Knowledge migration theory-based automatic driving evaluation key scene construction method

Abstract

The invention discloses a knowledge migration theory-based automatic driving evaluation key scene construction method. The method comprises the following steps of: 1, designing an automatic driving evaluation scene specification based on an ontology method; 2, constructing a basic automatic driving evaluation scene knowledge base; 3, expanding an automatic driving evaluation scene knowledge base based on knowledge migration; and 4, integrating and testing the key scene knowledge base. The method has the beneficial effects that a basic automatic driving evaluation scene knowledge base is constructed, the performance of an automatic driving system in a key scene is tested, and development and iteration of an algorithm are accelerated; a rich semantic information set is constructed; the coverage degree and generalization ability of the scene are improved; according to the method, the key scene knowledge base is constructed, the key scene knowledge base is sampled to obtain the representative robbing key scene, and the performance of the automatic driving system is verified.

Description

Technical Field

[0001] The present invention relates to a method for constructing key scenarios for autonomous driving evaluation, and in particular to a method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory. Background Art

[0002] At present, with the continuous development of autonomous driving technology, higher requirements are put forward for the safety and adaptability of autonomous driving systems. The implementation of safer and more reliable autonomous driving technology is inseparable from a large number of safety tests. In order to solve the problems of poor safety, high cost, time and labor caused by traditional open road tests or closed test sites for autonomous driving vehicles, the use of simulation modeling technology to generate autonomous driving test scenarios and improve the safety and reliability of autonomous driving systems is a hot topic of current research. By processing natural driving data to achieve scene reconstruction, a large amount of data can be generated to evaluate autonomous driving systems. However, the reproduction scenarios based on natural driving data have the problems of strong security and data redundancy, resulting in low test efficiency and increased computing power costs. Autonomous driving systems actually pay more attention to the evaluation performance of key scenarios. How to construct key scenarios is a difficulty that restricts the development of autonomous driving evaluation technology.

[0003] At present, the construction of key scenarios is mainly based on data-driven and expert experience knowledge scenario enumeration methods. The scenario generation model based on generative adversarial networks can generate some scenarios that do not exist in real data sets, increasing the breadth of data sources. However, due to the lack of scene characteristics caused by the actual training process, the authenticity and rationality of the scenarios may be insufficient, and reasonable indicators need to be determined to further evaluate them, which limits the application of this model. The scenario enumeration method based on expert experience knowledge can model typical key scenarios and perform causal relationship analysis on key scenarios. However, this method cannot enumerate key scenarios one by one, and the cost of building such scenarios is high, which is not suitable for the requirements of autonomous driving evaluation technology. The basic scenario knowledge base constructed by data-driven can introduce expert experience knowledge into the generation model through knowledge transfer, realize the expansion of the basic scenario knowledge base, and build key scenarios for autonomous driving evaluation to accelerate autonomous driving evaluation.

[0004] There are relatively few patent forms for methods based on knowledge transfer in key scenarios of autonomous driving evaluation. Chinese patent CN202211165810.3 is a key scenario generation method and system for simulation testing of autonomous driving systems, which generates key scenarios through the minimum expected collision time of the main vehicle. Chinese patent CN202211470078.0 discloses a method for generating key safety scenarios for autonomous driving based on deep reinforcement learning, which has high detection efficiency, diversity, and is suitable for most roads. Chinese patent CN202010386871.7 discloses an automatic generation method for simulation test scenarios for autonomous driving decision-making and planning systems, which uses keywords to semantically describe and digitize the described scenarios, and obtains semantic descriptions after keyword digitization to generate scene files in batches. The above three patents can realize the construction of some key scenarios, but they do not yet have strong coverage and representativeness. Summary of the invention

[0005] The main purpose of the present invention is to analyze the key objectives, key phenomena, key assumptions and key evidence of the autonomous driving evaluation scenario, learn a small number of sample data through neural networks and form a basic scenario knowledge base through expert knowledge, and expand the basic scenario knowledge base based on knowledge transfer, and use Monte Carlo sampling to obtain the key scenarios of autonomous driving evaluation;

[0006] Another purpose of the present invention is to solve the problem of low test efficiency and high cost caused by too many safety redundant scenarios in autonomous driving evaluation, and to accelerate autonomous driving evaluation;

[0007] In order to achieve the above objectives and solve the above problems, the present invention provides a method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory.

[0008] The present invention provides a method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory, and the method includes the following steps:

[0009] The first step is to design the autonomous driving evaluation scenario specifications based on the ontology method;

[0010] Step 2: Build a knowledge base of basic autonomous driving evaluation scenarios;

[0011] Step 3: Expand the knowledge base of autonomous driving evaluation scenarios based on knowledge transfer;

[0012] Step 4: Integration and testing of key scenario knowledge base.

[0013] The process of designing the autonomous driving evaluation scenario specification based on the ontology method in the first step is as follows:

[0014] Step 1: Extract key indicators. The specific steps are as follows:

[0015] Step 1: Key indicator screening: By screening the key indicators of the scene, the data collected from the real traffic scene or the data generated by the simulation engine are evaluated to quantify the criticality of the traffic conditions or scenes. Different key indicators are screened through different scenes. The key indicators include safety and ride comfort.

[0016] Step 2: Definition of key indicators. This step defines the key indicators of the scenario, as follows:

[0017] The safety index function is:

[0018]

[0019]

[0020] Where D (t) is the relative distance between the two vehicles, V i(t) Indicates the speed of the following vehicle, V i-1(t) Indicates the speed of the vehicle ahead;

[0021]

[0022] The comfort index function is:

[0023] E com =λ jerk J com +λ acc A com

[0024] Among them, λ jerk and λ acceleration They represent the weight coefficients of jerk and RMS acceleration, J com and A com They represent jerk and acceleration RMS scores respectively. The jerk definition formula is as follows:

[0025]

[0026] The acceleration RMS score is defined as follows:

[0027]

[0028] The jerk score formula is as follows:

[0029]

[0030] Step 3: Application of key indicators: Apply safety and comfort key indicators to sampling to enhance the probabilistic distribution of key scenarios, further solve the problem of scarcity of key test scenarios, obtain key scenarios, and improve the efficiency of autonomous driving evaluation;

[0031] Step 2: Extract key phenomena. The specific steps are as follows:

[0032] Step 1: Screening of key phenomena. First, screen the key phenomena of the test scene. The key phenomena related to natural environmental factors include: the scattering, absorption and refraction of the laser radar by the concentration of particulate matter in the air will cause the laser radar intensity to weaken, affecting the propagation path and propagation speed of the laser beam, and affecting the measurement precision and accuracy of the laser radar; rain on rainy days will interfere with the camera and laser radar sensor, which will reduce the detection accuracy of the sensor. Heavy rain will cause the sensor to short-circuit, thus affecting driving safety; snowy days or different light intensities will cause over- or under-exposure of the camera, which will lead to the inability to identify the target. Strong winds will produce lateral or longitudinal forces that affect the vehicle body, or cause collisions with other vehicles due to speed fluctuations. The key phenomena of the screening test scene are expressed as follows:

[0033] K phenomenon = {P c ,R,Li,W,f(P c ,R,Li,W)}

[0034] In the formula, K phenomenon represents the set of key phenomena in the test scenario, P c represents the concentration of particulate matter in the air, R represents the intensity of rain, L i represents light intensity, W represents wind force, f(P c ,R,L i ,W) is the coupling relationship function between each element;

[0035] Step 2: Definition of key phenomena;

[0036] The particle concentration is inverted from the extinction coefficient data measured by the lidar. The aerosol extinction coefficient κ is expressed as:

[0037]

[0038] Where: r is the radius of the particle size in the atmosphere, C ext is the extinction efficiency factor, is the particle size distribution, the aerosol density is ρ, and the particle mass concentration is It is expressed as:

[0039]

[0040] The effective radius of the particle is expressed as:

[0041]

[0042] The relationship between the aerosol extinction coefficient and the mass concentration of particulate matter is defined as:

[0043]

[0044] The rainfall intensity is divided into five levels, namely trace, light rain, moderate rain, heavy rain and rainstorm. The rainfall intensity is inverted by the mean echo difference value of the radar image perspective. The calculation formula for the mean echo intensity is:

[0045]

[0046] Where a, b are the ranges of the selected radar images, f(u p ,v q ) is the position on the radar image (u p ,v q )The echo intensity value at the pixel point;

[0047] The mean value of the echo difference value of the radar image viewing angle within the selected time is expressed as:

[0048]

[0049] E′(n) is the mean value of the echo difference value that has not been processed in the nth group selection time, E(n) is the mean value of the echo difference value that has not been processed within 1 minute on the left side of the nth group, E(n) is the mean value of the echo difference value that has not been processed within 1 minute on the left side of the nth group, and E(n+1) is the mean value of the echo difference value that has not been processed within 1 minute on the right side of the nth group. The outliers are removed, and a set number of data are selected as fitting data. The relationship between the mean echo difference value and the rainfall intensity is linearly fitted using the least squares method to obtain the fitting relationship:

[0050] R=f(E′(n))

[0051] The light intensity distribution calculation formula is:

[0052]

[0053] Where A0 is the light intensity of the light source, α is the incident angle of the light, λ is the atmospheric attenuation coefficient, d is the distance from the light source to the detection target, s is the reflectivity of the object surface, F is the aperture number, D is the object distance, l is the focal length of the camera, β is the field of view, μ is the light transmittance of the optical system, ξ is the vignetting coefficient, and F is the stray light system generated by the combined effect of the lens and the camera body.

