Driving risk analysis method and device, computer equipment and storage medium

Abstract

The invention provides a driving risk analysis method and device, computer equipment and a storage medium, and belongs to the field of driving risk analys.The method comprises the steps that a dynamic clustering method is adopted to divide a driver watching area into a vehicle front area, a vehicle left area, a vehicle right area and an in-vehicle instrument panel area; extracting eye movement behavior characteristics of each region, and determining a first change rule of the eye movement behavior characteristics in different fatigue states of each region; vehicle driving parameters are obtained, and a second change rule of the vehicle driving parameters in different fatigue states is determined; and constructing a Bayesian network, taking the eye movement behavior characteristics and the vehicle driving parameters as input variables, performing driving safety risk analysis according to the first change rule and the second change rule, and determining a risk assessment result. Therefore, according to the method, through multi-modal data fusion, dynamic region division and causal reasoning modeling, the accuracy, real-time performance and scene adaptability of driving risk assessment are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of driving risk analysis, and specifically relates to a driving risk analysis method, device, computer equipment and storage medium. Background Art

[0002] Fatigue driving plays a significant role in traffic accidents, and the changes in driving behavior and physiological characteristics caused by fatigue have been extensively studied both domestically and internationally. However, compared to other countries, quantitative assessments of fatigue driving risks are relatively scarce in China. Furthermore, the impact of fatigue driving is a cumulative process, and its potential dangers gradually increase over time. It is necessary to clarify which driving risk quantitative indicators can describe how driving risk changes over time. In actual driving, when a driver becomes fatigued, not only will they exhibit physiological and psychological changes, but it will also affect the "perception-judgment-decision-making-vehicle operation" process. Therefore, a fatigue driving risk assessment level that considers multiple risk factors is of great significance. By deeply exploring the mechanism by which fatigue affects driver behavior and accident risk, it will help to more effectively prevent and manage the potential threats posed by fatigue driving.

[0003] Every year, road traffic accidents pose a serious threat to human life and property. Research indicates that drivers are directly responsible for 70% to 90% of all accidents. Therefore, effectively controlling unsafe driving behavior, reducing the probability of accidents, minimizing traffic accidents, and achieving safe driving are crucial tasks for further implementing the primary responsibility of road transport companies for production safety and are a key regulatory priority for transportation management departments.

[0004] Existing technologies primarily rely on vehicle driving parameters as the primary data for risk analysis, such as speed, acceleration, steering wheel angle, and lateral offset. However, these parameters can lead to significant errors. For example, if the driver is distracted but the vehicle continues to move straight, sudden changes in driving parameters on densely populated roads can often be delayed, potentially leading to accidents. Summary of the Invention

[0005] In order to solve the above-mentioned problem of large errors in judgment based on vehicle driving parameters, the present invention provides a driving risk analysis method, device, computer equipment and storage medium.

[0006] In order to achieve the above object, the present invention provides the following technical solutions: A driving risk analysis method, comprising: The dynamic clustering method is used to divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side and the instrument panel inside the vehicle; Extracting eye movement behavior characteristics of each region and determining a first change pattern of the eye movement behavior characteristics under different fatigue states of each region; Acquire vehicle driving parameters and determine a second variation law of the vehicle driving parameters under different fatigue states; A Bayesian network is constructed, eye movement behavior characteristics and vehicle driving parameters are used as input variables, and driving safety risk analysis is performed according to the first change law and the second change law to determine the risk assessment result.

[0007] Optionally, the method of dividing the driver's gaze area into four areas, namely, the front of the vehicle, the left side, the right side, and the dashboard, by using a dynamic clustering method includes: The eye-tracking device is used to collect the driver's eye movement data during the simulated driving process, including the location of the gaze point and the duration of the gaze; ‌Preprocess the collected eye movement data to remove outliers and noise; The dynamic clustering algorithm is used to perform cluster analysis on the preprocessed eye movement data. The number of clusters is set to 4. Through iterative calculation, the driver's gaze area is divided into four areas: the front, left, right and dashboard of the vehicle.

[0008] Optionally, extracting eye movement behavior features based on the divided gaze areas and determining a first change pattern of the eye movement behavior features under different fatigue states in each area includes: Extracting eye movement behavior features from the divided gaze areas, including blink frequency, gaze duration, saccade duration, and pupil area. The driver's fatigue state is divided into three levels: alert, tired and very tired; For each gaze area, the first change pattern of blink frequency, gaze duration, saccade duration and pupil area under different fatigue states is analyzed, and the first change pattern represents the relationship between the eye movement behavior characteristics and the fatigue state.

[0009] Optionally, acquiring the vehicle driving parameter and determining a second variation rule of the vehicle driving parameter under different fatigue states includes: Collecting driving parameters of the driver under different fatigue states, including steering wheel angle, vehicle speed and vehicle lateral offset; Preprocess and analyze the collected vehicle driving parameters to calculate the mean and standard deviation of the steering wheel angle, the mean and standard deviation of the vehicle speed, and the mean and standard deviation of the vehicle lateral offset; Through statistical analysis and comparison, a second variation law of the steering wheel angle, vehicle speed and vehicle lateral offset under different fatigue states is determined, and the second variation law represents the influence of the fatigue state on the vehicle driving parameters.

