Freeway Incident Likelihood Prediction models: Development and application to traffic management systems

Pen-Chi Liu, Purdue University

Abstract

This thesis describes the concept of Freeway Incident Likelihood Prediction (ILP) and its application to the freeway incident clearance problem. ILP models take their input data from traffic and weather measurements and generate real-time incident likelihood estimates for various freeway sections. Unlike incident detection algorithms, ILP models provide incident likelihood predictions before any incident takes place. Data for ILP models are collected from three different sources. The development of ILP models is explained. Two binary logit ILP models are developed. One model describes overheating vehicle incidents, and the other model describes vehicle crashes. The ILP models are applied to traffic incident management systems. A single server, either patrolling the freeway or pre-positioned according to information provided by ILP models, is used as a probe vehicle and as a responding unit. A framework including mathematical models, in the case of perfect incident detection, and a heuristic, in the case of imperfect incident detection, is developed. The mathematical models are used to specify the location of a stationary server by minimizing total incident waiting time over the freeway network. The heuristic rules dictate how a server patrols the whole freeway network in order to minimize response time to potential incidents. To evaluate the ILP models in the case of imperfect incident detection, a series of simulations based on two scenarios (with and without the ILPs) and two sets of ILP values (two sets of incident scenarios), is performed. Results show that the reduction of incident waiting time is significant when ILPs are used. Two topics for future research are proposed. First, a conceptual framework is established for both the two-server and the multiple server freeway incident clearance problem. Second, a data fusion model, which combines the incident likelihood predictions with a Bayesian incident detection algorithm, is developed and discussed.

Degree

Ph.D.

Advisors

Madanat, Purdue University.

Subject Area

Civil engineering|Transportation

COinS