Driving safety behavior intelligent management method based on multi-modal fusion

Abstract

The invention discloses a driving safety behavior intelligent management method based on multi-modal fusion, and the method comprises the steps: collecting multi-source and multi-modal data of a driving scene, and carrying out the preprocessing of the multi-source and multi-modal data, and generating standardized multi-source and multi-modal data; constructing an improved Wiener process model, and forming a stochastic differential equation; a Milstein-Euler pre-estimation correction algorithm is executed, and a driving behavior state value of the next time step is obtained; performing continuous time sequence processing to obtain a processed driving behavior state sequence; constructing a behavior risk cumulant and performing index mapping to generate a stability index; and executing the driving safety behavior management strategy. According to the invention, multi-source and multi-modal data fusion is introduced, and a Wiener process model and a Milstein-Euler pre-estimation correction algorithm are improved, so that high-precision identification and risk trend intelligent management of driving safety behaviors in a complex traffic environment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation and vehicle safety, and particularly relates to a driving safety behavior intelligent management method based on multi-modal fusion. BACKGROUND

[0002] With the development of intelligent transportation systems, driving safety has become one of the focuses of research and technological development. Traditional driving behavior monitoring techniques mostly rely on vehicle sensor data such as speed, acceleration, throttle pedal pressure, and other physical data, which can provide basic driving behavior information of the driver. However, these traditional methods usually only use single sensor input, ignoring the driver's psychological state, fatigue level, and external environment factors. The application of a single data source leads to an incomplete assessment and risk prediction of driving behavior, especially in complex and unexpected driving situations, it is difficult to provide accurate risk warnings and timely safety interventions, and traditional risk assessment models often ignore the dynamic changes of behavior, leading to a delayed response to abnormal driving behavior of the driver, making it difficult to effectively prevent potential driving safety risks.

[0003] In recent years, multi-modal fusion technology has been gradually applied in the field of intelligent transportation, especially in driving behavior monitoring and risk assessment. By fusing data from multiple sensors such as vehicle dynamic information, driver physiological data, and environmental monitoring information, the behavior and state of the driver can be more comprehensively reflected. Traditional single-modal technology often struggles to cope with complex driving scenarios, while multi-modal fusion can provide more accurate real-time monitoring for driving safety and improve the system's response capability to changes in driving behavior. The driving behavior modeling and risk assessment method based on multi-modal data integrates different types of input information, not only improving the prediction accuracy of the model, but also adapting to more complex and dynamic driving environments. However, existing multi-modal fusion technology still has deficiencies in the dynamic modeling of driving behavior, lacking efficient response techniques for the continuity and suddenness of behavior changes.

[0004] Therefore, how to provide a driving safety behavior intelligent management method based on multi-modal fusion is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One purpose of the present application is to provide a multi-modal fusion-based intelligent management method for driving safety behavior, which fully utilizes multi-source multi-modal data fusion technology, improved Wiener process model and Milstein-Euler prediction correction algorithm, and describes in detail how to dynamically model driving behavior and real-time assess driving risk by fusing vehicle operating state, driver physiological data and external environment information. By utilizing the rich information provided by multi-modal data, combining behavior state deviation and risk assessment, and describing in detail how to generate stability index by optimizing risk prediction through numerical iteration and state deviation driving mechanism according to the continuous random evolution characteristics of driving behavior, the present application realizes accurate prediction and classification of driving risk. The method of the present application not only enhances the real-time monitoring and response capability for complex driving behavior, but also effectively avoids risk omission and delayed response caused by single data source, and has the advantages of high accuracy, high dynamic response and strong robustness, and can provide timely and effective safety intervention in complex and sudden driving environment.

[0006] According to the multi-modal fusion-based intelligent management method for driving safety behavior of the present application, the method comprises: Collecting multi-source multi-modal data of a driving scene, and pre-processing the multi-source multi-modal data to generate standardized multi-source multi-modal data; Constructing an improved Wiener process model to process the standardized multi-source multi-modal data, dynamically adjusting the drift term and diffusion term by setting behavior state threshold, behavior degradation adjustment factor and state deviation sensitivity coefficient, and constructing behavior state deviation according to the difference between behavior state variable and behavior state threshold to form a random differential equation; Discrete-time numerical solution of the random differential equation is performed by using Milstein-Euler prediction correction algorithm, and the driving behavior state value at the next time step is obtained by performing Euler prediction, double prediction of drift term and diffusion term and diffusion modulation containing state-related damping factor; Continuous-time sequence processing is performed on the driving behavior state value, including analyzing the state change rate of adjacent time steps, suppressing abnormal mutation points and smoothing the behavior state sequence to obtain the processed driving behavior state sequence; Based on the processed driving behavior state sequence, the behavior risk accumulation is constructed by using time accumulation method, and the risk accumulation is exponentially mapped to generate a stability index reflecting the safety of driving behavior; Based on the stability index and the behavior state deviation, the driving safety behavior management strategy including risk behavior identification, early warning prompt, driving intervention control and risk event recording is performed.

[0007] Optionally, the multi-source multi-modal data specifically includes vehicle operating state data, driver state data and external environment state data.

[0008] Optionally, the preprocessing of the multi-source multi-modal data specifically includes time synchronization, data cleaning, data missing completion, feature standardization, and data alignment.

