Generation of two-wheeler accident pre-crash scenarios based on multimodal data

In the Chinese traffic system, a crucial role is played by two-wheelers (TWs), yet significant road safety challenges are posed by them, with many TW drivers being killed or injured in traffic accidents annually. With the rapid advancement of autonomous driving technology, it is becoming feasible to reduce such accidents and minimize the number of casualties. Understanding accident scenarios is pivotal for developing and testing autonomous driving technologies. To ensure the reliability of these technologies, especially in the Chinese traffic environment, a method combining textual and numerical multimodal data is proposed based on the China In-Depth Accident Study (CIDAS) database. The total of 2,463 cases of vehicle-to-TW crashes are selected from the CIDAS database, and thorough extraction and analysis of textual and numerical data are conducted. The use of a Latent Dirichlet Allocation (LDA) topic analysis model based on Term Frequency Inverse Document Frequency (TF-IDF) effectively mines and understands driver behavior information from textual data. This information not only provides robust data support for understanding accident causes but also integrates subjective driver behaviors with objective road environment information in accident scenario generation for the first time. Finally, the integrated textual and numerical data undergo cluster analysis using spectral clustering algorithms, identifying six typical pre-crash scenarios between vehicles and TWs. Compared to traditional studies on generating traffic accident scenarios, this research not only offers deeper insights into driver behavior but also provides essential foundations for selecting typical scenarios and diversifying sources of scenario information for testing autonomous vehicles.