Method for analyzing influence of subjective workload on driving behavior risk

Abstract

The invention relates to a method for analyzing the influence of subjective workload on driving behavior risk, and the method comprises the steps: obtaining the multi-dimensional basic data of a to-be-detected target, constructing a risk driving behavior index system based on the original data of driving behaviors, carrying out the normalization and weighting of risk driving behavior indexes, and obtaining a risk driving behavior result. A driving adaptability evaluation decision matrix is constructed based on driving adaptability data; a comprehensive driving adaptability level is obtained through analysis; subjective workload evaluation indexes are processed; subjective workload grades are divided according to a comprehensive efficiency value; and constructing a collaborative analysis framework comprising a generalized linear hybrid model sub-module and a support vector machine sub-module, solving the SHAP value of each subjective workload evaluation dimension to quantify the feature contribution degree, and outputting an analysis result. The method reveals the degree of influence of subjective workload on the driving behavior risk, the key influence dimension and the nonlinear rule, and can effectively improve the pertinence and effectiveness of driving safety related products and strategies.

Description

Technical Field

[0001] This application relates to the field of driver assistance technology, and in particular to a method for analyzing the impact of subjective workload on the risk of driving behavior. Background Technology

[0002] As road traffic environments become increasingly complex, drivers face increasing cognitive, physical, and time pressures while driving. Subjective workload has become a key factor affecting driving safety. The risk of driving behavior is directly related to road traffic safety, and subjective workload, as the driver's subjective perceived load on the driving task, significantly affects the driver's judgment and decision-making abilities, operational stability, and risk response strategies. This can induce dangerous driving behaviors such as rapid acceleration, rapid deceleration, and unstable steering, increasing the probability of traffic accidents.

[0003] Currently, existing research on driving behavior risk largely focuses on analyzing the impact of objective factors such as road conditions, vehicle performance, and driver age / experience, without delving deeply into the intrinsic relationship between subjective workload and driving behavior risk. While some studies do involve the assessment of subjective workload, they suffer from the following shortcomings: First, the data collection scenarios are limited, often using general road scenarios without customized designs for complex traffic conflict scenarios involving motor vehicles, non-motorized vehicles, and pedestrians, resulting in data lacking specificity and authenticity. Second, the risk driving behavior indicator system is not comprehensive enough, focusing primarily on longitudinal driving operations (such as acceleration and deceleration) while neglecting lateral indicators such as steering stability and operational abruptness, making it difficult to fully characterize driving behavior risk. Third, traditional statistical methods (such as correlation analysis and regression analysis) cannot effectively quantify the influence weights of each dimension of subjective workload and are unable to reveal the potential nonlinear relationship between it and driving risk. Fourth, driving adaptability data (such as vision and reaction ability) are not included as control variables in the analysis, leading to imprecise analysis of the influencing mechanisms and difficulty in distinguishing the independent effects of "subjective load" and "personal ability" on driving risk.

[0004] The shortcomings of the existing technologies mentioned above have prevented the full understanding of the mechanism by which subjective workload affects the risk of driving behavior. This makes it impossible to provide scientific and accurate theoretical support for optimizing driver fatigue monitoring systems, formulating intelligent assisted driving strategies, and intervening in driver safety. Therefore, there is an urgent need for a method that can comprehensively and accurately analyze the impact of subjective workload on the risk of driving behavior. Summary of the Invention

[0005] To address or partially address the problems existing in related technologies, this application provides a method for analyzing the impact of subjective workload on driving behavior risk, thereby quantifying the degree of impact of subjective workload on driving behavior risk, ranking key impact dimensions, and revealing nonlinear patterns.

[0006] The first aspect of this application provides a method for analyzing the impact of subjective workload on the risk of driving behavior, including the following steps: Acquire multi-dimensional basic data of the target to be tested. The multi-dimensional basic data includes driver basic data, raw driving behavior data, subjective workload evaluation data and driving adaptability data. The raw driving behavior data is collected under 6 preset traffic conflict scenarios. Based on raw driving behavior data, a risk driving behavior index system is constructed. The risk driving behavior index is normalized and weighted by a multi-criteria comprehensive evaluation method to obtain a comprehensive score of driving behavior risk that represents the risk level of the target under various traffic conflict scenarios. A driving adaptability evaluation decision matrix is ​​constructed based on driving adaptability data. The comprehensive driving adaptability level is analyzed, subjective workload evaluation indicators are processed, and subjective workload is divided into three levels: low, medium, and high based on the comprehensive efficiency value. A collaborative analysis framework consisting of a generalized linear mixed model submodule and a support vector machine submodule is constructed. The SHAP value of each subjective workload evaluation dimension is solved to quantify the feature contribution. The generalized linear mixed model submodule takes the comprehensive score of driving behavior risk as the target variable and the subjective workload evaluation dimension, driving adaptability level and driver basic data as input features. The output includes the analytical results of the target under test, including the quantitative coefficient of the influence of subjective workload on the risk of driving behavior, the ranking of key influence dimensions based on the SHAP mean, and the nonlinear influence law of each dimension on driving risk.

