LSTV-V2V: A Large-Scale Traffic Virtual Dataset for Vehicle-to-Vehicle Cooperative Perception

Cooperative perception system is essential to address occlusion and small object perception issues in autonomous driving. However, autonomous vehicle (AV) perception methods based on deep learning rely on large-scale and accurately annotated traffic datasets. Compared to real datasets, which are difficult to collect and annotate and limited in quantity and diversity of scenes, the virtual dataset synthesis method is a convenient and efficient approach. To improve annotation accuracy while speeding up modeling, we propose a pipeline for constructing artificial traffic scenes and generating virtual datasets based on autonomous driving simulation software from a parallel vision perspective. Furthermore, to facilitate the development of cooperative perception, we propose a novel efficient traffic scene dataset for V2V cooperative perception named Large-Scale Traffic Virtual Dataset (LSTV-V2V). Our dataset contains 47 traffic scenarios, 8380 frames, and 191864 annotated 3D vehicle bounding boxes. Experimental results demonstrate the effectiveness of our virtual dataset in cooperative perception tasks.