[0054] The safe driving condition of a passenger car at a speed of 100 km / h under the action of wind speeds of level 6 (12 m / s), level 7 (15 m / s), level 8 (19 m / s), level 9 (22 m / s) and level 10 (25 m / s). The mechanical expression of the aerodynamic force F of the car body is:

[0055] W=0.5ρV 2 SC F ;

[0056] W is the aerodynamic force, ρ is the air density, V is the relative speed between the car and the air, S is the contact area, C F is the aerodynamic coefficient;

[0057] Decompose the aerodynamic force along the X, Y, and Z directions of the vehicle body coordinate system, and the aerodynamic force components are:

[0058]

[0059] Step 3: Define key assumptions. The specific steps are as follows:

[0060] Step 1: Comparison of simulation environment information and consistency with the real world;

[0061] The data collected in the real environment is converted into a test simulation scene. The high consistency between the sensor values ​​collected by the simulation scene information and the original data is taken as one of the key assumptions of the scene. The information collected by several sensors is synchronized in time to ensure a good match between the data. The lane line characteristic parameters are fitted and reconstructed. The mathematical model of the lane line is:

[0062] y=S0+S1x+S2x 2 +S3x 3 ;

[0063] Among them, x and y are the position coordinates of the lane line relative to the vehicle coordinate system, S0 represents the lateral distance between the lane line and the main vehicle, S1 represents the direction of the lane line, S2 represents the curvature of one-half of the lane line, and S3 represents the curvature change rate of one-sixth of the lane line. Create lanes, add roadside features, use the OpenDRIVE map format, use XML language to semantically describe the static road network, and generate .xodr static road files;

[0064] Step 2: Virtual simulation model mapping;

[0065] The real driving vehicle state information is synchronized with the virtual vehicle state in the simulation world, the position transformation and rotation angle transformation of the vehicle in the simulation world coordinate system are mapped, the vertical and horizontal coordinate information of the vehicle is obtained, and the real scene coordinates and virtual world coordinates of the vehicle are used as the second key assumption of the scene;

[0066] Step 4: Infer causal relationships. The specific steps are as follows:

[0067] Assuming that the basic event variables are independent of each other, the probability of each state occurring is positively correlated with its occurrence value, and the parent variable in the causal diagram of several key factors only contributes to the probability value of the child variable, and the relationship between the contributions is a simple addition;

[0068] Step 1: Identify key factors;

[0069] The causal relationship of the autonomous driving scene. Applying the causal relationship to the scene generation model can improve the efficiency of scene generation. There are two ways to determine the key factors of the autonomous driving test scene. One is that experts define the key factors in the scene, including the ghost scene. The target vehicle O with or without obstruction has a direct effect on the behavior of the vehicle E and the pedestrian P. Specifically, O will cause E's behavior, and O will also cause P's behavior. Then the target vehicle O is regarded as the key factor of the scene, and the scene is described by a graphical model. The second is to use causal discovery tools to continuously explore key factors and then describe the scene through a graphical model.

[0070] Link 2: Causal Reasoning: Reasoning about causal relationships through causal diagrams. Suppose O mn (n=1,2,…,i),θ(O mn ) represents the key factor O m The value of the nth state, the key factor O m Each state has a numerical value of occurrence, and there is no connection between the states, but the probability value of each state occurring is positively correlated with the probability value of its occurrence;

[0071] if

[0072]

[0073] So,

[0074]

[0075] therefore

[0076]

[0077] Among them, P(O mn ), (n=1,2,…,j) becomes the state probability distribution factor, and the parent variable O m The occurrence of the nth state leads to the subvariable O n The probability of the pth state occurring is expressed as P(O np,mn );

[0078]

[0079] O m All states are considered as a key factor event E, E = E m1 +E m2 +…+E mj , O n All states of are considered as a single event F, F = F n1 +F n2 +…+F nq, the connection event that E causes F is P, and the probability of P occurring is recorded as Pr{P};

[0080]

[0081] The default state of the event is introduced. When the sum of the probabilities of each state of the event is 1, the probability matrix satisfies the normalization property, so that the connection event probability can be realized by causal graph reasoning when it is in matrix form. At the same time, the introduction of the default state makes the system knowledge expression more complete and clear, and more in line with the logic of life. The causal graph and the behavior graph adjacency matrix are used to transfer the prior knowledge from the causal graph to the neural network.

[0082] Step 5: Collect key evidence. The specific steps are as follows:

[0083] In the testing of autonomous driving scenarios, critical evidence refers to evidence that can effectively support critical assumptions. Critical assumptions refer to important assumptions speculated in the test, which have a key impact on the performance and safety of autonomous driving vehicles.

[0084] Step 1: Screening of key evidence;

[0085] Perception scene description is an important piece of evidence in scene testing. The difference between the sensor perception results and the real scene is used as key evidence. The perception results of cameras, lidars, and millimeter-wave radar sensors are selected as key evidence to describe the characteristics of different perception targets and their relationships with each other. The camera perception results are used to associate data with the millimeter-wave radar and lidar perception targets, and then the millimeter-wave radar and lidar perception targets are associated with data.

[0086] Step 2: Recording key evidence. By taking the positional correlation degree and the vertical correlation degree of the perceived objects as two mass functions on the hypothesis space, the correlation probability is calculated. Then, the two mass functions are synthesized using evidence theory to obtain the final inference result of whether the two perceived targets are correlated. For a given perceived target V p and V q , judging whether they are related as key evidence, defining whether they are the same perception target recognition framework is:

[0087] V pq ={0,1};

[0088] Construct a method to consider the similarity of position p and V q The mass function of whether it is the same perception target is:

[0089]

[0090] Among them, ξ pos is the uncertainty factor, 1-ξpos Expressed that considering the position characteristics, V cannot be determined p and V q The probability of whether it is the same perceived target, Indicates the distance between perceived targets The probability density function to determine whether the two are the same target is expressed as follows:

[0091]

[0092] Among them, a pos To control the threshold of whether two perceived objects are the same target based on the position characteristics;

[0093] Construct a method to judge V by considering the similarity of speed p and V q The mass function for the same target is:

[0094]

[0095] Among them, ξ vel is the uncertainty factor, 1-ξ vel Expressed that considering the position characteristics, V cannot be determined p and V q The probability of whether it is the same perceived target, Indicates the distance between perceived targets The probability density function to determine whether the two are the same object is expressed as follows:

[0096]

[0097] Among them, a vel A threshold to control whether two objects are the same target based on position features;

[0098] The Dempster-Shafer synthesis formula is used to calculate the probability of each subset;

[0099]

[0100]

[0101] Calculate separately The probability that

[0102]

[0103] Then V p and V q For the same goal, determine the association success;

[0104] like Determine association failure;

[0105] like Unable to determine V p and V q Whether they are the same target or not, the association is also judged as failed, and the association result is recorded as a critical scenario.

[0106] The second step is to build a basic autonomous driving evaluation scenario knowledge base. The specific process is as follows:

[0107] Step 1: A knowledge base of autonomous driving evaluation scenarios based on natural driving data. The specific steps are as follows: Step 1: Extraction of comprehensive information of evaluation scenarios;

[0108] The autonomous driving evaluation scenario includes static information, dynamic information and natural environment information. Static information includes road topology, specifically: road type, curvature, slope, width and adhesion coefficient. Dynamic information includes the relative position and equivalent speed of the vehicle and target vehicle, pedestrians and other dynamic traffic participants, and the coding and classification of traffic participants. Natural environment information includes the key phenomena defined in the first step, step 2, link 2, specifically: particle concentration, whether there is rain, light intensity and wind speed.

