Machine Learning Framework for Improving Accuracy of Probe Speed Data

Abstract

A tremendous potential exists for using probe data to support various traffic operations activities. However, limited real-time probe data, especially on arterial roads, have become a barrier to realizing the full potential of this technology. In the absence of real-time probe data, traffic speeds are estimated via prediction engines trained on historical data. The accuracy of such traditional speed estimation approaches could be significantly improved if real-time data available through nearby infrastructure-mounted (IM) sensors were incorporated in the prediction process. This paper develops a machine learning framework for generating probe-like speed data from IM sensors with the aim of improving the accuracy of probe speed data during periods of low probe penetration. The framework includes using a pattern recognition system for extracting trends from historical traffic speed data. The extracted patterns together with historical temporal traffic flow data are used to prepare a representative training set for a deep learning–based model that can transform IM sensor data into probe-like data. The proposed approach successfully generated pseudo-probe data sets from nearby IM sensors with about 4.8 and 9.6 km/h mean absolute error on freeways and arterials, respectively. A comparative analysis with baseline methods proved the superiority of the methodology adopted.