[0007] Among them, the generalized linear mixed model submodule is a beta generalized linear mixed model based on the logit link function. The driver ID of the target to be tested is the random effect, the comprehensive score of driving behavior risk is the dependent variable, the subjective workload evaluation dimension is the independent variable, and the driver's basic data and comprehensive driving adaptability level are the control variables. The maximum likelihood estimation method is used to estimate the model parameters.

[0008] The Support Vector Machine submodule and the SHAP value analysis method construct a linear proxy model through the kernel SHAP framework, including: Kernel function support vector machine is used to model the nonlinear relationship between subjective workload evaluation dimensions, driving adaptability level, driver basic data and comprehensive driving behavior risk score; Based on the kernel SHAP method, a local linear surrogate model is constructed in the input feature space to approximate the output of the support vector machine; Based on the proxy model, the SHAP value corresponding to each subjective workload evaluation dimension is calculated, and the average contribution of each dimension is obtained by averaging the SHAP values ​​of all samples. The nonlinear influence between various subjective workload dimensions and driving behavior risk was analyzed by using the scatter distribution of SHAP values ​​as a function of feature values.

[0009] The risk driving behavior indicator system includes the percentage of rapid accelerations, the percentage of rapid decelerations, the standard deviation of acceleration, the standard deviation of steering wheel angle, the steering wheel angle entropy, and the percentage of sharp turns.

[0010] The subjective workload evaluation data includes at least six dimensions: psychological needs, physical needs, time needs, personal performance, effort level, and sense of frustration.

[0011] The normalization and weighting of the risky driving behavior indicators through a multi-criteria comprehensive evaluation method includes: The TOPSIS-entropy weight method was used to analyze the driving adaptability data to obtain the overall driving adaptability level. The subjective workload evaluation dimensions were processed using the data envelopment analysis method, and the subjective workload was divided into three levels: low, medium, and high. Furthermore, a multi-criteria compromise solution ranking method is used to comprehensively score the risk driving behavior indicators, resulting in a comprehensive score for driving behavior risk.

[0012] The TOPSIS-entropy weight method was used to analyze driving adaptability data, including the following steps: A driving adaptability decision matrix is ​​constructed with night vision, dynamic vision, deep vision, selective response, and operational ability as attributes. The decision matrix is ​​then standardized, and the weight of each attribute is calculated based on its information entropy. Determine the ideal solution and the negative ideal solution, and calculate the distance between the ideal solution and the negative ideal solution. The ideal solution and the negative ideal solution are the best and worst performance values ​​of each attribute among all alternative solutions. Calculate the overall score, which is a score based on the relative proximity of each alternative solution to the ideal solution and the negative ideal solution.

[0013] The technical solution provided in this application may include the following beneficial effects: This application provides a method for analyzing the impact of subjective workload on driving behavior risk. It collects raw driving behavior data under multiple pre-defined traffic conflict scenarios, constructing a four-dimensional data system based on basic information, behavioral data, subjective perception, and adaptability. Compared to the single-dimensional data collection methods of existing technologies, this method offers more targeted and comprehensive data sources, effectively avoiding the lack of authenticity in general scenario data and laying a reliable data foundation for subsequent impact mechanism analysis. It achieves a multi-dimensional characterization of driving behavior risk, more comprehensively and accurately reflecting the characteristics of dangerous driving behaviors, avoiding risk assessment bias caused by single-dimensional indicators. The method reveals the degree of influence, key influencing dimensions, and nonlinear laws of subjective workload on driving behavior risk, effectively improving the targeting and effectiveness of driving safety-related products and strategies, thereby reducing the traffic accident rate in complex traffic environments, and possessing significant engineering application value and social safety benefits.

[0014] 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

[0015] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0016] Figure 1 This is a flowchart illustrating the method in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the risk scenario design of the method shown in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the driving adaptability assessment results of an elderly driver using the method shown in the embodiments of this application; Figure 4 These are schematic diagrams of dangerous driving behaviors under different conflict scenarios of the method shown in the embodiments of this application (pink represents left turn, blue represents right turn). Figure 5 This is a schematic diagram of the subjective workload of driving under different conflict scenarios of the method shown in the embodiments of this application (pink represents left turn, blue represents right turn). Figure 6 This is a schematic diagram of the subjective workload evaluation results of elderly drivers using the method shown in the embodiments of this application (Note: arranged in ascending order of W value and counterclockwise). Figure 7 This is a schematic diagram showing the distribution of dangerous driving behavior indicators under different subjective workload levels of the method illustrated in the embodiments of this application; Figure 8This is a schematic diagram of VIKOR evaluation values ​​for different subjective driving workload levels illustrated in the embodiments of this application; Figure 9 This is a schematic diagram of the SHAP mean of the six dimensions of subjective driving workload of the method shown in the embodiments of this application; Figure 10 This is a SHAP scatter plot of the features of the method shown in the embodiments of this application; Figure 11 This is a schematic diagram of the VIF values ​​of the independent variables in the method shown in the embodiments of this application; Figure 12 This is a schematic diagram of the correlation coefficients between the variables in the method shown in the embodiments of this application. Detailed Implementation