[0109] Step 2: Scene slicing processing;

[0110] The original scene data is extracted by the above algorithm and divided into two categories: "longitudinal scene" and "lateral scene". The "lateral scene" is further divided into "cut-in scene" and "cut-out scene". The natural driving scene is composed of several driving scenes. The original driving scene dynamic information data is divided into N (N = 1, 2, ..., n) segments according to the time series;

[0111] Step 3: Scenario cluster analysis;

[0112] The longitudinal scene feature parameters are the vehicle speed, the front and rear target vehicle speeds, the front and rear target vehicle accelerations, the collision time between the vehicle and the front vehicle, and the headway. The lateral scene feature parameters are the lateral speed, the lateral distance, lateral speed, and collision time of the left and right target vehicles. Delete empty data and abnormal data with one or more values ​​within the range of plus or minus 3 times the standard deviation of the variable. Specifically, divide the data set and the test set with a ratio of 3:2, train the model, normalize the modeling data set, divide the data into N groups through the Kmeans clustering algorithm, randomly select N as the initial cluster center, calculate the distance between each object and each cluster center, and assume that the input sample P = a1, a2, ..., a n , select the initial k cluster centers, q1,q2,…,q k , for each sample x i , marked as distance from the class center:

[0113]

[0114] Each category center is updated to the mean of all samples belonging to that category:

[0115]

[0116] Repeat the following two steps until the change of the category center is less than a certain threshold, and select the number of iterations, the cluster center change rate and the minimum square error MSE as the termination conditions; the output cluster partition C = {C1, C2, ..., C k}, according to different clusters representing different risk levels of longitudinal driving scenes, the clustering model is tested, and the C i Cluster the center points, classify the test data, and obtain the true type set. The prediction accuracy calculation formula is as follows:

[0117] P=M / N·100%

[0118] M is the number of elements in the intersection of the actual set and the predicted set; N is the number of data in the entire test data set;

[0119] If the prediction accuracy P is less than 85%, the clustering model obtained by the current training is judged to be unqualified and the scene segmentation is re-performed. Otherwise, the current clustering model is deemed to be qualified and the clustering model is output. Similarly, clustering analysis is performed on the cut-in and cut-out scenes.

[0120] Step 4: Integration of scenario knowledge base based on natural driving data;

[0121] Integrate the clustering sets of step 1, step 3 in the second step to build a scene knowledge base based on natural driving data KB = {Longitudinal, Cut in, Cut out}, which are different levels of longitudinal driving scene sets, cut-in scene sets and cut-out scene sets respectively;

[0122] Step 2: Develop an autonomous driving evaluation scenario knowledge base based on expert knowledge. The specific steps are as follows:

[0123] Part 1: Screening of autonomous driving evaluation scenarios based on expert experience;

[0124] Build an autonomous driving test scenario through expert experience, introduce scenario modeling into ontology according to traffic rules and road traffic signs, and describe all entities in the scenario by setting effective parameter range, scenario constraints and expert experience, including target vehicles, pedestrians, signboards, lane lines and road types, as well as key phenomena related to the natural environment defined in step 1, step 2, link 2, specifically: concentration of particulate matter in the air, light intensity, wind and rain. Make a structured description of the entities and key phenomena in the scenario based on expert experience, describe the traffic conditions from the perspective of the vehicle, and generate computer-readable XML language;

[0125] Part 2: Extraction of driving experience knowledge based on data-driven;

[0126] Through neural networks, patterns and knowledge are discovered from small sample data. Sensors collect features including vehicle speed, accelerator pedal position, brake pedal position, steering wheel angle, and lane line position. Matching machine learning models are selected for training to extract driving experience knowledge, specifically the collision time, headway, following distance, and expected speed of drivers with different driving styles in longitudinal driving scenarios, and the lane change starting point, lane change time, and lateral distance to the adjacent lane in lateral driving scenarios.

[0127] Step 3: Integration of scenario knowledge base based on expert knowledge;

[0128] Integrate the autonomous driving scenarios based on expert experience in step 1 and the scenarios based on data-driven driving experience knowledge in step 2 to build a basic scenario knowledge base for autonomous driving evaluation;

[0129] Step 3: Integration of basic scenario knowledge base for autonomous driving evaluation;

[0130] The autonomous driving evaluation scenario knowledge base based on natural driving data in step one and the autonomous driving evaluation scenario knowledge base based on expert knowledge in step two are integrated to construct a basic scenario knowledge base for autonomous driving evaluation.

[0131] The third step is to expand the knowledge base of autonomous driving evaluation scenarios based on knowledge transfer. The specific process is as follows:

[0132] Step 1: Determine the source domain and target domain of the autonomous driving evaluation scenario. The specific steps are as follows:

[0133] Step 1: Determine the source domain of the autonomous driving evaluation scenario;

[0134] The autonomous driving evaluation scenario knowledge base based on expert knowledge in step 2 is used as the source domain, and the existing scenario knowledge is used to guide the new scenario target domain. The source domain is defined as To ensure that the tasks in the source domain and the target domain are similar, the driving scenarios in the scenario knowledge base are divided into longitudinal and lateral driving scenarios;

[0135] Step 2: Determine the target domain of the autonomous driving evaluation scenario;

[0136] The target domain is the knowledge base of autonomous driving evaluation scenarios based on natural driving data in step 2, which is defined as

[0137] Step 2: Determine the target tasks of knowledge transfer;

[0138] Scenario expansion based on knowledge transfer is to use External information D s and Internal information of t Learn the optimal policy π in the target domain * ;

[0139]

[0140] Among the strategies Is the target domain The function mapping relationship between states and actions in the strategy is and Learn to get;

[0141] Step 3: Transfer of knowledge of longitudinal and lateral driving scenarios based on strategies. The specific steps are as follows:

[0142] Step 1: Source domain pre-training based on natural driving data;

[0143] Learn policy representations from source domain data of natural driving scenarios, and generalize these policy representations to the target task by fixing the weights or parameters of the source domain policy to improve the performance of the model. Source domain pre-training can reduce the training time required for the target task. When training the target task, use the trained model as the initial parameter to reduce the risk of overfitting.

[0144] Stage 2: Strategy-based scenario knowledge transfer;

[0145] Migrate the strategic knowledge of the basic driving scenario knowledge base based on natural driving data to the new scenario, establish a general strategy, and conduct further training based on the characteristics of the task. Trained in the corresponding source domains respectively, the student strategy π uses The acquired knowledge is learned on the target domain. By integrating the knowledge of the teacher model into the student model, an intermediate strategy π0 is established. The new strategy is learned under the guidance of π0. The ultimate goal to be maximized is:

[0146]

[0147] The goal constraint strategy π i does not deviate from π0, where

[0148]

[0149]

[0150] c KL ≥0 determines the KL divergence, c Ent ≥0 is entropy;

[0151] α=c KL / (c KL +c Ent )

[0152] β=1 / (c KL +c Ent )

[0153] Step 3: Evaluate the performance of the migration model;

[0154] The performance of the migration model is evaluated, and the evaluation indicators include accuracy and recall. At the same time, the performance of the model is verified by cross-validation method. If the accuracy is greater than 90%, it is considered that the migration model meets the performance requirements. Otherwise, the parameters of the migration model are retrained.

[0155] Step 4: Integration and testing of key scenario knowledge base. The specific process is as follows:

[0156] Step 1: Integration of key scenario knowledge base. The specific steps are as follows:

[0157] Step 1: Screening key scenarios for autonomous driving evaluation;

[0158] First, the distribution characteristics of the key evaluation factors of the longitudinal driving scenario are determined, and the distribution probability of THW is used to evaluate the safety and experience of autonomous driving in the longitudinal scenario. The univariate normal distribution model is selected to fit the data for analysis:

[0159]

[0160] Among them, for the multi-dimensional sample data obtained, the probability density function of the vehicle speed is fitted:

[0161]

[0162] Select the quadratic function to fit the data of the vehicle distance and the vehicle speed, and the quadratic curve equation obtained by fitting is:

[0163] d=a1v2+a2v+a3

[0164] Where a1, a2, and a3 are the coefficients of the quadratic term, linear term, and constant term of the quadratic curve respectively;

[0165] The distribution density function of the longitudinal and lateral scene parameters is estimated by the kernel density estimation method:

[0166]

[0167] The Markov Monte Carlo sampling algorithm is used for sampling. The key events are estimated according to the above formula. The relative error of the estimation result is:

[0168]

[0169] Step 2: Sampling the importance of key scenarios;

[0170] The Monte Carlo method is used to randomly sample the original distribution p(x);

[0171]

[0172] Sample a large amount of data from key test scenarios and use an unbiased estimation method based on importance sampling. In this method, an easy-to-sample distribution s(x) is introduced. At the same time, for the key event occurrence function f(x), the importance distribution s(x) is selected for weighting to ensure that the extracted data is representative. In this way, an unbiased estimate can be obtained to analyze the required key test scenarios;

[0173]

[0174] Calculate the relative error of importance sampling estimation results;

[0175]

[0176] Step 2: Key scenario library test, the specific steps are as follows:

[0177] Step 1: Build a high-precision virtual simulation test environment;

[0178] By building a high-precision simulation test environment, the test of autonomous driving evaluation scenarios can be realized, including high-precision virtual sensor models, specifically: virtual camera, virtual millimeter wave radar, virtual lidar, and inertial navigation measurement unit with superimposed noise;

[0179] Stage 2: Key scenario testing;

[0180] The key scenarios selected in step 1, step 1, were tested. The simulation results were evaluated from the two indicators of safety and comfort. The evaluation factors were determined. The perception performance, decision-making performance and control performance were evaluated. The longitudinal and lateral scenarios were tested. The scenario pass rate was:

[0181]

[0182] Where test pass To test the number of key scenarios that meet safety and comfort requirements, test total is the total number of key scene items.

