This research applies computer vision techniques to extract road user trajectories from video data, from which surrogate safety measures (SSMs) are derived. Candidate extreme events, representing potential crash precursors, are identified through a combination of rule-based thresholds and unsupervised machine learning methods tailored to rare event detection, with the statistical modeling of extremes conducted separately using Extreme Value Theory (EVT). The aim is to improve the accuracy of crash estimation through this combined approach by constructing a Bayesian Hierarchical Model (BHM) that will take into account multi-site characteristics, providing a larger data pool during the modelling process. At the end, evidence based and infrastructure-oriented solutions are proposed alongside a decision support tool to perform traffic conflict analysis based on the improved approach.