[0009] Optionally, the random differential equation is formed specifically as follows: An improved Wiener process model is constructed, which is composed of an input mapping layer, a state-driven diffusion adjustment layer, a drift adjustment layer, and an equation construction layer; The input mapping layer constructs behavior state variables based on standardized multi-source multi-modal data, selects key features from vehicle operating state data, driver state data, and external environment state data, aligns, normalizes, and dimensionally unifies the key features according to a unified time reference, and fuses the processed multiple key features by weighted combination to obtain behavior state variables representing the change of driving behavior over continuous time; The state-driven diffusion adjustment layer calculates a basic diffusion amount based on the fluctuation range and intensity of the behavior state variables, introduces a behavior degradation adjustment factor and a state deviation sensitivity coefficient, jointly adjusts the reference amplitude of the basic diffusion amount and the response sensitivity to the behavior state deviation, differentiates the behavior state variables from the behavior state threshold to obtain the behavior state deviation, and introduces an exponential adjustment mechanism driven by the behavior state deviation diffusion term to combine the behavior state deviation, the behavior degradation adjustment factor, and the state deviation sensitivity coefficient, to nonlinearly adjust the basic diffusion amount in an exponential form, and obtain the dynamically adjusted diffusion term; The drift adjustment layer calculates a basic drift amount based on the change direction, amplitude, and trend of the behavior state variables at adjacent time instants, introduces a time-varying driving mechanism to dynamically process the basic drift amount, obtains the drift term, and inputs the behavior state deviation into the drift term adjustment process to dynamically adjust the drift term based on the behavior state deviation, and obtains the drift term adjusted based on the behavior state deviation; The equation construction layer combines the drift term adjusted based on the behavior state deviation with the dynamically adjusted diffusion term, calibrates their consistency in the time domain, introduces a dynamic weight adjustment based on the state change rate to adjust the proportion of the drift term adjusted based on the behavior state deviation and the dynamically adjusted diffusion term in the random differential equation, constrains and reconstructs the combined term-level structure to meet the dynamic response characteristics of continuous random evolution of driving behavior, and constructs a random differential equation describing the continuous random evolution of driving behavior.

[0010] Optionally, the obtaining of the driving behavior state value at the next time step includes: The discrete time step of the random differential equation is determined, the discrete time sequence is established, the driving behavior state at the initial time corresponding to the random differential equation is taken as the initial value, the number, time position, and corresponding driving behavior state storage structure of each calculation time step are specified; performing the Milstein-Euler prediction-correction algorithm, generating a random disturbance increment corresponding to the current time step based on the discrete time step, pairing the random disturbance increment with the current driving behavior state at the same time mark, and combining to form a random driving input data of the current time step; based on the driving behavior state of the current time step, the drift term and the diffusion term in the stochastic differential equation, the discrete time step and the random driving input data, performing the Euler prediction process, multiplying the drift term by the discrete time step to obtain a drift term time increment, multiplying the diffusion term by the random driving input data to obtain a diffusion term random increment, and performing an addition operation on the drift term time increment, the diffusion term random increment and the driving behavior state of the current time step to obtain a predicted driving behavior state of the next time step; introducing a double predictor structure, reading the drift term corresponding to the current time step as a first drift prediction, reading the behavior state variable corresponding to the predicted driving behavior state of the next time step, inputting the behavior state variable into the drift adjustment structure, re-calling the drift term adjustment parameter, generating a second drift prediction corresponding to the next time step by processing and operating on the behavior state change direction, change amplitude and change trend, and pairing and combining the first drift prediction and the second drift prediction with the drift term corresponding to the current time step to form a predicted correction of the drift term; reading the diffusion term corresponding to the current time step as a first diffusion prediction, reading the behavior state variable corresponding to the predicted driving behavior state of the next time step, inputting the behavior state variable into the diffusion term adjustment structure, calling the diffusion term adjustment parameter, generating a second diffusion prediction corresponding to the next time step by processing and operating on the behavior state fluctuation amplitude and fluctuation frequency, introducing a noise modulation factor, constructing a state-related damping structure in an exponential form according to the noise modulation factor, and generating a damping adjustment, weighting and combining the damping adjustment, the first diffusion prediction, the second diffusion prediction and the diffusion term corresponding to the current time step to form a predicted correction of the diffusion term; combining the predicted correction of the drift term and the predicted correction of the diffusion term with the driving behavior state of the current time step, the discrete time step and the random driving input data for numerical updating, and calculating to obtain the driving behavior state value of the next time step.

[0011] Optionally, the obtained processed driving behavior state sequence comprises: arranging the driving behavior state values in time sequence, and pairing the driving behavior state values of each time step with the corresponding time marks to form a behavior state time sequence to be processed; Differential processing is performed on the driving behavior state values of adjacent time steps in the behavior state time sequence, a state change rate of each time step is calculated, the state change rate is taken as basic data for identifying mutation points and fluctuation abnormalities, and a state change rate sequence is constructed; Based on the state change rate sequence, abnormal mutation point suppression processing is performed on the behavior state time sequence, time positions with a change rate exceeding a change threshold are marked, the amplitude of the driving behavior state value of the corresponding time step is limited, and a behavior state time sequence after abnormal suppression is generated; Smooth processing is performed on the behavior state time sequence after abnormal suppression, the driving behavior state values of consecutive time steps are weighted and averaged in a window, the short-term fluctuation amplitude of the driving behavior state sequence is reduced, and a processed driving behavior state sequence is generated.

[0012] Optionally, the generated stability index reflecting the safety of the driving behavior includes: Based on the processed driving behavior state sequence, the driving behavior state values of each time step are read in time sequence, each driving behavior state value is paired with the corresponding behavior state deviation to form a time sequence input; Cumulative processing is performed on the time sequence input in time steps, the behavior state deviations of each time step are gradually accumulated according to a time accumulation rule to obtain a risk accumulation amount reflecting the risk accumulation degree of the driving behavior in the entire time period; Exponential mapping processing is performed based on the risk accumulation amount, the risk accumulation amount is exponentially attenuated and converted according to an exponential mapping parameter to generate a stability mapping amount reflecting the stability degree of the driving behavior; The stability mapping amount is processed in time steps, the stability mapping amount of each time step is normalized and mapped in a numerical interval to generate a stability index value in a target numerical interval.

[0013] Optionally, the driving safety behavior management strategy including risk behavior identification, early warning prompt, driving intervention control and risk event recording includes: Based on the stability index and the corresponding behavior state deviation, the stability index of each time step is read in time sequence, the stability index is compared with a stability threshold, and the driving safety risk level of each time step is determined in combination with the behavior state deviation to form a driving safety risk level sequence; Based on the driving safety risk level sequence, time steps with a stability index lower than the stability threshold and a behavior state deviation exceeding a deviation threshold are screened, the driving behavior state of the corresponding time step is marked as a risk behavior time period, and the risk behavior time period is classified according to the behavior state deviation to generate a risk behavior identification result; According to the risk behavior identification result and the corresponding driving safety risk level, early warning information is generated and sent to a man-machine interaction interface; When the driving safety risk level reaches the intervention threshold, a driving intervention control command is generated and sent to the vehicle control execution module to adjust the vehicle's operating parameters. When generating driving intervention control commands, the stability index, behavioral state deviation, driving safety risk level, risk behavior identification results, and driving intervention control commands at the corresponding time step are recorded to form a driving risk event record, which is then stored in chronological order as a driving risk event sequence.