[0183] Beneficial effects of the present invention:

[0184] The method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory provided by the present invention realizes the construction of key scenarios for autonomous driving evaluation through four steps: designing autonomous driving evaluation scenario specifications based on ontology methods, constructing a basic autonomous driving evaluation scenario knowledge base, expanding the autonomous driving evaluation scenario knowledge base based on knowledge transfer, and integrating and testing the key scenario knowledge base, testing the performance of the autonomous driving system in key scenarios, and accelerating the development and iteration of algorithms; the present invention performs standardized design of evaluation scenarios based on ontology methods, obtains dynamic and static comprehensive scene information semantic information representation through key indicators and key phenomenon extraction, defines key assumptions, infers causal relationships and key evidence definitions, and constructs a rich semantic information set; the present invention obtains a basic scenario knowledge base for autonomous driving evaluation through small sample data, and expands the autonomous driving evaluation scenario knowledge base through strategy-based knowledge transfer, thereby improving the coverage and generalization ability of the scenarios; the present invention obtains representative key scenarios by constructing a key scenario knowledge base and sampling the key scenario knowledge base to verify the performance of the autonomous driving system. BRIEF DESCRIPTION OF THE DRAWINGS

[0185] Figure 1 Schematic diagram of the overall steps of the method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory described in the present invention.

[0186] Figure 2 Schematic diagram of the overall architecture of the method for constructing key scenarios for autonomous driving evaluation according to the present invention.

[0187] Figure 3 Schematic diagram of the overall architecture of the first step of the present invention.

[0188] Figure 4 Schematic diagram of the overall architecture of the second step of the present invention.

[0189] Figure 5 Schematic diagram of the overall architecture of the third step of the present invention.

[0190] Figure 6 It is a schematic diagram of the overall architecture of the fourth step described in the present invention. DETAILED DESCRIPTION

[0191] See also Figures 1 to 6 As shown:

[0192] The present invention provides a method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory, and the method includes the following steps:

[0193] The first step is to design the autonomous driving evaluation scenario specifications based on the ontology method;

[0194] Step 2: Build a knowledge base of basic autonomous driving evaluation scenarios;

[0195] Step 3: Expand the knowledge base of autonomous driving evaluation scenarios based on knowledge transfer;

[0196] Step 4: Integration and testing of key scenario knowledge base.

[0197] The process of designing the autonomous driving evaluation scenario specification based on the ontology method in the first step is as follows:

[0198] Step 1: Extract key indicators. The specific steps are as follows:

[0199] Step 1: Key indicator screening: By screening the key indicators of the scene, the data collected from the real traffic scene or the data generated by the simulation engine are evaluated to quantify the criticality of the traffic conditions or scenes. Different key indicators are screened through different scenes. The key indicators include safety and ride comfort.

[0200] Step 2: Definition of key indicators. This step defines the key indicators of the scenario, as follows:

[0201] The safety index function is:

[0202]

[0203]

[0204] Where D (t) is the relative distance between the two vehicles, V i(t) Indicates the speed of the following vehicle, V i-1(t) Indicates the speed of the vehicle ahead;

[0205]

[0206] The comfort index function is:

[0207] E com =λ jerk J com +λ acc A com

[0208] Among them, λ jerk and λ accelerationThey represent the weight coefficients of jerk and RMS acceleration, J com and A com They represent jerk and acceleration RMS scores respectively. The jerk definition formula is as follows:

[0209]

[0210] The acceleration RMS score is defined as follows:

[0211]

[0212] The jerk score formula is as follows:

[0213]

[0214] Step 3: Application of key indicators: Apply safety and comfort key indicators to sampling to enhance the probabilistic distribution of key scenarios, further solve the problem of scarcity of key test scenarios, obtain key scenarios, and improve the efficiency of autonomous driving evaluation;

[0215] Step 2: Extract key phenomena. The specific steps are as follows:

[0216] Step 1: Screening of key phenomena. First, screen the key phenomena of the test scene. The key phenomena related to natural environmental factors include: the scattering, absorption and refraction of the laser radar by the concentration of particulate matter in the air will cause the laser radar intensity to weaken, affecting the propagation path and propagation speed of the laser beam, and affecting the measurement precision and accuracy of the laser radar; rain on rainy days will interfere with the camera and laser radar sensor, which will reduce the detection accuracy of the sensor. Heavy rain will cause the sensor to short-circuit, thus affecting driving safety; snowy days or different light intensities will cause over- or under-exposure of the camera, which will lead to the inability to identify the target. Strong winds will produce lateral or longitudinal forces that affect the vehicle body, or cause collisions with other vehicles due to speed fluctuations. The key phenomena of the screening test scene are expressed as follows:

[0217] K phenomenon = {P c ,R,Li,W,f(P c ,R,Li,W)}

[0218] In the formula, K phenomenon represents the set of key phenomena in the test scenario, P c represents the concentration of particulate matter in the air, R represents the intensity of rain, L i represents light intensity, W represents wind force, f(P c ,R,L i ,W) is the coupling relationship function between each element;

[0219] Step 2: Definition of key phenomena;

[0220] The particle concentration is inverted from the extinction coefficient data measured by the lidar. The aerosol extinction coefficient κ is expressed as:

[0221]

[0222] Where: r is the radius of the particle size in the atmosphere, C ext is the extinction efficiency factor, is the particle size distribution, the aerosol density is ρ, and the particle mass concentration is P c t It is expressed as:

[0223]

[0224] The effective radius of the particle is expressed as:

[0225]

[0226] The relationship between the aerosol extinction coefficient and the mass concentration of particulate matter is defined as:

[0227]

[0228] The rainfall intensity is divided into five levels, namely trace, light rain, moderate rain, heavy rain and rainstorm. The rainfall intensity is inverted by the mean echo difference value of the radar image perspective. The calculation formula for the mean echo intensity is:

[0229]

[0230] Where a, b are the ranges of the selected radar images, f(u p ,v q ) is the position on the radar image (u p ,v q )The echo intensity value at the pixel point;

[0231] The mean value of the echo difference value of the radar image viewing angle within the selected time is expressed as:

[0232]

[0233] E′(n) is the mean value of the echo difference value that has not been processed in the nth group selection time, E(n) is the mean value of the echo difference value that has not been processed within 1 minute on the left side of the nth group, E(n) is the mean value of the echo difference value that has not been processed within 1 minute on the left side of the nth group, and E(n+1) is the mean value of the echo difference value that has not been processed within 1 minute on the right side of the nth group. The outliers are removed, and a set number of data are selected as fitting data. The relationship between the mean echo difference value and the rainfall intensity is linearly fitted using the least squares method to obtain the fitting relationship:

[0234] R=f(E′(n))

[0235] The light intensity distribution calculation formula is:

[0236]

[0237] Where A0 is the light intensity of the light source, α is the incident angle of the light, λ is the atmospheric attenuation coefficient, d is the distance from the light source to the detection target, s is the reflectivity of the object surface, F is the aperture number, D is the object distance, l is the focal length of the camera, β is the field of view, μ is the light transmittance of the optical system, ξ is the vignetting coefficient, and F is the stray light system generated by the combined effect of the lens and the camera body.

[0238] The safe driving condition of a passenger car at a speed of 100 km / h under the action of wind speeds of level 6 (12 m / s), level 7 (15 m / s), level 8 (19 m / s), level 9 (22 m / s) and level 10 (25 m / s). The mechanical expression of the aerodynamic force F of the car body is:

[0239] W=0.5ρV 2 SC F ;

[0240] W is the aerodynamic force, ρ is the air density, V is the relative speed between the car and the air, S is the contact area, C F is the aerodynamic coefficient;

[0241] Decompose the aerodynamic force along the X, Y, and Z directions of the vehicle body coordinate system, and the aerodynamic force components are:

[0242]

[0243] Step 3: Define key assumptions. The specific steps are as follows:

[0244] Step 1: Comparison of simulation environment information and consistency with the real world;

[0245] The data collected in the real environment is converted into a test simulation scene. The high consistency between the sensor values ​​collected by the simulation scene information and the original data is taken as one of the key assumptions of the scene. The information collected by several sensors is synchronized in time to ensure a good match between the data. The lane line characteristic parameters are fitted and reconstructed. The mathematical model of the lane line is:

[0246] y=S0+S1x+S2x 2 +S3x 3 ;

[0247] Among them, x and y are the position coordinates of the lane line relative to the vehicle coordinate system, S0 represents the lateral distance between the lane line and the main vehicle, S1 represents the direction of the lane line, S2 represents the curvature of one-half of the lane line, and S3 represents the curvature change rate of one-sixth of the lane line. Create lanes, add roadside features, use the OpenDRIVE map format, use XML language to semantically describe the static road network, and generate .xodr static road files;

[0248] Step 2: Virtual simulation model mapping;

[0249] The real driving vehicle state information is synchronized with the virtual vehicle state in the simulation world, the position transformation and rotation angle transformation of the vehicle in the simulation world coordinate system are mapped, the vertical and horizontal coordinate information of the vehicle is obtained, and the real scene coordinates and virtual world coordinates of the vehicle are used as the second key assumption of the scene;

[0250] Step 4: Infer causal relationships. The specific steps are as follows:

[0251] Assuming that the basic event variables are independent of each other, the probability of each state occurring is positively correlated with its occurrence value, and the parent variable in the causal diagram of several key factors only contributes to the probability value of the child variable, and the relationship between the contributions is a simple addition;

[0252] Step 1: Identify key factors;

[0253] The causal relationship of the autonomous driving scene. Applying the causal relationship to the scene generation model can improve the efficiency of scene generation. There are two ways to determine the key factors of the autonomous driving test scene. One is that experts define the key factors in the scene, including the ghost scene. The target vehicle O with or without obstruction has a direct effect on the behavior of the vehicle E and the pedestrian P. Specifically, O will cause E's behavior, and O will also cause P's behavior. Then the target vehicle O is regarded as the key factor of the scene, and the scene is described by a graphical model. The second is to use causal discovery tools to continuously explore key factors and then describe the scene through a graphical model.