[0010] Optionally, constructing a Bayesian network, taking eye movement behavior characteristics and vehicle driving parameters as input variables, and constructing a driving safety risk assessment model according to the first change rule and the second change rule includes: A Bayesian network is constructed, and the eye movement behavior characteristics of the four gaze areas and the vehicle driving parameters are used as observation nodes. The intermediate node of the fatigue state is determined based on the first change law, and the global node of the fatigue state is determined based on the second change law. The observation node of the eye movement characteristics is connected to the intermediate fatigue state node, and the observation node of the vehicle driving parameters is connected to the global fatigue state node. All state nodes are integrated for reasoning based on the preset weight distribution rules to output the risk assessment results.

[0011] A driving risk analysis device, comprising: A division module is used to divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side, and the instrument panel inside the vehicle using a dynamic clustering method; A determination module is configured to extract eye movement behavior characteristics of each region and determine a first change pattern of the eye movement behavior characteristics under different fatigue states of each region; obtain vehicle driving parameters and determine a second change pattern of the vehicle driving parameters under different fatigue states; A construction module is used to construct a Bayesian network, take eye movement behavior characteristics and vehicle driving parameters as input variables, perform driving safety risk analysis based on the first change law and the second change law, and determine the risk assessment result.

[0012] A computer-readable storage medium stores a computer program, which implements the above-mentioned driving risk analysis method when executed by a processor.

[0013] A computer device, characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned driving risk analysis method when executing the program.

[0014] The driving risk analysis method provided by the present invention has the following beneficial effects: First, the method extracts eye movement characteristics and vehicle driving parameters. Eye movement characteristics directly reflect the driver's physiological state, while vehicle driving parameters indirectly reflect the stability of driving behavior. This multimodal data approach overcomes the limitations of relying solely on vehicle driving parameters, simultaneously capturing both physiological abnormalities and the risk of behavioral loss, avoiding misjudgment based on a single indicator. Secondly, the method uses dynamic clustering to segment the driver's gaze area and conducts regional differentiation analysis to identify correlations between multimodal data such as fatigue, eye movement characteristics, and gaze area, further improving the accuracy of driving risk analysis. Finally, a Bayesian network is used to implement causal reasoning, determining joint probabilities based on multiple change patterns, further enhancing the accuracy of risk analysis. Thus, through multimodal data fusion, dynamic region segmentation, and causal reasoning modeling, the method significantly improves the accuracy, real-time nature, and scenario adaptability of driving risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0016] Figure 1 The figure is a flow chart of a driving risk analysis method provided by the present invention according to an exemplary embodiment.

[0017] Figure 2 A flowchart of a Bayesian algorithm is provided according to an exemplary embodiment of the present invention.

[0018] Figure 3 According to an exemplary embodiment of the present invention, a Figure 4 The present invention provides a block diagram of a driving risk analysis device according to an exemplary embodiment. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0021] First, the present invention provides a driving risk analysis method, specifically Figure 1 As shown, the following steps are included: S101. Divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side, and the instrument panel inside the vehicle using a dynamic clustering method.

[0022] In this step, the driver's eye movement data during the simulated driving process is collected using an eye tracker, including the location of the gaze point and the duration of gaze. The collected eye movement data is preprocessed to remove outliers and noise. The preprocessed eye movement data is clustered using a dynamic clustering algorithm. The number of clusters is set to 4, and through iterative calculations, the driver's gaze area is divided into four areas: the front of the vehicle, the left side, the right side, and the dashboard inside the vehicle.

[0023] While driving, the driver needs to rely on vision to capture information about the surrounding road and adjust the vehicle's posture. At this time, the driver's gaze will linger between different targets. These targets are divided according to certain patterns, and the divided areas are the driver's different gaze areas during driving. Three methods for dividing regions of interest are summarized:

[0024] (1) Visual field division method The visual field plane segmentation method divides the driver's field of view into multiple areas based on the vehicle's structure and the layout of the road ahead. Focusing on a particular area indicates attention to the target in that area. This method is applicable to visual field segmentation in static scenes. Based on the vehicle's front windshield, the driver's visual field plane is divided into nine areas of interest: left near view, front near view, right near view, left mid view, front mid view, right mid view, left far view, front far view, and right far view.

[0025] The advantage of the visual field segmentation method is its simplicity and ability to efficiently measure gaze activity in each region. However, this method only applies to static driver gaze behavior and assumes that the driver's head remains directly forward. However, in real-world driving, drivers' visual movements are constantly changing, so this method's accuracy and reliability are limited.

[0026] (2) Fixation point counting method The gaze point counting method involves replaying the video recorded during driving and counting the driver's gaze point frame by frame. The area where the gaze point falls represents the target area, and the driver's field of view is divided into two planes based on the statistical results. This method has accurate statistical results and does not need to consider the differences between different drivers. The error is small, so the division of the field of view plane is more accurate. However, due to the eye tracker's sampling frequency of 60HZ and the video recording speed of 20-30 frames per second, a large amount of data can be collected in a short period of time. Considering the eye movement data collected in units of hours for each experiment, this method is too labor-intensive, time-consuming, and slow.

[0027] (3) Dynamic clustering method This method clusters the coordinates of gaze points, grouping them into a single gaze area. This allows for gaze area segmentation. The advantage of clustering is that it effectively mitigates the influence of individual differences on driver gaze area segmentation, improving the accuracy of the results. Using MATLAB data analysis software to program the clustering analysis method, it can quickly process large amounts of driver gaze area data, ensuring the smooth progress of the experiment.

[0028] After a comprehensive comparison of various methods, this paper decided to use dynamic clustering to segment a driver's visual regions of interest. This method fully accounts for individual differences between drivers, is more accurate than the visual field plane segmentation method, and requires less work than the gaze point-by-point statistics method, but the calculation process is relatively complex.

