A dynamic scene boundary evaluation method based on a hybrid theory of physical mechanism and machine learning

Abstract

The application discloses a dynamic scene boundary evaluation method based on a physical mechanism and a machine learning hybrid theory, and the method comprises the following steps: first, solving the dynamic scene boundary through physical modeling; second, solving the dynamic scene dangerous region and the safe region boundary through a support vector machine; third, solving the dangerous degree boundary of different grades in the dynamic scene dangerous region through a support vector regression; fourth, dynamic scene boundary evaluation based on the physical mechanism and the machine learning hybrid theory; beneficial effects: a complete standard target scene physical model construction process is established, which is convenient for vehicle collision detection and scene danger degree solving. The dynamic scene boundary is solved, which is simple, fast and efficient. The division of the dangerous scene working condition and the safe scene working condition is realized. The division of the scene dangerous region and the safe region according to the training sample is realized.

Description

Technical Field

[0001] This invention relates to a scene boundary assessment method, and more particularly to a dynamic scene boundary assessment method based on a hybrid theory of physical mechanisms and machine learning. Background Technology

[0002] Currently, intelligent connectivity has become one of the mainstream automotive technology development directions, and intelligent connected vehicles represent the future trend. The prerequisite for intelligent connected vehicles to be allowed on the road is that their driving safety has been fully verified and meets relevant standards. To fully test the safety of intelligent connected vehicles and the reliability of their intelligent algorithms, simulation testing, track testing, and road testing must be conducted sequentially. Although these three testing methods differ in their stages, processes, and implementation, they all require the design of specific test scenarios to execute the testing process. Test scenarios should first be challenging to test the intelligent connected vehicle's ability to handle dangerous driving situations. Secondly, their difficulty should be predictable to quantify the level of the intelligent connected vehicle's ability to handle dangerous driving situations. Therefore, it is necessary to determine the boundaries of dynamic scenarios in the test scenario space, such as the boundaries between the danger domain and the safe domain, and the boundaries of different levels of danger within the danger domain. This will allow for the selection of dangerous and known-level test scenarios for testing intelligent connected vehicles, which will greatly improve the testing efficiency and reliability of intelligent connected vehicles and accelerate their deployment.

[0003] Methods for solving dynamic scene boundaries based on physical mechanisms establish a physical model of the target scene, combine vehicle kinematics, and analyze the motion and interaction between the vehicle and surrounding target vehicles. This approach, starting from kinematics and geometry, studies the problem theoretically. Its advantage lies in the relatively accurate dynamic scene boundaries obtained, minimizing misclassification between hazardous and safe zones. However, its disadvantage is that it doesn't consider real-world driving conditions; many scenarios in the hazardous zone have extremely low probability of occurrence on real roads, making testing intelligent connected vehicles with such scenarios less practical. Methods for solving dynamic scene boundaries based on machine learning use real-world driving data from a natural driving database. Employing machine learning algorithms, the boundaries of dynamic scenes are obtained through model self-learning. Its advantage is that it fully considers the probability of the target scene occurring on real roads, thus automatically filtering out many scenarios with extremely low probability of occurrence in the hazardous zone. However, this method heavily relies on the amount and quality of data in the natural driving database; when the data volume or quality is low, the learned scene boundaries may exhibit significant deviations. Physical mechanism-based and machine learning-based methods each have their advantages and can complement each other. Therefore, combining the two to form a hybrid theory based on physical mechanism and machine learning will greatly help solve the problem of scene boundary assessment.

[0004] Chinese patent CN202210941004.4 discloses a method, device, equipment, and storage medium for generating autonomous driving test scenarios. It can generate high-risk autonomous driving test scenarios based on target traffic participant data, but it fails to classify the risk level of these test scenarios or solve for the boundary of the hazard domain of the scenarios. Chinese patent CN202210804060.3 discloses a method, device, vehicle, and storage medium for generating autonomous driving test scenarios. It can generate simulation test scenarios by loading the key parameter ranges of the simulation test scenario into a scenario design document and using a preset script. Chinese patent CN202210741420.X discloses an automated simulation test system and related equipment for intelligent driving. Its scenario generation module can create test scenarios based on the driving parameters of the main vehicle and generalize the test scenarios to generate one or more test scenarios for testing intelligent driving algorithms, but it does not yet know the risk level of these scenarios or the boundary of the entire test scenario space. Summary of the Invention

[0005] The purpose of this invention is to determine the boundaries of dynamic scenarios in the test scenario space, such as the boundary between the danger domain and the safe domain, and the boundary of different levels of danger in the danger domain, by using a hybrid theory of physical mechanisms and machine learning. This lays the foundation for selecting dangerous scenarios with known danger levels to test intelligent connected vehicles, which will greatly improve the testing efficiency and reliability of intelligent connected vehicles and accelerate the deployment process of intelligent connected vehicles.

[0006] To achieve the above objectives, this invention provides a dynamic scene boundary evaluation method based on a hybrid theory of physical mechanisms and machine learning.

[0007] The present invention provides a dynamic scene boundary evaluation method based on a hybrid theory of physical mechanisms and machine learning, the method comprising the following steps:

[0008] The first step is to solve for the dynamic scene boundary through physical modeling. The specific process is as follows:

[0009] Step 1: Establish a physical model of the target scenario. First, determine the type of target scenario to be studied, including entry scenario, exit scenario, following scenario, highway exit scenario, and merging ramp scenario. Then, for the selected target scenario type, further design its operational design domain, that is, clarify what kind of scenario the target scenario is. In scenario research, the "own vehicle" refers to the intelligent connected vehicle to be tested, and the target vehicle refers to other vehicles in the scenario besides the own vehicle. When designing the ODD, determine all static scene elements of the target scenario. Vehicle-related static scene elements include the number and type of target vehicles, the initial... Starting from the initial position, static scene elements related to the lanes include the number and type of lanes, and the lanes to which the current vehicle and the target vehicle belong. Static scene elements related to the environment include the number and type of traffic signs, light intensity, and weather conditions. Finally, based on the interaction method between the current vehicle and surrounding target vehicles as specified by the target scene type, the dynamic scene elements of the target scene are discretized, such as the speed and acceleration of the current vehicle and the target vehicle, the triggering mode of the target scene, the triggering distance, and the triggering time. Arranging and combining the discretized dynamic scene elements yields the scene space of the target scene, thus establishing the physical model of the target scene.

[0010] Step 2: Modeling the trajectories of the vehicle and the target vehicle based on vehicle kinematics. Based on the physical model of the target scene obtained in Step 1, and combining the vehicle kinematics, the trajectories of the vehicle and the target vehicle are calculated. The static scene elements of the target scene contain the initial position information of the vehicle and the target vehicle, while the dynamic scene elements contain the velocity and acceleration sequences of the vehicle and the target vehicle over time. The trajectories of the vehicle and the target vehicle are calculated using the vehicle kinematic model. The modeling methods for the trajectories of the vehicle and the target vehicle are the same, both divided into two approaches: one is to use the center of the vehicle as the reference point to establish trajectory models for the left front contour point, right front contour point, left rear contour point, and right rear contour point; the second is to use one of the left front contour point, right front contour point, left rear contour point, or right rear contour point as the reference point to establish trajectory models for the other contour points. The reference points selected in these two approaches are different, but the modeling principles and methods are consistent. In actual modeling, the reference point chosen for ease of calculation and solution should be selected according to the specific problem.

[0011] Taking the right front contour point of the vehicle as the reference point for modeling as an example, and taking the position of the right front contour point of the vehicle at the initial time t0 as the origin of the coordinate system, the coordinates of the right front contour point after the vehicle has traveled for a period of time to time t1 are expressed as follows:

[0012]

[0013] In the formula, v e (t) represents the speed of the vehicle at time t, θ e(t) represents the heading angle of the vehicle at time t, p ex rf p represents the x-coordinate of the right front profile point of the vehicle at time t1. ey rf This represents the ordinate of the right front profile point of the vehicle at time t1;

[0014] Let the length of this car be a. e Width is b e If the vehicle outline is approximated as a rectangle, then the coordinates of the left front end outline point of the vehicle at time t1 (p ex lf p ey lf )for:

[0015]

[0016] The coordinates of the right rear profile point of the vehicle at time t1 (p ex rr p ey rr )for:

[0017]

[0018] The coordinates of the left rear end profile point of the vehicle at time t1 (p ex lr p ey lr )for:

[0019]

[0020] After solving for the coordinates of each contour point of the vehicle, connecting the coordinates of each contour point of the vehicle at different times during the cutting process will give the vehicle's driving trajectory during the cutting process.

[0021] Step 3: Solve the dynamic scene boundary using the separation axis theorem to obtain the driving trajectories of the vehicle and the target vehicle. Then, further analyze the interaction between the vehicle and the target vehicle using the separation axis theorem to solve the boundary state between the vehicle and the target vehicle. Based on the scene hazard index of the relative distance, relative speed, and relative acceleration between the vehicle and the target vehicle in the boundary state, determine the dynamic scene boundary.

[0022] Based on the type of the entry and target scene, and the established physical and driving trajectory models, the dynamic scene boundary is solved using the separating axis theorem. Assuming vehicle movement is within a two-dimensional plane, the vehicle is modeled as a directed bounding box, considering both its shape and direction of travel. According to the separating axis theorem, for any two separate convex polyhedra, there exists a separating axis such that the two polyhedra are spaced apart on the axis, and their projections onto the separating axis are also separated. For a single directed bounding box, it is sufficient to check whether at most two of its edge direction vectors satisfy the separating axis condition. For two directed bounding boxes, it is sufficient to check whether at most four edge direction vectors satisfy the separating axis condition. If any one of the four edge direction vectors is a separating axis, it is determined that the two directed bounding boxes do not intersect, meaning the vehicles do not collide. Let A represent the target vehicle and B represent the current vehicle. According to the separating axis theorem, if the following relationship is satisfied, it is determined that the two vehicles do not collide:

[0023] |s·l|>d A +d B ,l∈{a u ,a v ,b u ,b v} (5)

[0024] In the formula, s represents the distance vector between the center of the current vehicle and the center of the target vehicle, l represents the direction projection axis of the normalized vector, and a u a v Let b represent the normalized vectors along the two sides of the target vehicle. u b v Let d represent the normalized vectors along the two sides of the vehicle. A d represents the projected length of the target vehicle's center point on the projection axis. B This indicates the projected length of the vehicle's center point onto the projection axis.

