CN114970318A - Automatic driving system safety evaluation method based on behavior result double-agent model - Google Patents
Automatic driving system safety evaluation method based on behavior result double-agent model Download PDFInfo
- Publication number
- CN114970318A CN114970318A CN202210431404.0A CN202210431404A CN114970318A CN 114970318 A CN114970318 A CN 114970318A CN 202210431404 A CN202210431404 A CN 202210431404A CN 114970318 A CN114970318 A CN 114970318A
- Authority
- CN
- China
- Prior art keywords
- scene
- agent model
- result
- model
- test
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000011156 evaluation Methods 0.000 title claims abstract description 48
- 238000012360 testing method Methods 0.000 claims abstract description 142
- 238000005070 sampling Methods 0.000 claims abstract description 84
- 238000000034 method Methods 0.000 claims abstract description 33
- 230000006399 behavior Effects 0.000 claims description 60
- 238000011161 development Methods 0.000 claims description 13
- 238000003066 decision tree Methods 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000003542 behavioural effect Effects 0.000 claims description 7
- 238000009826 distribution Methods 0.000 claims description 7
- 230000001133 acceleration Effects 0.000 claims description 4
- 238000012163 sequencing technique Methods 0.000 claims description 4
- 238000012854 evaluation process Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000003068 static effect Effects 0.000 claims description 3
- 238000013136 deep learning model Methods 0.000 claims description 2
- 238000010276 construction Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 238000005192 partition Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000005728 strengthening Methods 0.000 description 2
- 238000012614 Monte-Carlo sampling Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000011076 safety test Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M17/00—Testing of vehicles
- G01M17/007—Wheeled or endless-tracked vehicles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/40—Business processes related to the transportation industry
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Human Resources & Organizations (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Medical Informatics (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Geometry (AREA)
- Game Theory and Decision Science (AREA)
- Educational Administration (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Development Economics (AREA)
- Computer Hardware Design (AREA)
- Primary Health Care (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention relates to a safety evaluation method of an automatic driving system based on a behavior result double-agent model, which comprises the following steps: test sampling, namely performing accelerated test based on a behavior agent model and a result agent model, and acquiring a historical test sample set of the safety performance of the tested automatic driving vehicle in a logic scene through the test sampling; and safety evaluation, namely evaluating the overall safety of the tested automatic driving vehicle in the logic scene based on the historical test sample set to obtain a comprehensive safety evaluation result. The test sampling stage comprises four steps of initialization sampling and cycle connection: constructing/updating an agent model, sampling an agent, calculating value and actually testing; the safety evaluation stage comprises four steps of tree model construction, danger subspace hyper-volume estimation, danger mode identification and evaluation result reliability estimation. Compared with the prior art, the method can efficiently and reliably carry out comprehensive test and evaluation on the safety of the automatic driving system in the logic scene space.
Description
Technical Field
The invention relates to the technical field of automatic driving, in particular to a safety evaluation method of an automatic driving system based on a behavior result double-proxy model.
Background
Safety is a difficult problem of falling on the ground of the automatic driving technology, and safety evaluation and demonstration are necessary ways for providing guarantee and confidence for the safety of an automatic driving system. The standard draft ISO 21448 proposes the concept of expected functional safety, which refers to the safety of an autonomous vehicle in the case of external environmental disturbances, insufficient expected functions or human mishandling. The "opinions on strengthening intelligent internet automobile production enterprises and product admission management" issued by Ministry of industry and belief in 2021 at 8 months emphasizes the importance of safety management in aspects of strengthening safety of expected functions of the automatic driving automobiles and the like so as to ensure the product quality and the production consistency. Therefore, the test and evaluation based on the scene are important links of the expected functional safety demonstration.
At present, for a high-grade automatic driving system, two main problems are mainly faced in scene-based safety assessment: firstly, neglecting the high complexity and uncertainty of the tested object and the operating environment thereof, continuing to use the evaluation thought of a high-level Driving Assistance System (ADAS), and replacing the evaluation of the function and logic scenes with typical specific scenes, wherein the 'point-by-point' mode brings misleading evaluation results; secondly, the complexity of the tested object and the operating environment thereof greatly increases the parameter dimension for describing the test scene, the traversal of the logic scene parameter space does not have practical feasibility, and the traditional test method (such as grid test, random test, orthogonal experiment and the like) has low efficiency and cannot process the high-dimensional test scene.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide the automatic driving system safety evaluation method based on the behavior result double-proxy model, which can perform comprehensive safety evaluation under the condition of determining a logic scene.