[0029] For example, this step can use the K-means clustering algorithm to divide the gaze area. The algorithm is to iteratively solve the non-convex optimization problem, based on the maximum value or its derivative method, according to the preset number of categories and initial center, divide similar data points, and iteratively update the data mean until the optimal clustering result is reached.

[0030] The algorithm steps are as follows: Step 1: Randomly select K test data samples as the initial centroid.

[0031] Step 2: Use the Euclidean distance method to calculate the distance between the data point and the initial centroid: ; Where: For the Cluster centers, for arrive distance, for point Arrive The k-th dimension coordinate of .

[0032] Since the coordinate axes of the collected eye movement data are two-dimensional coordinates, the formula can be further optimized: ; Step 3: After determining the distance from each data point to the initial center, calculate the similarity and then assign it to the category of the closest cluster center according to the proximity principle. The formula is as follows: ; Step 4: Group similar coordinate points into one category and update the cluster center coordinates of each category as follows: ; Where: For the The two-dimensional coordinates of the first point of the class, For the The number of data points in the class.

[0033] Step 5: Repeat Step 3 and Step 4 until the algorithm converges.

[0034] Following the principles and steps of the K-means clustering algorithm, we selected eye movement data collected from test drivers during driving, after abnormal data processing, for cluster analysis. We used the Matlab toolbox to develop the algorithm, setting the number of clusters to 4, 5, 6, and 7, and then ran the analysis for comparison.

[0035] The algorithm found that a clustering value of 4 is closer to the actual situation. When the clustered driver's gaze points are projected onto the front windshield of the actual vehicle being driven, the visual interest areas can be divided into four parts: the left side of the vehicle, the area in front of the vehicle, the interior instrument panel, and the right side of the vehicle. The meaning of each interest area number is shown in Table 1 below:

[0036] Table 1 Gaze areas corresponding to each area of ​​interest S102: Extract eye movement behavior characteristics of each region, and determine a first change pattern of the eye movement behavior characteristics under different fatigue states of each region.

[0037] In this step, eye movement behavior characteristics are extracted from the divided gaze areas, including blinking frequency, gaze duration, glance duration and pupil area; the driver's fatigue state is divided into three levels: awake, tired and very tired; for each gaze area, the first change pattern of blinking frequency, gaze duration, glance duration and pupil area under different fatigue states is analyzed, and the first change pattern characterizes the relationship between eye movement behavior characteristics and fatigue state.

[0038] Blink rate refers to the number of times a person closes and opens their eyes over a period of time. Therefore, blink rate can be used as a visual indicator of a driver's responsiveness to their environment. With increasing driving time and fatigue level, blink rate first increases and then decreases during the awake to very fatigued stages. Analysis shows that blink rates range from 17-19 blinks / minute, 20-23 blinks / minute, and 11-14 blinks / minute, respectively. While awake, blink rate fluctuates slightly but remains relatively stable within a certain range. This blinking is often due to driving habits and is unconscious. However, during fatigue, blink rate increases significantly, a voluntary response to relieve dry eyes and promote blood circulation after prolonged eye use. During the extreme fatigue stage, blink rate drops to approximately 12 blinks / minute. This is due to fatigue-induced weakening of neural activity, which in turn reduces eye function and significantly reduces eye movement, indicating a refusal to combat fatigue.

[0039] To further analyze whether there were significant differences between the two groups, one-way analysis of variance was used to conduct post hoc multiple comparisons of driving risk indicators that showed significant differences under different fatigue states. The results showed that the p value for blinking frequency between the awake-fatigue, awake-very tired, and tired-very tired groups was < 0.05, indicating that there were significant differences in the blinking frequency of drivers under different fatigue states. Therefore, this parameter can be used as a risk indicator to characterize eye movement changes in different fatigue states.

[0040] Individuals engage in fixation to acquire information about their current external environment. Fixation duration is closely related to the complexity and processing difficulty of the information. When a driver is performing a driving task, fixation duration can, to a certain extent, reflect the driver's energy level and directly impact driving safety risks. A fixation duration exceeding 100 milliseconds is considered a valid fixation. The duration of a driver's fixation on an area of ​​interest indicates the driver's attention to that area. During normal driving, fixation duration is relatively short. With increasing fatigue, fixation duration increases, primarily concentrated between 400 and 900 milliseconds. Furthermore, fixation duration increases steadily from the fatigued to very fatigued stages, from 400 to 800 milliseconds to 500 to 900 milliseconds. To investigate whether average fixation duration differs significantly across different fatigue states, a one-way analysis of variance and normality test were used. At the 0.05 level, the gaze durations across different fatigue states were significantly normal (p(awake)=0.065, p(fatigue)=0.207, p(very tired)=0.143). A one-way ANOVA showed an F=10.216, with a significance level of p=0.000<0.05, indicating significant differences in gaze duration across different fatigue states. Post hoc multiple comparisons of driving risk indicators that showed significant differences across fatigue states were performed using one-way ANOVA. The p<0.05 values ​​for gaze duration between the awake-fatigue, awake-very tired, and tired-very tired groups were significant, indicating significant differences between these two groups. This analysis suggests that this parameter can be used as a risk indicator for changes in eye movements across different fatigue states.

[0041] Furthermore, 76% of the driver's gaze is focused on the area directly in front of the vehicle (AOI 2), primarily focusing on the road ahead. The left side (AOI 1) and right side (AOI 4) account for a smaller proportion of the cumulative gaze time, while the instrument panel (AOI 3) accounts for the lowest cumulative gaze time. This is primarily due to the fact that the driver's gaze is primarily focused forward while driving, providing a basic field of view directly in front of the windshield. However, in complex driving scenarios such as turning, passing intersections, and changing lanes to overtake, the driver must pay close attention to the surrounding vehicle dynamics. In these situations, the driver's gaze will be more focused on the left or right side of the windshield to ensure driving safety.

[0042] Overall, drivers pay the most attention to the area ahead and the least to the dashboard. From alert to fatigued, drivers' gaze shifts toward the center primary visual field, further reducing their attention to the left and right sides of the vehicle. This prevents them from observing dynamic road conditions and increases driving risk. Drivers increase their visual search breadth in the area ahead, but decrease their attention to the area behind the vehicle.

[0043] Glance duration refers to the time it takes a driver's eyes to move from one point to another while observing a specific area, measured in milliseconds. Glance duration reflects a driver's attention span and ability to allocate attention. Drivers typically need to quickly and effectively scan between different areas, such as the dashboard and the road ahead, to obtain key information and make accurate driving decisions. Glance duration can be affected by fatigue. As driver fatigue increases, average glance duration decreases relative to alertness. Furthermore, average glance duration decreases with driving time. From alertness to extreme fatigue, average glance duration decreases from 47-56ms to 38-50ms. Glance duration reflects a driver's concentration and reaction speed. Short glances may indicate a driver's inattention and tendency to overlook important information, while long glances may indicate a slow reaction time and a lack of critical information.

[0044] One-way ANOVA and normality tests were used for analysis. The Shapiro-Wilk test showed that glance durations under different fatigue states were significantly normally distributed at the 0.05 level (p(awake)=0.055, p(fatigued)=0.806, p(very tired)=0.317). A one-way ANOVA showed an F=14.370, with a significance level of p=0.000<0.05, indicating significant differences in glance durations under different fatigue states.

[0045] To further analyze whether there were significant differences between the two groups, a one-way analysis of variance was used to conduct post hoc multiple comparisons of driving risk indicators that showed significant differences between the different fatigue states. The results showed that the driver's glance duration was significantly different between the awake-fatigue, awake-very fatigue, and fatigue-very fatigue groups (p < 0.05), indicating that these two groups were significantly different. This analysis suggests that this parameter can be used as a risk indicator for eye movement changes in different fatigue states.

[0046] Pupil area refers to the size of a driver's pupil as measured by an eye tracker. The eye tracker shines an infrared light onto the eye and calculates pupil area based on the reflected light. Pupil area can provide information about a driver's attention and cognitive load. A larger pupil area generally indicates increased attention or a higher cognitive load on a specific task or environment. A smaller pupil area, on the other hand, may indicate decreased attention or a lower cognitive load on a specific task or environment. This information can be used to assess driver condition and performance to improve driving training and safety.

[0047] To investigate whether average pupil area differed significantly across different fatigue states, a one-way analysis of variance and normality test were used. The Shapiro-Wilk test showed that pupil area across different fatigue states was significantly normally distributed at the 0.05 level (p(alert)=0.839, p(fatigue)=0.455, p(very fatigued)=0.148). A one-way analysis of variance showed an F=12.329, with a significance p=0.000<0.05, indicating significant differences in pupil area across different fatigue states.

[0048] To further analyze whether there were significant differences between the two groups, a one-way analysis of variance was used to conduct post hoc multiple comparisons of driving risk indicators that showed significant differences between the different fatigue states. The results showed that the average pupil area of ​​drivers was significantly different between the awake-fatigued, awake-very-fatigued, and fatigue-very-fatigued groups (p < 0.05), indicating that these two groups were significantly different. This analysis suggests that this parameter can be used as a risk indicator to characterize eye movement changes in different fatigue states.

[0049] S103: Acquire vehicle driving parameters and determine a second variation pattern of the vehicle driving parameters under different fatigue states.

[0050] In this step, the vehicle driving parameters of the driver under different fatigue states, including the steering wheel angle, vehicle speed and vehicle lateral offset, are collected through a simulated driver; the collected vehicle driving parameters are preprocessed and analyzed, and the mean and standard deviation of the steering wheel angle, vehicle speed and vehicle lateral offset are calculated respectively; through statistical analysis and comparison, the second change law of the steering wheel angle, vehicle speed and vehicle lateral offset under different fatigue states is determined, and the second change law represents the influence of the fatigue state on the vehicle driving parameters.

[0051] The steering wheel angle directly reflects the driver's ability to manipulate the vehicle's steering wheel during driving, as steering wheel rotation directly reflects the stability of the vehicle's trajectory and the likelihood of a collision. To ensure the vehicle stays on the intended trajectory and within its lane, the driver must continuously fine-tune the steering wheel based on real-time road conditions. Therefore, this step analyzes the steering wheel angle to examine its impact under fatigue conditions. The impact of the driver's driving style on the steering wheel angle is also analyzed.

[0052] When the driver is awake, they frequently adjust the steering wheel to keep the vehicle running smoothly, adjusting it based on changing road conditions, and the adjustments are small. When fatigued, the driver's steering adjustments become less frequent, and the steering wheel may occasionally remain stationary for extended periods. When the driver is extremely fatigued, these signs of fatigue become more pronounced. Due to reduced hand coordination and precision, the driver's ability to control the steering wheel decreases, resulting in prolonged periods of inactivity. Then, suddenly awakened, the driver may make emergency steering adjustments, resulting in a sudden increase in the magnitude of the adjustments, making the vehicle's turning and steering movements unstable and increasing the risk of traffic accidents.