[0254] Link 2: Causal Reasoning: Reasoning about causal relationships through causal diagrams. Suppose O mn (n=1,2,…,i),θ(O mn ) represents the key factor O m The value of the nth state, the key factor O m Each state has a numerical value of occurrence, and there is no connection between the states, but the probability value of each state occurring is positively correlated with the probability value of its occurrence;

[0255] if

[0256]

[0257] So,

[0258]

[0259] therefore

[0260]

[0261] Among them, P(O mn ), (n=1,2,…,j) becomes the state probability distribution factor, and the parent variable O m The nth state of the subvariable O n The probability of the pth state occurring is expressed as P(O np,mn );

[0262]

[0263] O m All states are considered as a key factor event E, E = E m1 +E m2 +…+E mj , O n All states of are considered as a single event F, F = F n1 +F n2 +…+F nq , the connection event that E causes F is P, and the probability of P occurring is recorded as Pr{P};

[0264]

[0265] The default state of the event is introduced. When the sum of the probabilities of each state of the event is 1, the probability matrix satisfies the normalization property, so that the connection event probability can be realized by causal graph reasoning when it is in matrix form. At the same time, the introduction of the default state makes the system knowledge expression more complete and clear, and more in line with the logic of life. The causal graph and the behavior graph adjacency matrix are used to transfer the prior knowledge from the causal graph to the neural network.

[0266] Step 5: Collect key evidence. The specific steps are as follows:

[0267] In the testing of autonomous driving scenarios, critical evidence refers to evidence that can effectively support critical assumptions. Critical assumptions refer to important assumptions speculated in the test, which have a key impact on the performance and safety of autonomous driving vehicles.

[0268] Step 1: Screening of key evidence;

[0269] Perception scene description is an important piece of evidence in scene testing. The difference between the sensor perception results and the real scene is used as key evidence. The perception results of cameras, lidars, and millimeter-wave radar sensors are selected as key evidence to describe the characteristics of different perception targets and their relationships with each other. The camera perception results are used to associate data with the millimeter-wave radar and lidar perception targets, and then the millimeter-wave radar and lidar perception targets are associated with data.

[0270] Step 2: Recording key evidence. By taking the positional correlation degree and the vertical correlation degree of the perceived objects as two mass functions on the hypothesis space, the correlation probability is calculated. Then, the two mass functions are synthesized using evidence theory to obtain the final inference result of whether the two perceived targets are correlated. For a given perceived target V p and V q , judging whether they are related as key evidence, defining whether they are the same perception target recognition framework is:

[0271] V pq ={0,1};

[0272] Construct a method to consider the similarity of position p and V q The mass function of whether it is the same perception target is:

[0273]

[0274] Among them, ξ pos is the uncertainty factor, 1-ξ pos Expressed that considering the position characteristics, V cannot be determined p and V q The probability of whether it is the same perceived target, Indicates the distance between perceived targets The probability density function to determine whether the two are the same target is expressed as follows:

[0275]

[0276] Among them, a pos To control the threshold of whether two perceived objects are the same target based on the position characteristics;

[0277] Construct a method to judge V by considering the similarity of speed p and V q The mass function for the same target is:

[0278]

[0279] Among them, ξ vel is the uncertainty factor, 1-ξvel Expressed that considering the position characteristics, V cannot be determined p and V q The probability of whether it is the same perceived target, Indicates the distance between perceived targets The probability density function to determine whether the two are the same object is expressed as follows:

[0280]

[0281] Among them, a vel A threshold to control whether two objects are the same target based on position features;

[0282] The Dempster-Shafer synthesis formula is used to calculate the probability of each subset;

[0283]

[0284]

[0285] Calculate separately The probability that

[0286]

[0287] Then V p and V q For the same goal, determine the association success;

[0288] like Determine association failure;

[0289] like Unable to determine V p and V q Whether they are the same target or not, the association is also judged as failed, and the association result is recorded as a critical scenario.

[0290] The second step is to build a basic autonomous driving evaluation scenario knowledge base. The specific process is as follows:

[0291] Step 1: A knowledge base of autonomous driving evaluation scenarios based on natural driving data. The specific steps are as follows: Step 1: Extraction of comprehensive information of evaluation scenarios;

[0292] The autonomous driving evaluation scenario includes static information, dynamic information and natural environment information. Static information includes road topology, specifically: road type, curvature, slope, width and adhesion coefficient. Dynamic information includes the relative position and equivalent speed of the vehicle and target vehicle, pedestrians and other dynamic traffic participants, and the coding and classification of traffic participants. Natural environment information includes the key phenomena defined in the first step, step 2, link 2, specifically: particle concentration, whether there is rain, light intensity and wind speed.

[0293] Step 2: Scene slicing processing;

[0294] The original scene data is extracted by the above algorithm and divided into two categories: "longitudinal scene" and "lateral scene". The "lateral scene" is further divided into "cut-in scene" and "cut-out scene". The natural driving scene is composed of several driving scenes. The original driving scene dynamic information data is divided into N (N = 1, 2, ..., n) segments according to the time series;

[0295] Step 3: Scenario cluster analysis;

[0296] The longitudinal scene feature parameters are the vehicle speed, the front and rear target vehicle speeds, the front and rear target vehicle accelerations, the collision time between the vehicle and the front vehicle, and the headway. The lateral scene feature parameters are the lateral speed, the lateral distance, lateral speed, and collision time of the left and right target vehicles. Delete empty data and abnormal data with one or more values ​​within the range of plus or minus 3 times the standard deviation of the variable. Specifically, divide the data set and the test set with a ratio of 3:2, train the model, normalize the modeling data set, divide the data into N groups through the Kmeans clustering algorithm, randomly select N as the initial cluster center, calculate the distance between each object and each cluster center, and assume that the input sample P = a1, a2, ..., a n , select the initial k cluster centers, q1,q2,…,q k , for each sample x i , marked as distance from the class center:

[0297]

[0298] Each category center is updated to the mean of all samples belonging to that category:

[0299]

[0300] Repeat the following two steps until the change of the category center is less than a certain threshold, and select the number of iterations, the cluster center change rate and the minimum square error MSE as the termination conditions; the output cluster partition C = {C1, C2, ..., C k}, according to different clusters representing different risk levels of longitudinal driving scenes, the clustering model is tested, and the C i Cluster the center points, classify the test data, and obtain the true type set. The prediction accuracy calculation formula is as follows:

[0301] P=M / N·100%

[0302] M is the number of elements in the intersection of the actual set and the predicted set; N is the number of data in the entire test data set;

[0303] If the prediction accuracy P is less than 85%, the clustering model obtained by the current training is judged to be unqualified and the scene segmentation is re-performed. Otherwise, the current clustering model is deemed to be qualified and the clustering model is output. Similarly, clustering analysis is performed on the cut-in and cut-out scenes.

[0304] Step 4: Integration of scenario knowledge base based on natural driving data;

[0305] Integrate the clustering sets of step 1, step 3 in the second step to build a scene knowledge base based on natural driving data KB = {Longitudinal, Cut in, Cut out}, which are different levels of longitudinal driving scene sets, cut-in scene sets and cut-out scene sets respectively;

[0306] Step 2: Develop an autonomous driving evaluation scenario knowledge base based on expert knowledge. The specific steps are as follows:

[0307] Part 1: Screening of autonomous driving evaluation scenarios based on expert experience;

[0308] Build an autonomous driving test scenario through expert experience, introduce scenario modeling into ontology according to traffic rules and road traffic signs, and describe all entities in the scenario by setting effective parameter range, scenario constraints and expert experience, including target vehicles, pedestrians, signboards, lane lines and road types, as well as key phenomena related to the natural environment defined in step 1, step 2, link 2, specifically: concentration of particulate matter in the air, light intensity, wind and rain. Make a structured description of the entities and key phenomena in the scenario based on expert experience, describe the traffic conditions from the perspective of the vehicle, and generate computer-readable XML language;

[0309] Part 2: Extraction of driving experience knowledge based on data-driven;

[0310] Through neural networks, patterns and knowledge are discovered from small sample data. Sensors collect features including vehicle speed, accelerator pedal position, brake pedal position, steering wheel angle, and lane line position. Matching machine learning models are selected for training to extract driving experience knowledge, specifically the collision time, headway, following distance, and expected speed of drivers with different driving styles in longitudinal driving scenarios, and the lane change starting point, lane change time, and lateral distance to the adjacent lane in lateral driving scenarios.