[0014] The beneficial effects of this invention are: This invention, based on an improved Wiener process model and the Milstein-Euler prediction and correction algorithm, accurately describes the continuous and stochastic changes in driving behavior by introducing the fusion of multimodal data and combining factors such as behavioral state deviation and risk accumulation. Through dynamic adjustment of drift and diffusion terms, particularly by introducing a diffusion term adjustment mechanism driven by behavioral state deviation and a time-varying drift term adjustment mechanism, accurate prediction and real-time response to driver behavior are achieved. By exponentially mapping the risk accumulation, the volatility of risk assessment is effectively reduced, enabling the system to provide timely warnings in emergency situations.

[0015] This invention can accurately reflect the stability and safety of driver behavior in real time. Through deep fusion of multimodal data, it overcomes the limitations of traditional single-data source technologies. By introducing an improved Wiener process model and the Milstein-Euler prediction and correction algorithm, it effectively improves the accuracy and response speed of dynamic modeling of driving behavior. Through stability mapping and risk behavior identification, it achieves efficient risk assessment and intervention measures, providing a more reliable guarantee for driving safety. This invention not only improves the accuracy and real-time performance of driving behavior prediction but also possesses the ability to respond quickly in complex and sudden driving situations, contributing to improved road safety and reducing the risk of traffic accidents. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of a driving safety behavior intelligent management method based on multimodal fusion proposed in this invention; Figure 2 This is a schematic diagram of the improved Wiener process model structure of a multimodal fusion-based intelligent management method for driving safety behavior proposed in this invention. Figure 3This is a flowchart of the Milstein-Euler prediction and correction algorithm for a multimodal fusion-based intelligent management method for driving safety behavior proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for intelligent management of driving safety behavior based on multimodal fusion, comprising: Collect multi-source, multi-modal data of driving scenarios, and preprocess the multi-source, multi-modal data to generate standardized multi-source, multi-modal data; An improved Wiener process model is constructed to process standardized multi-source multimodal data. By setting behavioral state thresholds, behavioral degradation adjustment factors, and state deviation sensitivity coefficients, the drift and diffusion terms are dynamically adjusted. The behavioral state deviation is constructed based on the difference between the behavioral state variables and the behavioral state thresholds, forming stochastic differential equations. The Milstein-Euler predictor-corrector algorithm is used to solve the stochastic differential equation in discrete time. By performing Euler prediction, dual prediction of drift and diffusion terms, and diffusion modulation including state-related damping factors, the driving behavior state value of the next time step is obtained. Continuous time series processing is performed on the driving behavior state values, including analyzing the rate of change of state between adjacent time steps, suppressing abnormal mutation points, and smoothing the behavior state sequence to obtain the processed driving behavior state sequence. Based on the processed driving behavior state sequence, a cumulative amount of behavioral risk is constructed using a time accumulation method, and an exponential mapping is performed on the cumulative amount of risk to generate a stability index that reflects the safety of driving behavior. Based on stability indicators and behavioral state deviations, a driving safety behavior management strategy is implemented, which includes risk behavior identification, early warning prompts, driving intervention control, and risk event recording.

[0020] In this embodiment, the multi-source multimodal data specifically includes vehicle operating status data, driver status data, and external environment status data.

[0021] In this embodiment, the preprocessing of multi-source multimodal data specifically includes time synchronization, data cleaning, data missing completion, feature standardization, and data alignment.

