CN115935642A - Intelligent vehicle extreme test scene automatic generation method and system based on accident information - Google Patents
Intelligent vehicle extreme test scene automatic generation method and system based on accident information Download PDFInfo
- Publication number
- CN115935642A CN115935642A CN202211546735.5A CN202211546735A CN115935642A CN 115935642 A CN115935642 A CN 115935642A CN 202211546735 A CN202211546735 A CN 202211546735A CN 115935642 A CN115935642 A CN 115935642A
- Authority
- CN
- China
- Prior art keywords
- scene
- information
- constraint
- accident
- traffic accident
- 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
- 238000012360 testing method Methods 0.000 title claims abstract description 72
- 238000000034 method Methods 0.000 title claims abstract description 41
- 206010039203 Road traffic accident Diseases 0.000 claims abstract description 61
- 230000003068 static effect Effects 0.000 claims abstract description 56
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 34
- 238000003058 natural language processing Methods 0.000 claims abstract description 24
- 238000007781 pre-processing Methods 0.000 claims abstract description 13
- 238000005516 engineering process Methods 0.000 claims abstract description 9
- 230000001133 acceleration Effects 0.000 claims description 16
- 238000013210 evaluation model Methods 0.000 claims description 16
- 230000006870 function Effects 0.000 claims description 16
- 238000011835 investigation Methods 0.000 claims description 15
- 230000008859 change Effects 0.000 claims description 8
- 238000012549 training Methods 0.000 claims description 8
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000012937 correction Methods 0.000 claims description 4
- 230000008030 elimination Effects 0.000 claims description 4
- 238000003379 elimination reaction Methods 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 4
- 238000012502 risk assessment Methods 0.000 claims description 4
- 238000005314 correlation function Methods 0.000 claims description 3
- 238000007689 inspection Methods 0.000 claims description 3
- SAZUGELZHZOXHB-UHFFFAOYSA-N acecarbromal Chemical compound CCC(Br)(CC)C(=O)NC(=O)NC(C)=O SAZUGELZHZOXHB-UHFFFAOYSA-N 0.000 claims 1
- 230000000875 corresponding effect Effects 0.000 description 9
- 230000009471 action Effects 0.000 description 8
- 238000000605 extraction Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 6
- 238000010276 construction Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 4
- 230000007547 defect Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 230000006378 damage Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 208000014674 injury Diseases 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000013441 quality evaluation Methods 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Images
Classifications
-
- 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
- Traffic Control Systems (AREA)
Abstract
The invention discloses an intelligent vehicle extreme test scene automatic generation method and system based on accident information, wherein the method comprises the following steps: acquiring a survey information text of an actual traffic accident, preprocessing the survey information text, and extracting scene static elements of the actual traffic accident from the survey information text by using a natural language processing technology, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information; initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm; and taking the extracted scene static elements of the actual traffic accidents as static elements of a scene to be generated, and combining the extracted scene static elements with the searched dynamic parameter combination set to generate an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accidents. The intelligent vehicle extreme test scene library is supported to build, and the test landing of the intelligent vehicle extreme test scene library is accelerated.
Description
Technical Field
The invention relates to the technical field of automatic driving simulation, in particular to an intelligent vehicle extreme test scene automatic generation method and system based on accident information.
Background
Simulation test is an important means for verifying the safety performance of the intelligent vehicle, and the construction of a huge test scene library is the primary basis of the test. For the construction of a test scene, research is mainly focused on a natural driving scene, a standard regulation scene and an accident extreme scene of a vehicle at present, and in order to make up for the defect that the overall test efficiency is low due to low dangerous scene triggering frequency in a test method based on the natural driving scene, a method for extracting the extreme scene from accident data to accelerate the overall test process has attracted extensive attention in the industry. The existing research of extracting the extreme scene of the accident focuses on the clustering extraction of the key type scene, some scene difference information can be ignored, and the automation degree of the building process of the extreme scene is not high finally. In order to solve the above problems, a method for automatically extracting an extreme test scenario from existing accident investigation information is needed.
Disclosure of Invention
The invention provides an intelligent vehicle extreme test scene automatic generation method and system based on accident information, which support the construction of an intelligent vehicle extreme test scene library and accelerate the test landing of the intelligent vehicle extreme test scene library.
