Scenario Modeling And Application Of Text-Based Traffic Accident Reports

Virtual scenario testing can effectively avoid the problems of high cost and high safety risk in real-world testing of autonomous vehicles,and using traffic accident data to drive the construction of test scenarios is one of the common methods for virtual scenario testing of autonomous vehicles.However,compared to structured accident data,unstructured accident data has been less studied,resulting in the waste of a large amount of data resources.In this paper,with the background of autonomous vehicle testing,we address the problems of low utilization rate of unstructured text-based accident data,low automation and simple dynamic scene modeling in the current scene construction,combine dynamic knowledge mapping technology to build a conceptual model of accident domain ontology,use the model as the basis for extracting the scene instance information in accident reports to generate a dynamic knowledge map of accident scenes,and finally build Finally,we construct an autonomous driving test scenario library based on the accident dynamic knowledge graph to provide a scenario basis for the testing of autonomous driving vehicles.The research work in this paper is carried out in the following three aspects:(1)Construction of ontology concept model of accident domainIn this paper,we combine the knowledge related to dynamic knowledge graph with the concept of accident domain,and treat the vehicle action as a relationship between specific concepts occurring in a certain period of time.Based on the bottom-up construction process of the knowledge graph,we first obtain the shortest form noun phrases from text utterances through component syntactic analysis,and then embed the obtained noun phrases through SBERT text pre-training model and cluster them by K-Means algorithm,and manually generalize the clustering results to obtain the concepts.In the next step,the relationship of connected concepts is obtained by dependent syntactic analysis with external semantic network,and finally the concept attributes are supplemented by dependent syntactic analysis.A total of11 categories of concepts,13 categories of relations and 7 categories of attributes are obtained from 100 textual accident reports to construct a conceptual model of the accident domain ontology.The comparison shows that the ontology concept model construction method of accident domain proposed in this paper has good scalability and can comprehensively represent the development process of accidents with good generality.(2)Accident dynamic knowledge mapping instance information extractionBy using SBERT semantic similarity + dependency syntax analysis,it replaces the problem of constructing keyword templates in the traditional information extraction method and has higher flexibility and scalability.By analyzing the static scene information and dynamic scene information,different instance information extraction schemes are designed and extracted separately.The experimental results prove that the static instance information extraction accuracy proposed in this paper is maximally improved by more than 20% compared with previous work,and the reports with the extraction accuracy of vehicle verb,direction and action timing information from 50% to 100% in a single text account for 85%,69% and 65% of all experimental reports,respectively.The problems of poor scalability of scene information extraction templates and incomplete scene information representation in previous work are solved.(3)Accident dynamic knowledge mapping applicationBased on the first two parts of extracting structured scene information and mapping to generate accident scene dynamic knowledge graphs,the extracted scene instance information is mapped to the accident domain ontology concept model to generate accident dynamic knowledge graphs.Using the Neo4 j graph database to store the scenario mapping data,the accident scenarios containing complex semantics can be queried to further realize the testing of specific types of scenarios.And the scenario reconstruction file is generated by MATLAB traffic simulation tool,which can be exported to OpenScenario format and used for testing of autonomous vehicle functions,and finally the scenarios obtained from accident reports can be tested for the effectiveness of autonomous vehicle algorithms by example,realizing the whole process from textual accident reports to scenario applications.