부산대학교 도서관

Road crashes cause more than 1.5 lakh deaths in India every year. The crash severity enables us to understand the road crash and design mitigation measures to reduce road crashes’ severity. While most states in India do not maintain a comprehensive database of road crashes, all states have to file First Information Reports (FIRs) on road crashes reported to the police. FIRs are recorded in text format by a police person, and they contain descriptive information related to the severity of the road crashes. This FIRs text data can be used for road crash severity prediction. In this study, we have designed the road crash severity modeling as a text classification problem. We labeled 2969 FIRs of Tamil Nadu state to pre-defined crash severity classes: fatal, grievous, and minor. The study developed a bi-directional LSTM model, and it achieved an F-1 score of 90 percent in measuring road crash severity. The bi-directional LSTM model outperformed the random forest model and baseline model. The model developed in this study can be applied to FIRs data of other states of India for road crash severity prediction and can be used as a quick tool by policymakers and road safety researchers. This study is a step towards automatically developing a database of road crash severity for all road crashes occurring in the country since most states do not have a road crash database.