Object detection plays an important role in intelligent transportation systems, especially due to the rise of autonomous vehicles and smart traffic management. Although object detection techniques in traffic scenes have been studied for decades, it is still challenging in addressing long-tail problems under complex and extreme conditions. In this paper, we first introduce long-tail problems in object detection for intelligent transportation systems and investigate the usage of cross-domain approaches, which are widely leveraged in dealing with these problems. Then we describe a general parallel theory based framework for enhancing the effectiveness of computer vision tasks in complex scenes. Combining ACP methodology with cross-domain object detection, we propose a novel architecture to perform long-tail problems concerned object detection by simulating the real transportation scenes with rare situations. Artificial system is constructed to simulate and represent longtail situations, making it possible to collect and annotate specific and available diversified datasets. Computational experiments are built on Graph Convolutional Network (GCN) based cross-domain detection method to learn and evaluate vision models from limited real samples, then improve the performance of detection tasks in both artificial and real systems. Parallel execution can be used to optimize the whole architecture and supports the proposed work to be a routine practice of visual computing for intelligent transportation systems.