In order to achieve the technical purpose, the invention adopts the following technical scheme:
an intelligent vehicle extreme test scene automatic generation method based on accident information comprises the following steps:
acquiring an investigation information text of an actual traffic accident, preprocessing the investigation information text, and extracting scene static elements of the actual traffic accident from the survey information text by using a natural language processing technology, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information;
initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm;
and taking the extracted scene static elements of the actual traffic accidents as static elements of a scene to be generated, and combining the static elements with the searched dynamic parameter combination set to generate an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accidents.
Further, the scene starting and ending point information comprises relative position information and relative driving direction information of the traffic accident participants, the collision basic information comprises a collision position and a collision angle of the traffic accident, and the collision position comprises an accident point geographical position and a position of a collision point on the vehicle body.
Further, the preprocessing the search information text comprises: secondary information elimination, word stem and word shape reduction and spelling error correction; the method comprises the steps of extracting scene static elements of actual traffic accidents from the text by utilizing a natural language processing technology, and specifically extracting scene static element information from the preprocessed accident description information text by utilizing a trained natural language processing model.
Further, the searching for the dynamic parameter combination set of the scene to be generated by using the depth deterministic gradient policy algorithm specifically includes: and gradually searching a dynamic parameter combination set of the scene to be generated by adopting the depth certainty gradient strategy algorithm and according to the initial speed, the acceleration and the position of the scene participant to be generated until the generated dynamic parameter combination meets the condition that the basic collision information of the actual traffic accident is the same, and finishing the search.
Further, a model training reward function of the depth certainty gradient strategy algorithm comprises a rationality constraint evaluation model and a risk evaluation model; the rationality constraint evaluation model comprises a traffic regulation constraint and a vehicle stability constraint; the risk evaluation model is used for guiding the traffic accident participants of the scene to be generated to gradually travel to the collision positions of the actual traffic accidents, and the rationality constraint evaluation model is used for preventing the irrationality of the generated dynamic parameter combination; the reward function expression is as follows:
w＝c 1 w 1 +c 2 w 2 +c 3 w 3
wherein w represents a reward function value; w is a 1 、w 2 、w 3 Respectively representing traffic regulation constraint reward, stability constraint reward and scene risk reward, c 1 、c 2 、c 3 The distribution coefficient for the corresponding prize.
Further, the traffic rule constraint and the vehicle stability constraint control whether the traffic rule constraint and the vehicle stability constraint respectively participate in the rationality constraint evaluation model or not through corresponding constraint switch fields, if the actual traffic accident is caused by the fact that the participator does not obey the traffic rule, the traffic rule constraint does not participate in the rationality constraint evaluation model, and if the actual traffic accident is caused by the fact that the participator does not obey the vehicle stability, the vehicle stability constraint does not participate in the rationality constraint evaluation model.
Further, the vehicle stability constrains yaw-rate from the vehicle
Transverse velocity v y Acceleration a and rotation angle θ are subjected to dynamic constraint:
|θ|≤θ lim
|a|≤a lim
in the formula, alpha r Indicating rear wheel side slip angle, v x And v y Respectively representing the longitudinal and lateral speed of the vehicle's center of mass,
to yaw angular velocity,/ r And l f The distances from the center of mass of the vehicle to the rear axle and the front axle respectively; alpha is alpha r,lim Is a rear wheel side deflection angle threshold value>
Linear cornering stiffness of the rear wheel; theta lim Indicating a turning angle limit of the vehicle, a lim Indicating the vehicle acceleration limit.
Further, the risk assessment model considers the distance between the participants and the distance from each participant to the actual traffic accident collision position, and is in negative correlation with each distance variation value respectively:
in the formula, w 3 A reward representing a scene risk constraint, Δ D representing a change in distance of the participant to the collision location, Δ D representing an amount of change in distance between the participants, n representing a number of scene participants, z 1 、z 2 The distribution coefficients of the respective portions are respectively expressed, and f and g respectively represent negative correlation functions with the variation of the distance between the participants and the variation of the distance from the participants to the collision point as arguments.
Further, the static elements and the dynamic parameter combination set are combined to generate an intelligent vehicle traffic accident extreme test scene, and specifically, the static elements and the dynamic parameter combination set are converted into an OpenX series file based on a scene general format in an XML format.
