A driving risk assessment method for autonomous vehicles in weaving areas on expressways

Abstract

The present invention belongs to the technical field of traffic control systems, and specifically relates to a method for assessing driving risks for autonomous vehicles in weaving zones on expressways. The method comprises the following steps: first, modeling the spatiotemporal risk field for obstacles; second, modeling the risk field for road boundaries and road lines; third, modeling the risk field for on- and off-ramps on the expressway; fourth, calculating the total risk field; fifth, extracting real-world vehicle trajectories on expressways; and sixth, calibrating parameters based on the theory of dynamic risk balance. The advantages of the present invention are: by innovatively introducing the physical quantity of spatiotemporal distance, the two-dimensional spatial risk field is upgraded to a three-dimensional temporal-spatial risk field, effectively quantifying the impact of predicted trajectories and time effects on risk. A special geometric configuration field for on- and off-ramps in weaving zones is also constructed, which accurately restricts the operation of specific vehicles. Furthermore, based on the theory of dynamic risk balance and machine vision technologies such as YOLO, the model parameters are calibrated using real aerial photography data.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic control systems, and specifically relates to a driving risk assessment method for autonomous driving vehicles in an expressway weaving area based on a spatiotemporal risk field. Background Art

[0002] Expressways, as the core arteries of urban transportation networks, bear a heavy traffic load, but traffic safety issues are particularly pronounced in weaving zones. Traffic flow in these zones exhibits a typical three-way interaction: on-ramps merge, off-ramps diverge, and vehicles on the main line travel through, creating multiple weaving flows. Frequent lane changes induce turbulence, significantly increasing accident rates and congestion indices, making them bottlenecks with high accident rates and major congestion. In recent years, the rapid development of automated driving (CAV) technology has provided a new technological path to addressing driving risks in weaving zones with its precise perception, decision-making, and control capabilities. However, in the current traffic environment where CAVs and human-driven vehicles (HDVs) coexist for a long time, the external complexity of driving risks has increased significantly. On the one hand, the inherent uncertainty of traditional traffic flow persists; on the other hand, the significant differences in driving behavior between CAVs and HDVs further exacerbate operational risks in weaving zones. This dual risk effect has, to a certain extent, limited the potential advantages of CAV technology. Therefore, conducting a systematic assessment and quantitative calibration study on the driving risks of autonomous vehicles in expressway weaving areas has important theoretical value and practical significance for ensuring the accurate operation of CAV lane-changing decisions and trajectory planning.

[0003] Currently, common driving risk assessment methods mainly include three types of deterministic methods: time-based (such as collision time TTC, headway time THW, and rear intrusion time PET), distance-based (such as minimum safe distance MSD), and acceleration-based (such as collision avoidance deceleration rate DRAC). A Chinese patent disclosed "A method for evaluating driving risks in underground tunnels based on dynamic Bayesian networks". Patent application number CN202411687491.1 proposed calculating the safe distance based on a Bayesian network, thereby realizing the assessment of vehicle rear-end collision risk; a Chinese patent disclosed "A dynamic driving risk assessment method and system based on DNC and NGBoost-PA-WOE". Patent application number: CN202411663328.1 proposed the use of machine learning methods to learn based on traffic accident-related data (such as time, distance, acceleration, etc.) to predict accident risks. These methods are simple to calculate and have high computational efficiency, and are widely used in risk identification and early warning systems. However, there are still significant deficiencies in practical applications: (1) The three types of methods generally assume that the vehicle state remains unchanged in a short period of time, which makes it difficult to reflect the uncertainty of vehicle movement and its surrounding environment, and easily leads to risk misjudgment; (2) Most methods focus on the identification of longitudinal rear-end collision risks, and it is difficult to effectively assess the more critical lateral collision risk during lane changing; (3) In scenarios with complex risk sources and diverse interactions, traditional methods tend to underestimate risks, which limits their application effect in highly complex traffic environments.

[0004] Risk field theory, an improved artificial potential field approach, originates from field theory in physics. Compared to traditional risk assessment methods, it offers advantages such as strong dynamic modeling capabilities, the ability to quantify risk distribution, and adaptability to the characteristics of diverse traffic participants. It can dynamically reflect the interactive risks between traffic participants across time and space, making it particularly suitable for assessing and predicting risk dynamics in complex, mixed traffic scenarios. For example, the article "Extension and Application of Driving Risk Field Models in a Collaborative Vehicle-Road Environment" proposes a risk assessment model for pure CAV scenarios that incorporates parameters such as TTC, vehicle geometry, and yaw angle. The Chinese patent application "Automated Driving Takeover Risk Assessment Model Considering Driving Risk Field and Its Construction Method" (patent application number: CN202211137773.5) proposes a method for constructing an automated driving takeover risk assessment model that considers driving risk fields, constructing a distance-based static risk field. The Chinese patent application "Vehicle Driving Risk Identification Method Based on Digital Twins and Predicted Risk Fields" (patent application number: CN202411018302.1) uses the predicted risk field to superimpose a series of potential static fields within the prediction time domain to obtain a risk field for the corresponding single-modal predicted trajectory, and uses this to identify driving risks. Although increasing research has improved risk fields, these constructed risk fields remain essentially two-dimensional static fields, accounting only for the risk impact of current risk sources and ignoring the impact of the future motion trends of dynamic obstacles on the current risk. Furthermore, the special fields in interweaving scenarios have rarely been studied. Due to the increased frequency of lane changes and the dominance of turbulence, the risk field in weaving areas is particularly unique, and the impact of the environment on risk is far greater than that of ordinary roads. However, there is still no relevant research to systematically explain this change. Summary of the Invention

[0005] In view of the above problems, the purpose of the present invention is to provide a driving risk assessment method for autonomous driving vehicles in the weaving area of ​​expressways. By introducing the physical quantity of space-time distance, the two-dimensional spatial risk field is upgraded to a three-dimensional time-space risk field, effectively quantifying the impact of predicted trajectory and time effect on risk, and constructing a special geometric configuration field of the on- and off-ramps in the weaving area, which imposes accurate restrictions on the operation of specific vehicles.