[0025] d A It can be obtained through the following formula:

[0026]

[0027] In the formula, These represent the target vehicle at a. u a v The length of the positive half-side in the direction;

[0028] d B It can be obtained through the following formula:

[0029]

[0030] In the formula, These respectively indicate that the vehicle is at b u bv The length of the positive half-side in the direction;

[0031] According to the separation axis theorem, the scene boundary corresponding to the boundary state where the right front contour point of the vehicle cuts into the rear of the target vehicle and does not collide with the right rear contour point of the target vehicle is:

[0032]

[0033] In the formula, t p1 This indicates the moment when the right front contour point of this vehicle and the right rear contour point of the target vehicle are on the same straight line. This moment is calculated by the longitudinal displacement of the trajectory of the right front contour point of this vehicle. o D1 represents the longitudinal velocity of the target vehicle, and D1 represents the initial distance between the current vehicle and the target vehicle at time t0.

[0034] Similarly, the scene boundary corresponding to the boundary state where the left rear contour point of the current vehicle cuts into front of the target vehicle and does not collide with the right front contour point of the target vehicle is:

[0035]

[0036] In the formula, t p2 This indicates the moment when the left rear profile point of this vehicle and the right front profile point of the target vehicle are on the same straight line. This moment is calculated by the longitudinal displacement of the driving trajectory of the left rear profile point of this vehicle.

[0037] Thus, the scene boundary of the target dynamic scene under the set ODD was obtained through physical modeling.

[0038] The second step is to solve for the boundaries of the danger domain and safe domain in the dynamic scene using support vector machines. The specific process is as follows:

[0039] Step 1: Scene Data Acquisition and Preprocessing. The scene data acquisition vehicle is equipped with LiDAR, millimeter-wave radar, GPS high-precision inertial navigation, high-definition camera, vehicle CAN bus, lane line sensor, rain sensor, and light sensor. Data from all sensors is collected according to a fixed acquisition cycle. The collected sensor data includes: spatial 3D point cloud in frames generated by LiDAR, obstacle status list in frames generated by millimeter-wave radar, positioning and attitude data in time series generated by GPS high-precision inertial navigation, color images in frames generated by high-definition camera and lane line sensor, vehicle handling and motion status data in time series generated by vehicle CAN bus, and voltage data in time series generated by rain sensor and light sensor. The specific content of data preprocessing includes: time and spatial alignment of each sensor data; verification of sensor data validity; generation of vehicle bus alignment signal, vehicle status alignment signal, and multimodal environment sensor alignment signal for use in subsequent steps.

[0040] Step 2: Extraction of hazardous and safe scene conditions. Based on the target scene type and its ODD type selected in Step 2 of Step 1, several segments of scene data consistent with the target scene and its ODD are extracted from all preprocessed data by manually watching the video. Each complete segment of target scene data is called a scene condition, representing a complete target scene event that occurs on a real road. In the scene condition extraction stage, the hazard of the scene is characterized by some vehicle state quantities, such as the vehicle's longitudinal speed, longitudinal acceleration, lateral acceleration, and yaw rate in the vehicle bus alignment signal and vehicle state alignment signal obtained in Step 1. Since the same driving operation will lead to different scene hazards at different longitudinal speeds, the vehicle's longitudinal speed is divided into several intervals. The longitudinal acceleration, lateral acceleration, and yaw rate are used as scene hazard indicators to establish hazardous scene condition extraction standards for the above indicators in different vehicle speed intervals.

[0041] Taking the entry scenario and its designed ODD as the target scenario, the longitudinal speed of the vehicle is divided into several intervals. Based on the characteristics of the entry scenario, dangerous operating condition standards for longitudinal acceleration, lateral acceleration, and yaw rate of the vehicle under different longitudinal speed intervals are designed. When any one of the longitudinal acceleration, lateral acceleration, and yaw rate of the vehicle reaches the dangerous operating condition standard at a certain moment in the operating condition, the operating condition is judged as a dangerous operating condition. Thus, all operating conditions are classified into dangerous and safe operating conditions.

[0042] Step 3: Solve the boundary between the hazard and safety regions of the dynamic scene using Support Vector Machine (SVM). First, select a moment that can characterize the hazard of the target scene as the critical moment and study the interaction between the vehicle and the target vehicle at this moment as the basis for dividing the hazard and safety regions of the scene. To improve the classification effect of SVM, instead of using the single vehicle state variable from Step 2 as the scene hazard index, select a comprehensive physical quantity that can integrate several vehicle and target vehicle state variables to characterize the scene hazard, including collision time, headway, and the relative velocity and relative acceleration of the vehicle and the target vehicle. Among them, TTC represents the time required for the two vehicles to maintain their current motion state from the current moment until a collision occurs. The smaller the value, the higher the hazard of the scene. TTC is calculated by the following formula:

[0043]

[0044] In the formula, ΔR represents the relative distance between the two vehicles, and v r The speed of the following vehicle is represented by v. f Indicates the speed of the vehicle in front;

[0045] THW represents the time it takes for the following vehicle to reach the position of the preceding vehicle while maintaining its current state of motion. The smaller the value, the higher the danger of the scenario. THW is calculated using the following formula:

[0046]

[0047] Then, calculate the TTC, THW, and the relative speed and relative acceleration of the vehicle and surrounding vehicles at critical moments for all dangerous and safe scenarios. Finally, input the calculated scenario hazard indicators and scenario hazard labels (i.e., whether the scenario is dangerous or safe) as training samples into the support vector machine algorithm, and output the dynamic boundary between the scenario hazard domain and the safe domain through model self-learning.

[0048] Taking the entry scenario and its designed ODD as the target scenario, and based on the characteristics of the entry scenario, the moment when the center of the vehicle coincides with the lane line is selected as the critical moment. The scenario hazard indicators for all scenario conditions at the critical moment are calculated, including TTC, THW, relative speed between the vehicle and the target vehicle, and relative acceleration between the vehicle and the target vehicle. Let the training sample be (x...). i ,y i ), i = 1,...,n, where n represents the total number of training samples, i.e., the total number of hazardous and safe operating scenarios, x i Let y represent the feature vector of the i-th scenario condition, composed of scenario hazard indicators. iLet represent the hazard label of the i-th scenario, with a hazard label of +1 for hazardous scenarios and a hazard label of -1 for safe scenarios. Let the decision surface equation that divides the hazardous and safe domains be:

[0049] ξx+c=0 (12)

[0050] In the formula, ξ represents the normal vector of the decision surface, which determines the direction of the decision surface; c represents the displacement term, which determines the distance between the decision surface and the origin; and x represents the feature vector of the scenario condition composed of scenario hazard indicators.

[0051] The distance r from the eigenvector x to the decision surface is:

[0052]

[0053] The decision boundary must be able to correctly classify the training samples; therefore, for any training sample, we have:

[0054]

[0055] The training samples closest to the decision boundary on either side are called support vectors. The sum of the distances γ from the two out-of-class support vectors to the decision boundary represents the classification margin.

[0056]

[0057] To find the decision boundary with the best classification performance, we need to maximize the classification margin, which means solving the following equation:

[0058]

[0059] At the same time, the following conditions must be met:

[0060] y i (ξx i +c)≥1,i=1,2,...,n. (17)

[0061] Solving for the optimal decision surface is a convex quadratic programming problem. The dual problem is obtained using the Lagrange multiplier method, by adding a Lagrange multiplier α to each constraint in formula (17). i The Lagrange function is constructed as follows:

[0062]

[0063] Setting the partial derivatives of the Lagrange function L with respect to ξ and c to zero, we get:

[0064]

[0065]

[0066] Therefore, the optimization problem of formula (18) is further transformed into the problem of optimizing the parameter α. i The dual problem of convex quadratic optimization is as follows:

[0067]

[0068] In the formula, α j x j and y j respectively by α i x i and y i Dual results;

[0069] Solving for α, we get i The optimal solution is The optimal solution ξ is obtained through formula (19). * The optimal solution c is obtained through formula (17). * ;

[0070] Therefore, the boundary equation f(x) between the danger zone and the safe zone of the dynamic scene output by the support vector machine is:

[0071] f(x)=ξ * x+c * (twenty two)

[0072] When the training samples are linearly inseparable, the training samples are mapped from the original space to a higher-dimensional feature space, making the training samples linearly separable in this feature space. Let φ(x) represent the feature vector corresponding to the mapping of x to the higher-dimensional feature space. The solution process involves calculating φ(x). i )φ(x j ), that is, x i With x j The inner product after mapping to the feature space is calculated by introducing a kernel function to compute the dot product between any two feature vectors mapped to the high-dimensional space:

[0073]

[0074] Wherein, κ(x) i ,x j () represents a kernel function. Commonly used kernel functions include the following types:

[0075] Linear kernel function:

[0076] κ(x i ,x j )=x i ·x j (twenty four)

[0077] Polynomial kernel function:

[0078] κ(xi ,x j )=(x i ·x j ) d (25)

[0079] Where d is the degree of the polynomial, and d≥1;

[0080] Gaussian kernel function:

[0081]

[0082] Where σ is the bandwidth of the Gaussian kernel function, and σ > 0;

[0083] Laplace kernel function:

[0084]

[0085] Sigmiod kernel function:

[0086] κ(x i ,x j )=tanh(βx i x j +θ) (28)

[0087] Where tanh is the hyperbolic tangent function, β>0, θ<0;

[0088] After introducing the kernel function, it is no longer necessary to directly calculate the inner product in the high-dimensional or even infinite-dimensional feature space. Thus, formula (21) is transformed into:

[0089]

[0090] After solving, the boundary equation F(x) for the dynamic scene's danger and safety regions, output by the support vector machine, is obtained as follows:

[0091]

[0092] The third step is to solve for the hazard boundaries of different levels in the dynamic scene hazard domain using support vector regression. The specific process is as follows:

[0093] Step 1: Pre-divide the hazard boundaries of different levels in the hazard domain according to TTC. In order to obtain accurate and complete scene hazard boundaries, firstly, pre-divide the hazard boundaries of different levels in the hazard domain according to TTC, calculate the TTC at critical moments of all hazardous scene conditions, and formulate the pre-division standard for the hazard boundaries of the hazard domain.

[0094] Taking the entry scenario and its ODD as the target scenario, and based on the characteristics of the entry scenario, the moment when the vehicle's center coincides with the lane line is selected as the critical moment. The TTC of all hazardous scenario conditions at the critical moment is calculated, and the pre-division standard of the hazard domain scenario hazard degree boundary is formulated as follows:

[0095] When TTC∈[0s,1s], the danger level of the scene is a collision scene; when TTC∈(1s,3s], the danger level of the scene is an emergency scene; when TTC∈(3s,5s], the danger level of the scene is a conflict scene; when TTC∈(5s,+∞), due to the large value of TTC, such scenes are judged as safe scenes that have been mistakenly classified as dangerous scenes and are directly discarded.

