CN116205024A - Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene - Google Patents

Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene Download PDF

Info

Publication number
CN116205024A
CN116205024A CN202211396377.4A CN202211396377A CN116205024A CN 116205024 A CN116205024 A CN 116205024A CN 202211396377 A CN202211396377 A CN 202211396377A CN 116205024 A CN116205024 A CN 116205024A
Authority
CN
China
Prior art keywords
scene
risk
space
scenes
elements
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211396377.4A
Other languages
Chinese (zh)
Inventor
张宇飞
孙博华
马芳武
赵帅
翟洋
吴量
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jilin University
Original Assignee
Jilin University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jilin University filed Critical Jilin University
Priority to CN202211396377.4A priority Critical patent/CN116205024A/en
Publication of CN116205024A publication Critical patent/CN116205024A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/04—Constraint-based CAD
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Geometry (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention discloses a general generation method of a self-adaptive automatic driving dynamic scene facing to a high-low dimension evaluation scene, which comprises the following steps: step one, optimizing the dimension of an evaluation scene element based on a chromatographic analysis method; secondly, constructing a scene space and a key long tail function; thirdly, generating a self-adaptive dynamic scene facing to the low-dimensional evaluation scene; fourthly, generating a self-adaptive dynamic scene facing to the high-dimensional evaluation scene; fifthly, self-adaptive dynamic scene assessment for high-low dimension assessment scenes; the beneficial effects are that: the step-by-step classification of the scene elements is realized, so that a hierarchical model of the scene elements is established. And selecting the scene elements with larger importance weight values as decision variables, thereby realizing the problem of optimizing the dimensions of the scene elements. A new ideal scene space construction method is established. And solving the occurrence probability of the scene in the low-dimensional ideal scene space. The search efficiency is greatly improved.

