| Rear-end collisions are frequent yet preventable crashes in the United States. Collision warning systems have the potential to prevent crashes and mitigate crash severity. However, their success depends on the algorithms used to trigger the warnings, and computational models of rear-end collision avoidance behaviors are critical for accurate calibration of warning algorithms. This dissertation addresses gaps in driver modeling research by developing models based on three different psychological perspectives. A driving simulator experiment was used to provide data of imminent collision with a stopped or decelerating lead vehicle in the presence or absence of a warning system. Three models were developed---two were based on the information-processing approach (static models) with different parameter associations, and one was based on the ecological approach and concepts of direct perception (dynamic model). The static model with independent stages considered parameters (reaction time, jerk, and deceleration) as independent; the static model with dependent stages used copula functions to construct trivariate distributions. Associations between variables suggests that assumptions of independence are invalid. Counterfactual analysis was used to perform benefits estimation, and results show that predicted benefits of the copula-based models and those of the static models with independent stages differ by 11%. The dynamic model used perceptual variables---visual angle and expansion rate---to model onset of braking (reaction time) and deceleration profiles. This dynamic model assumes that perception occurs in the light, i.e., ecological structures relevant to collision avoidance, such as visual looming of the lead vehicle, can be used to directly specify collision-avoidance actions. The dynamic models may represent drivers' braking responses more precisely, however, traditional statistical approaches cannot be used for the parameterization of such complex models. The Approximate Bayesian Computation technique was used to parameterize these models in this dissertation. Model parameters estimated with this technique indicate that combinations of perceptual variables generate dynamic collision-imminent deceleration profiles similar to those observed in the empirical data. Between-driver variances in deceleration were captured in the perceptual variable parameters. Taken together, the different models improve our understanding of the mechanisms governing drivers' rear-end collision avoidance and provide a basis for future behavioral modeling. |