The purpose of the invention can be realized by the following technical scheme: a safety evaluation method of an automatic driving system based on a behavior result double-agent model comprises the following steps:
s1, test sampling: performing an acceleration test based on the behavior agent model and the result agent model, and acquiring a historical test sample set of the safety performance of the tested automatic driving vehicle in a logic scene through test sampling;
s2, safety assessment: and based on the historical test sample set, evaluating the overall safety of the tested automatic driving vehicle in the logic scene to obtain a comprehensive safety evaluation result.
Further, the behavior agent model is specifically an agent model for performing predictive agent on a running track of the measured autonomous vehicle in a test scene, the running track comprises a track geometry and the time of the measured autonomous vehicle reaching each position on the track, and the input of the behavior agent model is an initial scene frame or an initial scene frame sequence; the output is a travel trajectory of the measured autonomous vehicle over a period of time in the future.
Further, the result agent model is specifically an agent model for performing mapping agent on the scene risk of the tested autonomous vehicle in the test scene, and the input of the result agent model is an initial scene frame; the output is the scenario risk for the test scenario.
Further, the scene frame refers to state information of the test scene at a certain instant, where the state information includes state information of static scene elements and dynamic scene elements.
Further, the step S1 specifically includes the following steps:
s11, initializing sampling, collecting a set number of sample points from a logic scene space for testing, and adding a test result into a historical test sample set to obtain initial information for starting an accelerated test cycle;
s12, constructing and updating an agent model, constructing or updating a behavior agent model and a result agent model according to a historical test sample set, and dividing a logic scene space into a plurality of subspaces by using the result agent model;
s13, proxy sampling, namely performing rapid proxy sampling on the behavior proxy model and the result proxy model to obtain a batch of temporary proxy sampling data;
s14, calculating value, namely calculating the observation value of each subspace according to the proxy sampling data, and sequencing the subspaces according to the observation value from high to low;
s15, actual testing, namely selecting a plurality of subspaces sequenced at the top N as target testing spaces, sampling specific scenes from the target testing spaces to perform actual testing, adding testing results into a historical testing sample set, and if the testing cost is exhausted, exiting the cycle and outputting the current historical testing sample set as results; otherwise, it returns to step S12.
Further, the step S12 specifically includes the following steps:
s121, constructing a behavior agent model, and learning a general driving behavior agent model as an initial model of the behavior agent model by using the field data and the deep learning model before the evaluation is started;
s122, updating the behavior agent model, calculating an agent error of the behavior agent model to the behavior of the tested automatic driving vehicle in the actual evaluation process, and finely adjusting model parameters by using the agent error to ensure that an agent result of the behavior agent model is gradually consistent with the specific tested automatic driving vehicle;
s123, constructing a result agent model, calling a unit space dividing step by using a logic scene space as a root node, recursively calling the unit space dividing step for divided subspaces in the same way, setting a stopping condition of recursive division, and creating a hierarchical division tree of the logic scene space as the result agent model;
and S124, updating the result proxy model, and reconstructing the result proxy model according to a set period.
Further, the step of dividing the cell space in step S123 specifically includes: the parent space is used as a root node, the curved surface is divided according to the historical test samples falling in the parent space, the parent space is divided into a dangerous subspace and a non-dangerous subspace, and two corresponding subspace nodes are created on the parent space root node.
Further, the observation value in the step S14 is composed of two parts, i.e., "development" value characterized by the scene risk degree and "exploration" value characterized by the observation information uncertainty.
Further, the calculation process of the "development" value is as follows: the development value of a specific scene is measured by the scene risk in the primary test process, and the development value of a scene subspace is represented and calculated by the expectation of all the specific scene risks in the subspace;
the calculation process of the exploration value is as follows: the exploration value of a scene subspace is measured by the uncertainty of the information of the agent model to the subspace and consists of uncertainty of two parts of sources, namely 'unobserved' and 'unobserved quasi'; calculating and representing uncertainty caused by 'unobserved' according to the distribution calculation sampling density of the historical test samples in the logic scene space; the "quasi-unobserved" uncertainty is computationally characterized by a behavioral agent model for the agent error of the actual autonomous vehicle under test.