Description

Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene
Technical Field
The invention relates to a general generation method of an automatic driving dynamic scene, in particular to a general generation method of a self-adaptive automatic driving dynamic scene facing to high-low dimension evaluation scenes.
Background
Currently, automatic driving has become one of the main technological development directions of automobiles, and the automatic driving of automobiles is a future development trend. The precondition that the automatic driving automobile can get on the road is that the driving safety is fully verified and reaches the relevant standard, and in order to fully test the safety of the automatic driving automobile and the reliability of an intelligent algorithm thereof, simulation test, site test and road test are required to be sequentially carried out on the automatic driving automobile. Although the three test methods belong to different stages, test flows and implementation modes, specific test scenes are required to be designed to execute the test process. Therefore, how to generate a representative simple self-adaptive automatic driving dynamic scene facing the low-dimensional evaluation scene and a complex self-adaptive automatic driving dynamic scene facing the high-dimensional evaluation scene becomes a very critical problem, which greatly improves the test efficiency and test reliability of the automatic driving automobile and accelerates the deployment process of the automatic driving automobile.
The current generation method for the self-adaptive automatic driving dynamic scene mainly has two defects, namely, only the danger of the scene is considered as the only criterion when the scene is generated, namely, only the scene which is dangerous for the automatic driving automobile is focused, the scene generation principle can test the safety boundary of the automatic driving automobile, but because the generated scene fails to consider the real road traffic condition, the probability of occurrence of a plurality of scenes with higher danger on the real road is extremely low or even not, and the practical significance of testing the automatic driving automobile by using the scene is not great; secondly, the current research on the generation of the self-adaptive automatic driving dynamic scene is the generation of the self-adaptive dynamic scene facing the low-dimensional evaluation scene, no perfect self-adaptive automatic driving dynamic scene generation method facing the high-dimensional evaluation scene exists, and the research on the generation of the self-adaptive automatic driving dynamic scene facing the high-dimensional evaluation scene is necessary because the automatic driving automobile is required to run on a road really and cannot be limited to be capable of coping with simple low-dimensional scenes and complex high-dimensional scenes.
Chinese patent CN202210941004.4 discloses a method, apparatus, device and storage medium for generating an autopilot test scene, which can generate an autopilot test scene with high risk according to data of a target traffic participant; chinese patent CN202210804060.3 discloses a method, apparatus, vehicle and storage medium for generating an automatic driving test scene, which can generate a simulation test scene by loading a key parameter interval of the simulation test scene into a scene design document and using a preset script; the Chinese patent CN202210741420.X discloses an automatic simulation test system for intelligent driving and related equipment, wherein a scene generation module can create a test scene according to the driving parameters of a main vehicle and generalize the test scene to generate one or more test scenes for testing an intelligent driving algorithm, but the three patents do not consider real road traffic conditions when generating scenes and do not relate to the problem of generating self-adaptive dynamic scenes facing high-dimensional evaluation scenes.
Disclosure of Invention
The invention aims to generate a key self-adaptive dynamic scene capable of accelerating an automatic driving automobile test, improve the test efficiency and the reliability of the automatic driving automobile and accelerate the deployment process of the automatic driving automobile, and provides a general self-adaptive automatic driving dynamic scene generation method for high-low dimension evaluation scenes.
The invention provides a self-adaptive automatic driving dynamic scene general generation method for a high-low dimension evaluation scene, which comprises the following steps:
the first step, the dimension optimization of the evaluation scene element based on the chromatographic analysis method comprises the following specific processes:
step one, constructing a scene element hierarchical model based on scene parameterization, describing a scene as a set of scene elements, classifying the scene elements step by step according to basic attributes of the scene elements and the geographic position conditions to which the scene elements belong, and thus constructing the scene element hierarchical model;
the scene is the overall dynamic description of the automatic driving automobile and each element of the driving environment thereof in a period of time, and the scene elements are divided into static elements and dynamic elements according to the difference of the geographic positions of the scene elements, wherein the static elements and the dynamic elements are divided into two types of off-road elements and in-road elements from the basic attributes of the scene elements;
the road exterior elements of the static elements mainly comprise static roadside objects such as buildings, trees, green belts and traffic marks, and the road interior elements of the static elements mainly comprise elements related to lanes and elements related to environments, wherein the elements related to the lanes such as lane types, lane line types, lane numbers, lane widths, lane gradients, lane curvatures, lane line types, lane line widths, road surface types and road surface attachment coefficients; environmental related elements such as weather and light status;
The off-road elements of the dynamic element mainly include traffic participants, such as traffic participant types and numbers of traffic participants, and the on-road elements of the dynamic element mainly include elements related to traffic participants, elements related to initial states, and elements related to running state sequences, elements related to traffic participants, such as traffic participant types and numbers of traffic participants; elements related to the initial state such as an initial position, an initial lane, and an initial speed; elements associated with the driving state sequence such as trigger pattern, relative distance sequence, relative velocity sequence, and relative acceleration sequence;
step two, scene element dimension optimization based on an analytic hierarchy process, solving the influence transfer times of scene elements of the same level in each level of an automatic driving system through an influence transfer model according to the scene element level model, calculating the difference value of the influence transfer times of the scene elements of the same level, converting the difference value of the influence transfer times into a scale value according to the corresponding relation between the difference value of the influence transfer times and a 1-9 scale method, thereby establishing a judgment matrix of the scene elements of the same level, solving the maximum characteristic value of the judgment matrix and the corresponding characteristic vector thereof, normalizing the characteristic vector to obtain the importance weight value of each scene element of the same level relative to the level, verifying the rationality of the importance weight value through consistency test, calculating the importance weight value of each scene element in all the scene elements according to the level model of the scene element and the importance weight value of each scene element of each level relative to the level, and selecting the scene element with larger importance weight value as a decision variable according to the type of a target scene and the dimension requirement of an evaluation scene to be established;
Taking the scene element hierarchical model in the first step as input, solving the influence transfer times of the scene elements of the same hierarchy in each hierarchy of the automatic driving system through an influence transfer model, wherein the influence transfer model has three assumptions, and the method comprises the following specific steps:
1) The influence of scene elements on the automatic driving system is transmitted step by step along with the hierarchy of the automatic driving system, and the influence is not attenuated along with the transmission of the hierarchy;
2) The influence of different types of scene elements on an automatic driving system is the same;
3) The influence of the scene element on the automatic driving system is represented by the influence transfer times of the scene element among layers of the automatic driving system, the more the transfer times are, the larger the influence of the scene element is, and the influence and the transfer times are in a linear relation;
the number of impact transfers of scene elements in the impact transfer model in each level of the autopilot system can be calculated by the following equation:
Figure BDA0003933187860000031
/>
wherein P (n) represents that a certain scene element is automatically drivenThe number of impact transfer times in each hierarchy of the system, n represents the number of element attributes of the scene element, E i Representing the influence transfer times of the ith element attribute of the scene element in each level of the automatic driving system;
After the influence transfer times of all the level scene elements in the scene element level model are obtained according to the influence transfer model, calculating the difference value of the influence transfer times of the scene elements of the same level, converting the difference value of the influence transfer times into a scale in a 1-9 scale method, after converting the difference value of the influence transfer times into a scale in a 1-9 scale method, establishing a judgment matrix of the scene elements of the same level, solving the maximum eigenvalue of the judgment matrix and the eigenvector corresponding to the maximum eigenvalue, and normalizing the eigenvector to obtain the importance weight value of the scene elements of the same level;
in order to verify the rationality of the importance weight value, consistency test is carried out on the judgment matrix, and the formula of the consistency test is as follows:
Figure BDA0003933187860000032
wherein CR represents a consistency index, RI is a standard value of a hierarchical total ordering average random consistency index, and different values are taken according to different orders of a judgment matrix;
CI represents a hierarchical total ordering consistency index, and the calculation formula of CI is as follows:
Figure BDA0003933187860000033
wherein lambda is max Represents the maximum eigenvalue of the judgment matrix, n is the order of the judgment matrix, and when CR<0.1, the consistency of the judgment matrix is good, and the consistency requirement is met;
finally, according to the hierarchical model of the scene elements and the importance weight value of each scene element of each hierarchy relative to the hierarchy, calculating the importance weight value of each scene element in all the scene elements, then determining the type of a target scene to be researched, and according to the type of the target scene and the dimension requirement of an evaluation scene to be built, selecting some scene elements with larger importance weight values as decision variables, namely finishing the dimension optimization of the scene elements;
Secondly, constructing a scene space and a key long tail function, wherein the specific process is as follows:
the method comprises the steps that firstly, an ideal scene space based on scene element dimension optimization is constructed, an ideal test scene space is obtained by discretizing decision variables obtained through the scene element dimension optimization in the first step, each decision variable is discretized into different values according to element attributes of the decision variable, the value range and the discrete step length of the decision variable are determined according to the importance weight value of the decision variable under the premise of fully considering the constraint of real road conditions, and the decision variable with higher importance weight value is set with a larger value range and a smaller discrete step length, so that the situation that as many scenes as possible can be contained in the ideal test scene space is ensured; the decision variables with low importance weights are selected according to the principle that key scenes can be contained in an ideal test scene space, wherein the value range and the discrete step length of the decision variables are selected;
step two, constructing a real scene space based on natural driving data, namely acquiring the scene data by acquiring the real scene space, wherein the vehicle runs on a real road through the scene data to acquire the scene data, preprocessing the acquired sensor data, screening target scene data from the preprocessed scene data, and screening data corresponding to decision variables from the target scene data to construct the real scene space;
Scene data acquisition should also follow the following two principles:
1) The scene data acquisition area should contain cities and areas with different roads and traffic characteristics, and the total acquisition mileage is enough;
2) The scene data acquisition environment comprises different weather types and illumination types;
the specific content of the sensor data preprocessing comprises: time alignment and spatial alignment of the sensor data; verifying the validity of the sensor data; generating a vehicle bus alignment signal, a vehicle state alignment signal and a multi-modal environmental sensor alignment signal;
according to the type of the target scene selected in the first step, capturing scene data of all target scenes from all preprocessed data in a mode of manually watching videos, wherein each segment of complete target scene data is called a scene working condition and represents a complete target scene event occurring on a real road, and data corresponding to decision variables in each scene working condition are screened out to form a real scene space;
step three, constructing a critical long-tail function based on the scene occurrence probability and the scene risk, and designing the critical long-tail function based on the scene occurrence probability and the scene risk as a basis for screening a critical adaptive dynamic scene according to an adaptive dynamic scene generation principle that the automatic driving automobile is challenging and has a certain occurrence probability on a real road, wherein the calculation formula of the critical long-tail function is as follows:
I(x)＝V(x)·P(x) (4)
Wherein, I (x) represents a critical long tail function value of a scene, P (x) represents a scene occurrence probability, and V (x) represents a scene risk;
thirdly, generating a self-adaptive dynamic scene facing to the low-dimensional evaluation scene, wherein the specific process is as follows:
firstly, solving the occurrence probability of a scene based on a convex combination algorithm, converting the scene in a low-dimensional real scene space from natural driving data into a scene in a low-dimensional ideal scene space through the convex combination algorithm, and solving the occurrence probability of the scene in the low-dimensional ideal scene space;
let the presence vector { x } 1 ,x 2 ,x 3 ...x n E.g. real lambda i Not less than 0, i=1, 2, 3..n, and λ 1 +λ 2 +...+λ n Let 1 be λ 1 x 1 +λ 2 x 2 +...λ n x n Is a vector { x } 1 ,x 2 ,x 3 ...x n A convex combination of };
solving the scene occurrence probability based on a convex combination algorithm has the following assumptions:
1) The scene occurrence probability does not change suddenly in a discrete step range of the ideal scene space, is continuously changed and follows a linear change rule;
2) In a discrete step range of an ideal scene space, the linear change rule of the scene occurrence probability can be represented by Euclidean distance between scenes in the ideal scene space;
3) In a discrete step range of an ideal scene space, the scene with a short Euclidean distance with a high occurrence probability has high occurrence probability, and the scene with a long Euclidean distance with the high occurrence probability has low occurrence probability;
Taking a scene space within a discrete step as an example, A 1 (R A1 ,Δv A1 ),A 2 (R A2 ,Δv A2 ),A 3 (R A3 ,Δv A3 ),A 4 (R A4 ,Δv A4 ) Is 4 uniform discrete scenes in a two-dimensional ideal scene space, B (R B ,Δv B ) Is a certain scene in the real scene space obtained by natural driving data, and B is A 1 ，A 2 ，A 3 ，A 4 Rectangular interior, L 11 Is B to A 2 、A 4 Euclidean distance of straight line, L 12 Is B to A 1 、A 3 Euclidean distance of straight line, L 21 Is B to A 3 、A 4 Euclidean distance of straight line, L 22 For B to A 1 、A 4 The Euclidean distance of the straight line can be converted into a scene A in the two-dimensional ideal scene space by a convex combination algorithm 1 ，A 2 ，A 3 ，A 4 ；
