基于AVstack和CARLA的合成数据集
收藏资源简介:
该数据集是一个基于AVstack框架和CARLA模拟器生成的大规模多传感器、多智能体合成数据集,由杜克大学研究团队创建。数据集内容涵盖地面车辆、空中平台和基础设施部署的同步感知数据,包括RGB图像、深度图、语义分割、激光雷达和雷达等多种模态,总规模可达TB级别,例如一个10分钟的多传感器实例即包含20万图像和310万标注对象。数据集通过可配置的生成流程创建,支持对智能体数量、传感器配置、天气条件和交通场景的灵活控制,并利用几何可见性算法进行精准标注。该数据集旨在支持自动驾驶领域中的特定应用感知训练、跨域评估以及多智能体协同算法研究,为解决真实数据稀缺性、视角多样性和协同感知复杂性等挑战提供可控实验平台。
This large-scale multi-sensor, multi-agent synthetic dataset is generated based on the AVstack framework and CARLA simulator, and developed by the research team from Duke University. It covers synchronized perception data collected from ground vehicles, aerial platforms and infrastructure deployments, including multiple modalities such as RGB images, depth maps, semantic segmentation outputs, LiDAR and radar data. The total scale of the dataset can reach the terabyte level; for example, a 10-minute multi-sensor instance contains 200,000 images and 3.1 million annotated objects. The dataset is built through a configurable generation pipeline, which enables flexible control over the number of agents, sensor configurations, weather conditions and traffic scenarios. It leverages geometric visibility algorithms to generate precise annotations. This dataset is designed to support application-specific perception training, cross-domain evaluation and multi-agent collaborative algorithm research in the autonomous driving field, providing a controllable experimental platform to address challenges including real-world data scarcity, viewpoint diversity and the complexity of collaborative perception.

- 1Scaling Datasets for Multi-Sensor, Multi-Agent, and Multi-Domain Learning in Autonomous Systems杜克大学·电气与计算机工程系 · 2026年



