Person2Drive
收藏资源简介:
Person2Drive是一个基于CARLA模拟器的数据集与基准测试平台,专注于个性化端到端自动驾驶研究。该数据集在闭环模拟环境中收集了人类驾驶记录,并按驾驶员级别进行组织,旨在支持驾驶风格建模、驾驶员个性化、路线级泛化以及端到端驾驶模型的闭环评估研究。与主要记录通用驾驶行为的传统数据集不同,Person2Drive提供了多名匿名人类驾驶员在共享或可比模拟设置下的多段驾驶记录。完整发布版本包含50名匿名驾驶员、4条不同的闭环CARLA环境驾驶路线。数据以驾驶员文件夹(drivers/driverXX/)形式组织,每条路线的驾驶记录存储为.tar.zst压缩档案。数据集内容涵盖个性化端到端驾驶研究所需的信息,包括自车状态、转向/油门/刹车等控制信号、自车轨迹、路线与导航信息、端到端驾驶模型常用的传感器数据,以及驾驶员级和地图级的组织元数据。该数据集既可用于开环行为分析,也可用于闭环个性化驾驶评估。典型应用场景包括驾驶员级个性化适配、路线级泛化能力测试、轨迹预测质量评估、驾驶风格相似性度量以及闭环驾驶性能评测。
Person2Drive is a dataset and benchmark platform based on the CARLA simulator, focusing on personalized end-to-end autonomous driving research. This dataset collects human driving recordings in closed-loop simulation environments and organizes them by individual driver, aiming to support research on driving style modeling, driver personalization, route-level generalization, and closed-loop evaluation of end-to-end driving models. In contrast to traditional datasets that primarily record general driving behaviors, Person2Drive provides multiple driving segments from numerous anonymous human drivers under shared or comparable simulation settings. The full released version includes 50 anonymous drivers and 4 distinct driving routes in the closed-loop CARLA environment. The data is organized in the form of driver folders (drivers/driverXX/), where driving recordings for each route are stored as .tar.zst compressed archives. The dataset covers all necessary information for personalized end-to-end driving research, including ego vehicle state, control signals such as steering, throttle and brake, ego vehicle trajectory, route and navigation information, sensor data commonly used in end-to-end driving models, as well as organizational metadata at the driver and map levels. This dataset can be used for both open-loop behavior analysis and closed-loop personalized driving evaluation. Typical application scenarios include driver-level personalization adaptation, route-level generalization capability testing, trajectory prediction quality evaluation, driving style similarity measurement, and closed-loop driving performance evaluation.
数据集概述
Person2Drive 是一个基于 CARLA 模拟器的闭环个性化端到端自动驾驶基准数据集,对应 ECCV 2026 论文《Driving like yourself: A Benchmark for Closed-Loop Personalized End-to-End Autonomous Driving》。
核心目标
- 支持人类驾驶风格建模、驾驶员级个性化、路线级泛化以及端到端驾驶模型的闭环评估。
- 区别于仅记录通用驾驶行为的数据集,Person2Drive 在共享或可比的模拟设置下提供多位匿名人类驾驶员的驾驶记录。
数据集统计
- 驾驶员数量:50 位匿名人类驾驶员。
- 驾驶路线:在闭环 CARLA 环境中包含 4 条驾驶路线。
- 数据结构:
- 每位驾驶员对应一个
driverXX/文件夹。 - 路线级数据以
.tar.zst压缩包形式存储。 - 包含驾驶员级和地图级的元数据文件。
- 每位驾驶员对应一个
数据集内容
提供用于个性化端到端驾驶研究所需的数据,包括:
- 自我车辆状态(ego-vehicle states)
- 控制信号(转向、油门、刹车)
- 自我轨迹(ego trajectories)
- 路线与导航信息
- 端到端驾驶模型使用的部分传感器数据
- 驾驶员级与路线级组织的元数据
数据集可用于开环行为分析和闭环个性化驾驶评估。
仓库结构
Person2Drive/ ├── drivers/ │ ├── driver01/ │ │ ├── Town04_drive_1.tar.zst │ │ ├── ... │ │ ├── Town04_drive_8.tar.zst │ │ ├── Town05_drive_1.tar.zst │ │ ├── ... │ │ ├── Town05_drive_8.tar.zst │ │ └── b2d_infos_train.pkl │ ├── driver02/ │ │ ├── Town04_drive_1.tar.zst │ │ ├── ... │ │ ├── Town05_drive_8.tar.zst │ │ └── b2d_infos_train.pkl │ └── ... ├── b2d_map_infos.pkl ├── RELEASE_MANIFEST.md ├── DATASET_DETAILS.md └── README.md
b2d_infos_train.pkl:存储对应驾驶员数据的元数据。b2d_map_infos.pkl:包含基准测试和评估流程所需的地图级元数据。- 详细数据组织和字段描述见
DATASET_DETAILS.md。
下载与解压
-
克隆仓库(含 Git LFS): bash git lfs install git clone https://huggingface.co/datasets/dongxr7/Person2Drive
-
Linux 解压: bash tar -I zstd -xvf Town04_drive_1.tar.zst
-
分步解压: bash zstd -d Town04_drive_1.tar.zst tar -xvf Town04_drive_1.tar
-
Windows 解压:可使用支持 Zstandard 压缩的归档工具,如 7-Zip、PeaZip、Bandizip。
基准测试用途
用于评估端到端驾驶模型在保持安全闭环性能的同时,能否适应个体人类驾驶风格。典型评估设置包括:
- 驾驶员级个性化
- 路线级泛化
- 轨迹预测质量
- 驾驶风格相似性
- 闭环驾驶性能
基准测试脚本、评估协议和使用说明将后续发布。
隐私与匿名化
驾驶员身份匿名化为 driverXX,公开版本不使用真实驾驶员名称。
许可协议
Apache-2.0
联系方式
如有疑问,请联系 Xiaoru Dong:xrdong@cs.hku.hk