[0311] Step 3: Integration of scenario knowledge base based on expert knowledge;

[0312] Integrate the autonomous driving scenarios based on expert experience in step 1 and the scenarios based on data-driven driving experience knowledge in step 2 to build a basic scenario knowledge base for autonomous driving evaluation;

[0313] Step 3: Integration of basic scenario knowledge base for autonomous driving evaluation;

[0314] The autonomous driving evaluation scenario knowledge base based on natural driving data in step one and the autonomous driving evaluation scenario knowledge base based on expert knowledge in step two are integrated to construct a basic scenario knowledge base for autonomous driving evaluation.

[0315] The third step is to expand the knowledge base of autonomous driving evaluation scenarios based on knowledge transfer. The specific process is as follows:

[0316] Step 1: Determine the source domain and target domain of the autonomous driving evaluation scenario. The specific steps are as follows:

[0317] Step 1: Determine the source domain of the autonomous driving evaluation scenario;

[0318] The autonomous driving evaluation scenario knowledge base based on expert knowledge in step 2 is used as the source domain, and the existing scenario knowledge is used to guide the new scenario target domain. The source domain is defined as To ensure that the tasks in the source domain and the target domain are similar, the driving scenarios in the scenario knowledge base are divided into longitudinal and lateral driving scenarios;

[0319] Step 2: Determine the target domain of the autonomous driving evaluation scenario;

[0320] The target domain is the knowledge base of autonomous driving evaluation scenarios based on natural driving data in step 2, which is defined as

[0321] Step 2: Determine the target tasks of knowledge transfer;

[0322] Scenario expansion based on knowledge transfer is to use External information D s and Internal information of t Learn the optimal policy π in the target domain * ;

[0323]

[0324] Among the strategies S t → t Is the target domain The function mapping relationship between states and actions in the strategy is and Learn to get;

[0325] Step 3: Transfer of knowledge of longitudinal and lateral driving scenarios based on strategies. The specific steps are as follows:

[0326] Step 1: Source domain pre-training based on natural driving data;

[0327] Learn policy representations from source domain data of natural driving scenarios, and generalize these policy representations to the target task by fixing the weights or parameters of the source domain policy to improve the performance of the model. Source domain pre-training can reduce the training time required for the target task. When training the target task, use the trained model as the initial parameter to reduce the risk of overfitting.

[0328] Stage 2: Strategy-based scenario knowledge transfer;

[0329] Migrate the strategic knowledge of the basic driving scenario knowledge base based on natural driving data to the new scenario, establish a general strategy, and conduct further training based on the characteristics of the task. Trained in the corresponding source domains respectively, the student strategy π uses The acquired knowledge is learned on the target domain. By integrating the knowledge of the teacher model into the student model, an intermediate strategy π0 is established. The new strategy is learned under the guidance of π0. The ultimate goal to be maximized is:

[0330]

[0331] The goal constraint strategy π i does not deviate from π0, where

[0332]

[0333]

[0334] c KL ≥0 determines the KL divergence, c Ent ≥0 is entropy;

[0335] α=c KL / (c KL +c Ent )

[0336] β=1 / (c KL +c Ent )

[0337] Step 3: Evaluate the performance of the migration model;

[0338] The performance of the migration model is evaluated, and the evaluation indicators include accuracy and recall. At the same time, the performance of the model is verified by cross-validation method. If the accuracy is greater than 90%, it is considered that the migration model meets the performance requirements. Otherwise, the parameters of the migration model are retrained.

[0339] Step 4: Integration and testing of key scenario knowledge base. The specific process is as follows:

[0340] Step 1: Integration of key scenario knowledge base. The specific steps are as follows:

[0341] Step 1: Screening key scenarios for autonomous driving evaluation;

[0342] First, the distribution characteristics of the key evaluation factors of the longitudinal driving scenario are determined, and the distribution probability of THW is used to evaluate the safety and experience of autonomous driving in the longitudinal scenario. The univariate normal distribution model is selected to fit the data for analysis:

[0343]

[0344] Among them, for the multi-dimensional sample data obtained, the probability density function of the vehicle speed is fitted:

[0345]

[0346] Select the quadratic function to fit the data of the vehicle distance and the vehicle speed, and the quadratic curve equation obtained by fitting is:

[0347] d=a1v2+a2v+a3

[0348] Where a1, a2, and a3 are the coefficients of the quadratic term, linear term, and constant term of the quadratic curve respectively;

[0349] The distribution density function of the longitudinal and lateral scene parameters is estimated by the kernel density estimation method:

[0350]

[0351] The Markov Monte Carlo sampling algorithm is used for sampling. The key events are estimated according to the above formula. The relative error of the estimation result is:

[0352]

[0353] Step 2: Sampling the importance of key scenarios;

[0354] The Monte Carlo method is used to randomly sample the original distribution p(x);

[0355]

[0356] Sample a large amount of data from key test scenarios and use an unbiased estimation method based on importance sampling. In this method, an easy-to-sample distribution s(x) is introduced. At the same time, for the key event occurrence function f(x), the importance distribution s(x) is selected for weighting to ensure that the extracted data is representative. In this way, an unbiased estimate can be obtained to analyze the required key test scenarios;

[0357]

[0358] Calculate the relative error of importance sampling estimation results;

[0359]

[0360] Step 2: Key scenario library test, the specific steps are as follows:

[0361] Step 1: Build a high-precision virtual simulation test environment;

[0362] By building a high-precision simulation test environment, the test of autonomous driving evaluation scenarios can be realized, including high-precision virtual sensor models, specifically: virtual camera, virtual millimeter wave radar, virtual lidar, and inertial navigation measurement unit with superimposed noise;

[0363] Stage 2: Key scenario testing;

[0364] The key scenarios selected in step 1, step 1, were tested. The simulation results were evaluated from the two indicators of safety and comfort. The evaluation factors were determined. The perception performance, decision-making performance and control performance were evaluated. The longitudinal and lateral scenarios were tested. The scenario pass rate was:

[0365]

[0366] Where test pass To test the number of key scenarios that meet safety and comfort requirements, test total is the total number of key scene items.

Claims

1. A method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory, characterized by: The method comprises the following steps: The first step is to design the autonomous driving evaluation scenario specifications based on the ontology method; Step 2: Build a knowledge base of basic autonomous driving evaluation scenarios; Step 3: Expand the knowledge base of autonomous driving evaluation scenarios based on knowledge transfer; Step 4: Integration and testing of key scenario knowledge base.

