Design and implementation of a single loop-detector filter for real-time traffic

Abstract

Inductive loop detectors (or ILDs) are electromagnetic sensors used to detect the passage of vehicles. They have been widely deployed on highways to measure traffic flow and occupancy. These measurements, combined with density models of traffic and state estimators, have been used in the past to estimate traffic density and thus infer the state of traffic (i.e. whether it is congested, in free flow, etc...) In recent years, new types of speed-measuring sensors have appeared (smartphones, Bluetooth sensors,radar data, etc...) and the large amounts of data made available by these sensors have motivated building large software systems which combine the different types of data with new speed models and estimators to produce more reliable estimates of the state of traffic. The goal of this thesis is to incorporate the raw data from ILDs into such a data fusion system. Two difficulties immediately arise from this objective: •ILD measurements are affected by systematic errors. To detect and discard these errors, a change-point detection algorithm has been implemented and applied to the correlation time-series of sensor pairs. •In order to be compatible with the speed model used in the data fusion system, the raw ILD data needs to be converted to speed. Models and physical relations from traffic theory are used to solve this subproblem; the use of a recursive Bayesian estimator, of the same family as the Kalman filter, is motivated (but not implemented) as a potentially better solution to this sub-problem.