B＝ω 1 ·A 1 +ω 2 ·A 2 +ω 3 ·A 3 +ω 4 ·A 4 (5)
Wherein omega is 1 ，ω 2 ，ω 3 ，ω 4 Respectively A 1 ，A 2 ，A 3 ，A 4 The weight coefficient of (2) is calculated as follows:
Figure BDA0003933187860000051
Figure BDA0003933187860000052
/>
Figure BDA0003933187860000053
Figure BDA0003933187860000054
according to the method, all scenes in the real scene space are converted into uniform discrete scenes in the ideal scene space, and the number of the uniform discrete scenes is counted by combining the weight coefficients of the uniform discrete scenes, so that the occurrence probability of the scenes in the two-dimensional ideal scene space is obtained;
taking a scene space within a discrete step as an example, A 1 (R A1 ,Δv A1 ,Δa A1 ),A 2 (R A2 ,Δv A2 ,Δa A2 ),A 3 (R A3 ,Δv A3 ,Δa A3 ),A 4 (R A4 ,Δv A4 ,Δa A4 ),A 5 (R A5 ,Δv A5 ,Δa A5 ),A 6 (R A6 ,Δv A6 ,Δa A6 ),A 7 (R A7 ,Δv A7 ,Δa A7 ),A 8 (R A8 ,Δv A8 ,Δa A8 ) Is 8 uniform discrete scenes in three-dimensional scene space, B (R B ,Δv B ,Δa B ) Is a certain scene in the real scene space obtained by natural driving data, and B is A 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 Inside the cube of the structure, L 11 Is B to A 2 ，A 4 ，A 6 ，A 8 Euclidean distance of plane, L 12 Is B to A 1 ，A 3 ，A 5 ，A 7 Euclidean distance of plane, L 21 Is B to A 3 ，A 4 ，A 7 ，A 8 Euclidean distance of plane, L 22 Is B to A 1 ，A 2 ，A 5 ，A 6 Euclidean distance of plane, L 31 Is B to A 5 ，A 6 ，A 7 ，A 8 Euclidean distance of plane, L 32 Is B to A 1 ，A 2 ，A 3 ，A 4 The Euclidean distance of the plane, the scene B in the real scene space can be converted into the scene A in the three-dimensional ideal scene space through a convex combination algorithm 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 ；
B＝ω 1 ·A 1 +ω 2 ·A 2 +ω 3 ·A 3 +ω 4 ·A 4 +ω 5 ·A 5 +ω 6 ·A 6 +ω 7 ·A 7 +ω 8 ·A 8 (10)
Wherein omega is 1 ，ω 2 ，ω 3 ，ω 4 ，ω 5 ，ω 6 ，ω 7 ，ω 8 Respectively A 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 The weight coefficient of (2) is calculated as follows:
Figure BDA0003933187860000055
Figure BDA0003933187860000056
Figure BDA0003933187860000061
Figure BDA0003933187860000062
Figure BDA0003933187860000063
Figure BDA0003933187860000064
Figure BDA0003933187860000065
Figure BDA0003933187860000066
according to the method, all scenes in the real scene space are converted into uniform discrete scenes in the ideal scene space, and the number of the uniform discrete scenes is counted by combining the weight coefficients of the uniform discrete scenes, so that the occurrence probability of the scenes in the three-dimensional ideal scene space is obtained;
step two, solving scene risk boundaries based on a support vector regression algorithm, pre-dividing scene risk boundaries in a real scene space according to collision time, constructing joint distribution of the collision time and relative acceleration, taking boundary points which belong to the same boundary line in the joint distribution as a training sample set, independently inputting the training sample set into the support vector regression algorithm, enabling each training sample set to correspond to one scene risk boundary, integrating all scene risk boundary lines to obtain all scene risk boundaries, mapping the scene risk boundaries in the real scene space into an ideal scene space, and setting risk coefficients for scenes with different risk degrees;
TTC represents the time required from the current moment to collision under the condition of keeping the current motion state unchanged, the smaller the value of the TTC is, the higher the danger of representing a scene is, and the TTC is calculated by the following formula:
Figure BDA0003933187860000067
wherein R represents the relative distance between two vehicles, and Deltav represents the relative speed between the vehicle and the target vehicle;
the pre-dividing standard for setting the scene risk boundary is as follows: when TTC is E [0s,1s ], the risk level of the scene is a collision scene; the scene risk level is an emergency scene when TTC (1 s,3 s), the scene risk level is a conflict scene when TTC (3 s,5 s), and the scene risk level is a safety scene when TTC (5 s, + -infinity) or (- + -infinity, 0 s);
constructing a joint distribution of TTC and delta a in a real scene space, wherein the joint distribution is equivalent to four scene risk boundaries determined by TTC=0, TTC=1, TTC=3 and TTC=5, respectively taking boundary points on the left side and the right side of the joint distribution as a single training sample set, inputting the single training sample set into a support vector regression algorithm for learning, selecting a linear kernel function for solving because the boundary of the joint distribution is nearly linear, obtaining scene risk boundaries on the left side and the right side of the joint distribution, combining the scene risk boundaries based on TTC pre-division with scene risk boundaries based on support vector regression, and obtaining the risk boundaries in the real scene space, wherein the risk boundaries are as follows:
Figure BDA0003933187860000071
Wherein, I l And/l r Scene risk boundaries on the left side and the right side of a joint distribution diagram obtained through support vector regression learning respectively, l m1 、l m2 、l m3 And l m4 Scene risk boundaries obtained by pre-dividing according to TTC;
mapping the scene risk boundary into an ideal scene space, setting the scene risk corresponding to a collision scene as 1, setting the scene risk corresponding to an emergency scene as 0.7, setting the scene risk corresponding to a collision scene as 0.3, and setting the scene risk corresponding to a safety scene as 0;
step three, generating a key self-adaptive dynamic scene based on a multi-starting-point optimization algorithm and a seed filling algorithm, calculating a key long-tail function value in an ideal scene space through the scene occurrence probability and the scene risk, inputting the key long-tail function, the ideal scene space and a key threshold value into the multi-starting-point optimization algorithm, solving to obtain a local key self-adaptive dynamic scene, inputting the local key scene, the key long-tail function, the ideal scene space and the key threshold value into the seed filling algorithm, and solving to obtain all the key self-adaptive dynamic scenes;
obtaining a critical long-tail function of a scene in an ideal scene space through the scene occurrence probability obtained in the first step and the scene risk obtained in the second step, inputting the critical long-tail function, the ideal scene space and a critical threshold gamma as inputs into a multi-starting-point optimization algorithm, sampling a plurality of points in the ideal scene space as starting points of the algorithm through a manual setting or random mode, solving the maximum value of the critical long-tail function of an attraction domain where each starting point is located, and outputting the scenes corresponding to the maximum value of all the critical long-tail functions larger than gamma to obtain a local key self-adaptive dynamic scene facing the low-dimensional evaluation scene;
The method comprises the steps of taking a local key scene, a key long tail function, an ideal scene space and a key threshold gamma which are output by a multi-starting-point optimization algorithm as inputs, inputting the inputs into a seed filling algorithm, taking each local key scene as a starting point, searching scenes with the key long tail function value larger than gamma in a neighborhood around the starting point in the ideal scene space, taking the scenes as new starting points, continuously searching the neighborhood around the scenes, repeating the steps until the key long tail function values of the neighborhood scenes around all the starting points are smaller than gamma, ending the algorithm, and outputting all the scenes marked as the starting points to generate all key self-adaptive dynamic scenes facing to a low-dimensional evaluation scene;
fourth, self-adaptive dynamic scene generation facing to high-dimensional evaluation scene comprises the following specific processes:
firstly, constructing a scene risk identification model based on a Hammerstein identification process and solving the scene occurrence probability based on a convex combination algorithm, wherein the Hammerstein identification process is formed by connecting a static nonlinear link and a dynamic linear link in series, taking the relative state quantity of a host vehicle and a target vehicle and the acceleration of the target vehicle as the input of the scene risk identification model, taking the acceleration of the host vehicle as the output of the model, training a large amount of scene data, taking the model parameters of the scene risk identification model as key parameters for representing the intrinsic attribute of the scene risk, decoupling and reducing the key parameters by adopting a principal component analysis method, clustering the scene risk by adopting an ant colony algorithm, selecting a plurality of representative 'state-action' pairs from each scene risk level, inputting the representative pairs into the scene risk identification model, establishing the mapping relation between the scene risk level and a clustering result according to the clustering category in which the parameters are positioned, and solving the scene occurrence probability based on the convex combination algorithm can directly solve the occurrence probability of the scene occurrence probability pair according to a method for low-dimensional evaluation scene;
The method comprises the steps that a high-dimensional evaluation scene is regarded as a Markov decision process, the relative state quantity of a host vehicle and a target vehicle is regarded as a state, the acceleration of the target vehicle is regarded as a motion, the state and the motion of the same time step are regarded as a 'state-motion' pair, the relative state quantity of the host vehicle and the target vehicle and the acceleration of the target vehicle at the same time are regarded as the input of a scene risk degree identification model, the acceleration of the host vehicle is regarded as the output of the model, the scene risk degree identification model is a multi-input single-output system, the Hammerstein identification process is formed by connecting a static nonlinear link and a dynamic linear link in series, the static nonlinear link is a dead zone function, an S-shaped function or a saturation function, and the z-transformation of the dynamic linear link is shown in the following formula:
A(z -1 )O p (k)＝B(z -1 )·z -d ·N(k) (21)
wherein O is p (k) Represents the set of acceleration of the main vehicle, N (k) represents the output set of the static nonlinear link, d represents the input delay order and is defined as an integer multiple of the sampling time, A (z) -1 ) And B (z) -1 ) Can be calculated by the following formula:
Figure BDA0003933187860000081
wherein (a) 1 ,…a q ) (b) 1 ,…b n ) Each coefficient is a dynamic linear link, and q and n are orders of the dynamic linear link;
after a large number of scene data from natural driving data are trained, model parameters contained in a static nonlinear link and a dynamic linear link are key data for representing the intrinsic attribute of the scene risk, so the model parameters are used as data samples for evaluating the scene risk, the space dimension of the data samples is reduced as much as possible to improve the calculation efficiency on the premise of expressing the same model characteristics, and a principal component analysis method is adopted for decoupling and dimension reduction of the key parameters in a scene risk identification model;
Let H represent the parameter dimension of the scene risk recognition model, E represent the number of training data, then the model parameter data set X can be expressed as:
Figure BDA0003933187860000082
wherein x is i An internal reference vector representing a scene risk identification model;
taking X as input, inputting into a principal component analysis algorithm, and defining the percentage of the sum of the characteristic values of the first M principal components and the sum of all the characteristic values as principal component contribution rate, and then accumulating the principal component contribution rate M m Calculated by the following formula:
Figure BDA0003933187860000083
wherein lambda is i Representing the feature vector;
to ensure the dimension reducing effect, take M m And (3) taking the m value which is more than or equal to 85% and corresponds to the m value as the dimension of the independent parameter of the model obtained by algorithm calculation, and finally obtaining an m multiplied by E dimension matrix L:
Figure BDA0003933187860000091
clustering scene dangers by adopting an ant colony algorithm, taking L as input of the ant colony algorithm, finding a division method for minimizing the sum of distances of each data sample to a clustering center with known clustering number in L, and classifying the scene dangers into 4 grades by referring to a scene danger classification mode of low-dimensional evaluation scenes in the step two of the third step, wherein the clustering number is 4, and the ant colony algorithm is represented by the following formula:
Figure BDA0003933187860000092
where J represents the sum of the distances of each data sample to 4 cluster centers, l ip The p-th model parameter feature representing the i-th data sample, c jp The p model parameter characteristic of the j class center is calculated by the following formula:
Figure BDA0003933187860000093
wherein E is j For the j-th class of the observed variables E, ω ij For the amount of the dependency mark between the observed variable and the category, the following formula is used for calculation:
Figure BDA0003933187860000094
the classification quality of the ant colony clusters can be improved by iterative updating equations, and the updating equations are shown as follows:
Figure BDA0003933187860000095
wherein P is ij For probability of data sample cross class conversion, τ ij Is a normalized pheromone between the data sample i and the belonging class j, and is obtained by the following formulaAnd (3) calculating to obtain:
Figure BDA0003933187860000096
wherein ρ represents the volatility of the pheromone and t represents the time step;
the method comprises the steps of obtaining a clustering result of a 'state-action' pair through iterative updating of an equation, and establishing a mapping relation between the clustering result and scene risk levels because the clustering result does not have physical significance, selecting a plurality of representative 'state-action' pairs from each scene risk level, inputting the representative 'state-action' pairs into a scene risk identification model, and establishing the mapping relation between the scene risk level and the clustering result according to the clustering category of the parameter;
the scene occurrence probability solving based on the convex combination algorithm directly solves the occurrence probability of a 'state-action' pair according to a method for solving the scene occurrence probability of a low-dimensional evaluation scene;
Step two, reconstructing a critical long-tail function based on a Markov decision process, obtaining an ideal scene space of a high-dimensional evaluation scene by discretizing a decision variable, and reconstructing the critical long-tail function in a form of a 'state-action' pair by regarding the high-dimensional evaluation scene as the Markov decision process;
taking the initial speed of the target vehicle, the initial relative distance between the vehicle and the target vehicle, the initial relative speed between the vehicle and the target vehicle and the acceleration sequence of the target vehicle as decision variables x:
x＝[v 0 ,R 0 ,Δv 0 ,a 01 ,a 02 ,...,a 0k ] (31)
in the formula, v o Representing the initial speed of the target vehicle, R o Representing the initial relative distance between the host vehicle and the target vehicle, deltav o Representing the initial relative speed of the host vehicle and the target vehicle, a 0k Representing the acceleration of the target vehicle at the kth time step;
when v is set o The value range of (2) is [20m/s,40m/s ]]The discrete step length is 2m/s; r is R o The range of the values is (0 m,90 m)]Discrete step sizeIs 2m; deltav o The range of the value of (C) is [ -20m/s,20 m/s)]The discrete step length is 2m/s; a, a 0k The range of the value of (C) is [ -4m/s 2 ,2m/s 2 ]Discrete step size of 0.2m/s 2 The method comprises the steps of carrying out a first treatment on the surface of the When k is 10s, the number of ideal scene space scenes is 21×45×21×31 10 In order to reduce the dimension of an ideal scene space and simplify the calculation complexity of a critical long-tail function, consider a high-dimension evaluation scene as a markov decision process, consider the relative distance and relative speed between a host vehicle and a target vehicle as states, consider the acceleration of the target vehicle as an action, the acceleration of the target vehicle at a certain time step only depends on the relative states of the host vehicle and the target vehicle at the time step, so that the states and the actions at the same time step are regarded as a whole, namely a 'state-action' pair, the number of the 'state-action' pairs in the ideal scene space is 21×45×21×31= 615195, compared with the number of the scenes in the prior ideal scene space, the number of the 'state-action' pairs in the ideal scene space is greatly reduced, and the critical long-tail function is reconstructed as follows:
