A method for real-time evaluation of the complexity of continuous test scenarios for autonomous vehicles

Abstract

The present invention belongs to the technical field of autonomous driving vehicles, and specifically, is a method for real-time evaluation of the complexity of continuous test scenarios of autonomous driving vehicles. The method comprises the following steps: step 1, establishing a scenario complexity evaluation framework; step 2, analyzing the feature weights of scenario elements; step 3, calculating the element-system complexity mapping relationship; step 4, synthesizing the scene complexity features obtained in steps 2 and 3 to obtain the real-time complexity evaluation results of the continuous test scenario. The present invention can evaluate the complexity of the surrounding environment in real time according to the vehicle driving process, and solves the non-real-time problem existing in the existing test scenario complexity evaluation method.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving vehicles, and specifically, is a method for real-time evaluation of complexity of continuous test scenarios of autonomous driving vehicles. Background Art

[0002] Continuous test scenarios are an important part of the scenario testing system for autonomous vehicles. Testing autonomous vehicles based on continuous test scenarios is an important means to discover performance defects of autonomous vehicles. However, due to the different behaviors of vehicles in continuous scenarios, the autonomous vehicles under test may encounter different tasks, which has led to much controversy over the fairness of continuous test scenarios. Existing methods mostly focus on independent test scenarios, and rarely consider the complexity evaluation methods of long-term driving processes in continuous test scenarios. Summary of the invention

[0003] The present invention provides a method for real-time evaluation of the complexity of continuous test scenarios of autonomous driving vehicles, which can evaluate the complexity of the surrounding environment in real time according to the vehicle driving process, and solves the non-real-time problem existing in the existing test scenario complexity evaluation method.

[0004] The technical solution of the present invention is described as follows in conjunction with the accompanying drawings:

[0005] A method for real-time evaluation of complexity of continuous test scenarios of an autonomous driving vehicle comprises the following steps:

[0006] Step 1: Establish a scenario complexity assessment framework;

[0007] Establish a real-time evaluation framework for the complexity of continuous test scenarios for autonomous vehicles, which is divided into three parts: scenario element feature weight analysis, element-system complexity mapping, and multi-system scenario complexity coupling;

[0008] Step 2: Analyze the feature weights of scene elements;

[0009] The importance of the types and states of different scene elements to the vehicle and traffic is analyzed one by one, and the corresponding scene element feature weights are established;

[0010] Step 3: Calculate the factor-system complexity mapping relationship;

[0011] Calculate the complexity mapping relationship between scene elements and autonomous driving vehicle subsystems from the perspective of perception, decision-making, and execution subsystems;

[0012] Step 4: Couple the scene complexity;

[0013] The complexity features of the scenarios obtained in steps 2 and 3 are combined to obtain the real-time complexity evaluation results of the continuous test scenarios.

[0014] Furthermore, the specific method of step one is as follows:

[0015] 11) Based on the six-layer model of the autonomous driving test scenario, the scene element categories are extracted, and the characteristic attribute weights of the scene elements are obtained through the hierarchical analysis method, which serves as the basis for the subsequent scene complexity evaluation process for the autonomous driving system;

[0016] 12) Combine the characteristic values ​​of scene elements with their impact on the system to establish the complexity mapping relationship between scene elements or element combinations and perception, control and execution systems;

[0017] 13) Calculate the complexity weight system of each subsystem of autonomous driving based on the influence transfer relationship, and couple to obtain the comprehensive evaluation results of the complexity of continuous test scenarios.

[0018] Furthermore, the specific method of step 2 is as follows:

[0019] 21) Based on the six-layer scene model, a three-layer scene element framework of static scene elements, dynamic scene elements and weather elements was established;

[0020] 22) The impact of scene elements on the driving process of autonomous vehicles is divided into two parts: one is the importance of scene elements in this category of elements; the other is the impact of scene element attributes on scene complexity; S is used to represent the feature weight score of scene elements that interact during the driving process of autonomous vehicles, as shown in formula (1):

[0021] (1)

[0022] In the formula, is the importance weight of the scene element in this category of elements; The attribute characteristics of scene elements are determined by using the hierarchical analysis method to determine the static scene elements and dynamic scene elements. w i , divided into five levels, with corresponding values ​​of 1-5 points; subscript i Indicates i Parameters or values ​​of scene elements; subscript j Represents different feature levels.

[0023] Furthermore, u j Calculation is performed using the extension analytic hierarchy process;

[0024] First, we construct the importance relationship judgment matrix of the scene element category set A =( a pq ) nj×nj ,in p ,q =1,2,…, n j ; n j For the j The number of scene element categories in each classification level; a pq =[ a + pq , a - pq ] is the number of extension intervals; a + pq and a - pq are the upper and lower endpoints of the interval caused by different cognition and preference; find the upper decision matrix composed of the upper endpoints of the interval respectively A + and the lower decision matrix consisting of the lower endpoints of the interval A - The corresponding normalized eigenvector x + and x - , and calculate the coefficient k and m , as shown in formulas (2) and (3):

[0025] (2)

[0026] (3)

[0027] When 0≤k≤1≤m is satisfied, it means that the extension decision matrix is ​​consistent; thus, we can find j The weight vector of the extension interval number of classification levels R k =[ kx - , mx + ]=( R 1, R 2,…, R nj ) T ; Then calculate the relative importance of the weight vectors of each interval number, using V ( R a ≥ R b )express R a Greater than R b possibility; if a , b =1,2,…, n j and a ≠ b , V ( R a ≥ R b )≥0, then the important weight coefficient ranking is obtained u a and u b , as shown in formula (4):

[0028] (4)

[0029] After normalization, the important weight coefficient of the element in the jth classification level is

[0030] ; , , , are the upper and lower endpoints of the extension interval number weight vector of element a and element b respectively; , are the extension interval number weight vectors of element a and element b respectively, It is k The total number of layer features.

[0031] Furthermore, the three-layer scene element framework is specifically as follows:

[0032] The first-level classification is divided into: static scene elements, dynamic scene elements and weather elements;

[0033] The secondary classification is divided into: road layer, traffic facility layer, temporary operation layer, target layer and natural environment layer;

[0034] The three-level classification is divided into: road type, road quality, number of lanes, road markings, signal lights, signs, other facilities, traffic accidents, temporary parking, road construction, motor vehicles, non-motor vehicles, pedestrians, light, rain, snow and fog;

[0035] Among them, when the first-level classification is static scene elements and the second-level classification is road layer, the third-level classification is road type, road quality, number of lanes and road markings;

[0036] When the first-level classification is static scene elements and the second-level classification is the traffic facility layer, the third-level classification is traffic lights, signboards, and other facilities;

[0037] When the first-level classification is static scene elements and the second-level classification is temporary operation layer, the third-level classification is traffic accidents, temporary parking and road construction;

[0038] When the first-level classification is dynamic scene elements and the second-level classification is the target layer, the third-level classification is motor vehicles, non-motor vehicles and pedestrians;

[0039] When the first level classification is weather elements and the second level classification is natural environment layers, the third level classification is light, rain, snow and fog.

[0040] Furthermore, the specific method of step three is as follows:

[0041] 31) Map the complexity of perception system elements; select three sensor types: camera, millimeter wave radar, and lidar to calculate the impact of scene elements on the perception system;

[0042] 32) Map the complexity of decision system elements; quantify the complexity mapping results between scene elements and decision systems using the concept of decision planning margin, which refers to the area within which an autonomous vehicle can plan to drive in a short period of time; the decision planning margin calculation process is divided into two parts: vehicle driving potential field and vehicle reachable domain. The former is the impact of the surrounding traffic environment on the driving decision of the car, and the latter is the driving range caused by the dynamic limitations of the autonomous vehicle itself;

[0043] 33) Map the complexity of the execution system; the execution system involves the movement of the vehicle itself and the interaction between the tire and the road. Refer to the backup tire force as the execution system complexity mapping E implement indicators.

[0044] Furthermore, the specific method of step 31) is as follows:

[0045] 311) For the camera system, in addition to the characteristics of the camera itself, the vehicle's motion state is considered, including two factors: relative distance and relative speed;

[0046] For relative distance interference, the relative distance between the target scene element and the vehicle is equivalent to the distribution of the target scene element in the image at a certain moment, and then the recognition difficulty of the camera system is characterized; as shown in formula (5):

[0047] (5)

[0048] In the formula, I dis It is the ratio of the actual size of the target scene element in the image captured by the camera sensor at a certain moment to the overall imaging result. The lower the ratio, the greater the impact of the actual size of the scene element at the current distance on the sensor recognition effect. Ptarget is the actual size of the target scene element; P a is the field of view of the camera sensor at the target scene element position at the current moment; d is the relative distance between the target scene element and the vehicle under test; FOV is the field of view of the camera sensor; for continuous scene elements, I dis Obtained by calculating the ratio of the visible length in the image to the image size;

[0049] For relative velocity disturbance, the elastic coefficient is introduced ε , analyze the difficulty of identifying dynamic scene elements that are close to the tested autonomous vehicle and moving at high speed;

[0050] Taking into account the interference of relative distance and relative speed, the total impact of the calculated scene elements on the camera sensor is shown in formula (6):

[0051] (6)

[0052] In the formula, I c is the complexity mapping result of scene elements to camera perception system, I c The larger the value, the more complex the scene element is for the camera perception system;

[0053] 312) For the millimeter wave radar system perception process, the coefficient is introduced The degree of impact is quantified as shown in formula (7): (7)

[0054] In the formula, The larger the value, the more complex the scene element is for the camera perception system; n is the number of close-range moving targets within the detection range of the millimeter-wave radar;

[0055] 313) For the lidar system, the point cloud density obtained by approximate fitting D s relationship to relative distance; and using coefficients based on the results I r1 The influence of relative distance on the lidar perception process is quantified as shown in formula (8):

[0056] (8)

[0057] In the formula, e is a natural constant;

[0058] Introduction coefficient I r2 The influence of the number of close-range moving targets within the detection range of the laser radar on the laser radar is shown in formula (9):

[0059] (9)

[0060] In the formula, n is the number of close-range moving targets within the laser radar detection range;

[0061] Taking into account the influence of the relative distance of scene elements and the occlusion of close-range targets, the total influence coefficient of scene elements on the lidar sensor is obtained as shown in formula (10):

[0062] (10)

[0063] In the formula, I r is the complexity mapping result of scene elements to camera perception system, I r The larger the value, the more complex the scene element is for the camera perception system;

[0064] 314) Comprehensively analyze the quantitative impact of weather factors on sensor perception characteristics, approximate the effects of different weather factors on vehicle sensors, and obtain the complexity mapping results of weather factors I w ;

[0065] When calculating the weighted scores of weather element characteristics, the assignment process of important weights is omitted, and only the specific scores are calculated based on the influence coefficients. S i(weather) , as shown in formula (11):

[0066] (11)

[0067] The influence of various scene elements on the camera, millimeter-wave radar, and lidar is comprehensively considered to obtain the complexity mapping relationship between scene elements and perception systems. E perception The result is shown in formula (12):

[0068] (12)

[0069] In the formula, T i For scene elements i Perceived System β The total duration captured; S i(weather)β For the perception system βThe weight score of the weather element characteristics that have an impact; if the scene element i The elements are identified by multiple types of perception systems through perception fusion, and the complexity of the perception system corresponding to the elements E perception Take the average value according to the number of fusion sensor categories; The impact of scene elements on different sensors, including cameras, lidar, and millimeter-wave radar, The values ​​of include c, m, r; The multiplication symbol.

[0070] Furthermore, the specific method of step 32) is as follows:

[0071] 321) Calculate the vehicle driving potential field;

[0072] The vehicle's operable space is calculated using a potential field model, which includes two parts: static scene element potential field and dynamic scene element potential field.

[0073] Static scene elements are divided into scene static objects and constraint facilities;

[0074] Among them, static objects in the scene have collision risks for autonomous vehicles, so they have a greater impact on traffic safety in the decision-making and planning process; C pi The potential field representing the static elements of the scene is calculated as shown in formula (13):

[0075] (13)

[0076] In the formula, S i Score the weights of scene element features; r i is the relative distance between the autonomous driving vehicle and the scene elements; k 1 is a fixed constant;

[0077] Constraint facilities tend to be constrained by traffic regulations and have a weak correlation with serious risks such as loss of life and property. Therefore, they have a relatively small impact on traffic safety in the decision-making and planning process. Constraint facilities are classified into speed constraint elements and position constraint elements, and potential field models are constructed for each. C dsi The potential field representing the position constraint elements such as lane lines is calculated as shown in formula (14):

[0078] (14)

[0079] use C dviThe potential field representing speed limit signs, traffic lights and other speed constraint elements is calculated as shown in formula (15):

[0080] (15)

[0081] In the formula, D i is the lane width; x i is the vertical position of the target scene element; v i is the driving speed of the autonomous vehicle; v di is the speed corresponding to the constraint element; k 2 , k 3 , k 4 is a fixed constant; Traffic regulations restrict the maximum speed of vehicles.

