Deep Learning of OpenStreetMap Images Labeled Using Road Traffic Accident Data

Identifying potentially high-risk areas is an important task for preventing road traffic accidents. In this paper, we present an attempt to train deep Convolutional Neural Networks (CNNs) on OpenStreetMap (OSM) images, each of which is labeled “high-risk” or “low-risk” using the data of road traffic accidents occurred in Okayama Prefecture, Japan from 2010 to 2021. It is shown that the trained CNNs can correctly predict whether a prescribed area is high-risk or low-risk only from the corresponding OSM image with a high probability. We also present some results of applications of Grad-CAM to the trained CNNs in order to visualize their decision making process.