An intelligent vehicle extreme test scene automatic generation system based on accident information comprises:
a text pre-processing module to: preprocessing the acquired survey information text of the actual traffic accident;
a natural language processing module to: extracting scene static elements of the actual traffic accident from the preprocessed investigation information text, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information;
a dynamic parameter search module to: initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm;
an extreme test scenario generation module to: the output results of the natural language processing module and the dynamic element searching module, namely the static and dynamic elements of the extreme scene, are butted, and the extreme test scene of the intelligent vehicle traffic accident similar to the actual traffic accident is generated in a combined manner, namely the extreme test scene is converted into a test scene general format file;
a storage module to: and storing the model hyper-parameters of the natural language processing module and the depth certainty gradient strategy algorithm, and converting to obtain various extreme test scene files.
Advantageous effects
On one hand, the invention promotes the deep extension from the accident investigation field to the intelligent vehicle testing field, and can expand the existing intelligent vehicle testing scene library; in the second aspect, the method focuses on building an extreme test scene which is highly similar to an actual traffic accident, the obtained test scene has obvious danger, and the test landing of the intelligent vehicle and related safety products can be accelerated to a certain extent; in a third aspect, the extreme test scene finally generated by the method is an OpenX series file in a scene general format based on an XML format, and the extracted extreme test scene information can be directly converted into a standard format which is general for simulation software, so that the universality of the result of the method is improved. Therefore, the invention is applied to the test application links of intelligent vehicles and related safety products, can test the safety performance of the intelligent vehicles in an extremely dangerous scene, finds the defects of the tested vehicles in time and accelerates the upgrading and perfecting of the safety function.
Drawings
FIG. 1 is an automatic generation method of an extreme test scenario of an intelligent vehicle based on accident information.
Fig. 2 is a scene static element extraction method based on natural language processing technology.
FIG. 3 is a method for generating extreme scene dynamic element confrontation based on DDPG algorithm.
FIG. 4 is a schematic diagram of a constraint model in the DDPG algorithm.
Figure 5 is a DDPG algorithm framework.
Fig. 6 is a schematic diagram of scene dynamic parameter search.
FIG. 7 is an intelligent vehicle extreme test scenario automation generation system architecture diagram based on accident information.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present application more comprehensible, embodiments accompanying the present application are described in detail below with reference to the accompanying drawings. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. This application is capable of embodiments in many different forms than those described herein and those skilled in the art will be able to make similar modifications without departing from the spirit of the application and it is therefore not intended to be limited to the embodiments disclosed below.
Example 1
The embodiment provides an automatic generation method of an extreme test scene of an intelligent vehicle based on accident information, and the method of the embodiment is shown in fig. 1 and comprises the following main steps:
and S10, acquiring a survey information text of the actual traffic accident, preprocessing the survey information text, and extracting scene static elements of the actual traffic accident from the survey information text by utilizing a natural language processing technology, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information.
In this embodiment, before importing the original complex accident investigation information text into the natural language processing model, certain preprocessing must be performed on the text information, and the specific steps of extracting static elements of a scene refer to fig. two, which mainly includes:
(1) Acquiring and arranging an accident investigation information text, wherein the information content generally relates to text information of thousands of fields, the content is various, and only part of key information is needed for extreme scene construction;
(2) Text secondary information elimination, which mainly eliminates unimportant element information for the analysis of scene elements at the next stage, such as punctuation marks which can not convey any effective information, such as some blank spaces, line feed marks and the like, and some description information which has no practical significance for static scene extraction, such as the year, month and day of an accident, the city of the accident and the like, so that a later natural language processing model can better identify key information;
(3) The text stem word form is restored, aiming at the problem that some text abbreviation modes in the accident investigation records can cause a natural language processing model to generate misinterpretation on the text meaning, the word stem word form is restored on the existing accident information text, and more accurate and more standard accident description text information can be obtained;
(4) Text spelling error correction, which is generally caused by the wrong entry of related investigators, makes the related scene elements in the incident description information more difficult to understand, so detecting and correcting the spelling errors can help to improve the accuracy of static scene element extraction.