Further, the step S2 specifically includes the following steps:
s21, constructing a tree model, binarizing labels of sample points in the historical test sample set obtained in the step S1 test sampling stage according to a given scene risk threshold value, and dividing the labels into dangerous sample points and non-dangerous sample points; then, a division tree of a logic scene space is constructed by using a decision tree algorithm, and the logic scene space is divided into a plurality of dangerous and non-dangerous subspaces;
s22, calculating the excessive volume of the dangerous subspace, extracting leaf nodes representing the dangerous subspace according to the decision tree model constructed in the step S21, calculating the excessive volume of the leaf nodes of each dangerous subspace and summing the excessive volumes to obtain the excessive volumes of all the dangerous subspaces, and calculating the proportion of the excessive volume of the dangerous subspace to the total volume of the space of the whole logic scene to obtain the total risk of the tested autonomous vehicle in the logic scene;
s23, dangerous mode identification, constructing an undirected graph according to the spatial connectivity of dangerous leaf nodes on the basis of the decision tree model constructed in the step S21, combining the connected dangerous subspace leaf nodes on the undirected graph to obtain one or more connected components, and using each connected component to represent a dangerous mode in the logic scene;
s24, evaluating the reliability of the result, and calculating the proxy error of each subspace by using the behavior proxy model updated in the last iteration turn of the test sampling stage of the step S1;
estimating the sampling density of each subspace by using the sample distribution of the historical test sample set updated in the last iteration round of the test sampling stage in the step S1;
and combining the proxy error and the sampling density to obtain the observation value of each subspace, and performing weighted summation on the observation values by using the subspace scene risk expectation as a weight to obtain the overall residual exploration value of the logic scene space, and representing the reliability of the evaluation result.
Compared with the prior art, the invention has the following advantages:
(1) the invention provides a behavior agent model and a result agent model, and combines the behavior agent model and the result agent model to realize an automatic driving system safety evaluation method based on scene test, which can effectively evaluate the safety of the expected function of a high-grade automatic driving system, comprehensively evaluate the safety of the automatic driving system in a logic scene, and improve the efficiency and reliability of safety evaluation by using the behavior agent model and the result agent model.
(2) The method is oriented to the goal of safety comprehensive evaluation of the automatic driving system in the logic scene space, effectively identifies the dangerous subspace in the logic scene space through the self-adaptive acceleration test, further analyzes and calculates the volume ratio and the dangerous mode of the dangerous subspace, estimates the reliability of the evaluation result through calculating the weighted residual exploration value, forms a comprehensive safety evaluation result, and can be well suitable for safety test, evaluation and demonstration of the automatic driving system.
(3) According to the invention, by introducing a behavior agent model and a 'pre-training-fine-tuning' technology, on the basis that the traditional accelerated test based on optimization only uses test result information, test process information and external field information are additionally introduced, so that a higher accelerated test effect is realized.
(4) The invention provides a comprehensive value model containing three values of scene risk, unobserved and unobserved, which can effectively balance exploration and development, accurately guide test sampling, avoid falling into local optimum and ensure the accuracy of acquiring a historical test sample set in a test sampling stage.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention;
fig. 2 is a schematic diagram of an application process of the embodiment.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments.
Examples
As shown in fig. 1, a safety evaluation method for an automatic driving system based on a behavior result dual-agent model includes the following steps:
s1, test sampling: performing an acceleration test based on the behavior agent model and the result agent model, and acquiring a historical test sample set of the safety performance of the tested automatic driving vehicle in a logic scene through test sampling;
s2, safety assessment: and based on the historical test sample set, evaluating the overall safety of the tested automatic driving vehicle in the logic scene to obtain a comprehensive safety evaluation result.
Wherein, the definition of the behavior agent model is as follows: carrying out predictive agent on a running track (comprising a track geometric shape and time for reaching each position on the track) of a tested automatic driving vehicle (hereinafter referred to as a tested object) in a test scene, wherein the input of a behavior agent model is an initial scene frame or an initial scene frame sequence; the output is the driving track of the measured object in a future period.
The definition of the resulting agent model is: carrying out mapping type proxy on the scene risk of the tested object under the test scene, wherein the input of a result proxy model is an initial scene frame; the output is a scenario risk indicator for the test scenario.
The scene frame refers to the state information of a scene at a certain instant, and the state information includes, but is not limited to, static roads, traffic sign lines, buildings, and the like, and dynamic position, posture, speed, and the like of traffic participants.
In the present embodiment, as shown in fig. 2, the above technical solution mainly includes two stages: a test sampling phase 1 and a security assessment phase 2.
1. And (5) testing a sampling stage. Performing accelerated test based on the behavior agent model and the result agent model, and acquiring related information about the safety performance of the tested object in a logic scene through test sampling, namely acquiring a historical test sample set finally obtained by accelerated test;
2. and (5) a safety evaluation stage. And based on the result (historical test sample set) of the test sampling stage, evaluating the overall safety of the tested object in the logic scene.
In the sampling test stage, the method mainly comprises four steps of initializing sampling, constructing/updating a circularly connected proxy model, proxy sampling, calculating value and actually testing. The details are as follows:
11. the sampling is initialized. Sampling methods including but not limited to random sampling, low-difference sequence sampling and the like are adopted, a small number of sample points are collected from a logic scene space for testing, and a test result is added into a historical test sample set to obtain initial information for starting an accelerated test cycle.
12. The proxy model is built/updated. And constructing or updating a behavior agent model and a result agent model according to the historical test sample set. The logical scene space is divided into a number of subspaces using a result agent model.
13. And (4) proxy sampling. And performing rapid proxy sampling on the behavior proxy model and the result proxy model by adopting a sampling method including but not limited to random sampling, low-difference sequence sampling or optimization-based sampling to obtain a batch of temporary proxy sampling data.
14. And (4) calculating the value. And calculating the observation value of each subspace according to the proxy sampling data, and sequencing the subspaces according to the observation value from high to low.
15. And (5) actual testing. And selecting a subspace with the front sequence as a target test space according to actual requirements, sampling a specific scene from the target test space, carrying out actual test, and adding a test result into a historical test sample set. If the testing cost is exhausted, the circulation is exited, and a final historical testing sample set is output as a result; if not, jump to step 12 proxy model build/update.
In the agent model building/updating step 12, the details of the method for building and updating the behavior agent model and the result agent model are as follows:
121. and constructing a behavior agent model. Before the evaluation is started, a general driving behavior agent model is learned by using field data including but not limited to natural driving data and the like in combination with a deep learning neural network model including but not limited to CNN, RNN, Transformer and the like as an initial model of the behavior agent model.
122. And updating the behavior agent model. In the actual evaluation process, the agent error of the behavior agent model on the running track of the object to be measured is calculated, and the model parameters are finely adjusted by using the agent error, so that the agent result of the behavior agent model is gradually consistent with the specific object to be measured.
123. And constructing a result agent model. Calling a unit space dividing step by using a logic scene space as a root node, recursively calling the unit space dividing step in the same way for divided subspaces, setting stop conditions of recursive division (including but not limited to the minimum number of historical test samples in the subspaces, the maximum branch depth of the division tree and the like), creating a hierarchical division tree of the logic scene space as a result proxy model
Wherein, a unit space dividing step is: the parent space is used as a root node, the curved surface is divided according to the historical test samples falling in the parent space, the parent space is divided into a dangerous subspace and a non-dangerous subspace, and two corresponding subspace nodes are created on the parent space root node.
124. The result is an update of the proxy model. The result agent model is reconstructed at a certain period (i.e., at certain test sampling times) and the reconstruction method is the same as the method for constructing the result agent model in step 123.
In the value calculation step 14, the composition and calculation method of the observed value are as follows:
141. the composition of the value was observed. The observation value is composed of two parts, namely a development value characterized by scene risk degree and an exploration value characterized by observation information uncertainty.
142. A method of calculating the "development" value. The development value of a specific scene is measured by the scene risk in the primary test process, and can be calculated by methods including but not limited to TTC, THW, feasible domain and the like according to actual conditions; the "development" value of a scene subspace is characterized and calculated by the expectation of all the specific scene risks within that subspace.
143. A method of calculating the "exploration" value. The 'exploration' value of a scene subspace is measured by the uncertainty of information of the agent model to the subspace and divided into two sources of 'unobserved' and 'unobserved quasi'. Wherein, the uncertainty caused by 'unobserved' is characterized by calculating sampling density by using a technology including but not limited to a Kernel Density Estimator (KDE) and the like according to the distribution of the historical test samples in the logic scene space; the uncertainty of 'unobserved quasi' is characterized by the behavior agent model to the agent error of the actual measured object.
