Intersection Type Classification from Connected Vehicle Data Using a Convolutional Neural Network

There are four broad types of traffic control at three- and four-legged intersections: traffic signals, four-way stops, two-way stops, and roundabouts. The scope and approach for mapping and labeling these intersections varies significantly by agency, making it difficult to obtain a geospatial inventory of the types of intersection control without field visits. This data can be used by autonomous vehicles to improve navigation, by insurance companies to evaluate driver behavior, or by transportation agencies to update inventories and determine performance measures to assess infrastructure. With road networks that change as much as ten percent each year, techniques to systematically classify the type of intersections on the roads need to be provided. This study applies a convolutional neural network (CNN) to high-resolution connected vehicle (CV) trajectory data to automatically classify intersections into four categories: signalized, four-way stop, two-way stop, and roundabout. Sampled demand, speeds, and accelerations around intersections, as well as geometric characteristics are extracted from CV data and used to train and evaluate the CNN model. The classification was applied to 600 intersections in Indiana, Ohio, and Pennsylvania using over 2,000,000 vehicles trajectories and 18,000,000 GPS points. An evaluation of the developed model shows an accuracy of 98% for the entire data set and 97% for the test data set. Since the proposed technique relies solely on commercial CV trajectory data and the location of intersection centers, intersection classifications can be systematically performed at the city, state, or national levels with minimum manual labor required.