Figure BDA0003933187860000101
Wherein s is i Representing the state at the ith time step, a i Represents the action at the ith time step, G (s i ,a i ) Can be calculated by the following formula:
G(s i ,a i )＝V(s i ,a i )·P(s i ,a i ) (33)
wherein V(s) i ,a i ) Represents s i And a i The corresponding risk of "state-action" pair, P (s i ,a i ) Represents s i And a i The probability of occurrence of the corresponding "state-action" pair;
thirdly, generating a key self-adaptive dynamic scene based on a Q-learning algorithm, constructing a Belman equation, solving an objective function of a Markov decision process through the Belman equation, and solving an optimal action sequence corresponding to each initial state through updating expected benefits, namely the key self-adaptive dynamic scene;
set Q%s , a) For the expected benefit of an agent taking action a in state s, r is the return that the agent gives in the environment with action a taken, the Belman equation is used to solve the optimal strategy for the Markov decision process:
V π (s)＝E π [R t+1 +ξV π (S t+1 )|S t ＝s] (34)
wherein pi represents policy, S represents state set, R represents return set, ζ represents discount factor, t represents time step, S represents state at t moment, V π (s) represents a cost function, and Q may be updated according to the following equation:
Figure BDA0003933187860000102
wherein A represents an action set, a represents an action at time t, and alpha represents a learning rate;
the optimal action sequence corresponding to each initial state, namely a key self-adaptive dynamic scene, can be obtained;
Fifthly, self-adaptive dynamic scene assessment for high-low dimension assessment scenes comprises the following specific processes:
step one, sampling self-adaptive dynamic scenes based on an E-greedy sampling strategy, sampling scenes from the self-adaptive dynamic scenes facing high-low dimensional evaluation scenes by adopting the same sampling method, taking a low-dimensional self-adaptive dynamic scene as an example, setting a small probability value E, randomly sampling scenes from the low-dimensional self-adaptive dynamic scene with the probability of 1-E, randomly sampling scenes in scenes except for the low-dimensional self-adaptive dynamic scene in a low-dimensional ideal scene space with the probability of E, forming a test scene library, namely a test group one, and naming the test scene library obtained under the high-dimensional condition as a test group two by adopting the same sampling method;
step two, based on the self-adaptive dynamic scene evaluation of the accident rate and the test times, the test group I and the test group II respectively test the automatic driving automobile, set a certain confidence coefficient, such as 80 percent, until the accident rate converges, respectively record the test times and the accident rate when the accident rates of the two test groups under the confidence coefficient converge; meanwhile, setting a comparison experiment, namely randomly sampling scenes from a low-dimensional ideal scene space and a high-dimensional ideal scene space, respectively marking the scenes as a first comparison group and a second comparison group, testing an automatic driving automobile, setting the same confidence level until the accident rate converges, respectively recording the test times and the accident rate when the accident rates of the two comparison groups converge under the confidence level, comparing the accident rates of the first comparison group, the second comparison group and the second comparison group with the test times when the accident rate converges, and if the accident rate of the first comparison group is far greater than the accident rate of the first comparison group and the test times when the accident rate converges is far less than the accident rate of the first comparison group, indicating that the self-adaptive dynamic scene evaluation for the low-dimensional test scene is effective; if the accident rate of the second test group is far greater than that of the second control group and the test times when the accident rate converges are far less than that of the second control group, the adaptive dynamic scene assessment for the high-dimensional test scene is effective.
The invention has the beneficial effects that:
the invention generates the key self-adaptive dynamic scene capable of accelerating the test of the automatic driving automobile by the self-adaptive automatic driving dynamic scene universal generation method facing the high-low dimension evaluation scene, which greatly improves the test efficiency and the credibility of the automatic driving automobile and accelerates the deployment process of the automatic driving automobile. The method has the following specific beneficial effects:
1) The invention provides a scene element level model construction method based on scene parameterization. The scene is described as a set of scene elements, and the scene elements are classified step by step according to the basic attributes of the scene elements, the affiliated geographic positions and other conditions, so that a hierarchical model of the scene elements is established.
2) The invention provides a scene element dimension optimization method based on an analytic hierarchy process. The method comprises the steps of quantifying the importance problem of the scene element which is difficult to quantify through an analytic hierarchy process, verifying the rationality of the importance weight value of the scene element, and selecting the scene element with larger importance weight value as a decision variable according to the type of a target scene and the dimension requirement of an evaluation scene to be established, so that the problem of optimizing the dimension of the scene element is solved.
3) The invention provides an ideal scene space construction method based on scene element dimension optimization. The ideal test scene space is obtained by discretizing decision variables, and a larger value range and a smaller discrete step length are set for the decision variables with higher importance weight values, so that as many scenes as possible are contained in the ideal test scene space; the selection principle of setting a value range and discrete step length for the decision variable with lower importance weight is that a key scene can be contained in an ideal test scene space, so that a new ideal scene space construction method is established.
4) The invention provides a real scene space construction method based on natural driving data. The method comprises the steps of collecting scene data by a scene data collecting vehicle, carrying out scene data collection on a real road, preprocessing the collected sensor data, screening target scene data from the preprocessed scene data, and screening data corresponding to decision variables from the target scene data so as to construct a real scene space.
5) The invention provides a key long tail function construction method based on scene occurrence probability and scene risk. A new self-adaptive dynamic scene generation principle which is challenging to automatically drive automobiles and has a certain occurrence probability on a real road is designed, and a key long tail function is designed based on the scene occurrence probability and the scene risk degree to serve as a basis for screening key self-adaptive dynamic scenes.
6) The invention provides a scene occurrence probability solving method based on a convex combination algorithm. And converting the scene in the low-dimensional real scene space from the natural driving data into the scene in the low-dimensional ideal scene space through a convex combination algorithm, so as to solve the occurrence probability of the scene in the low-dimensional ideal scene space.
7) The invention provides a scene risk boundary solving method based on a support vector regression algorithm and a scene risk identification model constructing method based on a Hammerstein identification process, which realize the division of scene risk boundaries and the solving of scene risk in a low-dimensional ideal scene space and a high-dimensional ideal scene space.
8) The invention provides a low-dimensional key self-adaptive dynamic scene screening method based on a multi-starting-point optimization algorithm and a seed filling algorithm and a high-dimensional key self-adaptive dynamic scene screening method based on a Q-learning algorithm, so that the rapid search of key self-adaptive dynamic scenes in the whole ideal scene space is realized, and the search efficiency is greatly improved.
9) The invention provides a key long tail function reconstruction method based on a Markov decision process. The high-dimensional evaluation scene is regarded as a Markov decision process, the relative state quantity of the vehicle and the target vehicle is regarded as a state, the acceleration of the target vehicle is regarded as an action, and the state and the action of the same time step are regarded as a 'state-action' pair, so that the key long-tail function is reconstructed through the 'state-action' pair.
10 The invention provides a self-adaptive dynamic scene assessment method for a high-low dimension assessment scene. Sampling of low-dimensional and Gao Weizi adaptive dynamic scenes is achieved through an E-greedy sampling strategy, and evaluation of the adaptive dynamic scenes facing Gao Diwei evaluation scenes is achieved through comparison of convergence accident rates and corresponding test times by means of a comparison group.
Drawings
Fig. 1 is a schematic diagram of overall steps of a general generation method for dynamic scenes according to the present invention.
Fig. 2 is a block diagram of a method architecture of a dynamic scene generic generation method according to the present invention.
Fig. 3 is a schematic diagram of an exemplary embodiment of step one of the first steps of the present invention.
FIG. 4 is a schematic diagram of an exemplary embodiment of an impact delivery model according to the present invention.
Fig. 5 is a schematic diagram of an exemplary embodiment of the second step according to the present invention.
FIG. 6 is a schematic diagram of an exemplary embodiment of the present invention when the ideal scene space is a two-dimensional space.
FIG. 7 is a schematic diagram of an exemplary embodiment of the present invention when the ideal scene space is a three-dimensional space.
FIG. 8 is a schematic diagram illustrating the result of an exemplary implementation of the third step II according to the present invention.
FIG. 9 is a schematic diagram of an exemplary embodiment of a cut-in scenario according to the present invention.
FIG. 10 is a schematic diagram of an exemplary embodiment of a multi-start optimization algorithm according to the present invention.
FIG. 11 is a schematic diagram of an exemplary embodiment of a seed filling algorithm according to the present invention.
Fig. 12 is a schematic diagram of an exemplary embodiment of a follow-up scenario according to the present invention.
Detailed Description
Please refer to fig. 1 to 12:
the invention provides a general generation method of a self-adaptive automatic driving dynamic scene facing to a high-low dimension evaluation scene, which comprises the following steps:
step one, optimizing the dimension of an evaluation scene element based on a chromatographic analysis method;
secondly, constructing a scene space and a key long tail function;
thirdly, generating a self-adaptive dynamic scene facing to the low-dimensional evaluation scene;
fourthly, generating a self-adaptive dynamic scene facing to the high-dimensional evaluation scene;
and fifthly, self-adaptive dynamic scene assessment for high-low dimension assessment scenes.
The first step is based on the chromatographic analysis method to evaluate the scene element dimension optimization process as follows:
step one, constructing a scene element level model based on scene parameterization. The scene is described as a set of scene elements, and the scene elements are classified step by step according to the basic attributes of the scene elements, the geographical position of the scene elements and other conditions, so that a hierarchical model of the scene elements is established.
An exemplary embodiment of the first step is shown in fig. 3. The scene is the overall dynamic description of the automatic driving automobile and each element of the driving environment thereof in a period of time, and the scene elements can be divided into static elements and dynamic elements from the basic attributes of the scene elements, and further, the static elements and the dynamic elements can be divided into off-road elements and on-road elements according to the different geographic positions of the scene elements.
The off-road elements of the static elements mainly comprise static roadside objects such as buildings, trees, green belts and traffic signs. The in-road elements of the static elements mainly include elements related to lanes and elements related to environments, such as lane types (expressways, urban roads, rural roads), lane types (straight roads, curves, intersections), the number of lanes, lane widths, lane gradients, lane curvatures, lane types, lane line widths, road types (asphalt roads, cement concrete roads), and road adhesion coefficients; elements associated with the environment such as weather (sunny, cloudy, rainy, snowy, foggy) and light (abundant, dull, variable).
Off-road elements of dynamic elements mainly include traffic participants, such as traffic participant types (pedestrians, bicycles, other living things) and the number of traffic participants. The in-road elements of the dynamic element mainly include elements related to traffic participants, elements related to initial states, and elements related to running state sequences, such as traffic participant types (car, truck, motorcycle, tricycle, passenger car) and the number of traffic participants; elements related to the initial state such as an initial position, an initial lane, and an initial speed; elements associated with the driving state sequence such as a trigger pattern (distance trigger, time trigger), a relative distance sequence, a relative velocity sequence, and a relative acceleration sequence.
Step two, optimizing the dimension of the scene element based on the analytic hierarchy process. According to the scene element hierarchical model, solving the influence transfer times of scene elements of the same hierarchy in each hierarchy of an automatic driving system through the influence transfer model, calculating the difference value of the influence transfer times of the scene elements of the same hierarchy, converting the difference value of the influence transfer times into a scale value according to the corresponding relation between the difference value of the influence transfer times and a 1-9 scale method, thereby establishing a judgment matrix of the scene elements of the same hierarchy, solving the maximum characteristic value of the judgment matrix and the corresponding characteristic vector thereof, normalizing the characteristic vector to obtain the importance weight value of each scene element of the same hierarchy relative to the current hierarchy, verifying the rationality of the importance weight value through consistency test, calculating the importance weight value of each scene element in all the scene elements according to the hierarchy model of the scene elements and the importance weight value of each scene element of each hierarchy relative to the current hierarchy, and selecting the scene element with larger importance weight value as a decision variable according to the type of a target scene and the dimension requirement of an evaluation scene to be established.
Taking the scene element hierarchical model of the first step as input, solving the influence transfer times of the scene elements of the same hierarchy in each hierarchy of the automatic driving system through the influence transfer model, and showing an exemplary embodiment of the influence transfer model in fig. 4. The impact transfer model has three assumptions:
1) The influence of the scene element on the automatic driving system is transmitted step by step along with the hierarchy of the automatic driving system, and the influence is not attenuated along with the transmission of the hierarchy;
2) The influence of different types of scene elements on an automatic driving system is the same;
