Neural Network Predicting Remote Vehicle Movement with Vehicle-to-Vehicle Data

Abstract

This paper presents a neural network developed for predicting the path of a remote vehicle using post facto created vehicle-to-vehicle (V2V) data and uses that prediction to determine whether it is safe for the host vehicle to change lanes. The data was collected in a 2013 experiment involving various drivers traveling on public roads in Ann Arbor, MI. The trips were on suburban roads, city roads and divided highways over a two-day period. The vehicular satellite global positioning system (GPS) data from movement over this period was gathered and post-processed to find vehicle paths within 10 meters of one another. The path traces of the two vehicles were combined to simulate what a V2V network would have provided to properly equipped vehicles if such a network and vehicles existed on real road networks demonstrating natural driving behavior. This research harnesses this data to determine the increased effectiveness of a neural network predicting the future path of remote vehicles and lane change safety when a V2V network is available. The most studied maneuver is overtaking. To a lesser extent, this paper also provides a view into how a neural network predicts remote vehicle behaviors using a host vehicle equipped with only perceptive hardware and no given information from the remote vehicle.