[0006] The present invention provides a method for assessing driving risks of an autonomous driving vehicle in an expressway weaving zone, comprising the following steps:

[0007] Step 1: Modeling the space-time risk field of obstacles, including: constructing a space-time coordinate system, calculating space-time distance, calculating the charge taking into account the body and motion properties, calculating the dielectric constant taking into account anisotropy and environmental complexity, and calculating the space-time risk of obstacles;

[0008] Step 2: Modeling the risk field of road boundaries and road lines, including: calculating the field strength of road boundaries, calculating the field strength of road solid lines, calculating the field strength of road dashed lines, and calculating the risk field of road boundaries and road lines;

[0009] Step 3: Modeling the risk field of on- and off-ramps on the expressway;

[0010] Step 4: Calculate the total risk field;

[0011] Step 5: Real-world freeway vehicle trajectory extraction, including: vehicle trajectory extraction based on YOLOv8, target tracking module based on Deep-SORT, video stabilization processing, model training and testing, and coordinate system conversion;

[0012] Step 6: Parameter calibration based on the risk dynamic balance theory, including: determining the objective function, determining the upper and lower risk limits of the risk dynamic balance theory, and using a hybrid genetic algorithm and a variable neighborhood search algorithm to solve the problem and realize the driving risk assessment of autonomous vehicles in the expressway weaving area.

[0013] As a preferred embodiment of the present invention, step one also includes the following steps:

[0014] A1. Construct a space-time coordinate system.

[0015] A three-dimensional space-time coordinate system is established based on the Frenet coordinate system, with the direction of travel along the road as the S axis, the direction perpendicular to the road as the D axis, and time as the T axis.

[0016] A2. Calculate spatiotemporal distance, including: obtaining predicted obstacle trajectories, calculating spatial distances that take into account vehicle geometry, calculating equivalent time-based distances that take into account risk anisotropy, and obtaining spatiotemporal distances.

[0017] A3. Calculate the amount of charge taking into account its own and motion properties.

[0018] Two representative parameters are selected from the comprehensive attributes of vehicle functions and driving behavior as key influencing factors of charge: braking performance and reaction time;

[0019] Charge q obs The calculation formula is defined as follows;

[0020]

[0021] Where m obs is the mass of the obstacle, v obs is the speed of the vehicle, a max is the maximum deceleration of the vehicle, T r is the vehicle reaction time, i.e., the sum of the driver's reaction time and the vehicle system delay, γ1 and γ2 are unknown coefficients, and exp is an exponential function;

[0022] A4. Calculate the dielectric constant taking into account anisotropy and environmental complexity.

[0023] The dielectric constant in the risk field is defined as the product of the anisotropic attenuation coefficient and the environmental attenuation coefficient. The dielectric constant ε is calculated as follows:

[0024]

[0025] Among them, η is the anisotropic attenuation coefficient, θ is the environmental complexity, and the specific calculation method is as follows:

[0026] η=exp[k·cos(ψ)·(β1·v obs +β2·a obs )];

[0027] θ=V·C;

[0028] Where ψ is the angle between the obstacle's direction of travel and d', d'=(s A -s obs ,d A -d obs ) is the position vector of the obstacle pointing to the observation point, v obs and a obs are the speed and acceleration of the vehicle, V is the ratio of the square difference of the speed in the weaving section to that on the normal road, C is the ratio of the lane change ratio in the weaving section to that on the normal road, k, β1 and β2 are unknown coefficients;

[0029] A5. Calculate the spatial and temporal risk of obstacles.

[0030] Obstacle space-time risk E obs,A The calculation formula is as follows:

[0031]

[0032] Among them, r is the space-time distance; q obs is the charge; ε is the dielectric constant.

[0033] As a preferred embodiment of the present invention, step A2 further includes the following steps:

[0034] A2.1. Obtain the predicted trajectory of the obstacle.

[0035] The obstacle trajectory scatter point set P is expressed as: P = {p i ,p i+1 ,…,p m}, where p i Indicates t i The space-time trajectory of the obstacle prediction at each moment, ti The position of the obstacle along the direction of travel of the road and the position perpendicular to the direction of travel of the road;

[0036] A2.2. Calculate the spatial distance considering the vehicle's geometric outline.

[0037] In the vehicle body coordinate system, the center of mass of the vehicle is taken as the origin, the direction along which the vehicle moves is the X axis, and the direction perpendicular to the direction in which the vehicle moves is the Y axis. In the space-time coordinate system, the coordinates of point A are (s A ,d A ), the coordinates of the obstacle mass center are (s obs ,d obs ), in the vehicle body coordinate system, the four vertices of the rectangle T E The coordinates are: (l / 2,w / 2), (l / 2,-w / 2), (-l / 2,w / 2), (-l / 2,-w / 2), In addition, the yaw angle of the vehicle The speed of the vehicle v obs , two types of coordinate systems are transformed using translation and rotation, as follows:

[0038]

[0039] The formula for solving the spatial distance d based on the vehicle body coordinate system is shown below;

[0040]

[0041] In the formula, (x A ,y A ) is the coordinate of the observation point in the vehicle body coordinate system; (x obs ,y obs ) is the coordinate of the observation point; (x′ A ,y′ A ) is the converted coordinate; w is the vehicle width; l is the vehicle length, (d1, d2, d3, d4) are the four vertices of the rectangle T E and the transformed observation point (x′ A ,y′ A )’s spatial distance;

[0042] A2.3. Calculate the equivalent time-base distance considering risk anisotropy.