[0022] In this embodiment, the formation of the stochastic differential equation specifically refers to: An improved Wiener process model is constructed, comprising an input mapping layer, a state-driven diffusion regulation layer, a drift regulation layer, and an equation construction layer. Specifically, the construction of the improved Wiener process model involves: An input mapping layer is added at the bottom of the Wiener process model. Behavioral state deviations are introduced to adjust the basic structure of the diffusion and drift terms of the Wiener process model, resulting in a state-driven diffusion adjustment layer and a drift adjustment layer. An equation construction layer is added after the drift adjustment layer to replace the original direct construction method of stochastic differential equations in the Wiener process model, thus obtaining the improved Wiener process model. The input mapping layer constructs behavioral state variables based on standardized multi-source multimodal data. Key features are selected from vehicle operation state data, driver state data, and external environment state data. These key features are aligned, normalized, and dimensionally standardized according to a unified time reference. The processed key features are then weighted and fused to obtain behavioral state variables representing the continuous-time changes in driving behavior. Key features include vehicle speed, acceleration, steering wheel angle, brake opening and lane departure from vehicle operating status data; eye direction, head posture, attention characteristics and fatigue characteristics from driver status data; and distance to the vehicle in front, road curvature, lighting conditions and weather conditions from external environmental status data. The state-driven diffusion regulation layer calculates the basic diffusion quantity based on the fluctuation range and intensity of the behavioral state variables. It introduces a behavioral degradation regulation factor and a state deviation sensitivity coefficient to jointly regulate the baseline amplitude of the basic diffusion quantity and the response sensitivity to behavioral state deviations. The behavioral state deviation is obtained by differentiating the behavioral state variables from the behavioral state threshold. A behavior-state deviation-driven diffusion term exponential regulation mechanism is then introduced, combining the behavioral state deviation, the behavioral degradation regulation factor, and the state deviation sensitivity coefficient to perform exponential nonlinear regulation of the basic diffusion quantity, resulting in a dynamically adjusted diffusion term, where: The calculation of the basic diffusion amount is as follows: The continuous values ​​of the behavioral state variable within the time window are read, and the difference between the maximum and minimum values ​​of the behavioral state variable within the time window is calculated as the behavioral state fluctuation range. The changes of the behavioral state variable between adjacent time steps are statistically analyzed, and the average value of the changes is calculated as the behavioral state fluctuation intensity. The fluctuation range and fluctuation intensity are weighted and summed to generate the basic diffusion amount. The degradation adjustment factor coefficients the baseline diffusion magnitude and multiplies it by the degradation adjustment factor. The baseline diffusion magnitude exhibits different initial fluctuation scales under different driving behavior states, forming the baseline diffusion magnitude adjustment result based on the behavior degradation adjustment factor. The state deviation sensitivity coefficient is combined with the behavioral state deviation information to gain or attenuate the response sensitivity of the diffusion term when the deviation changes. The behavioral state deviation is multiplied by the state deviation sensitivity coefficient to obtain the diffusion response adjustment amount that changes with the deviation. The diffusion term exponential adjustment mechanism forms an exponential adjustment quantity by weighting and summing the behavioral state deviation, behavioral degradation adjustment factor and state deviation sensitivity coefficient. The exponential adjustment quantity is then subjected to an exponential nonlinear transformation with the basic diffusion quantity. When the exponential adjustment quantity is positively offset, the basic diffusion quantity is amplified; when the exponential adjustment quantity is negatively offset, the basic diffusion quantity is attenuated, thus generating a dynamically adjusted diffusion term. The drift adjustment layer calculates the base drift based on the direction, magnitude, and trend of change of behavioral state variables at adjacent time points. A time-varying driving mechanism is introduced to dynamically process the base drift over time, yielding a drift term. The behavioral state deviation is then input into the drift term adjustment process, and the drift term is dynamically adjusted based on this deviation, resulting in a drift term adjusted for behavioral state deviation. Where: The calculation of the basic drift amount is as follows: The values ​​of the behavioral state variable at adjacent time points are read, the difference between adjacent time points is calculated, the sign of the difference represents the direction of change of the behavioral state variable, the absolute value of the difference is extracted to obtain the magnitude of the change of the behavioral state, the direction of change and the magnitude of change of multiple consecutive time steps are weighted and summed to form the trend of change of the behavioral state variable at the current time, and the direction of change, the magnitude of change and the trend of change are weighted and summed to generate the basic drift. The time-varying driving mechanism reads the time stamp of the current time step, combines the time stamp with the basic drift amount, scales the basic drift amount over time, multiplies the basic drift amount with the time weight corresponding to the time stamp, and scales the basic drift amount proportionally according to the time weight. The basic drift amount presents a change pattern corresponding to the time stamp at different time steps, generating a drift term that changes over time. The aforementioned deviation-driven dynamic adjustment of the drift term specifically includes: By weighted superposition of behavioral state deviation and drift term, the drift term increases or decreases accordingly based on the magnitude of behavioral state deviation. When the behavioral state deviation exceeds a threshold, the drift term is increased; when the behavioral state deviation is below the threshold, the drift term is decreased or increased, thus generating a drift term adjusted for deviation. The equation construction layer combines the drift term adjusted for behavioral state bias with the diffusion term adjusted for dynamic state bias, and performs consistency calibration in the time domain. It introduces dynamic weights based on the state change rate to adjust the proportions of the drift term adjusted for behavioral state bias and the diffusion term adjusted for dynamic state bias in the stochastic differential equation. This constrained reconstruction of the combined term-level structure satisfies the dynamic response characteristics of the continuous stochastic evolution of driving behavior, and constructs a stochastic differential equation describing the continuous stochastic evolution of driving behavior, where: The consistency calibration in the time domain specifically includes: By performing correspondence processing on the drift term adjusted for behavioral state deviation and the diffusion term adjusted for dynamic state deviation under the time mark of the current time step, the two are aligned according to the same time step, and the time increment of the drift term and the random increment of the diffusion term are weighted to form a calibrated combined result. The adjustment of the proportion of the drift term (adjusted for behavioral state deviation) and the diffusion term (adjusted for dynamic adjustment) in the stochastic differential equation is specifically as follows: By reading the changes in behavioral state variables between adjacent time steps, the rate of state change is calculated and converted into corresponding dynamic weights. When the rate of state change is large, the corresponding dynamic weight value is high, and when the rate of state change is small, the corresponding dynamic weight value is low. The dynamic weights are then applied to the drift term after behavioral state deviation adjustment and the diffusion term after dynamic adjustment. The proportions of the two in the stochastic differential equation are combined according to their weights. The contribution ratios of the drift term and the diffusion term in the stochastic differential equation are adjusted synchronously with the rate of state change. The specific steps for constraining and reconstructing the combined item-level structure are as follows: By restricting the numerical range, direction of change, and time update amplitude of the combined drift and diffusion terms, the drift and diffusion terms are truncated, normalized, and boundary adjusted according to the constraints. The combination of the two terms is then rearranged so that the combined term-level structure meets the update coherence requirements of continuous-time stochastic evolution, forming a reconstructed term that satisfies the characteristics of continuous stochastic evolution of driving behavior. The construction of the stochastic differential equations describing the continuous stochastic evolution of driving behavior is specifically as follows: By taking the drift term adjusted for behavioral state deviation as the deterministic change part of the equation and the diffusion term adjusted for dynamic state deviation as the random change part of the equation, and combining the two according to the term-level arrangement of stochastic differential equations, and connecting them with the time increment corresponding to the time step and the random increment corresponding to the random disturbance as the time input and random input of the equation, a stochastic differential equation is formed.