In the safety evaluation stage, the method mainly comprises the steps of constructing a tree model at the front, calculating the super volume of a danger subspace at the rear, identifying the danger mode and estimating the reliability of an evaluation result:
21. and (5) building a tree model. Carrying out binarization on labels of sample points in a historical test sample set obtained in a test sampling stage according to a given scene risk threshold value, and dividing the labels into dangerous sample points and non-dangerous sample points; and then, a division tree of the logical scene space is constructed by using decision tree algorithms including but not limited to C4.5, CART and the like, the division tree is divided into dangerous and non-dangerous subspaces, and each dimension of the divided subspaces is parallel to a certain dimension of the logical scene space.
22. And (4) calculating the over-volume of the danger subspace. And (3) according to the decision tree model obtained in the tree model construction step (21), extracting leaf nodes representing the dangerous subspace, calculating the over-volume of the leaf nodes of each dangerous subspace and summing the over-volumes to obtain the over-volumes of all dangerous subspaces. And calculating the proportion of the dangerous subspace hyper-volume to the total volume of the whole logic scene space as the total risk of the measured object in the logic scene.
23. And (4) dangerous mode identification. Constructing an undirected graph according to the spatial connectivity of the dangerous leaf nodes on the basis of the decision tree model obtained in the tree model construction step 21; and merging connected risk subspace leaf nodes by using algorithms including, but not limited to, Kosaraju, union and search set and the like on the undirected graph to obtain one or more connected components, wherein each connected component represents a risk mode in the logic scene.
24. And evaluating the reliability of the result. Calculating the proxy error of each subspace by using a behavior proxy model updated in the last iteration turn in the test sampling stage, estimating the sampling density of each subspace by using the sample distribution of a historical test sample set updated in the last iteration turn in the test sampling stage, and fusing the proxy error and the sampling density by using methods including but not limited to a weighted average method to obtain the observation value of each subspace. And weighting and summing the observation value of the subspace scene risk expectation serving as a weight to obtain the overall residual exploration value of the logic scene space, and representing the reliability of the evaluation result. The larger the residual exploration value is, the lower the reliability of the evaluation result is; the smaller the residual exploration value, the higher the reliability of the evaluation result.
In this embodiment, the logical scene is a three-lane straight road, and the vehicle is controlled by the tested automatic driving systemA vehicle (hereinafter, referred to as a "self vehicle") travels following another vehicle (hereinafter, referred to as a "preceding vehicle") ahead at a constant speed, and the preceding vehicle is braked at a constant deceleration at a certain time. The logical scene space is spanned by the following parameters: initial speed V0 of the host vehicle and the preceding vehicle, braking deceleration a of the preceding vehicle, and braking duration T of the preceding vehicle. The value ranges of the parameters are as follows: 70km/h<V0<100km/h，1m/s 2 <A<6m/s 2 ，0.5s<T<3s in the sequence. The safety evaluation of the self vehicle under the logic scene comprises the following two stages:
stage 1, testing and sampling stage. And guiding a test sampling position by using a pre-trained behavior agent model and a real-time constructed result agent model, and acquiring a historical test sample set of the self-vehicle in the logic scene. Specifically, the method comprises the following steps:
and step 11, initializing sampling. And generating 64 initial sampling point positions by using a Sobolev low difference sequence, and testing by using a specific scene corresponding to the sampling points. Obtaining a historical test sample set D;
and step 12, constructing/updating the agent model. Calculating a proxy error of the behavior proxy model by using the historical test sample set D, and finely adjusting and updating the behavior proxy model by using the error through a gradient back propagation algorithm; carrying out recursive division on the logic scene space according to the historical test sample set D, and dividing the logic scene space into 8 subspaces according to the stopping condition;
and step 13, proxy sampling. Generating 1000 sampling point positions in each subspace by using uniform distribution, performing numerical calculation on a behavior agent model and a result agent model by using the sampling points, and acquiring 1000 temporary agent sampling results, wherein each result comprises the scene risk of the point, the sampling density of the point and the agent error of the point;
and step 14, calculating the value. And respectively calculating the scene risk, the sampling density and the proxy error of all temporary proxy sampling points in each subspace to obtain the expected average value, so as to obtain the total scene risk, the sampling density and the proxy error of each subspace. And integrating the overall scene risk, the sampling density and the proxy error of each subspace by using a weighted average mode to obtain the observation value of each subspace. Sequencing the subspaces according to the observation values of the subspaces, and selecting 2 subspaces with the highest observation values as target test subspaces;