[0043] Introducing the equivalent time base distance T * , equivalent time base distance T *Meaning: When point A of the interactive object is directly in front of the obstacle's travel path, and point B of the interactive object is not directly in front of the vehicle's travel path, assuming that point A of the interactive object and point B of the interactive object have equal collision risk, then the time distance from the vehicle to point A of the interactive object is equal to the equivalent time distance from the vehicle to point B of the interactive object, the equivalent time base distance T * Used to deal with the problem of unequal risks in different directions, the equivalent time base distance T * It is based on the time dimension weighting of time distance and space-time distance. The specific expression is as follows:

[0044]

[0045] In the formula, (x′ A ,y′ A ) is the coordinate of the observation point in the vehicle body coordinate system; w is the vehicle width; l is the vehicle length; The four vertices T of the rectangle E The corresponding equivalent time base distance, μ is the conversion parameter; A2.4, combining time distance and time to obtain space-time distance,

[0046] In t i The space-time distance of a moment Solve according to the following formula:

[0047]

[0048] In the formula t i The temporal and spatial distance between the observation point and the obstacle at any moment, t i The equivalent time base distance between point A, the closest point to the observation point, and the observation point at the moment t i The point A closest to the observation point on the obstacle trajectory at the moment is denoted as p min , the time required for a vehicle to reach a designated location is called the time-based distance, t min t i Time trajectory scatter point p min The T-axis position of , α is the weight of the time effect;

[0049] The minimum value of the space-time distance of each scattering point is selected as the space-time distance from point A to the obstacle. The calculation formula is as follows:

[0050]

[0051] Where r i t i The temporal and spatial distance between the observation point and the obstacle at any moment, 2<n<m.

[0052] As a preferred embodiment of the present invention, step 2 further includes the following steps:

[0053] B1. Calculate the field strength at the road boundary.

[0054] Road boundary field strength E lane1,A The calculation formula is as follows:

[0055]

[0056] Where W is the road width; d lane 1 is the lateral position of the road boundary; d A is the horizontal position of the observation point, is the coefficient to be determined;

[0057] B2. Calculate the field strength of the solid road line.

[0058] Road solid line field strength E lane2,A The calculation formula is as follows:

[0059]

[0060] Where W is the road width; d lane2 is the horizontal position of the solid line of the road; d A is the horizontal position of the observation point, is the coefficient to be determined;

[0061] B3. Calculate the field strength of the dotted line of the road.

[0062] Field strength E of the road dotted line lane3，A The calculation formula is as follows:

[0063]

[0064] Where W is the road width; d lane3 is the horizontal position of the road dotted line; d A is the lateral position of the observation point, is the coefficient to be determined;

[0065] B4. Calculate the road boundary and road line risk field,

[0066] Road boundary and road line risk field E lane,A The calculation formula is as follows:

[0067] E lane,A =E lane1,A +E lane2,A +E lane3,A ;

[0068] Among them, E lane1,A is the field strength at the road boundary; E lane2,A is the field strength of the solid road line; Elane3,A is the field strength of the dotted line of the road.

[0069] As a preferred embodiment of the present invention, step three also includes the following steps:

[0070] Field strength E of the on- and off-ramp geo,A The calculation method is as follows:

[0071]

[0072] Where s end is the S-axis coordinate of the starting point of the down ramp, s start is the S-axis coordinate of the starting position of the forced lane change area, p A =(s A ,d A ) is the position of point A on the vehicle, P mand is the range of the forced lane change area, σ1 and σ2 are unknown parameters, χ is the set of vehicles on the ramp that intend to merge into the main line, δ is the set of vehicles on the main line that intend to get off the ramp, and γ is the set of remaining vehicles.

[0073] As a preferred embodiment of the present invention, step 4 further includes the following steps:

[0074] The total risk field is considered as the superposition of multiple traffic element fields. The total risk field is obtained by adding the spatial and temporal risk field of obstacles, the road boundary and road line field, and the risk field of on- and off-ramps on the expressway. The total risk field calculation formula for any point A is as follows:

[0075] E total,A =E obs,A +E lane,A +E geo,A ;

[0076] Where, E total,A is the total risk field of point A, E obs,A is the field strength exerted by the obstacle on point A, E lane,A is the field strength exerted by the road boundary and road line on point A, E geo,A is the field strength exerted on point A by the on- and off-ramps.

[0077] As a preferred embodiment of the present invention, step five also includes the following steps:

[0078] E1. Extract vehicle trajectory based on YOLOv8.

[0079] E1.1 Vehicle detection module based on YOLOv8,

[0080] The vehicle detection module based on YOLOv8 is used to directly predict the location, size and category of the target in a single network;

[0081] E1.2, Deep-SORT-based target tracking module,

[0082] The Deep-SORT-based target tracking module is used to describe the target's appearance information and distinguish different vehicles; E1.3, video stabilization processing,

[0083] Use the After Effects system to pre-process the video data. By tracking multiple fixed buildings in the video, the rotation and displacement information of the entire video is extracted, and then the inverse motion is applied to eliminate the transformation information.

[0084] E1.4, Model training and testing,

[0085] Construct a dedicated dataset for vertical view and use the image annotation tool LabelImg for data annotation;

[0086] E1.5, Coordinate system conversion,

[0087] By obtaining the position information of road boundaries and road lines, converting them into coordinate information in the Frenet coordinate system, we can obtain the real-world expressway vehicle trajectory.

[0088] As a preferred embodiment of the present invention, step six also includes the following steps:

[0089] F1. Determine the objective function,

[0090] Based on the risk dynamic balance theory, the mean square error of the distance between the sample risk after decision making and the risk dynamic balance curve is used as the objective function of the calibration model. The specific formula is as follows:

[0091]

[0092] Where RMSE is the objective function of the calibration model, that is, the distance between the risk after the nth sample decision and the acceptable risk space; ΔR n is the distance between the decision of the nth sample and the risk dynamic equilibrium curve; R n is the risk value at a specific moment after the decision of the nth sample; R max With R min are the upper limit risk and lower limit risk of the risk dynamic balance theory respectively; M is the number of samples;

[0093] F2. Determine the upper limit risk R of the risk dynamic balance theory max and the lower risk R min ;

[0094] Calculate the upper limit risk R using a proprietary dataset from a vertical top-down perspective max and the lower risk R min,The vertical top-view exclusive dataset includes high-risk dataset, low-risk dataset and decision dataset, which contains the trajectories of the observed vehicle and multiple surrounding vehicles. The high-risk dataset includes {R1, R2…R L}Includes collision time assessment, observing vehicles encountering high-risk scenarios, using 3 seconds as the TTC threshold, and the high-risk dataset is used to determine the upper limit risk R max ,On the contrary, the low-risk dataset contains scenarios where the observed vehicle always remains safe according to the judgment of TTC threshold, which is used to determine the optimal risk level R best , upper limit risk R max and the optimal risk level R best The calculation method is as follows,