[0017] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0018] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0020] Unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0021] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] like Figure 1 The method shown includes the following steps to analyze the impact of subjective workload on driving behavior risk: S1. Obtain multi-dimensional basic data of the target to be tested. The multi-dimensional basic data includes driver basic data, raw driving behavior data, subjective workload evaluation data and driving adaptability data. The raw driving behavior data is collected under 6 preset traffic conflict scenarios.

[0023] The driving simulation experiment was conducted using the KMRTDS driving simulation system developed by the Road Traffic Simulation Laboratory. The KMRTDS system consists of two main components: the cockpit and the console. The cockpit serves as the physical space for the driver to perform simulated driving tasks, while the console is used to monitor and guide the experimental process.

[0024] The experimental scenario included a representative two-way three-lane urban road with each lane 3.5 meters wide. All intersections on the experimental road were unsignalized. To ensure the effectiveness and safety of the experiment, participants were clearly instructed to obey traffic laws and regulations while driving. In the conflict scenarios of the driving simulation experiment, the conflict triggering zone was set 100 meters before the entrance. The following six traffic conflict risk scenarios were designed: Figure 2 As shown, the scene number serves as an identifier in the subsequent analysis: (a) The test vehicle turns left at an intersection, while a vehicle in the opposite lane travels straight across the intersection at a speed of 30 km / h.

[0025] (b) The test vehicle turns left at an intersection, while a motorcycle in the opposite lane travels straight across the intersection at a speed of 20 km / h.

[0026] (c) The experimental vehicle turns left at the intersection, and pedestrians cross from east to west.

[0027] (d) The experimental vehicle turns right at the intersection, while a vehicle in the left-hand intersection entrance lane travels straight through the intersection at a speed of 30 km / h.

[0028] (e) The test vehicle turns right at the intersection, while the motorcycle in the right-hand entrance lane of the intersection travels straight across the intersection at a speed of 20 km / h.

[0029] (f) The experimental vehicle turns right at the intersection, and pedestrians cross from east to west.

[0030] The number of recruited drivers should meet the sample size required for the experiment. The minimum sample size N required for the experiment is calculated based on the expected variance, target confidence level, and error magnitude, as shown in Equation 1. A 10% significance level reflects a 90% confidence level for the unknown parameter. When the confidence level is 90%, Z = 1.25, and σ ranges from 0.25 to 0.50. Taking σ = 0.35 and E = 10%, the minimum sample size N is calculated to be 19.

[0031] (1) In the formula, N is the sample size, σ is the standard deviation, Z is the standard normal distribution statistic, and E is the maximum error.

[0032] This experiment recruited 50 drivers, all of whom held driver's licenses, had accumulated 60,000 kilometers or more of driving mileage, and were in good health with no diseases that would affect normal driving. Due to reasons such as experimental instrument malfunction, power outages, and drivers requesting interruptions due to discomfort, driving data from 47 valid drivers were obtained, meeting the minimum sample size requirement for the experiment.

[0033] 3) Driving adaptability test The driving adaptability of test drivers was assessed by collecting data on night vision, dynamic visual acuity, depth visual acuity, selection reaction, and operational ability. Night vision, dynamic visual acuity, and depth visual acuity were measured using appropriate testing instruments, with the average of five tests used as the evaluation index. Selection reaction and operational ability were tested using a specially constructed testing platform, with three practice sessions, and the average of ten tests used as the evaluation index.

[0034] The indicators are selected as follows: 1) Subjective workload indicators The scores of six dimensions jointly represented by NASA-TLX are used as the evaluation indicators of driver subjective workload. The specific definitions of each dimension are as follows: Psychological needs: To what extent does the task require psychological and perceptual activities (e.g., thinking, decision-making, memory, searching, etc.)?

[0035] Physical requirements: How much physical activity (e.g., rotation, control, activation, etc.) is required for this task.

[0036] Time requirement: To what extent do you feel time pressure due to the speed or pace of the task elements? Your own performance: How successful do you feel when you complete the task goals set by the experimenter (or yourself)? Effort: How much effort (mental and physical) is required to reach your performance level? Frustration: During task execution, to what extent do you feel insecure, frustrated, stressed, or annoyed compared to feelings of security, satisfaction, relaxation, or pride? 2) Risk driving behavior indicators The following metrics were selected as indicators to assess the risk of driving behavior: the proportion of sudden acceleration incidents (PSAI), the proportion of sudden deceleration incidents (PSDI), the jerk standard deviation (JerkSD), the standard deviation of steering wheel angle (SWASD), the steering wheel angle entropy (SWAE), and the proportion of sharp turning incidents (PSTI). The thresholds for sudden acceleration, sudden deceleration, and sharp turning were based on the "Specifications for Safety Evaluation of Motor Vehicle Driver Behavior Based on Vehicle Trajectory Data." The selected indicators and their meanings are shown in Table 1.

[0037] Table 1 Indicators of Risky Driving Behavior S2. Based on the original driving behavior data, a risk driving behavior index system is constructed. The risk driving behavior index is normalized and weighted through a multi-criteria comprehensive evaluation method to obtain a comprehensive score of driving behavior risk that represents the risk level of the target under various traffic conflict scenarios.

[0038] S3. Construct a driving adaptability evaluation decision matrix based on driving adaptability data, analyze the comprehensive driving adaptability level, process the subjective workload evaluation index, and divide the subjective workload into three levels: low, medium, and high based on the comprehensive efficiency value.

[0039] The normalization and weighting of the risky driving behavior indicators through a multi-criteria comprehensive evaluation method includes: The TOPSIS-entropy weight method was used to analyze the driving adaptability data to obtain the overall driving adaptability level. The subjective workload evaluation dimensions were processed using the data envelopment analysis method, and the subjective workload was divided into three levels: low, medium, and high. Furthermore, a multi-criteria compromise solution ranking method is used to comprehensively score the risk driving behavior indicators, resulting in a comprehensive score for driving behavior risk.

[0040] The TOPSIS-entropy weight method was used to analyze driving adaptability data, including the following steps: A driving adaptability decision matrix is ​​constructed with night vision, dynamic vision, deep vision, selective response, and operational ability as attributes. The decision matrix is ​​then standardized, and the weight of each attribute is calculated based on its information entropy. Determine the ideal solution and the negative ideal solution, and calculate the distance between the ideal solution and the negative ideal solution. The ideal solution and the negative ideal solution are the best and worst performance values ​​of each attribute among all alternative solutions. Calculate the overall score, which is a score based on the relative proximity of each alternative solution to the ideal solution and the negative ideal solution.

[0041] TOPSIS - Entropy Weight Method is a comprehensive evaluation technique combining the entropy weight method and TOPSIS. The entropy weight method is an objective weighting method based on the concept of entropy in information theory, determining the weight of each indicator in the comprehensive evaluation by calculating the degree of variation of the indicators. TOPSIS, on the other hand, is a commonly used multi-attribute decision analysis method that evaluates the merits of alternative solutions by comparing ideal solutions and negative ideal solutions. The specific steps of the TOPSIS - Entropy Weight Method are as follows: Step 1: Data Standardization First, the decision matrix is ​​standardized. Let there be... One alternative and There are attributes, and the decision matrix is ​​denoted as . ,in express One alternative option Performance under a given attribute. Standardization is used to eliminate the influence of different units of measurement. The standardization formula is: (1) In the formula, 'a' is the standardized value. b and c are respectively The maximum and minimum values ​​of the d attribute.

[0042] Because the dimensions and attribute orientations ("the larger the better" or "the smaller the better") of each driving adaptability index are different, this embodiment performs dimensionless standardization on the decision matrix to achieve comparability: "Benefit-oriented" indicators such as dynamic vision and depth vision are normalized to their maximum / minimum values, so that the larger the value, the better the adaptability. "Cost-based" indicators such as night vision, number of response times, and number of operational errors are transformed into a "higher is better" form through reciprocal or linear transformations, and then normalized.

[0043] Step 2: Entropy weight calculation The entropy weighting method calculates the weight of each attribute based on its information entropy. Entropy reflects the amount of information provided by each attribute; the higher the information entropy, the greater its contribution to the decision. First, the proportions of each standardized value are calculated. : (2) Next, calculate each attribute Entropy value: (3) In the formula, It is a constant. Used to normalize the entropy value. Weight of each attribute. The calculation is as follows: (4) The entropy value and difference coefficient of each attribute are calculated based on a standardized matrix, and the information entropy weight of each driving adaptability index is obtained from this. Indicators with higher weights and more information content have a higher proportion in the comprehensive evaluation, so as to more fully reflect the discrimination of different subjects on that attribute.

[0044] Step 3: Determine the ideal solution and the negative ideal solution In the TOPSIS method, the ideal solution and negative ideal solution Each attribute is defined as the best and worst performance value among all alternatives. The ideal solution for a positive attribute is the maximum value of the attribute, and the ideal solution for a negative attribute is the minimum value; for a negative attribute, the ideal solution is the minimum value, and the ideal solution for a negative attribute is the maximum value.

[0045] (5) Step 4: Calculate the distance between the ideal solution and the negative ideal solution. Then, calculate the Euclidean distance between each alternative solution and the ideal and negative ideal solutions. For the th Alternative solutions to the ideal solution and negative ideal solution The distance is calculated as follows: (6) Step 5: Calculate the overall score Finally, a relative proximity score is calculated for each alternative solution based on its distance from the ideal solution and the negative ideal solution. This score reflects how closely the alternative solution approximates the ideal solution. The overall score is calculated using the following formula: (7) The ranking of the alternatives is based on their combined scores. The higher the score, the closer the alternative is to the ideal solution, and the higher its priority.

[0046] The VIKOR multi-criteria compromise solution ranking method is a ranking method for multi-attribute decision problems (MCDM), suitable for ranking alternatives and selecting the optimal solution when conflicting attributes exist. This method has been widely used in the field of driving behavior research, emphasizing compromise solutions that are "close to the ideal solution." The specific steps are as follows: There is a set of alternative solutions. Decision-making criteria set Weight vector ,satisfy Decision matrix That is, the plan In the guidelines The following is the evaluation value.

[0047] For each criterion : (1) For each scheme Calculate the following values: Total deviation That is, the group utility index: (2) Maximum deviation That is, the individual's maximum regret index: (3) The scores here represent the degree to which the solution deviates from the ideal solution under each criterion.

[0048] Further calculation of the VIKOR comprehensive evaluation index (4) In the formula, , , , , To weigh the factors, representing the decision-maker's preference between "group utility" and "maximum personal regret," a coefficient is often chosen. .

[0049] Finally, the solutions are ranked based on the values. A VIKOR recommended solution must meet one of the following two conditions to be considered an "acceptable compromise": Approximation to optimality: Solution of The value is the minimum, and it is consistent with the scheme. satisfy: (5) Consistency: Solution exist and At least one of them ranks first in the ranking.

[0050] If only one of the two conditions is met, multiple compromise solutions can be recommended.

[0051] S4. Construct a collaborative analysis framework including a generalized linear mixed model submodule and a support vector machine submodule. Solve the SHAP value of each subjective workload evaluation dimension to quantify the feature contribution. The generalized linear mixed model submodule takes the comprehensive score of driving behavior risk as the target variable and the subjective workload evaluation dimension, driving adaptability level and driver basic data as input features.

[0052] The generalized linear mixed model submodule is a beta generalized linear mixed model based on the logit link function. It uses the driver ID of the target as the random effect, the comprehensive score of driving behavior risk as the dependent variable, the subjective workload evaluation dimension as the independent variable, and the driver's basic data and comprehensive driving adaptability level as control variables. The maximum likelihood estimation method is used to estimate the model parameters, and the inclusion of random effects terms is evaluated by the likelihood ratio test, AIC and BIC and other indicators.

[0053] The Support Vector Machine submodule and SHAP value analysis method construct a linear surrogate model through the kernel SHAP framework, including: Kernel function support vector machine is used to model the nonlinear relationship between subjective workload evaluation dimensions, driving adaptability level, driver basic data and comprehensive driving behavior risk score; Based on the kernel SHAP method, a local linear surrogate model is constructed in the input feature space to approximate the output of the support vector machine; Based on the proxy model, the SHAP value corresponding to each subjective workload evaluation dimension is calculated, and the average contribution of each dimension is obtained by averaging the SHAP values ​​of all samples. The nonlinear influence between various subjective workload dimensions and driving behavior risk was analyzed by using the scatter distribution of SHAP values ​​as a function of feature values.

[0054] Generalized linear mixture models (GLMMs) are an extension of generalized linear models. They introduce random effects terms into the fixed effects component to characterize individual differences or correlations between groups. GLMMs are widely used in research on driving behavior, including analyzing the influencing factors of distracted driver speed adjustment behavior, acceleration, and braking behavior. The calculation steps are as follows: Let the observation data be ,in: For the first The response variables of each observation; For the corresponding fixed-effects covariate vector; For the corresponding random effects covariate vector; This is a vector of fixed effects parameters; This is the vector of random effects.

[0055] The general form of GLMM can then be expressed as: (6) In the formula, This is a link function. The conditional expectation is given; the random effects hypothesis is... ,in The covariance matrix of random effects; conditional distribution assumption. , This is the scale parameter.

[0056] To explore the impact of subjective workload on dangerous driving behavior in older drivers, this study employed a generalized linear mixed model (GLMM) with driver ID as a random effect. The dependent variable was the dangerous driving behavior score, and the independent variables were six dimensions of driving workload indicators. Control variables included age, driving experience, driving mileage, and driving adaptability score. Based on the distribution type of the dependent variable, a Beta-GLMM (linked by logit) was selected for analysis. Maximum likelihood estimation was used in the model, and the model performance was compared using AIC / BIC and likelihood ratio tests.

[0057] To improve the interpretability of machine learning models, the SHAP (SHapley Additive exPlanations) method is widely used to quantify the contribution of each input feature to the model output. SVM-SHAP refers to applying the SHAP method to an SVM model, aiming to reveal the contribution of individual features to the SVM prediction results. The specific process is as follows: Step 1: Train the Support Vector Machine There is a training dataset. The optimization problem of SVM is: (6) In the formula, It represents the feature space mapping, which is implicitly defined by the kernel function.

[0058] Step 2: Apply the kernel SHAP framework Since support vector machines are non-tree models, TreeSHAP cannot be applied directly. Instead, kernel SHAP is used to approximate Shapley values.

[0059] Approximate the SVM model output using a linear surrogate model. : (7) In the formula, It is a binary vector representing the presence or absence of features. It is similar to after training The weighted linear regression model. It is a feature The SHAP value.

[0060] Step 3: Sample feature subset For each subset Construct a simplified input : Keep The characteristics remain unchanged; use the mean or background distribution to input missing features. .

[0061] Evaluation of model prediction results: (8) Repeat the above steps for multiple subsets to generate training samples: (9) Step 4: Weighted Linear Regression Solve the following weighted least squares problem to estimate : (10) In the formula, the weights for: (11) These weights are derived from the original SHAP game theory formula to ensure a fair distribution of value.

[0062] S5. Output the analytical results of the target to be tested, including the quantitative coefficient of the influence of subjective workload on driving behavior risk, the ranking of key influence dimensions based on SHAP mean, and the nonlinear influence law of each dimension on driving risk.

[0063] 1) Analysis of driving adaptability and dangerous behavior characteristics of elderly drivers: The night vision, dynamic visual acuity, depth visual acuity, selection reaction, and operational ability indicators of elderly drivers based on driving adaptability tests were analyzed, and the mean and standard deviation of each indicator are shown in Table 1. According to the physical conditions and assessment requirements for motor vehicle drivers, the standard for night vision is less than or equal to 5 seconds, the standard for dynamic visual acuity is greater than or equal to 0.2 seconds, the normal range for depth visual acuity is -22mm to 22mm, the normal range for selection reaction is less than or equal to 5 times, and the normal range for the number of operational errors is less than or equal to 130 times. Therefore, it can be seen that some of the recruited elderly drivers have insufficient driving adaptability in terms of visual function, selection reaction, or operational ability.

[0064] Table 1. Descriptive statistical analysis results of driving adaptability indicators Based on driving adaptability indicators, TOPSIS was used to further assess the overall driving adaptability level of older drivers, and the results are as follows: Figure 3 As shown in the figure, there are differences in the driving adaptability levels of elderly drivers, and some of the test drivers have insufficient driving adaptability.

[0065] An analysis of the characteristic indicators of dangerous driving behavior of elderly drivers was conducted, and the results are shown in [the table below]. Figure 4 Specifically, the overall distribution of PSAI, SWASD, SWAE, and PSTI values ​​is higher when turning left than when turning right. Conversely, the distributions of PSDI and jerkSD are generally lower when turning left than when turning right. In pedestrian conflict scenarios, the means and distributions of PSAI, PSDI, jerkSD, SWASD, and SWAE are lower than in car and motorcycle conflict scenarios, and are more concentrated. The t-test results show significant differences in jerkSD and SWAE between right-turn and left-turn scenarios across all three types of conflict objects. This suggests that left-turn scenarios place higher demands on the perception and operation of older drivers, leading to reduced steering smoothness and stability. Furthermore, the lower variability of these indicators in pedestrian conflict scenarios implies that older drivers tend to adopt more cautious and controlled driving behaviors when interacting with vulnerable road users, possibly due to enhanced risk perception and compensatory driving strategies.

[0066] 2) Assessment of driving workload under different conflict scenarios: Further analysis of the subjective driving workload of elderly drivers (see) Figure 5The t-test showed significant differences in Mental Demands, Physical Demands, Own Performance, and Frustration reported by the test drivers in the three conflict scenarios involving cars, motorcycles, and pedestrians. Furthermore, the Temporal Demands and Effort reported by drivers in motorcycle and pedestrian conflicts differed significantly from those in car conflicts. Notably, among the three conflict participants, older drivers gave the highest self-performance ratings in car conflicts. Additionally, there was a significant difference in self-reported Effort when turning left and right, with Effort being significantly higher when turning left than when turning right.

[0067] The subjective driving workload assessment for older drivers combines six dimensions of driving workload, including Mental Demand, Physical Demand, Temporal Demand, Own Performance, Effort, and Frustration. The DEA assessment results are as follows: Figure 6 As shown, workload was comprehensively assessed based on Scale Efficiency and Technical Efficiency. Based on the scoring data, elderly drivers were divided into three groups—low, medium, and high—using 0.35 and 0.55 as subjective workload thresholds. The elderly drivers with low driving workload accounted for 40.4% of the total sample, while those with medium and high driving workloads accounted for 29.8%.

[0068] 3) Analysis of the impact of driving workload on dangerous driving behavior based on SVM-SHAP This study analyzed risk driving behavior indicators of elderly drivers under different subjective workload conditions, based on a classification of subjective workload levels (low, medium, and high), including longitudinal and lateral indicators (see...). Figure 7 First, all indicator data were normalized, and the average values ​​were calculated for comparison. The results showed that PSAI (Pressure Stability Assessment) increased significantly with increasing subjective driving workload, while PSDI (Pressure Stability Indicator), jerkSD (Jerk Slowness Detection), SWASD (Short Slowness Detection), and SWAE (Short Slowness Effect Detection) decreased. Analysis of variance (ANOVA) and effect size analysis were used to examine the differences in risk driving behavior characteristics of elderly drivers under different subjective workload levels. The results showed that jerkSD (F = 2.812, p = 0.028) was significantly lower in different workload groups. ≤0.05) and SWAE (F= 5.949, p = 0.005) There is a significant difference (≤0.05).

[0069] Based on the selected indicators representing risky driving behavior, the VIKOR method was used to comprehensively evaluate the risky driving behavior of elderly drivers. Then, the risky driving behavior scores at different levels of subjective workload were compared (see...). Figure 8 As shown in the figure, with the increase of subjective driving workload, the mean of the group utility measure (S) increases, while the individual regret measure (R) and the average risk score of driving behavior first increase and then decrease.

[0070] Based on the VIKOR model, a driving behavior risk score was assessed, and the SVM-SHAP method was used to analyze the impact of six dimensions of subjective driving workload on risky driving behavior. Figure 9 As shown, among the six dimensions of subjective workload indicators, psychological needs rank first, indicating that psychological needs are the biggest factor influencing the risky driving behavior of older drivers. In contrast, effort and frustration have a relatively smaller impact on risky driving behavior.

[0071] To further analyze the impact of various dimensions of subjective driving workload on risky driving behavior, the SHAP scatter plot generated by the algorithm (see...) Figure 10 The results showed that the relationship between self-reported subjective workload (SHAP) and the model-predicted driving risk was not strictly monotonic. Specifically, the SHAP value first decreased and then increased with increasing mental, physical, and effort demands. This suggests that when older drivers experience low to moderate levels of cognitive and physical workload, they may tend to adopt more cautious and conservative driving behaviors, resulting in lower predicted risk. However, when the workload further increases to a higher level, their ability to cope with driving tasks may decline, increasing the likelihood of risky driving behaviors and exhibiting a low-to-high nonlinear risk pattern. Furthermore, for time demands, self-performance, and frustration, the SHAP value first increased and then decreased with increasing eigenvalues. This suggests that moderate levels of perceived time pressure, decreased self-assessed performance, or negative emotional states may be associated with a higher likelihood of risk-driven behavior. However, when these subjective perceptions intensify to extreme levels, older drivers may adopt more cautious and defensive strategies, thereby reducing actual driving risk. This high-to-low non-monotonic pattern suggests that older drivers may have self-regulation mechanisms under certain subjective states to cope with complex driving environments.

[0072] 3) Analysis of the impact of driving workload on dangerous driving behavior based on GLMM To construct a GLMM model to examine the impact of subjective driving workload, driving adaptability, and demographic characteristics on dangerous driving behavior, multicollinearity and correlation among the independent variables were first analyzed. The results are as follows: Figure 11 and 12As shown in the figure. A VIF threshold > 10 and a correlation coefficient > 0.8 were used to exclude variables such as physical exertion, time demands, personal performance, frustration, and age. This preprocessing ensures the robustness and interpretability of the subsequent GLMM analysis, enabling a more accurate estimation of the relationship between key predictors and the outcomes of dangerous driving behaviors.

[0073] Further, dangerous driving behavior scores were set as the dependent variable, Mental Demands and Effort as independent variables, and driving experience, driving mileage, and driving adaptability scores as control variables for GLMM analysis. The control variables were categorical variables of type 4. Maximum likelihood estimation (MLE) was used to estimate the model, and the results are shown in Table 2. The coefficient of Driving_Mileage4 was -1.044, showing a significant negative impact (P=0.044<0.05), indicating that, all other things being equal, compared to the driving mileage reference group, the overall dangerous driving behavior score of this group of drivers was significantly lower, with an average score 1.044 units lower than the reference group, indicating that drivers in this mileage range engaged in less dangerous driving behavior. The Driving_Fitness_Score2 coefficient was 1.736, showing a significant positive effect (P=0.001<0.01), indicating that compared to the control group, this group had a significantly lower overall score for dangerous driving behavior, averaging 1.736 units lower than the control group. This coefficient had the largest impact among all significant variables, suggesting that drivers in this group engaged in significantly more dangerous driving behaviors than the control group. The Intercept coefficient was 1.426, showing a marginal positive effect (P=0.071<0.1), indicating that when all independent variables were at the control group level, the baseline value for the overall score for dangerous driving behavior was 1.426, suggesting a certain baseline level of dangerous driving behavior in the study sample. The Mental_Demands coefficient was 9.355, showing a marginal positive effect (P=0.0653<0.1), and had the largest absolute value, indicating that increased psychological demand significantly increased the overall score for dangerous driving behavior; for every 1 unit increase in psychological demand, the score increased by an average of 9.355 units.

[0074] Table A2 GLMM Analysis Results SE = Standard Error; LR Test-LRS = Likelihood Ratio Statistic; 0.001, 0.01, : 0.05, ^: 0.1. Finally, it should be noted that in this document, relationships such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations 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.

[0075] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for analyzing the impact of subjective workload on the risk of driving behavior, characterized in that, Includes the following steps: Acquire multi-dimensional basic data of the target to be tested. The multi-dimensional basic data includes driver basic data, raw driving behavior data, subjective workload evaluation data, and driving adaptability data. The raw driving behavior data is collected under six preset traffic conflict scenarios. Based on the original driving behavior data, a risk driving behavior index system is constructed. The risk driving behavior index is normalized and weighted by a multi-criteria comprehensive evaluation method to obtain a comprehensive driving behavior risk score that represents the risk level of the target under various traffic conflict scenarios. A driving adaptability evaluation decision matrix is ​​constructed based on driving adaptability data, and the comprehensive driving adaptability level is obtained by analysis. The subjective workload evaluation index is processed, and the subjective workload is divided into three levels: low, medium, and high according to the comprehensive efficiency value. A collaborative analysis framework including a generalized linear mixed model submodule and a support vector machine submodule is constructed to solve the SHAP value of each subjective workload evaluation dimension to quantify the feature contribution. The generalized linear mixed model submodule takes the comprehensive score of driving behavior risk as the target variable and the subjective workload evaluation dimension, driving adaptability level and driver basic data as input features. The output includes the analytical results of the target under test, including the quantification coefficient of the influence of subjective workload on driving behavior risk, the ranking of key influence dimensions based on SHAP mean, and the nonlinear influence law of each dimension on driving risk.

2. The method for analyzing the impact of subjective workload on driving behavior risk according to claim 1, characterized in that, The generalized linear mixture model submodule is a beta generalized linear mixture model based on the logit link function. The driver ID of the target to be tested is used as the random effect, the comprehensive driving behavior risk score is used as the dependent variable, and the subjective workload evaluation dimension is used as the independent variable. The driver's basic data and the comprehensive driving adaptability level are used as control variables. The maximum likelihood estimation method is used to estimate the model parameters.

3. The method for analyzing the impact of subjective workload on driving behavior risk according to claim 1, characterized in that, The Support Vector Machine submodule and SHAP value analysis method construct a linear proxy model through the kernel SHAP framework, including: Kernel function support vector machine is used to model the nonlinear relationship between subjective workload evaluation dimensions, driving adaptability level, driver basic data and comprehensive driving behavior risk score; Based on the kernel SHAP method, a local linear surrogate model is constructed in the input feature space to approximate the output of the support vector machine; Based on the aforementioned proxy model, the SHAP value corresponding to each subjective workload evaluation dimension is calculated, and the average contribution of each dimension is obtained by averaging the SHAP values ​​of all samples. The nonlinear influence between various subjective workload dimensions and driving behavior risk was analyzed by using the scatter distribution of SHAP values ​​as a function of feature values.

4. The method for analyzing the impact of subjective workload on driving behavior risk according to claim 1, characterized in that, The risk driving behavior indicator system includes the percentage of rapid accelerations, the percentage of rapid decelerations, the standard deviation of acceleration, the standard deviation of steering wheel angle, the steering wheel angle entropy, and the percentage of sharp turns.

5. The method for analyzing the impact of subjective workload on driving behavior risk according to claim 1, characterized in that, The subjective workload evaluation data includes at least six dimensions: psychological needs, physical needs, time needs, personal performance, effort level, and sense of frustration.

6. The method for analyzing the impact of subjective workload on driving behavior risk according to claim 1, characterized in that, The normalization and weighting of the risky driving behavior indicators using a multi-criteria comprehensive evaluation method includes: The TOPSIS-entropy weight method was used to analyze the driving adaptability data to obtain the comprehensive driving adaptability level. The subjective workload evaluation dimensions were processed using the data envelopment analysis method, and classified into three levels: low, medium, and high. The risk driving behavior indicators are comprehensively scored using a multi-criteria compromise solution ranking method to obtain a comprehensive score of driving behavior risk.

7. The method for analyzing the impact of subjective workload on driving behavior risk according to claim 6, characterized in that, The analysis of the driving adaptability data using the TOPSIS-entropy weight method includes the following steps: A driving adaptability decision matrix is ​​constructed with night vision, dynamic vision, deep vision, selective response, and operational ability as attributes. The decision matrix is ​​then standardized, and the weight of each attribute is calculated based on its information entropy. Determine the ideal solution and the negative ideal solution, and calculate the distance between the ideal solution and the negative ideal solution, where the ideal solution and the negative ideal solution are the best and worst performance values ​​of each attribute among all alternative solutions; Calculate a comprehensive score, which is a score based on the relative proximity of each alternative solution to the ideal solution and the negative ideal solution.

Images