Dynamic traffic flow prediction based on GPS Data

This paper presents a solution for traffic flow prediction in a city area. GPS devices offer new opportunities for short-term traffic prediction, especially in arterial road networks where traditional fixed-location sensors are sparse or expensive to install. However, GPS data is often sparse both temporally and spatially. On its own, it is often insufficient for real-time traffic prediction. We consider the fusion of two types of data for the purpose of dynamic traffic prediction: GPS data that is provided as point speeds, rather than trajectories, as well as traffic data that is available from previous tracking. Inspired by the observation that a driver often has its own route selection behavior, we define a mobility pattern as a consecutive series of road segment/link selections that exhibit frequent appearance along all the itineraries of the vehicle. We predict the traffic flow using a hybrid method based on Variable-order Markov Model and adding on top of it the average speed of all the vehicles passing through each road segment. Our solution comes with a highly scalable traffic simulator application that can be used to predict, manage and optimize car traffic in cities. The prediction accuracy is estimated according to various criteria.