DeepAccident
收藏资源简介:
DeepAccident是由香港大学等机构创建的首个支持V2X自动驾驶的大规模数据集,专注于真实世界中频繁发生的多样化事故场景。该数据集通过现实模拟器生成,包含57K标注帧和285K标注样本,远超其他大型数据集如nuScenes。DeepAccident不仅支持多种感知任务,还引入了端到端运动和事故预测新任务,旨在直接评估不同自动驾驶算法的事故预测能力。数据集中的每个场景都设计有四辆车和一个基础设施来记录数据,提供多视角的事故场景数据,支持V2X感知和预测研究。此外,数据集还包含详细的标注信息,如事故车辆ID和未来碰撞轨迹,以及正常无碰撞场景,以增强运动多样性。DeepAccident的应用领域主要集中在提升自动驾驶的安全性,解决事故预测和运动预测的关键问题。
DeepAccident is the first large-scale V2X-enabled autonomous driving dataset developed by institutions including the University of Hong Kong, focusing on diverse real-world accident scenarios that occur frequently. Generated using real-world simulators, it contains 57K annotated frames and 285K annotated samples, a scale exceeding that of other prominent datasets such as nuScenes. Beyond supporting multiple perception tasks, DeepAccident introduces novel end-to-end motion and accident prediction tasks, designed to directly evaluate the accident prediction performance of diverse autonomous driving algorithms. Each scenario in the dataset is configured with four vehicles and one infrastructure node for data acquisition, providing multi-view accident scene data to support V2X-oriented perception and prediction research. Furthermore, the dataset includes detailed annotation information such as accident vehicle IDs and future collision trajectories, as well as normal collision-free scenarios, to enhance the diversity of motion patterns. The primary application scope of DeepAccident centers on improving autonomous driving safety, addressing core challenges in accident prediction and motion prediction.

- 1DeepAccident: A Motion and Accident Prediction Benchmark for V2X Autonomous Driving香港大学 · 2023年



