the preset weights of each dimension to obtain the comprehensive confidence level corresponding to the candidate disability status level.

5. The method of claim 1, wherein, The control vehicle executes the intervention strategy corresponding to the target disability level, including: The disability detection duration is determined based on the comprehensive confidence level corresponding to the target disability status level; the comprehensive confidence level is negatively correlated with the disability detection duration. If the duration of monitoring the driver's operational response exceeds the disability detection duration, the intervention strategy corresponding to the target disability level is executed.

6. The method of claim 5, wherein, The disability detection duration satisfies the following relationship: wherein, is the disabled detection duration; is a preset detection duration, the preset detection duration being related to the target disabled state level; is the comprehensive confidence; is a preset confidence threshold; is a confidence sensitivity coefficient, used to control the influence degree of the comprehensive confidence on the window duration; is an evidence quality coefficient, determined based on a preset confidence of each dimension in the multi-source perception features.

7. The method of claim 1, wherein, The method further includes: Acquire the current multi-source sensing features; the current multi-source sensing features are the multi-source sensing features after executing the intervention strategy corresponding to the target disability state level; Based on the deviation between the multi-source sensing features and the current multi-source sensing features, a comprehensive score for the intervention effect is determined; With the goal of maximizing the comprehensive score of the intervention effect, the intervention control parameters are updated through a preset reinforcement learning algorithm; the intervention control parameters include at least one of the following: preset detection duration, preset reliability threshold, confidence sensitivity coefficient, and evidence quality coefficient.

8. The method of claim 5, wherein, The method further includes: When the change in the overall confidence level exceeds a preset change threshold, the disability detection duration is recalculated. If the duration of monitoring the driver's operational response is less than or equal to the disability detection duration, the vehicle is controlled to end the disability detection duration and the intervention strategy corresponding to the target disability level is canceled.

9. The method of claim 1, wherein, The multiple candidate disability status levels include a first disability level, a second disability level, and a third disability level; wherein, the severity of disability corresponding to the first disability level, the second disability level, and the third disability level increases sequentially. Intervention strategies corresponding to the first level of disability include: sending attention alerts to the pilot through at least one of the following cockpit alert operations: visual alerts, auditory alerts, tactile alerts, and olfactory alerts; The intervention strategies corresponding to the second level of disability include: adjusting vehicle dynamics through at least one of the vehicle chassis operations, namely speed limit control, lane keeping enhancement, and active braking preparation, to assist in stable vehicle driving; The intervention strategies corresponding to the third level of disability include: performing at least one of the following remote support operations: uploading fault data, notifying emergency contacts, and preparing for remote driving takeover.

10. A driver incapacitation intervention device characterized by, The driver disability intervention device includes: a feature acquisition module, a probability determination module, a confidence level determination module, a level determination module, and an execution strategy module; The feature acquisition module is used to acquire multi-source perception features; the multi-source perception features include features in the following dimensions: driver features, vehicle motion features, and environmental perception features; The probability determination module is used to determine the posterior probability value corresponding to any candidate disability state level from a preset causal graph model based on the multi-source perception features. The candidate disability state level is used to characterize the severity of driver disability. The preset causal graph model is used to characterize a directed acyclic graph of environment nodes, vehicle nodes, driver nodes, and disability state nodes connected by directed edges. Each directed edge represents the causal probability value between two nodes connected in the directed acyclic graph. The posterior probability value corresponding to any candidate disability state level is determined based on the causal probability value corresponding to at least one directed edge in the link pointing to the disability state node. The confidence determination module is used to, for any candidate disability status level among the plurality of candidate disability status levels: determine the causal contribution of each dimension in the multi-source perception features based on the posterior probability value corresponding to the candidate disability status level, and perform confidence weighting on the causal contribution of each dimension to determine the comprehensive confidence level; the causal contribution is used to characterize the degree of influence of each dimension on the posterior probability value corresponding to the candidate disability status level; the comprehensive confidence level is used to characterize the reliability of the posterior probability value corresponding to the candidate disability status level. The level determination module is used to select the candidate disability status level corresponding to the maximum posterior probability value from among the candidate disability status levels whose comprehensive confidence is greater than the corresponding preset confidence threshold, and determine it as the target disability status level. The execution strategy module is used to control the vehicle to execute the intervention strategy corresponding to the target disability level.

11. A vehicle, characterized in that, The vehicle includes: the driver disability intervention device as described in claim 10.

12. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the vehicle's processor, the vehicle is able to perform the driver disability intervention method as described in any one of claims 1 to 9.