And then, extracting scene static elements of the actual traffic accidents from the preprocessed survey information texts by utilizing a natural language processing technology. In this embodiment, a natural language processing model is constructed to screen representative information related to static elements of a scene, and the process mainly includes the following steps: firstly, keyword distribution is carried out, a relatively rich keyword lexicon is predefined, and the keyword lexicon generally comprises accident participant information (such as vehicle types, driving intentions and the like), accident participant relative positions (such as oncoming vehicles, equidirectional incoming vehicles and the like), collision directions (such as direction information that a main vehicle is subjected to frontal collision of a secondary vehicle or is subjected to right side collision of the secondary vehicle and the like), accident road network information (such as key information of intersections, three lanes and the like), weather environments (such as clear days, overcast and rainy days and the like) and other related characteristic words; and then, judging the matching degree of each type of keywords in the accident description text and the keyword lexicon by calculating the similarity, and when the matching degree reaches a preset threshold value, considering the word segment as scene static key information.
The extracted road network information comprises the current road type, the number of lanes, the lane width and the like, the weather environment information comprises illumination, rainfall, visibility and the like, the scene starting and ending point information comprises the starting relative position of a participant, the coordinates of a collision occurrence place, the driving intention of the participant and the like, the collision basic information comprises the collision position, the collision angle and the like of a traffic accident, and the collision position comprises the geographical position of the accident point and the position of the collision point on the vehicle body.
And S20, initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm.
The scene starting and ending point information extracted in the step S10 is used for setting the initial state of the traffic accident participant at the scene starting time to be generated, and the initial state comprises an initial position and an initial driving direction; and then exploring the possibility of various specific scenes similar to the actual traffic accident from the initial state to a specific end state (namely the state at the collision position of the actual traffic accident) by adopting a depth deterministic gradient strategy algorithm (DDPG algorithm for short).
In this implementation, before searching for dynamic parameter combinations using the DDPG algorithm, the model of the DDPG algorithm is trained. The Critic network and the Actor network in the algorithm model adopt a well-known, simple and efficient 4-layer fully-connected network structure, an activation function between layers is set as ReLU, and an output layer of the Actor network adopts a tanh function to ensure the boundedness of output speed and acceleration; the model training parameters are basically set as follows: the sampling step length is 0.1s, each training period is 25s, and the maximum training times are 10000 times.
In this embodiment, the observation state s of the scene environment mainly includes: vehicle speed (magnitude and direction), acceleration (magnitude and direction), vehicle coordinate position; the action parameters output by the action strategy network mainly refer to: operations such as acceleration and deceleration and steering; the value Q network is mainly used for acquiring the current state s i Take action a i The corresponding reward. Referring to fig. 5, this embodiment adopts a specific training process using the DDPG algorithm model:
a) Initializing an action strategy network, a value Q network and an experience playback pool with a certain capacity;
b) Initializing a scene according to the static element information of the scene;
c) Observing and outputting the current scene state s t (vehicle speed, acceleration, position)
d) According to the current state, the online action strategy network outputs the corresponding action a t And observe the next state s in the scene t+1 (ii) a Simultaneous online Q network according to a t 、s t 、s t+1 Outputting the reward r of the current action t+1 And guiding the updating of the online strategy network parameters;
e) Change the main wheel(s) t ,a t ,r t ,s t+1 ) Storing in an experience playback pool;
f) Sampling N sets of data N(s) from an empirical playback pool i ,a i ,r i ,s i+1 ) As training data for the online policy network and the online Q network;
g) Network mu's theta according to target policy μ’ And a target Q network Q' (s, a | θ) Q′ ) The following equation is obtained:
y i ＝r i +γQ′(s i+1 ,μ′s i+1 θ μ′ |θ Q′ )
h) Updating online Q network parameters according to the reverse transfer of the network gradient, wherein the mean square error loss function is as follows:
i) Updating online policy network parameters according to the policy gradient of the selected sample:
j) And periodically updating the parameters of the target policy network and the target Q network in a soft updating mode:
wherein tau is a soft update parameter used for adjusting the update speed of the network.
In implementation, in order to guide the model learning direction of the DDPG algorithm, the design of the reward function mainly considers two aspects of a rationality constraint evaluation model and a risk evaluation model; the risk index is used for guiding the participator to gradually drive to a collision point, and the rationality index is used for preventing the generation of an invalid test scene caused by the irrationality of parameter combination.
