Statistical and Behavioral Modeling of Driver Behavior on Signalized Intersection Approaches

The onset of a yellow indication is typically associated with the risk of vehicle crashes resulting from dilemma-zone and red-light-running problems. Such risk of vehicle crashes is greater for high-speed signalized intersection approaches. The research presented in this dissertation develops statistical as well as behavioral frameworks for modeling driver behavior while approaching high-speed signalized intersection approaches at the onset of a yellow indication. The analysis in this dissertation utilizes two sources of data. The main source is a new dataset that was collected as part of this research effort during the summer of 2008. This experiment includes two instructed speeds; 72.4 km/h (45 mph) with 1727 approaching trials (687 running and 1040 stopping), and 88.5 km/h (55 mph) with 1727 approaching trials (625 running and 1102 stopping). The complementary source is an existing dataset that was collected earlier in the spring of 2005 on the Virginia Smart Road facility. This dataset includes a total of 1186 yellow approaching trials (441 running and 745 stopping).

The adopted analysis approach comprises four major parts that fulfill the objectives of this dissertation. The first part is concerned with the characterization of different driver behavioral attributes, including driver yellow/red light running behavior, driver stop-run decisions, driver perception-reaction times (PRT), and driver deceleration levels. The characterization of these attributes involves analysis of variance (ANOVA) and frequency distribution analyses, as well as the calibration of statistical models. The second part of the dissertation introduces a novel approach for computing the clearance interval duration that explicitly accounts for the reliability of the design (probability that drivers do not encounter a dilemma zone). Lookup tables are developed to assist practitioners in the design of yellow timings that reflects the stochastic nature of driver PRT and deceleration levels. An extension of the proposed approach is presented that can be integrated with the IntelliDriveSM initiative.

Furthermore, the third part of the dissertation develops an agent-based Bayesian statistics approach to capture the stochastic nature of the driver stop-run decision. The Bayesian model parameters are calibrated using the Markov Chain Monte Carlo (MCMC) slice procedure implemented within the MATLAB® software. In addition, two procedures for the Bayesian model application are illustrated; namely Cascaded regression and Cholesky decomposition. Both procedures are demonstrated to produce replications that are consistent with the Bayesian model realizations, and capture the parameter correlations without the need to store the set of parameter realizations. The proposed Bayesian approach is ideal for modeling multi-agent systems in which each agent has its own unique set of parameters. Finally, the fourth part of the dissertation introduces and validates a state-of-the-art behavioral modeling framework that can be used as a tool to simulate driver behavior after the onset of a yellow indication until he/she reaches the intersection stop line. The behavioral model is able to track dilemma zone drivers and update the information available to them every time step until they reach a final decision. It is anticipated that this behavioral model will be implemented in microscopic traffic simulation software to enhance the modeling of driver behavior as they approach signalized intersections.