[0096] Step 2: Select appropriate scenario hazard indicators to construct a joint distribution, calculate the scenario hazard indicators of hazardous scenario conditions at critical moments, including TTC, THW, relative speed between the vehicle and the target vehicle, and relative acceleration between the vehicle and the target vehicle. Combine each scenario hazard indicator in pairs to construct a joint distribution. Combine the pre-division results of the scenario hazard boundary in Step 1 to observe the distribution of hazardous scenario conditions. Select the joint distribution of scenario hazard indicators with better classification effect as the input of the support vector regression algorithm.

[0097] Step 3: Solve for different levels of hazard boundaries in the dynamic scene hazard domain using support vector regression. First, extract the scene boundary points without boundary lines from the joint distribution of the scene hazard indicators with good classification results obtained in Step 2, and use them as training samples. Input them into the support vector regression algorithm. When extracting, pay attention to treating boundary points belonging to the same boundary line as a training sample set and inputting them separately into the support vector regression algorithm to ensure that each training sample set corresponds to only one scene hazard boundary. Finally, summarize all scene hazard boundary lines to obtain the different levels of hazard boundaries in the dynamic scene hazard domain.

[0098] Let the training samples in the training sample set be (t) i ,g i ), i = 1,...,m, where m represents the number of training samples in the training sample set, i.e., the number of boundary points, t i Let g represent the x-coordinate of the i-th boundary point in the joint distribution of scene hazard indices. i Let g represent the ordinate of the i-th boundary point in the joint distribution of scene hazard indicators. The support vector regression algorithm assumes that the model output f(t) and the ordinate of the boundary point g can have a maximum deviation of ε. That is, the loss is calculated only when the absolute value of the difference between f(t) and g is greater than ε. A gap band with a width of 2ε is constructed with f(t) as the center. If the training sample falls into this gap band, the prediction is considered to be correct.

[0099] Let the output f(t) of the support vector regression algorithm be:

[0100] f(t)=ηt+q (31)

[0101] In the formula, η represents the normal vector of the decision surface, which determines the direction of the decision surface; q represents the displacement term, which determines the distance between the decision surface and the origin; and t represents the abscissa of the boundary point in the joint distribution of the scene hazard index.

[0102] Therefore, the problem can be formalized as follows:

[0103]

[0104] In the formula, D represents the regularization constant, l ε The ε-insensitive function is expressed as:

[0105]

[0106] Introducing slack variable δ i and δ i Formula (32) is transformed into:

[0107]

[0108] Introducing the Lagrange multiplier μ i μ i '、α i and α i Construct the Lagrange function:

[0109]

[0110] Let the Lagrangian function L be applied to η, q, δ i and δ i Since the partial derivative of ' is zero, we can obtain:

[0111]

[0112]

[0113] α i +μ i =D (38)

[0114] α′ i +μ′ i =D (39)

[0115] Therefore, the dual problem of formula (35) is:

[0116]

[0117] In the formula, t j,α j α j 'By t respectively i α i and α i 'Dual results;'

[0118] α can be obtained by solving the above formula. i arbitrarily choose 0 < α i The training samples of D are used to solve for q:

[0119]

[0120] Furthermore, q selects multiple conditions that satisfy 0 < α. i The training samples <D are obtained by averaging the results after solving the problem;

[0121] Therefore, the scene hazard boundary equation f(t) output by support vector regression is:

[0122]

[0123] If the training samples are linearly inseparable, the kernel function method introduced in step three of the second step will be used. Assuming that t corresponds to φ(t) after mapping to the high-dimensional feature space, then η will become:

[0124]

[0125] After solving, the scene hazard boundary equation F(t) output by support vector regression is obtained as follows:

[0126]

[0127] Wherein, κ(t) i ,t j )=φ(t i )φ(t j ) represents the kernel function;

[0128] The scene hazard boundaries corresponding to all training sample sets are solved using the above method. These boundaries are then combined with the scene hazard boundaries pre-divided in step one to form the hazard boundaries of different levels in the dynamic scene hazard domain.

[0129] Step 4: Dynamic scene boundary assessment based on a hybrid theory of physical mechanisms and machine learning. The specific process is as follows:

[0130] Step 1: Consistency analysis between dynamic scene hazard boundary based on physical mechanism and dynamic scene hazard boundary based on machine learning. Calculate the scene hazard index at the dynamic scene boundary obtained by physical mechanism and machine learning, and calculate the difference between the scene hazard indexes corresponding to the two methods. Set a consistency standard based on the average difference of all hazard indexes.

[0131] Using the ODD designed for the entry scenario as the target scenario, the scenario boundary is described by establishing a functional relationship between scenario hazard indicators. Consistency analysis compares the degree of difference in the values ​​of each scenario hazard indicator at the dynamic scenario boundary. It calculates the scenario hazard indicators at the dynamic scenario boundary obtained in the first step based on physical mechanisms and the third step based on machine learning, and solves for the degree of difference in each scenario hazard indicator corresponding to the two methods, as well as the average degree of difference in all scenario hazard indicators. Based on the characteristics of the entry scenario, the consistency criteria are set as follows:

[0132] When the average difference of all scenario hazard indicators is less than 15%, the consistency is considered good; when the average difference of all scenario hazard indicators is less than 30%, the consistency is considered average; when the average difference of all scenario hazard indicators is greater than 30%, the consistency is considered poor.

[0133] Step 2: High-confidence dynamic scene boundary generation based on a hybrid theory of physical mechanisms and machine learning. According to the consistency criteria set in Step 1, the high-confidence scene boundary generation will adopt the following principles:

[0134] When the consistency analysis result is good, the offset weights are freely assigned to the dynamic scene boundaries obtained through two solution methods according to the actual needs of the research. That is, when the research problem needs to consider more real road traffic conditions, the high-confidence dynamic boundary is offset more towards the dynamic scene boundary obtained through machine learning; when the research problem focuses on theoretical analysis, the high-confidence dynamic scene boundary is offset more towards the dynamic scene boundary obtained through physical mechanisms. When the consistency analysis result is average, it indicates that the dynamic scene boundary obtained by the machine learning-based solution method has a large deviation due to the small amount or low quality of training sample data. In this case, more offset weights are assigned to the physical mechanism-based method based on the average difference. When the consistency analysis result is poor, it indicates that the dynamic scene boundary obtained by the machine learning-based solution method has a large deviation. In this case, the dynamic scene boundary obtained by the physical mechanism-based method is selected for the high-confidence dynamic scene boundary.

[0135] Step 3: Dynamic scene boundary verification based on hybrid theory of physical mechanism and machine learning. Based on the generation results of high-confidence dynamic scene boundaries, separate test scene libraries are set up for the scene spaces of the three levels of danger in the danger domain and the scene space of the safe domain. Scenes are continuously sampled from these four scene libraries to test intelligent connected vehicles until the accident rate of each scene library converges. If the accident rate of these four scene libraries increases with the increase of scene danger and shows a clear step-like distribution pattern, it indicates that the generated dynamic scene boundary has a good effect.

[0136] The beneficial effects of this invention are:

[0137] The dynamic scene boundary assessment method based on a hybrid theory of physical mechanisms and machine learning provided by this invention obtains the boundaries of dynamic scenes, such as the boundaries between danger and safety domains, and the boundaries of different levels of danger within the danger domain, by solving a hybrid theory based on physical mechanisms and machine learning. This lays the foundation for selecting dangerous scenarios with known danger levels for testing intelligent connected vehicles, which will greatly improve the testing efficiency and reliability of intelligent connected vehicles and accelerate the deployment process of intelligent connected vehicles. Specific beneficial effects are as follows:

[0138] 1) This invention provides a method for establishing a physical model of a target scene. First, the operation design domain is designed according to the type of the target scene. Then, the dynamic scene elements of the target scene are discretized. Finally, the physical model of the target scene is established by permutation and combination, thereby establishing a complete and standard construction process for the physical model of a target scene.

[0139] 2) This invention provides a method for modeling the driving trajectories of the vehicle and the target vehicle based on vehicle kinematics. This method, based on a vehicle kinematics model, solves the driving trajectory of each contour point of the vehicle in the target scene. Compared with ordinary methods for calculating the driving trajectory of the vehicle center, this method is more convenient for vehicle collision detection and solving scene hazard assessment.

[0140] 3) This invention provides a method for solving dynamic scene boundaries using the separating axis theorem. This method simplifies the vehicle as a directed bounding box, detects whether a collision occurs during vehicle interaction using the separating axis theorem, and thus solves for the dynamic scene boundaries. It is simple, fast, and efficient.

[0141] 4) This invention provides a method for extracting hazardous and safe operating conditions. This method uses the vehicle's driving state parameters as indicators and establishes a tiered evaluation standard for hazardous operating conditions based on the characteristics of the target scenario, thus realizing the division between hazardous and safe operating conditions.

[0142] 5) This invention provides a method for solving the boundary between the hazard and safety regions of a dynamic scene using a support vector machine. This method leverages the strength of support vector machines in solving binary classification problems, and solves for the boundary between the hazard and safety regions of a dynamic scene through model self-learning, thus achieving the division of the hazard and safety regions of the scene based on training samples.

[0143] 6) This invention provides a method for constructing a joint distribution by selecting appropriate scenario hazard indicators. This method constructs a joint distribution by combining hazard indicators in pairs, selecting joint distributions from which scenario condition distributions are favorable for classification. This provides effective input for the support vector regression algorithm and a basis for delineating hazard boundaries of different levels within the hazard domain.

[0144] 7) This invention provides a method for solving the hazard boundaries of different levels in the hazard domain of a dynamic scene using support vector regression. This method leverages the strength of support vector regression in machine learning to solve boundary fitting problems, and solves the hazard boundaries of different levels in the hazard domain of a dynamic scene through model self-learning, thus achieving the classification of different hazard levels in the scene hazard domain based on training samples.

[0145] 8) This invention provides a method for consistency analysis of dynamic scene hazard boundaries based on physical mechanisms and dynamic scene hazard boundaries based on machine learning. This method establishes a set of dynamic scene hazard consistency standards based on the average difference of scene hazard indicators and the characteristics of the target scene, thus achieving consistency analysis of dynamic scene boundaries.

[0146] 9) This invention provides a method for generating high-confidence dynamic scene boundaries based on a hybrid theory of physical mechanisms and machine learning. This method uses the results of dynamic scene boundary consistency analysis as a basis, integrates physical mechanism-based and machine learning-based methods, and establishes a hybrid theory of physical mechanisms and machine learning, thereby obtaining high-confidence dynamic scene boundaries.

[0147] 10) This invention provides a method for dynamic scene boundary verification based on a hybrid theory of physical mechanisms and machine learning. This method effectively evaluates the generation effect of dynamic scene boundaries based on a hybrid theory of physical mechanisms and machine learning by sampling scenes of different hazard levels for intelligent connected vehicle testing, using the accident rate converged after a large number of tests as an indicator. The method verifies the dynamic scene boundary generation method based on a hybrid theory of physical mechanisms and machine learning. Attached Figure Description

[0148] Figure 1 This is a schematic diagram of the overall process of the dynamic scene boundary evaluation method described in this invention.

