Analisi della sicurezza stradale basata sull'approccio dei conflitti di traffico

Traditionally, road safety analysis relies on the use of crash data. However, several issues may affect these data: lack of availability, lack of spatial and/or temporal precision, under-reporting, misclassification; moreover, crashes are relatively rare events, so data must be collected for several years and/or in several different locations to obtain enough data. An alternative approach consists of analyzing traffic conflicts, which can be defined, intuitively, as “near-crashes”. Despite an ever-growing interest in traffic conflicts in transportation research, there are still several open questions that this research aims to answer, for both long-term and real-time road safety applications. 1) How to model the probabilistic relationship between traffic conflicts and crashes? 2) Is it possible to predict crashes in real-time with a conflict-based approach? The first part of this dissertation aims to provide further insight on the probabilistic relationship between traffic conflicts and crashes, by applying univariate and bivariate extreme value theory, for long-term road safety analysis (i.e., prediction of annual crash rates in selected infrastructures). A real-world case study using data collected with radar sensors at several cross-sections of an Italian highway illustrates an operational-oriented procedure which aims to identify surrogate measures of safety threshold separating normal traffic interactions and conflicts, and to translate raw conflict data into annual crash rates. The results of the analysis are validated using historical crash rates: in about 90% of the highway cross-sections analyzed, the observed average annual number of crashes falls within the 95% confidence interval of the predicted number of crashes. The second part focuses on developing a conflict-based approach for real-time road safety analysis. Using data from the previous case study, conflicts are identified using the extreme value theory thresholds. Then machine learning-based methods are used to link real-time traffic variables to traffic conflicts that occur in the immediate future. The objective is using traffic conflicts to train an AI predictor able to forecast whether there will be an increased risk of crashes in the immediate future, based on the current traffic variable collected at several highway cross-sections. The results are compared to those of a traditional real-time crash prediction model (estimated with crash data), showing a significant improvement in terms of accuracy, sensitivity, and specificity. The application of conflict-based approaches has the potential to provide a positive broader impact on road safety. In particular, the possibility to avoid the use of crash data in practical applications can (i) allow to apply statistical methods to new scenarios in which crash data are unavailable or unreliable (e.g., rural roads, third-world countries, new infrastructures); (ii) proactively analyze safety, avoiding the ethical dilemma of crash-based approaches, in which injuries and fatalities are needed in order to correctly identify unsafe behaviors and locations; (iii) provide faster road-safety evaluations, since traffic conflicts are more frequent than crashes; for the same reason, provide more flexible and resilient road safety models, which can be based on more recent data. This dissertation combined the use of existing tools (traffic conflict theory, extreme value theory, machine learning) in order to answer in an innovative way to the research question expressed above. In particular, the application of extreme value theory to such a large-scale scenario is unique in the literature, whereas the combination of extreme value theory with real-time road safety is the most innovative contribution of the present research. The procedures described in the dissertation were developed with practical operations in mind and can be realistically applied in the real world.