[0082] When the scene element is a red traffic light, the speed limit is considered to be 0 m / s ,When the scene element is a green light, the speed limit is considered to be the current speed of the vehicle;

[0083] For dynamic scene elements, they are considered to be static objects in the scene with motion characteristics. The potential field representation of dynamic scene elements is shown in formula (16):

[0084] (16)

[0085] In the formula, v i is the moving speed of the target scene element; θ is the angle between the autonomous driving vehicle and the direction of motion of the scene element, i Refers to the sequence number of the scene element; k 5. k 6 is a fixed constant;

[0086] Superimpose the potential fields of static and dynamic scene elements to obtain the total potential field of the current scene F potential , as shown in formula (17):

[0087] (17)

[0088] In the formula, is a unified expression of the potential field of all elements in the scene, for , , The classification symbol of i is the counting number of different scene elements; the scene element potential field is classified into static scene element potential field and dynamic scene element potential field, and the static scene element potential field is further subdivided into scene static material potential field and constraint element potential field; among them, the scene static material potential field and the dynamic scene element potential field are described by the same expression, To characterize the potential field of these two types of elements C pi Classification corner mark; Refers to scene elements with position constraints such as lane lines C Dsi Classification corner mark; It refers to speed-restricted scene elements such as speed limit. C Dvi Classification corner mark;

[0089] 322) Calculate the reachable area of ​​the vehicle;

[0090] The vehicle reachable domain refers to the road area that the autonomous vehicle can reach in the short term due to vehicle dynamics and decision-making tasks. It includes two parts: the vehicle's operable space and the vehicle's decision-making event area.

[0091] The maneuverable space of the vehicle determined by dynamic constraints includes longitudinal and lateral parts. The minimum safe braking distance of a vehicle at a certain speed in the longitudinal following theory is defined as the longitudinal limit distance of the maneuverable space of the autonomous driving vehicle, as shown in formula (18):

[0092] (18)

[0093] In the formula, a is the maximum deceleration of the autonomous driving vehicle, a =10 m / s 2 ; v is the instantaneous speed of the vehicle before braking;

[0094] Considering the maximum wheel turning angle of the vehicle, the maximum static front wheel turning angle β =40° as the vehicle turning limit, thereby obtaining lateral maneuverable space;

[0095] Combined with the longitudinal and lateral operating limits, the theoretical operable space of the autonomous driving vehicle is calculated. During the calculation process, the influence of the position changes of the surrounding scene elements, especially the dynamic scene elements, on the potential field of the scene elements during this time period is considered, and the position offset of the dynamic scene elements during the time when the vehicle decelerates to a standstill and is also generated by braking at the maximum deceleration is selected as the final potential field center of the dynamic elements, and the subsequent decision margin is calculated.

[0096] The decision-making event area refers to the range of surrounding vehicles that the autonomous vehicle needs to consider in the short-term decision-making process. In the same scene, the vehicle's decision-making task area changes with the vehicle's driving process and is related to the relative distance between the target scene elements and the vehicle. TTC That is, the collision time is less than 2 s As the basis for judging dangerous scenes, select 2 s As the time boundary, the decision event area of ​​the autonomous driving vehicle is defined as the radius with the front of the vehicle as the center R =2| v |The semicircular range of;

[0097] Combine the vehicle's operable space and the vehicle's pending decision event area to obtain the reachable area of ​​the autonomous driving vehicle based on the actual road boundary F permission ;

[0098] 323) Calculate vehicle decision planning margin;

[0099] Set the danger level Rick Less than or equal to r The potential field area is used as the safe driving area of ​​the autonomous vehicle. The overlapping area with the vehicle reachable domain represents the area where the autonomous vehicle can achieve a safe driving path at the current moment, that is, the decision planning margin of the vehicle. E dicision ;

[0100] E dicision The calculation process is shown in formula (19):

[0101] (19)

[0102] In the formula, T is the total driving time of the vehicle.

[0103] Furthermore, the specific method of step 33) is as follows:

[0104] 331) The available reserve tire force is simplified to the product of the current vehicle acceleration and the weight score of the road quality factor characteristics. E implement , as shown in formula (20):

[0105] (20)

[0106] In the formula, S rp Score the characteristic weights of road quality elements in the road layer; v is the speed of the vehicle being tested.

[0107] Furthermore, the specific method of step 4 is as follows:

[0108] 41) The influence transmission relationship between perception, decision-making and execution systems is used as the weight coefficient The coupling process of the total scene complexity is introduced as shown in formula (21):

[0109] (twenty one)

[0110] In the formula, l i is the transfer coefficient, l 1 =3, l 2 =2, l 3 =1;

[0111] 42) Finally, the real-time complexity evaluation result of the continuous test scenario is obtained, as shown in formula (22):

[0112] (twenty two).

[0113] The beneficial effects of the present invention are:

[0114] The present invention can evaluate the interaction between the tested autonomous driving vehicle and the surrounding static obstacles and dynamic obstacles in real time according to the vehicle driving process, thereby generating real-time driving task complexity, solving the defect of the existing test scene complexity evaluation method that only relies on the scene initial value for complexity calculation, and ensuring the fairness of the test results in continuous test scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0116] Figure 1 This is a schematic diagram of the scene complexity assessment framework;

[0117] Figure 2 It is a schematic diagram of the relationship curve between point cloud density and relative distance;

[0118] Figure 3 Schematic diagram of the influence transfer relationship of autonomous driving subsystems. DETAILED DESCRIPTION

[0119] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0120] See also Figure 1 This embodiment provides a method for real-time evaluation of complexity of continuous test scenarios of an autonomous driving vehicle, comprising the following steps:

[0121] Step 1: Establish a scenario complexity assessment framework;

[0122] A real-time evaluation framework for the complexity of continuous test scenarios for autonomous vehicles is established, which is divided into three parts: scenario element feature weight analysis, element-system complexity mapping, and multi-system scenario complexity coupling, as follows:

[0123] 11) Based on the six-layer model of the autonomous driving test scenario, the scene element categories are extracted, and the characteristic attribute weights of the scene elements are obtained through the hierarchical analysis method, which serves as the basis for the subsequent scene complexity evaluation process for the autonomous driving system;

[0124] 12) Combine the characteristic values ​​of scene elements with their impact on the system to establish the complexity mapping relationship between scene elements or element combinations and perception, control and execution systems;

[0125] 13) Calculate the complexity weight system of each subsystem of autonomous driving based on the influence transfer relationship, and couple to obtain the comprehensive evaluation results of the complexity of continuous test scenarios.

[0126] Step 2: Analyze the feature weights of scene elements;

[0127] The importance of the types and states of different scene elements to the vehicle and traffic is analyzed one by one, and the corresponding scene element feature weights are established, as follows:

[0128] 21) Based on the six-layer scenario model, the interactive relationship between different scenario elements and the continuous driving process of autonomous driving is analyzed, and a three-layer scenario element framework of static scenario elements (road layer, infrastructure layer, temporary operation layer), dynamic scenario elements (target layer) and weather elements (natural environment layer) is re-established, as shown in Table 1.

[0129] Table 1 Classification of scene elements

[0130] 22) The impact of scene elements on the driving process of autonomous vehicles is divided into two parts. One part is the importance of scene elements in this category of elements, such as the importance of motor vehicles in the target layer; the other part is the impact of scene element attributes on scene complexity, such as the type of motor vehicle is a car or a truck. S is used to represent the weight score of scene element features that interact during the driving process of autonomous vehicles, as shown in formula (1): (1)

[0131] In the formula, is the importance weight of the scene element in this category of elements; The attribute characteristics of scene elements are determined by using the hierarchical analysis method to determine the static scene elements and dynamic scene elements. w i , divided into five levels, with corresponding values ​​of 1-5 points; subscript i Indicates i The parameter or value of a scene element, for example, the sign is a speed limit sign of 80km / h; j Represents different feature levels, such as static scene features (first layer), road layer (second layer), and number of lanes (third layer) in a table.

[0132] in, u j Calculation is performed using the extension analytic hierarchy process;

[0133] First, we construct the importance relationship judgment matrix of the scene element category set A =( a pq ) nj×nj ,in p , q =1,2,…, n j ; n j For the j The number of scene element categories in each classification level; a pq =[ a + pq , a - pq ] is the number of extension intervals; a + pq and a - pq are the upper and lower endpoints of the interval caused by different cognition and preference; find the upper decision matrix composed of the upper endpoints of the interval respectively A + and the lower decision matrix consisting of the lower endpoints of the interval A- The corresponding normalized eigenvector x + and x - , and calculate the coefficient k and m , as shown in formulas (2) and (3):

[0134] (2)

[0135] (3)

[0136] When 0≤k≤1≤m is satisfied, it means that the extension decision matrix has good consistency; thus, we can find j The weight vector of the extension interval number of classification levels R k =[ kx - , mx + ]=( R 1, R 2,…, R nj ) T ; Then calculate the relative importance of the weight vectors of each interval number, using V ( R a ≥ R b )express R a Greater than R b possibility; if a , b =1,2,…, n j and a ≠ b , V ( R a ≥ R b )≥0, then the important weight coefficient ranking is obtained u a and u b , as shown in formula (4):

[0137] (4)

[0138] After normalization, the important weight coefficient of the element in the jth classification level is ; , , , are the upper and lower endpoints of the extension interval number weight vector of element a and element b respectively; , are the extension interval number weight vectors of element a and element b respectively, It is k The total number of layer features.

[0139] Step 3: Calculate the factor-system complexity mapping relationship;

[0140] The complexity mapping relationship between scene elements and autonomous driving vehicle subsystems is calculated from the perspectives of the three subsystems of perception, decision-making, and execution, as follows:

[0141] 31) Mapping the complexity of perception system elements; select three sensor types: camera, millimeter wave radar, and lidar to calculate the impact of scene elements on the perception system;

[0142] 311) Camera perception system: For the camera system, in addition to the characteristics of the camera itself, the main considerations include the vehicle's motion state, including relative distance and relative speed.

[0143] For relative distance interference, the present invention equates the relative distance between the target scene element and the vehicle to the distribution of the target scene element in the image at a certain moment, and then characterizes the recognition difficulty of the camera system, as shown in formula (5):

[0144] (5)

[0145] In the formula, I dis It is the ratio of the actual size of the target scene element in the image captured by the camera sensor at a certain moment to the overall imaging result. The lower the ratio, the greater the impact of the actual size of the scene element at the current distance on the sensor recognition effect. P target is the actual size of the target scene element; P a is the field of view of the camera sensor at the target scene element position at the current moment; d is the relative distance between the target scene element and the vehicle under test; FOV is the field of view of the camera sensor; for continuous scene elements, I dis It is obtained by calculating the ratio of the visible length in the image to the image size.

[0146] For relative speed interference, the main consideration is the effect of image motion blur. The image blur length generated by scene elements in three-dimensional space can be expressed as a function of the camera's own parameters and the relative speed, relative distance, and motion direction angle between the target scene element and the camera. The elastic coefficient is introduced in this invention. ε , analyze the difficulty of identifying dynamic scene elements that are close to the tested autonomous vehicle and moving at high speed;

[0147] Taking into account the interference of relative distance and relative speed, the total impact of the scene elements on the camera sensor calculated by the present invention is shown in formula (6):

[0148] (6)

[0149] In the formula, I c The result of mapping the complexity of scene elements to the camera perception system; I c The larger the value, the more complex the scene element is for the camera perception system;

[0150] 312) Millimeter wave radar perception system;

[0151] For the perception process of the millimeter-wave radar system, the target measurement error is mainly related to the fixed parameters of the radar system itself, such as the frequency modulation bandwidth, frequency modulation linearity, and frequency modulation period, and is less related to the relative speed and relative distance of the target. However, when the number and density of targets increase, the number of false alarm targets of the radar will surge; therefore, the impact of close-range multiple targets on the detection accuracy of the millimeter-wave radar is mainly calculated. When the longitudinal relative distance of multiple moving targets within the detection range of the millimeter-wave radar is less than 5 m Or the horizontal relative distance is less than 2 m When , it is easy to affect the millimeter wave radar perception process and mark it as a close-range target. Introducing the coefficient I m The degree of impact is quantified as shown in formula (7):

[0152] (7)

[0153] In the formula, I m The larger the value, the more complex the scene element is for the camera perception system; n It is the number of close-range moving targets within the detection range of the millimeter-wave radar.

[0154] 313) Lidar;

[0155] For the LiDAR system, its perception results are less affected by scene elements and the relative speed between the vehicle and the vehicle, so the main consideration is the impact of relative distance and target occlusion on the LiDAR.

[0156] As the relative distance between the target scene elements and the vehicle gradually increases, the distribution point cloud will become increasingly sparse, and the contours of distant traffic targets will gradually become incomplete, which will affect the detection accuracy of the lidar. The point cloud density obtained by approximate fitting D s The relationship with relative distance is represented as Figure 2 shown.

[0157] based on Figure 2 The result using the coefficient I r1 The influence of relative distance on the lidar perception process is quantified as shown in formula (8), where e is a natural constant.

[0158] (8)

[0159] When the traffic volume is large, there may be a situation where some vehicles in the far lane in a frame or multiple frames of point cloud data are blocked by other targets in the near lane, which will cause the loss of some point cloud data of the target scene elements obtained by the laser radar, thereby affecting the target perception and recognition. Similar to the millimeter wave radar, the present invention introduces the coefficient I r2 The influence of the number of close-range moving targets within the detection range of the laser radar on the laser radar is shown in formula (9):

[0160] (9)

[0161] In the formula, n is the number of close-range moving targets within the laser radar detection range;

[0162] Taking into account the influence of the relative distance of scene elements and the occlusion of close-range targets, the total influence coefficient of scene elements on the lidar sensor is obtained as shown in formula (10):

[0163] (10)

[0164] In the formula, I r The result of mapping the complexity of scene elements to the camera perception system; I r The larger the value, the more complex the scene element is for the camera perception system;

[0165] 314) Weather element complexity mapping and feature weights;

[0166] Millimeter-wave radar has good adaptability to most weather conditions, is less affected by clutter, and has strong anti-interference capabilities for general bad weather. However, the impact of extreme weather such as heavy rain and snow on the detection performance of millimeter-wave radar still needs to be considered.

[0167] LiDAR is not affected by changes in lighting, but the sensor's performance will be greatly limited in adverse weather conditions such as rainy, snowy, and foggy days.

[0168] The present invention comprehensively analyzes the quantitative impact of weather factors on sensor perception characteristics, approximates the impact of different weather factors on vehicle sensors, and obtains the complexity mapping results of weather factors. I w As shown in Table 2.

[0169] Table 2 Distribution of weather factor influence coefficients

[0170] Since the weather scene element categories are relatively independent and almost never appear in the same scene at the same time, the assignment process of important weights is omitted when calculating the weather element feature weight score, and only the specific score is calculated by combining the influence coefficient in Table 2 S i(weather) , as shown in formula (11):

[0171] (11)

[0172] The influence of various scene elements on the camera, millimeter-wave radar, and lidar is comprehensively considered to obtain the complexity mapping relationship between scene elements and perception systems. E perception The result is shown in formula (12):

[0173] (12)

[0174] In the formula, T i For scene elements i Perceived System β The total duration captured; S i(weather)β For the perception system β The weight score of the weather element characteristics that have an impact; if the scene element i Recognition is performed by multiple types of perception systems through perception fusion. The perception system complexity corresponding to this factor is E perception Take the average value according to the number of fusion sensor categories; The impact of scene elements on different sensors, including cameras, lidar, and millimeter-wave radar, The values ​​of include c, m, r; The multiplication symbol.

[0175] 32) Map the complexity of the decision system elements; use the concept of decision planning margin to quantify the complexity mapping results between the scene elements and the decision system. The decision planning margin refers to the area where the autonomous driving vehicle can be planned to drive in a short time; the decision planning margin calculation process is divided into two parts: the vehicle driving potential field and the vehicle reachable domain. The former is the impact of the surrounding traffic environment on the driving decision of the car, and the latter is the driving range caused by the dynamic limitations of the autonomous driving vehicle itself, as follows:

[0176] 321) Calculate the vehicle driving potential field;

[0177] The vehicle's operable space is mainly calculated through the potential field model, which includes two parts: the static scene element potential field and the dynamic scene element potential field.

[0178] Static scene elements can be subdivided into scene static objects and constraint facilities.

[0179] Static objects in the scene, such as traffic accidents and temporary parking in the temporary control layer, pose collision risks to autonomous vehicles, so they have a greater impact on traffic safety in the decision-making and planning process. C pi The potential field representing the static elements of the scene is calculated as shown in formula (13):

[0180] (13)

[0181] In the formula, S i Score the weights of scene element features; r i is the relative distance between the autonomous driving vehicle and the scene elements; k 1 is a fixed constant;

[0182] (14)

[0183] (15)

[0184] In the formula, D i is the lane width; x i is the vertical position of the target scene element; v i is the driving speed of the autonomous vehicle; v di is the speed corresponding to the constraint element; k 2 , k 3, k 4 is a fixed constant; Traffic regulations restrict the maximum speed of vehicles.

[0185] For dynamic scene elements, they can be considered as static objects in the scene with motion characteristics. Referring to formula (13), the potential field representation of dynamic scene elements is shown in formula (16):

[0186] (16)

[0187] In the formula, v i is the moving speed of the target scene element; θ is the angle between the autonomous driving vehicle and the direction of motion of the scene element, i Refers to the sequence number of the scene element; k 5. k 6 is a fixed constant;

[0188] Superimpose the potential fields of static and dynamic scene elements to obtain the total potential field of the current scene F potential , as shown in formula (17):

[0189] (17)

[0190] In the formula, is a unified expression of the potential field of all elements in the scene, for , , The classification symbol of i is the counting number of different scene elements; the scene element potential field is classified into static scene element potential field and dynamic scene element potential field, and the static scene element potential field is further subdivided into scene static material potential field and constraint element potential field; among them, the scene static material potential field and the dynamic scene element potential field are described by the same expression, To characterize the potential field of these two types of elements C pi Classification corner mark; Refers to scene elements with position constraints such as lane lines C Dsi Classification corner mark; It refers to speed-restricted scene elements such as speed limit. C Dvi Classification corner mark;

[0191] 322) Calculate the reachable area of ​​the vehicle;

[0192] The vehicle reachable domain refers to the road area that the autonomous vehicle can reach in the short term due to vehicle dynamics and decision-making tasks. It includes two parts: the vehicle's operable space and the vehicle's decision-making event area.

[0193] The maneuverable space of the vehicle determined by dynamic constraints includes longitudinal and lateral parts. The minimum safe braking distance of a vehicle at a certain speed in the longitudinal following theory is defined as the longitudinal limit distance of the maneuverable space of the autonomous driving vehicle, as shown in formula (18):

[0194] (18)

[0195] In the formula, a is the maximum deceleration of the autonomous driving vehicle, a =10 m / s 2 ; v is the instantaneous speed of the vehicle before braking;

[0196] Considering the maximum wheel turning angle of the vehicle, the maximum static front wheel turning angle β =40° as the vehicle turning limit, thereby obtaining lateral maneuverable space;

[0197] Combined with the longitudinal and lateral operating limits, the theoretical operable space of the autonomous driving vehicle is calculated. During the calculation process, the influence of the position changes of the surrounding scene elements, especially the dynamic scene elements, on the potential field of the scene elements during this time period is considered, and the position offset of the dynamic scene elements during the time when the vehicle decelerates to a standstill and is also generated by braking at the maximum deceleration is selected as the final potential field center of the dynamic elements, and the subsequent decision margin is calculated.

[0198] The decision-making event area refers to the range of surrounding vehicles that the autonomous vehicle needs to consider in the short-term decision-making process. In the same scene, the vehicle's decision-making task area changes with the vehicle's driving process and is related to the relative distance between the target scene elements and the vehicle. TTC (Time To Collision) means the collision time is less than 2 s As the basis for judging dangerous scenes, select 2 s As the time boundary, the decision event area of ​​the autonomous driving vehicle is defined as the radius with the front of the vehicle as the center R =2| v |The semicircular range of;

[0199] Combine the vehicle's operable space and the vehicle's pending decision event area to obtain the reachable area of ​​the autonomous driving vehicle based on the actual road boundary F permission ;

[0200] 323) Calculate vehicle decision planning margin;

[0201] Set the danger level Rick Less than or equal to r The potential field area is used as the safe driving area of ​​the autonomous vehicle. The overlapping area with the vehicle reachable domain represents the area where the autonomous vehicle can achieve a relatively safe driving path at the current moment, that is, the decision planning margin of the vehicle. E dicision ;

[0202] E dicision The calculation process is shown in formula (19):

[0203] (19)

[0204] In the formula, T is the total driving time of the vehicle.

[0205] 33) Map the complexity of the execution system; the execution system involves the movement of the vehicle itself and the interaction between the tire and the road. Refer to the backup tire force as the execution system complexity mapping E implement The indicators are as follows:

[0206] 331) The available reserve tire force is simplified to the product of the current vehicle acceleration and the weight score of the road quality factor characteristics. E implement , as shown in formula (20):

[0207] (20)

[0208] In the formula, S rp Score the characteristic weights of road quality elements in the road layer, v is the speed of the vehicle being tested.

[0209] Step 4: Couple the scene complexity;

[0210] The complexity features of the scenarios obtained in steps 2 and 3 are combined to obtain the real-time complexity evaluation results of the continuous test scenarios, as follows:

[0211] Due to the inherent logical relationship between the perception system, decision-making system and execution system in the autonomous driving system, the mapping relationship between the scene and the complexity of each subsystem is not independent of each other. The scene elements indirectly affect the lower subsystem through the upper subsystem. The influence relationship transmission rules are as follows: Figure 3 shown.

[0212] 41) The influence transmission relationship between perception, decision-making and execution systems is used as the weight coefficient The coupling process of the total scene complexity is introduced as shown in formula (21):

[0213] (twenty one)

[0214] In the formula, l i is the transfer coefficient, l 1 =3, l 2 =2, l 3 =1;

[0215] 42) Finally, the real-time complexity evaluation result of the continuous test scenario is obtained, as shown in formula (22):

[0216] (twenty two).

[0217] In summary, the present invention can evaluate the complexity of the surrounding environment in real time according to the vehicle driving process, thereby solving the non-real-time problem existing in the existing test scene complexity evaluation method.

[0218] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for real-time evaluation of complexity of continuous test scenarios for autonomous vehicles, characterized in that: The following steps are involved: Step 1: Establish a scenario complexity assessment framework; Establish a real-time evaluation framework for the complexity of continuous test scenarios for autonomous vehicles, which is divided into three parts: scenario element feature weight analysis, element-system complexity mapping, and multi-system scenario complexity coupling; Step 2: Analyze the feature weights of scene elements; The importance of the types and states of different scene elements to the vehicle and traffic is analyzed one by one, and the corresponding scene element feature weights are established, as follows: 21) Based on the six-layer scene model, a three-layer scene element framework of static scene elements, dynamic scene elements and weather elements is established; 22) The impact of scene elements on the driving process of autonomous vehicles is divided into two parts: one is the important weight of scene elements in this category of elements; the other is the impact of scene element attributes on scene complexity; S is used to represent the feature weight score of scene elements that interact during the driving process of autonomous vehicles, as shown in formula (1): In the formula, u j is the importance weight of the scene element in this category of elements; w i The attribute characteristics of scene elements are determined by using the hierarchical analysis method to determine the w of static scene elements and dynamic scene elements. i , divided into five levels, with corresponding values ​​of 1-5 points; subscript i represents the parameter or value of the i-th scene element; subscript j represents different element levels; Step 3: Calculate the factor-system complexity mapping relationship; The complexity mapping relationship between scene elements and autonomous driving vehicle subsystems is calculated from the perspectives of the three subsystems of perception, decision-making, and execution, as follows: 31) Map the complexity of perception system elements; select three sensor types: camera, millimeter wave radar, and lidar to calculate the impact of scene elements on the perception system, as follows: 311) For the camera system, in addition to the characteristics of the camera itself, the vehicle's motion state is considered, including two factors: relative distance and relative speed; For relative distance interference, the relative distance between the target scene element and the ego vehicle is equivalent to the distribution of the target scene element in the image at a certain moment, and then the recognition difficulty of the camera system is characterized; as shown in formula (5): In the formula, I dis It is the ratio of the actual size of the target scene element in the image captured by the camera sensor at a certain moment to the overall imaging result. The lower the ratio, the greater the impact of the actual size of the scene element at the current distance on the sensor recognition effect. target is the actual size of the target scene element; P a is the field of view of the camera sensor at the target scene element position at the current moment; d is the relative distance between the target scene element and the vehicle under test; FOV is the field of view of the camera sensor; for continuous scene elements, I dis Obtained by calculating the ratio of the visible length in the image to the image size; For relative speed disturbance, the elastic coefficient ε is introduced to analyze the recognition difficulty of dynamic scene elements that are close to the tested autonomous vehicle and move at high speed; Taking into account the interference of relative distance and relative speed, the total impact of the calculated scene elements on the camera sensor is shown in formula (6): In the formula, I c is the complexity mapping result of scene elements to camera perception system, I c The larger the value, the more complex the scene element is for the camera perception system; 312) For the millimeter wave radar system perception process, the coefficient Im is introduced to quantify the degree of influence, as shown in formula (7): In the formula, I m The larger the value, the more complex the scene elements are for the camera perception system; n is the number of close-range moving targets within the detection range of the millimeter-wave radar; 313) For the laser radar system, the point cloud density D obtained by approximate fitting s relationship with relative distance; and based on the result, use coefficient I r1 The influence of relative distance on the lidar perception process is quantified as shown in formula (8): In the formula, e is a natural constant; Introduction coefficient I r2 The influence of the number of close-range moving targets within the laser radar detection range on the laser radar is shown in formula (9): Where n is the number of close-range moving targets within the detection range of the laser radar; Taking into account the influence of the relative distance of scene elements and the occlusion of close-range targets, the total influence coefficient of scene elements on the lidar sensor is obtained as shown in formula (10): I r =I r1 +I r2 (10) In the formula, I r is the complexity mapping result of scene elements to camera perception system, I r The larger the value, the more complex the scene element is for the camera perception system; 314) Comprehensively analyze the quantitative impact of weather factors on sensor perception characteristics, approximate the impact of different weather factors on vehicle sensors, and obtain the weather factor complexity mapping result I w ; When calculating the weighted scores of weather element characteristics, the assignment process of important weights is omitted, and only the specific score S is calculated by combining the influence coefficient. i(weather) , as shown in formula (11): S i(weather) =1+I w (11) The influence of each scene element on the camera, millimeter wave radar, and lidar is comprehensively considered to obtain the complexity mapping relationship E between the scene element and the perception system. perception The result is shown in formula (12): Where, T i is the total time that scene element i is captured by the perception system β; S i(weather)β Score the weights of weather element features that affect the perception system β; If scene element i is identified by multiple types of perception systems through perception fusion, the perception system complexity E corresponding to the element perception Take the average value according to the number of fusion sensor categories; I β is the impact of scene elements on different sensors, including cameras, lidars, and millimeter-wave radars. The values ​​of β include c, m, and r. ∏ is the multiplication symbol. 32) Map the complexity of decision-making system elements; The concept of decision planning margin is used to quantify the complexity mapping results between scene elements and decision systems. Decision planning margin refers to the area in which the autonomous driving vehicle can be planned to drive in a short time. The decision planning margin calculation process is divided into two parts: vehicle driving potential field and vehicle reachable domain. The former is the impact of the surrounding traffic environment on the driving decision, and the latter is the driving range caused by the dynamic limitations of the autonomous driving vehicle itself, as follows: 321) Calculate the vehicle driving potential field; The vehicle's operable space is calculated using a potential field model, which includes two parts: static scene element potential field and dynamic scene element potential field. Static scene elements are divided into scene static objects and constraint facilities; Among them, static objects in the scene have collision risks for autonomous vehicles, so they have a greater impact on traffic safety in the decision-making and planning process; pi The potential field representing the static elements of the scene is calculated as shown in formula (13): In the formula, S i Score the weight of scene element features; r i is the relative distance between the autonomous driving vehicle and the scene element; k1 is a fixed constant; Constraint facilities tend to be constrained by traffic regulations and have a weak correlation with serious risks such as loss of life and property. Therefore, they have a relatively small impact on traffic safety in the decision-making and planning process. Constraint facilities are classified into speed constraint elements and position constraint elements, and potential field models are constructed for each. Using C dsi The potential field representing the position constraint elements such as lane lines is calculated as shown in formula (14): Using C dvi The potential field representing speed limit signs, traffic lights and other speed constraint elements is calculated as shown in formula (15): Where D i is the lane width; x i is the vertical position of the target scene element; v i is the driving speed of the autonomous vehicle; v di is the speed corresponding to the constraint element; k2, k3, k4 are fixed constants; v d Traffic regulations restrict the maximum speed of vehicles. When the scene element is a red traffic light, the speed limit is considered to be 0m / s; when the scene element is a green traffic light, the speed limit is considered to be the current speed of the vehicle; For dynamic scene elements, they are considered to be static objects in the scene with motion characteristics. The potential field representation of dynamic scene elements is shown in formula (16): In the formula, v i is the moving speed of the target scene element; θ is the angle between the autonomous driving vehicle and the moving direction of the scene element; the subscript i refers to the sequence number of the scene element; k5 and k6 are fixed constants; Superimpose the potential fields of static and dynamic scene elements to obtain the total potential field F of the current scene potential , as shown in formula (17): In the formula, C iλ is the unified expression of the potential field of all elements in the scene, λ is the classification symbol of Dsi, Dvi, pi, and i is the counting number of different scene elements; the scene element potential field is classified into static scene element potential field and dynamic scene element potential field, and the static scene element potential field is further subdivided into scene static object potential field and constraint element potential field; among them, the scene static object potential field and the dynamic scene element potential field are described by the same expression, and pi is the potential field C that characterizes these two types of elements. pi The classification subscript; Dsi refers to the scene elements C such as lane lines and other position constraints Dsi Classification subscript; Dvi refers to speed limit and other speed constraint scene elements C Dvi Classification corner mark; 322) Calculate the vehicle reachable area; The vehicle reachable domain refers to the road area that the autonomous vehicle can reach in the short term due to vehicle dynamics and decision-making tasks. It includes two parts: the vehicle's operable space and the vehicle's decision-making event area. The maneuverable space of the vehicle determined by dynamics constraints includes longitudinal and lateral parts. The minimum safe braking distance of a vehicle at a certain speed in the longitudinal following theory is defined as the longitudinal limit distance of the maneuverable space of the autonomous driving vehicle, as shown in formula (18): Where a is the maximum deceleration of the autonomous driving vehicle, and a=10m / s 2 ; v is the instantaneous speed of the vehicle before braking; Considering the maximum wheel turning angle of the vehicle, the maximum static front wheel turning angle β = 40° is taken as the vehicle turning limit, thereby obtaining the lateral maneuverable space; Combined with the longitudinal and lateral operating limits, the theoretical operable space of the autonomous driving vehicle is calculated. During the calculation process, the influence of the position changes of the surrounding scene elements and dynamic scene elements on the potential field of the scene elements within this time period is considered, and the position offset of the dynamic scene elements during the time when the vehicle decelerates to a standstill and is also generated by braking at the maximum deceleration is selected as the final potential field center of the dynamic elements, and the subsequent decision margin is calculated. The pending decision event area refers to the range of surrounding vehicles that the autonomous vehicle needs to consider in its short-term decision-making process. In the same scene, the vehicle's pending decision task area changes with the vehicle's driving process and is related to the relative distance between the target scene elements and the vehicle. Taking TTC, i.e., collision time less than 2s, as the basis for judging dangerous scenes, 2s is selected as the time boundary, and the pending decision event area of ​​the autonomous vehicle is defined as a semicircular range with a radius of R = 2|v| with the front of the vehicle as the center. The vehicle's operable space and the vehicle's pending decision event area are combined to obtain the reachable domain F of the autonomous driving vehicle based on the actual road boundary. permission ; 323) Calculate vehicle decision planning margin; The potential field area with a danger level Rick less than or equal to r is set as the safe driving area of ​​the autonomous vehicle. The overlapping area with the vehicle reachable domain represents the area where the autonomous vehicle can achieve a relatively safe driving path at the current moment, that is, the vehicle's decision planning margin E dicision ; E dicision The calculation process is shown in formula (19): In the formula, T is the total driving time of the vehicle; 33) Map the complexity of the execution system; the execution system involves the movement of the vehicle itself and the interaction between the tire and the road. Refer to the backup tire force as the execution system complexity map E implement The indicators are as follows: 331) The available reserve tire force is simplified to the product of the current vehicle acceleration and the weight score of the road quality factor characteristic, which is expressed as E implement , as shown in formula (20): In the formula, S rp is the characteristic weight score of the road quality elements in the road layer; v is the speed of the tested vehicle; Step 4: Couple the scene complexity; The complexity features of the scenarios obtained in steps 2 and 3 are combined to obtain the real-time complexity evaluation results of the continuous test scenarios, as follows: 41) The influence transmission relationship between perception, decision-making and execution systems is used as the weight coefficient The coupling process of the total scene complexity is introduced as shown in formula (21): In the formula, l i is the transfer coefficient, l1=3, l2=2, l3=1; 42) Finally, the real-time complexity evaluation result of the continuous test scenario is obtained, as shown in formula (22):

2. The method for real-time evaluation of complexity of continuous test scenarios of an autonomous driving vehicle according to claim 1, characterized in that: The specific method of step one is as follows: 11) Based on the six-layer model of the autonomous driving test scenario, the scene element categories are extracted, and the characteristic attribute weights of the scene elements are obtained through the hierarchical analysis method, which serves as the basis for the subsequent scene complexity evaluation process for the autonomous driving system; 12) Combine the characteristic values ​​of scene elements with their impact on the system to establish the complexity mapping relationship between scene elements or element combinations and perception, control and execution systems; 13) Calculate the complexity weight system of each subsystem of autonomous driving based on the influence transfer relationship, and couple to obtain the comprehensive evaluation results of the complexity of continuous test scenarios.

3. The method for real-time evaluation of complexity of continuous test scenarios of an autonomous driving vehicle according to claim 1, characterized in that: u j The calculation is performed by using the extension hierarchical analysis method, as follows: First, we construct the importance relationship judgment matrix A of the scene element category set = (a pq ) nj×nj , where p, q = 1, 2, ..., n j ;n j is the number of scene element categories in the jth classification level; a pq =[a + pq , a - pq ] is the number of extension intervals; a + pq and a - pq are the upper and lower endpoints of the interval caused by different cognition and preference; find the upper decision matrix A formed by the upper endpoints of the interval respectively + and the lower decision matrix A formed by the lower endpoints of the interval - The corresponding normalized eigenvector x + and x - , and calculate the coefficients k and m, as shown in formulas (2) and (3): When 0≤k≤1≤m is satisfied, it means that the extension judgment matrix is ​​consistent; thus, the extension interval number weight vector R of the jth classification level is obtained. k =[kx - , mx + ]=(R1,R2,…,R nj ) T ; Then calculate the relative importance of each interval number weight vector, using V(R a ≥R b ) represents R a Greater than R b possibility; if And a≠b, V(R a ≥R b )≥0, then the important weight coefficient ranking u is obtained a and u b , as shown in formula (4): After normalization, the important weight coefficient of the element in the jth classification level is U j =(u j 1 ,u j 2 ,…u j nj ) T ; are the upper and lower endpoints of the extension interval weight vector of element a and element b respectively; R a , R b are the extension interval number weight vectors of element a and element b, n k is the total number of elements in the kth layer.

4. The method for real-time evaluation of complexity of continuous test scenarios of an autonomous driving vehicle according to claim 1, characterized in that: The three-layer scene element framework is as follows: The first-level classification is divided into: static scene elements, dynamic scene elements and weather elements; The secondary classification is divided into: road layer, traffic facility layer, temporary operation layer, target layer and natural environment layer; The three-level classification is divided into: road type, road quality, number of lanes, road markings, signal lights, signs, other facilities, traffic accidents, temporary parking, road construction, motor vehicles, non-motor vehicles, pedestrians, light, rain, snow and fog; Among them, when the first-level classification is static scene elements and the second-level classification is road layer, the third-level classification is road type, road quality, number of lanes and road markings; When the first-level classification is static scene elements and the second-level classification is the traffic facility layer, the third-level classification is traffic lights, signboards, and other facilities; When the first-level classification is static scene elements and the second-level classification is temporary operation layer, the third-level classification is traffic accidents, temporary parking and road construction; When the first-level classification is dynamic scene elements and the second-level classification is the target layer, the third-level classification is motor vehicles, non-motor vehicles and pedestrians; When the first level classification is weather elements and the second level classification is natural environment layers, the third level classification is light, rain, snow and fog.

Images