Firstly, the dynamic parameter rationality constraint comprises a traffic regulation constraint and a vehicle stability constraint, and because the particularity of certain accidents, namely the accidents are directly caused by the irrationality of the constraints, a constraint switch field is correspondingly arranged before the rationality constraint evaluation and is specially used for controlling the on-off of a rationality constraint evaluation module; the constraint switch field is obtained by analyzing keywords related to accident reasons in the accident basic information in the process of extracting the static elements of the scene, if the accident is caused by that one party of the participators does not comply with the traffic rules, the traffic rules switch field is regarded as a closed state, namely, when the dynamic parameters are searched, the algorithm model can be searched towards the direction of violating the traffic rules without constraint; similarly, the operation principle of the stability constraint switch field is basically similar.
The vehicle stability constraint mainly considers that the unreasonable matching of speed and acceleration in dynamic parameters causes vehicle rollover, so that the yaw velocity of the vehicle is required
Transverse velocity v y Acceleration a and rotation angle θ are subjected to dynamic constraint:
|θ|≤θ lim
|a|≤a lim
in the formula, alpha r Indicating rear wheel side slip angle, v x And v y Respectively representing the longitudinal and lateral velocity of the vehicle's center of mass,
as yaw rate, /) r And l f The distances from the center of mass of the vehicle to the rear axle and the front axle respectively; alpha is alpha r,lim Is a rear wheel side deflection angle threshold value>
Linear yaw stiffness of the rear wheel; theta lim Indicating a turning angle limit of the vehicle, a lim Representing a vehicle acceleration limit;
in summary, the reward expression associated with the traffic regulation constraint and the vehicle stability constraint is:
secondly, the risk assessment model mainly considers the distance between the participants and the distance between each participant and the recorded accident collision point, and is respectively in negative correlation with each distance change value:
where Δ D represents a change in distance from the participant to the collision point, Δ D represents a change in distance between the participants, n represents the number of participants in the scene design, and z represents the number of participants in the scene design 1 、z 2 Respectively representing the distribution coefficient of each part, f and g respectively representing the variation in the distance between the participants and the parametersThe square-to-collision point distance varies as a negative correlation function of the independent variables.
The final algorithm model reward function expression is as follows:
w＝c 1 w 1 +c 2 w 2 +c 3 w 3
in the formula, w 1 、w 2 、w 3 Respectively representing a traffic regulation constraint reward, a stability constraint reward, and a scenario risk reward, c 1 、c 2 、c 3 The specific setting for the distribution coefficient corresponding to the reward is determined according to the training effect of the specific model.
The searching of the dynamic parameter combination set of the scene to be generated by adopting the DDPG algorithm specifically comprises the following steps: the method comprises the steps of taking the speed, the acceleration and the position of a scene participant to be generated as states in a depth certainty gradient strategy algorithm, taking the acceleration, the deceleration and the steering operation of the scene participant to be generated as actions in the depth certainty gradient strategy algorithm, circularly outputting the optimal action of a current vehicle according to the current scene state by adopting the depth certainty gradient strategy algorithm and according to the initial speed, the acceleration and the position of the scene participant to be generated, promoting the current scene to be converted into the next state, realizing the gradual search of a dynamic parameter combination set of the scene to be generated until the generated dynamic parameter combination meets the condition that the dynamic parameter combination is the same as collision basic information of an actual traffic accident, and finishing the search. Such as the search diagram shown in fig. 6.
And S30, taking the extracted scene static elements of the actual traffic accidents as static elements of a scene to be generated, and combining the static elements with the searched dynamic parameter combination set to generate an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accidents.
In this embodiment, the static and dynamic elements of the scene obtained in steps S10 and S20 have a regular format, i.e., "attribute-attribute value"; the format conversion script is mainly responsible for correspondingly modifying the attribute values of the extracted scene elements, and if element information which is not related in other accident descriptions or is not acquired by the method is not related in other accident descriptions, the corresponding default values are continuously selected and unchanged.