[0149] Figure 2This is a block diagram of the method architecture for the dynamic scene boundary evaluation method described in this invention.

[0150] Figure 3 This is an exemplary embodiment diagram of step one of the first steps of the present invention.

[0151] Figure 4 This is a schematic diagram of an exemplary embodiment of step two of the first step described in this invention.

[0152] Figure 5 This is a schematic diagram of an exemplary embodiment of step three of the first step described in this invention.

[0153] Figure 6 This is an exemplary architecture block diagram of step one of the second step described in this invention.

[0154] Figure 7 This is an exemplary architecture block diagram of step three of the second step described in this invention.

[0155] Figure 8 This is an exemplary calculation result of step two of the third step described in this invention.

[0156] Figure 9 This is an exemplary calculation result of step three of the third step described in this invention. Detailed Implementation

[0157] Please see Figures 1 to 9 As shown:

[0158] The dynamic scene boundary evaluation method based on a hybrid theory of physical mechanisms and machine learning provided by this invention is as follows:

[0159] The first step is to solve the dynamic scene boundary through physical modeling;

[0160] The second step is to solve for the boundaries of the danger domain and the safe domain in the dynamic scene using support vector machines.

[0161] The third step is to solve for the hazard boundaries of different levels in the dynamic scene hazard domain using support vector regression.

[0162] Step 4: Dynamic scene boundary assessment based on hybrid theory of physical mechanisms and machine learning.

[0163] The process of solving the dynamic scene boundary through physical modeling in the first step is as follows:

[0164] Step 1: Establish the physical model of the target scenario. First, determine the type of target scenario to be studied, such as an entry scenario, exit scenario, following scenario, highway exit scenario, merging ramp scenario, etc. Then, for the selected target scenario type, further design its Operational Design Domain (ODD), that is, clarify what kind of scenario the target scenario specifically is. In scenario research, "this vehicle" generally refers to the intelligent connected vehicle to be tested, and "target vehicle" refers to other vehicles in the scenario besides this vehicle. When designing the ODD, all static scene elements of the target scenario should be determined. Vehicle-related static scene elements include the number and type of target vehicles, the initial positions of this vehicle and target vehicles, etc. Lane-related static scene elements include the number and type of lanes, the lanes to which this vehicle and target vehicles belong, etc. Environment-related static scene elements include the number and type of traffic signs, light intensity, and weather conditions, etc. Finally, based on the interaction method between the vehicle and surrounding target vehicles as specified by the target scene type, the dynamic scene elements of the target scene are discretized, such as the speed and acceleration of the vehicle and target vehicles, the triggering mode, triggering distance and triggering time of the target scene, etc. By arranging and combining the discretized dynamic scene elements, the scene space of the target scene can be obtained, and the physical model of the target scene can be established.

[0165] exist Figure 3 The diagram illustrates an exemplary embodiment of step one, where the target scene is the type of the entry scene, and the design is as follows: Figure 3 The ODD shown depicts a scenario where the current vehicle and the target vehicle are in two adjacent lanes. The target vehicle maintains a straight line throughout the process. After the cut-in action is triggered, the current vehicle leaves its lane and enters the target vehicle's lane. A1 represents the initial state when the cut-in action is triggered, and B1 represents the target vehicle's state at that moment. A2 and A3 represent the two boundary states of the current vehicle during the cut-in process, and B2 represents the state of the green car (the target vehicle) at the boundary states. The black dashed lines represent the current vehicle's cut-in trajectory corresponding to the two boundary states during the cut-in process, and the white dashed lines represent the target vehicle's trajectory. The area enclosed by the black dashed lines represents the danger zone where a collision could occur, and the area outside the black dashed lines is the safe zone. This allows for the establishment of a physical model with the cut-in scenario as the target scenario.

[0166] Step Two: Modeling the Trajectories of the Vehicle and the Target Vehicle Based on Vehicle Kinematics. Based on the physical model of the target scene obtained in Step One, the trajectories of the vehicle and the target vehicle are solved by combining the vehicle's kinematic relationships. The static scene elements of the target scene contain the initial position information of the vehicle and the target vehicle, while the dynamic scene elements contain information such as the velocity and acceleration sequences of the vehicle and the target vehicle over time. Therefore, the trajectories of the vehicle and the target vehicle can be calculated by combining the vehicle kinematic model. The modeling methods for the trajectories of the vehicle and the target vehicle are the same, and can be divided into two approaches: one is to use the center of the vehicle as the reference point to establish the trajectory model of the left front contour point, right front contour point, left rear contour point, and right rear contour point; the second is to use one of the left front contour point, right front contour point, left rear contour point, or right rear contour point as the reference point to establish the trajectory model of the other contour points. The reference points selected in these two approaches are different, but the modeling principles and methods are consistent. In actual modeling, the reference point that is easy to calculate and solve can be selected according to the specific problem.

[0167] exist Figure 4 The diagram illustrates an exemplary embodiment of the first step and the second step. In this embodiment, the right front profile point of the vehicle is taken as the reference point, and the position of the right front profile point of the vehicle at the initial time t0 is taken as the origin of the coordinate system. Then, the coordinates of the right front profile point after the vehicle has traveled for a period of time to time t1 can be expressed as follows:

[0168]

[0169] In the formula, v e (t) represents the speed of the vehicle at time t, θ e (t) represents the heading angle of the vehicle at time t, p ex rf p represents the x-coordinate of the right front profile point of the vehicle at time t1. ey rf This represents the ordinate of the right front profile point of the vehicle at time t1.

[0170] Let the length of this car be a. e Width is b e If the vehicle outline is approximated as a rectangle, then the coordinates of the left front end outline point of the vehicle at time t1 (p ex lf p ey lf )for:

[0171]

[0172] The coordinates of the right rear profile point of the vehicle at time t1 (p ex rr p ey rr )for:

[0173]

[0174] The coordinates of the left rear end profile point of the vehicle at time t1 (p ex lr p ey lr )for:

[0175]

[0176] After solving for the coordinates of each contour point of the vehicle, connecting the coordinates of each contour point at different times during the cutting process will give the vehicle's trajectory during the cutting process.

[0177] Step 3: Solve for the dynamic scene boundary using the Separation Axis Theorem. After obtaining the driving trajectories of the vehicle and the target vehicle, the interaction between the two vehicles can be further analyzed using the Separation Axis Theorem to solve for the following: Figure 3 The boundary state between the vehicle and the target vehicle is used to determine the dynamic scene boundary based on the scene hazard indicators such as the relative distance, relative speed, and relative acceleration between the vehicle and the target vehicle.

[0178] exist Figure 5 The diagram illustrates an exemplary embodiment of step three in the first step. The embodiment still uses the entry scenario as the type of the target scenario. Figure 3 The established physical model and Figure 4 Based on the established driving trajectory model, the dynamic scene boundary is solved using the separating axis theorem. Assuming vehicle movement is within a two-dimensional plane, the vehicle is modeled as a directed bounding box, considering both its shape and direction of travel. According to the separating axis theorem, for any two separate convex polyhedra, there exists a separating axis such that the two polyhedra are spaced apart on the axis, and their projections onto the separating axis are also separated. For a single directed bounding box, it is sufficient to check whether at most two of its edge direction vectors satisfy the separating axis condition. For two directed bounding boxes, it is sufficient to check whether at most four edge direction vectors satisfy the separating axis condition. If any one of the four edge direction vectors is a separating axis, it can be determined that the two directed bounding boxes do not intersect, meaning the vehicles do not collide. Figure 5 In this diagram, A represents the target vehicle and B represents the current vehicle. According to the split axle theorem, the two vehicles can be determined not to collide if the following relationship is satisfied:

[0179] |s·l|>d A +d B ,l∈{a u ,a v ,b u ,b v} (5)

[0180] In the formula, s represents the distance vector between the center of the current vehicle and the center of the target vehicle, l represents the direction projection axis of the normalized vector, and a u a v Let b represent the normalized vectors along the two sides of the target vehicle. u b v Let d represent the normalized vectors along the two sides of the vehicle. A d represents the projected length of the target vehicle's center point on the projection axis. B This indicates the projected length of the vehicle's center point on the projection axis.

[0181] d A It can be obtained through the following formula:

[0182]

[0183] In the formula, These represent the target vehicle at a. u a v The length of the positive half-side in the direction.

[0184] d B It can be obtained through the following formula:

[0185]

[0186] In the formula, These respectively indicate that the vehicle is at b u b v The length of the positive half-side in the direction.

[0187] According to the separation axis theorem, we can obtain Figure 3 The scene boundary corresponding to the boundary state where the front right contour point of the vehicle cuts behind the target vehicle and just does not collide with the rear right contour point of the target vehicle is:

[0188]

[0189] In the formula, t p1 This indicates the moment when the right front contour point of the current vehicle is exactly on the same straight line as the right rear contour point of the target vehicle. This moment can be easily calculated by the longitudinal displacement of the trajectory of the right front contour point of the current vehicle. o D1 represents the longitudinal velocity of the target vehicle, and D1 represents the initial distance between the current vehicle and the target vehicle at time t0.

[0190] Similarly, we can obtain Figure 3 The scene boundary corresponding to the boundary state where the left rear contour point of the vehicle cuts in front of the target vehicle and the right front contour point of the target vehicle just does not collide is:

[0191]

[0192] In the formula, tp2 This indicates the moment when the left rear profile point of this vehicle is exactly on the same straight line as the right front profile point of the target vehicle. This moment can be easily calculated by the longitudinal displacement of the trajectory of the left rear profile point of this vehicle.

[0193] Thus, the scene boundary of the target dynamic scene under the set ODD was obtained through physical modeling.

[0194] The process of solving for the boundaries of the danger zone and safe zone in the dynamic scene using support vector machines in the second step is as follows:

[0195] Step 1: Scene Data Acquisition and Preprocessing. The scene data acquisition vehicle is equipped with sensors including LiDAR, millimeter-wave radar, GPS high-precision inertial navigation, high-definition cameras, an onboard CAN bus, lane line sensors, rain sensors, and light sensors. Data from all sensors is collected at a fixed acquisition cycle. The collected sensor data includes: spatial 3D point clouds in frames from LiDAR, obstacle status lists in frames from millimeter-wave radar, positioning and attitude data in time series from GPS high-precision inertial navigation, color images in frames from the high-definition camera and lane line sensors, vehicle handling and motion status data in time series from the onboard CAN bus, and voltage data in time series from the rain and light sensors. Data preprocessing includes: time and spatial alignment of the sensor data; verification of sensor data validity; and generation of onboard bus alignment signals, vehicle status alignment signals, and multimodal environment sensor alignment signals for subsequent steps. Figure 6 An exemplary architecture block diagram of step one of the second step is shown in the figure.