and step 15, actual testing. And (3) generating 10 sampling point positions in 2 subspaces by using a Sobolev sequence in combination with acceptance-rejection sampling, and performing actual test by using the sampling points to obtain 20 pieces of new test sample data, and adding the 20 pieces of new test sample data into a historical test sample set. If the number of the samples in the historical test sample set exceeds 1000, entering a stage 2; otherwise, jumping to step 12;
stage 2, safety evaluation stage. And integrally evaluating the safety of the self vehicle under the following logic scene according to more than 1000 historical test samples obtained in the stage 1. Specifically, the method comprises the following steps:
and step 21, constructing a tree model. Setting a threshold value of a scene risk index as 1, screening historical test samples, marking scene risks lower than 1 as non-dangerous, marking scene risks higher than 1 as dangerous, and obtaining a marked historical test sample set D label . Using C4.5 decision Tree as model Algorithm, D label As training data, learning to obtain a partition tree of a logic scene space;
and step 22, calculating the over-volume of the dangerous subspace. And traversing all leaf nodes of the partition tree, and selecting the nodes with the danger marking sample proportion of more than 50% as danger child nodes. And calculating and summing the super volumes of the hypercubes represented by each dangerous sub-node to obtain the total volume of the dangerous sub-space. Calculating the proportion of the total volume of the dangerous subspaces to the total volume of the logical scene space, wherein the proportion is 42%, and the overall risk index of the self-vehicle in the logical scene is 0.42;
and step 23, dangerous mode identification. And (4) inspecting the spatial connectivity of all the dangerous sub-nodes in the partition tree, building edges between the directly connected dangerous sub-nodes, and constructing an undirected graph. Applying a Kosaraju algorithm to the undirected graph, and finally obtaining 2 disjoint connected components. The dangerous modes of the self vehicle in the logic scene are 2 and are characterized by the identified connected components.
And 24, evaluating the reliability of the result. Sampling is carried out on the behavior agent model by using a Monte Carlo sampling method to estimate the agent error of each subspace, the sample density distribution of the historical test sample set is calculated by using a kernel density estimator, the sampling density of each subspace is estimated, and the agent error and the sampling density are fused by using a weighted average method to obtain the observation value of each subspace. And weighting and summing the observation values of the subspace scene risk expectations by using the subspace scene risk expectations as weights, so that the obtained logic scene space overall residual exploration value is low, and the reliability of the safety evaluation result is proved to be acceptable.
In summary, the technical scheme provides the automatic driving system safety evaluation method based on the behavior result double-agent model, which is used for comprehensively evaluating the safety of the automatic driving system in a logic scene.
Claims (10)
1. A safety evaluation method of an automatic driving system based on a behavior result double-agent model is characterized by comprising the following steps:
s1, test sampling: performing an acceleration test based on the behavior agent model and the result agent model, and acquiring a historical test sample set of the safety performance of the tested automatic driving vehicle in a logic scene through test sampling;
s2, safety assessment: and based on the historical test sample set, evaluating the overall safety of the tested automatic driving vehicle in the logic scene to obtain a comprehensive safety evaluation result.
2. The automated driving system safety evaluation method based on the behavior result double-agent model is characterized in that the behavior agent model is specifically an agent model for predictively acting a running track of a tested automated driving vehicle under a test scene, the running track comprises a track geometric shape and the time of the tested automated driving vehicle reaching each position on the track, and the input of the behavior agent model is an initial scene frame or an initial scene frame sequence; the output is a travel trajectory of the measured autonomous vehicle over a period of time in the future.
3. The automated driving system safety evaluation method based on the behavior result double-agent model is characterized in that the result agent model is specifically an agent model for carrying out mapping type agent on the scene risks of the tested automated driving vehicle in a test scene, and the input of the result agent model is an initial scene frame; the output is the scenario risk for the test scenario.
4. The automated driving system safety assessment method based on the behavioral result dual-agent model according to claim 2 or 3, wherein the scene frame refers to the state information of the test scene at a certain instant, and the state information includes the state information of static scene elements and dynamic scene elements.