After obtaining the standard format file of the extreme test scene according to the method, sequentially storing the scene related configuration file and the corresponding model parameters by taking the scene number as a title; during later-stage scene inspection, only standard OpenX series format files corresponding to scenes are called in simulation visualization software, scene quality evaluation is carried out through three-dimensional visualization expression of the scenes, and corresponding adjustment is carried out on the model according to an evaluation result. The generated scenes can be automatically classified and stored according to different indexes, and the indexes can be 'the injury degree of scene personnel recorded in accident information', 'the type of scene participants', 'the cause of the accident scene', 'the specific form of the scene', and the like, so that the automatic classification function of the scenes is realized.
In the method for automatically generating the intelligent vehicle extreme test scene based on the accident information, the extreme scene is extracted and divided into two main parts. The first part adopts a natural language processing technology to extract scene static key elements from accident investigation text information, so that the manual extraction process of the existing scene features is omitted, and high automation is realized; before information extraction, necessary operations such as secondary information elimination, word stem and word shape reduction, spelling error correction and the like are carried out on the accident investigation text, and the accuracy and the rapidity of information extraction are guaranteed to a great extent. And in the second part, initializing scene parameters according to scene starting and ending point information and collision key information extracted in the previous part, exploring scene dynamic parameters through a DDPG algorithm, and finally building a plurality of extreme test scenes similar to the accident. The reward function of the algorithm model mainly considers two indexes of constraint rationality and scene risk, can guide dynamic parameter exploration to develop towards the direction of accident high reduction and risk exploration, prevents the generation of invalid test scenes caused by the irrational combination of parameters, and ensures the efficiency and the correctness of the generation of extreme test scenes. Finally, scene static and dynamic key information obtained by the first two parts is converted into a simulation software universal format through a data conversion script, so that the universal applicability of the method is improved, and the later-stage scene inspection and maintenance are facilitated.
In conclusion, the method can realize automatic extraction of relevant extreme test scenes from basic information of actual traffic accident investigation, and make up for the defect that scenes are built after accident features are extracted manually in the prior art. The reasonability constraint and the risk evaluation item considered in the reward function can ensure the rapidity and the correctness of the scene building process. Finally, the applicability and maintainability of the method are enhanced by writing a data conversion script.
Example 2
The embodiment provides an intelligent vehicle extreme test scenario automatic generation system based on accident information, which is shown in fig. 7 and includes:
a text pre-processing module to: preprocessing the acquired survey information text of the actual traffic accident;
a natural language processing module to: extracting scene static elements of the actual traffic accident from the preprocessed investigation information text, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information;
a dynamic parameter search module to: initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm;
an extreme test scenario generation module to: the output results of the natural language processing module and the dynamic element searching module, namely the static and dynamic elements of the extreme scene, are butted, and the extreme test scene of the intelligent vehicle traffic accident similar to the actual traffic accident is generated in a combined manner, namely the extreme test scene is converted into a test scene general format file;
a storage module to: and storing the model hyper-parameters of the natural language processing module and the depth certainty gradient strategy algorithm, and converting to obtain various extreme test scene files. The storage module can classify and store the generated scenes according to different indexes, and the indexes can be 'the damage degree of scene personnel recorded in accident information', 'the type of scene participants', 'the cause of an accident scene', 'the specific form of the scene', and the like, so that the automatic classification function of the scenes is realized.
The specific implementation of each module included in the intelligent vehicle extreme test scenario automatic generation system based on accident information in this embodiment is the same as that described in the method in embodiment 1, and is not repeated here.
The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the claims of the present application.
Claims (10)
1. An intelligent vehicle extreme test scene automatic generation method based on accident information is characterized by comprising the following steps:
acquiring an investigation information text of an actual traffic accident, preprocessing the investigation information text, and extracting scene static elements of the actual traffic accident from the survey information text by using a natural language processing technology, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information;
initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm;
and taking the extracted scene static elements of the actual traffic accidents as static elements of a scene to be generated, and combining the extracted scene static elements with the searched dynamic parameter combination set to generate an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accidents.
2. The intelligent vehicle extreme test scene automatic generation method according to claim 1, wherein the scene starting and ending point information comprises relative position information and relative driving direction information of traffic accident participants, the collision basic information comprises a collision position and a collision angle of the traffic accident, and the collision position comprises an accident point geographic position and a position of a collision point on a vehicle body.