2. The method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory according to claim 1, characterized in that: The process of designing the autonomous driving evaluation scenario specification based on the ontology method in the first step is as follows: Step 1: Extract key indicators. The specific steps are as follows: Step 1: Key indicator screening: By screening the key indicators of the scene, the data collected from the real traffic scene or the data generated by the simulation engine are evaluated to quantify the criticality of the traffic conditions or scenes. Different key indicators are screened through different scenes. The key indicators include safety and ride comfort. Step 2: Definition of key indicators. This step defines the key indicators of the scenario, as follows: The safety index function is: Where D (t) is the relative distance between the two vehicles, V i(t) Indicates the speed of the following vehicle, V i-1(t) Indicates the speed of the vehicle ahead; The comfort index function is: E com =λ jerk J com +λ acc A com Among them, λ jerk and λ acceleration They represent the weight coefficients of jerk and RMS acceleration, J com and A com They represent jerk and acceleration RMS scores respectively. The jerk definition formula is as follows: The acceleration RMS score is defined as follows: The jerk score formula is as follows: Step 3: Application of key indicators: Apply safety and comfort key indicators to sampling to enhance the probabilistic distribution of key scenarios, further solve the problem of scarcity of key test scenarios, obtain key scenarios, and improve the efficiency of autonomous driving evaluation; Step 2: Extract key phenomena. The specific steps are as follows: Step 1: Screening of key phenomena. First, screen the key phenomena of the test scene. The key phenomena related to natural environmental factors include: the scattering, absorption and refraction of the laser radar by the concentration of particulate matter in the air will cause the laser radar intensity to weaken, affecting the propagation path and propagation speed of the laser beam, and affecting the measurement precision and accuracy of the laser radar; rain on rainy days will interfere with the camera and laser radar sensor, which will reduce the detection accuracy of the sensor. Heavy rain will cause the sensor to short-circuit, thus affecting driving safety; snowy days or different light intensities will cause over- or under-exposure of the camera, which will lead to the inability to identify the target. Strong winds will produce lateral or longitudinal forces that affect the vehicle body, or cause collisions with other vehicles due to speed fluctuations. The key phenomena of the screening test scene are expressed as follows: <h2 style=";text-align:left;direction:ltr">K<h2 style=";text-align:left;direction:ltr"> phenomenon <h2 style=";text-align:left;direction:ltr"> ={P<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> ,R,Li,W,f(P<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> ,R,Li,W)} In the formula, K phenomenon represents the set of key phenomena in the test scenario, P c represents the concentration of particulate matter in the air, R represents the intensity of rain, L i represents light intensity, W represents wind force, f(P c ,R,L i ,W) is the coupling relationship function between each element; Step 2: Definition of key phenomena; The particle concentration is inverted from the extinction coefficient data measured by the lidar. The aerosol extinction coefficient κ is expressed as: Where: r is the radius of the particle size in the atmosphere, C ext is the extinction efficiency factor, is the particle size distribution, the aerosol density is ρ, and the particle mass concentration is P c t It is expressed as: The effective radius of the particle is expressed as: The relationship between the aerosol extinction coefficient and the mass concentration of particulate matter is defined as: The rainfall intensity is divided into five levels, namely trace, light rain, moderate rain, heavy rain and rainstorm. The rainfall intensity is inverted by the mean echo difference value of the radar image perspective. The calculation formula for the mean echo intensity is: Where a, b are the ranges of the selected radar images, f(u p ,v q ) is the position on the radar image (u p ,v q )The echo intensity value at the pixel point; The mean value of the echo difference value of the radar image viewing angle within the selected time is expressed as: E′(n) is the mean value of the echo difference value that has not been processed in the nth group selection time, E(n) is the mean value of the echo difference value that has not been processed within 1 minute on the left side of the nth group, E(n) is the mean value of the echo difference value that has not been processed within 1 minute on the left side of the nth group, and E(n+1) is the mean value of the echo difference value that has not been processed within 1 minute on the right side of the nth group. The outliers are removed, and a set number of data are selected as fitting data. The relationship between the mean echo difference value and the rainfall intensity is linearly fitted using the least squares method to obtain the fitting relationship: R=f(E′(n)) The light intensity distribution calculation formula is: Where A0 is the light intensity of the light source, α is the incident angle of the light, λ is the atmospheric attenuation coefficient, d is the distance from the light source to the detection target, s is the reflectivity of the object surface, F is the aperture number, D is the object distance, l is the focal length of the camera, β is the field of view, μ is the light transmittance of the optical system, ξ is the vignetting coefficient, and F is the stray light system generated by the combined effect of the lens and the camera body. The safe driving condition of a passenger car at a speed of 100 km / h under the action of wind speeds of level 6 (12 m / s), level 7 (15 m / s), level 8 (19 m / s), level 9 (22 m / s) and level 10 (25 m / s). The mechanical expression of the aerodynamic force F of the car body is: W=0.5ρV 2 SC F ; W is the aerodynamic force, ρ is the air density, V is the relative speed between the car and the air, S is the contact area, C F is the aerodynamic coefficient; Decompose the aerodynamic force along the X, Y, and Z directions of the vehicle body coordinate system, and the aerodynamic force components are: Step 3: Define key assumptions. The specific steps are as follows: Step 1: Comparison of simulation environment information and consistency with the real world; The data collected in the real environment is converted into a test simulation scene. The high consistency between the sensor values ​​collected by the simulation scene information and the original data is taken as one of the key assumptions of the scene. The information collected by several sensors is synchronized in time to ensure a good match between the data. The lane line characteristic parameters are fitted and reconstructed. The mathematical model of the lane line is: y=S0+S1x+S2x 2 +S3x 3 ; Among them, x and y are the position coordinates of the lane line relative to the vehicle coordinate system, S0 represents the lateral distance between the lane line and the main vehicle, S1 represents the direction of the lane line, S2 represents the curvature of one-half of the lane line, and S3 represents the curvature change rate of one-sixth of the lane line. Create lanes, add roadside features, use the OpenDRIVE map format, use XML language to semantically describe the static road network, and generate .xodr static road files; Step 2: Virtual simulation model mapping; The real driving vehicle state information is synchronized with the virtual vehicle state in the simulation world, the position transformation and rotation angle transformation of the vehicle in the simulation world coordinate system are mapped, the vertical and horizontal coordinate information of the vehicle is obtained, and the real scene coordinates and virtual world coordinates of the vehicle are used as the second key assumption of the scene; Step 4: Infer causal relationships. The specific steps are as follows: Assuming that the basic event variables are independent of each other, the probability of each state occurring is positively correlated with its occurrence value, and the parent variable in the causal diagram of several key factors only contributes to the probability value of the child variable, and the relationship between the contributions is a simple addition; Step 1: Identify key factors; The causal relationship of the autonomous driving scene. Applying the causal relationship to the scene generation model can improve the efficiency of scene generation. There are two ways to determine the key factors of the autonomous driving test scene. One is that experts define the key factors in the scene, including the ghost scene. The target vehicle O with or without obstruction has a direct effect on the behavior of the vehicle E and the pedestrian P. Specifically, O will cause E's behavior, and O will also cause P's behavior. Then the target vehicle O is regarded as the key factor of the scene, and the scene is described by a graphical model. The second is to use causal discovery tools to continuously explore key factors and then describe the scene through a graphical model. Link 2: Causal Reasoning: Reasoning about causal relationships through causal diagrams. Suppose O mn (n=1,2,…,i),θ(O mn ) represents the key factor O m The value of the nth state, the key factor O m Each state has a numerical value of occurrence, and there is no connection between the states, but the probability value of each state occurring is positively correlated with the probability value of its occurrence; if So, therefore Among them, P(O mn ), (n=1,2,…,j) becomes the state probability distribution factor, and the parent variable O m The nth state of the subvariable O n The probability of the pth state occurring is expressed as P(O np,mn ); O m All states are considered as a key factor event E, E = E m1 +E m2 +…+E mj , O n All states of are considered as a single event F, F = F n1 +F n2 +…+F nq , the connection event that E causes F is P, and the probability of P occurring is recorded as Pr{P}; The default state of the event is introduced. When the sum of the probabilities of each state of the event is 1, the probability matrix satisfies the normalization property, so that the connection event probability can be realized by causal graph reasoning when it is in matrix form. At the same time, the introduction of the default state makes the system knowledge expression more complete and clear, and more in line with the logic of life. The causal graph and the behavior graph adjacency matrix are used to transfer the prior knowledge from the causal graph to the neural network. Step 5: Collect key evidence. The specific steps are as follows: In the testing of autonomous driving scenarios, critical evidence refers to evidence that can effectively support critical assumptions. Critical assumptions refer to important assumptions inferred during the test, which have a key impact on the performance and safety of the autonomous driving vehicle. Step 1: Screening of key evidence; Perception scene description is an important piece of evidence in scene testing. The difference between the sensor perception results and the real scene is used as key evidence. The perception results of cameras, lidars, and millimeter-wave radar sensors are selected as key evidence to describe the characteristics of different perception targets and their relationships with each other. The camera perception results are used to associate data with the millimeter-wave radar and lidar perception targets, and then the millimeter-wave radar and lidar perception targets are associated with data. Step 2: Recording key evidence. By taking the positional correlation degree and the vertical correlation degree of the perceived objects as two mass functions on the hypothesis space, the correlation probability is calculated. Then, the two mass functions are synthesized using evidence theory to obtain the final inference result of whether the two perceived targets are correlated. For a given perceived target V p and V q , judging whether they are related as key evidence, defining whether they are the same perception target recognition framework is: V pq ＝{0,1}； Construct a method to consider the similarity of position p and V q The mass function of whether it is the same perception target is: Among them, ξ pos is the uncertainty factor, 1-ξ pos Expressed that considering the position characteristics, V cannot be determined p and V q The probability of whether it is the same perceived target, Indicates the distance between perceived targets The probability density function to determine whether the two are the same target is expressed as follows: Among them, a pos To control the threshold of whether two perceived objects are the same target based on the position characteristics; Construct a method to judge V by considering the similarity of speed p and V q The mass function for the same target is: Among them, ξ vel is the uncertainty factor, 1-ξ vel Expressed that considering the position characteristics, V cannot be determined p and V q The probability of whether it is the same perceived target, Indicates the distance between perceived targets The probability density function to determine whether the two are the same object is expressed as follows: Among them, a vel A threshold to control whether two objects are the same target based on position features; The Dempster-Shafer synthesis formula is used to calculate the probability of each subset; Calculate separately The probability that Then V p and V q For the same goal, determine the association success; like Determine association failure; like Unable to determine V p and V q Whether they are the same target or not, the association is also judged as failed, and the association result is recorded as a critical scenario.