5. The automated driving system safety evaluation method based on the behavioral result dual-agent model according to claim 1, wherein the step S1 specifically comprises the following steps:
s11, initializing sampling, collecting a set number of sample points from a logic scene space for testing, and adding a test result into a historical test sample set to obtain initial information for starting an accelerated test cycle;
s12, constructing and updating an agent model, constructing or updating a behavior agent model and a result agent model according to a historical test sample set, and dividing a logic scene space into a plurality of subspaces by using the result agent model;
s13, proxy sampling, namely performing rapid proxy sampling on the behavior proxy model and the result proxy model to obtain a batch of temporary proxy sampling data;
s14, calculating value, namely calculating the observation value of each subspace according to the proxy sampling data, and sequencing the subspaces according to the observation value from high to low;
s15, actual testing, namely selecting a plurality of subspaces sequenced at the top N as target testing spaces, sampling specific scenes from the target testing spaces to perform actual testing, adding testing results into a historical testing sample set, and if the testing cost is exhausted, exiting the cycle and outputting the current historical testing sample set as results; otherwise, it returns to step S12.
6. The automated driving system safety evaluation method based on the behavioral result dual-agent model according to claim 5, wherein the step S12 specifically comprises the following steps:
s121, constructing a behavior agent model, and learning a general driving behavior agent model as an initial model of the behavior agent model by using the field data and the deep learning model before the evaluation is started;
s122, updating the behavior agent model, calculating an agent error of the behavior agent model to the behavior of the tested automatic driving vehicle in the actual evaluation process, and finely adjusting model parameters by using the agent error to ensure that an agent result of the behavior agent model is gradually consistent with the specific tested automatic driving vehicle;
s123, constructing a result agent model, calling a unit space dividing step by using a logic scene space as a root node, recursively calling the unit space dividing step for divided subspaces in the same way, setting a stopping condition of recursive division, and creating a hierarchical division tree of the logic scene space as the result agent model;
and S124, updating the result proxy model, and reconstructing the result proxy model according to the set period.
7. The automated driving system safety assessment method based on the behavioral result dual-agent model according to claim 6, wherein the step of dividing the cell space in the step S123 specifically comprises: the parent space is used as a root node, the curved surface is divided according to the historical test samples falling in the parent space, the parent space is divided into a dangerous subspace and a non-dangerous subspace, and two corresponding subspace nodes are created on the parent space root node.
8. The method as claimed in claim 6, wherein the observation value in step S14 is composed of a "development" value characterized by scene risk degree and an "exploration" value characterized by observation information uncertainty.
9. The automated driving system safety assessment method based on behavior result double-agent model according to claim 8, characterized in that the calculation process of the "development" value is as follows: the development value of a specific scene is measured by the scene risk in the primary test process, and the development value of a scene subspace is represented and calculated by the expectation of all the specific scene risks in the subspace;
the calculation process of the exploration value is as follows: the exploration value of a scene subspace is measured by the uncertainty of the information of the agent model to the subspace and consists of uncertainty of two parts of sources, namely 'unobserved' and 'unobserved quasi'; calculating and representing uncertainty caused by 'unobserved' according to the distribution calculation sampling density of the historical test samples in the logic scene space; the "quasi-unobserved" uncertainty is computationally characterized by a behavioral proxy model to the proxy error of the actual measured autonomous vehicle.