3) The magnitude of the influence of the scene element on the automatic driving system can be represented by the number of influence transfer times of the scene element among layers of the automatic driving system, the more the number of transfer times is, the larger the influence of the scene element is, and the influence and the number of transfer times are in a linear relation.
The number of impact transfers of scene elements in the impact transfer model in each level of the autopilot system can be calculated by the following equation:
Figure BDA0003933187860000141
wherein P (n) represents the number of times of influence transfer of a scene element in each level of the automatic driving system, n represents the number of element attributes of the scene element, E i Representing the number of times the scene element ith element attribute is transferred in each level of the automatic driving system.
And after the influence transfer times of each level scene element in the scene element level model are obtained according to the influence transfer model, calculating the difference value of the influence transfer times of the scene elements of the same level, and converting the difference value of the influence transfer times into scales in a 1-9 scale method according to the table 1.
Table 1: correspondence table for influencing transfer times difference and 1-9 scale method
Figure BDA0003933187860000142
The meaning of each scale representation in the 1-9 scale is shown in Table 2:
table 2:1-9 scale meaning table
Figure BDA0003933187860000151
After the difference value of the influence transfer times is converted into a scale in a 1-9 scale method, a judging matrix of the scene elements of the same level can be established, the maximum eigenvalue of the judging matrix and the eigenvector corresponding to the maximum eigenvalue are solved, and the eigenvector is normalized to obtain the importance weight value of the scene elements of the same level.
In order to verify the rationality of the importance weight value, consistency test is required to be performed on the judgment matrix, and the formula of the consistency test is as follows:
Figure BDA0003933187860000152
wherein CR represents the consistency index, RI is the standard value of the average random consistency index of the total hierarchical sequence, and different values are taken according to different orders of the judgment matrix, as shown in Table 3:
table 3: average random consistency index table for hierarchical total ordering
Figure BDA0003933187860000153
CI represents a hierarchical total ordering consistency index, and the calculation formula of CI is as follows:
Figure BDA0003933187860000154
wherein lambda is max Represents the maximum eigenvalue of the judgment matrix, and n is the order of the judgment matrix. When CR is<And 0.1, the consistency of the judgment matrix is good, and the consistency requirement is met.
Finally, according to the hierarchical model of the scene elements and the importance weight value of each scene element of each hierarchy relative to the hierarchy, calculating the importance weight value of each scene element in all the scene elements, then determining the type of the target scene to be researched, and according to the type of the target scene and the dimension requirement of the evaluation scene to be built, selecting some scene elements with larger importance weight values as decision variables, namely finishing the dimension optimization of the scene elements.
In the second step, the process of constructing the evaluation scene space and the key long tail function is as follows:
step one, constructing an ideal scene space based on scene element dimension optimization. The ideal test scene space is obtained by discretizing decision variables obtained by optimizing the element dimensions of the first step, each decision variable can be discretized into different values according to the element attributes of the decision variable, and the value range and the discrete step length of the decision variable are determined according to the importance weight value of the decision variable on the premise of fully considering the constraint of the real road condition. The decision variable with higher importance weight value should be set with larger value range and smaller discrete step length, so as to ensure that as many scenes as possible can be contained in the ideal test scene space; the decision variables with low importance weights are selected according to the principle that key scenes can be contained in an ideal test scene space, wherein the value range and the discrete step length of the decision variables are selected.
And secondly, constructing a real scene space based on natural driving data. The method comprises the steps of collecting scene data by a scene data collecting vehicle, driving the vehicle on a real road to collect the scene data, preprocessing the collected sensor data, screening target scene data from the preprocessed scene data, and screening data corresponding to decision variables from the target scene data to construct a real scene space.
An exemplary embodiment of the second step is shown in fig. 5. The method comprises the steps of carrying sensors such as a laser radar, a millimeter wave radar, a GPS high-precision inertial navigation, a high-definition camera, a vehicle-mounted CAN bus, a lane line sensor, a rainfall sensor, an illumination sensor and the like on a scene data acquisition vehicle, acquiring all sensor data according to a fixed acquisition period, wherein the acquired sensor data comprises the following data forms: the system comprises a space three-dimensional point cloud which is generated by a laser radar and takes a frame as a unit, an obstacle state list which is generated by a millimeter wave radar and takes a frame as a unit, positioning and gesture data which is generated by GPS high-precision inertial navigation and takes a time sequence as a unit, a color image which is generated by a high-definition camera and a lane line sensor and takes a frame as a unit, vehicle steering and motion state data which is generated by a vehicle-mounted CAN bus and takes a time sequence as a unit, and voltage data which is generated by a rainfall sensor and an illumination sensor and takes a time sequence as a unit. Furthermore, scene data acquisition should follow the following two principles:
1) The scene data acquisition area should contain cities and areas with different roads and traffic characteristics, and the total acquisition mileage should be large enough;
2) The scene data collection environment should include different weather types and illumination types.
The specific content of the sensor data preprocessing comprises: time alignment and spatial alignment of the sensor data; verifying the validity of the sensor data; an onboard bus alignment signal, a vehicle status alignment signal, and a multi-modal environmental sensor alignment signal are generated.
According to the type of the target scene selected in the first step, scene data of all target scenes are intercepted from all preprocessed data in a mode of manually watching videos, each segment of complete target scene data is called a scene working condition and represents a complete target scene event occurring on a real road, and data corresponding to decision variables in each scene working condition are screened out, so that a real scene space is formed.
And thirdly, constructing a key long tail function based on the scene occurrence probability and the scene risk. According to the self-adaptive dynamic scene generation principle that the automobile is challenging to automatically drive and has certain occurrence probability on a real road, a critical long tail function is designed based on the scene occurrence probability and the scene risk degree to serve as the basis for screening the critical self-adaptive dynamic scene. The calculation formula of the key long tail function is as follows:
I(x)＝V(x)·P(x) (4)
wherein, I (x) represents the critical long tail function value of the scene, P (x) represents the scene occurrence probability, and V (x) represents the scene risk.
The process of generating the self-adaptive dynamic scene facing the low-dimensional evaluation scene in the third step is as follows:
step one, solving the scene occurrence probability based on a convex combination algorithm. And converting the scene in the low-dimensional real scene space from the natural driving data into the scene in the low-dimensional ideal scene space through a convex combination algorithm, and solving the occurrence probability of the scene in the low-dimensional ideal scene space.
Let the presence vector { x } 1 ,x 2 ,x 3 ...x n E.g. real lambda i Not less than 0, i=1, 2, 3..n, and λ 1 +λ 2 +...+λ n Let 1 be λ 1 x 1 +λ 2 x 2 +...λ n x n Is a vector { x } 1 ,x 2 ,x 3 ...x n A convex combination of }.
Solving the scene occurrence probability based on a convex combination algorithm has the following assumptions:
1) In a discrete step range of an ideal scene space, the scene occurrence probability does not change suddenly, is continuously changed and follows a linear change rule;
2) In a discrete step range of an ideal scene space, the linear change rule of the scene occurrence probability can be represented by Euclidean distance between scenes in the ideal scene space;
3) In a discrete step range of an ideal scene space, the scene with the small Euclidean distance from the scene with high occurrence probability has high occurrence probability, and the scene with the large Euclidean distance from the scene with high occurrence probability has low occurrence probability.
An exemplary embodiment is shown in fig. 6 when the ideal scene space is a two-dimensional space. When the ideal scene space is a two-dimensional space, namely, the two decision variables are formed by discretization, taking the relative distance R and the relative speed Deltav as the decision variables as examples, the solving method of the scene occurrence probability in the two-dimensional ideal scene space is introduced. Taking a scene space within a discrete step as an example, A 1 (R A1 ,Δv A1 ),A 2 (R A2 ,Δv A2 ),A 3 (R A3 ,Δv A3 ),A 4 (R A4 ,Δv A4 ) Is 4 uniform discrete scenes in a two-dimensional ideal scene space, B (R B ,Δv B ) Is a certain scene in the real scene space obtained by natural driving data, and B is A 1 ，A 2 ，A 3 ，A 4 A rectangular interior is constructed. L (L) 11 Is B to A 2 、A 4 Euclidean distance of straight line, L 12 Is B to A 1 、A 3 Euclidean distance of straight line, L 21 Is B to A 3 、A 4 Euclidean distance of straight line, L 22 For B to A 1 、A 4 Euclidean distance of the straight line. Scene B in the real scene space can be converted into scene a in the two-dimensional ideal scene space by the convex combination algorithm 1 ，A 2 ，A 3 ，A 4 ：
B＝ω 1 ·A 1 +ω 2 ·A 2 +ω 3 ·A 3 +ω 4 ·A 4 (5)
Wherein omega is 1 ，ω 2 ，ω 3 ，ω 4 Respectively A 1 ，A 2 ，A 3 ，A 4 The weight coefficient of (2) is calculated as follows:
Figure BDA0003933187860000171
Figure BDA0003933187860000172
Figure BDA0003933187860000173
Figure BDA0003933187860000174
according to the method, all scenes in the real scene space can be converted into uniform discrete scenes in the ideal scene space, and the occurrence probability of the scenes in the two-dimensional ideal scene space can be obtained by counting the number of the uniform discrete scenes by combining the weight coefficients of the uniform discrete scenes.
An exemplary embodiment is shown in fig. 7 when the ideal scene space is a three-dimensional space. When the ideal scene space is a three-dimensional space, namely, three decision variables are formed by discretization, the solution method of scene occurrence probability in the three-dimensional ideal scene space is introduced by taking the relative distance R, the relative speed Deltav and the relative acceleration Deltaa as decision variables as examples. Taking a scene space within a discrete step as an example, A 1 (R A1 ,Δv A1 ,Δa A1 ),A 2 (R A2 ,Δv A2 ,Δa A2 ),A 3 (R A3 ,Δv A3 ,Δa A3 ),A 4 (R A4 ,Δv A4 ,Δa A4 ),A 5 (R A5 ,Δv A5 ,Δa A5 ),A 6 (R A6 ,Δv A6 ,Δa A6 ),A 7 (R A7 ,Δv A7 ,Δa A7 ),A 8 (R A8 ,Δv A8 ,Δa A8 ) Is 8 uniform discrete scenes in three-dimensional scene space, B (R B ,Δv B ,Δa B ) Is a certain scene in the real scene space obtained by natural driving data, and B is A 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 Inside the built-up cube. L (L) 11 Is B to A 2 ，A 4 ，A 6 ，A 8 Euclidean distance of plane, L 12 Is B to A 1 ，A 3 ，A 5 ，A 7 Euclidean distance of plane, L 21 Is B to A 3 ，A 4 ，A 7 ，A 8 Euclidean distance of plane, L 22 Is B to A 1 ，A 2 ，A 5 ，A 6 Euclidean distance of plane, L 31 Is B to A 5 ，A 6 ，A 7 ，A 8 Euclidean distance of plane, L 32 Is B to A 1 ，A 2 ，A 3 ，A 4 Euclidean distance of the plane in which the device is located. Scene B in the real scene space can be converted into scene a in the three-dimensional ideal scene space by the convex combination algorithm 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 ：
B＝ω 1 ·A 1 +ω 2 ·A 2 +ω 3 ·A 3 +ω 4 ·A 4 +ω 5 ·A 5 +ω 6 ·A 6 +ω 7 ·A 7 +ω 8 ·A 8 (10)
Wherein omega is 1 ，ω 2 ，ω 3 ，ω 4 ，ω 5 ，ω 6 ，ω 7 ，ω 8 Respectively A 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 The weight coefficient of (2) is calculated as follows:
Figure BDA0003933187860000181
Figure BDA0003933187860000182
Figure BDA0003933187860000183
Figure BDA0003933187860000184
Figure BDA0003933187860000185
Figure BDA0003933187860000186
Figure BDA0003933187860000187
Figure BDA0003933187860000188
according to the method, all scenes in the real scene space can be converted into uniform discrete scenes in the ideal scene space, and the occurrence probability of the scenes in the three-dimensional ideal scene space can be obtained by counting the number of the uniform discrete scenes by combining the weight coefficients of the uniform discrete scenes.
And secondly, solving scene risk boundaries based on a support vector regression algorithm. The scene risk boundaries in the real scene space are pre-divided according to the collision time, joint distribution of the collision time and the relative acceleration is constructed, boundary points which belong to the same boundary line in the joint distribution are used as a training sample set, the training sample set is independently input into a support vector regression algorithm, each training sample set corresponds to one scene risk boundary, all scene risk boundary lines are collected to obtain all scene risk boundaries, finally the scene risk boundaries in the real scene space are mapped into an ideal scene space, and risk coefficients are set for scenes with different risk degrees.
An exemplary implementation result of the third step is shown in fig. 8, taking the cut-in scene as the target scene as an example, an exemplary embodiment of the cut-in scene is shown in fig. 9, when the center of the target vehicle coincides with the lane line, the relative distance R, the relative velocity Δv and the relative acceleration Δa between the host vehicle and the target vehicle are selected as decision variables, an ideal scene space and a real scene space based on natural driving data are constructed, and the collision time (Time To Collision, TTC) of the scene in the real scene space is calculated. TTC represents the time required from the current moment to the collision of two vehicles while maintaining the current motion state, and the smaller the value, the higher the risk of representing a scene. TTC can be calculated by the following formula:
Figure BDA0003933187860000191
Where R represents the relative distance between the two vehicles, and Δv represents the relative speed between the host vehicle and the target vehicle.
The pre-dividing standard for setting the scene risk boundary is as follows: when TTC epsilon [0s,1s ], the risk level of the scene is collision scene (crash scenarios); when TTC epsilon (1 s,3 s), the danger level of the scene is an emergency scene (emergency scenarios); when TTC epsilon (3 s,5 s), the risk level of the scene is a conflict scene (conflict scenarios), and when TTC epsilon (5 s, ++ infinity) or (-infinity, 0 s), the risk level of the scene is a safety scene.
Constructing joint distribution of TTC and delta a in a real scene space, wherein the joint distribution is equivalent to four scene risk boundaries determined by TTC=0, TTC=1, TTC=3 and TTC=5, respectively taking boundary points on the left side and the right side of the joint distribution as a single training sample set, inputting the single training sample set into a support vector regression algorithm for learning, and selecting a linear kernel function for solving because the boundary of the joint distribution is approximately linear, so that the scene risk boundaries on the left side and the right side of the joint distribution can be obtained. Combining the scene risk boundary based on TTC pre-division with the scene risk boundary based on support vector regression to obtain the risk boundary in the real scene space as follows:
Figure BDA0003933187860000192
Wherein, I l And/l r Scene risk boundaries on the left side and the right side of a joint distribution diagram obtained through support vector regression learning respectively, l m1 、l m2 、l m3 And l m4 Scene risk boundaries obtained by pre-dividing according to TTC are respectively obtained.