[0196] Step Two: Extraction of Hazardous and Safe Scenarios. Based on the target scenario type and its ODD type selected in Step Two of Step One, multiple segments of scenario data consistent with the target scenario and its ODD are extracted from all preprocessed data by manually viewing the video. Each complete segment of target scenario data is called a scenario condition, representing a complete target scenario event that occurs on a real road. In the scenario condition extraction stage, the hazard of the scenario can be characterized by some vehicle state variables, such as the vehicle's longitudinal speed, longitudinal acceleration, lateral acceleration, and yaw rate in the vehicle bus alignment signal and vehicle state alignment signal obtained in Step One. Since the same driving operation may lead to different scenario hazards at different longitudinal speeds, the vehicle's longitudinal speed is divided into multiple intervals. Using longitudinal acceleration, lateral acceleration, and yaw rate as scenario hazard indicators, a hazardous scenario condition extraction standard is established for the above indicators in different vehicle speed intervals.

[0197] Table 1: Overview of Scenarios Risk Level Indicators

[0198]

[0199] As shown in Table 1, the embodiment still uses the entry scenario and its designed ODD as the target scenario. The embodiment divides the vehicle's longitudinal speed into multiple intervals and, based on the characteristics of the entry scenario, designs hazardous operating condition standards for longitudinal acceleration, lateral acceleration, and yaw rate under different longitudinal speed intervals of the vehicle. When, at a certain moment in the scenario, any one of the vehicle's longitudinal acceleration, lateral acceleration, and yaw rate reaches the hazardous operating condition standard, the scenario is determined to be a hazardous operating condition. In this way, all scenario operating conditions can be classified by hazard level, and hazardous and safe scenario operating conditions can be extracted.

[0200] Step 3: Solve for the boundaries of the hazard and safety domains of the dynamic scene using Support Vector Machines (SVM). First, select a moment that characterizes the hazard of the target scene as the critical moment and study the interaction between the vehicle and the target vehicle at that moment as the basis for dividing the hazard and safety domains. To improve the classification performance of the SVM, instead of using the single vehicle state variable from Step 2 as the hazard indicator, a comprehensive physical quantity that integrates multiple vehicle and target vehicle state variables is selected to characterize the hazard of the scene, such as Time To Collision (TTC), Time Headway (THW), and the relative velocity and relative acceleration between the vehicle and the target vehicle. TTC represents the time required for the two vehicles to maintain their current states of motion from the current moment until a collision occurs; the smaller the value, the higher the hazard of the scene. TTC can be calculated using the following formula:

[0201]

[0202] In the formula, ΔR represents the relative distance between the two vehicles, and v r The speed of the following vehicle is represented by v. f Indicates the speed of the vehicle in front.

[0203] THW represents the time it takes for the following vehicle to reach the position of the preceding vehicle while maintaining its current state of motion. A smaller THW value indicates a higher level of danger in the scenario. THW can be calculated using the following formula:

[0204]

[0205] Then, the TTC, THW, and relative speed and acceleration between the vehicle and surrounding vehicles are calculated at critical moments for all hazardous and safe driving scenarios. Finally, the calculated scenario hazard indicators and scenario hazard labels (i.e., whether the scenario is hazardous or safe) are used as training samples and input into the support vector machine algorithm. Through model self-learning, the dynamic boundaries of the scenario hazard domain and safety domain are output.

[0206] exist Figure 7 The diagram illustrates an exemplary implementation of step two of the second step. The embodiment still uses the entry scenario and its designed ODD as the target scenario. Based on the characteristics of the entry scenario, the moment when the vehicle's center coincides with the lane line is selected as the critical moment. The scenario hazard indicators for all scenario conditions at the critical moment are calculated, including TTC, THW, relative speed between the vehicle and the target vehicle, and relative acceleration between the vehicle and the target vehicle. Let the training sample be (x... i ,y i ), i = 1,...,n, where n represents the total number of training samples, i.e., the total number of hazardous and safe operating scenarios, x i Let y represent the feature vector of the i-th scenario condition, composed of scenario hazard indicators. i Let represent the hazard label of the i-th scenario, with a hazard label of +1 for hazardous scenarios and a hazard label of -1 for safe scenarios. Let the decision surface equation that divides the hazardous and safe domains be:

[0207] ξx+c=0 (12)

[0208] In the formula, ξ represents the normal vector of the decision surface, which determines the direction of the decision surface; c represents the displacement term, which determines the distance between the decision surface and the origin; and x represents the feature vector of the scenario condition composed of scenario hazard indicators.

[0209] The distance r from the eigenvector x to the decision surface is:

[0210]

[0211] The decision boundary should be able to correctly classify the training samples; therefore, for any training sample, we have:

[0212]

[0213] The training samples closest to the decision boundary on either side are called support vectors. The sum of the distances γ from the two out-of-class support vectors to the decision boundary represents the classification margin.

[0214]

[0215] To find the decision boundary with the best classification performance, we need to maximize the classification margin, which means solving the following equation:

[0216]

[0217] At the same time, the following conditions must be met:

[0218] y i (ξx i +c)≥1,i=1,2,...,n. (17)

[0219] Since solving for the optimal decision surface is a convex quadratic programming problem, the Lagrange multiplier method is used to obtain its dual problem. A Lagrange multiplier α is added to each constraint in formula (17). i The Lagrange function is constructed as follows:

[0220]

[0221] Setting the partial derivatives of the Lagrange function L with respect to ξ and c to zero, we get:

[0222]

[0223]

[0224] Therefore, the optimization problem of formula (18) can be further transformed into the problem of optimizing the parameter α. i The dual problem of convex quadratic optimization is as follows:

[0225]

[0226] In the formula, α j x j and y j respectively by α i x i and y i The result of duality.

[0227] Solving for α, we get i The optimal solution is The optimal solution ξ can be obtained through formula (19). * The optimal solution c can be obtained through formula (17). * .

[0228] Therefore, the boundary equation f(x) between the danger zone and the safe zone of the dynamic scene output by the support vector machine is:

[0229] f(x)=ξ * x+c * (twenty two)

[0230] When training samples are linearly inseparable, they can be mapped from the original space to a higher-dimensional feature space, making them linearly separable in this feature space. Let φ(x) represent the eigenvector corresponding to the mapping of x to the higher-dimensional feature space. The solution process involves calculating φ(x). i )φ(x j ), that is, x i With x j The inner product after mapping to the feature space. Since the dimension of the feature space can be very high, or even infinitely so, direct computation is often difficult. To circumvent this obstacle, a kernel function is introduced to calculate the dot product between any two feature vectors mapped to the high-dimensional space:

[0231]

[0232] Wherein, κ(x) i ,x j The ) represents a kernel function. Commonly used kernel functions include the following types:

[0233] Linear kernel function:

[0234] κ(x i ,x j )=x i ·x j (twenty four)

[0235] Polynomial kernel function:

[0236] κ(x i ,x j )=(x i ·x j ) d (25)

[0237] Where d is the degree of the polynomial, and d≥1.

[0238] Gaussian kernel function:

[0239]

[0240] Where σ is the bandwidth of the Gaussian kernel function, and σ > 0.

[0241] Laplace kernel function:

[0242]

[0243] Sigmiod kernel function:

[0244] κ(x i ,x j )=tanh(βx i x j+θ) (28)

[0245] Where tanh is the hyperbolic tangent function, β>0, θ<0.

[0246] After introducing the kernel function, it is no longer necessary to directly calculate the inner product in the high-dimensional or even infinite-dimensional feature space. Therefore, formula (21) can be transformed into:

[0247]

[0248] After solving, the boundary equation F(x) for the dynamic scene's danger region and safe region, output by the support vector machine, can be obtained as follows:

[0249]

[0250] The process of solving for the hazard boundaries of different levels in the dynamic scene hazard domain using support vector regression in the third step is as follows:

[0251] Step 1: Pre-define the hazard boundaries of different levels within the hazard domain based on TTC. TTC, as the most commonly used scenario hazard indicator, can effectively assess the hazard level of simple, low-dimensional scenarios. However, for complex, high-dimensional scenarios, TTC only considers the relative distance and relative speed between the two vehicles, thus failing to provide accurate, comprehensive, and complete scenario hazard boundaries. Furthermore, relying solely on TTC for scenario hazard boundary delineation has a drawback: considering the probability of such scenarios occurring on real roads, some hazardous scenarios with very low TTC values ​​are unlikely to occur in real-world driving. These scenarios are unsuitable for inclusion in the hazard domain for intelligent connected vehicle testing. Therefore, to obtain accurate and complete scenario hazard boundaries, pre-define the hazard boundaries of different levels within the hazard domain based on TTC. Calculate the TTC at critical moments for all hazardous scenario conditions and establish pre-delineation standards for the hazard domain scenario hazard boundaries.

[0252] Table 2: Overview of Scene Hazard Levels

[0253] TTC value Scene danger level [0，1] Crash scenarios (1，3] Emergency scenarios (3，5] Conflict scenarios (5，+∞) Safety scenarios, discard

[0254] As shown in Table 2, the implementation still uses the entry scenario and its ODD as the target scenario. Based on the characteristics of the entry scenario, the moment when the center of the vehicle coincides with the lane line is selected as the critical moment. The TTC of all dangerous scenario conditions at the critical moment is calculated, and the pre-division standard of the danger domain scenario danger degree boundary is formulated as follows: when TTC∈[0s, 1s], the danger level of the scenario is a crash scenario; when TTC∈(1s, 3s], the danger level of the scenario is an emergency scenario; when TTC∈(3s, 5s], the danger level of the scenario is a conflict scenario; when TTC∈(5s, +∞), since the TTC value is large, such scenarios are judged as safe scenarios that have been mistakenly classified as dangerous scenario conditions and are directly discarded.

[0255] Step 2: Select appropriate scenario hazard indicators to construct a joint distribution. Calculate scenario hazard indicators for critical moments in hazardous scenarios, such as TTC, THW, relative speed between the vehicle and the target vehicle, and relative acceleration between the vehicle and the target vehicle. Combine each scenario hazard indicator in pairs to construct a joint distribution. Combine this with the pre-division results of the scenario hazard boundary from Step 1, observe the distribution of hazardous scenario conditions, and select the joint distribution of scenario hazard indicators with better classification performance as the input for the support vector regression algorithm.

[0256] exist Figure 8 The example shown is a calculation result of step two in the third step. The embodiment still takes the entry scenario and the ODD designed for it as the target scenario. According to the characteristics of the entry scenario, the moment when the center of the vehicle coincides with the lane line is selected as the critical moment. The method in step one is used to pre-divide the boundary of different levels of hazard in the hazard domain according to TTC. The hazard indicators of each scenario are combined in pairs to construct a joint distribution. The distribution of hazard scenario conditions is observed. Finally, it is found that the classification effect of hazard scenario conditions is the best in the joint distribution of TTC and relative acceleration between the two vehicles.