10. The automated driving system safety evaluation method based on the behavioral result dual-agent model according to claim 1, wherein the step S2 specifically comprises the following steps:
s21, constructing a tree model, binarizing the labels of the sample points in the historical test sample set obtained in the step S1 test sampling stage according to a given scene risk threshold value, and dividing the labels into dangerous sample points and non-dangerous sample points; then, a division tree of a logic scene space is constructed by using a decision tree algorithm, and the logic scene space is divided into a plurality of dangerous and non-dangerous subspaces;
s22, calculating the excessive volume of the dangerous subspace, extracting leaf nodes representing the dangerous subspace according to the decision tree model constructed in the step S21, calculating the excessive volume of the leaf nodes of each dangerous subspace and summing the excessive volumes to obtain the excessive volumes of all the dangerous subspaces, and calculating the proportion of the excessive volume of the dangerous subspace to the total volume of the space of the whole logic scene to obtain the total risk of the tested autonomous vehicle in the logic scene;
s23, dangerous mode identification, constructing an undirected graph according to the spatial connectivity of dangerous leaf nodes on the basis of the decision tree model constructed in the step S21, combining the connected dangerous subspace leaf nodes on the undirected graph to obtain one or more connected components, and using each connected component to represent a dangerous mode in the logic scene;
s24, evaluating the reliability of the result, and calculating the proxy error of each subspace by using the behavior proxy model updated in the last iteration turn of the test sampling stage of the step S1;
estimating the sampling density of each subspace by using the sample distribution of the historical test sample set updated in the last iteration round of the test sampling stage in the step S1;
and integrating proxy errors and sampling density to obtain the observation value of each subspace, weighting and summing the observation values by using subspace scene risk expectation as weight to obtain the overall residual exploration value of the logic scene space, and representing the reliability of the evaluation result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210431404.0A CN114970318A (en) | 2022-04-22 | 2022-04-22 | Automatic driving system safety evaluation method based on behavior result double-agent model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210431404.0A CN114970318A (en) | 2022-04-22 | 2022-04-22 | Automatic driving system safety evaluation method based on behavior result double-agent model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114970318A true CN114970318A (en) | 2022-08-30 |
Family
ID=82980237
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210431404.0A Pending CN114970318A (en) | 2022-04-22 | 2022-04-22 | Automatic driving system safety evaluation method based on behavior result double-agent model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114970318A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117435558A (en) * | 2023-12-20 | 2024-01-23 | 杭州硕磐智能科技有限公司 | Metadata management method, computing device and storage medium |
-
2022
- 2022-04-22 CN CN202210431404.0A patent/CN114970318A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117435558A (en) * | 2023-12-20 | 2024-01-23 | 杭州硕磐智能科技有限公司 | Metadata management method, computing device and storage medium |
| CN117435558B (en) * | 2023-12-20 | 2024-03-29 | 杭州硕磐智能科技有限公司 | Metadata management method, computing device and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yu et al. | Real-time traffic speed estimation with graph convolutional generative autoencoder | |
| Kasiviswanathan et al. | Comparison of methods used for quantifying prediction interval in artificial neural network hydrologic models | |
| KR102492205B1 (en) | Apparatus and method for detecting delivery vehicle based on Inverse Reinforcement Learning | |
| Torrens | Calibrating and validating cellular automata models of urbanization | |
| Kang et al. | Urban traffic travel time short‐term prediction model based on spatio‐temporal feature extraction | |
| CN114036135A (en) | Method and system for estimating urban mobile source pollution emission by using incomplete information | |
| CN112150304A (en) | Power grid running state track stability prejudging method and system and storage medium | |
| Platho et al. | Predicting velocity profiles of road users at intersections using configurations | |
| CN114970318A (en) | Automatic driving system safety evaluation method based on behavior result double-agent model | |
| Bagheri et al. | Simulation of vehicles’ gap acceptance decision at unsignalized intersections using SUMO | |
| Wen et al. | Modeling the effects of autonomous vehicles on human driver car-following behaviors using inverse reinforcement learning | |
| Jiang et al. | Efficient and unbiased safety test for autonomous driving systems | |
| Tibaldi et al. | Bayesian optimization for qaoa | |
| Howard et al. | Evaluating temporal observation-based causal discovery techniques applied to road driver behaviour | |
| CN113379156A (en) | Speed prediction method, device, equipment and storage medium | |
| Yuan et al. | Efficient traffic state estimation and prediction based on the ensemble Kalman filter with a fast implementation and localized deterministic scheme | |
| CN115691140B (en) | Analysis and prediction method for space-time distribution of automobile charging demand | |
| Porfyri et al. | Calibration of a second-order traffic flow model using a metamodel-assisted Differential Evolution algorithm | |
| Sirohi et al. | Uncertainty-aware lidar panoptic segmentation | |
| CN116244974A (en) | Method and device for evaluating authenticity of simulated traffic flow data | |
| Zhong et al. | Estimating link flows in road networks with synthetic trajectory data generation: Inverse reinforcement learning approach | |
| Pérez et al. | Model testing and assessment: perspectives from a swarm intelligence, agent-based model of forest insect infestations | |
| Morgan et al. | Prediction Intervals of Machine Learning Models for Taxi Trip Length | |
| Coelho et al. | Efficient protocol for solving combinatorial graph problems on neutral-atom quantum processors | |
| Maghsood et al. | Detection of steering events using hidden Markov models with multivariate observations |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |