Joint parameter and state estimation algorithms for real - time traffic monitoring

A common approach to traffic monitoring is to combine a macroscopic traffic flow model with traffic sensor data in a process called state estimation, data fusion, or data assimilation. The main challenge of traffic state estimation is the integration of various types of sensor data (e.g. speed, flow, travel time, etc.) into the flow model due to the nonlinearities of the traffic model. When parameters are also estimated, the nonlinearity of the estimation problem increases, motivating the development of advanced estimation algorithms to handle the additional nonlinearity. To improve performance of traffic state estimation algorithms this work investigates the problem of simultaneously or jointly estimating both the traffic state and the parameters of the traffic model. It uses two new traffic parameter and state estimation algorithms based on multiple model particle filtering, and multiple model particle smoothing. Because incidents on freeways can be modeled through parameter changes in the traffic model, this work applies both algorithms to the problem of incident detection. Findings The main contributions of this work are as follows. First, it is shown that the problem of detecting incidents can be posed as a traffic flow model parameter estimation problem. A family of parameters known as regime variables are used to model the location and severity of the incident. The regime variable is used to indicate the number of open (unobstructed) lanes everywhere along the roadway. A reduction of the number of open lanes, as indicated by the regime variable, indicates that an incident has occurred. Second, two new algorithms are developed to jointly estimate the traffic state, and the regime variable parameters. Both algorithms are able to estimate the traffic conditions based on speed data obtained from GPS equipped vehicles, and/or flow and density data from inductive loop detectors. Both algorithms are online, sequential algorithms, and the primary distinction between the filter and the smoother is that the smoother runs with a slight lag compared to the filter. In other words, the smoother uses all data up to the current time to estimate traffic conditions in the past (e.g. two minutes ago), while the filter uses the same data to estimate the traffic at the current time. Third, we test both algorithms in two numerical environments. In the first set of experiments, synthetic data is generated from a similar macroscopic model to the one assumed in both estimation algorithms. These experiments represent an upper bound on …