[0095] R max =min{R1,R2…R L};

[0096]

[0097] where R m To observe the level of risk experienced by the vehicle after making a decision;

[0098] Based on R max and R best , determine the lower limit risk R min , the calculation method is as follows,

[0099] R min =2R best -R max ;

[0100] F3, using hybrid genetic algorithm and variable neighborhood search algorithm to solve,

[0101] A hybrid genetic algorithm and a variable neighborhood search algorithm are used to solve the risk field parameters, and according to steps one to three, the driving risk assessment of autonomous vehicles in the expressway weaving area is realized.

[0102] The beneficial effects of this invention are as follows: By innovatively introducing the physical quantity of space-time distance, the two-dimensional spatial risk field is upgraded to a three-dimensional time-space risk field, effectively quantifying the impact of predicted trajectory and time effects on risk. A special geometric configuration field for on- and off-ramps in weaving areas is also constructed, which accurately restricts the operation of specific vehicles. Furthermore, based on the theory of dynamic risk balance and machine vision technologies such as YOLO, the model parameters are calibrated using real aerial data. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] By referring to the following description in conjunction with the accompanying drawings, and with a more complete understanding of the present invention, other objects and results of the present invention will become more clear and easy to understand. In the accompanying drawings:

[0104] Figure 1 A schematic diagram of the vehicle geometric profile space distance of the present invention;

[0105] Figure 2 A schematic diagram of a vehicle weaving field on an expressway on-ramp or off-ramp according to the present invention;

[0106] Figure 3 This is a trajectory recognition effect diagram of the present invention;

[0107] Figure 4 This is a flow chart for solving the hybrid genetic algorithm and variable neighborhood search algorithm of the present invention. DETAILED DESCRIPTION

[0108] See Figure 1-4 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0109] An embodiment of the present invention provides a method for assessing driving risk of an autonomous driving vehicle in an expressway weaving area, comprising the following steps:

[0110] Step 1: Modeling the space-time risk field of obstacles.

[0111] A1. Construct a space-time coordinate system.

[0112] A three-dimensional space-time coordinate system is established based on the Frenet coordinate system, with the direction of travel along the road as the S axis, the direction perpendicular to the road as the D axis, and time as the T axis;

[0113] A2. Calculate the time and space distance,

[0114] Since the obstacle trajectory does not follow a specific function, the time-space distance cannot be calculated directly. i Taking the time-space distance at the moment as an example, the time-space distance calculation method is used, including: obtaining the predicted trajectory of the obstacle, the spatial distance calculation method considering the geometric contour of the vehicle, the calculation method of the equivalent time-based distance considering the risk anisotropy, and obtaining the time-space distance;

[0115] A2.1. Obtain the predicted trajectory of the obstacle.

[0116] The obstacle trajectory scatter point set P is expressed as: P = {p i ,p i+1 ,…,p m}, where p i Indicates t i The space-time trajectory of the obstacle prediction at each moment, t i The position of the obstacle along the direction of road travel and the position perpendicular to the direction of road travel, that is, the position in the time and space coordinate system;

[0117] A2.2. Calculate the spatial distance considering the vehicle's geometric outline.

[0118] Treating obstacles as point masses with mass but no shape does not fully consider the impact of vehicle geometry, reducing the accuracy of risk quantification. To address this issue, a spatial distance solution method based on coordinate transformation is designed. In this method, the obstacle shape is regarded as a rectangle with fixed length and width values. This is because most obstacles during driving are interfering vehicles.

[0119] Since the T axis does not change, the space-time coordinate system and the vehicle body coordinate system are established with the T axis ignored. In the vehicle body coordinate system, the center of mass of the vehicle is the origin, the direction of vehicle travel is the X axis, and the direction perpendicular to the direction of vehicle travel is the Y axis. In the space-time coordinate system, the coordinates of point A are (s A ,d A ), the coordinates of the obstacle mass center are (s obs ,d obs ), in the vehicle body coordinate system, the four vertices of the rectangle T E The coordinates are: (l / 2,w / 2), (l / 2,-w / 2), (-l / 2,w / 2), (-l / 2,-w / 2), In addition, the yaw angle of the vehicle The speed of the vehicle v obs , using the two steps of translation and rotation, we can get two types of coordinate systems, as follows:

[0120]

[0121] The formula for solving the spatial distance d based on the vehicle body coordinate system is shown below;

[0122]

[0123]

[0124] In the formula, (x A ,y A ) is the coordinate of the observation point in the vehicle body coordinate system; (x obs ,y obs ) is the coordinate of the observation point; (x′ A ,y′ A ) is the converted coordinate; w is the vehicle width; l is the vehicle length, (d1, d2, d3, d4) are the four vertices of the rectangle T E and the transformed observation point (x′ A ,y′ A )’s spatial distance;

[0125] A2.3. Calculate the equivalent time-base distance considering risk anisotropy.

[0126] Considering that the Euclidean distance in A2.2 has two major flaws: first, it assumes that obstacles pose equal risks to objects at different locations but equal distances, which obviously violates common sense; second, the Euclidean distance and time are not the same physical properties, and there is a disadvantage of being physically meaningless when weighting the time-space distance; in order to improve the above problems, the equivalent time-based distance T is introduced. * The concept of equivalent time base distance T * Meaning: When point A of the interactive object is directly in front of the obstacle's travel path, and point B of the interactive object is not directly in front of the vehicle's travel path, assuming that point A of the interactive object and point B of the interactive object have equal collision risk, then the time distance from the vehicle to point A of the interactive object is equal to the equivalent time distance from the vehicle to point B of the interactive object, the equivalent time base distance T * It is used to deal with the problem of unequal risks in different directions, and in terms of physical properties, the equivalent time base distance T * It is based on the time dimension weighting of time distance and space-time distance. The specific expression is as follows:

[0127]

[0128]

[0129] In the formula, (x′ A ,y′ A ) is the coordinate of the observation point in the vehicle body coordinate system; w is the vehicle width; l is the vehicle length; The four vertices T of the rectangle E The corresponding equivalent time base distance, μ is the conversion parameter; A2.4, combining time distance and time to obtain space-time distance,

[0130] In t i The space-time distance of a moment (i.e. the distance from the observation point to the vehicle trajectory scatter point) is solved according to the following formula:

[0131]

[0132] In the formula t i The temporal and spatial distance between the observation point and the obstacle at any moment, t i The equivalent time base distance between point A, the closest point to the observation point, and the observation point at the moment t i The point A closest to the observation point on the obstacle trajectory at the moment is denoted as p min , the time required for a vehicle to reach a designated location is called the time-based distance, t min t i Time trajectory scatter point p minThe T-axis position of , α is the weight of the time effect;

[0133] After obtaining the space-time distances of all scattering points, the minimum value of the space-time distance of each scattering point is selected as the space-time distance from point A to the obstacle. The calculation formula is as follows:

[0134]

[0135] Where r i t i The temporal and spatial distance between the observation point and the obstacle at any moment, 2<n<m;

[0136] A3. Calculate the amount of charge taking into account its own and motion properties.

[0137] In the electrostatic field, charge is a fundamental physical quantity that characterizes the electrical properties of an object (positive or negative) and its relationship with potential energy. Similarly, in the risk field, charge represents the comprehensive attributes of vehicle functions and driving behavior. Two representative parameters are selected from the comprehensive attributes of vehicle functions and driving behavior as key influencing factors of charge: braking performance and reaction time. The reasons for selecting these two parameters are: (1) Braking performance reflects the ability of the vehicle to avoid risks when faced with risks; (2) Reaction time, including driver response time and vehicle system delay, determines the driver and vehicle's perception and response efficiency to potential risks; (3) Charge is measured by combining the vehicle's own properties, motion properties, mass, and acceleration.

[0138] Charge q obs The calculation formula is defined as follows;

[0139]

[0140] Where m obs is the mass of the obstacle, v obs is the speed of the vehicle, a max is the maximum deceleration of the vehicle, T r is the vehicle reaction time, i.e., the sum of the driver's reaction time and the vehicle system delay, γ1 and γ2 are unknown coefficients, and exp is an exponential function;

[0141] A4. Calculate the dielectric constant taking into account anisotropy and environmental complexity.

[0142] In electrostatic fields, the dielectric constant is a physical quantity that describes the ability of a material to enhance the charge storage capacity of a capacitor. It is defined as the product of the absolute dielectric constant in a vacuum and the relative dielectric constant in a given environment. In the context of risk fields, an analogy is established between the dielectric constant in risk fields and the dielectric constant in electrostatic fields.

[0143] The dielectric constant in the risk field is defined as the product of the anisotropic attenuation coefficient and the environmental attenuation coefficient. The dielectric constant ε is calculated as follows:

[0144]

[0145] Among them, η is the anisotropic attenuation coefficient, θ is the environmental complexity, and the specific calculation method is as follows:

[0146] η=exp[k·cos(ψ)·(β1·v obs +β2·a obs )];

[0147] θ=V·C;

[0148] Where ψ is the angle between the obstacle's direction of travel and d', d'=(s A -s obs ,d A -d obs ) is the position vector of the obstacle pointing to the observation point, v obs and a obs are the speed and acceleration of the vehicle, V is the ratio of the square difference of the speed in the weaving section to that on the normal road, C is the ratio of the lane change ratio in the weaving section to that on the normal road, k, β1 and β2 are unknown coefficients;

[0149] A5. Calculate the spatial and temporal risk of obstacles.

[0150] Obstacle space-time risk E obs,A The calculation formula is as follows:

[0151]

[0152] Among them, r is the space-time distance; q obs is the charge; ε is the dielectric constant;

[0153] Step 2: Modeling the road boundary and road line risk field.

[0154] Road boundaries and lane markings impose restrictions on vehicles traveling normally. To avoid penalties associated with crossing lane markings, drivers constantly adjust the vehicle's lateral position throughout the driving process. Therefore, the risk field of road boundaries and lane markings varies with relative distance. The size of the risk field is only related to the line shape and lateral position of the road boundaries and lane markings.

[0155] B1. Calculate the field strength at the road boundary.

[0156] Road boundary field strength E lane1,A The calculation formula is as follows:

[0157]

[0158] Where W is the road width; d lane1 is the lateral position of the road boundary; d A is the lateral position of the observation point, is the coefficient to be determined;

[0159] B2. Calculate the field strength of the solid road line.

[0160] Road solid line field strength E lane2,A The calculation formula is as follows:

[0161]

[0162] Where W is the road width; d lane2 is the horizontal position of the solid line of the road; d A is the lateral position of the observation point, is the coefficient to be determined;

[0163] B3. Calculate the field strength of the dotted line of the road.

[0164] Field strength E of the road dotted line lane3，A The calculation formula is as follows:

[0165]

[0166] Where W is the road width; d lane3 is the horizontal position of the road dotted line; d A is the lateral position of the observation point, is the coefficient to be determined;

[0167] B4. Calculate the road boundary and road line risk field,

[0168] Road boundary and road line risk field E lane,A The calculation formula is as follows:

[0169] E lane,A =E lane1,A +E lane2,A +E lane3,A ;

[0170] Among them, E lane1,A is the field strength at the road boundary; E lane2,A is the field strength of the solid road line; E lane3,A is the field strength of the dotted line of the road;

[0171] Step 3: Modeling the risk field of on- and off-ramps on the expressway.

[0172] In the weaving area of ​​an expressway, vehicles on the ramp merging onto the main line must move from the acceleration lane to the inner lane within specified spatial constraints, while vehicles on the ramp off must move from the main line lane to the outer lane within specified spatial constraints. For a typical weaving area geometry, the forced lane change behavior on the ramps is considered an adaptive decision influenced by the risk field of the longitudinal acceleration and deceleration lanes.

[0173] Field strength E of the on- and off-ramp geo,A The calculation method is as follows:

[0174]

[0175] Where s end is the S-axis coordinate of the starting point of the down ramp, s start is the S-axis coordinate of the starting position of the forced lane change area, p A =(s A ,d A ) is the position of point A on the vehicle, P mand is the range of the forced lane change area, σ1 and σ2 are unknown parameters, χ is the set of vehicles on the ramp that intend to merge into the main line, δ is the set of vehicles on the main line that intend to get off the ramp, and γ is the set of remaining vehicles;

[0176] Step 4: Calculate the total risk field.

[0177] The total risk field is considered as the superposition of multiple traffic element fields. The total risk field is obtained by adding the spatial and temporal risk field of obstacles, the road boundary and road line field, and the risk field of on- and off-ramps on the expressway. The total risk field calculation formula for any point A is as follows:

[0178] E total,A =E obs,A +E lane,A +E geo,A ;

[0179] Where, E total,A is the total risk field of point A, E obs,A is the field strength exerted by the obstacle on point A, E lane,A is the field strength exerted by the road boundary and road line on point A, E geo,A is the field strength exerted on point A by the on- and off-ramps;

[0180] Step 5: Real-world expressway vehicle trajectory extraction,

[0181] E1. Extract vehicle trajectory based on YOLOv8.

[0182] E1.1 Vehicle detection module based on YOLOv8 (vehicle detection algorithm),

[0183] The YOLO family of algorithms can directly predict the location, size, and category of an object within a single network. Their high detection speed makes real-time detection easy. The YOLOv8 algorithm further improves detection accuracy and speed based on the YOLO family of algorithms, enabling more effective detection of vehicles in videos.

[0184] The vehicle detection module based on YOLOv8 is used to directly predict the location, size and category of the target in a single network;

[0185] E1.2, Deep-SORT-based target tracking module (target tracking algorithm),

[0186] After completing vehicle detection, the detection results need to be combined with the Multi-Object Tracking (MOT) algorithm to achieve continuous tracking of vehicle trajectories. Deep-SORT is an improvement on the SORT (Simple Online and Realtime Tracking) algorithm. It introduces deep learning features to describe the appearance information of the target, improving the accuracy of target matching and the stability of tracking. When the vehicle is occluded or its appearance changes, Deep-SORT can use appearance features to better distinguish different vehicles, reduce mismatches and target loss, and has stronger robustness in complex scenarios.

[0187] The Deep-SORT-based target tracking module is used to describe the target's appearance information and distinguish different vehicles; E1.3, video stabilization processing,

[0188] When drones shoot at high altitudes, they are affected by wind direction and air currents, causing the captured video to rotate and shift, inevitably resulting in slight jitter. This change in the video image can alter the vehicle's pixel coordinates, causing them to deviate from their actual position. To stabilize the video and extract the vehicle's trajectory, we used the After Effects system to pre-process the video data. By tracking multiple fixed buildings in the video, we extracted the overall rotation and displacement information of the video, then applied an inverse motion to eliminate the transformation information, thereby achieving video stabilization.

[0189] E1.4, Model training and testing,

[0190] Many existing object detection models rely on training datasets acquired from oblique perspectives, reflecting common shooting scenarios. Focusing on vehicle detection from a vertical bird's-eye view, while this perspective provides more comprehensive and unique vehicle information, existing models trained on oblique-perspective datasets are unable to adapt to these image characteristics, resulting in poor detection performance from this perspective. To address this issue, a dedicated vertical bird's-eye view dataset was constructed and annotated using the image annotation tool LabelImg. The selected altitude was approximately 350 meters, the minimum required to fully cover the intersecting area and help capture global features. The results of aerial vehicle recognition and tracking were also analyzed.

[0191] E1.5, Coordinate system conversion,

[0192] Due to limitations in the camera's battery life and human control, the shooting height and position of each video set cannot be completely consistent. As a result, the same position in different video sets may have discrepancies. Therefore, by obtaining the position information of road boundaries and road lines, and converting it into the Frenet coordinate system, we can obtain the real-world vehicle trajectory on the expressway.

[0193] Step 6: Parameter calibration based on risk dynamic balance theory.

[0194] F1. Determine the objective function,

[0195] Based on the risk dynamic balance theory, the mean square error of the distance between the sample risk after decision making and the risk dynamic balance curve is used as the objective function of the calibration model. The specific formula is as follows:

[0196]

[0197] Where RMSE is the objective function of the calibration model, that is, the distance between the risk after the nth sample decision and the acceptable risk space; ΔR n is the distance between the decision of the nth sample and the risk dynamic equilibrium curve; R n is the risk value at a specific moment after the decision of the nth sample; R max With R min are the upper limit risk and lower limit risk of the risk dynamic balance theory respectively; M is the number of samples;

[0198] F2. Determine the upper limit risk R of the risk dynamic balance theory max and the lower risk R min ;

[0199] Calculate the upper limit risk R using a proprietary dataset from a vertical top-down perspective max and the lower risk R minThe vertical viewpoint exclusive dataset includes high-risk dataset, low-risk dataset and decision dataset, which contains the trajectories of the observed vehicle (the vehicle itself) and multiple surrounding vehicles. The high-risk dataset includes {R1, R2…R L The use of time-to-collision (TTC) evaluation is included to observe the vehicle encountering high-risk scenarios, using 3 seconds as the TTC threshold, and the high-risk dataset is used to determine the upper limit risk R max ,On the contrary, the low-risk dataset contains scenarios where the observed vehicle always remains safe according to the judgment of TTC threshold, which is used to determine the optimal risk level R best , upper limit risk R max and the optimal risk level R best The calculation method is as follows,

[0200] R max =min{R1,R2…R L};

[0201]

[0202] where R m To observe the level of risk experienced by the vehicle after making a decision;

[0203] Based on R max and R best , determine the lower limit risk R min , the calculation method is as follows,

[0204] R min =2R best -R max ;

[0205] F3, using hybrid genetic algorithm and variable neighborhood search algorithm to solve,

[0206] A hybrid genetic algorithm and a variable neighborhood search algorithm are used to solve the risk field parameters (

[0207] α, β1, β2, k, γ1, γ2, ), and according to steps one to three, realize the driving risk assessment of autonomous driving vehicles in the expressway weaving area.

[0208] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for assessing driving risk of an autonomous vehicle in an expressway weaving zone, characterized in that: The following steps are involved: Step 1: Modeling the space-time risk field of obstacles, including: constructing a space-time coordinate system, calculating space-time distance, calculating the charge taking into account the vehicle's own and motion properties, and calculating the dielectric constant taking into account anisotropy and environmental complexity, to achieve the calculation of space-time risk of obstacles; Step 2: Modeling the risk field of road boundaries and road lines, including: calculating the field strength of road boundaries, calculating the field strength of road solid lines, and calculating the field strength of road dotted lines, to achieve the calculation of the risk field of road boundaries and road lines; Step 3: Modeling the risk field of on- and off-ramps on the expressway; Step 4: Calculate the total risk field. The total risk field is the superposition of multiple traffic element fields. The total risk field is obtained by adding the obstacle spatiotemporal risk field, road boundary and road line field, and expressway on-ramp and off-ramp risk fields. Step 5: Real-world freeway vehicle trajectory extraction, including: vehicle trajectory extraction based on YOLOv8, target tracking module based on Deep-SORT, video stabilization processing, model training and testing, and coordinate system conversion; Step 6: Parameter calibration based on the risk dynamic balance theory, including: determining the objective function, determining the upper and lower risk limits of the risk dynamic balance theory, and solving the problem using a hybrid genetic algorithm and a variable neighborhood search algorithm to implement driving risk assessment for autonomous vehicles in the freeway weaving area; In determining the objective function, based on the risk dynamic balance theory, the mean square error of the distance between the sample risk after decision making and the risk dynamic balance curve is used as the objective function of the calibration model. The specific formula is as follows: Where RMSE is the objective function of the calibration model, that is, the distance between the risk after the nth sample decision and the acceptable risk space; ΔR n is the distance between the decision of the nth sample and the risk dynamic equilibrium curve; R n is the risk value at a specific moment after the decision of the nth sample; R max With R min are the upper limit risk and lower limit risk of the risk dynamic balance theory; M is the number of samples.

2. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step 1 also includes the following steps: A1. Construct a space-time coordinate system. A three-dimensional space-time coordinate system is established based on the Frenet coordinate system, with the direction of travel along the road as the S axis, the direction perpendicular to the road as the D axis, and time as the T axis. A2. Calculate spatiotemporal distance, including: obtaining predicted obstacle trajectories, calculating spatial distances that take into account vehicle geometry, calculating equivalent time-based distances that take into account risk anisotropy, and obtaining spatiotemporal distances. A3. Calculate the amount of charge taking into account its own and motion properties. Two representative parameters are selected from the comprehensive attributes of vehicle functions and driving behavior as key influencing factors of charge: braking performance and reaction time; Charge q obs The calculation formula is defined as follows; Where m obs is the mass of the obstacle, v obs is the speed of the vehicle, a max is the maximum deceleration of the vehicle, T r is the vehicle reaction time, i.e., the sum of the driver's reaction time and the vehicle system delay, γ1 and γ2 are unknown coefficients, and exp is an exponential function; A4. Calculate the dielectric constant taking into account anisotropy and environmental complexity. The dielectric constant in the risk field is defined as the product of the anisotropic attenuation coefficient and the environmental attenuation coefficient. The dielectric constant ε is calculated as follows: Where η is the anisotropic attenuation coefficient, is the environmental complexity, which is calculated as follows: η=exp[k·cos(ψ)·(β1·v obs +β2·a obs )]; Where ψ is the angle between the obstacle's direction of travel and d', d'=(s A -s obs ,d A -d obs ) is the position vector of the obstacle pointing to the observation point, v obs and a obs are the speed and acceleration of the vehicle, V is the ratio of the square difference of the speed in the weaving section to that on the normal road, C is the ratio of the lane change ratio in the weaving section to that on the normal road, k, β1 and β2 are unknown coefficients; A5. Calculate the spatial and temporal risk of obstacles. Obstacle space-time risk E obs,A The calculation formula is as follows: Among them, r is the space-time distance; q obs is the charge; ε is the dielectric constant.

3. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step A2 also includes the following steps: A2.

1. Obtain the predicted trajectory of the obstacle. The obstacle trajectory scatter point set P is expressed as: P = {p i ,p i+1 ,…,p m }, where p i Indicates t i The space-time trajectory of the obstacle prediction at each moment, 1≤i≤m, t i The position of the obstacle along the direction of travel of the road and the position perpendicular to the direction of travel of the road; A2.

2. Calculate the spatial distance considering the vehicle's geometric outline. In the vehicle body coordinate system, the center of mass of the vehicle is taken as the origin, the direction along which the vehicle moves is the X axis, and the direction perpendicular to the direction in which the vehicle moves is the Y axis. In the space-time coordinate system, the coordinates of point A are (s A ,d A ), the coordinates of the obstacle mass center are (s obs ,d obs ), in the vehicle body coordinate system, the four vertices of the rectangle T E The coordinates are: (l / 2,w / 2), (l / 2,-w / 2), (-l / 2,w / 2), (-l / 2,-w / 2), In addition, the yaw angle of the vehicle The speed of the vehicle v obs , two types of coordinate systems are transformed using translation and rotation, as follows: The formula for solving the spatial distance d based on the vehicle body coordinate system is shown below; In the formula, (x A ,y A ) is the coordinate of the observation point in the vehicle body coordinate system; (x obs ,y obs ) is the coordinate of the observation point; (x ′ A ,y ′ A ) is the converted coordinate; w is the vehicle width; l is the vehicle length, (d1, d2, d3, d4) are the four vertices of the rectangle T E Compared with the transformed observation point (x A ′,y A ′) spatial distance; A2.

3. Calculate the equivalent time-base distance considering risk anisotropy. Introducing the equivalent time base distance T * , equivalent time base distance T * Meaning: When point A of the interactive object is directly in front of the obstacle's travel path, and point B of the interactive object is not directly in front of the vehicle's travel path, assuming that point A of the interactive object and point B of the interactive object have equal collision risk, then the time distance from the vehicle to point A of the interactive object is equal to the equivalent time distance from the vehicle to point B of the interactive object, the equivalent time base distance T * Used to deal with the problem of unequal risks in different directions, the equivalent time base distance T * It is based on the time dimension weighting of time distance and space-time distance. The specific expression is as follows: In the formula, (x ′ A ,y ′ A ) is the coordinate of the observation point in the vehicle body coordinate system; w is the vehicle width; l is the vehicle length; The four vertices T of the rectangle E The corresponding equivalent time base distance, μ is the conversion parameter; A2.

4. Combining temporal distance and time to obtain spatial-temporal distance, In t i The space-time distance of a moment Solve according to the following formula: In the formula t i The temporal and spatial distance between the observation point and the obstacle at any moment, t i The equivalent time base distance between point A, the closest point to the observation point, and the observation point at the moment t i The point A closest to the observation point on the obstacle trajectory at the moment is denoted as p min , the time required for a vehicle to reach a designated location is called the time-based distance, t min t i Time trajectory scatter point p min The T-axis position of , α is the weight of the time effect; The minimum value of the space-time distance of each scattering point is selected as the space-time distance from point A to the obstacle. The calculation formula is as follows: Where r i t i The temporal and spatial distance between the observation point and the obstacle at any moment, 2<n<m.

4. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step 2 also includes the following steps: B1. Calculate the field strength at the road boundary. Road boundary field strength E lane1,A The calculation formula is as follows: Where W is the road width; d lane1 is the lateral position of the road boundary; d A is the lateral position of the observation point, is the coefficient to be determined; B2. Calculate the field strength of the solid road line. Road solid line field strength E lane2,A The calculation formula is as follows: Where W is the road width; d lane2 is the horizontal position of the solid line of the road; d A is the lateral position of the observation point, is the coefficient to be determined; B3. Calculate the field strength of the dotted line of the road. Field strength E of the road dotted line lane3，A The calculation formula is as follows: Where W is the road width; d lane3 is the horizontal position of the road dotted line; d A is the lateral position of the observation point, is the coefficient to be determined; B4. Calculate the road boundary and road line risk field, Road boundary and road line risk field E lane,A The calculation formula is as follows: AND lane,A =And lane1,A +E lane2,A +E lane3,A ; Among them, E lane1,A is the field strength at the road boundary; E lane2,A is the field strength of the solid road line; E lane3,A is the field strength of the dotted line of the road.

5. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step three also includes the following steps: Field strength E of the on- and off-ramp geo,A The calculation method is as follows: Where s end is the S-axis coordinate of the starting point of the down ramp, s start is the S-axis coordinate of the starting position of the forced lane change area, p A =(s A ,d A ) is the position of point A on the vehicle, P mand is the range of the forced lane change area, σ1 and σ2 are unknown parameters, χ is the set of vehicles on the ramp that intend to merge into the main line, δ is the set of vehicles on the main line that intend to get off the ramp, and γ is the set of remaining vehicles.

6. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step 4 also includes the following steps: In the total risk field, the total risk field calculation formula for any point A is as follows: AND total,A =And obs,A +E lane,A +E geo,A ; Where, E total,A is the total risk field of point A, E obs,A is the field strength exerted by the obstacle on point A, E lane,A is the field strength exerted by the road boundary and road line on point A, E geo,A is the field strength exerted on point A by the on- and off-ramps.

7. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step five also includes the following steps: E1. Extract vehicle trajectory based on YOLOv8. E1.1 Vehicle detection module based on YOLOv8, The vehicle detection module based on YOLOv8 is used to directly predict the location, size and category of the target in a single network; E1.2, Deep-SORT-based target tracking module, The Deep-SORT-based target tracking module is used to describe the target's appearance information and distinguish different vehicles; E1.3, video stabilization processing, Use the After Effects system to pre-process the video data. By tracking multiple fixed buildings in the video, the rotation and displacement information of the entire video is extracted, and then the inverse motion is applied to eliminate the transformation information. E1.4, Model training and testing, Construct a dedicated dataset for vertical view and use the image annotation tool LabelImg for data annotation; E1.5, Coordinate system conversion, By obtaining the position information of road boundaries and road lines, converting them into coordinate information in the Frenet coordinate system, we can obtain the real-world expressway vehicle trajectory.

8. The method for risk assessment of autonomous driving vehicles in an expressway weaving zone according to claim 1 is characterized in that: Step six also includes the following steps: F2. Determine the upper limit risk R of the risk dynamic balance theory max and the lower risk R min ; Calculate the upper limit risk R using a proprietary dataset from a vertical top-down perspective max and the lower risk R min ,The vertical top-view exclusive dataset includes high-risk dataset, low-risk dataset and decision dataset, which contains the trajectories of the observed vehicle and multiple surrounding vehicles. The high-risk dataset includes {R1, R2…R L }Includes collision time assessment, observing vehicles encountering high-risk scenarios, using 3 seconds as the TTC threshold, and the high-risk dataset is used to determine the upper limit risk R max ,On the contrary, the low-risk dataset contains scenarios where the observed vehicle always remains safe according to the judgment of TTC threshold, which is used to determine the optimal risk level R best , upper limit risk R max and the optimal risk level R best The calculation method is as follows, R max =min{R1,R2…R L }; where R m To observe the level of risk experienced by the vehicle after making a decision; Based on R max and R best , determine the lower limit risk R min , the calculation method is as follows, R min =2R best -R max ; F3, using hybrid genetic algorithm and variable neighborhood search algorithm to solve, A hybrid genetic algorithm and a variable neighborhood search algorithm are used to solve the risk field parameters, and according to steps one to three, the driving risk assessment of autonomous vehicles in the expressway weaving area is realized.

Images