Statistical signal processing algorithms for estimation of vehicle trajectories from magneto-inductive traffic sensors.

摘要

Transportation agencies have invested in an extensive network of magneto-inductive sensors for vehicle detection and speed estimation. These vehicle detector sensors are installed at the vast majority of signalized intersections and are also used on highways for traffic monitoring. This dissertation presents several ways to enhance the data collected from these sensors to collect better vehicle trajectory information including algorithms to estimate the travel time and acceleration of vehicles.;A magneto-inductive vehicle detector produces a time varying waveform corresponding to the change in the sensor's inductance due to a passing vehicle. This waveform is referred to as the vehicle's signature. Using a statistical model for these signatures, this dissertation develops algorithms to estimate the speed and acceleration of passing vehicles. Acceleration estimates are important to yellow-interval, intersection safety, and vehicle emissions studies. The acceleration estimates are used to create position indexed signatures. By matching vehicle signatures from two geographically separated locations, travel time estimates can be generated. The vehicle accelerations and travel times allow for better characterization of vehicle trajectories.;The travel time estimation algorithm is evaluated by comparing travel times estimates to the ground truth travel times from 10,000 vehicles captured by video. The travel time estimation algorithm is found to correctly identify approximately 64.5% of the passing vehicles. Using 200 acceleration estimates from four GPS probe vehicles, the root mean squared error (RMSE) of the acceleration estimates is found to be 0.02g for inductive loop speed traps and 0.10g for Microloop speed traps.;A new framework is also developed for non-parametric comparison of travel time distributions. Methods that are currently in use focus on the comparison of travel time means or possibly standard deviations for studies of travel time reliability. The new framework is based on the Kullback-Leibler (KL) divergence between the travel time distributions. Even though the KL-divergence is not strictly a distance metric, it is shown to reliably characterize the similarity of two distributions and to provide conservative estimates for the sample size required for probe vehicle travel time studies.