[0023] In this embodiment, obtaining the driving behavior state value at the next time step includes: Determine the discrete time step of the stochastic differential equation, establish a discrete time series, and use the driving behavior state corresponding to the stochastic differential equation at the initial time as the initial value. Define the number, time position, and corresponding driving behavior state storage structure for each computation time step, where: The establishment of the discrete time series specifically involves: By using a determined discrete time step as the starting point, time markers corresponding to each time step are generated sequentially by accumulating the time markers in advance according to the discrete time step, and the continuously generated time markers are arranged in chronological order to form a discrete time sequence. The specified number, time position, and corresponding driving behavior state storage structure for each calculation time step are as follows: By assigning serial numbers to each time step in the discrete time series according to the generation order, recording the time markers corresponding to each time step as time positions, and establishing a driving behavior state storage unit corresponding to the number for each time step, each time step has a unique number, a clear time position, and a data storage structure for recording driving behavior state values. The Milstein-Euler prediction and correction algorithm is executed. Based on the discrete time step, a random disturbance increment corresponding to the current time step is generated. This random disturbance increment is paired with the current driving behavior state under the same time marker and combined to form the random drive input data for the current time step, where: The generation of the random perturbation increment is specifically as follows: By calling a random number generation method within the discrete time step range corresponding to the current time step, a random value matching the scale is generated based on the value of the discrete time step, and the random value is used as the random perturbation increment of the current time step. The combination forms the random driving input data for the current time step, specifically: By matching the random disturbance increment corresponding to the current time step with the current driving behavior state according to the same time mark and merging the two, the random disturbance increment and the current driving behavior state form a data pair in a one-to-one correspondence, forming random driving input data; Based on the driving behavior state at the current time step, the drift and diffusion terms in the stochastic differential equation, the discrete time step, and the stochastic driving input data, Euler prediction processing is performed. The drift term is multiplied by the discrete time step to obtain the drift term time increment, and the diffusion term is multiplied by the stochastic driving input data to obtain the diffusion term random increment. The drift term time increment, the diffusion term random increment, and the driving behavior state at the current time step are then summed to obtain the predicted driving behavior state for the next time step. The Euler prediction process is performed as follows: By reading the driving behavior state, drift term, diffusion term, discrete time step, and random drive input data in the current time step, the drift term and diffusion term are processed in the order of Euler operations. The drift term is multiplied by the discrete time step to obtain the drift term time increment, and the diffusion term is multiplied by the random drive input data to obtain the diffusion term random increment. The drift term time increment, diffusion term random increment, and current driving behavior state are summed and superimposed to form the predicted driving behavior state for the next time step. A dual predictor structure is introduced. The drift term corresponding to the current time step is read as the first drift prediction. The behavior state variable corresponding to the estimated driving behavior state of the next time step is read and input into the drift adjustment structure. The drift term adjustment parameters are re-invoked. By processing and calculating the direction, magnitude, and trend of behavior state changes, a second drift prediction corresponding to the next time step is generated. The first and second drift predictions are then paired and combined with the drift term corresponding to the current time step to form the estimated correction value for the drift term. Where: The specific processing and calculation of the direction, magnitude, and trend of behavioral state changes are as follows: The direction of state change is obtained by calculating the difference between the current time step and the next time step. The magnitude of change is obtained by calculating the difference between the current time step and the next time step. It is compared with a threshold to determine whether it exceeds the range. If the magnitude of change exceeds the threshold, it is considered that the change has a large fluctuation and proceeds to the next step. Based on the state data of multiple consecutive time steps, the trend of state change is obtained by calculating the moving average of multiple time steps. The direction of change, the magnitude of change, and the trend of change are combined and weighted to obtain the adjustment amount. The diffusion term corresponding to the current time step is read as the first diffusion prediction. The behavior state variable corresponding to the estimated driving behavior state of the next time step is read and input into the diffusion term adjustment structure. The diffusion term adjustment parameters are called, and the calculation is performed based on the fluctuation amplitude and frequency of the behavior state to generate the second diffusion prediction for the next time step. A noise modulation factor is introduced, and a state-related damping structure is constructed exponentially using the noise modulation factor, generating a damping adjustment. The damping adjustment, the first diffusion prediction, the second diffusion prediction, and the diffusion term corresponding to the current time step are weighted and combined to form the estimated correction of the diffusion term, where: The processing and calculation based on the fluctuation amplitude and frequency of the behavioral state are as follows: Based on the behavioral state data of the current time step and the next time step, the behavioral state difference between adjacent time steps is calculated and compared with the fluctuation threshold to determine the intensity of the fluctuation. Data from multiple consecutive time steps is analyzed to calculate the frequency of state fluctuations. The fluctuation amplitude and fluctuation frequency are combined. When the fluctuation amplitude is large, the sensitivity of the diffusion term needs to be increased, and the adjustment coefficient is increased accordingly. When the fluctuation frequency is high, the rate of change of the diffusion term needs to be increased, and the adjustment coefficient will also increase. The fluctuation amplitude and fluctuation frequency are weighted and synthesized to generate a comprehensive adjustment coefficient. The response intensity and rate of change of the diffusion term are adjusted based on the comprehensive adjustment coefficient. When the comprehensive adjustment coefficient is large, the response intensity and rate of change of the diffusion term are increased. When the comprehensive adjustment coefficient is small, the response intensity and rate of change of the diffusion term are decreased, generating a second diffusion prediction quantity. The noise modulation factor is adjusted based on the difference between the current behavioral state and the risk threshold. By using an exponential adjustment function, the influence of noise is effectively suppressed when the behavioral state deviation is large, and the strength of the diffusion term is weakened. When the behavioral state deviation is close to the risk threshold, the noise modulation factor decreases, increasing the influence of noise on the diffusion term. The noise modulation factor generates a damping adjustment amount by multiplying the noise modulation factor with the diffusion term. The estimated correction values ​​for the drift and diffusion terms are combined with the driving behavior state at the current time step, the discrete time step size, and the random driving input data to perform numerical updates, thereby calculating the driving behavior state value for the next time step. Specifically, the calculation of the driving behavior state value for the next time step involves: Numerical updates are performed by combining the estimated correction values ​​of the drift term and the diffusion term with the driving behavior state, discrete time step, and random driving input data at the current time step. The driving behavior state and random driving input data at the current time step are weighted and combined with the time step. The drift term is multiplied by the time step to obtain the drift term time increment, and the diffusion term is multiplied by the random driving input data to obtain the diffusion term random increment. The drift term time increment and the diffusion term random increment are superimposed on the driving behavior state at the current time step to obtain the driving behavior state at the next time step. In this