Development of a Method for Predicting the Probability of Traffic Accidents Using a Multimodal AI Model of Structured Data and Satellite Images

Abstract

Predicting traffic accidents is an important issue for improving public safety. In this study, we developed an AI model which predicts the probability of traffic accidents in Fukuoka City with spatial and temporal resolutions of ~130m and 1 month, respectively. A multimodal machine-learning model was developed by combining various structured data (roads, demographics and economic statistics, weather, events, etc.) and satellite imagery, and was trained using personal injury accident data between 2016 and 2017 in Fukuoka City. As a result, the model succeeded in predicting the outcome for Fukuoka City in 2018 at an AUC of the ROC curve of 77%.