Although visual perception algorithms have made significant progress in most normal scenes, it is still challenging for autonomous driving systems to accurately perceive long-tail scenes that occur less frequently, which can lead to serious traffic safety issues. However, existing open-source datasets do not systematically collect sufficient long-tail scenes. To fill this gap, we propose a pipeline for designing large-scale, diverse long-tail traffic scenes and generating virtual datasets based on the parallel vision approach. A virtual dataset named Vir-LTTS (virtual long-tail traffic scenes) is built, comprising various scenes such as extreme weather conditions, adverse lighting conditions, traffic accidents, unique forms of traffic objects, and blurry images caused by camera defects. We investigate the potential of training models using the Vir-LTTS dataset in long-tail traffic scenes. Experimental results show that pre-training with Vir-LTTS significantly improves the performance of visual models in long-tail traffic scenes.