Assisted and automated road transport requires extensive testing procedures to verify and validate safety performance under varying conditions. Virtual test drives with computer simulation models are being used by the automotive industry, because they can decrease costs in the development cycle and can comprise high numbers of scenarios with combinations of varying factors. Finding the key driving scenarios for the simulations is currently one of the main challenges. This paper presents a new data analysis method to generate such “benchmark” scenarios from historical crash data. It applies the k-medoids clustering method to partition the crashes into distinct groups. Then, association rule mining is used to define further parameters for each cluster, which constitute the key scenarios for simulation experiments. The method is demonstrated by analyzing 1326 junction accidents from an English in-depth database, which resulted in nine clusters for T-junctions, and six clusters for crossroads. Association rules revealed common crash characteristics, which were the basis for the scenario description. In total, 34 scenarios were identified, which constitute the core population of crash situations at UK junctions.