Identification of Road Network Intersection Types from Vehicle Telemetry Data Using a Convolutional Neural Network

GPS trajectories collected from automotive telematics for insurance purposes go beyond being a collection of points on the map. They are in fact powerful data source that we can use to extract map and road network properties. While location junctions is readily available, information about traffic control element regulating intersection typically unknown. However, this would be helpful, e.g., contextualizing driver’s behavior. Our focus map-matched OBD-dongle dataset provided by Canadian company classify intersections into three classes according type present: light, stop sign, or no sign. We design convolutional neural (CNN) classifying intersections. The takes as entries, defined number trips, speed acceleration profiles over each segment one meter window around intersection. method outperforms two other competing approaches, achieving 99% overall accuracy. Furthermore, our CNN model infer even with few 25 trips.