3. The intelligent vehicle extreme test scenario automatic generation method of claim 1, wherein preprocessing the inspection information text comprises: secondary information elimination, word stem and word shape reduction and spelling error correction; the method comprises the steps of extracting scene static elements of actual traffic accidents from the text by utilizing a natural language processing technology, and specifically extracting scene static element information from the preprocessed accident description information text by utilizing a trained natural language processing model.
4. The method for automatically generating the intelligent vehicle extreme test scenario according to claim 1, wherein the searching for the dynamic parameter combination set of the scenario to be generated by using a depth deterministic gradient strategy algorithm specifically comprises: and gradually searching a dynamic parameter combination set of the scene to be generated by adopting the depth certainty gradient strategy algorithm and according to the initial speed, the acceleration and the position of the scene participant to be generated until the generated dynamic parameter combination meets the condition that the basic collision information of the actual traffic accident, and finishing the search.
5. The automatic generation method of the intelligent vehicle extreme test scene according to claim 1, characterized in that a model training reward function of the depth certainty gradient strategy algorithm comprises a rationality constraint evaluation model and a risk evaluation model; the rationality constraint evaluation model comprises a traffic regulation constraint and a vehicle stability constraint; the risk assessment model is used for guiding a traffic accident participant of a scene to be generated to gradually drive to a collision position of an actual traffic accident, and the rationality constraint assessment model is used for preventing the irrationality of the generated dynamic parameter combination; the reward function expression is as follows:
w＝c 1 w 1 +c 2 w 2 +c 3 w 3
wherein w represents a value of the reward function; w is a 1 、w 2 、w 3 Respectively representing traffic regulation constraint reward, stability constraint reward and scene risk reward, c 1 、c 2 、c 3 The distribution coefficient for the corresponding prize.
6. The method for automatically generating the intelligent vehicle extreme test scene according to claim 5, wherein the traffic rule constraint and the vehicle stability constraint are controlled through corresponding constraint switch fields to determine whether the traffic rule constraint and the vehicle stability constraint respectively participate in the rationality constraint evaluation model, if the actual traffic accident is caused by the fact that the participants do not obey the traffic rules, the traffic rule constraint does not participate in the rationality constraint evaluation model, and if the actual traffic accident is caused by the fact that the participants are in the vehicle stability, the vehicle stability constraint does not participate in the rationality constraint evaluation model.
7. The intelligent vehicle extreme test scenario automatic generation method of claim 5, wherein the vehicle stability constraint is derived from a yaw rate of the vehicle
Transverse velocity v y Acceleration a and rotation angle θ are subjected to dynamic constraint: />
|θ|≤θ lim
|a|≤a lim
In the formula, alpha r Indicating rear wheel side slip angle, v x And v y Respectively representing the longitudinal and lateral speed of the vehicle's center of mass,
as yaw rate, /) r And l f The distances from the center of mass of the vehicle to the rear axle and the front axle respectively; alpha is alpha r,lim Is a rear wheel side deflection angle threshold value>
Linear cornering stiffness of the rear wheel; theta lim Indicates the turning angle limit of the vehicle, a lim Indicating the vehicle acceleration limit.
8. The method for automatically generating the intelligent vehicle extreme test scene as claimed in claim 5, wherein the risk assessment model considers the distance between the participants and the distance between each participant and the actual traffic accident collision position, and is in negative correlation with each distance variation value respectively:
in the formula, w 3 A reward representing a scene risk constraint, Δ D representing a change in distance of the participant to the collision location, Δ D representing an amount of change in distance between the participants, n representing a number of scene participants, z 1 、z 2 The distribution coefficients of the respective portions are respectively expressed, and f and g respectively represent negative correlation functions with the variation of the distance between the participants and the variation of the distance from the participants to the collision point as arguments.
9. The method for automatically generating the intelligent vehicle extreme test scenario as claimed in claim 1, wherein the static elements and the dynamic parameter combination set are combined to generate the intelligent vehicle traffic accident extreme test scenario, and specifically, the static elements and the dynamic parameter combination set are converted into an OpenX series file based on XML format.