[0257] Step 3: Solve for different levels of hazard boundaries in the dynamic scene hazard domain using Support Vector Regression. First, extract the scene boundary points without boundary lines from the joint distribution of the scene hazard indicators with good classification results obtained in Step 2. Use these points as training samples and input them into the Support Vector Regression algorithm. When extracting, be careful to treat boundary points belonging to the same boundary line as a single training sample set and input them separately into the Support Vector Regression algorithm to ensure that each training sample set corresponds to only one scene hazard boundary. Finally, sum up all scene hazard boundary lines to obtain the different levels of hazard boundaries in the dynamic scene hazard domain.

[0258] Since the solution method for support vector regression is basically the same for each training sample set, the following uses a single training sample set as an example to introduce the method of using the support vector regression algorithm to solve for different levels of hazard boundaries in a scene's hazard domain. Let the training samples in the training sample set be (t... i g i ), i = 1, ..., m, where m represents the number of training samples in the training sample set, i.e., the number of boundary points, t i Let g represent the x-coordinate of the i-th boundary point in the joint distribution of scene hazard indices. i Let represent the ordinate of the i-th boundary point in the joint distribution of scene hazard indicators. The support vector regression algorithm assumes that the deviation between the model output f(t) and the ordinate g of the boundary point can be at most ε. That is, the loss is calculated only when the absolute value of the difference between f(t) and g is greater than ε. This is equivalent to constructing a gap band with a width of 2ε centered on f(t). If the training sample falls into this gap band, the prediction is considered correct.

[0259] Let the output f(t) of the support vector regression algorithm be:

[0260] f(t)=ηt+q (31)

[0261] In the formula, η represents the normal vector of the decision surface, which determines the direction of the decision surface; q represents the displacement term, which determines the distance between the decision surface and the origin; and t represents the abscissa of the boundary point in the joint distribution of the scene hazard index.

[0262] Therefore, the problem can be formalized as follows:

[0263]

[0264] In the formula, D represents the regularization constant, l ε The ε-insensitive function is expressed as:

[0265]

[0266] Introducing slack variable δ i and δ i Formula (32) can be transformed into:

[0267]

[0268] Introducing the Lagrange multiplier μ i μ i '、α i and α i Construct the Lagrange function:

[0269]

[0270] Let the Lagrangian function L be applied to η, q, δ i and δ i Since the partial derivative of ' is zero, we can obtain:

[0271]

[0272]

[0273] α i +μ i =D (38)

[0274] α′ i +μ′ i =D (39)

[0275] Therefore, the dual problem of formula (35) is:

[0276]

[0277] In the formula, t j ,α j α j 'By t respectively i α i and α i 'The result of duality'.

[0278] α can be obtained from the above formula. i arbitrarily choose 0 < α i The training samples of D can be used to solve for q:

[0279]

[0280] Furthermore, q can also be achieved by selecting multiple conditions that satisfy 0 < α. i The training samples <D are obtained by averaging the results after solving the problem.

[0281] Therefore, the scene hazard boundary equation f(t) output by support vector regression is:

[0282]

[0283] If the training samples are linearly inseparable, the method of introducing a kernel function in step three of the second step can be used. Assuming that t, after being mapped to the high-dimensional feature space, corresponds to φ(t), then η will become:

[0284]

[0285] After solving, the scene hazard boundary equation F(t) output by support vector regression can be obtained as follows:

[0286]

[0287] Wherein, κ(t) i ,t j )=φ(t i )φ(t j ) represents the kernel function.

[0288] The scene hazard boundaries corresponding to all training sample sets are solved using the above method. These boundaries are then combined with the scene hazard boundaries pre-divided in step one to obtain the different levels of hazard boundaries in the dynamic scene hazard domain.

[0289] exist Figure 9 The example shown is a calculation result of step three in the third step. The embodiment still uses the entry scenario and its designed ODD as the target scenario. Based on the pre-division result of the scenario hazard boundary in step one, the joint distribution of scenario hazard indicators with better classification effect obtained in step two, i.e., the joint distribution of TTC and the relative acceleration of the two vehicles, is equivalent to having four hazard boundaries determined by TTC=0, TTC=1, TTC=3, and TTC=5. However, because the joint distribution contains more scenario hazard indicators, coupled with the distribution characteristics of the scenario conditions themselves, these three hazard boundaries determined by TTC alone cannot completely separate the scenarios in the hazard domain. There are still some scenario boundary points without boundary lines, such as... Figure 8 The boundary points on both sides of the joint distribution map shown are used to further solve for the scene hazard boundaries at the boundary points without boundary lines using support vector regression. The boundary points on both sides of the joint distribution map are treated as separate training sample sets and input into the support vector regression algorithm for learning. Since the boundaries of the joint distribution map are nearly linear, a linear kernel function is chosen to solve for them, thus obtaining the scene hazard boundaries on both sides of the joint distribution map. Combining the obtained hazard boundaries with the hazard boundaries pre-defined according to TTC, we can obtain the hazard boundaries for different levels in this hazard domain as follows:

[0290]

[0291] In the formula, l1 and l r These represent the hazard boundaries on the left and right sides of the joint distribution map obtained through support vector regression, respectively. m1 l m2 l m3 and l m4 These are the hazard boundaries obtained from the TTC pre-classification.

[0292] The process of dynamic scene boundary evaluation based on the hybrid theory of physical mechanisms and machine learning in the fourth step is as follows:

[0293] Step 1: Consistency Analysis of Dynamic Scene Hazard Boundaries Based on Physical Mechanisms and Machine Learning. Calculate the hazard indices for each scene at the dynamic scene boundary obtained from both physical mechanism and machine learning methods, and calculate the degree of difference between the hazard indices for each scene according to the two methods. Set a consistency standard based on the average degree of difference of all hazard indices.

[0294] Table 3: Summary of Consistency Analysis Results

[0295] Average difference Consistency analysis results [0，15％] good (15％，30％] generally (30％，100％] Poor

[0296] As shown in Table 3, the implementation examples still use the entry scenario and its designed ODD as the target scenario. Although the solution methods based on physical mechanisms and machine learning are different, the final solution result, i.e., the expression of the dynamic scene boundary, is basically consistent. Both describe the scene boundary by establishing a functional relationship between scene hazard indicators. Therefore, the consistency analysis compares the degree of difference between the values ​​of each scene hazard indicator at the dynamic scene boundary. The scenario hazard indicators at the dynamic scene boundary obtained in the first step based on physical mechanisms and the third step based on machine learning are calculated. The difference between the scene hazard indicators corresponding to the two methods and the average difference of all scene hazard indicators are solved. Based on the characteristics of the entry scenario, the consistency criteria are set as follows: when the average difference of all scene hazard indicators is less than 15%, the consistency is considered good; when the average difference of all scene hazard indicators is less than 30%, the consistency is considered average; when the average difference of all scene hazard indicators is greater than 30%, the consistency is considered poor.

[0297] The above consistency criteria are determined by the solution principles of the two solution methods: dynamic scene boundary solution methods based on physical mechanisms and machine learning. These reflect two mainstream research ideas in scene research: one is to derive the solution through pure theoretical research, and the other is to learn the rules from real road driving data. The physics-based solution method establishes a physical model of the target scene, combines vehicle kinematics, and analyzes the motion and interaction between the vehicle and surrounding target vehicles. From a kinematic and geometric perspective, it theoretically obtains the hazard boundary of the dynamic scene. Its advantage is that the obtained dynamic scene boundary is relatively accurate, and there is virtually no misclassification between hazard and safe domain scenes. However, its disadvantage is that because it does not consider real-world driving conditions, many scenarios in the hazard domain have an extremely low probability of occurring on real roads, making it not very meaningful to test intelligent connected vehicles using such scenarios. The machine learning-based solution method uses real-world driving data from a natural driving database as its foundation, employing support vector machines and support vector regression algorithms to obtain the hazard boundary of the dynamic scene through model self-learning. Its advantage is that it fully considers the probability of the target scene occurring on real roads, thus automatically filtering out many scenarios with extremely low probability of occurrence in the hazard domain. However, this method is highly dependent on the amount and quality of data in the natural driving database. When the amount of data in the natural driving database is small or the data quality is low, the learned scene boundary may have a large deviation. Therefore, the dynamic scene boundaries obtained by solving these two methods will inevitably have a certain degree of deviation. So a relatively lenient consistency judgment threshold was set. However, under the premise that the two methods are solved correctly, the dynamic scene boundaries obtained by the two methods should be basically consistent in a large range.

[0298] Step 2: High-confidence dynamic scene boundary generation based on a hybrid theory of physical mechanisms and machine learning. Based on the consistency criteria set in Step 1, the following principles can be adopted for generating high-confidence scene boundaries: When the consistency analysis result is good, the offset weights can be freely assigned to the dynamic scene boundaries obtained by the two solution methods according to the actual needs of the research. That is, when the research problem needs to consider more real road traffic conditions, the high-confidence dynamic boundary can be shifted more towards the dynamic scene boundary obtained by machine learning; when the research problem focuses on theoretical analysis, the high-confidence dynamic scene boundary can be shifted more towards the dynamic scene boundary obtained by physical mechanism. When the consistency analysis result is average, it indicates that the dynamic scene boundary obtained by the machine learning-based solution method may have a large deviation due to the small amount or low quality of training sample data. In this case, more offset weights should be assigned to the physical mechanism-based method according to the average difference. When the consistency analysis result is poor, it indicates that the dynamic scene boundary obtained by the machine learning-based solution method has a large deviation. In this case, the high-confidence dynamic scene boundary should be the dynamic scene boundary obtained by the physical mechanism-based method.

[0299] Step 3: Dynamic Scene Boundary Validation Based on a Hybrid Theory of Physical Mechanisms and Machine Learning. Based on the generated high-confidence dynamic scene boundaries, separate test scene libraries are established, comprising scene spaces with three levels of hazard in the hazard domain and a scene space in the safe domain. Scenes are sequentially sampled from these four scene libraries to test the intelligent connected vehicle until the accident rate of each scene library converges. If the accident rate after convergence of these four scene libraries increases with increasing scene hazard and exhibits a clear stepwise distribution, it indicates that the generated dynamic scene boundaries are effective.

