Vehicular traffic flow intensity detection and prediction through mobile data usage

Abstract

A novel approach, consisting of an ensemble of data-mining and machine learning techniques, is proposed to prove that it is possible to extract and predict vehicular traffic patterns from mobile usage data. An anonymized mobile phone usage dataset from a telecommunications provider in Malta was used to generate an origin-destination (OD) matrix that defines the top two locations towards which each user travels to through clustering. The OD matrix was used to infer user trips over fastest routes between these top two locations across time. We then applied spatial binning techniques to deduce the aggregate distribution of traffic load on the traffic network. A predictive model based on an artificial neural network was trained with grid nodes’ traffic levels in a time series to predict traffic level for specific nodes. Our findings are promising and show that the built models are more effective to measure and predict traffic flow demand for specific locations rather than the actual traffic flow rate. The proposed solution needs improvement by adding a dynamic traffic assignment to the whole algorithm. This would give more accurate results, especially for traffic flow points that tend to be congested, by capturing user route selection changes and get more precise localization of delay causes.