embodiment, the processed driving behavior state sequence includes: The driving behavior state values ​​are arranged in chronological order, and the driving behavior state values ​​at each time step are paired with the corresponding time markers to form a time sequence of behavior states to be processed. The driving behavior state values ​​at adjacent time steps in the behavioral state time series are differentially processed to calculate the state change rate at each time step. The state change rate is used as the basic data for identifying abrupt changes and fluctuation anomalies, and a state change rate sequence is constructed. The construction of the state change rate sequence specifically involves: By performing differential processing on the driving behavior state values ​​of adjacent time steps in the behavior state time series, the state change rate of each time step is calculated. The driving behavior state value of the current time step is compared with the state value of the previous time step, the difference is calculated, and the difference is divided by the time step length to obtain the state change rate. Based on the state change rate of each time step, a state change rate sequence is formed. Anomaly suppression processing is performed on the behavioral state time series based on the state change rate sequence. This involves marking time points where the rate of change exceeds a threshold, limiting the amplitude of the driving behavior state value at the corresponding time step, and generating an anomaly-suppressed behavioral state time series. The anomaly suppression processing is specifically performed as follows: By marking and identifying abnormal mutation points, when the rate of change of the state at a certain time step exceeds the change threshold, it is considered that there is a mutation or abnormal fluctuation at the current time point. For the driving behavior state value at the current time step, amplitude limiting processing is applied. When the driving behavior state value is out of range, the driving behavior state value is automatically limited to the maximum and minimum values ​​to avoid the impact of abnormal fluctuations on the entire behavior state time series. The time series of driving behavior states after anomaly suppression is smoothed by averaging the driving behavior state values ​​of consecutive time steps using a window, thereby reducing the short-term fluctuations of the driving behavior state series and generating a processed driving behavior state series. Specifically, the smoothing process for the time series of driving behavior states after anomaly suppression involves: The time series of behavioral states after anomaly suppression is smoothed to reduce the impact of short-term fluctuations. The driving behavior state values ​​of consecutive time steps are weighted and averaged. The driving behavior state values ​​of each time step within the corresponding window are weighted, and the more stable driving behavior state values ​​are given higher weights. In this embodiment, generating a stability index reflecting driving behavior safety includes: Based on the processed driving behavior state sequence, the driving behavior state values ​​of each time step are read in chronological order, and each driving behavior state value is paired with the corresponding behavior state deviation to form a time series input; The time series input is accumulated step by step, and the behavioral state deviation at each time step is gradually accumulated according to the time accumulation rule to obtain the risk accumulation amount, which reflects the degree of risk accumulation of driving behavior over the entire time period. An exponential mapping process is performed based on the accumulated risk amount. This process involves exponentially decaying the accumulated risk amount according to the exponential mapping parameters to generate a stability mapping quantity that reflects the stability of driving behavior. Specifically, the exponential mapping process based on the accumulated risk amount involves: The risk accumulation index is converted into an exponential decay value according to the mapping parameter. As the risk accumulation increases, the index mapping parameter grows exponentially. When the risk of driving behavior increases, the index mapping parameter increases and the stability mapping value decreases. When the risk accumulation is low, the index mapping parameter decreases and the stability mapping value increases. Through exponential mapping, the risk accumulation is converted into the stability mapping value. The stability mapping is processed according to time steps. The stability mapping at each time step is normalized and mapped to a numerical range to generate a stability index value within the target numerical range. Specifically, the normalization and numerical range mapping of the stability mapping at each time step involves: The normalization process transforms the stability mapping quantity into a standardized value that falls within a uniform range. Based on the maximum and minimum values ​​of the stability mapping quantity, the stability mapping quantity at each time step is adjusted proportionally. All stability mapping quantities are within the range of [0,1]. Numerical range mapping is then performed to map the normalized stability mapping quantity to the target numerical range. In this embodiment, the execution of the driving safety behavior management strategy, which includes risk behavior identification, early warning prompts, driving intervention control, and risk event recording, includes: Based on stability indices and corresponding behavioral state deviations, stability indices are read sequentially at each time step. These indices are compared with stability thresholds, and combined with behavioral state deviations, the driving safety risk level at each time step is determined, forming a driving safety risk level sequence. The method of determining the driving safety risk level at each time step by combining behavioral state deviations is as follows: The stability index of each time step is read in chronological order and compared with the stability threshold. When the stability index is lower than the stability threshold, it indicates that the driving behavior is less stable and there is a higher risk. By combining the behavioral state deviation, the risk level of each time step is further adjusted. When the behavioral state deviation is large, it indicates that the driving behavior deviates significantly from the normal standard. At this time, the corresponding risk level should be increased, and vice versa. The driving safety risk level of each time step is determined. The formation of the driving safety risk level sequence is specifically as follows: The driving safety risk levels at each time step are arranged and combined to form a driving safety risk level sequence; Based on the driving safety risk level sequence, time steps where the stability index is below the stability threshold and the behavioral state deviation exceeds the deviation threshold are filtered. The driving behavior state at the corresponding time step is marked as a risky behavior time period, and the risky behavior time periods are classified according to the behavioral state deviation to generate risky behavior identification results. The classification of risky behavior time periods according to the behavioral state deviation is as follows: Based on the driving safety risk level sequence, time steps where the stability index is below the stability threshold and the behavioral state deviation exceeds the deviation threshold are selected as risky behavior time periods. Based on the magnitude of the behavioral state deviation, the selected risky behavior time periods are classified. When the behavioral state deviation is large, the driving behavior is more abnormal and there is a higher risk, which is classified as high-risk behavior. When the behavioral state deviation is small, the driving behavior is closer to the normal standard and is classified as medium-risk behavior. Based on the results of risk behavior identification and the corresponding driving safety risk level, a warning message is generated and sent to the human-machine interface; When the driving safety risk level reaches the intervention threshold, a driving intervention control command is generated and sent to the vehicle control execution module to adjust the vehicle's operating parameters. When generating driving intervention control commands, the stability index, behavioral state deviation, driving safety risk level, risk behavior identification results, and driving intervention control commands at the corresponding time step are recorded to form a driving risk event record, which is then stored in chronological order as a driving risk event sequence.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to the entire process of intelligent driving safety management for a ride-hailing fleet. Multimodal driving safety monitoring terminals were installed in 36 GAC Aion A60 vehicles in the pilot fleet to collect real-time vehicle dynamics data, driver physiological and behavioral data, external road environment data, and emotion recognition video images. All data is aggregated to a cloud-based driving safety behavior analysis platform via an onboard 5G communication module. After standardized preprocessing, the data enters the improved Wiener process model and Milstein-Euler prediction and correction solution process to calculate driving behavior states and dynamic stability indicators in real time.