Claims

1. A dynamic scene boundary evaluation method based on a hybrid theory of physical mechanisms and machine learning, characterized in that: The method includes the following steps: The first step is to solve for the dynamic scene boundary through physical modeling. The specific process is as follows: Step 1: Establish a physical model of the target scenario. First, determine the type of target scenario to be studied, including entry scenario, exit scenario, following scenario, highway exit scenario, and merging ramp scenario. Then, for the selected target scenario type, further design its operational design domain, that is, clarify what kind of scenario the target scenario is. In scenario research, the "own vehicle" refers to the intelligent connected vehicle to be tested, and the target vehicle refers to other vehicles in the scenario besides the own vehicle. When designing the ODD, determine all static scene elements of the target scenario. Vehicle-related static scene elements include the number and type of target vehicles, and the initial positions of the own vehicle and target vehicles. Static scene elements related to lanes include the number and type of lanes, and the lanes to which the vehicle and the target vehicle belong. Static scene elements related to the environment include the number and type of traffic signs, light intensity, and weather conditions. Finally, based on the interaction method between the vehicle and surrounding target vehicles as specified by the target scene type, the dynamic scene elements of the target scene are discretized. The dynamic scene elements are the speed and acceleration of the vehicle and the target vehicle, the triggering mode of the target scene, the triggering distance, and the triggering time. By arranging and combining the discretized dynamic scene elements, the scene space of the target scene can be obtained, and the physical model of the target scene can be established. Step 2: Modeling the trajectories of the vehicle and the target vehicle based on vehicle kinematics. Based on the physical model of the target scene obtained in Step 1, and combining the vehicle kinematics, the trajectories of the vehicle and the target vehicle are calculated. The static scene elements of the target scene contain the initial position information of the vehicle and the target vehicle, while the dynamic scene elements contain the velocity and acceleration sequences of the vehicle and the target vehicle over time. The trajectories of the vehicle and the target vehicle are calculated using the vehicle kinematic model. The modeling methods for the trajectories of the vehicle and the target vehicle are the same, both divided into two approaches: one is to use the center of the vehicle as the reference point to establish trajectory models for the left front contour point, right front contour point, left rear contour point, and right rear contour point; the second is to use one of the left front contour point, right front contour point, left rear contour point, or right rear contour point as the reference point to establish trajectory models for the other contour points. The reference points selected in these two approaches are different, but the modeling principles and methods are consistent. In actual modeling, the reference point chosen for ease of calculation and solution should be selected according to the specific problem. Taking the right front contour point of the vehicle as the reference point for modeling as an example, and taking the position of the right front contour point of the vehicle at the initial time t0 as the origin of the coordinate system, the coordinates of the right front contour point after the vehicle has traveled for a period of time to time t1 are expressed as follows: （1） In the formula, v e (t) represents the speed of the vehicle at time t, θ e (t) represents the heading angle of the vehicle at time t, p ex rf p represents the x-coordinate of the right front profile point of the vehicle at time t1. ey rf This represents the ordinate of the right front profile point of the vehicle at time t1; Let the length of this car be a. e Width is b e If the vehicle outline is approximated as a rectangle, then the coordinates of the left front end outline point of the vehicle at time t1 (p ex lf p ey lf )for: （2） The coordinates of the right rear profile point of the vehicle at time t1 (p ex rr p ey rr )for: （3） The coordinates of the left rear end profile point of the vehicle at time t1 (p ex lr p ey lr )for: （4） After solving for the coordinates of each contour point of the vehicle, connecting the coordinates of each contour point of the vehicle at different times during the cutting process will give the vehicle's driving trajectory during the cutting process. Step 3: Solve the dynamic scene boundary using the separation axis theorem to obtain the driving trajectories of the vehicle and the target vehicle. Then, further analyze the interaction between the vehicle and the target vehicle using the separation axis theorem to solve the boundary state between the vehicle and the target vehicle. Based on the scene hazard index of the relative distance, relative speed, and relative acceleration between the vehicle and the target vehicle in the boundary state, determine the dynamic scene boundary. Taking the entry scene as the target scene type, based on the established physical model and driving trajectory model, the dynamic scene boundary is solved using the separating axis theorem. Assuming vehicle movement is within a two-dimensional plane, the vehicle is modeled as a directed bounding box, considering both its shape and direction of travel. According to the separating axis theorem, for any two separate convex polyhedra, there exists a separating axis such that the two polyhedra are spaced apart on the axis, and their projections onto the separating axis are also separated. For a single directed bounding box, it is sufficient to check whether at most two of its edge direction vectors satisfy the separating axis condition. For two directed bounding boxes, it is sufficient to check whether at most four edge direction vectors satisfy the separating axis condition. If any one of the four edge direction vectors is a separating axis, it is determined that the two directed bounding boxes do not intersect, meaning the vehicles do not collide. Let A represent the target vehicle and B represent the current vehicle. According to the separating axis theorem, if the following relationship is satisfied, it is determined that the two vehicles do not collide: （5） In the formula, s represents the distance vector between the center of the current vehicle and the center of the target vehicle, l represents the direction projection axis of the normalized vector, and a u a v Let b represent the normalized vectors along the two sides of the target vehicle. u b v Let d represent the normalized vectors along the two sides of the vehicle. A d represents the projected length of the target vehicle's center point on the projection axis. B This indicates the projected length of the vehicle's center point onto the projection axis. d A It can be obtained through the following formula: （6） In the formula, h u A h v A These represent the target vehicle at a. u a v The length of the positive half-side in the direction; d B It can be obtained through the following formula: （7） In the formula, h u B h v B These respectively indicate that the vehicle is at b u b v The length of the positive half-side in the direction; According to the separation axis theorem, the scene boundary corresponding to the boundary state where the right front contour point of the vehicle cuts into the rear of the target vehicle and does not collide with the right rear contour point of the target vehicle is: （8） In the formula, t p1 This indicates the moment when the right front contour point of this vehicle and the right rear contour point of the target vehicle are on the same straight line. This moment is calculated by the longitudinal displacement of the trajectory of the right front contour point of this vehicle. o D1 represents the longitudinal velocity of the target vehicle, and D1 represents the initial distance between the current vehicle and the target vehicle at time t0. Similarly, the scene boundary corresponding to the boundary state where the left rear contour point of the current vehicle cuts into front of the target vehicle and does not collide with the right front contour point of the target vehicle is: （9） In the formula, t p2 This indicates the moment when the left rear profile point of this vehicle and the right front profile point of the target vehicle are on the same straight line. This moment is calculated by the longitudinal displacement of the driving trajectory of the left rear profile point of this vehicle. Thus, the scene boundary of the target dynamic scene under the set ODD was obtained through physical modeling. The second step is to solve for the boundaries of the danger domain and safe domain in the dynamic scene using support vector machines. The specific process is as follows: Step 1: Scene Data Acquisition and Preprocessing. The scene data acquisition vehicle is equipped with LiDAR, millimeter-wave radar, GPS high-precision inertial navigation, high-definition camera, vehicle CAN bus, lane line sensor, rain sensor, and light sensor. Data from all sensors is collected according to a fixed acquisition cycle. The collected sensor data includes: spatial 3D point cloud in frames generated by LiDAR, obstacle status list in frames generated by millimeter-wave radar, positioning and attitude data in time series generated by GPS high-precision inertial navigation, color images in frames generated by high-definition camera and lane line sensor, vehicle handling and motion status data in time series generated by vehicle CAN bus, and voltage data in time series generated by rain sensor and light sensor. The specific content of data preprocessing includes: time and spatial alignment of each sensor data; verification of sensor data validity; generation of vehicle bus alignment signal, vehicle status alignment signal, and multimodal environment sensor alignment signal for use in subsequent steps. Step 2: Extraction of hazardous and safe scene conditions. Based on the target scene type and its ODD type selected in Step 2 of Step 1, several segments of scene data consistent with the target scene and its ODD are extracted from all preprocessed data by manually watching the video. Each complete segment of target scene data is called a scene condition, representing a complete target scene event that occurs on a real road. In the scene condition extraction stage, the hazard of the scene is characterized by some vehicle state variables. The vehicle state variables are the longitudinal speed, longitudinal acceleration, lateral acceleration, and yaw rate of the vehicle in the vehicle bus alignment signal and vehicle state alignment signal obtained in Step 1. Since the same driving operation will lead to different scene hazards at different longitudinal speeds, the longitudinal speed of the vehicle is divided into several intervals. The longitudinal acceleration, lateral acceleration, and yaw rate are used as scene hazard indicators to establish hazardous scene condition extraction standards for the above indicators in different vehicle speed intervals. Taking the entry scenario and its designed ODD as the target scenario, the longitudinal speed of the vehicle is divided into several intervals. Based on the characteristics of the entry scenario, dangerous operating condition standards for longitudinal acceleration, lateral acceleration, and yaw rate of the vehicle under different longitudinal speed intervals are designed. When any one of the longitudinal acceleration, lateral acceleration, and yaw rate of the vehicle reaches the dangerous operating condition standard at a certain moment in the operating condition, the operating condition is judged as a dangerous operating condition. Thus, all operating conditions are classified into dangerous and safe operating conditions. Step 3: Solve the boundary between the hazard and safety regions of the dynamic scene using Support Vector Machine (SVM). First, select a moment that can characterize the hazard of the target scene as the critical moment and study the interaction between the vehicle and the target vehicle at this moment as the basis for dividing the hazard and safety regions of the scene. To improve the classification effect of SVM, instead of using the single vehicle state variable from Step 2 as the scene hazard index, select a comprehensive physical quantity that can integrate several vehicle and target vehicle state variables to characterize the scene hazard, including collision time, headway, and the relative velocity and relative acceleration of the vehicle and the target vehicle. Among them, TTC represents the time required for the two vehicles to maintain their current motion state from the current moment until a collision occurs. The smaller the value, the higher the hazard of the scene. TTC is calculated by the following formula: （10） In the formula, ∆R represents the relative distance between the two vehicles, and v r v represents the speed of the following vehicle. f Indicates the speed of the vehicle in front; THW represents the time it takes for the following vehicle to reach the position of the preceding vehicle while maintaining its current state of motion. The smaller the value, the higher the danger of the scenario. THW is calculated using the following formula: （11） Then, calculate the TTC, THW, and the relative speed and relative acceleration of the vehicle and surrounding vehicles at critical moments for all dangerous and safe scenarios. Finally, input the calculated scenario hazard indicators and scenario hazard labels (i.e., whether the scenario is dangerous or safe) as training samples into the support vector machine algorithm, and output the dynamic boundary between the scenario hazard domain and the safe domain through model self-learning. Taking the entry scenario and its designed ODD as the target