3. The method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory according to claim 1, characterized in that: The second step is to build a basic autonomous driving evaluation scenario knowledge base. The specific process is as follows: Step 1: Develop a knowledge base of autonomous driving evaluation scenarios based on natural driving data. The specific steps are as follows: Step 1: Extraction of comprehensive information of evaluation scenarios; The autonomous driving evaluation scenario includes static information, dynamic information and natural environment information. Static information includes road topology, specifically: road type, curvature, slope, width and adhesion coefficient. Dynamic information includes the relative position and equivalent speed of the vehicle and target vehicle, pedestrians and other dynamic traffic participants, and the coding and classification of traffic participants. Natural environment information includes the key phenomena defined in the first step, step 2, link 2, specifically: particle concentration, whether there is rain, light intensity and wind speed. Step 2: Scene slicing processing; The original scene data is extracted by the above algorithm and divided into two categories: "longitudinal scene" and "lateral scene". The "lateral scene" is further divided into "cut-in scene" and "cut-out scene". The natural driving scene is composed of several driving scenes. The original driving scene dynamic information data is divided into N (N = 1, 2, ..., n) segments according to the time series; Step 3: Scenario cluster analysis; The longitudinal scene feature parameters are the vehicle speed, the front and rear target vehicle speeds, the front and rear target vehicle accelerations, the collision time between the vehicle and the front vehicle, and the headway. The lateral scene feature parameters are the lateral speed, the lateral distance, lateral speed, and collision time of the left and right target vehicles. Delete empty data and abnormal data with one or more values ​​within the range of plus or minus 3 times the standard deviation of the variable. Specifically, divide the data set and the test set with a ratio of 3:2, train the model, normalize the modeling data set, divide the data into N groups through the Kmeans clustering algorithm, randomly select N as the initial cluster center, calculate the distance between each object and each cluster center, and assume that the input sample P = a1, a2, ..., a n , select the initial k cluster centers, q1,q2,…,q k , for each sample x i , marked as distance from the class center: Each category center is updated to the mean of all samples belonging to that category: Repeat the next two steps until the category center change is less than a certain threshold, and select the number of iterations, cluster center change rate and minimum square error MSE as the termination conditions; Output cluster partition C = {C1, C2, ..., C k }, according to different clusters representing different risk levels of longitudinal driving scenes, the clustering model is tested, and the C i Cluster the center points, classify the test data, and obtain the true type set. The prediction accuracy calculation formula is as follows: P=M / N·100% M is the number of elements in the intersection of the actual set and the predicted set; N is the number of data in the entire test data set; If the prediction accuracy P is less than 85%, the clustering model obtained by the current training is judged to be unqualified and the scene segmentation is re-performed. Otherwise, the current clustering model is deemed to be qualified and the clustering model is output. Similarly, clustering analysis is performed on the cut-in and cut-out scenes. Step 4: Integration of scenario knowledge base based on natural driving data; Integrate the clustering sets of step 1, step 3 in the second step to build a scene knowledge base based on natural driving data KB = {Longitudinal, Cut in, Cut out}, which are different levels of longitudinal driving scene sets, cut-in scene sets and cut-out scene sets respectively; Step 2: Develop an autonomous driving evaluation scenario knowledge base based on expert knowledge. The specific steps are as follows: Part 1: Screening of autonomous driving evaluation scenarios based on expert experience; Build an autonomous driving test scenario through expert experience, introduce scenario modeling into ontology according to traffic rules and road traffic signs, and describe all entities in the scenario by setting effective parameter range, scenario constraints and expert experience, including target vehicles, pedestrians, signboards, lane lines and road types, as well as key phenomena related to the natural environment defined in step 1, step 2, link 2, specifically: concentration of particulate matter in the air, light intensity, wind and rain. Make a structured description of the entities and key phenomena in the scenario based on expert experience, describe the traffic conditions from the perspective of the vehicle, and generate computer-readable XML language; Part 2: Extraction of driving experience knowledge based on data-driven; Through neural networks, patterns and knowledge are discovered from small sample data. Sensors collect features including vehicle speed, accelerator pedal position, brake pedal position, steering wheel angle, and lane line position. Matching machine learning models are selected for training to extract driving experience knowledge, specifically the collision time, headway, following distance, and expected speed of drivers with different driving styles in longitudinal driving scenarios, and the lane change starting point, lane change time, and lateral distance to the adjacent lane in lateral driving scenarios. Step 3: Integration of scenario knowledge base based on expert knowledge; Integrate the autonomous driving scenarios based on expert experience in step 1 and the scenarios based on data-driven driving experience knowledge in step 2 to build a basic scenario knowledge base for autonomous driving evaluation; Step 3: Integration of basic scenario knowledge base for autonomous driving evaluation; The autonomous driving evaluation scenario knowledge base based on natural driving data in step one and the autonomous driving evaluation scenario knowledge base based on expert knowledge in step two are integrated to construct a basic scenario knowledge base for autonomous driving evaluation.

4. The method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory according to claim 1, characterized in that: The third step is to expand the knowledge base of autonomous driving evaluation scenarios based on knowledge transfer. The specific process is as follows: Step 1: Determine the source domain and target domain of the autonomous driving evaluation scenario. The specific steps are as follows: Step 1: Determine the source domain of the autonomous driving evaluation scenario; The autonomous driving evaluation scenario knowledge base based on expert knowledge in step 2 is used as the source domain, and the existing scenario knowledge is used to guide the new scenario target domain. The source domain is defined as To ensure that the tasks in the source domain and the target domain are similar, the driving scenarios in the scenario knowledge base are divided into longitudinal and lateral driving scenarios; Step 2: Determine the target domain of the autonomous driving evaluation scenario; The target domain is the knowledge base of autonomous driving evaluation scenarios based on natural driving data in step 2, which is defined as Step 2: Determine the target tasks of knowledge transfer; Scenario expansion based on knowledge transfer is to use External information D s and Internal information of t Learn the optimal policy π in the target domain * ; Among the strategies Is the target domain The function mapping relationship between states and actions in the strategy is and Learn to get; Step 3: Transfer of knowledge of longitudinal and lateral driving scenarios based on strategies. The specific steps are as follows: Step 1: Source domain pre-training based on natural driving data; Learn policy representations from source domain data of natural driving scenarios, and generalize these policy representations to the target task by fixing the weights or parameters of the source domain policy to improve the performance of the model. Source domain pre-training can reduce the training time required for the target task. When training the target task, use the trained model as the initial parameter to reduce the risk of overfitting. Stage 2: Strategy-based scenario knowledge transfer; Migrate the strategic knowledge of the basic driving scenario knowledge base based on natural driving data to the new scenario, establish a general strategy, and conduct further training based on the characteristics of the task. Trained in the corresponding source domains respectively, the student strategy π uses The acquired knowledge is learned on the target domain. By integrating the knowledge of the teacher model into the student model, an intermediate strategy π0 is established. The new strategy is learned under the guidance of π0. The ultimate goal to be maximized is: The goal constraint strategy π i does not deviate from π0, where c KL ≥0 determines the KL divergence, c Ent ≥0 is entropy; α=c KL / (c KL +c Ent ) β=1 / (c KL +c Ent ) Step 3: Evaluate the performance of the migration model; The performance of the migration model is evaluated, and the evaluation indicators include accuracy and recall. At the same time, the performance of the model is verified by cross-validation method. If the accuracy is greater than 90%, it is considered that the migration model meets the performance requirements. Otherwise, the parameters of the migration model are retrained.

5. The method for constructing key scenarios for autonomous driving evaluation based on knowledge transfer theory according to claim 1, characterized in that: The fourth step is the integration and testing of the key scenario knowledge base. The specific process is as follows: Step 1: Integration of key scenario knowledge base. The specific steps are as follows: Step 1: Screening key scenarios for autonomous driving evaluation; First, the distribution characteristics of the key evaluation factors of the longitudinal driving scenario are determined, and the distribution probability of THW is used to evaluate the safety and experience of autonomous driving in the longitudinal scenario. The univariate normal distribution model is selected to fit the data for analysis: Among them, for the multi-dimensional sample data obtained, the probability density function of the vehicle speed is fitted: Select the quadratic function to fit the data of the vehicle distance and the vehicle speed, and the quadratic curve equation obtained by fitting is: d=a1v2+a2v+a3 Where a1, a2, and a3 are the coefficients of the quadratic term, linear term, and constant term of the quadratic curve respectively; The distribution density function of the longitudinal and lateral scene parameters is estimated by the kernel density estimation method: The Markov Monte Carlo sampling algorithm is used for sampling. The key events are estimated according to the above formula. The relative error of the estimation result is: Step 2: Sampling the importance of key scenarios; The Monte Carlo method is used to randomly sample the original distribution p(x); Sample a large amount of data from key test scenarios and use an unbiased estimation method based on importance sampling. In this method, an easy-to-sample distribution s(x) is introduced. At the same time, for the key event occurrence function f(x), the importance distribution s(x) is selected for weighting to ensure that the extracted data is representative. In this way, an unbiased estimate can be obtained to analyze the required key test scenarios; Calculate the relative error of importance sampling estimation results; Step 2: Key scenario library test, the specific steps are as follows: Step 1: Build a high-precision virtual simulation test environment; By building a high-precision simulation test environment, the test of autonomous driving evaluation scenarios can be realized, including high-precision virtual sensor models, specifically: virtual camera, virtual millimeter wave radar, virtual lidar, and inertial navigation measurement unit with superimposed noise; Stage 2: Key scenario testing; The key scenarios selected in step 1, step 1, were tested. The simulation results were evaluated from the two indicators of safety and comfort. The evaluation factors were determined. The perception performance, decision-making performance and control performance were evaluated. The longitudinal and lateral scenarios were tested. The scenario pass rate was: Where test pass To test the number of key scenarios that meet safety and comfort requirements, test total is the total number of key scene items.

Images