Innovative non-crash-based safety estimation: An extreme value theory approach
Abstract
Crash-based safety analysis is hampered by several shortcomings, such as randomness and rarity of crash occurrences, lack of timeliness, and inconsistency in crash reporting. Safety analysis based on observable traffic characteristics more frequent than crashes is one promising alternative. A traditional approach to alternative safety analysis relies on the assumption of constant risk across locations. In addition, the current practice of collecting surrogate data often suffers from the inherent subjectivity of the humans involved in the task. In this research, we propose a novel application of the extreme value theory to a non-crash-based safety estimation that no longer relies on the assumption of constant risk. We evaluate the proposed method by applying it to right-angle collisions at signalized intersections. The feasibility of facilitating the measurement of traffic characteristics with digital video and image processing technology is also examined. Eight-hour traffic movements at selected intersections were recorded using a mobile traffic laboratory. The risk of right-angle collisions was estimated using so-called post-encroachment times (PET). Evaluated video image processing techniques were not sufficiently accurate for the purpose of our research. Therefore, post-processing of digitized video clips using a manual method was selected. The Poisson and negative binomial regression analyses of short PETS and observed crash counts indicate a significant relationship between these two. A series of negated PET observations was discretized into fixed time intervals and the maximum values in each interval were treated as extremes. This approach elegantly handles the dependence of extremes in comparison to an alternative approach that defines threshold excesses as extremes. A distribution of extreme values was modeled with a generalization of the generalized extreme value distribution as the non-stationary r largest order statistic model. Based on the premise that PETS being zero or less define a collision situation, safety levels were determined from the model in terms of crash frequency and return level estimates. Evaluation of the safety estimates against historical crash counts indicates a promising relationship between these two. However, the proposed method still yields large-variance estimates due to an insufficient observation period. A semi-empirical simulation experiment revealed that a few weeks of PET observation were needed to obtain crash frequency estimates with confidence intervals comparable to those being obtained from three-year observed crash counts. The proposed method can be applied to other types of locations and collisions as well.
Degree
Ph.D.
Advisors
Tarko, Purdue University.
Subject Area
Civil engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.