10. The utility model provides an extreme test scene automatic generation system of intelligent car based on accident information which characterized in that includes:
a text pre-processing module to: preprocessing the acquired survey information text of the actual traffic accident;
a natural language processing module to: extracting scene static elements of the actual traffic accident from the preprocessed investigation information text, wherein the scene static elements comprise scene road network information, weather environment information, scene starting and ending point information and collision basic information;
a dynamic parameter search module to: initializing a scene to be generated based on the extracted static elements of the scene, and searching a dynamic parameter combination set of the scene to be generated by adopting a depth certainty gradient strategy algorithm;
an extreme test scenario generation module configured to: the output results of the natural language processing module and the dynamic element searching module, namely the static and dynamic elements of the extreme scene, are butted, and an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accident is generated by combining, namely the intelligent vehicle traffic accident extreme test scene is converted into a test scene general format file;
a storage module to: and storing the model hyperparameters of the natural language processing module and the depth certainty gradient strategy algorithm, and various extreme test scene files obtained through conversion.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211546735.5A CN115935642A (en) | 2022-12-05 | 2022-12-05 | Intelligent vehicle extreme test scene automatic generation method and system based on accident information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211546735.5A CN115935642A (en) | 2022-12-05 | 2022-12-05 | Intelligent vehicle extreme test scene automatic generation method and system based on accident information |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115935642A true CN115935642A (en) | 2023-04-07 |
Family
ID=86655433
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211546735.5A Pending CN115935642A (en) | 2022-12-05 | 2022-12-05 | Intelligent vehicle extreme test scene automatic generation method and system based on accident information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115935642A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117033562A (en) * | 2023-10-09 | 2023-11-10 | 北京中机车辆司法鉴定中心 | Dangerous prediction method and system based on scene element matching |
-
2022
- 2022-12-05 CN CN202211546735.5A patent/CN115935642A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117033562A (en) * | 2023-10-09 | 2023-11-10 | 北京中机车辆司法鉴定中心 | Dangerous prediction method and system based on scene element matching |
| CN117033562B (en) * | 2023-10-09 | 2024-02-23 | 北京中机车辆司法鉴定中心 | Dangerous prediction method and system based on scene element matching |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106874597B (en) | highway overtaking behavior decision method applied to automatic driving vehicle | |
| CN107169567B (en) | Method and device for generating decision network model for automatic vehicle driving | |
| Li et al. | Lane changing intention recognition based on speech recognition models | |
| CN111008293A (en) | Visual question-answering method based on structured semantic representation | |
| CN111241279B (en) | Natural language relation extraction method based on multi-task learning mechanism | |
| CN111966812B (en) | Automatic question answering method based on dynamic word vector and storage medium | |
| CN114358128A (en) | Method for training end-to-end automatic driving strategy | |
| CN110619042B (en) | Neural network-based teaching question and answer system and method | |
| KR20220060780A (en) | Knowledge based dialogue system and method for language learning | |
| CN115935642A (en) | Intelligent vehicle extreme test scene automatic generation method and system based on accident information | |
| CN117787342A (en) | Adaptive labeling and attention pruning method for visual language model | |
| CN118379755A (en) | Multi-mode data anomaly identification method and device, electronic equipment and storage medium | |
| Li et al. | Research on lane change prediction model based on GBDT | |
| CN113743466B (en) | Road type identification method and system based on decision tree | |
| CN113033180B (en) | Automatic generation service system for Tibetan reading problem of primary school | |
| Khan et al. | Safety Testing of Automated Driving Systems: A Literature Review | |
| CN107045639B (en) | Following behavior generation method based on time automaton | |
| CN114701517A (en) | Multi-target complex traffic scene automatic driving solution based on reinforcement learning | |
| CN115096305A (en) | Intelligent driving automobile path planning system and method based on generation of countermeasure network and simulation learning | |
| CN113296500B (en) | Local path planning method and system | |
| CN110619118B (en) | Automatic text generation method | |
| CN114818381A (en) | Method for constructing test scene library of automatic driving automobile | |
| CN110667597B (en) | Driving style state identification method based on vehicle controller local area network data information | |
| CN114117944A (en) | Model updating method, device, equipment and readable storage medium | |
| CN118364905B (en) | Wrong question labeling data generation method and device |
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 |