[0053] To further analyze the changes in steering wheel angle characteristics under different fatigue states, this step conducts a statistical analysis of steering wheel angle-related indicators. International research indicates that in a monotonous driving environment with minimal information stimulation, the average amplitude of the steering wheel, the maximum steering wheel angle, and the standard deviation increase over time while the driver is driving in a straight line. Therefore, the absolute mean (SWA_mean) of the steering wheel angle and the standard deviation (SWA_std) of the steering wheel angle are used as characteristic parameters of driving behavior risk.

[0054] SWA_mean and SWA_std can indirectly reflect the driver's stability in steering wheel operation. In the present invention, SWA_mean is the average value of the absolute value of the steering wheel angle. Only its size is considered during calculation without considering the direction. The calculation formula is as follows.

[0055] ; ; Where, is the number of sampling points in the sample; is the steering wheel angle at the i-th moment; is the mean of the actual steering wheel angle.

[0056] Driving speed is a key indicator of driver fatigue. The driver's ability to control the accelerator and brake pedals directly impacts their ability to adjust and maintain vehicle speed. Speed ​​objectively reflects the driver's mental state and reaction time. It is influenced by factors such as traffic conditions, driving style, and driving conditions.

[0057] Vehicle speed is a commonly used metric for evaluating a driver's control over their vehicle. Driving conditions and traffic conditions significantly influence the driver's speed choices. Speed ​​is clearly correlated with the potential for a collision and its severity. As vehicle speed increases, the likelihood of a traffic accident increases significantly, and the energy released by collisions with obstacles also increases, significantly increasing driving risk. Studies have shown a close correlation between average speed and traffic accident rates; as average speed increases, the likelihood of a traffic accident also increases. The standard deviation of speed, which reflects the discrete distribution of speed, is positively correlated with the occurrence of traffic accidents. To further explore the characteristics of speed variation under different fatigue states, the mean and standard deviation of speed were analyzed. The mean speed (V_mean) and the standard deviation (V_std) were used as characteristic parameters to characterize fatigue driving, using the following calculation formula.

[0058] ; ; Where, is the number of sampling points in the sample; is the velocity at the i-th moment.

[0059] From alert to very fatigued, average speed and speed standard deviation increased with driving time and fatigue level. In the early stages of the test, drivers demonstrated strong control over the vehicle, maintaining speed within a certain range, resulting in a more concentrated data distribution. Furthermore, a smaller speed standard deviation indicated smoother driving. In the later stages of the test, due to reduced driving behavior, the vehicle sometimes maintained a consistent speed and sometimes accelerated or decelerated, but overall the trend was upward. The data distribution became more dispersed, with a wider range of fluctuations and a larger speed standard deviation.

[0060] To investigate whether the mean and standard deviation of the average vehicle speed time series data differed significantly across different fatigue states, one-way analysis of variance and normality tests were used. The Shapiro-Wilk test showed that V_mean was significantly normalized across all fatigue states at the 0.05 level (p(alert)=0.066, p(fatigue)=0.378, p(very fatigue)=0.119). A one-way analysis of variance showed an F=8.214, with a significance level of p=0.000<0.05, indicating significant differences in vehicle speed across different fatigue states. The Shapiro-Wilk test showed that V_std was significantly normalized across all fatigue states at the 0.05 level (p(alert)=0.243, p(fatigue)=0.592, p(very fatigue)=0.111). The one-way ANOVA showed F=5.882, with a significance of p=0.000<0.05, indicating that the standard deviation of vehicle speeds was significantly different under different driver fatigue states. In summary, the vehicle speed characteristic index can be used as an indicator to characterize driver fatigue state.

[0061] To further analyze whether there were significant differences between the two groups, one-way analysis of variance was used to conduct post hoc multiple comparisons of driving risk indicators that showed significant differences under different fatigue states. The p value of V_std between the awake-fatigue, awake-very tired, and tired-very tired groups was < 0.05, indicating that the standard deviation of vehicle speed was significantly different under different fatigue states and could be used as a risk indicator to distinguish driving behavior in different fatigue states.

[0062] A vehicle's lateral offset refers to the degree to which it deviates sideways from the road centerline. It is a key indicator of vehicle stability and handling, and is crucial for accurate vehicle control and safe driving. Variations in lateral offset can be affected by a variety of factors, including road conditions, driving style, driving time, and driver input. Accurately assessing and controlling lateral offset can improve vehicle stability and handling, thereby enhancing driving safety and comfort.

[0063] A vehicle's lateral offset is one of the key indicators of a driver's lane-keeping ability and is the most direct indicator of the vehicle's operating status. Furthermore, during extended driving, it can be difficult for a vehicle to remain centered in its lane. Lateral offset can be positive or negative, with a positive value indicating a vehicle's deviation to the right of the road and a negative value indicating a deviation to the left.

[0064] Intuitive analysis shows a positive correlation between driver fatigue and the volatility of the vehicle's lateral offset. Different levels of fatigue exhibit significant differences in lateral position. The absolute mean (LP_mean) and standard deviation (LP_std) of the vehicle's lateral offset are used as characteristic parameters to characterize fatigue driving. LP_mean is the average of the absolute values ​​of the lateral offset, considering only its magnitude, not its direction. The calculation formula is as follows.

[0065] (4-7) (4-8) Where, is the number of sampling points in the sample; is the steering wheel angular velocity at the i-th moment; is the actual mean value of the vehicle's lateral offset.

[0066] The mean and standard deviation of vehicle lateral excursion are positively correlated with driving time and driver fatigue. They are relatively stable in the initial stages of the test, due to drivers' relatively cautious and vigilant driving. Consequently, lateral excursion is relatively stable and small. Drivers are better able to control the vehicle, maintain the center of the lane, and have a lower probability of lane deviation. Furthermore, a smaller standard deviation of lateral excursion indicates smoother driving. In the later stages of the test, drivers are more likely to fall asleep, which can lead to unstable steering, vehicle swaying from side to side along the lane centerline, and increased lateral excursion. Some drivers have even left their lane during the test. As the test progresses, the fluctuation range and standard deviation of lateral excursion increase, indicating increased driving risk. In summary, the mean and standard deviation of lateral excursion reflect lane position fluctuations during driving and demonstrate the driver's ability to control the vehicle's lateral motion. The standard deviation of lateral excursion increases when the driver is fatigued.

[0067] To investigate whether the mean and standard deviation of the time-series data for vehicle lateral offset differ significantly across different fatigue states, one-way analysis of variance and normality tests were used. The Shapiro-Wilk test showed that LP_mean was significantly normalized across all fatigue states at the 0.05 level (p(awake)=0.099, p(fatigue)=0.155, p(very tired)=0.059). A one-way analysis of variance showed an F=18.135, with a significance level of p=0.000<0.05, indicating significant differences in vehicle lateral offset across different fatigue states. The Shapiro-Wilk test showed that LP_std was significantly normalized across all fatigue states at the 0.05 level (p(awake)=0.099, p(fatigue)=0.109, p(very tired)=0.262). A one-way ANOVA showed F=32.572, with a significance of p=0.000<0.05, indicating that the standard deviation of vehicle lateral offset was significantly different under different driver fatigue states. In summary, the vehicle lateral offset characteristic index can be used as a representation indicator of driver fatigue state.

[0068] To further analyze whether there were significant differences between the two groups, a one-way analysis of variance was used to conduct post hoc multiple comparisons of driving risk indicators that showed significant differences between different fatigue states. The results showed that LP_mean had a p value < 0.05 between the awake-fatigue and awake-very-fatigue groups, while p = 0.098 > 0.05 between the fatigue-very-fatigue groups. This indicates that there were no significant differences between the two groups.

[0069] The present invention also considers the impact of driving style on vehicle stability. Taking the impact of driving style on steering wheel angle as an example, one-way analysis of variance and normality tests were used to examine whether driving style significantly affects steering wheel angle. Normality tests were first performed on all data. The Shapiro-Wilk test showed that at the 0.05 level, cautious drivers with different fatigue states all showed a significant normal distribution (p(alert) = 0.092, p(fatigue) = 0.171, p(very fatigue) = 0.158). Steady drivers with different fatigue states all showed a significant normal distribution (p(alert) = 0.186, p(fatigue) = 0.191, p(very fatigue) = 0.232). Aggressive drivers with different fatigue states all showed a significant normal distribution (p(alert) = 0.061, p(fatigue) = 0.247, p(very fatigue) = 0.087).

[0070] As fatigue increases, steering wheel angle also increases, a conclusion consistent with the above findings. A one-way ANOVA analysis of steering wheel angles for cautious, steady, and aggressive drivers across different fatigue states showed significant p values ​​less than 0.05. This indicates that steering wheel angle changes for cautious, steady, and aggressive drivers all increase significantly with increasing fatigue, suggesting that drivers with different driving styles exhibit certain differences in their steering wheel handling. While awake, driving style does not significantly affect steering wheel angle, but the differences in steering wheel angles between cautious and steady drivers, steady and aggressive drivers, and aggressive and cautious drivers gradually increase. In the fatigue and very fatigue states, the differences in steering wheel angles between cautious and aggressive drivers are significant, while the differences between cautious and steady drivers, or between steady and aggressive drivers, are not significant.

[0071] In addition, driving style also affects vehicle speed and lateral offset, such as the impact on steering wheel angle mentioned above, which will not be repeated here.

[0072] S104: Construct a Bayesian network, use eye movement behavior characteristics and vehicle driving parameters as input variables, perform driving safety risk analysis based on the first change rule and the second change rule, and determine a risk assessment result.

[0073] In this step, the node variables of the Bayesian network are determined. These node variables include input nodes and output nodes. The input nodes include blink frequency, fixation duration, saccade duration, pupil area, steering wheel angle mean, steering wheel angle standard deviation, vehicle speed standard deviation, and vehicle lateral offset standard deviation. The output node includes the driving safety risk level. Specifically, a Bayesian network is constructed, using the eye movement behavior characteristics of the four fixation areas and vehicle driving parameters as observation nodes. Based on the first variation rule, the intermediate fatigue state node is determined, and based on the second variation rule, the global fatigue state node is determined. The observation node of the eye movement characteristics is connected to the intermediate fatigue state node, and the observation node of the vehicle driving parameters is connected to the global fatigue state node. All state nodes are integrated and reasoned based on a preset weight distribution rule to output the risk assessment result. The parameters of the Bayesian network are learned using maximum likelihood estimation or Bayesian estimation. The model is validated through cross-validation and optimized and adjusted based on the validation results.

[0074] This step utilizes the professional Bayesian analysis software Netica, introducing the K2 algorithm, to construct a fatigue driving risk assessment model. Once the model is constructed, node parameter learning is required. The core learning task of a Bayesian network is determining the network node parameters, referred to as parameter learning. Because the present invention acquires and collects a complete dataset through a simulated driving platform, it can combine maximum likelihood estimation and Bayesian estimation to achieve parameter learning.

[0075] The construction of a Bayesian network mainly includes three steps. First, it is necessary to clarify the purpose of building the network and determine the relevant nodes and their values ​​based on the purpose and the actual data. Next, based on the research purpose and node conditions, fully combine expert knowledge and data information to formulate the Bayesian network structure and perform parameter learning to form a Bayesian network model. Finally, the effectiveness of the formed model is verified and the model construction is completed after the test. Based on the relatively complete data set of this paper, the process of constructing a Bayesian network is as follows: Figure 2 shown.

[0076] Before establishing a fatigue driving risk assessment model, the input nodes, implicit nodes, and output nodes of the Bayesian network are first determined. Based on the aforementioned one-way analysis of variance, the various indicators under different fatigue states are statistically analyzed, and the optimal feature subset is extracted. The driver's eye movement characteristics include blink frequency, gaze duration, glance duration, and pupil area. The driving behavior characteristic indicators include steering wheel angle, vehicle speed, and vehicle lateral offset. SPSS data analysis software is used to divide the indicators of different fatigue states into three states: low risk, medium risk, and high risk, as shown in Table 2 below:

[0077] Table 2 Classification of fatigue driving risk indicators like Figure 3 As shown in Figure 2, except for observation nodes, both eye movement risk representation and driving behavior risk have only two states: risk and non-risk, represented by yes and no, respectively. Fatigue driving risk is divided into three states.

[0078] Eye movement representation risk (ER) = {yes, no}; Driving behavior risk (DR) = {yes, no}; Fatigue driving risk (FR) = {Soberness, Fatigue, Severe fatigue}.

[0079] Netica software provides graphical modeling tools and visualization capabilities for probabilistic networks, such as Bayesian networks, decision networks, continuous networks, and hybrid networks. It supports a variety of node types and attributes, accurately capturing different types of uncertainty and relationships. Furthermore, it features an efficient inference engine, enabling precise probabilistic reasoning and conditional probability queries on probabilistic models. It supports multiple inference algorithms, including forward, backward, and sequential reasoning, enabling rapid probabilistic results. Therefore, Netica software was selected to develop a Bayesian network model for driver risk assessment under different fatigue conditions.

[0080] In this step, the parameters of the Bayesian network can be determined using Bayesian estimation. Bayesian estimation first summarizes prior knowledge of the parameters using probability distributions, then uses the likelihood function to extract deeper meaning from the data. Finally, the Bayesian estimation formula cleverly combines the prior distribution and the likelihood function to derive the posterior distribution of the parameters.

[0081] Bayesian parameter learning iteratively uses posterior probabilities as new prior probabilities. Building on prior probabilities, Bayesian methods can fully utilize sample data, reducing reliance on datasets and sensitivity to data size. Bayesian estimation considers both prior and posterior probabilities, avoiding the subjectivity of relying solely on prior probabilities while reducing the impact of dataset noise, effectively addressing the shortcomings of maximum likelihood parameter learning algorithms.

[0082] Furthermore, the driving risk data for different fatigue states used in this invention is fully recorded by the driver simulator, with no gaps in information. Therefore, the acquired sample data set is imported into Netica software, where parameter learning methods are used to learn the required complete conditional probability table, thereby forming the Bayesian network parameters. Once parameter learning is complete, if its reliability can be verified through validity testing, the Bayesian network structure derived through parameter learning can be used to assess the risk level of fatigue driving and conduct inference analysis on the relationships between node variables.

[0083] The Bayesian network model enables a quantitative assessment of the fuzzy concept of driving risk caused by fatigue driving. In the experimental dataset, the probability of fatigue driving risk being in a very fatigued state was 39.1%, the probability of being in a fatigued state was 31.3%, and the probability of being in an alert state was 29.6%. The parameter learning results indicate that both eye movement characteristics and driving behavior characteristics are high risk factors, with probabilities of occurrence of 56.3% and 57.3%, respectively. Among visual characteristics, gaze duration has a high probability of being in a high-risk state of 40.3%, followed by blink frequency at 35.5%. Among driving behaviors, the probability of the vehicle's average steering wheel angle being in a high-risk state was 50.0%. The results demonstrate that both the risk level and the level of fatigue driving risk are quantitatively reflected.

[0084] Using this method, eye movement characteristics and vehicle driving parameters are first extracted. Eye movement characteristics directly reflect the driver's physiological state, while vehicle driving parameters indirectly reflect the stability of driving behavior. This multimodal data approach overcomes the limitations of relying solely on vehicle driving parameters, simultaneously capturing both physiological abnormalities and the risk of behavioral loss, avoiding misjudgment based on a single indicator. Secondly, dynamic clustering is used to segment the driver's gaze area, and regional differentiation analysis is performed to determine the correlation between multimodal data such as fatigue, eye movement characteristics, and gaze area, helping to further improve the accuracy of driving risk analysis. Finally, causal reasoning is implemented using a Bayesian network, determining joint probabilities based on multiple change patterns, further enhancing the accuracy of risk analysis. Thus, through multimodal data fusion, dynamic region segmentation, and causal reasoning modeling, this method significantly improves the accuracy, real-time nature, and scenario adaptability of driving risk assessment.

[0085] Secondly, the present invention also provides a driving risk analysis device, such as Figure 4 As shown, including: The division module 401 is used to divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side, and the instrument panel inside the vehicle using a dynamic clustering method.

[0086] Determination module 402 is used to extract the eye movement behavior characteristics of each area and determine the first change law of the eye movement behavior characteristics under different fatigue states in each area; obtain vehicle driving parameters and determine the second change law of the vehicle driving parameters under different fatigue states.

[0087] The construction module 403 is used to construct a Bayesian network, take the eye movement behavior characteristics and vehicle driving parameters as input variables, perform driving safety risk analysis according to the first change rule and the second change rule, and determine the risk assessment result.

[0088] Using the above device, eye movement behavior characteristics and vehicle driving parameters are first extracted. Eye movement behavior characteristics directly reflect the driver's physiological state, while vehicle driving parameters indirectly reflect the stability of driving behavior. This multimodal data application overcomes the limitation of relying solely on vehicle driving parameters, and can simultaneously capture physiological abnormalities and the risk of behavioral loss, avoiding misjudgment based on a single indicator. Secondly, the driver's gaze area is divided using a dynamic clustering method, and regional differentiation analysis is performed to determine the correlation between multimodal data such as fatigue, eye movement behavior characteristics, and gaze area, which helps to further improve the accuracy of driving risk analysis. Finally, causal reasoning is implemented through a Bayesian network, and joint probabilities are determined based on multiple change patterns, further improving the accuracy of risk analysis. In this way, through multimodal data fusion, dynamic region division, and causal reasoning modeling, the above method significantly improves the accuracy, real-time nature, and scenario adaptability of driving risk assessment.

[0089] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 The steps of the driving risk analysis method are provided.

[0090] The present invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The steps of the driving risk analysis method are provided.

[0091] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0093] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0095] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are included in the scope of protection of the patent for the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A driving risk analysis method, characterized in that: The method comprises: The dynamic clustering method is used to divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side and the instrument panel inside the vehicle; Extracting eye movement behavior characteristics of each region and determining a first change pattern of the eye movement behavior characteristics under different fatigue states of each region; Acquire vehicle driving parameters and determine a second variation law of the vehicle driving parameters under different fatigue states; A Bayesian network is constructed, eye movement behavior characteristics and vehicle driving parameters are used as input variables, and driving safety risk analysis is performed according to the first change law and the second change law to determine the risk assessment result.

2. A driving risk analysis method according to claim 1, characterized in that: The dynamic clustering method is used to divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side, and the instrument panel inside the vehicle. The eye-tracking device is used to collect the driver's eye movement data during the simulated driving process, including the location of the gaze point and the duration of the gaze; ‌Preprocess the collected eye movement data to remove outliers and noise; The dynamic clustering algorithm is used to perform cluster analysis on the preprocessed eye movement data. The number of clusters is set to 4. Through iterative calculation, the driver's gaze area is divided into four areas: the front, left, right and dashboard of the vehicle.

3. A driving risk analysis method according to claim 1, characterized in that: Extracting eye movement behavior features based on the divided gaze areas and determining a first change pattern of the eye movement behavior features under different fatigue states in each area includes: Extracting eye movement behavior features from the divided gaze areas, including blink frequency, gaze duration, saccade duration, and pupil area. The driver's fatigue state is divided into three levels: alert, tired and very tired; For each gaze area, the first change pattern of blink frequency, gaze duration, saccade duration and pupil area under different fatigue states is analyzed, and the first change pattern represents the relationship between the eye movement behavior characteristics and the fatigue state.

4. The driving risk analysis method according to claim 1, characterized in that: The obtaining of the vehicle driving parameters and determining the second variation rule of the vehicle driving parameters under different fatigue states includes: Collecting driving parameters of the driver under different fatigue states, including steering wheel angle, vehicle speed and vehicle lateral offset; Preprocess and analyze the collected vehicle driving parameters, and calculate the mean and standard deviation of the steering wheel angle, vehicle speed and vehicle lateral offset respectively; Through statistical analysis and comparison, a second variation law of the steering wheel angle, vehicle speed and vehicle lateral offset under different fatigue states is determined, and the second variation law represents the influence of the fatigue state on the vehicle driving parameters.

5. The driving risk analysis method according to claim 1, characterized in that: The Bayesian network is constructed, eye movement behavior characteristics and vehicle driving parameters are used as input variables, and a driving safety risk assessment model is constructed according to the first change law and the second change law, including: A Bayesian network is constructed, and the eye movement behavior characteristics of the four gaze areas and the vehicle driving parameters are used as observation nodes. The intermediate node of the fatigue state is determined based on the first change law, and the global node of the fatigue state is determined based on the second change law. The observation node of the eye movement characteristics is connected to the intermediate fatigue state node, and the observation node of the vehicle driving parameters is connected to the global fatigue state node. All state nodes are integrated for reasoning based on the preset weight distribution rules to output the risk assessment results.

6. A driving risk analysis device, characterized in that: The device comprises: A division module is used to divide the driver's gaze area into four areas: the front of the vehicle, the left side, the right side, and the instrument panel inside the vehicle using a dynamic clustering method; A determination module is configured to extract eye movement behavior characteristics of each region and determine a first change pattern of the eye movement behavior characteristics under different fatigue states of each region; obtain vehicle driving parameters and determine a second change pattern of the vehicle driving parameters under different fatigue states; A construction module is used to construct a Bayesian network, take eye movement behavior characteristics and vehicle driving parameters as input variables, perform driving safety risk analysis based on the first change law and the second change law, and determine the risk assessment result.

7. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 5 when executing the program.

Images