RING-Net: road inference from GPS trajectories using a deep segmentation network

Abstract

Accurate and rich representation of roads in a map is critical for safe and efficient navigation experience. Often, open source road data is incomplete and manually adding roads is labor intensive and consequently expensive. In this paper, we propose RING-Net, an approach for Road INference from GPS trajectories using a deep image segmentation Network. Previous work on road inference is either focused on satellite images or GPS trajectories, but they are not compatible with each other when there is a lack of high quality data from either of the source types. Even though it is primarily focused on using GPS trajectories as its input, RING-Net architecture is flexible enough to be used with multiple data sources with minimal effort. More specifically, RING-Net converts raw GPS trajectories into multi-band raster images with trip related features, and infers roads with high precision. Experiments on public data show that Ring-Net can be used to improve the completeness of a road network. Our approach is promising to bring us one step closer to fully automated map updates.