scenario, and based on the characteristics of the entry scenario, the moment when the center of the vehicle coincides with the lane line is selected as the critical moment. The scenario hazard indicators for all scenario conditions at the critical moment are calculated, including TTC, THW, relative speed between the vehicle and the target vehicle, and relative acceleration between the vehicle and the target vehicle. Let the training sample be (x...). i , y i ), i=1,...,n, where n represents the total number of training samples, i.e., the total number of hazardous and safe operating scenarios, x i Let y represent the feature vector of the i-th scenario condition, composed of scenario hazard indicators. i Let represent the hazard label of the i-th scenario, with a hazard label of +1 for hazardous scenarios and a hazard label of -1 for safe scenarios. Let the decision surface equation that divides the hazardous and safe domains be: （12） In the formula, ξ represents the normal vector of the decision surface, which determines the direction of the decision surface; c represents the displacement term, which determines the distance between the decision surface and the origin; and x represents the feature vector of the scenario condition composed of scenario hazard indicators. The distance r from the eigenvector x to the decision surface is: （13） The decision boundary must be able to correctly classify the training samples; therefore, for any training sample, we have: （14） The training samples closest to the decision boundary on either side are called support vectors. The sum of the distances φ from the two out-of-class support vectors to the decision boundary represents the classification margin. （15） To find the decision boundary with the best classification performance, we need to maximize the classification margin, which means solving the following equation: （16） At the same time, the following conditions must be met: （17） Solving for the optimal decision surface is a convex quadratic programming problem. The dual problem is obtained using the Lagrange multiplier method, by adding a Lagrange multiplier α to each constraint in formula (17). i The Lagrange function is constructed as follows: （18） Setting the partial derivatives of the Lagrange function L with respect to ξ and c to zero, we get: （19） （20） Therefore, the optimization problem of formula (18) is further transformed into the problem of the parameter α. i The dual problem of convex quadratic optimization is as follows: （21） In the formula, α j x j and y j respectively by α i x i and y i Dual results; Solving for α, we get i The optimal solution is α i * Then the optimal solution ξ obtained by formula (19) is ξ. * The optimal solution c is obtained through formula (17). * ; Therefore, the boundary equation f(x) between the danger zone and the safe zone of the dynamic scene output by the support vector machine is: （22） When the training samples are linearly inseparable, the training samples are mapped from the original space to a higher-dimensional feature space, making the training samples linearly separable in this feature space. Let φ(x) represent the feature vector corresponding to the mapping of x to the higher-dimensional feature space. The solution process involves calculating φ(x). i )φ(x j ), that is, x i With x j The inner product after mapping to the feature space is calculated by introducing a kernel function to compute the dot product between any two feature vectors mapped to the high-dimensional space: （23） Wherein, κ(x) i , x j () represents a kernel function. Commonly used kernel functions include the following types: Linear kernel function: （24） Polynomial kernel function: （25） Where d is the degree of the polynomial, and d≥1; Gaussian kernel function: （26） Where σ is the bandwidth of the Gaussian kernel function, and σ > 0; Laplace kernel function: （27） Sigmiod kernel function: （28） Where tanh is the hyperbolic tangent function, β>0, θ<0; After introducing the kernel function, it is no longer necessary to directly calculate the inner product in the high-dimensional or even infinite-dimensional feature space. Thus, formula (21) is transformed into: （29） After solving, the boundary equation F(x) for the dynamic scene's danger and safety regions, output by the support vector machine, is obtained as follows: （30）； The third step is to solve for the hazard boundaries of different levels in the dynamic scene hazard domain using support vector regression. The specific process is as follows: Step 1: Pre-divide the hazard boundaries of different levels in the hazard domain according to TTC. In order to obtain accurate and complete scene hazard boundaries, firstly, pre-divide the hazard boundaries of different levels in the hazard domain according to TTC, calculate the TTC at critical moments of all hazardous scene conditions, and formulate the pre-division standard for the hazard boundaries of the hazard domain. Taking the entry scenario and its designed ODD as the target scenario, and based on the characteristics of the entry scenario, the moment when the vehicle's center coincides with the lane line is selected as the critical moment. The TTC of all hazardous scenario conditions at the critical moment is calculated, and the pre-division standard of the hazard domain scenario hazard degree boundary is formulated as follows: When TTC∈[0s,1s], the danger level of the scene is a collision scene; when TTC∈(1s,3s], the danger level of the scene is an emergency scene; when TTC∈(3s,5s], the danger level of the scene is a conflict scene; when TTC∈(5s,+∞), due to the large value of TTC, such scenes are judged as safe scenes that have been mistakenly classified as dangerous scenes and are directly discarded. Step 2: Select appropriate scenario hazard indicators to construct a joint distribution, calculate the scenario hazard indicators of hazardous scenario conditions at critical moments, including TTC, THW, relative speed between the vehicle and the target vehicle, and relative acceleration between the vehicle and the target vehicle. Combine each scenario hazard indicator in pairs to construct a joint distribution. Combine the pre-division results of the scenario hazard boundary in Step 1 to observe the distribution of hazardous scenario conditions. Select the joint distribution of scenario hazard indicators with better classification effect as the input of the support vector regression algorithm. Step 3: Solve for different levels of hazard boundaries in the dynamic scene hazard domain using support vector regression. First, extract the scene boundary points without boundary lines from the joint distribution of the scene hazard indicators with good classification results obtained in Step 2, and use them as training samples. Input them into the support vector regression algorithm. When extracting, pay attention to treating boundary points belonging to the same boundary line as a training sample set and inputting them separately into the support vector regression algorithm to ensure that each training sample set corresponds to only one scene hazard boundary. Finally, summarize all scene hazard boundary lines to obtain the different levels of hazard boundaries in the dynamic scene hazard domain. Let the training samples in the training sample set be (t) i ,g i ), i=1,...,m, where m represents the number of training samples in the training sample set, i.e., the number of boundary points, t i Let g represent the x-coordinate of the i-th boundary point in the joint distribution of scene hazard indices. i Let g represent the ordinate of the i-th boundary point in the joint distribution of scene hazard indicators. The support vector regression algorithm assumes that the model output f(t) and the ordinate of the boundary point g can have a maximum deviation of φ, that is, the loss is calculated only when the absolute value of the difference between f(t) and g is greater than φ. A gap band with a width of 2φ is constructed with f(t) as the center. If the training sample falls into this gap band, the prediction is considered to be correct. Let the output f(t) of the support vector regression algorithm be: (31) In the formula, 𝜂 represents the normal vector of the decision surface, which determines the direction of the decision surface; q represents the displacement term, which determines the distance between the decision surface and the origin; t represents the abscissa of the boundary point in the joint distribution of the scene hazard index. Therefore, the problem can be formalized as follows: (32) In the formula, D represents the regularization constant, l 𝜀 The ε-insensitive function is expressed as: (33) Introducing slack variable δ i and δ i ’ Formula (32) is transformed into: (34) Introducing Lagrange multipliers 𝜇 i , i ’ , i and 𝛼 i ’ Construct the Lagrange function: （35）； Let the Lagrangian function L be applied to , q, δ i and δ i ’ Since the partial derivative is zero, we can obtain: (36) (37) (38) (39) Therefore, the dual problem of formula (35) is: (40) In the formula, t j , j , j ’ Each by t i , i and 𝛼 i ’ Dual results; Solving the above formula yields 𝛼 i Choose any value 0 < 𝛼 i The training samples of D are used to solve for q: (41) Furthermore, q selects multiple conditions that satisfy 0 < 𝛼 i The training samples <D are obtained by averaging the results after solving the problem; Therefore, the scene hazard boundary equation f(t) output by support vector regression is: (42) If the training samples are linearly inseparable, the method of introducing a kernel function in step three of the second step will be adopted. Assuming that t corresponds to φ(t) after being mapped to the high-dimensional feature space, then 𝜂 will become: (43) After solving, the scene hazard boundary equation F(t) output by support vector regression is obtained as follows: (44) Wherein, κ(t) i , t j )= φ(t i )φ(t j ) represents the kernel function; The scene hazard boundaries corresponding to all training sample sets are solved using the above method. These boundaries are then combined with the scene hazard boundaries pre-divided in step one to form the hazard boundaries of different levels in the dynamic scene hazard domain. Step 4: Dynamic scene boundary assessment based on a hybrid theory of physical mechanisms and machine learning. The specific process is as follows: Step 1: Consistency analysis between dynamic scene hazard boundary based on physical mechanism and dynamic scene hazard boundary based on machine learning. Calculate the scene hazard index at the dynamic scene boundary obtained by physical mechanism and machine learning, and calculate the difference between the scene hazard indexes corresponding to the two methods. Set a consistency standard based on the average difference of all hazard indexes. Taking the entry scenario and its designed ODD as the target scenario, the scenario boundary is described by establishing a functional relationship between scenario hazard indicators. Consistency analysis compares the degree of difference between the values ​​of each scenario hazard indicator at the dynamic scenario boundary. The scenario hazard indicators at the dynamic scenario boundary obtained in the first step based on physical mechanisms and the third step based on machine learning are calculated. The degree of difference between the scenario hazard indicators corresponding to the two methods and the average degree of difference of all scenario hazard indicators are solved. Based on the characteristics of the entry scenario, the consistency criteria are set as follows: When the average difference of all scenario hazard indicators is [0, 15%], the consistency is considered good; when the average difference of all scenario hazard indicators is (15%, 30%), the consistency is considered moderate; when the average difference of all scenario hazard indicators is (30%, 100%), the consistency is considered poor. Step 2: High-confidence dynamic scene boundary generation based on a hybrid theory of physical mechanisms and machine learning. According to the consistency criteria set in Step 1, the high-confidence scene boundary generation will adopt the following principles: When the consistency analysis result is good, the offset weights are freely assigned to the dynamic scene boundaries obtained through two solution methods according to the actual needs of the research. That is, when the research problem needs to consider more real road traffic conditions, the high-confidence dynamic boundary is offset more towards the dynamic scene boundary obtained through machine learning; when the research problem focuses on theoretical analysis, the high-confidence dynamic scene boundary is offset more towards the dynamic scene boundary obtained through physical mechanisms. When the consistency analysis result is average, it indicates that the dynamic scene boundary obtained by the machine learning-based solution method has a large deviation due to the small amount or low quality of training sample data. In this case, more offset weights are assigned to the physical mechanism-based method based on the average difference. When the consistency analysis result is poor, it indicates that the dynamic scene boundary obtained by the machine learning-based solution method has a large deviation. In this case, the dynamic scene boundary obtained by the physical mechanism-based method is selected for the high-confidence dynamic scene boundary. Step 3: Dynamic scene boundary verification based on hybrid theory of physical mechanism and machine learning. Based on the generation results of high-confidence dynamic scene boundaries, separate test scene libraries are set up for the scene spaces of the three levels of danger in the danger domain and the scene space of the safe domain. Scenes are continuously sampled from these four scene libraries to test intelligent connected vehicles until the accident rate of each scene library converges. If the accident rate of these four scene libraries increases with the increase of scene danger and shows a clear step-like distribution pattern, it indicates that the generated dynamic scene boundary has a good effect.

Images