Training a deep learning model for identifying dangerous vehicles requires a large amount of labeled accident data. However, it is difficult to collect a sufficient amount of accident data in the real world. To address this challenge, we introduce a driving-simulator-based data generator that can arbitrarily produce a wide variety of accident scenarios. Furthermore, in order to reduce the gap between synthetic data and real data, we propose a new domain adaptation algorithm that refines both features and labels. We conduct extensive real-data experiments to demonstrate that our dangerous vehicle classifier can reduce the missed detection rate by at least 23.2%, as compared to those trained only with scarce real data, for an interested scenario in which time-to-collision is 1.6-1.8 s. We also find that our algorithm can identify various accident-related factors (such as wheel angles, vehicle orientations, and velocities of nearby vehicles) to enable high prediction accuracy for complex accident scenes.