[0025] The pilot period was from March to August 2025, coinciding with the spring return-to-work rush and weekday traffic peaks, resulting in severe urban road congestion. Drivers were generally under high pressure and high-frequency operation, making it difficult for traditional fatigue detection or OBD single-parameter trigger-based risk assessment methods to detect potential risk patterns such as the coupling of inattention and distance fluctuations, and emotional driving cycles in a timely manner. Under the multimodal fusion mechanism of this invention, the platform can automatically focus on the feature subspace most sensitive to risk, filter out local interference in high-noise environments through a diffusion term adjustment mechanism, and replace single threshold judgment with the evolution trend of driving behavior, so that risks can be identified and quantified in the early stages of formation. On the 12th day of operation, the system detected three drivers with short-term inattention and distance fluctuation linkage risks through the stability time-domain decay characteristics, and located seven key road sections in just 90 seconds. On the 21st day, it successfully identified two drivers exhibiting repeated cycles of emotional acceleration and sudden braking at night, providing early warnings before any dangerous situations occurred, and helping driving behavior return to stability.

[0026] Compared to traditional manual inspections and OBD single-parameter risk control, the pilot fleet's missed detection rate decreased from 17.9% to 2.4%, and the false alarm rate decreased from 9.7% to 2.8%. The average time for a single risk location was shortened from 28 minutes to approximately 2.7 minutes, and the average warning lead time before a risk trend occurs increased from approximately 6.5 seconds to 41.2 seconds, enabling drivers and the management platform to decelerate, adjust distances, and intervene before the risk actually materializes. During the 6-month pilot period, a total of 197,000 kilometers were driven. The number of emergency braking incidents decreased from 1.87 per 1,000 kilometers to 0.66, and the number of rear-end collisions decreased from 61 to 22, with the overall incident rate decreasing by 58.4% compared to the control group. The average driving behavior stability score increased from 0.54 to 0.71, an overall improvement of approximately 31.5%, with the most significant improvement observed during weekday morning and evening rush hours. A driver survey conducted after the pilot program concluded showed that 92% of drivers believed the system intervention was not overly disruptive but helped reduce risk, and driving comfort was not significantly affected.

[0027] Table 1. Comparison of key indicators between the method of this invention and traditional driving safety management methods during the pilot period.

[0028] As shown in Table 1, during the pilot period from March to August 2025, the method of this invention significantly improved risk identification capabilities compared to traditional manual inspection and single-parameter OBD risk control. The success rates for identifying fatigue, distracted driving, and emotional driving increased from 71.6%, 68.2%, and 57.5% to 93.4%, 92.1%, and 88.7%, respectively. Simultaneously, the false negative rate decreased from 17.9% to 2.4%, and the false positive rate decreased from 9.7% to 2.8%. This indicates that multimodal fusion and behavioral evolution modeling effectively reduced instances of missed or false alarms, resulting in more comprehensive and accurate risk identification.

[0029] In terms of risk handling efficiency, this invention reduces the average time for a single risk location from 28 minutes to approximately 2.7 minutes, increases the average warning lead time before a risk trend occurs from approximately 6.5 seconds to 41.2 seconds, and improves the cumulative risk intervention success rate from 63.4% to 91.8%. This demonstrates that the system can not only identify problems faster but also reserve sufficient intervention windows before risks truly develop into dangerous situations, significantly enhancing the initiative of fleet dispatching and driver self-correction.

[0030] From the perspectives of driving safety and subjective experience, the number of emergency braking incidents decreased from 1.87 per 1,000 kilometers to 0.66, rear-end collision incidents decreased from 61 to 22 within 6 months, safety complaints per vehicle per 10,000 kilometers decreased from 0.42 to 0.11, and the average driving behavior stability score increased from 0.54 to 0.71, representing an overall improvement of approximately 31.5%. Simultaneously, the driver acceptance rate of warnings and interventions increased from 76% to 92%, and the comfort score improved from 7.4 to 8.1. This indicates that the invention improves safety while essentially not increasing the driving burden, demonstrating its engineering practicality and scalability in real-world operating environments.

[0031] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent management of driving safety behavior based on multimodal fusion, characterized in that, include: Collect multi-source, multi-modal data of driving scenarios, and preprocess the multi-source, multi-modal data to generate standardized multi-source, multi-modal data; An improved Wiener process model is constructed to process standardized multi-source multimodal data. By setting behavioral state thresholds, behavioral degradation adjustment factors, and state deviation sensitivity coefficients, the drift and diffusion terms are dynamically adjusted. The behavioral state deviation is constructed based on the difference between the behavioral state variables and the behavioral state thresholds, forming stochastic differential equations. The Milstein-Euler predictor-corrector algorithm is used to solve the stochastic differential equation in discrete time. By performing Euler prediction, dual prediction of drift and diffusion terms, and diffusion modulation including state-related damping factors, the driving behavior state value of the next time step is obtained. Continuous time series processing is performed on the driving behavior state values, including analyzing the rate of change of state between adjacent time steps, suppressing abnormal mutation points, and smoothing the behavior state sequence to obtain the processed driving behavior state sequence. Based on the processed driving behavior state sequence, a cumulative amount of behavioral risk is constructed using a time accumulation method, and an exponential mapping is performed on the cumulative amount of risk to generate a stability index that reflects the safety of driving behavior. Based on stability indicators and behavioral state deviations, a driving safety behavior management strategy is implemented, which includes risk behavior identification, early warning prompts, driving intervention control, and risk event recording.

2. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The multi-source multimodal data specifically includes vehicle operating status data, driver status data, and external environment status data.

3. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The preprocessing of multi-source, multi-modal data specifically includes time synchronization, data cleaning, data missing completion, feature standardization, and data alignment.

4. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The formation of the stochastic differential equation specifically involves: An improved Wiener process model is constructed, which consists of an input mapping layer, a state-driven diffusion regulation layer, a drift regulation layer, and an equation construction layer. The input mapping layer constructs behavioral state variables based on standardized multi-source multimodal data. It selects key features from vehicle operation state data, driver state data, and external environment state data, aligns, normalizes, and unifies the dimensions of the key features according to a unified time reference, and then fuses the weighted combination of multiple key features to obtain behavioral state variables that characterize the changes in driving behavior over continuous time. The state-driven diffusion regulation layer calculates the basic diffusion amount based on the fluctuation range and intensity of the behavioral state variables. It introduces a behavioral degradation regulation factor and a state deviation sensitivity coefficient to jointly regulate the baseline amplitude of the basic diffusion amount and the response sensitivity to behavioral state deviation. It obtains the behavioral state deviation by differentiating the behavioral state variables from the behavioral state threshold, and introduces a diffusion term exponential regulation mechanism driven by behavioral state deviation. It combines the behavioral state deviation, the behavioral degradation regulation factor, and the state deviation sensitivity coefficient to perform exponential nonlinear regulation on the basic diffusion amount, resulting in a dynamically adjusted diffusion term. The drift adjustment layer calculates the basic drift amount based on the direction, magnitude and trend of change of the behavioral state variable at adjacent time points. It introduces a time-varying driving mechanism to dynamically process the basic drift amount as it changes over time, thereby obtaining the drift term. The behavioral state deviation is then input into the drift term adjustment process to dynamically adjust the drift term driven by the deviation, resulting in the drift term adjusted by the behavioral state deviation. The equation construction layer combines the drift term adjusted by behavioral state deviation with the diffusion term adjusted by dynamic adjustment, and performs consistency calibration in the time domain. It introduces dynamic weight adjustment based on the state change rate to adjust the proportion of the drift term adjusted by behavioral state deviation and the diffusion term adjusted by dynamic adjustment in the stochastic differential equation. The combined term-level structure is constrained and reconstructed to meet the dynamic response characteristics of the continuous stochastic evolution of driving behavior, and a stochastic differential equation describing the continuous stochastic evolution of driving behavior is constructed.

5. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The process of obtaining the driving behavior state value for the next time step includes: Determine the discrete time step of the stochastic differential equation, establish a discrete time series, take the driving behavior state corresponding to the stochastic differential equation at the initial time as the initial value, and specify the number, time position and corresponding driving behavior state storage structure of each calculation time step. The Milstein-Euler prediction and correction algorithm is executed to generate random disturbance increments corresponding to the current time step based on the discrete time step. The random disturbance increments are paired with the current driving behavior state under the same time mark and combined to form the random driving input data of the current time step. Based on the driving behavior state at the current time step, the drift term and diffusion term in the stochastic differential equation, the discrete time step, and the stochastic driving input data, Euler prediction processing is performed. The drift term is multiplied by the discrete time step to obtain the drift term time increment, and the diffusion term is multiplied by the stochastic driving input data to obtain the diffusion term random increment. The drift term time increment, the diffusion term random increment, and the driving behavior state at the current time step are summed to obtain the predicted driving behavior state for the next time step. A dual predictor structure is introduced. The drift term corresponding to the current time step is read as the first drift prediction. The behavior state variable corresponding to the estimated driving behavior state of the next time step is read. The behavior state variable is input to the drift adjustment structure, and the drift term adjustment parameters are called back. By processing and calculating the change direction, change magnitude and change trend of the behavior state, the second drift prediction corresponding to the next time step is generated. The first drift prediction and the second drift prediction are paired and combined with the drift term corresponding to the current time step to form the estimated correction amount of the drift term. The diffusion term corresponding to the current time step is read as the first diffusion prediction quantity. The behavior state variable corresponding to the estimated driving behavior state of the next time step is read. The behavior state variable is input into the diffusion term adjustment structure. The diffusion term adjustment parameters are called. The fluctuation amplitude and fluctuation frequency of the behavior state are processed and calculated to generate the second diffusion prediction quantity corresponding to the next time step. A noise modulation factor is introduced. The noise modulation factor is used to construct a state-related damping structure in an exponential form and generate a damping adjustment quantity. The damping adjustment quantity, the first diffusion prediction quantity, the second diffusion prediction quantity and the diffusion term corresponding to the current time step are weighted and combined to form the predicted correction quantity of the diffusion term. The estimated correction values ​​for the drift term and the diffusion term are combined with the driving behavior state at the current time step, the discrete time step size, and the random driving input data to update the values ​​and calculate the driving behavior state value for the next time step.

6. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The processed driving behavior state sequence includes: The driving behavior state values ​​are arranged in chronological order, and the driving behavior state values ​​at each time step are paired with the corresponding time markers to form a time sequence of behavior states to be processed. The driving behavior state values ​​of adjacent time steps in the behavior state time series are differentially processed to calculate the state change rate of each time step. The state change rate is used as the basic data for identifying abrupt change points and fluctuation anomalies, and a state change rate sequence is constructed. Anomaly suppression processing is performed on the behavioral state time series based on the state change rate sequence. By marking the time positions where the change rate exceeds the change threshold, the amplitude of the driving behavior state value at the corresponding time step is limited, and anomaly suppression is generated for the behavioral state time series. Smoothing is performed on the time series of driving behavior states after anomaly suppression. The driving behavior state values ​​of continuous time steps are weighted by windows to reduce the short-term fluctuation amplitude of the driving behavior state series and generate the processed driving behavior state series.

7. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The generation of stability indicators reflecting driving behavior safety includes: Based on the processed driving behavior state sequence, the driving behavior state values ​​of each time step are read in chronological order, and each driving behavior state value is paired with the corresponding behavior state deviation to form a time series input; The time series input is accumulated step by step, and the behavioral state deviation at each time step is gradually accumulated according to the time accumulation rule to obtain the risk accumulation amount, which reflects the degree of risk accumulation of driving behavior over the entire time period. Based on the risk accumulation, an exponential mapping process is performed, and the risk accumulation is transformed by exponential decay according to the exponential mapping parameters to generate a stability mapping quantity that reflects the stability of driving behavior. The stability mapping is processed according to time steps. The stability mapping at each time step is normalized and mapped to a numerical range to generate a stability index value within the target numerical range.

8. The intelligent management method for driving safety behavior based on multimodal fusion according to claim 1, characterized in that, The implementation of the driving safety behavior management strategy includes risk behavior identification, early warning prompts, driving intervention control, and risk event recording, including: Based on stability indices and corresponding behavioral state deviations, stability indices at each time step are read sequentially over time. The stability indices are compared with stability thresholds, and the driving safety risk level at each time step is determined by combining the behavioral state deviations, thus forming a driving safety risk level sequence. Based on the driving safety risk level sequence, time steps with stability indices below the stability threshold and behavioral state deviations exceeding the deviation threshold are filtered out. The driving behavior state of the corresponding time step is marked as a risk behavior time period, and the risk behavior time periods are classified according to the behavioral state deviation to generate risk behavior identification results. Based on the results of risk behavior identification and the corresponding driving safety risk level, a warning message is generated and sent to the human-machine interface; When the driving safety risk level reaches the intervention threshold, a driving intervention control command is generated and sent to the vehicle control execution module to adjust the vehicle's operating parameters. When generating driving intervention control commands, the stability index, behavioral state deviation, driving safety risk level, risk behavior identification results, and driving intervention control commands at the corresponding time step are recorded to form a driving risk event record, which is then stored in chronological order as a driving risk event sequence.

Images