Mapping the scene risk boundary into an ideal scene space, setting the scene risk corresponding to a collision scene as 1, setting the scene risk corresponding to an emergency scene as 0.7, setting the scene risk corresponding to a collision scene as 0.3, and setting the scene risk corresponding to a safety scene as 0.
And thirdly, generating a key self-adaptive dynamic scene based on a multi-starting-point optimization algorithm and a seed filling algorithm. And calculating a critical long tail function value in an ideal scene space through the scene occurrence probability and the scene risk, inputting the critical long tail function, the ideal scene space and a critical threshold value into a multi-start-point optimization algorithm, solving to obtain a local critical self-adaptive dynamic scene, inputting the local critical scene, the critical long tail function, the ideal scene space and the critical threshold value into a seed filling algorithm, and solving to obtain all the critical self-adaptive dynamic scenes.
An exemplary embodiment of a multi-start optimization algorithm is shown in fig. 10. And obtaining a critical long tail function of the scene in the ideal scene space through the scene occurrence probability obtained in the first step and the scene risk obtained in the second step. And (3) taking the critical long-tail function, the ideal scene space and the critical threshold gamma as inputs, inputting the inputs into a multi-start-point optimization algorithm, sampling a plurality of points in the ideal scene space as the start points of the algorithm in a manual setting or random mode, solving the maximum value of the critical long-tail function of the attraction domain where each start point is located, and outputting the scenes corresponding to the maximum values of all the critical long-tail functions larger than gamma to obtain the local key self-adaptive dynamic scene facing the low-dimensional evaluation scene.
An exemplary embodiment of a seed filling algorithm is shown in fig. 11. And taking the local key scenes, the key long tail functions, the ideal scene space and the key threshold gamma which are output by the multi-starting-point optimization algorithm as inputs, inputting the inputs into the seed filling algorithm, taking each local key scene as a starting point, searching scenes with the key long tail function value larger than gamma in a neighborhood around the starting point in the ideal scene space, taking the scenes as new starting points, continuously searching the neighborhood around the scenes, repeating the steps until the key long tail function values of the neighborhood scenes around all the starting points are smaller than gamma, ending the algorithm, and outputting all the scenes marked as the starting points, thereby generating all the key self-adaptive dynamic scenes facing the low-dimensional evaluation scene.
The fourth step is that the process of self-adaptive dynamic scene generation facing to the high-dimensional evaluation scene is as follows:
step one, constructing a scene risk identification model based on a Hammerstein identification process and solving scene occurrence probability based on a convex combination algorithm. The Hammerstein identification process is formed by connecting static nonlinear links and dynamic linear links in series, taking the relative state quantity of the vehicle and the target vehicle and the acceleration of the target vehicle as the input of a scene risk identification model, taking the acceleration of the vehicle as the output of the model, training a large amount of scene data, taking the model parameters of the scene risk identification model as key parameters for representing the intrinsic attribute of the scene risk, decoupling and dimension reduction of the key parameters by adopting a principal component analysis method, clustering the scene risk by adopting an ant colony algorithm, selecting a plurality of representative 'state-action' pairs from each scene risk level, inputting the representative 'state-action' pairs into the scene risk identification model, and establishing the mapping relation between the scene risk level and a clustering result according to the clustering category where the parameters are located. The scene occurrence probability solving based on the convex combination algorithm can directly solve the occurrence probability of the 'state-action' pair according to the scene occurrence probability solving method facing the low-dimensional evaluation scene.
The high-dimensional evaluation scene is regarded as a Markov decision process, the relative state quantity of the vehicle and the target vehicle is regarded as a state, the acceleration of the target vehicle is regarded as an action, the state and the action of the same time step are regarded as a 'state-action' pair, the relative state quantity of the vehicle and the target vehicle at the same time and the acceleration of the target vehicle are regarded as inputs of a scene risk degree identification model, the acceleration of the vehicle is regarded as outputs of the model, and therefore the scene risk degree identification model is a multi-input single-output system. The Hammerstein identification process is formed by connecting a static nonlinear link and a dynamic linear link in series, wherein the static nonlinear link can be a dead zone function, an S-shaped function or a saturation function and other functions, and the z-transformation of the dynamic linear link is shown in the following formula:
A(z -1 )O p (k)＝B(z -1 )·z -d ·N(k) (21)
wherein O is p (k) Represents the acceleration set of the main vehicle, N (k) represents the output set of the static nonlinear link, d
Represents the input delay order and is defined as an integer multiple of the sampling time, a (z -1 ) And B (z) -1 ) Can be calculated by the following formula:
Figure BDA0003933187860000211
wherein (a) 1 ,…a q ) (b) 1 ,…b n ) Each coefficient is a dynamic linear link, and q and n are orders of the dynamic linear link.
After a large amount of scene data from natural driving data are trained, model parameters contained in the static nonlinear link and the dynamic linear link are key data for representing the intrinsic attribute of the scene risk, so that the model parameters are used as data samples for evaluating the scene risk. In order to ensure that the space dimension of the data sample is reduced as much as possible to improve the calculation efficiency on the premise of expressing the same model characteristics, a principal component analysis method is adopted to decouple and reduce the dimension of key parameters in the scene risk identification model.
Let H represent the parameter dimension of the scene risk recognition model, E represent the number of training data, then the model parameter data set X can be expressed as:
Figure BDA0003933187860000212
wherein x is i And an internal reference vector representing the scene risk identification model.
X is taken as an input and is input into a principal component analysis algorithm. Defining the percentage of the sum of the characteristic values of the first M principal components and the sum of all the characteristic values as the principal component contribution rate, the principal component cumulative contribution rate M m Can be calculated by the following formula:
Figure BDA0003933187860000213
wherein lambda is i Representing the feature vector.
To ensure the dimension reducing effect, take M m And (3) taking the m value which is more than or equal to 85% and corresponds to the m value as the dimension of the independent parameter of the model obtained by algorithm calculation, and finally obtaining an m multiplied by E dimension matrix L:
Figure BDA0003933187860000214
clustering scene dangers by adopting an ant colony algorithm, taking L as input of the ant colony algorithm, finding a division method for minimizing the sum of distances of each data sample to a clustering center with known clustering number in L, and classifying the scene dangers into 4 grades by referring to a scene danger classification mode of low-dimensional evaluation scenes in the step two of the third step, wherein the clustering number is 4, and the ant colony algorithm can be represented by the following formula:
Figure BDA0003933187860000215
where J represents the sum of the distances of each data sample to 4 cluster centers, l ip The p-th model parameter feature representing the i-th data sample, c jp The p model parameter characteristic of the j class center can be calculated by the following formula:
Figure BDA0003933187860000221
in the method, in the process of the invention,E j for the j-th class of the observed variables E, ω ij For the amount of the dependency mark between the observed variable and the class, the following formula can be used for calculation:
Figure BDA0003933187860000222
the classification quality of the ant colony clusters can be improved by iterative updating equations, and the updating equations are shown as follows:
Figure BDA0003933187860000223
wherein P is ij For probability of data sample cross class conversion, τ ij The normalized pheromone between the data sample i and the belonging class j can be obtained by the following formula:
Figure BDA0003933187860000224
where ρ represents the pheromone volatility and t represents the time step.
The clustering result of the state-action pair can be obtained by iteratively updating the equation, and the mapping relation between the clustering result and the scene risk level needs to be established because the clustering result has no physical meaning yet. And selecting a plurality of representative 'state-action' pairs from each scene risk level, inputting the representative 'state-action' pairs into a scene risk identification model, and establishing a mapping relation between the scene risk level and a clustering result according to the clustering category of the parameters of the representative 'state-action' pairs.
The scene occurrence probability solving based on the convex combination algorithm can directly solve the occurrence probability of the 'state-action' pair according to the scene occurrence probability solving method facing the low-dimensional evaluation scene.
And step two, reconstructing a key long tail function based on a Markov decision process. The ideal scene space of the high-dimensional evaluation scene is obtained by discretizing the decision variables, and the critical long-tail function is reconstructed in the form of a 'state-action' pair by regarding the high-dimensional evaluation scene as a Markov decision process.
An exemplary embodiment of a follow-up scenario is shown in fig. 12. Taking the following scene as a target scene for example, since the following scene cannot represent the working condition of the whole scene at key moments any more, but the process of dynamic change of the scene in a period of time is to be researched, when an ideal scene space is constructed by decision variables, the number of scenes in the ideal scene space increases exponentially with the increase of the dimension of the scene. Taking the initial speed of the target vehicle, the initial relative distance between the vehicle and the target vehicle, the initial relative speed between the vehicle and the target vehicle and the acceleration sequence of the target vehicle as decision variables x:
x＝[v 0 ,R 0 ,Δv 0 ,a 01 ,a 02 ,...,a 0k ] (31)
in the formula, v o Representing the initial speed of the target vehicle, R o Representing the initial relative distance between the host vehicle and the target vehicle, deltav o Representing the initial relative speed of the host vehicle and the target vehicle, a 0k Representing the acceleration of the target vehicle at the kth time step.
When v is set o The value range of (2) is [20m/s,40m/s ] ]The discrete step length is 2m/s; r is R o The range of the values is (0 m,90 m)]The discrete step length is 2m; deltav o The range of the value of (C) is [ -20m/s,20 m/s)]The discrete step length is 2m/s; a, a 0k The range of the value of (C) is [ -4m/s 2 ,2m/s 2 ]Discrete step size of 0.2m/s 2 The method comprises the steps of carrying out a first treatment on the surface of the When k is 10s, the number of ideal scene space scenes is 21×45×21×31 10 . To reduce the dimension of the ideal scene space and simplify the calculation complexity of the critical long-tail function, the high-dimension evaluation scene can be regarded as a Markov decision process, the relative distance and the relative speed between the host vehicle and the target vehicle are regarded as states, the acceleration of the target vehicle is regarded as motion, the acceleration of the target vehicle at a certain time step only depends on the relative states of the host vehicle and the target vehicle at the time step, therefore, the states and the motion at the same time step can be regarded as a whole, namely a 'state-motion' pair, the number of the 'state-motion' pairs in the ideal scene space is 21×45×21×31= 615195, compared with the prior ideal sceneThe number of scenes in the scene space is greatly reduced. The key long tail function is reconstructed as:
Figure BDA0003933187860000231
wherein s is i Representing the state at the ith time step, a i Represents the action at the ith time step, G (s i ,a i ) Can be calculated by the following formula:
G(s i ,a i )＝V(s i ,a i )·P(s i ,a i ) (33)
wherein V(s) i ,a i ) Represents s i And a i The corresponding risk of "state-action" pair, P (s i ,a i ) Represents s i And a i Probability of occurrence of the corresponding "state-action" pair.
And thirdly, generating a key self-adaptive dynamic scene based on a Q-learning algorithm. Constructing a Belman equation, solving an objective function of a Markov decision process through the Belman equation, and solving an optimal action sequence corresponding to each initial state through updating expected benefits, namely, a key self-adaptive dynamic scene.
Let Q(s) , a) For the expected benefit of an agent taking action a in state s, r is the return that the agent gives in the environment with action a taken, the Belman equation is used to solve the optimal strategy for the Markov decision process:
V π (s)＝E π [R t+1 +ξV π (S t+1 )|S t ＝s] (34)
wherein pi represents policy, S represents state set, R represents return set, ζ represents discount factor, t represents time step, S represents state at t moment, V π (s) represents a cost function. Q may be updated according to the following equation:
Figure BDA0003933187860000232
in the formula, A represents an action set, a represents an action at time t, and alpha represents a learning rate.
Thus, the optimal action sequence corresponding to each initial state, namely the key self-adaptive dynamic scene, can be obtained.
In the fifth step, the process of self-adaptive dynamic scene assessment facing to the high-low dimension assessment scene is as follows:
Step one, self-adaptive dynamic scene sampling based on an E-greedy sampling strategy. Sampling scenes from the self-adaptive dynamic scenes facing the high-low dimension evaluation scenes by adopting the same sampling method, taking the low-dimension self-adaptive dynamic scenes as an example, setting a small probability value epsilon, randomly sampling scenes from the low-dimension self-adaptive dynamic scenes by using the probability of 1 epsilon, randomly sampling scenes from the scenes except the low-dimension self-adaptive dynamic scenes in the low-dimension ideal scene space by using the probability of epsilon to form a test scene library, namely a test group I, and obtaining the test scene library under the high-dimension condition by using the same sampling method to form a test group II.
And step two, self-adaptive dynamic scene evaluation based on accident rate and test times. The first test group and the second test group respectively test the automatic driving automobile, a certain confidence coefficient is set, for example, 80 percent, until the accident rate converges, the test times and the accident rate when the accident rates of the two test groups under the confidence coefficient converge are respectively recorded; meanwhile, setting a comparison experiment, randomly sampling scenes from a low-dimensional ideal scene space and a high-dimensional ideal scene space respectively, marking the scenes as a first comparison group and a second comparison group respectively, testing the automatic driving automobile, setting the same confidence level until the accident rate converges, and recording the test times and the accident rate when the accident rates of the two comparison groups converge under the confidence level respectively. Comparing the accident rates of the first test group with the first comparison group, the second test group with the second comparison group and the number of times of testing when the accident rate is converged, if the accident rate of the first test group is far greater than the accident rate of the first comparison group and the number of times of testing when the accident rate is converged is far less than the accident rate of the first comparison group, the self-adaptive dynamic scene assessment for the low-dimensional test scene is effective; if the accident rate of the second test group is far greater than that of the second control group and the test times when the accident rate converges are far less than that of the second control group, the adaptive dynamic scene assessment for the high-dimensional test scene is effective.

