LoT-nuScenes: A Virtual Long-Tail Scenario Dataset for Parallel Vision and Parallel Vehicles
Yang Mi, Yunjie Ji, Keqiu Wang, Yadong Wang, Tianyu Shen, Kunfeng Wang · 2024
Autonomous driving systems rely on massive amounts of high-quality data, and the long-tail problem is a major challenge for their development. The long-tail problem involves a large number of rare, special, or complex driving scenarios, which are difficult to be comprehensively covered by traditional data collection methods. But the long-tail scenarios can be reduced to simulation software to create diverse and controllable driving environments. In this paper, we construct accident scenarios under different states in the Carla simulator, which includes six types of motor vehicle accidents, one type of pedestrian accidents and combines three extreme weathers, three time periods and five types of locations. At the same time, we collect accident events in the format of the nuScenes dataset, equipped with multi-sensors and 360° field-of-view. This dataset not only fills the gap of accident scenario data and achieves long-tailed normalized distribution, but also provides resources for target detection and tracking task testing and validation of autonomous driving systems. In the future, we plan to further expand the diversity of events and explore more application scenarios. Our dataset will be available at https://github.com/BUCT-IUSRC/Dataset_LoT-nuScene.