Claims (1)

1. A general generation method of a self-adaptive automatic driving dynamic scene facing to high-low dimension evaluation scenes is characterized by comprising the following steps: the method comprises the following steps:
the first step, the dimension optimization of the evaluation scene element based on the chromatographic analysis method comprises the following specific processes:
step one, constructing a scene element hierarchical model based on scene parameterization, describing a scene as a set of scene elements, classifying the scene elements step by step according to basic attributes of the scene elements and the geographic position conditions to which the scene elements belong, and thus constructing the scene element hierarchical model;
the scene is the overall dynamic description of the automatic driving automobile and each element of the driving environment thereof in a period of time, and the scene elements are divided into static elements and dynamic elements according to the difference of the geographic positions of the scene elements, wherein the static elements and the dynamic elements are divided into two types of off-road elements and in-road elements from the basic attributes of the scene elements;
the road exterior elements of the static elements mainly comprise static roadside objects such as buildings, trees, green belts and traffic marks, and the road interior elements of the static elements mainly comprise elements related to lanes and elements related to environments, wherein the elements related to the lanes such as lane types, lane line types, lane numbers, lane widths, lane gradients, lane curvatures, lane line types, lane line widths, road surface types and road surface attachment coefficients; environmental related elements such as weather and light status;
The off-road elements of the dynamic element mainly include traffic participants, such as traffic participant types and numbers of traffic participants, and the on-road elements of the dynamic element mainly include elements related to traffic participants, elements related to initial states, and elements related to running state sequences, elements related to traffic participants, such as traffic participant types and numbers of traffic participants; elements related to the initial state such as an initial position, an initial lane, and an initial speed; elements associated with the driving state sequence such as trigger pattern, relative distance sequence, relative velocity sequence, and relative acceleration sequence;
step two, scene element dimension optimization based on an analytic hierarchy process, solving the influence transfer times of scene elements of the same level in each level of an automatic driving system through an influence transfer model according to the scene element level model, calculating the difference value of the influence transfer times of the scene elements of the same level, converting the difference value of the influence transfer times into a scale value according to the corresponding relation between the difference value of the influence transfer times and a 1-9 scale method, thereby establishing a judgment matrix of the scene elements of the same level, solving the maximum characteristic value of the judgment matrix and the corresponding characteristic vector thereof, normalizing the characteristic vector to obtain the importance weight value of each scene element of the same level relative to the level, verifying the rationality of the importance weight value through consistency test, calculating the importance weight value of each scene element in all the scene elements according to the level model of the scene element and the importance weight value of each scene element of each level relative to the level, and selecting the scene element with larger importance weight value as a decision variable according to the type of a target scene and the dimension requirement of an evaluation scene to be established;
Taking the scene element hierarchical model in the first step as input, solving the influence transfer times of the scene elements of the same hierarchy in each hierarchy of the automatic driving system through an influence transfer model, wherein the influence transfer model has three assumptions, and the method comprises the following specific steps:
1) The influence of scene elements on the automatic driving system is transmitted step by step along with the hierarchy of the automatic driving system, and the influence is not attenuated along with the transmission of the hierarchy;
2) The influence of different types of scene elements on an automatic driving system is the same;
3) The influence of the scene element on the automatic driving system is represented by the influence transfer times of the scene element among layers of the automatic driving system, the more the transfer times are, the larger the influence of the scene element is, and the influence and the transfer times are in a linear relation;
the number of impact transfers of scene elements in the impact transfer model in each level of the autopilot system can be calculated by the following equation:
Figure FDA0003933187850000021
wherein P (n) represents the number of times of influence transfer of a scene element in each level of the automatic driving system, n represents the number of element attributes of the scene element, E i Representing the influence transfer times of the ith element attribute of the scene element in each level of the automatic driving system;
After the influence transfer times of all the level scene elements in the scene element level model are obtained according to the influence transfer model, calculating the difference value of the influence transfer times of the scene elements of the same level, converting the difference value of the influence transfer times into a scale in a 1-9 scale method, after converting the difference value of the influence transfer times into a scale in a 1-9 scale method, establishing a judgment matrix of the scene elements of the same level, solving the maximum eigenvalue of the judgment matrix and the eigenvector corresponding to the maximum eigenvalue, and normalizing the eigenvector to obtain the importance weight value of the scene elements of the same level;
in order to verify the rationality of the importance weight value, consistency test is carried out on the judgment matrix, and the formula of the consistency test is as follows:
Figure FDA0003933187850000031
wherein CR represents a consistency index, RI is a standard value of a hierarchical total ordering average random consistency index, and different values are taken according to different orders of a judgment matrix;
CI represents a hierarchical total ordering consistency index, and the calculation formula of CI is as follows:
Figure FDA0003933187850000032
wherein lambda is max Represents the maximum eigenvalue of the judgment matrix, n is the order of the judgment matrix, and when CR<0.1, the consistency of the judgment matrix is good, and the consistency is satisfiedRequirements;
finally, according to the hierarchical model of the scene elements and the importance weight value of each scene element of each hierarchy relative to the hierarchy, calculating the importance weight value of each scene element in all the scene elements, then determining the type of a target scene to be researched, and according to the type of the target scene and the dimension requirement of an evaluation scene to be built, selecting some scene elements with larger importance weight values as decision variables, namely finishing the dimension optimization of the scene elements;
Secondly, constructing a scene space and a key long tail function, wherein the specific process is as follows:
the method comprises the steps that firstly, an ideal scene space based on scene element dimension optimization is constructed, an ideal test scene space is obtained by discretizing decision variables obtained through the scene element dimension optimization in the first step, each decision variable is discretized into different values according to element attributes of the decision variable, the value range and the discrete step length of the decision variable are determined according to the importance weight value of the decision variable under the premise of fully considering the constraint of real road conditions, and the decision variable with higher importance weight value is set with a larger value range and a smaller discrete step length, so that the situation that as many scenes as possible can be contained in the ideal test scene space is ensured; the decision variables with low importance weights are selected according to the principle that key scenes can be contained in an ideal test scene space, wherein the value range and the discrete step length of the decision variables are selected;
step two, constructing a real scene space based on natural driving data, namely acquiring the scene data by acquiring the real scene space, wherein the vehicle runs on a real road through the scene data to acquire the scene data, preprocessing the acquired sensor data, screening target scene data from the preprocessed scene data, and screening data corresponding to decision variables from the target scene data to construct the real scene space;
Scene data acquisition should also follow the following two principles:
1) The scene data acquisition area should contain cities and areas with different roads and traffic characteristics, and the total acquisition mileage is enough;
2) The scene data acquisition environment comprises different weather types and illumination types;
the specific content of the sensor data preprocessing comprises: time alignment and spatial alignment of the sensor data; verifying the validity of the sensor data; generating a vehicle bus alignment signal, a vehicle state alignment signal and a multi-modal environmental sensor alignment signal;
according to the type of the target scene selected in the first step, capturing scene data of all target scenes from all preprocessed data in a mode of manually watching videos, wherein each segment of complete target scene data is called a scene working condition and represents a complete target scene event occurring on a real road, and data corresponding to decision variables in each scene working condition are screened out to form a real scene space;
step three, constructing a critical long-tail function based on the scene occurrence probability and the scene risk, and designing the critical long-tail function based on the scene occurrence probability and the scene risk as a basis for screening a critical adaptive dynamic scene according to an adaptive dynamic scene generation principle that the automatic driving automobile is challenging and has a certain occurrence probability on a real road, wherein the calculation formula of the critical long-tail function is as follows:
I(x)＝V(x)·P(x) (4)
Wherein, I (x) represents a critical long tail function value of a scene, P (x) represents a scene occurrence probability, and V (x) represents a scene risk;
thirdly, generating a self-adaptive dynamic scene facing to the low-dimensional evaluation scene, wherein the specific process is as follows:
firstly, solving the occurrence probability of a scene based on a convex combination algorithm, converting the scene in a low-dimensional real scene space from natural driving data into a scene in a low-dimensional ideal scene space through the convex combination algorithm, and solving the occurrence probability of the scene in the low-dimensional ideal scene space;
let the presence vector { x } 1 ,x 2 ,x 3 ...x n E.g. real lambda i Not less than 0, i=1, 2, 3..n, and λ 1 +λ 2 +...+λ n Let 1 be λ 1 x 1 +λ 2 x 2 +...λ n x n Is a vector { x } 1 ,x 2 ,x 3 ...x n A convex combination of };
solving the scene occurrence probability based on a convex combination algorithm has the following assumptions:
1) The scene occurrence probability does not change suddenly in a discrete step range of the ideal scene space, is continuously changed and follows a linear change rule;
2) In a discrete step range of an ideal scene space, the linear change rule of the scene occurrence probability can be represented by Euclidean distance between scenes in the ideal scene space;
3) In a discrete step range of an ideal scene space, the scene with a short Euclidean distance with a high occurrence probability has high occurrence probability, and the scene with a long Euclidean distance with the high occurrence probability has low occurrence probability;
Taking a scene space within a discrete step as an example, A 1 (R A1 ,Δv A1 ),A 2 (R A2 ,Δv A2 ),A 3 (R A3 ,Δv A3 ),A 4 (R A4 ,Δv A4 ) Is 4 uniform discrete scenes in a two-dimensional ideal scene space, B (R B ,Δv B ) Is a certain scene in the real scene space obtained by natural driving data, and B is A 1 ，A 2 ，A 3 ，A 4 Rectangular interior, L 11 Is B to A 2 、A 4 Euclidean distance of straight line, L 12 Is B to A 1 、A 3 Euclidean distance of straight line, L 21 Is B to A 3 、A 4 Euclidean distance of straight line, L 22 For B to A 1 、A 4 The Euclidean distance of the straight line can be converted into a scene A in the two-dimensional ideal scene space by a convex combination algorithm 1 ，A 2 ，A 3 ，A 4 ；
B＝ω 1 ·A 1 +ω 2 ·A 2 +ω 3 ·A 3 +ω 4 ·A 4 (5)
Wherein omega is 1 ，ω 2 ，ω 3 ，ω 4 Respectively A 1 ，A 2 ，A 3 ，A 4 The weight coefficient of (2) is calculated as follows:
Figure FDA0003933187850000061
/>
Figure FDA0003933187850000062
Figure FDA0003933187850000063
Figure FDA0003933187850000064
according to the method, all scenes in the real scene space are converted into uniform discrete scenes in the ideal scene space, and the number of the uniform discrete scenes is counted by combining the weight coefficients of the uniform discrete scenes, so that the occurrence probability of the scenes in the two-dimensional ideal scene space is obtained;
taking a scene space within a discrete step as an example, A 1 (R A1 ,Δv A1 ,Δa A1 ),A 2 (R A2 ,Δv A2 ,Δa A2 ),A 3 (R A3 ,Δv A3 ,Δa A3 ),A 4 (R A4 ,Δv A4 ,Δa A4 ),A 5 (R A5 ,Δv A5 ,Δa A5 ),A 6 (R A6 ,Δv A6 ,Δa A6 ),A 7 (R A7 ,Δv A7 ,Δa A7 ),A 8 (R A8 ,Δv A8 ,Δa A8 ) Is 8 uniform discrete scenes in three-dimensional scene space, B (R B ,Δv B ,Δa B ) Is a certain scene in the real scene space obtained by natural driving data, and B is A 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 Inside the cube of the structure, L 11 Is B to A 2 ，A 4 ，A 6 ，A 8 Euclidean distance of plane, L 12 Is B to A 1 ，A 3 ，A 5 ，A 7 Euclidean distance of plane, L 21 Is B to A 3 ，A 4 ，A 7 ，A 8 Euclidean distance of plane, L 22 Is B to A 1 ，A 2 ，A 5 ，A 6 Euclidean distance of plane, L 31 Is B to A 5 ，A 6 ，A 7 ，A 8 Euclidean distance of plane, L 32 Is B to A 1 ，A 2 ，A 3 ，A 4 The Euclidean distance of the plane, the scene B in the real scene space can be converted into the scene A in the three-dimensional ideal scene space through a convex combination algorithm 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 ；
B＝ω 1 ·A 1 +ω 2 ·A 2 +ω 3 ·A 3 +ω 4 ·A 4 +ω 5 ·A 5 +ω 6 ·A 6 +ω 7 ·A 7 +ω 8 ·A 8 (10)
Wherein omega is 1 ，ω 2 ，ω 3 ，ω 4 ，ω 5 ，ω 6 ，ω 7 ，ω 8 Respectively A 1 ，A 2 ，A 3 ，A 4 ，A 5 ，A 6 ，A 7 ，A 8 The weight coefficient of (2) is calculated as follows:
Figure FDA0003933187850000071
Figure FDA0003933187850000072
Figure FDA0003933187850000073
Figure FDA0003933187850000074
Figure FDA0003933187850000075
Figure FDA0003933187850000076
Figure FDA0003933187850000077
Figure FDA0003933187850000078
according to the method, all scenes in the real scene space are converted into uniform discrete scenes in the ideal scene space, and the number of the uniform discrete scenes is counted by combining the weight coefficients of the uniform discrete scenes, so that the occurrence probability of the scenes in the three-dimensional ideal scene space is obtained;
step two, solving scene risk boundaries based on a support vector regression algorithm, pre-dividing scene risk boundaries in a real scene space according to collision time, constructing joint distribution of the collision time and relative acceleration, taking boundary points which belong to the same boundary line in the joint distribution as a training sample set, independently inputting the training sample set into the support vector regression algorithm, enabling each training sample set to correspond to one scene risk boundary, integrating all scene risk boundary lines to obtain all scene risk boundaries, mapping the scene risk boundaries in the real scene space into an ideal scene space, and setting risk coefficients for scenes with different risk degrees;
TTC represents the time required from the current moment to collision under the condition of keeping the current motion state unchanged, the smaller the value of the TTC is, the higher the danger of representing a scene is, and the TTC is calculated by the following formula:
Figure FDA0003933187850000081
wherein R represents the relative distance between two vehicles, and Deltav represents the relative speed between the vehicle and the target vehicle;
the pre-dividing standard for setting the scene risk boundary is as follows: when TTC is E [0s,1s ], the risk level of the scene is a collision scene; the scene risk level is an emergency scene when TTC (1 s,3 s), the scene risk level is a conflict scene when TTC (3 s,5 s), and the scene risk level is a safety scene when TTC (5 s, + -infinity) or (- + -infinity, 0 s);
constructing a joint distribution of TTC and delta a in a real scene space, wherein the joint distribution is equivalent to four scene risk boundaries determined by TTC=0, TTC=1, TTC=3 and TTC=5, respectively taking boundary points on the left side and the right side of the joint distribution as a single training sample set, inputting the single training sample set into a support vector regression algorithm for learning, selecting a linear kernel function for solving because the boundary of the joint distribution is nearly linear, obtaining scene risk boundaries on the left side and the right side of the joint distribution, combining the scene risk boundaries based on TTC pre-division with scene risk boundaries based on support vector regression, and obtaining the risk boundaries in the real scene space, wherein the risk boundaries are as follows:
Figure FDA0003933187850000082
Wherein, I l And/l r Scene risk boundaries on the left side and the right side of a joint distribution diagram obtained through support vector regression learning respectively, l m1 、l m2 、l m3 And l m4 Scene risk boundaries obtained by pre-dividing according to TTC;
mapping the scene risk boundary into an ideal scene space, setting the scene risk corresponding to a collision scene as 1, setting the scene risk corresponding to an emergency scene as 0.7, setting the scene risk corresponding to a collision scene as 0.3, and setting the scene risk corresponding to a safety scene as 0;
step three, generating a key self-adaptive dynamic scene based on a multi-starting-point optimization algorithm and a seed filling algorithm, calculating a key long-tail function value in an ideal scene space through the scene occurrence probability and the scene risk, inputting the key long-tail function, the ideal scene space and a key threshold value into the multi-starting-point optimization algorithm, solving to obtain a local key self-adaptive dynamic scene, inputting the local key scene, the key long-tail function, the ideal scene space and the key threshold value into the seed filling algorithm, and solving to obtain all the key self-adaptive dynamic scenes;
obtaining a critical long-tail function of a scene in an ideal scene space through the scene occurrence probability obtained in the first step and the scene risk obtained in the second step, inputting the critical long-tail function, the ideal scene space and a critical threshold gamma as inputs into a multi-starting-point optimization algorithm, sampling a plurality of points in the ideal scene space as starting points of the algorithm through a manual setting or random mode, solving the maximum value of the critical long-tail function of an attraction domain where each starting point is located, and outputting the scenes corresponding to the maximum value of all the critical long-tail functions larger than gamma to obtain a local key self-adaptive dynamic scene facing the low-dimensional evaluation scene;
The method comprises the steps of taking a local key scene, a key long tail function, an ideal scene space and a key threshold gamma which are output by a multi-starting-point optimization algorithm as inputs, inputting the inputs into a seed filling algorithm, taking each local key scene as a starting point, searching scenes with the key long tail function value larger than gamma in a neighborhood around the starting point in the ideal scene space, taking the scenes as new starting points, continuously searching the neighborhood around the scenes, repeating the steps until the key long tail function values of the neighborhood scenes around all the starting points are smaller than gamma, ending the algorithm, and outputting all the scenes marked as the starting points to generate all key self-adaptive dynamic scenes facing to a low-dimensional evaluation scene;
fourth, self-adaptive dynamic scene generation facing to high-dimensional evaluation scene comprises the following specific processes:
firstly, constructing a scene risk identification model based on a Hammerstein identification process and solving the scene occurrence probability based on a convex combination algorithm, wherein the Hammerstein identification process is formed by connecting a static nonlinear link and a dynamic linear link in series, taking the relative state quantity of a host vehicle and a target vehicle and the acceleration of the target vehicle as the input of the scene risk identification model, taking the acceleration of the host vehicle as the output of the model, training a large amount of scene data, taking the model parameters of the scene risk identification model as key parameters for representing the intrinsic attribute of the scene risk, decoupling and reducing the key parameters by adopting a principal component analysis method, clustering the scene risk by adopting an ant colony algorithm, selecting a plurality of representative 'state-action' pairs from each scene risk level, inputting the representative pairs into the scene risk identification model, establishing the mapping relation between the scene risk level and a clustering result according to the clustering category in which the parameters are positioned, and solving the scene occurrence probability based on the convex combination algorithm can directly solve the occurrence probability of the scene occurrence probability pair according to a method for low-dimensional evaluation scene;
The method comprises the steps that a high-dimensional evaluation scene is regarded as a Markov decision process, the relative state quantity of a host vehicle and a target vehicle is regarded as a state, the acceleration of the target vehicle is regarded as a motion, the state and the motion of the same time step are regarded as a 'state-motion' pair, the relative state quantity of the host vehicle and the target vehicle and the acceleration of the target vehicle at the same time are regarded as the input of a scene risk degree identification model, the acceleration of the host vehicle is regarded as the output of the model, the scene risk degree identification model is a multi-input single-output system, the Hammerstein identification process is formed by connecting a static nonlinear link and a dynamic linear link in series, the static nonlinear link is a dead zone function, an S-shaped function or a saturation function, and the z-transformation of the dynamic linear link is shown in the following formula:
A(z -1 )O p (k)＝B(z -1 )·z -d ·N(k) (21)
wherein O is p (k) Represents the set of acceleration of the main vehicle, N (k) represents the output set of the static nonlinear link, d represents the input delay order and is defined as an integer multiple of the sampling time, A (z) -1 ) And B (z) -1 ) Can be calculated by the following formula:
Figure FDA0003933187850000101
wherein (a) 1 ,…a q ) (b) 1 ,…b n ) Each coefficient is a dynamic linear link, and q and n are orders of the dynamic linear link;
after a large number of scene data from natural driving data are trained, model parameters contained in a static nonlinear link and a dynamic linear link are key data for representing the intrinsic attribute of the scene risk, so the model parameters are used as data samples for evaluating the scene risk, the space dimension of the data samples is reduced as much as possible to improve the calculation efficiency on the premise of expressing the same model characteristics, and a principal component analysis method is adopted for decoupling and dimension reduction of the key parameters in a scene risk identification model;
Let H represent the parameter dimension of the scene risk recognition model, E represent the number of training data, then the model parameter data set X can be expressed as:
Figure FDA0003933187850000111
wherein x is i An internal reference vector representing a scene risk identification model;
taking X as input, inputting into a principal component analysis algorithm, and defining the percentage of the sum of the characteristic values of the first M principal components and the sum of all the characteristic values as principal component contribution rate, and then accumulating the principal component contribution rate M m Calculated by the following formula:
Figure FDA0003933187850000112
wherein lambda is i Representing the feature vector;
to ensure the dimension reducing effect, take M m And (3) taking the m value which is more than or equal to 85% and corresponds to the m value as the dimension of the independent parameter of the model obtained by algorithm calculation, and finally obtaining an m multiplied by E dimension matrix L:
Figure FDA0003933187850000113
clustering scene dangers by adopting an ant colony algorithm, taking L as input of the ant colony algorithm, finding a division method for minimizing the sum of distances of each data sample to a clustering center with known clustering number in L, and classifying the scene dangers into 4 grades by referring to a scene danger classification mode of low-dimensional evaluation scenes in the step two of the third step, wherein the clustering number is 4, and the ant colony algorithm is represented by the following formula:
Figure FDA0003933187850000121
where J represents the sum of the distances of each data sample to 4 cluster centers, l ip The p-th model parameter feature representing the i-th data sample, c jp The p model parameter characteristic of the j class center is calculated by the following formula:
Figure FDA0003933187850000122
wherein E is j For the j-th class of the observed variables E, ω ij For the amount of the dependency mark between the observed variable and the category, the following formula is used for calculation:
Figure FDA0003933187850000123
the classification quality of the ant colony clusters can be improved by iterative updating equations, and the updating equations are shown as follows:
Figure FDA0003933187850000124
/>
wherein P is ij For probability of data sample cross class conversion, τ ij The normalized pheromone between the data sample i and the belonging class j is obtained by the following calculation:
Figure FDA0003933187850000125
wherein ρ represents the volatility of the pheromone and t represents the time step;
the method comprises the steps of obtaining a clustering result of a 'state-action' pair through iterative updating of an equation, and establishing a mapping relation between the clustering result and scene risk levels because the clustering result does not have physical significance, selecting a plurality of representative 'state-action' pairs from each scene risk level, inputting the representative 'state-action' pairs into a scene risk identification model, and establishing the mapping relation between the scene risk level and the clustering result according to the clustering category of the parameter;
the scene occurrence probability solving based on the convex combination algorithm directly solves the occurrence probability of a 'state-action' pair according to a method for solving the scene occurrence probability of a low-dimensional evaluation scene;
Step two, reconstructing a critical long-tail function based on a Markov decision process, obtaining an ideal scene space of a high-dimensional evaluation scene by discretizing a decision variable, and reconstructing the critical long-tail function in a form of a 'state-action' pair by regarding the high-dimensional evaluation scene as the Markov decision process;
taking the initial speed of the target vehicle, the initial relative distance between the vehicle and the target vehicle, the initial relative speed between the vehicle and the target vehicle and the acceleration sequence of the target vehicle as decision variables x:
x＝[v 0 ,R 0 ,Δv 0 ,a 01 ,a 02 ,...,a 0k ] (31)
in the formula, v o Representing the initial speed of the target vehicle, R o Representing the initial relative distance between the host vehicle and the target vehicle, deltav o Representing the initial relative speed of the host vehicle and the target vehicle, a 0k Representing the acceleration of the target vehicle at the kth time step;
when v is set o The value range of (2) is [20m/s,40m/s ]]The discrete step length is 2m/s; r is R o The range of the values is (0 m,90 m)]The discrete step length is 2m; deltav o The range of the value of (C) is [ -20m/s,20 m/s)]The discrete step length is 2m/s; a, a 0k The range of the value of (C) is [ -4m/s 2 ,2m/s 2 ]Discrete step size of 0.2m/s 2 The method comprises the steps of carrying out a first treatment on the surface of the When k is 10s, the number of ideal scene space scenes is 21×45×21×31 10 In order to reduce the dimension of an ideal scene space and simplify the calculation complexity of a critical long-tail function, consider a high-dimension evaluation scene as a markov decision process, consider the relative distance and relative speed between a host vehicle and a target vehicle as states, consider the acceleration of the target vehicle as an action, the acceleration of the target vehicle at a certain time step only depends on the relative states of the host vehicle and the target vehicle at the time step, so that the states and the actions at the same time step are regarded as a whole, namely a 'state-action' pair, the number of the 'state-action' pairs in the ideal scene space is 21×45×21×31= 615195, compared with the number of the scenes in the prior ideal scene space, the number of the 'state-action' pairs in the ideal scene space is greatly reduced, and the critical long-tail function is reconstructed as follows:
Figure FDA0003933187850000141
Wherein s is i Representing the state at the ith time step, a i Represents the action at the ith time step, G (s i ,a i ) Can be calculated by the following formula:
G(s i ,a i )＝V(s i ,a i )·P(s i ,a i ) (33)
wherein V(s) i ,a i ) Represents s i And a i The corresponding risk of "state-action" pair, P (s i ,a i ) Represents s i And a i The probability of occurrence of the corresponding "state-action" pair;
thirdly, generating a key self-adaptive dynamic scene based on a Q-learning algorithm, constructing a Belman equation, solving an objective function of a Markov decision process through the Belman equation, and solving an optimal action sequence corresponding to each initial state through updating expected benefits, namely the key self-adaptive dynamic scene;
let Q (s, a) be the expected benefit of an agent taking action a in state s, r be the return that the agent gives in the environment with action a taken, the bellman equation is used to solve the optimal strategy for the markov decision process:
V π (s)＝E π [R t+1 +ξV π (S t+1 )|S t ＝s] (34)
wherein pi represents policy, S represents state set, R represents return set, ζ represents discount factor, t represents time step, S represents state at t moment, V π (s) represents a cost function, and Q may be updated according to the following equation:
Figure FDA0003933187850000142
wherein A represents an action set, a represents an action at time t, and alpha represents a learning rate;
the optimal action sequence corresponding to each initial state, namely a key self-adaptive dynamic scene, can be obtained;
Fifthly, self-adaptive dynamic scene assessment for high-low dimension assessment scenes comprises the following specific processes:
step one, sampling self-adaptive dynamic scenes based on an E-greedy sampling strategy, sampling scenes from the self-adaptive dynamic scenes facing high-low dimensional evaluation scenes by adopting the same sampling method, taking a low-dimensional self-adaptive dynamic scene as an example, setting a small probability value E, randomly sampling scenes from the low-dimensional self-adaptive dynamic scene with the probability of 1-E, randomly sampling scenes in scenes except for the low-dimensional self-adaptive dynamic scene in a low-dimensional ideal scene space with the probability of E, forming a test scene library, namely a test group one, and naming the test scene library obtained under the high-dimensional condition as a test group two by adopting the same sampling method;
step two, based on the self-adaptive dynamic scene evaluation of the accident rate and the test times, the test group I and the test group II respectively test the automatic driving automobile, set a certain confidence coefficient, such as 80 percent, until the accident rate converges, respectively record the test times and the accident rate when the accident rates of the two test groups under the confidence coefficient converge; meanwhile, setting a comparison experiment, namely randomly sampling scenes from a low-dimensional ideal scene space and a high-dimensional ideal scene space, respectively marking the scenes as a first comparison group and a second comparison group, testing an automatic driving automobile, setting the same confidence level until the accident rate converges, respectively recording the test times and the accident rate when the accident rates of the two comparison groups converge under the confidence level, comparing the accident rates of the first comparison group, the second comparison group and the second comparison group with the test times when the accident rate converges, and if the accident rate of the first comparison group is far greater than the accident rate of the first comparison group and the test times when the accident rate converges is far less than the accident rate of the first comparison group, indicating that the self-adaptive dynamic scene evaluation for the low-dimensional test scene is effective; if the accident rate of the second test group is far greater than that of the second control group and the test times when the accident rate converges are far less than that of the second control group, the adaptive dynamic scene assessment for the high-dimensional test scene is effective.
CN202211396377.4A 2022-11-09 2022-11-09 Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene Pending CN116205024A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211396377.4A CN116205024A (en) 2022-11-09 2022-11-09 Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211396377.4A CN116205024A (en) 2022-11-09 2022-11-09 Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene

Publications (1)

Publication Number Publication Date
CN116205024A true CN116205024A (en) 2023-06-02

Family

ID=86516195

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211396377.4A Pending CN116205024A (en) 2022-11-09 2022-11-09 Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene

Country Status (1)

Country Link
CN (1) CN116205024A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116665456A (en) * 2023-07-27 2023-08-29 交通运输部科学研究院 Method for evaluating traffic state by combining high-dimensional index dimension reduction processing
CN116977810A (en) * 2023-09-25 2023-10-31 之江实验室 Multi-mode post-fusion long tail category detection method and system
CN117975736A (en) * 2024-03-29 2024-05-03 北京市计量检测科学研究院 Unmanned vehicle road cooperative application scene test method and system

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116665456A (en) * 2023-07-27 2023-08-29 交通运输部科学研究院 Method for evaluating traffic state by combining high-dimensional index dimension reduction processing
CN116665456B (en) * 2023-07-27 2023-12-19 交通运输部科学研究院 Method for evaluating traffic state by combining high-dimensional index dimension reduction processing
CN116977810A (en) * 2023-09-25 2023-10-31 之江实验室 Multi-mode post-fusion long tail category detection method and system
CN116977810B (en) * 2023-09-25 2024-01-09 之江实验室 Multi-mode post-fusion long tail category detection method and system
CN117975736A (en) * 2024-03-29 2024-05-03 北京市计量检测科学研究院 Unmanned vehicle road cooperative application scene test method and system
CN117975736B (en) * 2024-03-29 2024-06-07 北京市计量检测科学研究院 Unmanned vehicle road cooperative application scene test method and system

Similar Documents

Publication Publication Date Title
US20210197851A1 (en) Method for building virtual scenario library for autonomous vehicle
CN116205024A (en) Self-adaptive automatic driving dynamic scene general generation method for high-low dimension evaluation scene
Wang et al. A Vision‐Based Video Crash Detection Framework for Mixed Traffic Flow Environment Considering Low‐Visibility Condition
US20190179979A1 (en) Simulated Sensor Testing
CN110009765A (en) A kind of automatic driving vehicle contextual data system and scene format method for transformation
CN111709381A (en) Road environment target detection method based on YOLOv3-SPP
CN111460919A (en) Monocular vision road target detection and distance estimation method based on improved YO L Ov3
CN110356412A (en) The method and apparatus that automatically rule for autonomous driving learns
CN107985189A (en) Towards driver&#39;s lane change Deep Early Warning method under scorch environment
US11887324B2 (en) Cross-modality active learning for object detection
CN111428558A (en) Vehicle detection method based on improved YO L Ov3 method
DE102021124913A1 (en) METRIC BACKPROPAGATION FOR EVALUATION OF SUBSYSTEMS PERFORMANCE
CN115662166B (en) Automatic driving data processing method and automatic driving traffic system
CN112489072B (en) Vehicle-mounted video perception information transmission load optimization method and device
CN110427993A (en) Bullet train navigational gap localization method based on meteorologic parameter
EP4257443A1 (en) Method and system for automatic driving data collection and closed-loop management
CN114926984B (en) Real-time traffic conflict collection and road safety evaluation method
CN115544888A (en) Dynamic scene boundary assessment method based on physical mechanism and machine learning hybrid theory
Chuanxia et al. Machine learning and IoTs for forecasting prediction of smart road traffic flow
CN116597690B (en) Highway test scene generation method, equipment and medium for intelligent network-connected automobile
CN117521389A (en) Vehicle perception test method based on vehicle-road collaborative perception simulation platform
CN116630702A (en) Pavement adhesion coefficient prediction method based on semantic segmentation network
CN114926823B (en) WGCN-based vehicle driving behavior prediction method
US20220188621A1 (en) Generative domain adaptation in a neural network
CN115451987A (en) Path planning learning method for automatic driving automobile

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination